(Springer Tracts in Modern Physics 173) Gernot Alber, Thomas Beth, Michał Horodecki, Paweł Horodecki, Ryszard Horodecki, Martin Rötteler, Harald Weinfurter, Reinhard Werner, Anton Zeilinger (auth.) - .pdf

1 From the Foundations of Quantum Theory
to Quantum Technology – an Introduction
Gernot Alber
Nowadays, the new technological prospects of processing quantum informa-

tion in quantum cryptography [1], quantum computation [2] and quantum
communication [3] attract not only physicists but also researchers from other
scientific communities, mainly computer scientists, discrete mathematicians
and electrical engineers. Current developments demonstrate that character-
istic quantum phenomena which appear to be surprising from the point of
view of classical physics may enable one to perform tasks of practical interest
better than by any other known method. In quantum cryptography, the no-
cloning property of quantum states [4] or the phenomenon of entanglement
[5] helps in the exchange of secret keys between various parties, thus en-
suring the security of one-time-pad cryptosystems [6]. Quantum parallelism
[7], which relies on quantum interference and which typically also involves
entanglement [8], may be exploited for accelerating computations. Quantum
algorithms are even capable of factorizing numbers more efficiently than any
known classical method is [9], thus challenging the security of public-key cryp-
tosystems such as the RSA system [6]. Classical information and quantum
information based on entangled quantum systems can be used for quantum
communication purposes such as teleporting quantum states [10, 11].
Owing to significant experimental advances, methods for processing quan-
tum information have developed rapidly during the last few years.1 Basic
quantum communication schemes have been realized with photons [10, 11],
and basic quantum logical operations have been demonstrated with trapped
ions [13, 14] and with nuclear spins of organic molecules [15]. Also, cavity
quantum electrodynamical setups [16], atom chips [17], ultracold atoms in
optical lattices [18, 19], ions in an array of microtraps [20] and solid-state
devices [21, 22, 23] are promising physical systems for future developments
in this research area. All these technologically oriented, current developments
rely on fundamental quantum phenomena, such as quantum interference, the
measurement process and entanglement. These phenomena and their distinc-
tive differences from basic concepts of classical physics have always been of
central interest in research on the foundations of quantum theory. However,
in emphasizing their technological potential, the advances in quantum infor-
1
Numerous recent experimental and theoretical achievements are discussed in [12].
G. Alber, T. Beth, M. Horodecki, P. Horodecki, R. Horodecki, M. Rötteler, H. Weinfurter,

R. Werner, A. Zeilinger: Quantum Information, STMP 173, 1–13 (2001)

c Springer-Verlag Berlin Heidelberg 2001
2 Gernot Alber
mation processing reflect a profound change in the general attitude towards

these fundamental phenomena. Thus, after almost two decades of impressive
scientific achievements, it is time to retrace some of those significant early de-
velopments in quantum physics which are at the heart of quantum technology
and which have shaped its present-day appearance.
1.1 Early Developments
Many of the current methods and developments in the processing of quantum

information have grown out of a long struggle of physicists with the foun-
dations of modern quantum theory. The famous considerations by Einstein,
Podolsky and Rosen (EPR) [24] on reality, locality and completeness of phys-
ical theories are an early example in this respect. The critical questions raised
by these authors inspired many researchers to study quantitatively the essen-
tial difference between quantum physics and the classical concepts of reality
and locality. The breakthrough was the discovery by J.S. Bell [25] that the
statistical correlations of entangled quantum states are incompatible with the
predictions of any theory which is based on the concepts of reality and lo-
cality of EPR. The constraints imposed on statistical correlations within the
framework of a local, realistic theory (LRT) are expressed by Bell’s inequality
[25]. As the concept of entanglement and its peculiar correlation properties
have been of fundamental significance for the development of quantum infor-
mation processing, it is worth recalling some of its most elementary features
in more detail.
1.1.1 Entanglement and Local, Realistic Theories
In order to clarify the characteristic differences between quantum mechan-

ical correlations originating from entangled states and classical correlations
originating from local, realistic theories, let us consider the following basic
experimental setup (Fig. 1.1). A quantum mechanical two-particle system,
α2
β β1
2
ϕ ϕ
ϕ
α1
A B
+1 +1
αi s βi
-1 -1
Fig. 1.1. Basic experimental setup for testing Bell’s inequality; the choices of the
directions of polarization on the Bloch sphere for optimal violation of the CHSH
inequality (1.3) correspond to ϕ = π/4 for spin-1/2 systems
1 From the Foundations of Quantum Theory to Quantum Technology 3
such as a photon pair, is produced by a source s. Polarization properties of

each of these particles are measured subsequently by two distant observers A
and B. Observers A and B perform polarization measurements by randomly
selecting one of the directions α1 or α2 , and β1 or β 2 , respectively, in each
experiment. Furthermore, let us assume that for each of these directions only
two measurement results are possible, namely +1 or −1. In the case of pho-
tons these measurement results would correspond to horizontal or vertical
polarization.
What are the restrictions imposed on correlations of the measurement
results if the physical process can be described by an underlying LRT with
unknown (hidden) parameters? For this purpose, let us first of all summarize
the minimal set of conditions any LRT should fulfill.
1. The state of the two-particle system is determined uniquely by a parame-
ter λ, which may denote an arbitrary set of discrete or continuous labels.
Thus the most general observable of observer A or B for the experimental
setup depicted in Fig. 1.1 is a function of the variables (αi , β j , λ). If the
actual value of the parameter λ is unknown (hidden), the state of the
two-particle system has to be described by a normalized probability dis-
tribution P (λ), i.e. Λ dλ P (λ) = 1, where Λ characterizes the set of all
possible states. The state variable λ determines all results of all possible
measurements, irrespective of whether these measurements are performed
or not. It represents the element of physical reality inherent in the ar-
guments of EPR: “If, without in any way disturbing a system, we can
predict with certainty the value of a physical quantity, then there exists
an element of physical reality corresponding to this physical quantity”
[24].
2. The measurement results of each of the distant (space-like separated)
observers are independent of the choice of polarizations of the other ob-
server. This assumption reflects the locality concept inherent in the argu-
ments of EPR: “The real factual situation of the system A is independent
of what is done with the system B, which is spatially separated from the
former” [24]. Thus, taking into account also this locality requirement, the
most general observable of observer A for the experimental setup depicted
in Fig. 1.1 can depend on the variables αi and λ (for B, β j and λ) only.
These two assumptions, which reflect fundamental notions of classical physics
as used in the arguments of EPR, restrict significantly the possible cor-
relations of measurements performed by both distant observers. Accord-
ing to these assumptions, the following measurement results are possible:
a(αi , λ) ≡ ai = ±1 (i = 1, 2) for observer A, and b(β i , λ) ≡ bi = ±1
(i = 1, 2) for observer B. For a given value of the state variable λ, all these
possible measurement results of the dichotomic (two-valued) variables ai and
bi (i = 1, 2) can be combined in the single relation
|(a1 + a2 )b1 + (a2 − a1 )b2 | = 2 . (1.1)
4 Gernot Alber
It should be mentioned that this relation is counterfactual [26] in the sense

