Quantum Computer Science

— course lecture notes HT 2008 (with updates 2010)∗—

Bob Coecke
Oxford University Computing Laboratory

March 13, 2010

Contents 6.2 The Deutch-Jozsa algorithm . . . . . . . . . 27

6.3 Grover’s search algorithm . . . . . . . . . . . 28
1 Historical and physical context 1
6.4 Shor’s factoring algorithm . . . . . . . . . . 29
1.1 The birth of quantum mechanics . . . . . . . 1
6.4.1 Period finding . . . . . . . . . . . . . 29
1.2 The status of quantum mechanics . . . . . . . 2
6.4.2 Factoring and code-breaking . . . . . 30
1.3 The birth of quantum informatics . . . . . . . 3
6.5 Quantum key distribution . . . . . . . . . . . 31
1.4 The status of quantum informatics . . . . . . 3
7 Mixed states 31
2 Qubits vs. bits 4
2.1 Acting on qubits . . . . . . . . . . . . . . . . 4 8 Quantum logic and Gleason’s theorem 33
2.2 Describing a qubit with complex numbers . . 5
9 Mixed operations 35
2.3 Describing two qubits . . . . . . . . . . . . . 7
10 More on tensors 37
3 von Neumann’s pure state formalism 7
3.1 Hilbert space . . . . . . . . . . . . . . . . . 7 11 Semantics for quantum informatics 39
3.2 Matrices . . . . . . . . . . . . . . . . . . . . 9 11.1 Symmetric monoidal categories . . . . . . . 39
3.3 Tensor structure . . . . . . . . . . . . . . . . 12 11.2 Naturality implies basis-independence . . . . 40
3.4 Dirac notation . . . . . . . . . . . . . . . . . 15 11.3 †-compact categories . . . . . . . . . . . . . 40
11.4 Classical uncertainty and open systems . . . . 41
4 Protocols from entanglement 17
4.1 Bell-basis and Bell-matrices . . . . . . . . . 18
4.2 Teleportation and entanglement swapping . . 18
1 Historical and physical context
5 The structure of entanglement 20
1.1 The birth of quantum mechanics
5.1 Map-state duality and compositionality . . . . 20
5.2 The logic of bipartite entanglement . . . . . . 22 There is no agreed clear date attached to the ‘birth of quan-
tum theory’, as is for example the case of Sir Isaac Newton’s
5.3 Quantifying entanglement . . . . . . . . . . 24
1686 theory of (classical) mechanics and Albert Einstein’s
5.4 Trace from Bell-states . . . . . . . . . . . . . 25 1905/1917 theories of special/general relativity. Quantum
Theory came about rather by several discoveries, insights and
6 Algorithms and gates 26 developments which ultimately lead to John von Neumann’s
6.1 Special gates . . . . . . . . . . . . . . . . . 26 Hilbert space quantum mechanical formalism which is cur-
∗
rently still in use. Some of the most important of these dis-
Please report any error, typo or omission found in these
notes both to the lectures and the class problems tutor, Ja-
coveries, insights and developments are (e.g. [1]):
cob Biamonte, via email: [email protected] and
[email protected]. • In 1900 Max Planck noted that the physically observed
1
 frequency dependence of so-called black body radiation quantum mechanics and applies to any pair of non-compatible
requires energy to be quantized and come in the form of quantities.
finite chunks.

• In 1905 Albert Einstein explained the photoelectric ef- (non-local) Entanglement. There are states, e.g. EPR-
fect by postulating that light is also quantized and comes states and Bell-states, which yield statistical correlations be-
in packets which he called photons. tween systems separated by a large physical distance, and
these correlations ‘must travel faster than the speed of light’,
• In 1913 Niels Bohr explained the spectral lines of the in fact, they are instantaneous. Surprisingly however, these
hydrogen atom emission spetra by a new model for the correlations cannot be used to send information faster than
atom which, a priori, involved discrete (i.e. quantized) light, and hence special relativity is not violated. These cor-
energy levels. relations have been experimentally observed many times dur-
ing the past 30 years, and are typically referred to as quantum
• In 1924 Louis de Broglie suggested that, dually to the
entanglement.
discrete ‘particle’ nature of light, matter should also be
thought of as having ‘wave’-like behavior.
Intrinsic probabilities. When we perform a quantum mea-
• Around 1925 Werner Heisenberg constructed matrix
surement, i.e. verify some physical property of a superposi-
mechanics and Erwin Schrödinger constructed wave me-
tion state, then the outcome will occur in a probabilistic man-
chanics including the Schrödinger equation.
ner. E.g. if we verify whether the cat is either dead or alive, or,
• John von Neumann developed the mathematically rig- whether the bit is either 0 or 1, then if we are in a superposi-
orous Hilbert space formalism for quantum mechanics tion state — both outcomes can occur with some probability.
which is now still in use — first published in 1932 [2]. There are mathematical theorems which state that by assign-
Also, Paul Dirac’s bra-ket notation, which appeared in ing additional statistical variables to the quantum system we
his 1930 book, remains in popular use today [5]. cannot get rid of these probabilities [6, 7].

Before we turn to a formal development of quantum me-

1.2 The status of quantum mechanics
chanics in the next chapter, we now informally discuss some
crucial structural features of quantum mechanics. Once we 1. Probably the most successful physical theory ever in
have a mathematical formalism at our disposal, it will become terms of predictions e.g. quantum electrodynamics pre-
much clearer how they inter-relate. dicts correct results up to ±10 digits!

Superposition. When a system (e.g. a particle) admits 2. It has many important applications such as:
some distinct properties e.g. ‘being either here or there’, ‘be- • The description of individual particles such as
ing either 1, 2, 3, ..., or 111 years old’, ‘being either dead molecules, atoms, photons, electrons, protons and
or alive’, ‘being either 0 or 1’, (etc.) then a superposition neutrons, and hence all the obvious applications
state stands for a situation — where a kind of combination in many fields ranging from chemistry, nuclear
of these alternatives applies with is different than a probabil- physics, and in the future possibly high-energy
ity distribution over the states. In the case of the so-called physics.
Schrödinger’s cat paradox the cat is neither dead or alive,
but somewhere in between, and in computer science terms a • New technologies such as the laser and the elec-
quantum bit (qubit) can take both the values 0 and 1 concur- tron microscope, and in particular, the transistor
rently. In a sense, while the utterance ‘quantum’ indicates ‘a as a replacement for valves, enabling the scale at
passage from the continuous to the discrete’, from an infor- which micro-electronics (including computer hard-
matic perspective it is rather ‘a passage from the discrete to ware) can currently be built.
the continuous’. But unfortunately, that continuous space is • Important medical tools such as magnetic reso-
not evidently accessible, due to the nature of quantum mea- nance imaging techniques.
surement. • Actual quantum informatic devices such as quan-
tum communication systems and quantum proces-
Uncertainty. A physical quantum system cannot admit sors?
both a sharp (= not in superposition) position and sharp mo-
mentum at the same time, a principle known as Heisenberg’s 3. Big conceptual and philosophical questions, initially
uncertainty principle. An analogous principle arises often in raised by Einstein, remained unanswered. The biggest
2
 of these is the so-called measurement problem: it is con- The first algorithm of that kind was the Deutch-Jozsa al-
ceptually not clear at all what causes ‘a measurement gorithm which exploits quantum parallelism—computing a
to take place’. But most of the physics community has function for several values at once, but uses this to solve a
moved forward — ignoring most conceptual problems problem of little practical interest. What is often considered
and accepting quantum mechanics as a cook-book which to be the start of quantum algorithmics was Peter Shor’s 1994
provides ‘weird’ recipes on how to handle and interpret factoring algorithm [13] which provided exponential speed-
matter, and more recently information. up as compared to all known classical factoring methods. An-
other well known quantum algorithm is Lov Grover’s 1995
4. The formalism is still mathematically unsatisfactory for search algorithm [14] for unstructured data of size N which
many reasons: it contains redundancies such as global reduces the search-time from N to O(√N ).
phases, yields re-normalization problems in quantum
field theory, lacks high-levelness etc. The formalism But quantum informatics is not only about algorithmics.
in fact hasn’t changed since it’s creation by John von There are several intriguing quantum protocols which expose
Neumann, who actually denounced it three years after fascinating physical phenomena, some of which turn out to
creating it! This lead to so-called quantum logic — a have applications which are most likely to be the first real-
field of study launched by Birkhoff and von Neumann in life incarnations of a quantum informatic revolution. Among
1936 [4], but there are serious doubts that this has given these conceptually intriguing protocols are quantum telepor-
much insights either in quantum theory or in logic, and tation and entanglement swapping. At the practical side there
in no way did it have the capabilities to replace von Neu- are the many variants of quantum key-distribution, within the
mann’s quantum mechanical formalism. field of protocol security. This is a nice example of how quan-
tum informatics constitutes both a danger and provides the
corresponding solution to the security of communication pro-
tocols: An actual quantum computer runing Shor’s algorithm
1.3 The birth of quantum informatics
would provide a danger to many cryptographic protocols cur-
Most of the scientific activity on ‘pure quantum mechanics’ rently in use which typically rely on hardness of factoring. On
which took place in the second half of the 20th century was the other hand, quantum key-distribution provides a solution
either on its experimental confirmation, on its philosophi- to any such attack!
cal justification, or on generalizing/modifying its formalism.
The passage to quantum informatics can be seen as a mat-
ter of change of attitude towards the so-called ‘quantum- 1.4 The status of quantum informatics
weirdness’:
1. Many different experimental devices of a small number
It’s a feature, not a bug! of qubits (< 10) are operational but scalability is still a
major problem. This problem is due mainly to the de-
The first to mention quantum computing was Paul Benioff in coherence of quantum data due to interaction with the
1980 who studied how particular kinds of quantum evolutions environment, but (at least) theoretical solutions do exist.
could simulate classical Turing machines. Richard Feynman
on the other hand asked the dual question i.e. whether clas- 2. The search for new kinds of algorithms and applications
sical computers can simulate quantum evolution, and conjec- continues including resent efforts to use quantum pro-
tured that such a simulation came with an exponential slow- cessors to simulate chemical reactions at the quantum
down, while, in principle, quantum systems could simulate mechanical level!
themselves without this exponential slow-down by simply re-
lying on the natural quantum evolution they are already gov- 3. It is commonly accepted that information security will
erned by. Hence, the first advantage of considering a quan- likely be the first practicable application of quantum
tum mechanical system as a computational device had been informatics, and quantum communication devices are
exposed. The key to this speed-up is that quantum evolu- available from commercial companies (MagiQ and ID
tion physically computes a function for several inputs at the Quantique). An actual quantum key distribution proto-
same time, which are in superposition. But the nature of the col has taken place between a Swiss bank and Geneva
quantum measurement process doesn’t allow the state of the City Hall [21]. However, while the quantum component
quantum system to be read without actually altering it, and of the experiment worked perfectly, the authentication
converting this potential of intrinsic parallelism within quan- protocol failed to be secure due to flaws in the analysis
tum evolution into concrete examples of algorithmic speed- of its classical component (e.g. [22]). The irony here is
up of quantum computers as compared to classical computers that a true danger to classical security protocols is posed
turned out to be a highly non-trivial matter. by Shor’s algorithm, i.e. by quantum informatics, while
3
 |1i |e1 i
it is again quantum informatics which provides the solu- |ψi
tion.

4. At a very fundamental level questions still remain e.g.: |e0 i

|0i

• What are the true origins of a quantum algorithmic

• only admits special transformations which preserve the
computational speed-up?
angles (and hence opposites) on the sphere, and hence
• What are the limits of quantum computation? which are in particular reversible.
• What is a model for general quantum computing? • only admits ‘reading’ through so-called quantum mea-
surements M (|e0 i, |e1 i) which
New computational models are being proposed e.g. the
one-way model [19] which radically challenges all – only have two possible outcomes |e0 i and |e1 i,
known methods related to the circuit or gate array model. – change the initial state |ψi to either |e0 i or |e1 i,
5. There is no real high-level quantum computer science. hence one could say that a measurement M (|e0 i, |e1 i)
The current methods, from a programming perspec- does not tell us |ψi = α|e0 i + β|e1 i but destroys |ψi!
tive, are comparable with hacking with bits (but instead
with complex numbers rather than with bits). There is A metaphor: quantum measurement of Colors. Assume
also a convincing argument that current high-level com- the points of the sphere, i.e. the possible states of the system,
puter science structures for distributed computing, hy- correspond to colors. We can for example ask if it is blue
brid computing, embedded computing and tools for ver- or red (= 2 colors), but not if it is either blue, red or green
ification might actually be extremely useful for the the- (= 3 colors). Assume now the system is green, and the mea-
oretical side of the quantum computational endeavor, surement we perform asks if it is either blue or red. Then the
both at the level of the quantum mechanical formal- outcome will either be blue or red, meaning that the system
ism and those posed by the new quantum computational has indeed become blue or red respectively, but we will never
paradigms. It seems that there is a true need for quantum get to know that the initial color was actually green.
computer science in the British/European sense.
Here’s a more detailed look at quantum measurements:
6. The quantum computational activity has already pro-
vided some fresh insights into the domain of foundations |e1 i
of physics, providing new concepts and paradigms from
informatics.
|ψi

θ1
2 Qubits vs. bits θ0
2.1 Acting on qubits

A bit:

• admits two distinct values 0 and 1, |e0 i

• admits arbitrary transformations (can erase, copy etc. at The two transitions
ease).
Pe0 :: |ψi 7→ |e0 i Pe1 :: |ψi 7→ |e1 i
• is freely readable (hence, the state of the bit is left un- have respective chance prob(θ0 ) and prob(θ1 ) with
changed if you measure it),
prob(θ0 ) + prob(θ1 ) = 1
A qubit: since quantum theory dictates that for θ the angle on the
sphere between the initial state and a possible outcome state
• a sphere of values (which in some particular manner is (cf. the above picture) we have
‘spanned’ by two quantum states |0i and |1i or |e0 i and θ
|e1 i), prob(θ) = cos2 ,
2
4
and in particular do we have The state
 of  a qubit is described by a pair of complex num-
z1
1 bers = z1 |0i + z2 |1i up to a non-zero complex multi-
prob(0) = 1 prob(90o ) = prob(180o ) = 0 z2
2 ple, which means that for any z ∈ C0 (= C without zero C/0)
and in general the pairs
     
0 < prob(θ) < 1 for 0 < θ < 180o . z1 z1 z · z1
and z· :=
z2 z2 z · z2
Since there are impossible transitions (cf. prob(180o ) = 0),
we obtain two ‘partial constant maps’ on the sphere Q both define the same state. Typically one writes these pairs as

Pe0 : Q \ {| e1 i} → Q :: |ψi 7→ | e0 i. |ψi := z1 · |0i + z2 · |1i

Pe1 : Q \ {| e0 i} → Q :: |ψi 7→ | e1 i to make a connection with bits. Ignoring the global redun-
dancy of the non-zero complex number z, a qubit state is a
capturing the dynamics of measurement, which can be used as complex linear combination of two reference states |0i and
a dynamic resource when designing algorithms and protocols |1i.
— as we shall see further, for so-called degenerate measure-
ments these maps are not always constant. In fact, restricting When representing the complex numbers in the 2D com-
to states and measurements which are such that measurements plex plane, passing from cartesian to polar coordinates yields
behave deterministically and don’t change the state is equiv- the amplitude and phase of a complex number, respectively
alent to doing classical reversible computing! Hence: p y
r = x2 + y 2 , tan(θ) = ,
x
• bad: quantum measurements destroy most data
and conversely, the real and complex parts re-emerge as
• good: quantum measurements expose some data
x = r · cos(θ) , y = r · sin(θ) .
• good: quantum measurements act on data
Hence a complex number can also be written as
Conclusively, designing quantum algorithms and protocols
boils down to exploiting the enlarged state space by acting z = r · eiθ since eiθ := cos(θ) + i sin(θ) .
on quantum data either with:
Hence, when representing a qubit by a pair of complex num-
bers (z1 , z2 ) there is both a redundant global phase and global
• a particular kind of reversible operations — which for
amplitude. Concerning the redundant global amplitude, one
example do not admit cloning as well as deleting, or,
usually only considers normalized vectors i.e.
• irreversible measurements, for which we have to per-  
form acrobatics between ‘the good’ and ‘the bad’. z1
(z̄1 z̄2 ) ◦ = z̄1 · z1 + z̄2 · z2 = x21 + x22 + y12 + y22 = 1 ,
z2

since those are the ones which occur in the expressions for
2.2 Describing a qubit with complex numbers calculating the probabilities. Note here in particular that a
pair of complex numbers has four ‘real degrees of freedom’,
Let R denote the real numbers and C denote the complex
and hence that a pair of complex numbers up to a non-zero
numbers i.e. numbers z = x + i y where x, y ∈ R and i
complex multiple has two ‘real degrees of freedom’, what in-
is implicitly
√ defined within i · i = −1 so i can be thought of
deed corresponds with points on a sphere. More generally,
as −1. Hence for addition and multiplication of complex
‘n-tuples of complex numbers up to a non-zero complex mul-
numbers z1 = x1 + i y1 and z2 = x2 + i y2 we have
tiple’ have 2n − 2 ‘real degrees of freedom’.
z1 + z2 = (x1 + y1 ) + i (y1 + y2 )
Exercise 2.1 If we take ‘all pairs of real numbers up to a non-
z1 · z2 = (x1 x2 − y1 y2 ) + i (x2 y1 + y2 x1 ) . zero real multiple’ to be the states of some system, which
geometric object do we obtain as the state space? How would
The complex conjugate of z = x + i y is z̄ = x − i y hence you define opposite states? Representing real number pairs in
the XY -plane, when do two such pairs yield opposite states?
z̄ + z = 2x and z̄ · z = x2 + y 2 .
5
  
1 which induce a change of state
Special examples of states are |0i := and |1i :=
0       
 
0 1 0 z1 z1 1
which constitute the so-called computational basis |ψi 7→ P0 (|ψi) = = ∼
1 0 0 z2 0 0
corresponding to the classical bit values 0 and 1. The states       
of the computational basis are indeed opposite 0 0 z1 0 0
 states,
 which |ψi 7→ P1 (|ψi) =
0 1 z2
=
z2
∼
1
z1′
 
z1
in terms of pairs of complex numbers and
z2 z2′ i.e. the possible outcome states indeed constitute the compu-
requires tational basis. All the other measurements on a qubit can be
z̄1 · z1′ + z̄2 · z2′ = 0 , obtained by rotations of the sphere using the same transfor-
or equivalently, in terms of an inner or matrix product, mations which characterize the logic gates, resulting in
 ′ 
z1 U ◦ Pβ ◦ U −1
(z̄1 z̄2 ) ◦ = 0.
z2′ i.e. using U −1 we first rotate ‘backwards’ to the computa-
tional basis, then preform the measurement in the computa-
In practice however, calculations can be performed within tional basis, and then using U to rotate forward again. From
standard linear algebra, ignoring these redundancies. For ex- the above requirements on U we obtain for
ample, quantum logic gates are 2 × 2-matrices of complex    
numbers U11 U12 −1 Ū11 Ū21

U11 U12
 U= and U := ,
U= , U21 U22 Ū12 Ū22
U21 U22
that they are indeed inverses i.e. both
and induce a change of the state „ «„ « „ «„ «
Ū11 Ū21 U11 U12 U11 U12 Ū11 Ū21
     and
z1 U11 U12 z1 Ū12 Ū22 U21 U22 U21 U22 Ū12 Ū22
|ψi = 7→ U (|ψi) =
z2 U21 U22 z2 yield the identity  
1 0
 
U11 · z1 + U12 · z2
= , .
U21 · z1 + U22 · z2 0 1

which both preserves normalization

 and opposites i.e. the
 Exercise 2.2 In the lectures it was explained that we can rep-
1 0
‘canonical opposites’ and should stay oppo- resent pairs of complex numbers up to a complex number on a
0 1
sphere, with |0i := (1, 0) as the sphere top and |1i := (0, 1)
site
 andpreserve their global amplitude, resulting in both
U11 U12 as the sphere bottom, and with the states √12 (1, eiθ ) on the
and being normalized, i.e. ‘equator’, where we singled out |+i := √1 (1, 1), |−i :=
U21 U22 2
√1 (1, −1), |y+ i := √12 (1, i) and |y− i := √1 (1, −i). What
  2 2
U11 is the action of the following operations on these 6 special
(Ū11 Ū21 ) ◦ = Ū11 · U11 + Ū21 · U21 = 1
U21 points (depict on the sphere):
and
     
1 1 1 1 0 1 0
H := √ S := T := π

U12
 2 1 −1 0 i 0 ei 4
(Ū12 Ū22 ) ◦ = Ū12 · U12 + Ū22 · U22 = 1 ,
U22 respectively called the Hadamard-, phase-, and π/8-gate (also
and called the T-gate), and
 

U12
 0 1
(Ū11 Ū21 ) ◦ = Ū11 · U12 + Ū21 · U22 = 0 . X :=
U22 1 0
 
In other words, measurements are representable by particu- 0 −i
Y :=
lar families of projectors. The measurement with respect to i 0
the computational basis is described by the following pair of 
1 0

projectors Z :=
0 −1
   
1 0 0 0 typically called the Pauli X, Y and Z matrices. In particular,
P0 := and P1 := which of the 6 states mentioned above are either invariant or
0 0 0 1
6
permuted by these gates. Can you discover any special rela- freedom we obtain (n × n) − 1 complex degrees of freedom,
tions between these operations? (e.g. do some commute, are that is, for n large enough, approximately
idempotent (U ◦ U = U ), involutive (U ◦ U = 1), or do we
have a relation like U1 ◦ U2 = U3 for certain triples? (Note: 2n 7→ n2 ,
please spend as much time as needed on this important ques-
and it is the resulting additional degrees of freedom which
tion.)
enable communication. Actually, as we will see it what will
come, it is not completely wrong to think of the n × n-matrix
representing a communication channel through which infor-
2.3 Describing two qubits mation can flow and be processed.