that it involves both results of actually performed measurements and pos-
sible results of unperformed measurements. All these measurement results
are determined uniquely by the state variable λ. If this state variable is un-
known (hidden), (1.1) has to be averaged over the corresponding probability
distribution P (λ). This yields an inequality for the statistical mean values,

ai bj LRT = d λ P (λ)a(αi , λ)b(β j , λ) (i, j = 1, 2), (1.2)
Λ
which is a variant of Bell’s inequality and which is due to Clauser, Horne,

Shimony and Holt (CHSH) [27], namely
| a1 b1 LRT + a2 b1 LRT + a2 b2 LRT − a1 b2 LRT | ≤ 2 . (1.3)
This inequality characterizes the restrictions imposed on the correlations be-
tween dichotomic variables of two distant observers within the framework of
any LRT. There are other, equivalent forms of Bell’s inequality, one of which
was proposed by Wigner [28] and will be discussed in Chap. 3.
Quantum mechanical correlations can violate this inequality. For this pur-
pose let us consider, for example, the spin-entangled singlet state
1
|ψ = √ (| + 1A | − 1B − | − 1A | + 1B ) , (1.4)
2
where | ± 1A and | ± 1B denote the eigenstates of the Pauli spin operators
σzA and σzB , with eigenvalues ±1. Quantum mechanically, the measurement
of the dichotomic polarization variables ai and bi is represented by the spin
operators âi = αi ·σ A and b̂i = β i ·σ B . (σ A , for example,
denotes the vector of
Pauli spin operators referring to observer A, i.e. σ A = i=x,y,z σiA ei , where
ei are the unit vectors.) The corresponding quantum mechanical correlations
entering the CHSH inequality (1.3) are given by
âi b̂j QM = ψ|âi b̂j |ψ = −αi · β j . (1.5)

Choosing the directions of the polarizations (α1 , β 1 ), (β 1 , α2 ), (α2 , β2 ) on
the Bloch sphere so that they involve an angle of π/4 (see Fig. 109), one finds
a maximal violation of inequality (1.3), namely
√
| â1 b̂1 QM + â2 b̂1 QM + â2 b̂2 QM − â1 b̂2 QM | = 2 2 > 2 . (1.6)
Thus, for this entangled state, the quantum mechanical correlations between
the measurement results of the distant observers A and B are stronger than
any possible correlation within the framework of an LRT. Obviously, these
correlations are incompatible with the classical notions of reality and local-
ity of any LRT. It is these peculiar quantum correlations originating from
entanglement which have been of central interest in research on the founda-
tions of quantum theory and which are also of central interest for quantum
information processing.
So far, numerous experiments testing and supporting violations of Bell’s

inequality [29, 30, 31] have been performed.2 However, from a strictly logical
point of view, the results of all these experiments could still be explained
by an LRT, owing to two loopholes, namely the locality and the detection
loopholes. The locality loophole concerns violations of the crucial locality as-
sumption underlying the derivation of Bell’s inequality. According to this as-
sumption one has to ensure that any signaling between two distant observers
A and B is impossible. The recently performed experiment of G. Weihs et
al. [31] succeeded in fulfilling this locality requirement by choosing the sep-
aration between these observers to be sufficiently large. However, so far all
experiments have involved low detection efficiencies, so that in principle the
observed correlations which violate Bell’s inequality can still be explained by
an LRT [32, 33]. This latter detection loophole constitutes a major experi-
mental challenge, and it is one of the current experimental aims to close both
the detection loophole and the locality loophole simultaneously [34, 35, 36].
The concepts of physical reality and locality which lead to inequality (1.3)
can also lead to logical contradictions with quantum theory which are not of
statistical origin. This becomes particularly apparent when one considers an
entangled three-particle state of the form
1
|ψGHZ = √ (| + 1A | + 1B | + 1C − | − 1A | − 1B | − 1C ) , (1.7)
2
a so-called Greenberger–Horne–Zeilinger (GHZ) state [37]. Again | ± 1A ,
| ± 1B , and | ± 1C denote the eigenstates of the Pauli spin operators σzA ,
σzB , and σzC , with eigenvalues ±1. Similarly to Fig. 109, let us assume that
the polarization properties of this entangled quantum state are investigated
by three distant (space-like separated) observers A, B and C. Each of these
observers chooses his or her direction of polarization randomly along either
the x or the y axis.
What are the consequences an LRT would predict? As the three observers
are space-like separated, the locality assumption implies that a polarization
measurement by one of these observers cannot influence the results of the
other observers. Following the notation of Fig. 109, the possible results of the
polarization measurements of observers A, B and C along directions αi , β j
and γ k are ai = ±1, bj = ±1 and ck = ±1. Let us now consider four pos-
sible coincidence measurements of these three distant observers, with results
(ax , bx , cx ), (ax , by , cy ), (ay , bx , cy ) and (ay , by , cx ). As we are dealing with
dichotomic variables, within an LRT the product of all these measurement
results is always given by
RLRT = (ax bx cx )(ax by cy )(ay bx cy )(ay by cx ) = a2x b2x c2x a2y b2y c2y = 1 . (1.8)
What are the corresponding predictions of quantum theory? In quantum
theory the variables ai , bj and ck are replaced by the Pauli spin operators
2
For a comprehensive discussion of experiments performed before 1989, see [29].
6 Gernot Alber
âi = αi · σ A , b̂j = β j · σ B and ĉk = γ k · σ C . The GHZ state of (1.7) fulfills