Since, for the case of qubits, the fact that 2 + 2 = 2 × 2 might

cause some confusion in the argument we wish to make, we 3 von Neumann’s pure state formalism
will consider ‘n-tuples of complex numbers up to a non-zero
complex multiple’, called ‘qudits’ (d is for digit). Consider We will only consider finite-dimensional Hilbert spaces, that
the following situation. We start with 3 qudits, the first one is, in physical terms, all measurements have a finite number
being in an arbitrary state |ψi and the other two being in a of outcomes. While for informatic purposes this suffices, in-
particular joint state |Ψi. Then we apply a particular joint finite spectra do play an important role in general quantum
measurement to the first two qudits. Using quantum theory, it mechanics e.g. position and momentum observables.
can then be shown that after performing a particular logic gate
on the third qunit it will be exactly in the same state |ψi as the
3.1 Hilbert space
first qunit initially was. This involves sending the measure-
ment outcome which ‘witnesses which projector Pα actually Definition 3.1 A (finite-dimensional) Hilbert space is a vec-
took place’ to the third qunit, such that the appropriate logic tor space H over the complex number field C which also
gate Uα can be applied. comes with an inner-product, i.e. a map

h− | −i : H × H → C ,
|ψi
satisfying
Uα hψ|c1 · ψ1 + c2 · ψ2 i = c1 hψ|ψ1 i + c2 hψ|ψ2 i
{Pα }α hc1 · ψ1 + c2 · ψ2 |ψi = c̄1 hψ1 |ψi + c̄2 hψ2 |ψi
hψ|φi = hφ|ψi hψ|ψi ∈ R+ hψ|ψi = 0 ⇔ ψ = 0
for all c1 , c2 ∈ C and all φ, ψ, ψ1 , ψ2 ∈ H.
|Ψi
A linear operator between Hilbert spaces H1 and H2 is a map
|ψi f : H1 → H2 which satisfies

f (c1 · ψ1 + c2 · ψ2 ) = c1 · f (ψ1 ) + c2 · f (ψ2 ) ,

All together something very weird has happened here:
for all c1 , c2 ∈ C and all ψ1 , ψ2 ∈ H1 , hence the inner-
• We were able to teleport the quantum state of the product is linear in the second variable while being anti-linear
first qunit to the third qunit, that is, sending quantum in the first variable i.e. a so-called sesquilinear form. Two
data, but we only communicated finitary classical data, vectors ψ, φ ∈ H are called orthogonal iff
namely the measurement outcome! So what causes this
hψ | φi = 0
magic?
and a vector ψ ∈ H is normalized iff
The magic is hidden in the particular nature of the particular
initial joint state describing the second and the third qunit. |ψ|2 := hψ | ψi = 1 .
Indeed, quantum theory tells us that a pair of qudits is not de-
scribed by assigning a state to each of them, but by assigning
Exercise 3.2 Prove that C is itself a Hilbert space over C
a n × n-matrix (up to a complex multiple) to the pair of them,
i.e. show that there exists an inner-product on C.
and hence, rather than (n − 1) + (n − 1) complex degrees of
7
If f : H1 → H2 is a linear map then it always has a unique Special examples of projectors on H are the identity
adjoint f † : H2 → H1 which is implicitly defined within
1H : H → H :: ψ 7→ ψ
hf † (φ)|ψi = hφ|f (ψ)i
and the zero-operator
for all ψ ∈ H1 and all φ ∈ H2 . In Exercise 3.11.i we will
construct this adjoint, hence or otherwise prove its existence, OH : H → H :: ψ 7→ 0 .
and uniqueness also follows in a straightforward manor.
Proposition 3.7 Each self-adjoint operator H : H → H ad-
Exercise 3.3 Show that (g ◦ f )† = f † ◦ g † . mits a so-called ‘spectral decomposition’
X
A linear operator U is unitary if its inverse U −1 exists and, H= ai · Pi
equivalently, i

• U −1 = U † , where all ai ∈ R and all Pi : H → H are projectors which

are ‘mutually orthogonal’ i.e. Pi ◦ Pj = OH , ∀i 6= j.
• U (and also U † ) preserves the inner-product.
The proof of this proposition can be performed relying on the
matrix calculus and the fact that each self-adjoint operator
Exercise 3.4 Show that the two definitions of unitarity given admits a diagonal form (see Exercise 3.14 below).
above are indeed equivalent.
Postulate 3.8 [measurements] A measurement on a quan-
A subset of vectors A ⊆ H is called a subspace of a vec- tum system is described by a self-adjoint operator. The set
tor space H if it is closed under linear combinations of the {a } in the operator’s spectral decomposition are the mea-
i
vectors it contains i.e. surement outcomes while the set of projectors {Pi } describes
ψ1 , ψ2 ∈ A ⇒ c1 · ψ1 + c2 · ψ2 ∈ A . the change of the state that takes place during a measurement.
In particular, when a measurement takes place:
Special types of subspaces are those formed by rays . Rays
are the subspaces spanned (or generated) by a single vector 1. The initial state ψ undergoes one of the transitions
i.e.
span(ψ) = {c · ψ | c ∈ C} . Pi :: ψ 7→ Pi (ψ)
We are now ready to state a first postulate of von Neumann’s and the probability of the possible transitions is
formulation of quantum theory.
prob(Pi , ψ) = hψ|Pi (ψ)i
Postulate 3.5 [states and transformations] The state of a
quantum system is described by a ray in a Hilbert space. De- where ψ needs to be normalized.
terministic transformations of quantum systems are described
by unitary operators acting on that Hilbert space. 2. The observer which performs the measurement receives
the value ai as a token-witness of that fact.
Hence, from a computational perspective, the deterministic
logic gates which we can apply to quantum data are exactly
the unitary transformations. Besides unitary transformations, It should be clear that from a structural perspective the ac-
other linear endo-operators which play a special role in quan- tual values of the measurement outcomes {ai } are of no sig-
tum theory are self-adjoint operators i.e. nificance, and in a sense the respective measurements repre-
sented by the self-adjoint operators
hH(φ)|ψi = hφ|H(ψ)i X X
† ai · Pi and i · Pi
that is for all ψ ∈ H1 and all φ ∈ H2 , H = H. Self- i i
adjoint endo-operators P : H → H which are also idempo-
tent, i.e. P ◦ P = P, are called projectors. can be considered as equivalent, and in particular, the latter is
completely determined by the set {Pi }i . On the other hand
Exercise 3.6 i. If U is unitary and H self-adjoint show that however, the measurement outcomes are typically physical
U −1 ◦ H ◦ U is also self-adjoint. ii. If U is unitary and P is a quantities such as position, momentum and energy, which of
projector show that U −1 ◦ P ◦ U is a projector. course play an important quantitive role in physical theories.

8
 so we obtain the usual formula for application of the matrix
Exercise 3.9 Show that, equivalently, we could have set of f to the column of vector coordinates of ψ i.e.

prob(Pi , ψ) = |Pi (ψ)|2

 P    
j m1j cj m11 · · · m1j · · · m1m c1
 ..   .. .. ..   .. 
 P .
  . . . 
for the probability of each possible transition.  . 
  
 
j mij cj  =  mi1 · · · mij · · · mim   cj  .
    

 ..   .. .. ..   .. 
We have thus far only considered the description of individual    . . .  . 
P .
quantum systems. A postulate on compound systems is still j mnj cj mn1 · · · mnj · · · mnm cm
missing, but for this we need to introduce the tensor product
′
of Hilbert spaces — we postpone this discussion until Sub- On the other hand, for (mij )ij the matrix of another linear
section 3.3. operator g : H2 → H3 with respect to the basis {e′′i }i — we
then apply the above result twice and obtain
X 
3.2 Matrices (g ◦ f )(ek ) = g mjk · e′j
j
A basis for a vector space H is a set of vectors {ei }i which is X
such that each ψ ∈ H can, in a unique manner, be written as = mjk · g(e′j )
j
X X X 
ψ= ci · ei = mjk · m′ij · e′′i
i j i
X X 
for some set of complex numbers {ci }i , which we call the co- = m′ij mjk · e′′i
ordinates of ψ with respect to the basis {ei }i , and the number i j
of basis vectors is the dimension of the vector space. Given a
fixed basis {ei }i of H1 any linear operator f : H1 → H2 is so we obtain the usual formula for post-composing the matrix
completely determined by its action on the basis vectors since of g with the matrix of f , i.e.
 ′
m11 · · · m′1j · · · m′1m
 
m11 · · · m1k · · · m1µ
X  X
f (ψ) = f ci · ei = ci · f (ei ) .  .. .. ..   .. .. .. 
i i  . . .   . . . 
 ′
 m · · · m′ · · · m′   mj1 · · · mjk · · · mjµ 
  
i1 ij im
Moreover, since given a basis {e′i }i for H2 for each f (ei )   .. ..

..   .. ..

.. 
there exists a unique set {mij }i such that  . . .   . . . 
′ ′
mn1 · · · mnj · · · mnm ′ mm1 · · · mmk · · · mmµ
X
′
f (ej ) = mij · ei ,  P ′ P ′ P ′
j m1j mj1 · · · j m1j mjk · · ·

i j m1j mjµ
 .. .. .. 
f is completely determined by its matrix (mij )ij with respect  P .′ P .′ P .′
 

′
to the basis {ei }i and {ei }i . By convention, the index i runs = 
 j m ij m j1 · · · j m ij mjk · · · j m ij mjµ
.

.. .. ..
over rows and the index j runs over columns i.e.
 

P ′ . P ′ . P ′ . 
j mnj mj1 · · · j mnj mjk · · · j mnj mjµ
 
m11 · · · m1j · · · m1m
 .. .. ..  In fact, application of a linear operator to a vector can itself
 . . . 

 mi1 · · · mij · · · mim  .
 also be seen as a composition of functions, noting that there

 ..
 is a one-to-one correspondence between the vectors ψ ∈ H
.. .. 
 . . .  and the linear functions
mn1 · · · mnj · · · mnm
C → H :: 1 7→ ψ ,
P
When applying f to a vector ψ = j cj · ej we have
the matrix of the latter being the column of coordinates of ψ.
Note here also that having a matrix calculus is a property also
X  X
f cj · ej = cj · f (ej )
satisfied by relations — although not over a field but over
j j
X X  a semiring i.e. a ring without inverses — cf. {0, 1}-valued
= cj · mij · e′i matrices exactly encode all relations, hence in particular also
j i all multi-valued ones. In fact, structurally, linear operators are
X X 
= mij cj · ei ′ mathematically much closer to being relations than to being
i j functions — a fact which we will return to later in this course.
9
 Having an inner-product on top of a vector space structure, and hence the matrix of a general self-adjoint operator is
i.e. having a Hilbert space as in Definition 3.1, turns out to be  
r11 · · · min · · · mn1
the same thing as fixing a basis in that vector space.  .. .. .. 
 . . . 
 
Exercise 3.10 For a vector space H with basis {ei }i show  mi1 · · · rii · · · min 
 
that  .. .. .. 
   . . . 
c1
 ..  mn1 · · · min · · · rnn
  . 
hφ | ψi := c′1 · · · c′j · · · c′n 
  where r1 , . . . , ri , . . . , rn ∈ R.
 cj 

 ..  As is typical, we adopt the notation of writing a † in su-
 . 
cn perscript after an operator (e.g. A† ) to denote both the ad-
P P ′ joint of a linear map and the conjugate transpose of a matrix.
with ψ = i ci · ei and φ = c
i i · ei indeed defines an Since unitary operators preserve the inner-product, a unitary
inner-product in the sense of Definition 3.1. transformation sends an orthonormal basis {ei }i to another
orthonormal basis {U (ei )}i , and conversely, as is the case for
In a Hilbert space a basis is called orthonormal if
any linear operator, a unitary transformation is completely de-
hei |ej i = δij termined by its action on an orthonormal basis. So when fix-
ing an orthonormal basis, there is a bijective correspondence
where δij = 0 for i 6= j and δii = 1. In this case, since
between unitary operators and the set of all orthonormal ba-
D E X
sis.
X
he |ψi = e
i i j c ·e =
j j c he |e i = c
i j i
j j
Exercise 3.12 i. Describe the matrix of a unitary operator
we obtain the coordinates through the inner-product. Since U : H1 → H2 with respect to the basis {ei }i of H1 and the
DX X E X X basis {U (ei )}i of H2 . ii. Describe the matrix of a unitary
hφ|ψi = c′i ·ei cj ·ej = c̄′i cj hei |ej i = c̄′i ci operator U : H → H with respect to the basis {e } of H
i i
i j ij i
— hint: do this in terms of the vectors in {U (ei )}i . iii. For
in any Hilbert space the inner-product always coincides with unitary operators of type H1 → H2 , explicitly describe the
the one defined in Exercise 3.10 whenever {ei }i is an or- bijective correspondence with orthonormal basis of H2 .
thonormal basis of H. It can be shown that each Hilbert space
admits an orthonormal basis so what we can do with Hilbert Let (mjk )jk be the matrix of f : H1 → H2 for orthonormal
′
spaces can be done in matrix calculus over C with the inner- basis {ek }k of H1 and {ej }k of H2 i.e.
product of Exercise 3.10. X
f (ek ) = mjk · e′j .
Exercise 3.11 If (mij )ij is the matrix of f : H1 → H2 in j

some basis for H1 and some basis for H2 show that (mji )ij We would like to know what the matrix of f is for basis
is the matrix of its adjoint f † with respect to those basis. {U (ek )}k of H1 and {U ′ (e′j )}j of H2 . Let
X X
So when we have agreed on a fixed basis for each Hilbert U (el ) = ukl · ek and e′j = u′ji · U ′ (e′i )
space, by Exercise 3.11.i the matrix corresponding to the ad- k i
joint of the linear operator with matrix where (ukl )kl is the matrix of U for the basis {ei }i and (u′ij )ij
is the matrix of U ′ for the basis {e′i }i , and hence (ūji )ij the
 
m11 · · · m1j · · · m1m
 ..
 .
.. ..  matrix of U −1 = U † in that basis. We obtain
 . . 
 X 
 mi1 · · · mij · · · mim  f (U (e l )) = f u kl · ek
 
 .. .. ..  k
 . . .  X
mn1 · · · mnj · · · mnm = ukl · f (ek )
k
is its conjugate transposed X X 
  = ukl · mjk · e′j
m11 · · · mi1 · · · mn1 k j
 .. .. ..  X X X 
 .
 . . 
 = ukl · mjk · u′ji · U ′ (e′i )
 m1j · · · mij · · · mnj , k j l
 
 .. .. ..  X X 
 . . .  = u′ji mjk ukl · U ′ (e′i )
m1m · · · mim · · · mnm i jk

10
  
so the resulting matrix for f in the basis {U (ek )}k of H1 and for P := 1 0
0 iv. Give the probabilities for the input
{U ′ (e′j )}k of H2 is the matrix product 0 0
states e− , e+ , eθ and e⊥
θ for the measurements
′ †
(uij )ij (mjk )jk (ukl )kl .
{P+ , P⊥−} and {Pθ , P⊥θ }.
′
When denoting the matrices of f , U and U when expressed
in the basis {ei }i and {e′i }i (slightly abusively) also as f , U
and U ′ this expression simplifies to
†
Quantum mechanics in matrix terms. We provided both
U′ ◦ f ◦ U . unitary operators and quantum measurements (as families of
mutually orthogonal projectors arising from a self-adjoint op-
Proposition 3.13 For each self-adjoint operator H : H → H erator through the spectral decomposition theorem) with an
there exists an orthonormal basis in which its matrix is ‘diag- easy matrix representation, given a fixed basis {e } :
i i
onal ’ i.e. all its non-diagonal elements become 0.
• Unitary operators are in one-to-one correspondence with
Fixing an orhonormal basis {ei }i in which we express all ma- ONBs, and represent in matrix terms as the ONB
trices let {U (ei )}i be the basis in which the matrix of H is di- {U (ei )}i written as a list of column vectors
agonal i.e., continuing our abuse of notation for the matrices

of f and U in {ei }i , U (e1 ) · · · U (en ) .

U† ◦ f ◦ U . • Non-degenerate quantum measurements — i.e. quan-

is diagonal, so for the matrix of f expressed in {ei }i we have tum measurements for which the spectral decomposition
only contains of projectors with rays as range, or equiva-
f = (U ◦ U † ) ◦ f ◦ (U ◦ U † ) lently, for which the number of mutually orthogonal pro-
= U ◦ (U † ◦ f ◦ U ) ◦ U † jectors is equal to the dimension of the Hilbert space —
are completely determined by a unitary operator, which
= U ◦ M ◦ U†
itself is completely determined by a basis {U (ei )}i , with
respect to which the quantum measurement is given by
where M is some diagonal matrix. Conversely, each matrix
N = U ◦ M ◦ U † with M diagonal defines a self-adjoint {U ◦ P1 ◦ U † , . . . , U ◦ Pn ◦ U † }
operator by Exercise 3.6.i, namely the one which has matrix
N in the basis {ei }i , since we can interpret M as the matrix where
of a linear operator expressed in the basis {U (ei )}i . Note that
   
1 0 ··· 0 0 ··· 0 0
this argument also provides a converse to Proposition 3.13. 0 0 ··· 0   .. .. . .
. .. ..
 
P1 =  .  · · · Pn =  . .
   
.. .. ..
 . . . ..   0 ... 0 0 
Exercise 3.14 i. Is the orthonormal basis in which the ma- 0 0 ··· 0 0 ··· 0 1
trix of a self-adjoint operator becomes diagonal unique? ii.
Describe the matrix of general projectors in an orthonormal The outcome state for the projector U ◦ Pi ◦ U † is ex-
basis in which its matrix is diagonal. iii. Relying on Propo- actly the state described by the basis vector U (ei ). This
sition 3.13 explicitly construct the spectral decomposition justifies the extremely handy slogan:
(cf. Proposition 3.7) of a self-adjoint operator H : H → H.
non-degenerate measurement = orthonormal basis!