the relations
âx b̂x ĉx |ψGHZ = −|ψGHZ ,
âx b̂y ĉy |ψGHZ = ây b̂x ĉy |ψGHZ = ây b̂y ĉx |ψGHZ = |ψGHZ . (1.9)
Therefore the quantum mechanical result for the product of (1.8) is given by
RQM |ψGHZ = (âx b̂x ĉx )(âx b̂y ĉy )(ây b̂x ĉy )(ây b̂y ĉx )|ψGHZ
= (−1)|ψGHZ (1.10)
and contradicts the corresponding result of an LRT. These peculiar quantum
mechanical predictions have recently been observed experimentally [38]. The
entanglement inherent in these states offers interesting perspectives on the
possibility of distributing quantum information between three parties [39].
1.1.2 Characteristic Quantum Effects for Practical Purposes
According to a suggestion of Feynman [40], quantum systems are not only of

interest for their own sake but might also serve specific practical purposes.
Simple quantum systems may be used, for example, for simulating other, more
complicated quantum systems. This early suggestion of Feynman emphasizes
possible practical applications and thus indicates already a change in the
attitude towards characteristic quantum phenomena.
In the same spirit, but independently, Wiesner suggested in the 1960s the
use of nonorthogonal quantum states for the practical purpose of encoding
secret classical information [41].3 The security of such an encoding procedure
is based on a characteristic quantum phenomenon which does not involve
entanglement, namely the impossibility of copying (or cloning) nonorthogonal
quantum states [4]. This impossibility becomes apparent from the following
elementary consideration. Let us imagine a quantum process which is capable
of copying two nonorthogonal quantum states, say |0 and |1, with 0 <
|0|1| < 1. This process is assumed to perform the transformation
|0|ϕ|a → |0|0|a0 ,
|1|ϕ|a → |1|1|a1 , (1.11)

where |ϕ represents the initial quantum state of the (empty) copy and
|a, |a0 , |a1 denote normalized quantum states of an ancilla system. This
ancilla system describes the internal states of the copying device. As this
copying process has to be unitary, it has to conserve the scalar product be-
tween the two input and the two output states. This implies the relation
0|1(1 − 0|1a0 |a1 ) = 0. This equality can be fulfilled only if either states
3
Though this article was written in the 1960s, it was not published until 1983.
|0 and |1 are orthogonal, i.e. 0|1 = 0, or if 0|1 = 1 = a0 |a1 . Both pos-
sibilities contradict the original assumption of nonorthogonal, nonidentical
initial states. Therefore a quantum process capable of copying nonorthogo-
nal quantum states is impossible. This is an early example of an impossible
quantum process.
Soon afterwards, Bennett and Brassard [42] proposed the first quan-
tum protocol (BB84) for secure transmission of a random, secret key using
nonorthogonal states of polarized photons for the encoding (see Table 1.1).
In the Vernam cipher, such a secret key is used for encoding and decoding
messages safely [6, 43]. In this latter encoding procedure the message and
the secret key are added bit by bit, and in the decoding procedure they are
subtracted again. If the random key is secret, the safety of this protocol is
guaranteed provided the key is used only once, has the same length as the
message and is truly random [44]. Nonorthogonal quantum states can help in
transmitting such a random, secret key safely. For this purpose A(lice) sends
photons to B(ob) which are polarized randomly either horizontally (+1) or
vertically (−1) along two directions of polarization. It is convenient to choose
the magnitude of the angle between these two directions of polarization to be
π/8. B(ob) also chooses his polarizers randomly to be polarized along these
directions. After A(lice) has sent all photons to B(ob), both communicate to
each other their choices of directions of polarization over a public channel.
However, the sent or measured polarizations of the photons are kept secret.
Whenever they chose the same direction (yes), their measured polarizations
are correlated perfectly and they keep the corresponding measured results
for their secret key. The other measurement results (no) cannot be used for
the key. Provided the transmission channel is ideal, A(lice) and B(ob) can
use part of the key for detecting a possible eavesdropper because in this case
some of the measurements are not correlated perfectly. In practice, however,
the transmission channel is not perfect and A(lice) and B(ob) have to process
their raw key further to extract from it a secret key [45]. It took some more
Table 1.1. Part of a possible idealized protocol for transmitting a secret key,
according to [42]
A(lice)’s direction i 1 2 1 1 2 1 2 2 1 2 ···

A(lice)’s polarization +1 −1 −1 +1 +1 +1 −1 −1 −1 +1 ···
B(ob)’s direction i 2 1 1 2 2 1 2 1 1 2 ···
B(ob)’s measured polarization +1 −1 −1 −1 +1 +1 −1 +1 −1 +1 ···
Public test of common direction No No Yes No Yes Yes Yes No Yes Yes ···
Secret key −1 +1 +1 −1 −1 +1 ···
years to realize that an exchange of secret keys can be achieved with the
help of entangled quantum states [46]. Thereby, the characteristic quantum
correlations of entangled states and the very fact that they are incompat-
8 Gernot Alber
ible with any LRT can be used for ensuring security of the key exchange.
After the first proof-of-principle experiments [47, 48], the first practical im-
plementation of quantum cryptography over a distance of about 1 km was
realized at the University of Geneva using single, polarized photons trans-
mitted through an optical fiber [49]. These developments launched the whole
new field of quantum cryptography. Now, this field represents the most devel-
oped part of quantum information processing. Quantum cryptography based
on the BB84 protocol has already been realized over a distance of 23 km
[50]. Recent experiments [30, 31] have demonstrated that photon pairs can
also be entangled over large distances, so that entanglement-based quantum
cryptography over such large distances might become accessible soon. Some
of these experiments are discussed in Chap. 3.
Simultaneously with these developments in quantum cryptography, nu-
merous other physical processes were discovered which were either enabled
by entanglement or in which entanglement led to an improvement of perfor-
mance. The most prominent examples are dense coding [51], entanglement-
assisted teleportation [10, 11, 52] and entanglement swapping [52, 53]. (These
processes are discussed in detail in Chaps. 2 and 3.) In the spirit of Feynman’s
suggestion, all these developments demonstrate that characteristic quantum
phenomena have practical applications in quantum information processing.
1.1.3 Quantum Algorithms
Feynman’s suggestion also indicates interesting links between quantum phy-