• Degenerate quantum measurements require, in addition

Exercise 3.15 i. Which projectors  P+ and P− have the states
 to a basis, specification of a partition
1 1
respectively described by e+ := and e− :=
1 −1 {1, . . . , n} = I1 ∪ . . . ∪ Ik
as their ‘only outcome states’ i.e. the range of the
 projector
 is
1 which provides (as projectors) a family diagonal matri-
the ray spanned by that vector. ii. Given eθ := , pick
eiθ ces which have 0s everywhere except for the ith diagonal
a vector e⊥θ which is orthogonal eθ and give the projectors Pθ elements for i ∈ Ij where there are 1s i.e.
⊥
and Pθ which have the states described by these vectors as  j  
their only outcome states. iii. Give the matrices of the unitary r 11 · · · 0 j
.  1 ≤ j ≤ k , rii ∈ {0, 1}

 

operators U+ and Uθ which are such that  .. . . .
 . . .  j
 j rii = 1 ⇔ i ∈ Ij 
0 · · · rnn
 
P+ = U+ ◦ P0 ◦ U+† and Pθ = Uθ ◦ P0 ◦ Uθ† .
11
 • Physically speaking, basisd on the above we can A key mathematical application of the direct sum is that it
make two choices for implementing a particular (non- provides Hilbert spaces which come with a preferred basis.
degenerate) measurement {U ◦ P− ◦ U † }i :
Exercise 3.17 Show that for each n ∈ N the direct sum pro-
1. We perform a measurement in the basis {U (ei )}i
vides a Hilbert space of dimension n together with a canoni-
i.e. we implement the projectors
cal choice for a basis — hint: recall that C is itself a Hilbert
† † space which comes with a special element 1 ∈ C (= the mul-
{U ◦ P1 ◦ U , . . . , U ◦ Pn ◦ U } ;
tiplicative inverse of the underlying field C).
2. We first perform the unitary transformation U † ,
then the measurement in the basis {ei }i i.e. we im- Physically, the direct sum describes pairs of states, and hence
plement the projectors would constitute a likely candidate to describe compound sys-
{P1 , . . . , Pn } , tems, but quantum theory disagrees! This means we will need
to introduce the tensor product of two Hilbert spaces.
and then (provided we are not just interested in
the measurement outcome but also in the resulting Postulate 3.18 [compound systems] The joint state of a
state) we perform the unitary transformation U . compound quantum system consisting of two subsystems is
The advantage of the second implementation is that we described by the tensor product of the Hilbert spaces which
only need to rely on one particular quantum measure- describe the two subsystems.
ment, independent on which measurement we actually
want to implement, namely the one in the fixed basis. So now we will define the tensor product of Hilbert spaces H1
and H2 , which allows several formulations that are equivalent
‘up to isomorphism’. Initially we define it as
3.3 Tensor structure   
 c11 . . . c1m
 

 .. . .. .
..  ∀i, j : cij ∈ C ,
We define the direct sum of Hilbert spaces H1 and H2 as H1 ⊗ H2 :=  .

 
cn1 . . . cnm
n o  
H1 ⊕ H2 := (ψ1 , ψ2 ) ψ1 ∈ H1 , ψ2 ∈ H2
where n is the dimension of H1 and m is the dimension of
together with two linear injections
H2 , together with the map
ι1 : H1 → H1 ⊕ H2 :: ψ 7→ (ψ, 0)
ξ : H1 ⊕ H2 → H1 ⊗ H2 :: (ψ, φ) 7→ (ci c′j )ij
ι2 : H2 → H1 ⊕ H2 :: ψ 7→ (0, ψ) . P ′ ′
P
where ψ = i ci · ei and φ = j cj · ej for orthonormal
It is simple to extend the definition of the direct sum beyond
L basis {ei }i of H1 and {e′i }i of H2 , that is, using coordinates,
the binary case — i.e. we can also consider i Hi together
with a family of linear injections 
c1
  ′ 
c1

c1


ξ ::   ...  ,  ...   7→  ...  c′1 . . . c′m .

M

{ιj : H1 → Hi }j .
     
i cn c′m cn
We define the direct sum of two linear maps f : H1 → H1 ′
Note that the elements of H1 ⊗ H2 are exactly the matrices
and g : H2 → H2′ component-wise i.e. as the linear map
of the linear maps of type H2 → H1 (which by the taking
f ⊕ g : H1 ⊕ H2 → H1′ ⊕ H2′ :: (ψ, φ) 7→ (f (ψ), g(φ)) . the transpose or the adjoint are equivalent to the matrices of
the linear maps of type H1 → H2 ),1 and when thinking of
Exercise 3.16 i. Show that H1 ⊕H2 is indeed a Hilbert space 1
We could of course also have defined the tensor product such that it is
i.e. show that it comes with an inner-product — write this exactly the matrices of the linear maps of type H1 → H2 , but in that case
inner-product down without referring to a basis. ii. Describe the presentation would have been slightly less clean due to the fact that
matrix composition goes backward as compared to the (western) reading
a basis of H1 ⊕ H2 in terms of basis for the Hilbert spaces direction i.e. we would have
H1 and H2 , and describe the coordinates of the elements of 80 1˛ 9
c11 . . . c1n
H1 ⊕ H2 in that basis. iii. Show that >
< ˛ >
=
H1 ⊗ H2 := @ ... .. .. C ˛ ∀i, j : c ∈ C ,
B ˛
. . A˛ ij

ιi ◦ ι†i = 1⊕i Hi and ι†j ◦ ιi = δij .

X > >
: ˛ ;
cm1 . . . cmn
i
together with
where δij = O : Hi → Hj :: ψ 7→ 0 iff i 6= j and δii = 1Hi .
ξ : H1 ⊕ H2 → H1 ⊗ H2 :: (ψ, φ) 7→ (c′j c′i )ji

12
vectors as linear functions of type C → H (cf. the discussion H there exists a unique linear map h : H1 ⊗ H2 → H such
above), for the corresponding matrices we have that ζ = h ◦ ξ i.e. in a commutative diagram:

ξ :: (ψ, φ) 7→ ψ ◦ φT ξ
H1 ⊕ H2 ✲ H1 ⊗ H2
where (−)T denotes the transpose. One easily verifies that
H1 ⊗ H2 is a Hilbert space for the inner-product ∃!h

ζ
❄
X
h(cij )ij | (c′ij )ij i := cij c′ij .

✲
ij
H

A basis for the tensor product is provided by all m × n-

matrices which only contain a single 1 while all the other el- To stress that ξ is not ‘globally’ linear one rather writes its
ements are 0. Notice that when applying ξ to a pair (ei , e′j ) domain as the cartesian product H1 × H2 than as the direct
for basis {ei }i of H1 and {e′j }j of H2 we obtain the matrix sum H1 ⊕ H2 . The so-called universal property of the tensor
with 1 in the jth row and ith column spot, and which is 0 product expressed in Exercise 3.19.iii is in many cases taken
everywhere else, in particular, as the definition of the tensor product, which defines it up to
an ‘isomorphism of vector spaces’ — cf. [?] §8.7 & §8.8.
ξ :: (ei , e′j ) 7→ ei ◦ e′T
j While our definition for the tensor product is very straight-
forward, it does depend on a choice of basis. We will work
for the matrices of the linear maps representing these basis toward a basis-independent (but slightly less straightforward)
vectors. So roughly speaking, while we obtain a basis for construction. First replace the matrices in the above definition
the direct sum by taking the ‘(disjoint) union’ of the basis of the tensor product by ‘formal linear combinations’ with re-
vectors of the underlying spaces, we obtain a basis for the spect to a basis for the tensor product i.e.
tensor product by taking the ‘cartesian product’ of the basis
nX o
vectors of the underlying spaces. In particular: H1 ⊗ H2 := cij · (ei , e′j ) ∀i, j : cij ∈ C
ij
dim(H1 ⊕ H2 ) = dim(H1 ) + dim(H2 ),
together with the map
dim(H1 ⊗ H2 ) = dim(H1 ) × dim(H2 ).
ξ : H1 ⊕ H2 → H1 ⊗ H2 :: (ψ, φ) 7→ (ψ, φ)
Finally, the above defined tensor product easily extends be-
yond the binary case by setting where for (ψ, φ) ∈ H1 ⊗ H2 we set
i=k X X  X
′ ′
ci c′j · (ei , e′j ) .
n o
c · e , c · ej :=
O
Hi := (ci1 ...ik )i1 ...ik ∀i1 , . . . , ik : ci1 ...ik ∈ C i i j (1)
i=1 i j ij

together with the map This seems to allow us to think of ‘pairs of states’ as possible
i=k i=k
states of compound systems, but one still needs to be a bit
cautious as is demonstrated by the following exercise.
M O
ξ: Hi → Hi :: (ψ1 , . . . , ψk ) 7→ (ci1 . . . cik )i1 ...ik
i=1 i=1
P P Exercise 3.20 Show that ξ is not injective, and in particular,
where ψ1 = i1 ci1 · ei1 , . . . , and ψk = ik cik · eik . which ψ 6= φ are equalized under the action of ξ.

Exercise 3.19 i. Show that ξ is not linear. ii. Show that ξ is On the other hand, besides pairs of states there are many other
bilinear i.e. linear when conceived as a two-variable function ones too, the so-called entangled states, which are superposi-
where one variable takes its values in H1 and the other one in tions (= sums) of the pairs of states.
H2 . iii. Show that for any other bilinear map ζ : H1 ⊕ H2 →
Exercise 3.21 Show that not all elements of H1 ⊗ H2 can be
which becomes
written as (ψ, φ) for some ψ ∈ H1 and some φ ∈ H2 .
c′1 c′1
00 1 0 11 0 1
c1
BB . C B . CC B . C` ´
ξ :: @ @ .. A , @ .. A A 7→ @ .. A c1 ... cn .
Usually the pairs (ψ, φ) ∈ H1 ⊗ H2 are denoted by ψ ⊗ φ to
cn c′m c′m indicate that we ‘live in the tensor product’ — and not in the
But the true ‘natural’ structural connection between the tensor product and directsum/cartesian product. Hence we obtain
linear maps is actually far more subtile than either of these two matricial
candidates, as we will see in Section 11.2. ξ : H1 ⊕ H2 → H1 ⊗ H2 :: (ψ, φ) 7→ ψ ⊗ φ,
13
so ξ = − ⊗ −, for which we moreover have where all summations are finitary. Next we introduce a con-
X  X  X gruence (= equivalence relation) on Ω, namely
ci · ei ⊗ c′i · e′i := ci c′j · ei ⊗ e′j . X X  X
i i ij
αi · βj · ψi,j ⊗ φi ∼ αi βj · ψi,j ⊗ φi
i j ij
We can actually think of ψ ⊗ φ as a special case of the tensor X X  X
of two linear maps, where we define the tensor of two linear αi · ψi ⊗ βj · φi,j ∼ αi βj · ψi ⊗ φi,j
i j ij
maps f : H1 → H1′ and g : H2 → H2′ as the linear map
— which is clearly basis-independent. Expressing this con-
f ⊗ g : H1 ⊗ H2 → ⊗H1′ H2′
:: ψ ⊗ φ 7→ f (ψ) ⊗ g(φ) . gruence in respective basis for H1 and H2 yields
X X
Hence you should be aware of the triple use of −⊗−, namely αi · ψi ⊗ φi ∼ α̃i · ψ̃i ⊗ φ̃i
i i

• as a map from pairs in the tensor product, m

α̃i c̃ij d˜ik

X X
• as a connective on Hilbert spaces, ∀j, k : αi cij dik =
i i
• as a tensor of two linear maps. where
X X
ψi = cij · ej , φi = dik · e′k ,
j k
Exercise 3.22 i. First convince yourself that the above pre-
d˜ik · e′k ,
X X
scription for f ⊗ g is well-defined. Now rely on Exercise ψ̃i = c̃ij · ej , φ̃i =
3.19.iii to show (again) that the prescription of f ⊗ g indeed j k
induces a unique linear map. ii. Show that
which boils down to the matrices (cjk )jk := ( i αi cij dik )jk
P
we started with. Conversely, each such matrix can be realised
hψ ⊗ ψ ′ | φ ⊗ φ′ i = hψ|φihψ ′ |φ′ i .
by setting i := (j, k), ψjk := ej , φjk := e′k and αjk := cjk ,
indeed defines an inner-product on the tensor product of two
Hilbert spaces. iii. Let f, g : H → H be linear maps with The no-cloning theorem [8]. The passage from pairs of
respective matrices states to the tensor product, on which we only are allowed
   ′ ′  to act with unitary operations, comes with some drastic con-
a b a b sequences. Assume we start with two quantum systems in
and
c d c′ d′ states ψ ⊗ φ0 ∈ H and we which, by means of some unitary
operator U : H ⊗ H → H ⊗ H, to copy the state of the first
in the basis {e1 , e2 } of H. Give the matrix of f ⊕ g in the
one to the second one i.e. obtain ψ ⊗ ψ. Assume we are able
basis
to do this both for ψ1 and for ψ2 i.e. we have
{(e1 , 0), (e2 , 0), (0, e1 ), (0, e2 )}
U (ψ1 ⊗ φ0 ) = ψ1 ⊗ ψ1 and U (ψ2 ⊗ φ0 ) = ψ2 ⊗ ψ2
and the matrix of f ⊗ g in the basis
— note here that it is crucial that in both cases U should be
{e1 ⊗ e1 , e1 ⊗ e2 , e2 ⊗ e1 , e2 ⊗ e2 } . the same, since in general the state of the system is unknown,
and measurement would alter it. Taking the inner-product of
iv. In these basis describe the 4 × 4 matrices for operations the above equalities yields

σ⊕ : H ⊕ H → H ⊕ H :: (ψ, φ) 7→ (φ, ψ) hU (ψ1 ⊗ φ0 )|U (ψ2 ⊗ φ0 )i = hψ1 ⊗ ψ1 |ψ2 ⊗ ψ2 i ,

that is, by U † = U −1 and Exercise 3.22.ii,
σ⊗ : H ⊗ H → H ⊗ H :: ψ ⊗ φ 7→ φ ⊗ ψ .
hψ1 |ψ2 ihψ0 |ψ0 i = hψ1 |ψ2 ihψ1 |ψ2 i
Are these operations unitary?
and hence, assuming that all vectors are normalized,
All the above results in a basis-independent construction of hψ1 |ψ2 i = hψ1 |ψ2 i2
the tensor product of H1 and H2 . First we ‘freely’ introduce
the following set of formal expressions forcing hψ1 |ψ2 i = 0 or hψ1 |ψ2 i = 1 i.e. ψ1 and ψ2 need
to be either equal or orthogonal, so we cannot copy arbitrary
nX o
Ω := αi · ψi ⊗ φi ∀i : αi ∈ C, ψi ∈ H1 , φi ∈ H2 states! There is also a corresponding no-deleting theorem [9],
i
but that is slightly more subtle in its formulation.
14
Bell- and EPR-correlations. Two historically important will be a steping-stone to a compositional high-level formal-
examples of entangled states are the Bell-state ism, and hence from the start we will formally justify it, and
we will also need to slightly restrict it. Ultimately, we will
Bell := e1 ⊗ e1 + e2 ⊗ e2 extend it into a graphical notation in order to be able to cope
with the intrinsic two-dimensional compositional structure of
and the EPR-state quantum mechanics cf. sequential composition −◦− and par-
allel composition − ⊗ −. To justify this notation formally we
EPR := e1 ⊗ e2 − e2 ⊗ e1
want to think of vectors ψ ∈ H as linear maps
which respectively correspond to the matrices
C → H :: 1 7→ ψ ,
   
1 0 0 −1
and which we (slightly abusively) also denote by ψ. The Dirac
0 1 1 0
notation is formally justified by letting
i.e. the Bell-state corresponds to the identity, and the EPR-
state seems hardly any more interesting. However, let’s see • |ψi := ψ,
what’s happens when we measure them. First note that indeed • hψ| := ψ † ,
there are no a1 , a2 , a3 , a4 ∈ C such that either
„
a1
«
` ´
„
1 0
« „
a1
«
` ´
„
0 −1
«
in a table:
b1 b2 = b1 b2 =
a2 0 1 a2 1 0
linear map matrix Dirac
so the Bell-state and the EPR-state are truly entangled, that is,
cannot be written in the form ψ ⊗ φ. But the real magic starts 
c1

when we measure them in the computational basis. If we  .. 
ψ:C→H  .  |ψi
measure the left system i.e. we apply the (degenerate) mea-
surement cm
{P1 ⊗ id, P2 ⊗ id}
to the whole system we obtain ψ† : H → C hψ|


c̄1 . . . c̄m

(P1 ⊗ id)(Bell) = e1 ⊗ e1 (P1 ⊗ id)(EPR) = e1 ⊗ e2

Hence in particular we have
(P2 ⊗ id)(Bell) = e2 ⊗ e2 (P2 ⊗ id)(EPR) = e2 ⊗ e1
that is, we will now get a certain answer for a measurement |ψi† = hψ| and hψ|† = |ψi .
on the second system i.e. we now apply
We call |ψi a ket and hψ| a bra. When writing one symbol
{id ⊗ P1 , id ⊗ P2 } after another we think of it as composition, either of linear
functions or of the corresponding matrices. Hence an inner-
to the whole system, since in the case of the Bell-state we product is a bra-ket(pronounced bracket):
will always obtain the same outcome as we obtained when
measuring the first system, while in the case of the EPR-state linear map matrix Dirac notation
we will always obtain the opposite to what we obtained when
c′1
 
measuring the first system (cf. identity). Typically these two
 . 
systems are far apart so we witness a non-local effect. The ψ† ◦ φ c̄1 . . . c̄m  ..  hψ | φi
experimental proof of this non-local effect requires making c′m
measurements in more than a single basis and the measure-
ment outcomes must be shown to violate Bell’s Inequality.
A projector on the ray spanned by ψ is a ket-bra:
3.4 Dirac notation
linear map matrix Dirac
A very popular notation in quantum mechanics and quantum  
informatics is the so-called Dirac notation or bra-ket notation c1
 .. 
ψ ◦ ψ† Pψ := |ψihψ|


[5]. Interestingly, while in most textbooks this is declared to  .  c̄1 . . . c̄m
be ‘merely’ a convenient notation, and sometimes even as- cm
sumed as too informal and mathematically unsound, for us it
15
A projector on a ray indeed takes the shape ψ◦ψ † . To see this, measure between states. In Dirac notation one usually consid-
first note that with respect to a fixed basis we indeed have ers a privileged computational basis denoted by

Pi = ei ◦ e†i . { | i i | 0 ≤ i ≤ n − 1}

Hence, for a general projector for each Hilbert space of dimension n. Mathematically speak-
ing, such an n-dimensional Hilbert space which comes with a
Pψ = U ◦ Pi ◦ U † with ψ = U ◦ ei privileged basis can be produced as in Exercise 3.17, that is,

— one verifies that each projector on a ray indeed admits such H := C ⊕ . . . ⊕ C ,

a representation using Pi ◦ φ = c · ei for some c ∈ C —
and in particular for a qubit we have
Pψ = U ◦ (ei ◦ e†i ) ◦ U †
Q := C ⊕ C ,
= (U ◦ ei ) ◦ (e†i ◦ U † )
= (U ◦ ei ) ◦ (U ◦ ei )† for which the computational basis vectors are | 0 i and | 1 i.
= ψ◦ψ . † The vectors of the computational basis for H ⊗ H′ are in the
literature denoted in several ways:
As an application of Dirac notation observe that |ii ⊗ |j i | i i| j i | ij i
Pψ ◦ Pφ = |ψihψ|φihφ| = OH but we will not be using the second which might cause con-
fusion and even insinuates inconsistencies e.g. should
so two projectors on rays are orthogonal hψ|φi = 0 i.e. if
and only if the rays ψ and φ on which they project are or- (|ψihψ|)(|φihφ|)
thogonal — the underlined bra-ket is an inner-product, hence
a scalar, so we obtain c · |ψihφ| for c := hψ|φi which can be interpreted either as a composition or as a tensor i.e. as
indeed only be 0 if either c, ψ or φ would be 0/0. Here are
some more examples of expressions in Dirac notation which |ψihψ|φihφ| or |ψ ⊗ φihψ ⊗ φ| ?
illustrate the compositional nature of Dirac notation (here it
is assumed that f = f † ):
Exercise 3.23 Show that for projectors on rays we have
linear map matrix Dirac
Pψ ⊗ Pφ = Pψ⊗φ
c′1
  
m11 . . . m1m by first showing that
f ◦ψ
 .. ..   ..  f |ψi
 . .  . 
mn1 . . . mnm c′m (| ψi ⊗ | φi) ◦ λC = | ψ ⊗ φi ,

where we used the unitary ‘isomorphism’ (check this!)

 
m11 . . . m1m λC : C → C ⊗ C :: 1 7→ 1 ⊗ 1 ,
φ† ◦f c̄1 . . . c̄m

 .. ..  hφ|f
 . . 
mn1 . . . mnm by then showing that in general we have

(f1 ⊗ f2 ) ◦ (g1 ⊗ g2 ) = (f1 ◦ g1 ) ⊗ (f2 ◦ g2 ) ,

φ† ◦ f ◦ ψ ′
hφ|f |ψi
P
... = ij ci mij cj ∈C and finally using these two facts to infer the above claim.

Conclusively, we have the following change of notation:

As a special case we have probabilities for which we have (
e1 ❀ |0i e(i+1) ⊗ e(j+1) ❀ |iji
hψ|Pφ |ψi = hψ|φihφ|ψi = |hφ|ψi|2
e2 ❀ |1i e(i1 +1) ⊗ . . . ⊗ e(in +1) ❀ |i1 . . . in i
whenever the projector is of the form Pφ i.e. projects on a
and hence an arbitrary state now takes the form
one-dimensional subspace. Hence we have again a very triv-
ial computation which exposes an interesting feature: we can
X
Ψ= αi | i1 . . . in i .
really think of quantum probabilities as some kind of distance i
16
If we assume that ‘larger Hilbert spaces’ H always arise as for a partition I1 ∪ . . . ∪ Ik = {1, . . . , n}.

H := Q ⊗ . . . ⊗ Q From now on, we will use the same notation for a linear op-
erator and its matrix in the computational basis except when
then, using the notation it is explicitly stated to be otherwise. Consider for example
arbitrary qubit measurementwith as set of projectors
|i1 . . . in i with i1 . . . in ∈ {0, 1} n o
PU0 := U | 0ih0 |U †
, PU
1 := U | 1ih1 |U †
for the basis, this is actually nothing more then writing

|ii with i ∈ {0, . . . , 2n − 1}physically either being

n o
in binary rather than in decimal (or anything else). Note that PU0 := (U | 0i)(h0 |U †
) , PU
1 := (U | 1i)(h1 |U †
) ,
this binary representation also allows for ‘easy comparison’
with classical computing with bit-strings. Important exam- or
ples are the Bell-state and the EPR-state n o
PU 0 := U (| 0ih0 |)U †
, PU
1 := U (| 1ih1 |)U †
.
Bell = | 00i + | 11i and EP R = | 01i − | 10i .
If we perform this measurement then two possible outcome
or, for three qubits, the GHZ-state and the W-state states U | 0i and U | 1i can be obtained, ‘up to normalization’,
GHZ := | 000i+| 111i and W = | 100i+| 010i+| 001i . by post-composing the input state ψ with the respective pro-
jectors i.e.
Usually one introduces a normalization constant resulting in
U | 0ih0 |U † |ψi and U | 1ih1 |U † |ψi
1 1
√ (| 00i + | 11i) and √ (| 000i + | 111i)
2 2 The underlined scalars yield the corresponding probabilities
when multiplying them with their conjugate since
assuming that the basis vectors are normalized, but for both
esthetic and ecological reasons we will drop these. The stan- hψ|PU † † †
i |ψi = hψ|U | iihi |U |ψi = hi |U |ψihi |U |ψi .
dard single qubit and two-qibit computational basis are

{|0i , |1i} {|00i , |01i , |10i , |11i} Non-local correlations in Dirac Notation. To illustrate
Dirac notation in action we now redo this calculation. Mea-
with corresponding measurement projectors
surement in the computational basis of the first qubit of a
{|0ih0| , |1ih1|} {|00ih00| , |01ih01| , |10ih10| , |11ih11|} . qubit pair in the Bell-state either yields

General self-adjoint operators take the form2 (P0 ⊗ 1Q )(| 00i + | 11i)
i=n
X  i=n = (| 0ih0 | ⊗ 1Q )(| 00i + | 11i)
X
†
H=U αi · | iihi | U = αi · U | iihi |U † = (| 0ih0 | ⊗ 1Q )| 00i + (| 0ih0 | ⊗ 1Q )| 11i
i=n i=n = (| 0ih0 | 0i) ⊗ | 0i + (| 0ih0 | 1i) ⊗ | 1i = | 00i
and for projectors we have ai ∈ {0, 1}, hence they are
X  X or
P=U | iihi | U † = U | iihi |U † (P1 ⊗ 1Q )(| 00i + | 11i) = | 11i
i∈I i∈I
which indeed yields a certain answer for a measurement on
for I ⊆ {0, . . . , n}, so quantum measurements take the shape the second qubit, hence a non-local correlation.
nX X o
U | iihi |U † , . . . , U | iihi |U †
i∈I1 i∈Ik 4 Protocols from entanglement
2
Note that, of course, in general
˛ X ED X ˛ X With the machinery of tensor products and Dirac notation
ψi ψi ˛ 6= |ψi ihψi |
˛ ˛
˛ at hand we are now able to expose certain protocols, which
i i i
surprisingly have only been recently discovered, and lift the
e.g. for qubits (|0i + |1i)(h0| + h1|) is a projector on the state |0i + |1i,
while |0ih0| + |1ih1| is the identity! Hence the sum seems to play two
‘weirdness’ of non-local correlations one level higher — to
roles, and this will enable us to accomodate so-called mixed states to be outer space if you wish, where they meet the crew abord Star
introduced and studied in Section 11.4. Trek’s USS Enterprise.
17
4.1 Bell-basis and Bell-matrices for which we have (note the above matrix factors into a prod-
uct H ⊗ I of and the CNOT gate)
While the standard 2-qubit quantum measurement is with re-
spect to the computational basis |ΨBell-basis i = UBell-basis |Ψcomputational basis i

|00i, |01i, |10i, |11i and

a very important measurment basis is the Bell-basis: PBell = |ΨBell ihΨBell |

= |ΨBell i (|ΨBell i)†
|00i + |11i, |00i − |11i, |01i + |10i, |01i − |10i .
= UBell |Ψcomp. i (UBell |Ψcomp. i)†
This basis is obtained by respectively applying the unitaries †
= UBell |Ψcomp. ihΨcomp. |UBell
†
       
1 0 1 0 0 1 0 −1 = UBell Pcomp. UBell .
0 1 0 −1 1 0 1 0
to the second qubit of the Bell-state, i.e. applying 1Q ⊗ U to
the whole system. For example,
 when we apply the fourth 4.2 Teleportation and entanglement swapping
0 −1
Bell matrix U := to the qubit basis we obtain Quantum teleportation [11]. We are now ready to prove
1 0
quantum teleportation, which we have already described
| 0i 7→ | 1i and | 1i 7→ −| 0i above.
hence when we apply id ⊗ U we have
| 00i 7→ | 01i and | 11i 7→ −| 10i
Ui
so id ⊗ U applied to the Bell-state yields
| 00i + | 11i 7→ | 01i − | 10i
i.e. the fourth Bell-basis vector. We call these matrices the
Bell-matrices, and they are exactly (the transposed of) the
i P
matrices encoding the Bell-basis in our matricial definition
of the tensor product, cf.
≃
X
cij | iji ←→ (cij )ij Denoting the state of the input qubit as
ij
| ψi = c0 · | 0i + c1 · | 1i ,
Alternatively, we can apply the transposed Bell-matrices to
the first qubit of the Bell-state, i.e. applying U T ⊗ 1Q to the we assume that the second and third qubit are in the Bell-state.
whole system, and again we exactly obtain the Bell-basis. Then we perform a Bell-basis measurement on the first and
Hence, denoting the Bell-matrices by {Ui }i and the Bell-basis the second qubit. If the outcome state of this measurement is
by {Ψi }i we have the i-th Bell-basis vector, then we act with the transposed i-th
Bell-matrix on the third qubit. Explicitly, we start with
(UiT ⊗ 1Q )|ΨBell i = Ψi = (1Q ⊗ Ui )|ΨBell i
| ψi ⊗ (|00i + |11i)
Note also that we produce a Bell-basis measurement
n = (c0 · | 0i + c1 · | 1i) ⊗ (|00i + |11i)
(|00i + |11i)(h00| + h11|) , = c0 · (|000i + |011i) + c1 · (|100i + |111i)
(|00i − |11i)(h00| − h11|) , = |00i ⊗ (c0 · |0i) + |01i ⊗ (c0 · |1i)

(|01i + |10i)(h01| + h10|) , +|10i ⊗ (c1 · |0i) + |11i ⊗ (c1 · |1i)

o
(|01i − |10i)(h01| − h10|) There are four ‘cases’ corresponding to the possible outcome
states of the Bell-basis measurement, so we need to concider
from a measurement in the computational basis together with   
  (|ΨihΨ|) ⊗ 1Q | ψi ⊗ (|00i + |11i)
1 1 0 0
 0 0 1 1 
UBell-basis :=   for each Bell-basis vector |Ψi. E.g. for |01i − |10i we obtain
 0 0 1 −1 
1 −1 0 0 (|01i − |10i) ⊗ (c0 · |1i − c1 · |0i)
18
  
0 1
and applying to this indeed yields
−1 0
(|01i − |10i) ⊗ |ψi .
The other three cases proceed analogously.
to the different Bell-state entanglements:
Exercise 4.1 i. Given that you only have the ability to per-
form measurements in the computational basis, Hadamard
gates (see above), and CNOT-gates, i.e. gates with matrix
 
1 0 0 0
 0 1 0 0  Using a different geometry, we passed from
 0 0 0 1 ,
 

0 0 1 0
how would you perform a Bell-basis measurement? ii. When
performing such a Bell-basis measurement involving a mea-
surement in the computational basis, let boolean b1 ∈ B be
the outcome of measuring the first qubit, and let boolean to
b2 ∈ B be the outcome of measuring the second qubit. Let
U1 : B → U and U2 : B → U be functions with U the four
element set consisting of the Pauli-matrices and the identity.
Can you choose U1 and U2 such that U1 (b1 ) ◦ U2 (b2 ) pro-
vides the required correction of the third qubit in the tele-
portation protocol? (i.e. depending on the outcome b2 of the
measurement of the second qubit we perform unitary U2 (b2 ), Exercise 4.2 i. Verify the entanglement swapping protocol.
and then, depending on the outcome b1 of the measurement ii. If in the entanglement swapping protocol we start with two
of the first qubit we perform unitary U1 (b1 )) EPR-states rather than with two Bell-states, do we obtain the
same result, i.e. do we still obtain two Bell-states, or, do we
This implementation of the teleportation protocol is the one
instead obtain two EPR-states, or something else? iii. Can
you’ll find in most textbooks.
you modify the measurement dependent ‘unitary corrections’
such that we do end up with two Bell-states, with two EPR-
Entanglement swapping [12]. We start with four qubits, to states, or instead with one EPR-state and one Bell-state?
which we refer as a, b, c, d, where a and b are in a Bell-state,
and also c and d are in a Bell-state. Then we perform a Bell- When analyzing this protocol (and its proof) we can see that
basis measurement on c and d, and depending on the mea- it includes two crucial components:
surement outcome, analogously to what we did in the telepor-
tation protocol, we apply the transpose of the corresponding 1. The measurement destroys correlations between pairs
Bell-matrix both to qubit c and d. and creates new ones between other pairs.

2. The unitary ‘corrections’ using the Bell-matrices which

guarantee that the resulting correlated pairs are in-
Ui Ui deed all in the Bell-state, and hence ‘reverse’ the non-
deterministic differences due to the measurement.

A similar analysis also applies to the teleportation protocol.

We end this section with some additional comments on
Pi teleportation and entanglement swappping. While in the tele-
portation protocol we have

((|Ψi ihΨi |) ⊗ 1Q ) (| ψi ⊗ |ΨBell i) = |Ψi i ⊗ Ui |ψi

Now qubits a and d are in a Bell-state, and also qubits b and c (before applying U † to the third qubit) we also have
i
are in a Bell-state. So, we ‘swapped’ the entanglement from
the Bell-state entanglements: (hΨi | ⊗ 1Q ) (| ψi ⊗ |ΨBell i) = Ui |ψi
19
so it seems that we actually only need the ‘bra-part’ of the and hence does not really correspond with a primitive
 physi-

projector |ΨihΨ| to achieve teleportation. This seems to in- T c00 c10
cal operation. Indeed, when we apply f :=
dicate that there are truly ‘two components’ to a projector, c01 c11
one, the ket-part, producing an outcome states, and one, the to the qubit basis we obtain
bra-part, being the ‘action’ which (for example) yields tele-
portation. Hence the Dirac-representation of a projector as a | 0i 7→ c00 · | 0i + c01 · | 1i and | 1i 7→ c10 · | 0i + c11 · | 1i
bra followed by a ket really reflects two distinct components
in what such a projector actually does. Note that the ‘unitary hence when we apply 1Q ⊗ f T we have
corrections’ need not necessarily be the ones we used here,
but our choice will again be motivated by conceptual analy- | 00i 7→ c00 ·| 00i+c01 ·| 01i and | 11i 7→ c10 ·| 10i+c11 ·| 11i
sis. For entanglement swapping we have
so 1Q ⊗ f T applied to the Bell-state yields
(1Q ⊗ (|Ψi ihΨi |) ⊗ 1Q ) ((|00i + |11i) ⊗ (|00i + |11i))
| 00i + | 11i 7→ c00 · | 00i + c01 · | 01i + c10 · | 10i + c11 · | 11i
nXo
= (1Q⊗Q⊗Q ⊗ Ui )(| 0i⊗|Ψi i⊗| 0i + | 1i⊗|Ψi i⊗| 1i)
nXo
= (1Q⊗Q ⊗ Ui ⊗ Ui ) that is
X
(| 0i⊗|ΨBell i⊗| 0i + | 1i⊗|ΨBell i⊗| 1i) | 00i + | 11i 7→ cij · | iji .
ij

Alternatively we can use the swap maps σ : |iji 7→ |jii to

 
c00 c01
avoid explicit use of basis vectors yielding Applying f := to the qubit basis yields
c10 c11
(1Q ⊗ (|Ψi ihΨi |) ⊗ 1Q ) (|ΨBell i ⊗ |ΨBell i)
nXo | 0i 7→ c00 · | 0i + c10 · | 1i and | 1i 7→ c01 · | 0i + c11 · | 1i
= (1Q⊗Q ⊗ Ui ⊗ Ui ) Uswap (|ΨBell i ⊗ |ΨBell i)
hence when we apply f ⊗ 1Q we have
where
| 00i 7→ c00 ·| 00i+c10 ·| 10i and | 11i 7→ c01 ·| 01i+c11 ·| 11i
Uswap := (1Q⊗Q ⊗ σ)(1Q ⊗ σ ⊗ 1Q ) .
so f ⊗ 1Q applied to the Bell-state again yields
— it should be clear that it becomes quite problematic to write
things down nicely! When we consider only the ‘bra-part’ of | 00i + | 11i 7→ c00 · | 00i + c01 · | 01i + c10 · | 10i + c11 · | 11i.
the projector we obtain
The role which the Bell-state plays in this pseudo-operational
(1Q ⊗ hΨi | ⊗ 1Q ) (|ΨBell i ⊗ |ΨBell i) = (1Q ⊗ Ui )|Ψi i
correspondence is of course due to the fact that in the for-
i.e. the first and fourth qubit become entangled while no en- mal correspondence, the Bell-state corresponds to the Tiden-
tanglement on the middle qubits have been created, any any tity. Hence a bipartite state Ψf represented by a matrix f can
entanglement on these qubits actually has been destroyed. always be written down in two distinct pseudo-operational
manners

5 The structure of entanglement (f T ⊗ 1Q )|ΨBell i = |Ψf i = (1Q ⊗ f )|ΨBell i.

We already saw that we can obtain the Bell-basis by acting We invite the reader to make a picture of this. Another useful
with the Bell-matrices on the second qubit on a Bell-state. property follows from
But actually we have much more.
(f1 ⊗ f2 ) ◦ (g1 ⊗ g2 ) = (f1 ◦ g1 ) ⊗ (f2 ◦ g2 )

5.1 Map-state duality and compositionality as in Exercise 3.23. We have

Acting on the second qubit of a Bell-state with any linear op- (f ⊗ id) ◦ (id ⊗ g) = (f ◦ id) ⊗ (id ◦ g)
erator f exactly yields the bipartite state encoded by the trans-
= (id ◦ f ) ⊗ (g ◦ id)
posed to the matrix of that operator in our matricial definition
of the tensor product, providing the linear map-bipartite state = (id ⊗ g) ◦ (f ⊗ id).
correspondence with a true (pseudo-)operational significance
— we say pseudo-operational since in general f is not unitary Again we invite the reader to make a picture of this.
20
Logic-gate teleportation [15]. So what happens if in the This procedure is called logic-gate teleportation and turns
teleportation protocol we decide not to start with a Bell-state out to be a universal quantum computational primitive. Of
but with some other entangled state Ψf with matrix f T ? Ex- course, logic gates rarely commute, but in fact it suffices to
plicitly, since for ordinary teleportation the input state was find unitary operations {Ũi }i such that

| ψi ⊗ |ΨBell i Ũi ◦ f = f ◦ Ui
now the input state is for each Bell-matrix, then applying Ũi† to the last qubit we
| ψi ⊗ Ψf = | ψi ⊗ ((1Q ⊗ f )|ΨBell i) do obtain f |ψi. In fact, with some appropriate acrobatics one
can even go beyond the above considered situation [15].
= (1Q | ψi) ⊗ ((1Q ⊗ f )|ΨBell i)
 
= (1Q ⊗ 1Q ⊗ f ) | ψi ⊗ |ΨBell i
  Swapping from teleportation. In fact, entanglement swap-
= (1Q⊗Q ⊗ f ) | ψi ⊗ |ΨBell i ping can be seen as a consequence of teleportation. For tele-
portation we have
For ordinary teleportation the four ‘cases’ corresponding to   
the possible outcome states of the Bell-basis measurement we (|Ψ i ihΨi |) ⊗ 1Q | ψi ⊗ |ΨBell i = |Ψi i ⊗ Ui |ψi
needed to concider were
   hence in particular, (i) considering |ψi := | ji, (ii) apply-
(|Ψi ihΨi |) ⊗ 1Q | ψi ⊗ |ΨBell i ing | ji ⊗ − to both sides of the equation, and (iii) using the
(◦, ⊗)-exchange properties, we obtain
for each Bell-basis vector |Ψi i, so now we have
  
1Q ⊗ (|Ψi ihΨi |) ⊗ 1Q | jji ⊗ |ΨBell i
   
(|Ψi ihΨi |) ⊗ 1Q (1Q⊗Q ⊗ f ) | ψi ⊗ |ΨBell i
   = | ji ⊗ |Ψi i ⊗ Ui | ji,
= (1Q⊗Q ⊗ f ) (|Ψi ihΨi |) ⊗ 1Q | ψi ⊗ |ΨBell i .
hence adding for j = 0, 1 yields
But we know from our study of teleportation that
X  
   1Q ⊗ (|Ψi ihΨi |) ⊗ 1Q | jji ⊗ |ΨBell i
(1Q⊗Q ⊗ Ui† ) (|Ψi ihΨi |) ⊗ 1Q | ψi ⊗ |ΨBell i , j=0,1
X
where Ui is the ith Bell-matrix, yields the state |ψi for that = | ji ⊗ |Ψi i ⊗ Ui | ji
third qubit, hence j=0,1
  
(|Ψi ihΨi |) ⊗ 1Q | ψi ⊗ |ΨBell i
P
and by linearity and |ΨBell i = j=0,1 | jji we indeed ob-
tain entanglement swapping. Again, analyzing this calcula-
yields Ui |ψi for the third qubit and hence tion in a picture can be very instructive. Compositionality
   in computer science means breaking a big problem down in
(1Q⊗Q ⊗ f ) (|Ψi ihΨi |) ⊗ 1Q | ψi ⊗ |ΨBell i
smaller, hopefully already known ones. Above we did this:
yields f Ui |ψi for the third qubit. Assuming that f and Ui we derived logic-gate teleportation and entanglement swap-
ping from the teleportation protocol. On the other hand, the
would commute, it then suffices to apply Ui† to the last qubit
following protocol which is sometimes (wrongly?) referred
to obtain f |ψi i.e. we teleported the state and at the same time
to as some kind of converse to teleportation does have a nice
applied a (possibly unknown) operation f to this state |ψi.
conceptual derivation in terms of map-state duality.