sics and computer science. After the demonstration [54] that quantum sys-
tems can simulate reversible Turing machines [55], the first quantum gener-
alization of Turing machines was developed [7]. (Turing machines are general
models of computing devices and will be discussed in detail in Chap. 4.) Fur-
thermore, it was pointed out that one of the remarkable properties of such
a quantum Turing machine is quantum parallelism, by which certain tasks
may be performed faster than by any classical computing device. Deutsch’s
algorithm [7, 56] was the first quantum algorithm demonstrating how the
interplay between quantum interference, entanglement and the quantum me-
chanical measurement process could serve this practical purpose.
The computational problem solved by Deutsch’s algorithm is the follow-
ing. We are given a device, a so-called oracle, which computes a Boolean
function f mapping all possible binary n-bit strings onto one single bit.
Therefore, given a binary n-bit string x as input, this oracle can compute
f (x) ∈ {0, 1} in a single step. Furthermore, let us assume that this function
is either constant or balanced. Thus, in the first case the 2n possible input
values of x are all mapped onto 0 or all onto 1. In the second case half of
the input values are mapped onto either 0 or 1 and the remaining half are
mapped onto the other value. The problem is to develop an algorithm which
determines whether f is constant or balanced.
Let us first of all discuss briefly the classical complexity of this problem.
In order to answer the question in the worst possible case, the oracle has to be
queried more than 2n−1 times. It can happen, for example, that the first 2n−1
queries all give the same result, so that at least one more query of the oracle
is required to decide whether f is constant or balanced. Thus, classically, it
is apparent that the number of steps required grows exponentially with the
number of bits.
|x> |x>
Uf
|a> |a f(x)>
Fig. 1.2. Basic operation of a quantum oracle Uf which evaluates a Boolean func-
tion f : x ∈ Zn 1
2 → f (x) ∈ Z2 ≡ {0, 1}; |x
is the input state of an n-qubit quantum
system; |a
is a one-qubit state and ⊕ denotes addition modulo 2
Quantum mechanically, the situation is different. The 2n possible binary

n-bit strings x can be represented by quantum states |x, which form a ba-
sis in a 2n -dimensional Hilbert space H2n , which is the state space of n
qubits. Furthermore, we imagine that the classical oracle is replaced by a
corresponding quantum oracle (Fig. 1.2). This is a unitary transformation
Uf which maps basis states of the form |x|a, where a ∈ {0, 1}, to output
states of the form |x|a ⊕ f (x) in a single step. Here, |a denotes the quan-
tum state of an ancilla qubit and ⊕ denotes addition modulo 2. If the initial
state is |x|0, for example, the quantum oracle performs an evaluation of
f (x), resulting in the final state |x|f (x). However, as this transformation
is unitary, it can perform this task also for any linear combination of possi-
ble basis states in a single step. This is the key idea of quantum parallelism
[7]. Deutsch’s quantum algorithm obtains the solution to the problem posed
above by the following steps (Fig. 1.3):
1. The n-qubit quantum system √ and the ancilla system are prepared in
states |0 and (|0 − |1)/ 2. Then a Hadamard transformation
1
H : |0 → √ (|0 + |1) ,
2
1
|1 → √ (|0 − |1) (1.12)
2
10 Gernot Alber
| ψ1 > |ψ2 >

measurement
|0> H H |0><0|
. . .
n . Uf . .
. . .
|0> H H |0><0|
1
(| 0 > - | 1 >)
2
Fig. 1.3. Schematic representation of Deutsch’s quantum algorithm
is applied to all of the first n qubits. We denote by H (i) the application

of H to the ith qubit. Thus, the separable quantum state
n
1 1
|ψ1 ≡ √ ⊗H (i) |0 (|0 − |1) = √ |x(|0 − |1) (1.13)
2 i=1 2n+1 x∈2n
is prepared.
2. A single application of the quantum oracle Uf to state |ψ1 yields the
quantum state
1
|ψ2 ≡ Uf |ψ1 = √ (−1)f (x) |x(|0 − |1) . (1.14)
2n+1 x∈2n
3. Subsequently a quantum measurement is performed to determine whether
the system is in state |ψ1 or not. With the help of n Hadamard trans-
formations (as in step 1), this quantum measurement can be reduced to
a measurement of whether the first n qubits of the quantum system are
in state |0 or not.
If in step 3 the quantum system is found in state |ψ1 , f is constant, otherwise

f is balanced. One of these two possibilities is observed with unit probability.
The probability p of observing the quantum system in state |ψ1 is given by
1
p ≡ | ψ1 |ψ2 |2 = n | (−1)f (x) |2 . (1.15)
2 x∈2n
Taking into account the single application of the quantum oracle in step 2
and the application of the Hadamard transformations in the preparation and
measurement processes, Deutsch’s quantum algorithm requires O(n) steps to
obtain the final answer, in contrast to any classical algorithm, which needs
an exponential number of steps. Thus Deutsch’s quantum algorithm leads to
an exponential speedup.
A key element of this quantum algorithm and of those discovered later is
the quantum parallelism involved in step 2, where the linear superposition
of the first n qubits comprises the requested global information about the
function f . For most of the possible functions f this intermediate quantum
state is expected to be entangled. An exception is the case of a constant func-
tion f , for which the quantum state |ψ2 is separable. Furthermore, it is also
crucial for the success of this quantum algorithm that the final measurement
in step 3 yielding the required answer can be implemented by a fast quantum
measurement whose complexity is polynomial in n. This is a requirement
fulfilled by all other known fast quantum algorithms. The quantum algo-
rithm described above was the first example demonstrating that quantum
phenomena may speed up computations in such a way that an exponential
gap appears between the complexity class of the quantum problem and the
complexity class of the corresponding classical probabilistic problem.
Continuing this development initiated by Deutsch, other, new fast quan-
tum algorithms were discovered in the subsequent years. The most prominent
examples are Simon’s quantum algorithm [57], Shor’s celebrated algorithm
[9] for factorizing numbers, and Grover’s search algorithm [58]. (Quantum al-
gorithms are discussed in detail in Chap. 4.) In addition, possible realizations
of quantum computing devices were suggested which were based on trapped
ions [59] and on cavity quantum electrodynamical setups [60]. These devel-
opments called for new methods for stabilizing quantum algorithms against
perturbing environmental influences, which tend to destroy quantum inter-
ference and quantum entanglement [61]. This led to the development of the
first error-correcting codes [62, 63, 64, 65, 66] by adaptation of classical error-
correcting techniques to the quantum domain. An introduction to the theory
of quantum error correction is presented in Chap. 4.
1.2 Quantum Physics and Information Processing