Ui Superdense coding [10]. It is our aim to use one quan-

tum bit to communicate two classical bits i.e. in some way
a kind of converse to quantum teleportation. We start with
two qubits in a Bell-state, the two parties involved in the pro-
tocol each possessing one of the two qubits. Depending on
Pi which pair of classical bits we want to communicate one ap-
plies one of the four Bell-matrices to the first qubit which is

Ψf
then sent to the other party. The other party then performs a
Bell-measurement on the pair of qubits and the outcome to
that measurement reveals the encoded two bits.
21
 discussed protocols.
(hΨf | ⊗ 1Q )(1Q ⊗ |Ψg i)
nXo
P i nXo
= ((hΨBell |(1Q ⊗ f † )) ⊗ 1Q )(1Q ⊗ ((1Q ⊗ g)|ΨBell i))

= (hΨBell |⊗1Q )(1Q ⊗f † ⊗1Q )(1Q⊗Q ⊗g)(1Q ⊗|ΨBell i)

nXo
T = (hΨBell |⊗1Q )(1Q⊗Q ⊗g)(1Q ⊗f † ⊗1Q )(1Q ⊗|ΨBell i)
U i nXo
= g(hΨBell | ⊗ 1Q )(1Q ⊗ f † ⊗ 1Q )(1Q ⊗ |ΨBell i)
nXo
= g(hΨBell | ⊗ 1Q )(1Q ⊗ |Ψf¯i)
nXo
To verify this protocol it suffices to recall that if we apply any = g(hΨBell | ⊗ 1Q )(1Q⊗Q ⊗ f¯)(1Q ⊗ |ΨBell i)
of the Bell-matrices to the second qubit then we obtainnthe Xo
corresponding Bell-state, which we then will observe in the = (g ◦ f¯)(hΨBell | ⊗ 1Q )(1Q ⊗ |ΨBell i)
Bell-basis measurement which follows.
— following this calculation is quasi impossible without the
support of a picture. This result is quite intriguing: since
5.2 The logic of bipartite entanglement
(hΨBell | ⊗ 1Q )(1Q ⊗ |ΨBell i)
Recalling that a bipartite state Ψf represented by a matrix f T is just an instance of the teleportation protocol, namely condi-
can always be written down either as tioned on the fact that the measurement on the first two qubits
yields the Bell-state, so
(f T ⊗ 1Q )|ΨBell i = |Ψf i = (1Q ⊗ f )|ΨBell i.
(hΨBell | ⊗ 1Q )(1Q ⊗ |ΨBell i) = 1Q ,
Using the property hence we obtain

(f ⊗ g)† = f † ⊗ g † (hΨf | ⊗ 1Q )(1Q ⊗ |Ψg i) = g ◦ f¯.

which can be represented in a picture as:
which straightforwardly follows from the ‘pointwise’ defini-
tion of −⊗− on linear operators, we can now take the adjoint
of these resulting in Ψf g
 †
hΨf | = (f T ⊗ 1Q )|ΨBell i = hΨBell |(f¯ ⊗ 1Q )
= f
Ψg
with f¯ the conjugate of f and
 † What is particularly interesting about this expression is the
hΨf | = (1Q ⊗ f )|ΨBell i) = hΨBell |(1Q ⊗ f † ). fact that while on the left, respecting compositional order, we
first have an expression involving g and then one involving f ,
All this results in four alternative representations for a bipar- on the right we first have f¯ and only then g. This seems as if
tite projector, namely there would be some weird reversal in the causal order!
nXo
Pf = |Ψf ihΨf | = (1Q ⊗ f )|ΨBell ihΨBell |(f¯ ⊗ 1Q ) Exercise 5.1 Show that we also have
nXo (1Q ⊗ hΨg | ⊗ 1Q )(|Ψf i ⊗ |Ψh i) = Ψh◦ḡ◦f
= (1Q ⊗ f )|ΨBell ihΨBell |(1Q ⊗ f † )
nXo which can be represented in a picture as:
= (f T ⊗ 1Q )|ΨBell ihΨBell |(f¯ ⊗ 1Q )

Ψg
nXo
= (f T ⊗ 1Q )|ΨBell ihΨBell |(1Q ⊗ f † )

=
And again, to get a feel for what this actually stands for, you
Ψf Ψh Ψh o go f
want to represent these formulae in a picture. In fact, we
can now formulate the structural crux behind all of the above
22
 which moreover provides a straightforward extension of the
Exercise 5.2 [BA exam 2006] Four qubits are such that the proof of logic-gate teleportation (and hence teleportation it-
first and second are in the joint state a · |00i + b · |11i and the self) beyond the case of qubits. We proceed in two steps:
third and fourth are also in the joint state a · |00i + b · |11i.
Then we perform a Bell-basis measurement on the second and 1. First we show that we have
third qubits. i. For each of the possible outcomes of the mea-
surement, assuming a, b ∈]0, 1[, what is the resulting state of (hΨBell | ⊗ 1H )(1H ⊗ |ΨBell i) = 1H (2)
the qubits? (Write these states in Dirac notation in the com-
putational basis.) ii. Can you find values for a and b such where ΨBell stands for the state corresponding to the
that after applying some well-chosen unitary corrections the identity matrix on H i.e.
states of the qubits do not depend on the measurement out- X
come anymore? (Explicitly give these corrections.) iii. What ΨBell := | iii.
is the probability for the second and the third qubit to end up i

in a Bell-state (when doing no corrections)? for any dimension of the underlying Hilbert space.

2. Next, given a family of n := dim(H) unitary operators

Exercise 5.3 We wish to design a slightly more sophisticated {Ui : H → H}i
version of entanglement swapping involving six qubits, and
yielding the passage from Bell-state entanglements which are such that the states

{|ΨUi i : C → H}i

are mutually orthogonal, by the above we know that

to Bell-state entanglements (hΨUi | ⊗ 1Q )(1Q ⊗ |Ψf i) = f ◦ Ūi

to which we can now apply the reasoning we made above

for qubit-gate teleportation.

We can push all this a bit further within the domain of the
that is, using a different geometry, from absurd, as is illustrated by the following exercise.

Exercise 5.4 Assume we start with three qubits in a state

|ψi ⊗ |Ψf1 i .

To these we will apply five measurements, respectively in-

to cluding the projectors

Pf2 Pf3 Pf4 Pf5 Pf6

which we assume to have taken place in these five measure-

ments, and

Can you come up with a protocol which does that? • first Pf2 takes place on the first two qubits,

• next Pf3 takes place on the last two qubits,

Correctness of the logic-gate teleportation protocol is in
• next Pf4 takes place on the first two qubits,
fact fully captured by the slightly elaborated variant3
• next Pf5 takes place on the last two qubits,
(Pf ⊗ 1Q )(1Q ⊗ |Ψg i)
= (1Q⊗Q ⊗ (g ◦ f¯))((|Ψf ihΨBell |) ⊗ 1Q )(1Q ⊗ |ΨBell i) • finally Pf6 takes place on the first two qubits.
= |Ψf i ⊗ (g ◦ f¯)
What is the state of the third qubit after all this?
3
In course-notes v.7 there was a clumsy typo here.

23
5.3 Quantifying entanglement Since we know that bipartite states are in bijective correspon-
dence with linear maps we can now use the above result to
Clearly, the Bell-state is more entangled than any (separable) classify and quantify entanglement. For any bipartite state
tensor |ψi⊗|φi. But what exactly do we mean by ‘being more we have
entangled’? There are many different proposals for a precise
conception of this, but consensus has yet to be reached in the |Ψf i = (1H ⊗ f )|ΨBell i = (1H ⊗ (U ◦ g))|ΨBell i.
research community. One way to attack this problem is to ask
†
to which extent an entangled state enables typical quantum Applying U to the last qubit we obtain
phenomena, such as non-local correlations and teleportation
(1H ⊗ U † )|Ψf i = (1H ⊗ g)|ΨBell i
and how ‘efficient’ we can do that. For example, given a Bell-
state we can do full teleportation, with as crux i.e. we can always undo the effect of the unitary component,
(hΨBell | ⊗ 1Q )(1Q ⊗ |ΨBell i) = 1Q . and we can do that in a reversible manner. Hence it is the
(unique) positive component g which determines the ‘degree
On the other hand of entanglement’. But since g is itself self-adjoint it admits
(hΨBell | ⊗ 1Q )(1Q ⊗ |ψ ⊗ φi) diagonalization i.e. can be written as
= (hΨBell |(1Q ⊗ |ψi)) ⊗ |φi
. = |φihψ̄| g = U′ ◦ h ◦ U′
†

i.e. the third qubit will end up in the state |φi independent with h diagonal, so using Subsection 5.2 we obtain
of what the initial state of the first qubit was. In general we
would end up somewhere between these two extremes of ‘full (U ′T ⊗ (U ◦ U ′ )† )|Ψf i = (1H ⊗ h)|ΨBell i
teleportation’ and ‘no teleportation’. To understand this better
we need some more linear algebra. so we can again reversibly undo the effect of the unitaries,
so the ‘degree of entanglement’ for bipartite states is now re-
A linear operator f : H → H is positive iff
duced to diagonal positive matrices i.e. a list of n positive
†
a. it can be written as f = g ◦ g for some other linear reals. Again using map-state duality, but now in the converse
operator g : H → H′ . direction, we straightforwardly obtain the following.

Each positive operator is obviously always self-adjoint, hence Proposition 5.8 Each bipartite state Ψ ∈ H ⊗ H′ admits a
its eigenvalues will always be real. But we have more: ‘Schmidt decomposition’ i.e. it can be written as
X
Proposition 5.5 A linear operator f : H → H is positive if Ψ= ri · ei ⊗ e′i with {ri }i ⊆ R+
and only if, equivalently, i

b. for all ψ ∈ H we have that hψ|f |ψi ∈ R+ ; for some well-chosen ONBs {ei }i of H and {e′i }i of H′ .
c. f admits a self-adjoint square-root i.e. f can be decom- So for the qubit case the problem is reduced to comparing a
posed as f = g ◦ g with g self-adjoint. pair of real numbers ‘up to a real number’ (cf. normalization)
i.e. a one-dimensional problem. Extreme cases are (1, 0) (and
Exercise 5.6 Prove Proposition 5.5 — you can proceed by equivalently (0, 1)) which corresponds to pure tensors ψ ⊗ φ
respectively showing that a ⇒ b, a&b ⇒ c, c ⇒ a where while (1, 1) captures for example both the EPR-state and the
for proving a&b ⇒ c you can use the fact that positive oper- Bell-state. In general, we will say that (a, b) for a ≥ b capture
ators are self-adjoint and hence all diagonalize in some basis. ‘more entanglement’ than (c, d) for c ≥ d iff
a c
Next we show that each linear opertor factors into a unitary ≤ .
b d
one and a positive one, something which is known as the po-
lar decomposition of linear operators. For normalized coefficients a2 + b2 = c2 + d2 = 1 this boils
down to a ≤ c or equivalently d ≤ b, or again equivalently
Theorem 5.7 Each linear operator f can be written as a2 ≤ c2 or d2 ≤ b2 . So we can now assign a quantitative
f = U ◦ g = g′ ◦ U measure to bipartite qubit states by setting

where U is unitary and g and g ′ are both positive, and in Ψ 7→ (a, b) 7→ 2 · b2

′
particular, g and g are uniquely determined. Explicitly,
p p which is 0 whenever the state is disentangled and one when-
g = f† ◦ f and g′ = f ◦ f † . ever it is ‘maximally entangled’ e.g. in the Bell-state.
24
 For higher dimensions the problem becomes more delicate, 5.4 Trace from Bell-states
but a very natural candidate is the so-called majorization or-
der [17] which is defined as follows. Let (a1 , . . . , an ) and For a linear operator f : H → H there exists a unique scalar
(b1 , . . . , bn ) be decreasing lists of positive reals such that X X
tr(f ) = hi |f | ii = mii
i i
X X
ai = bi = 1 .
i i which turns out to be independent of the choice of the ba-
sis. From Exercise 5.10i. it indeed folows that the basis-
We say that dependent expression can be rewritten as a basis-independent
one.
(a1 , . . . , an ) ≥ (b1 , . . . , bn )
Exercise 5.10 i. Show for that
if and only if
tr(f ) = hΨBell |(1H ⊗ f )|ΨBell i
a1 ≤ b1 a1 +a2 ≤ b1 +b2 a1 +a2 +a3 ≤ b1 +b2 +b3 etc.
where we considered the generalized Bell-state
So given a bipartite state with {ri }i we can consider the de- X
2
creasing list (riP 2 2
≥ ri2 ≥ . . . ≥ rin ), since for normalised Ψ Bell := | iii ∈ H ⊗ H ,
1
i
states we have i ri2 = 1, and by comparing these lists we
can compare their degree of entanglement. It can be shown (with norm pdim(H) !), that is, in a picture,
that this order has an operational physical meaning [18]:

• One bipartite state Ψ can be converted into another bi- f

partite state Φ by means of local operations and classical
communication iff Ψ is above Φ w.r.t the above order
i.e. iff Ψ is more entangled than Φ.
ii. Show that

hΨBell |(1H ⊗(f ◦g))|ΨBell i = hΨBell |(1H ⊗(g ◦f ))|ΨBell i .

Exercise 5.9 i. Show that the majorization order is a partial
order (= reflexive, anti-symmetric, transitive) on decreasing Can you represent this equation in a picture? iii. For a linear
n-lists of positive reals, and hence induces a pre-ordering (= operator c : C → C we clearly have that tr(c) = c. Can you
reflexive, transitive) on bipartite entangled states. ii. Draw a use part (ii) of this exercise to conclude from this that
Hasse-diagram for the bipartite states
tr(P ◦ ρψ ) = hψ|P|ψi
r r
1 3 1
|00i √ (|00i + |22i) |11i + |22i where ρψ := |ψihψ|?
2 4 4
Exercise 5.10iii. exposes yet another variant on the rule for
1 1 1 1 calculating probabilities. The advantage of this rule is that
√ |00i + √ |11i + √ |22i √ (|00i + |11i + |22i) .
4 2 4 3 it extends to mixed states which we will discuss in the last
section of this course. Another major advantage of the rep-
1
(|00i + |01i + |10i + |11i) resentation of the trace as in Exercise 5.10 is that it enables
2 a straightforward definition of the partial trace, which, given
That is, if a state is above another in the majorization order, a linear map f : H ⊗ H1 → H ⊗ H2 is defined as a linear
then you depict it above it and connect them by a line, but if operator of type
a ≤ b ≤ c then you only draw a line between a&b and b&c,
but not between a&c. iii. Characterise those states which are trHH1 ,H2 (f ) : H1 → H2
either maximal or minimal in the majorization order
with as prescription

All this is restricted to bipartite states. Understanding multi- trHH1 ,H2 (f ) := (hΨBell | ⊗ 1H2 )(1H ⊗ f )(|ΨBell i ⊗ 1H1 )
partite entanglement is still very much an open problem both
with respect to quantitative and qualitative understanding. i.e., in a picture,
25
 6.1 Special gates

The the ‘controled not’ or CNOT-gate is an important gate in

f the quantum computing literature is because of its entangling
capabilities. Whenever the first qubit is in state | 0i it doesn’t
alter the second qubit, but if the first qubit is in state | 0i then
the third Bell-matrix is applied to the second qubit, hence

Exercise 5.11 Show that for | 0i i 7→ | 0i i | 1i i 7→ (| 1i ⊗ U2 | i i)

f : H → H1 and g : H → H2 where U2 is the third Bell-matrix or NOT-gate. The main

application of this gate is preparation of Bell-states. We have
we have
CNOT ((| 0i + | 1i) ⊗ | 0i) = | 00i + | 11i
trH †
H1 ,H2 (|Ψg ihΨf |) = g ◦ f .
and more general we have
Can you represent this equation in a picture?
CNOT ((c0 · | 0i + c1 · | 1i) ⊗ | 0i) = c0 · | 00i + c1 · | 11i .

σ
Exercise 5.12 [BA exam 2006] Let PGHZ : Q⊗Q → Q⊗Q Exercise 6.1 Define CNOT as the gate obtained by ex-
be the projector which projects on the ray spanned by the state changing the role played by the first and the second qubit.
What is the effect of first applying CNOT, then CNOTσ , and
ΨGHZ := |000i + |111i ∈ Q ⊗ Q ⊗ Q then again CNOT to a pair of qubits?

i. Compute
Exercise 6.2 Consider the following eight equations:
f := trQ
Q⊗Q,Q⊗Q (PGHZ ) : Q ⊗ Q → Q ⊗ Q .
CNOT ◦ (Ui ⊗ 1Q ) = (ξ ⊗ ξ ′ ) ◦ CNOT
i.e. provide the matrix of f in the computational basis. ii. As- CNOT ◦ (1Q ⊗ Ui ) = (ξ ⊗ ξ ′ ) ◦ CNOT
sume now that in the above we substitute ΨGHZ ∈ Q⊗Q⊗Q
by any arbitray state Ψ ∈ Q ⊗ H. Prove positivity of where Ui can be any of the four Bell-matrices. Verify for each
of these eight equations whether there exist operations ξ and
Q
f := trH,H (P) : H → H ξ ′ such that it holds. What can you conclude from this for
logic-gate teleportation of a CNOT-gate?
by showing that it always factors as f = g † ◦g for some linear
operator g : H → Q. Proposition 6.3 Each n-qubit gate can be obtained by com-
postion, tensor, 1-qubit gates and the CNOT-gate.