What are the common features of these early developments? The common
element of these early developments in quantum cryptography and quan-
tum computation is that they all involve the practical processing of informa-
tion and they are all founded on and facilitated by characteristic quantum
phenomena. These phenomena, among which the most prominent is entan-
glement, are in conflict with the classical concepts of physical reality and
locality. Obviously, these early developments hint at a profound connection
between the concept of information and some fundamental concepts of quan-
tum theory, which is also promising from the technological point of view.
It is these technologically oriented aspects of quantum information theory
[67, 68, 69] which are at the heart of quantum information processing.
Methods for processing quantum information have developed rapidly dur-
ing the last few years [12]. Owing to significant experimental advances, ba-
sic interference and entanglement phenomena which are of central interest
for processing quantum information have been realized in the laboratory in
various physical systems. Basic schemes for quantum communication have
12 Gernot Alber
been demonstrated with photons [10, 11, 49, 70]. Realizations of elementary
quantum logical operations have been based on trapped ions [13, 14] and on
nuclear magnetic resonance [15]. Recent experiments indicate that besides
cavity quantum electrodynamical setups [16], trapped neutral atoms which
are guided along magnetic wires (atom chips) might also be useful for quan-
tum information processing [17]. There have also been theoretical proposals
on using ultracold atoms in optical lattices [18, 19], on ions in an array of
microtraps [20] and on solid-state devices [21, 22, 23] for the implementation
of quantum logical gates.
By now, quantum information processing has become an interdisciplinary
subject which attracts not only physicists but also researchers from other
communities. The common interest is the practical, technologically oriented
application of characteristic quantum phenomena. At this stage of develop-
ment, it appears necessary to examine recent achievements and to emphasize
the underlying, general, basic concepts, which have been developing gradu-
ally and which are now commonly adopted by all researchers in this field.
This is one of the main intentions of the rest of the book.
In Chap. 2, Werner introduces the basic concepts of quantum information
theory and describes the fundamental mathematical structures underlying re-
cent and current developments. In particular, this chapter addresses a natural
question appearing in connection with Feynman’s suggestion, namely what
can be done with the help of quantum systems and what cannot be done. A
first example of an impossible quantum process, the copying of nonorthogonal
quantum states, has already been mentioned. Other examples of possible and
impossible quantum processes are discussed in detail in this contribution.
First experimental realizations of basic quantum communication schemes
based on entangled photon pairs are discussed in Chap. 3 by Weinfurter and
Zeilinger. These first experiments on entanglement-based quantum cryptog-
raphy, dense coding and quantum teleportation demonstrate the important
role photons play in current experiments. Furthermore, these experiments
also emphasize once again the fundamental significance of entanglement for
quantum information processing.
The basic theoretical concepts of quantum computation and the mathe-
matical structure underlying quantum algorithms are discussed in Chap. 4
by Beth and Rötteler. In particular, it is demonstrated how recent results in
the theory of signal processing can be used for the development of new fast
quantum algorithms. A short introduction to the theory of quantum error
correction is also presented.
A comprehensive account of the mathematical structure of entanglement
and of the significance of mixed entangled states for quantum information
processing is presented in Chap. 5 by M. Horodecki, P. Horodecki and R.
Horodecki. One of the most surprising recent developments in this context
has been the discovery of bound entanglement [71]. Though much is still
unknown, this section gives a state-of-the-art presentation of what is known
about this new form of entanglement and its implications for processing quan-
tum information.
2 Quantum Information Theory
– an Invitation
Reinhard F. Werner
2.1 Introduction
Quantum information and quantum computers have received a lot of public

attention recently. Quantum computers have been advertised as a kind of
warp drive for computing, and indeed the promise of the algorithms of Shor
and Grover is to perform computations which are extremely hard or even
provably impossible on any merely “classical” computer. On the experimental
side, perhaps the most remarkable feat of quantum information processing
was the realization of “quantum teleportation”, which once again has science
fiction overtones.
In some sense these miracles are an extension of the strangeness of quan-
tum mechanics – those unresolved questions in the foundations of quantum
mechanics, which most physicists know about, but few try to tackle directly
in their research. However, trying to build an explanation of quantum in-
formation on the literature about the foundations of quantum mechanics is
more likely to mystify than to clarify. It would also give a wrong idea of
how discussions in this new field are conducted. Because, just as physicists
with widely differing convictions on foundational matters can usually agree
quite easily on what the predictions of quantum mechanics are in a partic-
ular experimental setup, researchers in quantum information can agree on
whether a device should work, no matter what they may think about the
deeper meaning of the wave function. For example, one of the founders of the
field is an outspoken proponent of the many-worlds interpretation of quan-
tum mechanics (which I, personally, find useless and bizarre). But, whatever
the intuitions leading him to his discoveries about quantum computing may
have been, these discoveries make sense in every other interpretation.
In this article I shall give an account of the basic concepts of quantum
information theory, staying as much as possible in the area of general agree-
ment. So, in order to enter this new field, plain quantum mechanics is enough,
and no new, perhaps obscure, views are needed. There is, of course, a charac-
teristic shift in emphasis expressed by the word “information”, and we shall
have to explore the consequences of this shift.
The article is divided into two parts. The first (up to the end of Sect. 2.5)
is mostly in plain English, centered around the exploration of what can or
cannot be done with quantum systems as information carriers. The second

2 Quantum Information Theory – an Invitation 15
part, Sect. 2.6, then gives a description of the mathematical structures and
of some of the tools needed to develop the theory.
2.2 What is Quantum Information?

Let us start with a preliminary definition:
Quantum information is that kind of information which is carried

by quantum systems from the preparation device to the measuring
apparatus in a quantum mechanical experiment.
So a “transmitter” of quantum information is nothing but a device preparing

quantum particles, and a “receiver” is just a measuring device. Of course, this
is not saying much. But even so, it is a strange statement from the point of
view of classical information theory: in that theory one usually does not care
about the physical carrier of the information, or else one would have to dis-
tinguish “electrodynamical information”, “printed information”, “magnetic
information” and many more. In fact, the success of (classical) information
theory depends largely on abstracting from the physical carrier, and going
instead for the general principles underlying any information exchange. So
why should “quantum information” be any different?
A moment’s reflection makes clear why the abstraction from the physical
carrier of information leads to a successful theory: the reason is that it is so
easy to convert information between all such carriers. The conversion from
bytes on a hard disk, to currents in a chip, to signals on a cable, to radio
waves via satellite and maybe, finally, to an image on a computer screen in
another continent all happens essentially without loss, and if there are losses,
they are well understood, and it is known how to correct for them. Therefore,
the crucial question is: can “quantum information” in the above loose sense
also be converted to those standard classical kinds of information, and back,
without loss? Or: are there fundamental limitations to such a translation,
and is quantum information hence really a new kind of information?
This book would not have been written if the answer to the last question
were not affirmative: quantum information is indeed a new kind of informa-
tion. But to make this precise, let us see what would be required of a suc-
cessful translation. Let us begin with the conversion of quantum information
to classical information: a device for this conversion would take a quantum
system and produce as its output some classical information. This is nothing
but a complicated way of saying “measurement”. The reverse translation,
from classical to quantum information, obviously involves some preparation
of quantum systems. The classical input to such a device is used to control
the settings of the preparing device, and any dependence of the preparation
process on classical information is admissible. There are two kinds of de-
vices we can combine from these two elements. Let us first consider a device
going from classical to quantum to classical information. This is a rather
16 Reinhard F. Werner
M P
Fig. 2.1. Classical teleportation. Here and in the following diagrams, a wavy arrow
stands for quantum systems, and a straight arrow for the flow of classical informa-
tion
commonplace operation. For example, one can encode one classical bit in
the polarization degree of freedom of a photon (clearly a quantum system),
by choosing one of two orthogonal polarizations for the photon, depending
on the value of the classical bit. The readout is done by a photomultiplier
combined with a polarization filter in one of the corresponding directions. In
principle, this allows a perfect transmission. In some sense every transmis-
sion of classical information is of this kind, because every physical system
ultimately obeys the laws of quantum mechanics, even if we can often dis-
regard this fact and treat it classically. Hence classical information can be
translated into quantum information (and back).
But what about the converse? This hypothetical (and in fact, impossi-
ble) process has come to be known as classical teleportation (see Fig. 2.1). It
would involve a measuring device M, operating on some input quantum sys-
tems. The results of the measurements are subsequently fed into a preparing
device P, which produces the final output of the combined device. The task
is to set things up such that the outputs of the combined device are indistin-
guishable from the quantum inputs. Of course, we have to say precisely what
“indistinguishable” should mean. Clearly, this cannot mean that “the same”
system comes out at the other end. In the classical case this is not demanded
either. What can only be meant in quantum mechanics is that no statistical
test will see the difference. In other words, no matter what the preparation
of the input systems is and no matter what observable we measure on the
outputs of the teleportation device, we shall always get the same probability
distribution of results as if the inputs had been directly measured. Note also
that this criterion does not involve the states of individual systems, but only
states in the form of the distribution parameters of ensembles of identically
prepared systems.
The impossibility of classical teleportation will be treated extensively in
the following section, where it is related to a hierarchy of impossible machines.
For a mathematical statement of this impossibility in the standard quantum
formalism of quantum mechanics, see the remark after (2.7). For the moment,
however, let us take it for granted, and see what all this says about the new
concept of quantum information.
First of all, we are concerned here with problems of transmission, not with
content or meaning. This is exactly the same as in classical information the-
ory. There, too, it is often not easy to avoid confusion with a different concept
of “information” used in everyday language, namely the kind available at an

information desk. Information theory does not care whether a TV channel
is used for “misinformation”, but can say everything about what it takes
to ensure the technical quality of the final images. Hence the quantitative
measures of “information” all relate to storage and transmission capacity, to
the possibilities of compression and error correction and so on. In the same
vein, quantum information theory will not tell us what the meaning of a
“quantum message” is, and this is probably meaningless anyway, because a
message that has been “read” is classical almost by definition. But quantum
information theory has precise notions of the resources needed to transmit
such information faithfully.
Secondly, transmission of quantum information is not at all an exotic
concept in the context of modern physics. It can be paraphrased in various,
perhaps more familiar ways, for example as “transmission of intact quantum
states”, as “coherent transmission of quantum systems” or as transmission
“preserving all interference possibilities” of the system. Nevertheless the in-
formation metaphor is useful, not only because it suggests new applications,
but also because it leads one to ask new questions, and leads to quantitative
notions where previously there was only a qualitative understanding. And
possibly this even provides a way to see in a sharper light the old conun-
drums of the foundations of quantum mechanics.
2.3 Impossible Machines
The usefulness of considering impossible machines is well known from ther-

modynamics: the second law of thermodynamics is often stated as the impos-
sibility of a perpetual-motion machine. The theorem of the impossibility of
classical teleportation is likewise a fundamental law of quantum mechanics,
and a lot can be learned from analyzing it. Typically, the impossible ma-
chines of quantum theory are perfectly possible in classical physics, so their
impossibility does not follow superficially from their description, but rather
carries a connotation of paradox.
We shall discuss a range of impossible tasks, consisting of
• teleportation
• copying (“cloning”)
• joint measurement
• Bell’s telephone.
As we shall see, teleportation is the most powerful of these, in the sense that
if we had a teleportation device, we could build a quantum copier, from which
we could in turn construct joint measurements and, finally, a device known
as Bell’s Telephone, by which we could set up superluminal communication.
Hence, if we uphold the principle of causality, which forbids the weakest
machine in this hierarchy, we are certain that teleportation is likewise im-

possible. In this section we shall follow this line of reasoning to prove the
impossibility of teleportation. Of course, there are other, more direct ways of
proving this result from the structure of quantum mechanics. However, these
usually require more of the quantum formalism and give less insight into the
differences between classical and quantum information.
2.3.1 The Quantum Copier
This is the machine referred to in the well-known paper of Wootters and