Proof: See [17] pp.191–194. ✷

6 Algorithms and gates
Other important gates are
We now go over to the standard quantum computing stuff,
     
1 1 1 0 1 0
a story about trying to design quantum algorithms by play- H := 1 −1
S :=
0 i
T :=
0 ei 4
π

ing around with quantum logic gates. In order to compare

the complexity of quantum algorithms with those of classical respectively called Hadamard-, phase-, and Tofoli-gate.
algorithms we need to have a particular model for quantum
computing which straightforwardly compares to a model for Proposition 6.4 Each n-qubit gate can be approximated
classical computing. This model is the so-called circuit model with arbitrary accuracy using compostion, tensor, the H-
or gate-array model basisd on the following three steps: gate, the S-gate, the T -gate and the CNOT-gate — i.e. the
difference in probabilities when performing measurements af-
ter applying the gates can be kept arbitrary small, explicitly
preparation ❀ logic gates ❀ measurement.
hΨ|U † ◦ P ◦ U |Ψi − hΨ|Uapprox
†
◦ P ◦ Uapprox |Ψi
Comparing algorithms is now a matter of comparing the num-
ber of gates that need to be applied. can be kept arbitrary small for all Ψ and P.
26
Proof: See [17] pp.194–197. ✷ We are now all set to exploit quantum parallelism:4

In the standard literature, rather than the Bell-matrices, the Uf (−, | 0i) :: | 0i + | 1i 7→ | 0 f (0)i + | 1 f (1)i
Pauli-matrices occur, that is
  i.e. with one execution of Uf we actually obtain the image
0 1 under f both for 0 and 1. This idea moreover easily extends
X := U2 =
1 0 to functions f : Bn → B, setting
 
0 −i Uf : Q⊗n ⊗ Q → Q⊗n ⊗ Q :: | i ji 7→ | i (j + f (i) mod 2)i ,
Y := i U3 =
i 0
  and then considering
1 0
Z := U1 =
0 −1 X X
Uf (−, | 0i) :: | ii 7→ | i f (i)i .
which of course do not have exactly same the ‘nice’ corre- i∈B n i∈Bn

spondence with the Bell-basis as the Bell-matrices have. But

they are self-adjoint and give rise to a group, the so-called But there is a major problem! While our state encodes all pos-
Pauli group. In the representation of a qubit as a sphere, they sible argument/image pairs (or if you prefer I/O-pairs), this
represent a 180◦ rotation respectively around the X-axis, the data is not accessible. Indeed, when we measure in the basis
Y -axis and the Z-axis. n o
| 0i , . . . , | 2n+1− 1i

6.2 The Deutch-Jozsa algorithm then the outcome states with non-zero probability are
n o
The aim of this ‘pedagogical’ example is to illustrate how | i f (i)i
easy it is try to exploit quantum parallelism, but how hard it is i∈Bn

to actually succeed. Consider a Boolean function f : B → B but measurement

where B := {0, 1}. Since in general f is not injective, hence
not reversible, we extend it in a way that after its execution • will only expose one of these, and,
we still ‘remember’ the argument, by setting
• destroy all the other components in the superposition.
f˜ : B × B → B × B :: (i, j) 7→ (i, j + f (i) mod 2) .
So the only thing we achieved so far is introducing random-
The function f can be recovered as f (i) = π2 (f˜(i, 0)) where
ness concerning which outcome we are actually calculating,
π2 : B × B → B is the second projection, while its argument
which we could as well have done by flipping a coin. And
can be retained as i = π1 (f˜(i, 0)). In fact, we produce pairs
in fact, this problem can not be overcome i.e. we will never
f˜(−, 0) : B × B → B :: i 7→ (i, f (i)) . be able to extract all the desired data from the superposition
state. Does this mean our endeavor was a waste of time?
consiting of both the argument and the image. The second Actually, as David Deutch and (later) Richard Jozsa
argument of f˜ enables it to be bijective: we can rewrite f˜ as showed, while we are not able to extract more than one argu-
( ment/image pair from the superposition state it turns out that
(i, j) 7→ (i, j) f (i) = 0 alternatively we can extract a particular bit of data from it,
f˜ ::
(i, j) 7→ (i, σ(j)) f (i) = 1 encoding a certain property of f , which classically would re-
quire knowledge of many argument/image pairs, in the worst
where σ : B → B permutes 0 and 1, from which bijectiv- case n + 1 pairs. For a function f which is either
2
ity clearly follows either by considering cases, or by putting
everything in one line as • constant i.e. for all arguments the image is the same,
f˜ :: (i, j) 7→ (i, σ f (i) (j)) . • balanced i.e. the number of 0- and 1-images are equal,

But besides enabling bijectivity, the second input seems to be we will be able to verify with certainty which of the two it is.
of no use at all, but as we will see further it will turn out to Let us consider
play a crucial role. Next we consider the by f˜ induced unitary
permutation of basis vectors Uf (−, | 0i − | 1i) : Qn → Qn ⊗ Q

Uf : Q ⊗ Q → Q ⊗ Q :: | i ji 7→ | i (j + f (i) mod 2)i . 4

Note that a matrix representation is not very useful in this case.

27
(so we exploit the 2nd input!) that is5
P
• If f is constant we obtain i | ii
with certainty.
P
| ii 7→ | i (0 + f (i) mod 2)i − | i (1 + f (i) mod 2)i • If f is balanced we cannot obtain i | ii.

and since ( Since the number of required evaluations of f (encodes as Uf )

0 + f (i) mod 2 = f (i) is classically proportional to 2n , while in the quantum case a
1 + f (i) mod 2 = 1 − f (i) single evaluation suffices, we have a true example of substan-
we obtain tial algorithmic speed-up. On the other hand, the striking arti-
  ficiality of this example confirms the hardness of the quantum
| ii 7→ | ii ⊗ | f (i)i − |1 − f (i)i informatic endeavor and the need for better/high-level meth-
ods to study quantum algorithms.
yielding, for f (i) = 0 and f (i) = 1 respectively,

| ii 7→ | ii ⊗ (|0i − |1i) and | ii 7→ | ii ⊗ (|1i − |0i) , The text book version of the facts. Typically e.g. [17] one
extends the presentation by making also the preparation of
in short, states explicit, as well as the generation of arbitrary measure-
| ii 7→ (−1)f (i) | ii ⊗ (|0i − |1i) . ments using some ‘standard’ logic gates and measurement in
For the simple f : B → B case this becomes the computational basis. For creating superpositions within
  the computational basis we apply the so-called Hadamard
| 0i + | 1i 7→ (−1)f (0) | 0i + (−1)f (1) | 1i ⊗ (|0i − |1i) . gate  
1 1
H :=
In this case constant means f (0) = f (1) yielding 1 −1
  to first basis vector yielding
Uf (| 0i + | 1i) ⊗ (| 0i − | 1i) = ±(| 0i + | 1i) ⊗ (| 0i − | 1i)
H✲
while balanced means f (0) 6= f (1) yielding | 0i | 0i + | 1i
 
Uf | 0i + | 1i) ⊗ (| 0i − | 1i) = ±(| 0i − | 1i) ⊗ (| 0i − | 1i). which we can generalize to
H ⊗n
✲
X
Hence it suffices to measure the first qubit in the basis | 0i⊗n (| 0i + | 1i)⊗n = | ii .
i∈Bn
{| 0i + | 1i , | 0i − | 1i}
To get the input which induces inference we set
to achieve our goal. Crucial is the fact that we obtained two
mutually orthogonal states | 0i + | 1i and | 0i − | 1i, each rep- H✲
| 1i | 0i − | 1i .
resenting one of the two alternatives we wish to distinguish.
In the general case we obtain Measuring in the basis
X X 
| ii 7→ (−1)f (i) | ii ⊗ (|0i − |1i) {| 0i + | 1i , | 0i − | 1i}
i i
is achieved by first applying a Hadamard gate and then mea-
that is, for f constant, suring inPthe computational basis, while a measurement in-
X   X  cluding i∈Bn | ii as an outcome state is achieved by apply-
Uf | ii ⊗ (| 0i − | 1i) = ± | ii ⊗ (| 0i − | 1i) .
ing H ⊗n before a measurement in the computational basis.
i i

f (j) ∈ {0, 1} we have

P
For a state j ǫj | ji with P ǫj := (−1) 6.3 Grover’s search algorithm
that it is orthogonal to i | ii if and only if
X X This algorithm
√ searches an unsorted databasis of size N
0= ǫj hi | ji = ǫj ,
in O( N ) time while classically this takes O(N ) time.
ij j
Grover’s is provably the fastest quantum algorithm that does
that is, exactly when f is balanced. Hence the case of a con- search. Consider a boolean function
stant f can be distinguished by measurement from that of a
balanced f , when performing a measurement on the first n f : Bn → B
P
qubits which includes i | ii as an outcome state:
which assigns 1 to the searched entry and 0 else. Search boils
5
Note that | i + ji =
6 | ii + | ji e.g. | 1i = | 0 + 1i 6= | 0i + | 1i. down to evaluating such a function for different input values
28
until the right one is found. Referring to the original title of 6.4.1 Period finding
Lov Grover’s paper [14] we may consider
Given a function
haystack → {needle, straw} . f : Bn → Bm
we intend to find its period. To do this we will first produce a
We can also calculate the inverse to f and then f −1 (1) is the big state which encodes all input-output pairs i.e.
desired value. This is what Grover’s algorithm does. The X
intuition behind this algorithm is purely geometrical. Let | α f (α)i ∈ Bn ⊗ Bm
ω := f −1 (1) and set α∈Bn

and as we know from the previous section, we can produce

(
⊗n ⊗n
+|ωi 7→ −|ωi
Uf : Q → Q :: such a state by a single unitary operation. The next step is a
+| i i 7→ +| i i (i 6= ω) so-called discrete quantum Fourier transform
Let ǫ := √1 (|0i + |1i)⊗n and set X n
2n | αi 7→ ei2παβ/2 | βi
β∈Bn
Uǫ = 2|ǫihǫ| − 1Q⊗n

Consider the plane spanned by |ωi and |ǫi and let |ωi⊥ be the performed on the first n qubits, which can shown both to ad-
orthocomplement to |ωi in this plane. When restricting to this mit an inverse and to be unitary. Hence our state now becomes
plane one easily verifies that: X X n
ei2παβ/2 · | β f (α)i ∈ Bn ⊗ Bm .
α∈Bn β∈Bn
• Uf ∼ refection against ray(|ωi⊥ )

• Uǫ ∼ refection against ray(|ǫi) Remarkably, we are now done and it suffices to measure the
state of the first n qubits to ‘with a high probability’ find the
• Uǫ ◦ Uf ∼ 2θ rotation towards |ωi for period of the function! Indeed, suppose that f has period ω
so we have for all α that
1
sinθ = hǫ|ωi = √
2n f (α + ω) = f (α) .
Hence for r rounds with Now fix a value for β and we are interested in the probability
π
π θ −2 of obtaining this outcome in a measurement of the first qubit,
− θ = 2θr i.e. r =
2 4 so we are interested in the weight of the term
the vectors |ωi and (Uǫ ◦ Uf )n (|ǫi) become ‘almost’ aligned. X n
X n
ei2παβ/2 · | β f (α)i = | βi ⊗ ei2παβ/2 · | f (α)i
Hence the protocol: n n
α∈B α∈B

• The input state is ǫ. which, for | βi normalized, is determined by the length of

π
−2
• Apply Uǫ ◦ Uf “the closest integer to θ
” times. n
X
4 ei2παβ/2 · | f (α)i.
• Measure in computational basis. α∈Bn

Since f is periodic, several values of α, namely

With very high probability we will obtain the answer. The
π
−2
probability arises from the fact that θ 4 is not an integer. α , α + ω , α + 2ω , α + 3ω , . . .
This probability increases with N since θ decreases. For N =
2n >> 1 we have θ ≃ sinθ = √1N so will contribute to the coefficient of same term | f (α)i. Since
sometimes these coifficients are positive and sometimes they
π √
− 2 are negative, averagely they will more or less annihilate each
r= θ ∼ O( N ) .
4 other giving that term a very small weight. However, if

2n
6.4 Shor’s factoring algorithm β =k· for k∈N
ω
This was the first quantum algorithm, which, if it could be then the components contributing to the same term have
efficiently implemented on a quantum computer, would have
unavoidably an important impact. ei2kπα/ω , ei2kπ ei2kπα/ω , ei2·2kπ ei2kπα/ω , ei3·2kπ ei2kπα/ω , . . .
29
as coefficients, which by eil·2kπ = 1 for l ∈ N are in fact all period finding algorithm, while the ‘classical part’ of his al-
n
the same. That is, for β = k · 2ω we get constructive interfer- gorithm relates period-finding to factoring, more specifically,
ence, hence will have a very high probability, and in principle to finding a single factor.
we could then compute ω if it wasn’t for the presence of k. So we wish to factor a number N0 := N . First choose
To find the actual period with high probability it turns out that
a number N1 < N0 and we use Euclid’s algorithm to
that it suffices to repeat the above roughly log log ωn times. look for common factors i.e. we perform a number of divi-
The bad thing is that we only obtain a probabilistic result. sions Ni /Ni+1 yielding quotients q1 , q2 , . . . qx and remain-
The good thing is that with a growing number of qubits (i.e. n) ders N2 , N3 , . . . , Nx−1 , 0 i.e.
the number of computations (i.e. for each of the possible input
values α ∈ 2n of f ) and the corresponding stored data grows N 0 = q1 · N1 + N2
exponentially, and it turns out that for a ‘small computer’ with N1 = q2 · N2 + N3
only 270 qubits we would actually have performed more com- ..
.
putations and stored more results than the estimated number
of particles in the universe. Nx−2 = qx−1 · Nx−1 + Nx
Nx−1 = qx · Nx (+0)
6.4.2 Factoring and code-breaking
with the greatest common divisor being Nx . To see that
To quantify ‘how long’ it takes a certain algorithm to do a Nx is indeed a common divisor of N0 and N1 it suffices
task, it is natural to ask how the needed time increases with to substitute the last equation in the one just before yield-
the size of the input e.g. if we want to factor a large number ing Nx−2 = (. . . q . . .) · Nx , next substituting the two last
N , how does the required time grows when we let N grow. in the one before that yielding Nx−3 = (. . . q . . .) · Nx , etc.
Alternatively, we can measure the input in the number of re- So ultimately we will both obtain N1 = (. . . q . . .) · Nx and
quired (qu)bits, i.e. n = log2 N , enabling a direct comparison N0 = (. . . q . . .) · Nx . If this procedure yield a non-trivial
between classical and quantum algorithms. On conventional common devisor, we have our desired factor of N = N0 . On
computers the best known factoring algorithm runs in the other hand, if the greatest common devisor is 1 we have
established that N0 and N1 are co-prime.
 1 !!
64 3 1 2 Consider the sequence of N1 -powers modulo N0 i.e.
O exp (lnN ) 3 (lnlnN ) 3
9
f (0), f (1), f (2), . . . for f (α) = N1α mod N0 .
time, so the required time grows exponentially with the num-
ber of digits n = log2 N of the number we wish to factor. For ω the smallest number such that N1ω mod N0 = 1 this
As an example, in 1994 a 129 digit number was success- series becomes
fully factored on 1600 workstations in parallel in a period of
1, N1 , N12 , . . . , N1ω−1 , 1, N1 , N12 , . . . , N1ω−1 , . . .
8 months. The same computer setup would however require
800.000 years to factor a 250 digit number, and significantly Indeed, if
longer than the age of the universe to factor a 1000 digit num-
ber. Of course, computers do become faster, but everytime the N1ω mod N0 = 1 then N1ω − k · N0 = 1
speed doubles ‘we can just add a digit’ to maintain the above
mentioned ‘absurdly long’ required computation times. for some k ∈ N, so we have N1ω = 1 + k · N0 and hence

The hardness of factoring large numbers is crucial for pub- N1ω+l = N1l ·(1+k ·N0 ) = N1l +(k ·N1l )·N0 = N1l mod N0 .
lic key crypto-systems, e.g. those used in banks, of which the
secrecy typically relies on the assumed difficulty (not to say Thus, the sequence which we obtain is periodic, and we will
impossibility) to factor a number of approximately 250 dig- use a quantum computer to find this period ω in the way it
its. But in 1994 Peter Shor of AT&T proposed an algorithm was described in the previous subsection. If ω is even we
for factoring which on a gate-array-type quantum computer proceed as discussed below. If ω is odd we need to start over
would run in   again and pick another number N1 — this only happens in
O (log2 N )3 50% of the cases so not to many runs are needed in general.
Rewriting N1ω mod N0 = 1 as
time, which means that it is only polynomial in the number of
ω
qubits, and factoring a 250 bit number turns out to only take (N12 )2 − 1 = 0 mod N0
a few billion computational steps — e.g. Microsoft’s Xbox
360 game console does about a hundred billion in a second. we obtain ω ω
The ‘quantum part’ of his algorithm is the above discussed (N12 − 1)(N12 + 1) = k · N0
30
 ω
for some k ∈ N i.e. we obtain two factors (N12 − 1) and and Bob at its turn tells Alice in which basis he measured
ω
(N 2 + 1) of which the product is equal to some multiple them. They only retain the outcome-digits
1 ω ω
of N0 , so if N0 is not prime either (N12 − 1) or (N12 + 1) 0 ∼ | 0i 1 ∼ | 1i 0 ∼ | 0i + | 1i 1 ∼ | 0i − | 1i
should have a non-trivial factor in common with it. We can
extract this common factor using Euclid’s algorithm. So from for those states of which their basis match, and the resulting
two numbers N0 and N1 with no common factor we have string of bits is a secret shared only by the two of them.
build a pair N1′ and N1′′ of which at least one has a non-trivial The safety of this protocol follows from the fact that an
common factor with N0 whenever the latter is not prime, and eavesdropper intercepting the states cannot measure them
this procedure involved period-finding for which we can use without in 50% of the cases altering them, causing a very
a quantum computer. If this non-trivial common factor ends high number of mismatches in the key shared by Alice and
up being N0 itself we start over again with a different N1 — Bob, which can easily be detected by them if they compare
but there are more optimal ways to deal with this issue. a small number of their key-digits. So the crucial quantum-
feature which guarantees secrecy in this protocol is the fact
6.5 Quantum key distribution that measurements in general alter the states, in the case of
the protocol being in 50% of the cases.
So bye-bye to all our money in the banks? In fact, for the A mild variant of this scheme is the Ekert91 quantum key
positive-minded who beliefs that someday quantum comput- distribution protocol in which Alice and Bob now share Bell-
ers will be fact, there already is a solution available to the states. Each measures their Bell-states in a basis of their
above stated problem, a solution somewhat ironically also choice, and after having measured all their qubits they again
provided by quantum informatics as it was the case for the compare their basis, only retaining those outcome-digits for
problem itself, and in fact, you can already buy it online ei- which their basis match. In fact, the difference between BB84
ther at MagiQ: and Ekert91 essentially boils down to interpreting the ‘iden-
tity’ 1Q : QAlice → QBob either as:
http://www.magiqtech.com/
• effectively sending a qubit;
or at the Swiss basisd ID quantique:
• sharing a Bell-pair.

http://www.idquantique.com/

The BB84 quantum key distribution protocol is a simple pro- 7 Mixed states
tocol which goes as follows. There are two parties, namely
Alice and Bob. Alice prepares her qubits in either of the states Thus far we defined a state to be a ray in a Hilbert space. It
turs out to be useful to have a more general notion of state
| 0i | 1i | 0i + | 1i | 0i − | 1i which will enable us to describe:

in a randomly distributed manner, and sends them one by one 1. Situations where there is a (probabilistic) lack of com-
to Bob. Bob, in order to know their content has to choose plete knowledge on the actual state of a quantum system.
between measuring them in either of the basis
2. Large statistical ensembles of quantum systems.
{| 0i , | 1i} or {| 0i + | 1i , | 0i − | 1i} .
3. Subsystems of a bigger (entangled) quantum system.
Choosing ‘the wrong’ basis for measuring a qubit would of 4. Non-isolated (=open) quantum systems; decoherence.
course destroy its data and yield an outcome unrelated to its
actual initial state, and there is no way for Bob to know (with- All of these can be represented by the same mathematical ob-
out Alice’s help) whether an outcome reflects the true initial ject, namely a density operator, also referred to as a mixed
state or not. So Bob can’t do anything else but measuring the state. As compared to a pure state which is a ray in a Hilbert
qubits in a randomly picked basis of his choice and records space H, a density operator is defined to be a linear operator
for each of the qubits which basis he used and what the out- ρ : H → H which is:
come was. After all this is done, Alice tells Bob publicly for
all of the qubits she has send whether they either belonged to • positive (and hence self-adjoint);

{| 0i , | 1i} or {| 0i + | 1i , | 0i − | 1i} , • has trace equal to one.