Zurek entitled “A single quantum cannot be cloned” [4]. By definition, a
copier would be a device taking one quantum system as input and turning out
two systems of the same type. The condition for calling this a (faithful) copier
is that we would not be able to distinguish a system coming from the output
from the input system by any statistical test, i.e. by means of the probabilities
measured for any observable, and for any preparation of the initial state.
Hence the device has to operate on arbitrary “unknown” states. It is clear
that a copier in the ordinary sense, e.g. a mail relay distributing email to
several recipients, indeed satisfies this condition in the domain of classical
information. Note that we are not so unreasonable as to demand what the
paper quoted above suggests, namely that we could test this device on single
events, or even assume some ontological “identity” of input and output: the
criterion for faithful copying is flatly statistical, and can be verified by a
straightforward collection of statistical tests.
Given a teleportation device, building a copier is quite easy (see Fig. 2.2).
All we have to do is to remember that the classical information obtained in
the intermediate stage of the teleportation process can be copied perfectly.
Hence we can apply the measuring device of the teleportation line to the
input system, copy the results, and simply run the reconstructing preparation
process on each of these copies.
P
M = C
P
Fig. 2.2. Making a copier from a “classical teleportation” line
2.3.2 Joint Measurement
This is the task of combining two separate measuring devices into a single
device, or the “simultaneous measurement” of two quantum observables A
and B. Thus, a joint measuring device “A&B” is a device giving a pair (a, b)
of classical outputs each time it is operated, such that a is a possible output
of A, and b is a possible output of B. (We use the symbol A to denote both an
observable and a device that measures this observable, and similar for B.) We
require that the statistics of the a outcomes alone are the same as for device
A, and similarly for B. Note that once again our criterion is statistical, and
can be tested without recourse to counterfactual conditionals such as “the
result which would have resulted if B rather than A had been measured on
this particular quantum particle”.
Many quantum observables are not jointly measurable in this sense. The
most famous examples, position and momentum, different components of
angular momentum, and positions of a free particle at different times, are
probably contained in every quantum mechanics course. Hence the impossi-
bility of joint measurements is nothing but a precise statement of an aspect
of “complementarity”.
Nevertheless, a joint measurement device for any of these could readily
be constructed given a functioning quantum copier (see Fig. 2.3): one would
simply run the copier C on the quantum system, and then apply the two given
measuring devices, A and B, to the copies. It is easy to see that the definition
of the copier then guarantees that the statistics of a and b separately come
out right. In other words, a copier can be seen as a universal joint measuring
device.
C =
Fig. 2.3. Obtaining joint measurements from a copier
2.3.3 Bell’s Telephone
This is not named after a certain phone company, but after John S. Bell,
who never proposed it in this form, but might have. It refers to a project of
performing superluminal communication using only correlations of the type

tested by Bell’s inequalities. Without going into details for the moment, the
basic setup would consist of a source producing pairs of particles and sending
one member of each pair to each of the two communicating parties, conven-
tionally named “Alice” and “Bob”. Each of them has a collection of different
measuring devices to choose from, and the idea is for Alice to do some-
thing which creates a noticeable change in the probabilities measured by
Bob. Clearly, this is a paradoxical task, because no particle or other physi-
cal carrier of information actually goes from Alice to Bob. Therefore, if the
particles move sufficiently far apart from one another, this device transmits
superluminally.
It is maybe useful to point out here a common confusion concerning such
superluminal effects, which sometimes even afflicts otherwise reliable profes-
sional writers. The mistake can usually be spotted easily by a device I call
the “ping-pong ball test”. It goes like this:
Take an author’s explanation of Bell’s inequalities, and substitute
“ping-pong balls” for every quantum particle. Then if whatever the
author is selling as paradoxical remains true, he/she hasn’t under-
stood a thing.
Here is an example: imagine a box containing a ping-pong ball; the box can
be separated into two parts, without anyone looking at the ball. One part is
shipped to Tokyo or Alpha Centauri, without anyone looking inside. Then if
I open the other box I know instantly, i.e. “at superluminal speed”, whether
the ball is at the distant location or not. Of course, this is true but hardly
paradoxical, and is totally useless for sending a message either way. To repeat:
there is nothing paradoxical in statistical correlations per se between distant
systems with a common past, even if the correlation is perfect.
If Alice wants to send a message to Bob, correlations between two mea-
suring devices are useless, because they cannot even be detected without
comparing the results, which requires exactly the communication the Tele-
phone was intended for. Only if something Alice does has an effect on the
measurement results at Bob’s end can we speak of communication. The only
thing Alice can do in the standard setup is to choose a measuring device, and
Bell’s Telephone can be said to work if these choices have an influence on the
probabilities measured by Bob (who has no access to Alice’s measurement
results). If there is no physical system traveling from Alice to Bob, however,
this will be impossible.
To be fair, this can hardly be counted as an impossible machine of quan-
tum mechanics, since the argument has nothing to do with quantum theory.
What makes it fit into the hierarchy described here is the following: if we
assume that Bob has a joint measuring device for two yes/no measurements,
and Bell’s inequalities are violated, we can design a strategy for Alice to send
signals to Bob with better than chance results. Hence the joint measurement
of suitable observables can provide a device sufficiently strong to achieve a
S ?
=
Fig. 2.4. Building Bell’s Telephone from a joint measurement
task forbidden by causality, and hence is impossible in general. This is the

link between the last two elements in the hierarchy of impossible machines
mentioned at the beginning of Sect. 2.3.
The proof of this step amounts to yet another derivation of Bell’s inequal-
ities, but since it emphasizes the communication aspect it fits well into our
context, and we shall at least sketch it. This step will be rather more technical
than the rest of this section, but does not require any quantum theory. The
argument can be skipped without loss as far as later sections are concerned.
So let us assume that Alice and Bob each have at their disposal two
measuring devices, say A1 , A2 and B1 , B2 , respectively. Each of these can give
a result of either +1 or −1. We shall denote by P(a, b | Ai , Bj ) the probability
for Alice to obtain a and Bob to obtain b in a correlation experiment in which
Alice uses measuring device Ai and Bob uses Bj . By