31
We should now reformulate the axioms of quantum mechan- Conversely, all mixed states clearly arise in this way. The
ics with respect to this new generalised notion of state. transition under unitaries and measurement in Postulate 7.1
is also induced by this heuristics. Exactly the same argument
Postulate 7.1 [extension to mixed states] The state of a holds for statistical ensembles of pure states.
quantum system is a density operator ρ : H → H. Deter-
ministic transformations correspond to ρ 7→ U ◦ ρ ◦ U † where Exercise 7.3 Does a density matrix always uniquely repre-
U : H → H is unitary. Pure measurements P are described by
sent a particular set of pure states together with a correspond-
a set of projectors {Pi : H → H}i with i Pi = 1H 6 and ing set of probabilistic weights? If not, characterise all pairs
they cause a state transition of states (φ1 , φ2 ) and weights (ω1 , ω2 ) with

Pi ◦ ρ ◦ Pi  1
0

ρ 7→ 2
Tr(Pi ◦ ρ) 0 1
2

and this transition happens with probability Tr(Pi ◦ ρ). as density matrix. Can you find an example of (φ1 , φ2 , φ3 )
with ω1 = ω2 = ω3 = 13 which again is described by the
same density matrix? Repeat this question but now for
Exercise 7.2 Show that when setting ρ := |ψihψ| these pos-  1
  
tulates boil down to those for pure states. 4 0 1 0
3 and .
0 4 0 0
These postulates can in fact be derived from those for pure
states given the above heuristics. Consider a probabilistic
lack of knowledge on a set of states i.e. we have a fam- Now consider the situation that we have |Φi ∈ K ⊗ H. A
ily of pure states {ψi }i together with respective probabilis- measurement of H ‘alone’ is formally realised by considering
tic weights {ωi }i of the system actually being in that state. {1K ⊗ Pi }i where {Pi : H → H}i a measurement of H.
The probability for a certain outcome in a measurement is the Hence the respective probabilities are given by
weighted sum of the individual probabilities i.e.
X X hΦ|(1K ⊗ Pi )|Φi = hΨBell |(1K ⊗ (f † ◦ Pi ◦ f ))|ΨBell i
ωj hψj |Pi |ψj i = ωj Tr (Pi ◦ |ψj ihψj |) = Tr(f † ◦ Pi ◦ f )
j j
 X  = Tr(Pi ◦ f ◦ f † )
= Tr Pi ◦ ωj |ψj ihψj | = Tr(Pi ◦ ρ)
j
= Tr(Pi ◦ ρ) . — this can again easily be seen in a picture. We claim that
P f ◦ f † is indeed a density matrix:
We claim that j ωj |ψj ihψj | is indeed a density matrix:
• It is positive by Proposition 5.5.
• For all |φi we have that hφ|ψj ihψj |φi = |hφ|ψj i|2 is
positive and hence so is • We have
X X  Tr(f ◦ f † ) = hΦ|Φi = 1
ωj hφ|ψj ihψj |φi = hφ| ωj |ψj ihψj | |φi
j j whenever |Φi is normalised.

which establishes positivity. √

Again, all mixed states arise in this way by setting f := ρ.
The transition under unitaries and measurement in Postulate
• We moreover have
X    7.1 is again induced by this heuristics. Exactly the same ar-
gument holds for open systems where K now describes the
X
Tr ωj |ψj ihψj | = ωj Tr |ψj ihψj |
j j ‘unknown’ environment. Note also that we can extract ρ in a
X more direct manner from:
= ωj = 1
j
ρ = trK
H,H (|ΦihΦ|)

what completes the claim. i.e. we consider the density matrix of the pure state of the
6
Recall that this is the same thing as saying that these projectors arise large system and trace out the component we are not inter-
as the spectral decomposition of a self-adjoint operator. ested in. In a picture we have:
32
 Decoherence. As a result of interaction between a quantum
f system andP the environment the state of a system decoheres.
Ψf = = f Let |ψi = i ci |ii be the state of the quantum system and let
Ψf f† |ǫi be the state of the environment. We have
f†
decohere
X X
|ǫi ⊗ |ψi = |ǫi ⊗ ci |ii ❀ ci |ǫi i ⊗ |ii
i i
The result of the converse derivation, i.e. finding |Φi given ρ,
is called a purification of ρ. ⇒ the system becomes entangled with the environment
⇒ the system becomes part of a bigger system
Exercise 7.4 Are purifications always unique? If not, can
you characterise all possible purifications of type H⊗H given ⇒ the system’s state becomes a mixed state
a density matrix ρ : H → H. ⇒ we loose lack of knowledge on the state
⇒ the state becomes less informative.
Comparing degrees of mixedness. We can use the majori- Ultimately the state could become the maximally mixed state
sation order to compare the mixedness of mixed states. This i.e. the system could be in any state with equal probability.
can be done in two ways: Obviously this is extremely bad for computational purposes.
Decoherence is a major problem in the experimental realisa-
• Compare respective purifications in majorisation order. tion of quantum informatic devices. Avoiding decoherence
requires error-correction [17].
• Diagonalise the density matrices and compare the lists
of ordered diagonal elements in majorisation order.
8 Quantum logic and Gleason’s theorem
These two are easily seen to be equivalent. Pure states are
minimal elements in this order while there is a unique top In 1932 von Neumann published the current quantum me-
element, namely the maximally mixed state chanical formalism [2], but already in 1935 he wrote in a
 1 letter to G. Birkhoff [3]:
0

N
1 ..
⊥N := · 1H =  .. “I would like to make a confession which may seem
 
.
N
0 1 immoral: I do not believe absolutely in Hilbert
N
space no more.” [von Neumann, 1935]
This state represents a situation where there is no information
on the actual state of the system. Its classical counterpart is This resulted in a joint paper entitled ‘The logic of quantum
the uniform probability distribution since it behaves as such mechanics’ [4], in which the order-theoretic structure which
w.r.t. to any possible basis (cf. it’s an identity up to a scalar). exists on the subspaces of a Hilbert space is taken to be the
logical/structural feature which provides the key difference
Exercise 7.5 For a maximally mixed state ⊥N : H → H between classical and quantum behavior. The subspaces of a
what are the possible purifications? What can you tell about Hilbert space indeed come with order structure:
their degree of entanglement?
A≤B⇔A⊆B

What makes this partial order special is the fact that not all
Exercise 7.6 What is the mixed state describing a single subsets of H are subspaces, nor are the subsets of the set
qubit within the GHZ-state? What is the mixed state describ-
Σ = {| ψi | ψ ∈ H}
ing a pair of qubits within the GHZ-state?
of all rays.

Exercise 7.7 Assume we perform the teleportation protocol Birkhoff-von Neumann quantum logic. Classically, the
but Alice does not communicate the measurement outcome to algebra of (observable) properties that can be attributed to a
Bob. Hence Bob cannot perform the unitary correction. What physical or computational system consists of the subsets of
is the resulting state at Bob’s end? the state space i.e. it is the powerset P(Σ) . Any proposition
on truth of an observable can be expressed in terms of such
33
a subset. Indeed, let f : Σ → R be a physical observable in L(H) e.g. in Q for | +i := | 0i + | 1i we have
e.g. energy values, color, location, speed, etc. Then
| +i ∩ (| 0i ∨ | 1i) = | +i =
6 0 = (| +i ∩ | 0i) ∨ (| +i ∩ | 1i).
f −1 [E] ∈ P(Σ)
This example also shows that the suprema are not disjunctive
expresses the property due te fact that there are superposition states. This turns out
to have dramatic consequences for its logical status e.g. there
“the value of f is within E ⊆ R”. is notion of deduction, nor of modus ponense (etc.) i.e. we do
not have
The basic idea in Birkhoff and von Neumann (1936) is the
same thing. All statements of the form A∧B ⊢C A⊢B ∧ B⇒C
.
A⊢B⇒C A⊢C
“the value of H is within E ⊆ R”
People have played around with with the Sasaki hook
for a self-adjoint operator H can be represented by the pro-
A ⇒ B := A⊥ ∨ (B ∧ A)
jector PH E in the spectral decomposition, or the subspace A H
E
of its fixed points. Conversely, each subspace A ⊆ H is the
but this requires replacing ∧ by the non-commutative and
eigenspace for some projector. Hence the algebra of observ-
non-associative binary connective
able properties of a quantum system seems to be
PA (B) := A ∧ (B ∨ A⊥ ) ,
L(H) := {A ⊂ H | A is a subspace} .

Since L(H) ⊆ P(H) set-theoretic inclusion provides a which cannot be interpreted as a conjunction — an interpre-
partial order and set-theoretic intersection provides greatest tation for this which does seem to make sense is one in terms
lower bounds. We also have least upper bounds, namely the of a Hoare-style weakest precondition/ strongest postcondi-
linear span, which is definable as tion semantics. Therefore it makes more sense to call this an
algebra. There are some important results of this setting.
_ \
Ai = {A ∈ L(H) | ∀i : Ai ⊆ A} ,
i Theorem 8.1 (Gleason 1957) Let dim(H) ≥ 3. For each
state | ψi there exists exactly one function
and orthocomlements
ω| ψi : L(H) → [0, 1]
A⊥ = {ψ ∈ H | ∀φ ∈ A : hψ|φi = 0} .
such that
which satisfies
ω| ψi (A) = 1 ⇔ | ψi ⊆ A
A ≤ B ⇒ B ⊥ ≤ A⊥ A⊥⊥ = A X _
ω| ψi (Ai ) = ω| ψi ( Ai )
A ∧ A⊥ = 0 A ∨ A⊥ = H i i

Hence one obtains a structure where we assume all {Ai }i to be always mutually orthogonal.
 
⊥
L(H) , ∩ , ∨ , , H, 0 Existence is not a surprise since we know such a map, namely

which very strongly resembles a classical Boolean algebra ω| ψi :: A 7→ hψ|PA |ψi .

i.e. an ordinary propositional logic
The fact that this is the only one is quite astonishing. It means
(L , ∧ , ∨ , ¬ , tt , ff ) that in some manner the quantum probability structure is al-
ready encoded in the partial order. Another fascinating result
which comes with conjunction, disjunction, negation, true
is the Mackey-Piron-Solèr Theorem [25], which gives the ex-
and false, and for which we can set
act assumptions one needs to impose on an order-theoretic
a ⇒ b := ¬a ∨ b . structure for it to be the lattice of ‘closed’ subspaces of an
infinite-dimensional Hilbert space — key to the proof is the
There is however a major difference since fundamental theorem of projective geometry. Some of the ax-
ioms admit reasonable physical interpretations. The follow-
A and (B or C) 6= (A and B) or (A and C) , ing is an import consequence of the previous result.
34
Theorem 8.2 (Gleason 1957) Let dim(H) ≥ 3. The collec- Complete positivity is essential. Consider the transpose
tion of all functions
(−)T : F : Σ(Q) → Σ(Q) :: ρ 7→ ρT .
ω : L(H) → [0, 1]
If ρ ∈ Σ(Q) then indeed ρT ∈ Σ(Q) but when applying
which are such that ((−)T ⊗ 1Σ(Q) ) to the Bell-state the result is not positive.
X
ω(Ai ) = 1 Exercise 9.1 Attempt to calculate this using matrices just to
i convince yourself how much fun this is.
W
where {Ai }i are mutually orthogonal and i Ai = H, are in We will show this in a more structural manner. For
bijective correspondence with density matrices. In particular X
f= fij |iihj| : Ha → Hb
ωρ :: A 7→ Tr(ρ PA ) . ij

we have
This result shows that all possible probabilistic behaviors, for X   X 
whatever kind of system including both classical and quantum hii| ⊗ 1Ha ◦ (1Hb ⊗ f ⊗ 1Ha ) ◦ 1Hb ⊗ |jji
uncertainty, can always be described by a density operator. i j
X
However, this order-theoretic approach dramatically failed = hi|f |ji|jihi|
in capturing the Hilbert space tensor product at any alge- ij
braic or conceptual level. That is, any attempt to axiomatize
X X
= fij |jihi| = fji |iihj| = f T
it essentially requires the full-blown Hilbert space structure. ij ij
So we have good reasons to abandon the order-theoretic ap-
proach, but is there any other candidate mathematical struc- that is, in a picture
ture which would enable us to capture the tensor product at a
higher level of abstraction/conceptualization.
fT
= f

9 Mixed operations
This can also be purely graphically derived by ‘sliding the
This is a very involved topic which has its roots in C ∗ -algebra f -box’. Although we have (f ◦ f † )T = f ◦ f T , graphically
[23]. We will mainly mention the key concepts and results.
Similarly as density operators describe a more general notion f = f
of state there are more general notions of operation (contra † T
f f
unitaries) and of measurement.
Denote the set of all mixed states of type H → H as Σ(H).
The type of a generalised notion of operation is i.e. the transposed of something positive is always positive,
F : Σ(H) → Σ(H) ((−)T ⊗ 1Σ(Q) )(|ΨBell ihΨBell |) is not positive:

and from the heuristics of mixed states it follows that:

=
• It is a convex-linear map i.e.
X X
F( pi ρi ) = pi F(ρi ) .
i i
Indeed, we have
Hence, thinking of density matrices as vectors in H ⊗ H,
(h01| − h10|) σ⊗ (|01i − |10i)
these generalised operations have a matrix of type H ⊗
H → H ⊗ H. If H is n-dimensional then ρ is an n × n- = (h10| − h01|) (|01i − |10i) = −2 ≤ 0 .
matrix and F has an (n×n)×(n×n) matrix i.e. 4×4 for
a qubit and 16 × 16 for a pair of qubits i.e. 256 entries.
Exercise 9.2 What does the following expression stands for:
• It preserves positivity in a ‘strong sense’ i.e. whenever X   X 
hii| ⊗ 1Ha ◦ (1Hb ⊗ f † ⊗ 1Ha ) ◦ 1Hb ⊗ |jji
ρ ∈ Σ(H ⊗ K) then (F ⊗ 1Σ(K) )(ρ) has to be positive
i j
too. This condition is called complete positivity.
35
 where ρ is the state of the system, f ◦ f † is the state of the
Exercise 9.3 What is the result of applying the operation ancilla, g is the linear map applied to system + ancilla, and
1H→H ⊗ (−)T : (H → H) → (H → H) the yellow part is the resulting completely positive map of
the form h ◦ (1K ⊗ ρ) ◦ h† with h and h† the dotted parts.
to the CNOT-gate? Moreover, since it is known that each linear map can itself
be realised by applying a unitary to a larger system it follows
The operation 1H→H ⊗ (−)T is called the partial transpose.
that each completely positive map can be realised by applying
Exercise 9.4 How much is: a unitary operator to an extended system. Hence it provides
operational meaning for completely positive maps in terms of
describing ‘open system dynamics’, or equivalently, ‘being
part of a bigger operation’.
In quantum information theory one defines channels to be
Examples of completely positive maps are: completely positive maps which are, depending on the au-
thor, either trace-preserving or trace-decreasing. The aim is
F : Σ(H) → Σ(H) :: ρ 7→ g ◦ (1K ⊗ ρ) ◦ g †
to have an as ‘clean’ as possible channel i.e. with minimal
where g : K ⊗ H → H is linear. Since noise. Noise is a result of interaction with the environment
(g ⊗ 1K′ ) ◦ (1K ⊗ (f ◦ f † )) ◦ (g ⊗ 1K′ )† so we indeed need our generalised operations to model it. A
quantity which measures the level of noise due to the environ-
= (g ⊗ 1K′ ) ◦ (1K ⊗ f ) ◦ (1K ⊗ f † ) ◦ (g ⊗ 1K′ )†
ment is the channel fidelity:
= ((g ⊗ 1K′ ) ◦ (1K ⊗ f )) ◦ ((g ⊗ 1K′ ) ◦ (1K ⊗ f ))†

we have that F is indeed completely positive. This again can hΨBell | (F ⊗ 1H→H ) (|ΨBell ihΨBell |)) |ΨBell i ∈ R+
be immediately seen in a picture:
g g Note that this expression is quite hard to read.

f = f
Exercise 9.6 Show that channel fidelity is a positive number.
f† f†
g† g† In a picture it becomes:

where the yellow part represent the completely positive map

g
while the blue bit is the mixed state of the extended system in
which it acts. By the following theorem it follows that each
completely positive map arises like this.

Theorem 9.5 (Stinespring) Each completely positive map g†

can be realised by applying some linear operator to an ex-
tended system and then tracing this extended system out.
where the yellow part is F = g ◦ (1K ⊗ −) ◦ g † and the dotted
This indeed assures that all completely positive maps admit
part the Bell-state. Via the trace we ‘compare’ the channel’s
the shape g ◦ (1K ⊗ −) ◦ g † since we have:
input with its output. When the channel is perfectly clean,
i.e. it is an identity, then we have:
g
g f
f = =
ρ ρ
f†
g† f†
g†
If it is utterly dirty on the other hand, i.e. nothing is transmit-
ted and it only produces noise, we have:
36
 system in state ψ ⊗ ψ ⊗ Φ0 while Udelete should be such that
for any ψ we have
= Udelete (ψ ⊗ ψ ⊗ Φ0 ) = ψ ⊗ φ0 ⊗ Φψ ,

where again Φψ can depend on ψ and you are allowed to

choose φ0 , Φ0 and Φψ . Does such a unitary operator Udelete
exist? Warning: this setting is not the setting of the so-called
There are many other similar quantities. no-deleting theorem, since we could still be able to re-extract
a copy of ψ out of Φψ , something which is explicitly for-
Theorem 9.7 (Krauss) For each completely positive map F bidden in the notion of deleting which gives rise to the no-
there exist linear maps {Ek : H → H}k such that deleting theorem [9].

Ek ◦ ρ ◦ Ek† .
X
F : Σ(H) → Σ(H) :: ρ 7→ The following exercise is closely related to ex.10.1 ii but puts
k it in a ‘less decorated’ formal context.
A generalised measurement is described by a family of lin-
Exercise 10.2 [MSc mini-project 2006] We study maps be-
ear maps {Ek : H → H}k and induces a transition
tween a Hilbert space H and H ⊗ H. Is the map
ρ 7→ Ek ◦ ρ ◦ Ek†
γ : H → H ⊗ H :: | ψi 7→ | ψi ⊗ | ψi
with probability Tr(Ek ◦ ρ ◦ Ek† )
= Tr(Ek†
◦ Ek ◦ ρ). A
POVM is described by a family of positive linear operators linear, and for which arguments does the linear map
{Mk : H → H}k and produces outcomes with probability
Tr(Mk ◦ ρ). No change of state is associated with a POVM. ∆ : H → H ⊗ H :: | ii 7→ | iii

Theorem 9.8 (Naimark) Each POVM can be realised by ap- coincide with γ?
plying some projective measurement to an extended system
and then tracing this extended system out. The following exercise investigates the structure of the pure
tensors within the whole tensor product of two qubits.

10 More on tensors Exercise 10.3 [MSc mini-project 2006] i. Provide a state

Φ ∈ Q ⊗ Q which is orthogonal to the subspace of Q ⊗ Q
We already proved the no-cloning theorem. The following spanned by the set
exercises indicate how resource sensitivity is imposed by the n o
tensor product, and how it is ‘entangled’ with entanglement. PTQ := |ψi ⊗ |ψi |ψi ∈ Q .

Exercise 10.1 [BA exam 2006] i. We modify the setting in Next, find a basis BQ for the subspace of Q ⊗ Q spanned
which we proved the no-cloning theorem by introducing a by PT, and prove that this is indeed a basis. (Hint: You can
state of the “environment” i.e. we start with a system in state pick Φ and BQ from the union of the elements of the com-
ψ ⊗ φ0 ⊗ Φ0 and we wish to find a unitary operator Uclone putational basis and the Bell-basis.) Is PT itself a subspace
such that for any “unknown” state ψ we have of Q ⊗ Q? ii. We will now generalize the above to a Hilbert
Uclone (ψ ⊗ φ0 ⊗ Φ0 ) = ψ ⊗ ψ ⊗ Φψ space H of arbitrary dimension. Provide a basis BH for the
subspace of H ⊗ H spanned by the set
where Φψ is allowed to depend on ψ while φ0 and Φ0 are n o
constants, but you are allowed to choose φ0 , Φ0 and Φψ . Does PTH := |ψi ⊗ |ψi |ψi ∈ H .
such a unitary operator Uclone exist? ii. Why is the operation
Next, find a set A of vectors which are all orthogonal to PTH
δ : H ⊗ H → H ⊗ H :: | iji 7→ | iii
and such that A ∪ BH is a basis of H ⊗ H. (Hint: Are
not a copying operation? More specifically, characterize the there some obvious ways in which you can embed each of
initial states for which δ does not map the input state |ψi ⊗ the vectors included in the Bell-basis for Q ⊗ Q in the higher-
|φ0 i to the output state |ψi ⊗ |ψi. iii. We reverse this cloning dimensional space H ⊗ H?)
setting into some “pseudo-deleting” setting by starting with a
37
Compositionality with relations. We now present a sur- what boils down to exactly being the cartesian product i.e.
prising analogue to the compositionality result which enabled “pairs of pairs” R1 × R2 =
us to derive several quantum protocols. Let’s move from the
world of Hilbert space to sets, ... {((x1 , y1 ), (x2 , y2 )) ∈ (X1 ×Y1 )×(X2 ×Y2 ) |
(x1 , y1 ) ∈ R1 , (x2 , y2 ) ∈ R2 }
Hilbert space ❀ set
For two relations R ⊆ X ×Y and S ⊆ Y ×Z their sequential
linear map ❀ relation
composition is defined as
tensor product ❀ cartesian product
x(R; S)z ′′ ⇔ ∃y ∈ Y : xRy & ySz
f unction composition ❀ relational composition
Where do we end up in this way: something classical-like that is
or something quantum-like? The obvious guess would be to
R; S := {(x, z) ∈ X × Z | y ∈ Y : xRy , ySz} .
think that we obtain something classical-like. However!
A relation ‘from X to Y ’ is some R ⊆ X × Y . Hence We can now calculate what a ket-bra is. Since r : {∗} → X
it is a family of pairs (x, y) ∈ R. When we write xRy for is some ({∗} × A) ⊆ {∗} × X we obtain
(x, y) ∈ R we can write:
rc ; r = (A × {∗}); ({∗} × A) = A × A
R = {(x, y) | x ∈ X, y ∈ Y, xRy} .
where rc stands for the converse relation so
We have an analogue to scalars:
adjoint ❀ converse.
H ⊗ C ≃ H ❀ X × {∗} ≃ X
By the trivial map state duality bipartite ket-bras are:
and hence an analogue to state:
PR = R × R ⊆ (X × Y ) × (X × Y )
|ψi : C → H ❀ r : {∗} → X
in analogy with Pf := |Ψf ihΨf |. We wish to study
so the ‘relational states’ are the subsets of X, i.e. elements of
the powerset P(X). A notion of Superposition emerges: (1X × PS ); (PR × 1Z )

✻
where R ⊆ X × Y and S ⊆ Y × Z. We have

(x, y, z)(1X × PS )(x′ , y ′ , z ′ )

{2} {1,2}
if and only if
{2,3} {1,2,3}
∅ {1} ✲ x = x′ y ′ Sz ′ (ySz)

{3} {1,3}
and we have

✠ (x′ , y ′ , z ′ )(PR × 1Z )(x′′ , y ′′ , z ′′ )

Map state duality is a tautology: if and only if

P(X × Y ) = P(X × Y ) z ′ = z ′′ x′ Ry ′ (x′′ Ry ′′ )

where so
(x, y, z)(1X × PS ))(PR × 1Z )(x′′ , y ′′ , z ′′ )
• lefthandside are the states of the space X × Y
if there exists (x′ , y ′ , z ′ ) such that x = x′ Ry ′ Sz ′ = z ′′ i.e.
• righthandside are all relations ‘from X to Y ’
x(R; S)z
For two relations R1 ⊆ X1 × Y1 and R2 ⊆ X2 × Y2 their
which exactly yields our well-known compositionality result
parallel composition (cf. tensor) is defined as
for Hilbert spaces with corresponding ‘seemingly backward
(x , x )(R “×”R )(y , y ) ⇔ x R y & x R y in time information flow’.
1 2 1 2 1 2 1 1 1 2 2 2

38
 11.1 Symmetric monoidal categories
Exercise 10.4 i. What is a Bell-state in the world of rela-
tions? ii. Show that in the world of relations there also exists This Section is available as §1,2,3,4,5 in a paper at
an analogue to Exercise 5.1. iii. Can you come up with a
web.comlab.ox.ac.uk/internal/courses/
notion of full and partial trace for the world of relations? iv.
materials05-06/qcs/Cats.pdf
Using this notion of trace prove an analogue to Exercise 5.11
on partial traces in the world of relations. In the symmetric monoidal category (Rel, ×) with

There are two ways to pass from Hilbert space to the purely • sets as objects;
quantitative level of relations namely via • relations as morphisms;
• the cartesian product as its monoidal bifunctor with the
• matrix calculus where the Boolean semiring B (i.e. 1 +
singleton set {∗} as its unit;
1 = 1 and not Z2 ) replaces the complex field C.
and the symmetric monoidal category (FdHilb, ⊗) with
• Dirac notation where the singleton set {∗} replaces C in
the definition of bras and kets. • finite dimensional Hilbert spaces as objects;
• linear maps as morphisms;
Can you figure out why in one case we have a two-element
set while in the other case we have a singleton set — this is • the tensor product as its monoidal bifunctor with the
quite a hard one to solve! one-dimensional Hilbert space C as its unit;
What do sets, relations and cartesian product on-the-one- there is no natural diagonal i.e. no family of morphisms
hand and Hilbert spaces, linear maps and tensor product on-
the-other-hand have in common? Their category-theoretic {∆A : A → A ⊗ A}A
structure are very similar! The fact that we picked sets wasn’t such that for all f : A → B we have commutation of
important, but it was crucial to pick relations and not func-
f ✲
tions, and to pick cartesian product and not disjoint union. A B

Exercise 10.5 If we would have picked functions and carte- ∆A ∆B

sian product would there be some notion of superposition?
❄ ❄
A⊗A ✲ B⊗B
f ⊗f
This while there are obvious candidates for such an operation.
11 Semantics for quantum informatics In FdHilb a first candidate would be

Goals of this semantics: H → H × H :: |ψi 7→ |ψi ⊗ |ψi

but this map fails to be linear. A second candidate would be
• We want a formal counterpart to the picture language
∆{| ii}i : H → H × H :: | ii 7→ | iii .
which seems to help understanding entanglement.
The obvious candidate for Rel is the ‘function’
• We want to do quantum theory in a more conceptual
∆X : X → X × X :: x 7→ (x, x)
manner than just playing around with matrices.
which written as a relation is
• We hope to produce a better quantum formalism than the R∆ := {(x, x) | x ∈ X} ⊆ X × X .
one around, a search which is by now a more that 70 year
old lasting endeavor. But it turns out that neither ∆{| ii}i nor R∆ make the required
diagrams commute. Counterexamples can be found in §7 of
• Revealing a structure which lives on Hilbert spaces, but the above mentioned reference. Analysis of these counterex-
rather at the level of linear maps than at the level of the amples for the FdHilb-case clearly show that the existence of
vectors in the Hilbert space since category theory reveals superposition states is what causes the problem. In the case
some structural connections which ordinary mathemati- of relation the corresponding problem is that relations can be
cal structures don’t. multi-valued, and this is why for the symmetric monoidal cat-
egory (Set, ×) with
39
 • sets as objects; Another important example of a map which depends on the
choice of basis is
• functions as morphisms; | ei i 7→ hei |
• the cartesian product as its monoidal bifunctor with the since we have
singleton set {∗} as its unit;
ci · | ei i = | ci · ei i 7→ hci · ei | = c̄i · hei | .
the function ∆X does provide a natural diagonal. Note that That is, {| e i 7→ he |} and {| c · e i 7→ hc · e |} define
i i i i i i i i
this absence of a natural diagonal in FdHilb is strongly con- non-equal linear maps whenever one of the c has a non-trivial
i
nected to the No-Cloning theorem. Moreover, while No- imaginary part. Hence it follows that also
Cloning required unitarity in its proof, here we don’t rely on
unitarity whatsoever. | ei i ⊗ | ej i 7→ | ej ihei |

depends on the choice of basis since

11.2 Naturality implies basis-independence
ci · |ei i ⊗ |ej i = | ci · ei i ⊗ |ej i 7→ |ej ihci · ei | = c̄i · |ej ihei |.
Generally speaking, given expression Λ(−, . . . , −) and ′ ′
Ξ(−, . . . , −) — where we only use brackets and the monoidal Let H ⊸ H be all linear maps f : H → H . The assignment
bifunctor — naturality of a family of morphisms H ⊗ H′ → H ⊸ H′ :: | i ji 7→ | iihj |
{ξA1 ,...,An : Λ(A1 , . . . , An ) → Ξ(A1 , . . . , An )}A1 ,...,An is exactly how we related the elements of the tensor product to
linear maps i.e. map-state duality. Hence it follows that this
means that we have commutation of: correspondence is not basis-independent, hence not natural.
The solution to this problem consists of defining for each
Λ(f1 , . . . , fn✲) Hilbert space H the conjugate Hilbert space H∗ which has
Λ(A1 , . . . , An ) Λ(B1 , . . . , Bn )
the same vectors as H but in which each complex number is
interpreted as the complex conjugate of a complex number
ξA1 ,...,An ξB1 ,...,Bn for H, that is, for c ∈ C and ψ, φ ∈ H∗ we have
❄ ❄ c •H∗ ψ := c̄ •H ψ hψ | φiH∗ := hψ | φiH = hφ | ψiH .
Ξ(A1 , . . . , An ) ✲ Ξ(B1 , . . . , Bn )
Ξ(f1 , . . . , fn )
It turns out that now we do have a natural isomorphism
for all fi : Ai → Bi . Restricting this requirement to all the fi
being unitarity for the case of Hilbert spaces we obtain H ⊗ H′ ≃ H∗⊸ H′ .

Λ(U1 , . . . , Un✲) Moreover

Λ(H1 , . . . , Hn ) Λ(H1′ , . . . , Hn′ ) H∗∗ = H .
This seems to indicate a logical interpretation where we take ∗
ξH1 ,...,Hn ξH1′ ,...,Hn′ to be negation, ⊸ to be implication and ⊗ to be ‘coinciding’
❄ ❄ conjunction-disjunction since
Ξ(H1 , . . . , Hn ) ✲ Ξ(H′ , . . . , H′ )
1 n
Ξ(U1 , . . . , Un ) (H ⊗ H′ )∗ ≃ H∗ ⊗ H′∗ .

We can take each Ui : Hi → Hi′ to be a change of basis. In fact, we have a ‘degenerate’ so-called Linear Logic in
Hence we obtain that naturality implies basis-independency. which conjunction and disjunction coincide.
This then also immediately makes clear why ∆{| ii}i : H →
H × H couldn’t have been natural.
11.3 †-compact categories
Exercise 11.1 Show that ∆{| ii}i indeed depends on the We define a †-compact category as a symmetric monoidal cat-
choice of basis i.e. there exists a basis {| ei i}i such that egory which comes with the following additional data
∆{| ei i}i : H → H × H :: |ei i 7→ |ei i ⊗ |ei i • involution dual A 7→ A∗ ;
does not coincide with ∆{| ii}i . Can you find the necessary †
• contravariant ⊗-involution7 adjoint fA→B 7→ fB→A ;
condition on {| ei i}i such that ∆{| ii}i and ∆{| ei i}i coincide? 7
I.e. the ‘abstract’ adjoint has to preserve the tensor-structure.

40
 • Bell-states ηA : I → A∗ ⊗ A; Both f∗ and f ∗ admit a nice purely diagrammatic characteri-
zation, respectively:
for which we have ηA∗ = σA∗,A ◦ ηA and

≃ η † ⊗ 1A
A✛ I ⊗ A ✛A
∗
(A ⊗ A∗ ) ⊗ A =:
✻ ✻ f f†
*
1A ≃

A ✲ A⊗I ✲ A ⊗ (A∗ ⊗ A)
≃ 1A ⊗ ηA
As discussed in =:
f* f
http://arxiv.org/abs/quant-ph/0510032

such a category also admits a corresponding graphical calcu-

lus in which the above commuting diagram admits a yanking
We can abstractly show that these are factors of the adjoint:
interpretation:

= (f )*
*
= f†
= f†

An example are Hilbert spaces, with their conjugates, adjoint

of linear maps, and “natural” Bell-states. First define conju- and analogous we can prove that (f ∗ ) = f † .
∗
gate kets from kets by making the passage

| ψi : C → H ❀ | ψi∗ : C → H∗ 11.4 Classical uncertainty and open systems

for which we in particular have But there is in fact an even more stunning presence of com-
| c · ψi ⊗ | ψi∗ = c · (| ψi ⊗ | ψi∗ ) = | ψi ⊗ | c̄ · ψi∗ plex numbers at the level of abstract picture calculi of the kind
we consider for which we refer to §4.d of the above mentioned
and the “natural” Bell-state are reference. Exactly the same reasoning as done there holds for
X
C → H∗ ⊗ H :: 1 7→ | ii∗ ⊗ | ii . f 7→ f ⊗ f∗ and f 7→ pf q ⊗ (pf q)†
i
as for
Note that this Bell-state arises through the “natural” map-state
f 7→ f ⊗ f † .
duality from the identity on H. Also, each linear map f :
H → H′ induces a conjugate one In fact, all of these represent the passage from pure states to
density matrices, for kets respectively
f∗ : H∗ → H′∗ :: | ψi∗ 7→ | f (ψ)i∗
|ψi 7→ |ψi ⊗ |ψi∗ and |ψi 7→ |ψihψ|
for which for
X E X mixed states arise as R+ -weighted sums of these
| ψi∗ = ci | ii = c̄i | ii∗
∗
i i
X X
ri · |ψi i ⊗ |ψi i∗ and ri · |ψi ihψi |
we have i i
X E X
| ψi∗ 7→ | f (ψ)i∗ = ci mij | ji = c̄i mij | ji∗ they capture all probabilistic behaviors, including
∗
ij ij
• ‘lack of knowledge’;
i.e. we obtain the conjugate matrix. Each such conjugate map
admits an adjoint, which we will denote by f∗ : H′∗ → H∗ . • ‘statisical mixture’;
41
 • ‘looking at part of a system’. [8] Wootters, W. and Zurek, W. (1982) A single quantum cannot
be cloned. Nature 299, 802–803.
[9] Pati, A. K. and Braunstein, S. L. (2000) Impossibility of delet-
Exercise 11.2 Show that when we measure the second sys- ing an unknown quantum state. Nature 404, 164–165.
tem of a compound system which is globally in state |Ψf i [10] Bennett, C. H. and Wiesner, S. J. (1992) Communication via
that the probabilities are given by Tr(ρ P) where ρ := f ◦ f † . one- and two-particle operators on EPR states. Physical Re-
view Letters 69, 2881–2884.
[11] Bennett, C. H., Brassard, G., Crépeau, C., Jozsa, R., Peres, A.
Since ρ := f ◦ f † is positive, it is also self-adjoint, and hence
and Wooters, W. K. (1993) Teleporting an unknown quantum
admits a spectral decomposition so it can indeed be written as state via dual classical and Einstein-Podolsky-Rosen chan-
a density matrix of the above kind. nels. Physical Review Letters 70, 1895–1899.
Besides mixed states there are of course also mixed opera- [12] Żukowski, M., Zeilinger, A., Horne, M. A. and Ekert, A. K.
tions. We end with a very recent result in which both mixed (1993) ‘Event-ready-detectors’ Bell experiment via entangle-
states and mixed operations are constructed at the abstract ment swapping. Physical Review Letters 71, 4287–4290.
level of the picture calculus [?]: [13] Shor, P. W. (1994) Algorithms for quantum computation: dis-
crete logarithms and factoring. Proceedings of the 35th An-
B B* nual IEEE Symposium on Foundations of Computer Science,
IEEE Computer Science Press.
f f
* [14] Grover, L. (1997) Quantum mechanics helps in searching for
A C C* A* a needle in a haystack. Physical Review Letters 79, 325–328.
arXiv:quant-ph/9706033
[15] Gottesman, D. and Chuang, I. L. (1999) Quantum teleportation
and admit covariant composition: is a universal computational primitive. Nature 402, 390–393.
E E* [16] Gruska, J. (1999) Quantum Computing. McGraw-Hill.
[17] Nielsen, M. A. and Chuang, L. (2000) Quantum Computation
g g and Quantum Information. Cambridge University Press.
*
D D* [18] Nielsen, M. A. (1999) Conditions for a class of entanglement
B B* transformations. Physical Review Letters 83, 436–439.
[19] Raussendorf, R. and Briegel, H.-J. (2001) A one-
f f way quantum computer. Physical Review Letters 86,
* 5188. Raussendorf, R., Browne, D. E. and Briegel,
A C C* A*
H.-J. (2003) Measurement-basisd quantum computa-
tion on cluster states. Physical Review A 68, 022312.
arXiv:quant-ph/0301052
[20] Kitaev, A. Yu., Shen, A. H. and Vyalyi, M. N. (2001) Classi-
References cal and Quantum Computing. Garduate Studies in Mathemat-
ics 47, American Mathematical Society.
[1] Kleppner, D. and Jackiw, R. (2000) One Hundred Years of [21] Quantum Technologies Group at ARC Seibersdorf research
Quantum Physics. Science 289, 893–898. arXiv:quant- GmbH and the group Quantum Experiments and the Foun-
ph/0008092 dations of Physics of the University of Vienna (April
[2] von Neumann, J. (1932) Mathematische Grundlagen der 21, 2004) World Premiere: Bank Transfer via Quantum
Quantenmechanik. Springer-Verlag. English translation Cryptography Based on Entangled Photons. Press release.
(1955): Mathematical Foundations of Quantum Mechanics. www.quantenkryptographie.at/
Princeton University Press. [22] Beth, Th., Mueller-Quade, J., and Steinwandt, R. (2004)
[3] Rédei, M. (1997) Why John von Neumann did not like the Cryptanalysis of a practical quantum key distribu-
Hilbert space formalism of quantum mechanics (and what he tion with polarization-entangled photons. To appear in
liked instead). Studies in History and Philosophy of Modern Quantum Information and Computation. Available at
Physics 27, 493–510. arXiv:quant-ph/0407130
[4] Birkhoff, G. and von Neumann, J. (1936) The logic of quan- [23] Paulsen, V. (2002) Completely Bounded Maps and Operator
tum mechanics. Annals of Mathematics 37, 823–843. algebras. Cambridge University Press.
[5] Dirac, P. A. M. (1947) The Principles of Quantum Mechanics, [24] Mackey, G. W. (1963) Mathematical Foundations of Quantum
3rd edition. Oxford University Press. Mechanics. W.A. Benjamin Inc.
[6] Kochen, S. and Specker, E.P. (1967) The Problem of Hid- [25] Piron, C. (1964) Axiomatique quantique. Helvetica Physica
den Variables in Quantum Mechanics. Journal Mathathemat- Acta 37, 439–468. Piron, C. (1976) Foundations of Quan-
ics and Mechanics 17, 59–87. tum Physics, W.A. Benjamin Inc. Solèr, M. P. (1995) Charac-
[7] Belinfante, F. J. (1973) A Survey of Hidden-Variables Theo- terization of Hilbert Spaces by Orthomodular Spaces’. Com-
ries. Pergamon Press. (in Mathematical Institute’s library) muncations in Algebra 23, 219–243.
42
[26] Coecke, B. (2003) The Logic of entanglement. An invitation.
PRG-RR-03-12. web.comlab.ox.ac.uk/oucl/pub-
lications/tr/rr-03-12.html & arXiv:quant-
ph/0402014
[27] Abramsky, S. and Coecke, B. (2004) A categorical seman-
tics of quantum protocols. Proceedings of the 19th Annual
IEEE Symposium on Logic in Computer Science, IEEE
Computer Science Press. arXiv:quant-ph/0402130
Coecke, B. (2005) Quantum information-flow, concretely,
and axiomatically. In: Proceedings of Quantum Informat-
ics 2004, pp. 15–29, Y. I. Ozhigov, Ed., Proceedings of
SPIE Vol. 5833. arXiv:quant-ph/0506132 Coecke, B.
(2005) Kindergarten quantum mechanics — lecture notes.
In: Quantum Theory: Reconsiderations of the Foundations
III. A. Khrennikov, Ed., American Institute of Physics Press.
arXiv:quant-ph/0510032
[28] Abramsky, S. and Duncan, R. (2005) A categorical quan-
tum logic. Mathematical Structures in Computer Science.
arXiv:quant-ph/0512114

43