C(Ai , Bj ) = ab P(a, b | Ai , Bj )
a,b
we shall denote the correlation coefficient, which lies between −1 and +1.
The combination
β = C(A1 , B1 ) + C(A1 , B2 ) + C(A2 , B1 ) − C(A2 , B2 ) (2.1)
carries special significance, as we shall see below. Because the inequality β ≤ 2

is known as the Bell inequality (see Sect. 1.1.1), we shall call β the Bell corre-
lation for this choice of four observables. It is a quantity directly accessible to
experiment. Note that Bob usually cannot tell from his data which apparatus
(A1 or A2 ) Alice chose. This is reflected by the equation

P(a, b | A1 , Bj ) = P(a, b | A2 , Bj ) ≡ P(b | Bj ) ,
a a
and is borne out by all known experimental data. Now suppose Bob has a
joint measuring device for his B1 and B2 , which we shall denote by B1 &B2 ,
which produces pair outcomes (b1 , b2 ) (see Fig. 2.4). We can then determine
the probabilities pi (ai , b1 , b2 ) = P(ai , (b1 , b2 ) | Ai , B1 &B2 ). The condition

that this is really a joint measurement is expressed by the equations

pi (ai , b1 , b2 ) = P(ai , b2 | Ai , B2 ) and (2.2)
b1

pi (ai , b1 , b2 ) = P(ai , b1 | Ai , B1 ) , (2.3)
b2
each for i = 1, 2. The basic rule for the information transmission is the
following:
Alice encodes the bit she wants to send by choosing either apparatus
A1 or apparatus A2 . Then Bob looks at his readout and interprets it
as “A1 ” whenever the two displays coincide (b1 = b2 ), and as “A2 ”
if they are different.
We can then estimate the probability pok for Bob to be right, assuming
that the choices A1 and A2 are made with the same frequency. Assume first
that Alice chooses A1 . Then Bob is right with probability
b1 + b2

2 |a1 | p1 (a1 , b1 , b2 ) ,
a1 ,b1 ,b2
where the first factor takes into account the condition b1 = b2 , and the second
is introduced for later convenience. Combining this with a second term of
similar kind for Alice’s choice A2 , and taking into account the probability
1/2 for each of these choices, we obtain the overall probability pok for Bob to
be correct as

1 b1 + b2
pok = 2 |a1 | p1 (a1 , b1 , b2 )
2
a1 ,b1 ,b2

1 b1 − b2
+
2 2 |a2 | p2 (a2 , b1 , b2 )
a2 ,b1 ,b2
1
≥ (b1 + b2 )a1 p1 (a1 , b1 , b2 )
4
a1 ,b1 ,b2
1
+ (b1 − b2 )a1 p2 (a2 , b1 , b2 )
4
a2 ,b1 ,b2
1
= C(A1 , B1 ) + C(A1 , B2 ) + C(A2 , B1 ) − C(A2 , B2 )
4
β
= . (2.4)
4
Bob is right with a better probability than chance if pok > 1/2, which, by this
computation, can be guaranteed if β > 2, i.e. if the classical Bell inequality
(in Clauser–Horne–Shimony–Holt form [72]) is violated. But this is indeed the
case in√experiments conducted to determine β (e.g. [73]), which give roughly

β ≈ 2 2 ≈ 2.8. If we believe these experiments, the only conclusion can be
that the joint measurability of the B1 and B2 used in the experiment would
be sufficient to make Bell’s Telephone work, which was our claim.
2.3.4 Entanglement, Mixed-State Analyzers

and Correlation Resolvers
Violations of Bell’s inequalities can also be seen to prove the existence of a

new class of correlations between quantum systems, known as entanglement.
This concept is as fundamental to the field of quantum information theory
as the idea of quantum information itself. So rather than organizing this
introduction as an answer to the the question “why quantum information is
different from classical information”, we could have followed the line “why
entanglement is different from classical correlation”. There are impossible
machines in this line of approach, too, and we shall now describe briefly how
they fit in.
Consider a correlation experiment of the kind used in the study of Bell’s
inequalities (see Sect. 2.3.3). If Bob looks at his particles, and makes mea-
surements on them without any communication from Alice, he will find that
their statistics are described by a certain mixed state. The state must be
mixed, because if he now listens to Alice and sorts his particles according
to Alice’s measurement results, he will get two subensembles, which are in
general different. In the usual ideal 2-qubit situation, in which one obtains
the maximal violation of Bell’s inequalities, these subensembles are described
by pure states.
This is very satisfying for people who see the occurrence of mixed states
in quantum mechanics merely as a result of ignorance, as opposed to the
deeper kind of randomness encoded in pure states. This view usually comes
with an individual-state interpretation of quantum mechanics, in which each
individual system can be assigned a pure state (a single vector in Hilbert
space), and a general preparation procedure is given not just by its density
matrix, but by a specific probability distribution of pure states. Let us use
the term mixed-state analyzer for a hypothetical device which can see the
difference, i.e. a measuring device whose output after many measurements
on a given ensemble is not just a collection of expectations of quantum ob-
servables, but the distribution of pure states in the ensemble. In the case of a
correlation experiment, where Bob sees a mixed state only because he is igno-
rant about Alice’s results, this machine would find for him the decomposition
of his mixed state into two pure states.
The problem is, of course, that Alice has several choices of measuring de-
vices, and that the decomposition of Bob’s mixed state depends, accordingly,
on Alice’s choice. Hence she could signal to Bob, and we would have another
instance of Bell’s Telephone. There would be a way out if joint measurements
were available (to Alice in this case): then we could say that the two decom-
positions were just the first step in an even finer decomposition, a further
reduction of ignorance, which would be brought to light if Alice were to ap-
ply her joint measurement. Presumably the mixed-state analyzer would then
yield this finer decomposition, because the operation of this device would not
depend on how closely Alice cared to look at her particles.
But just as two quantum observables are often not jointly measurable, two
decompositions of mixed states often have no common refinement (actually,
in the formalism of quantum theory, these are two variants of the same the-
orem). In particular, the two decompositions belonging to Alice’s choices in
an experiment demonstrating a violation of Bell’s inequalities have no com-
mon refinement, and any mixed-state analyzer could be used for superluminal
communication in this situation.
Another device, which is suggested by the individual-state interpretation,
arises from a naive extrapolation of this view to the parts of a composite
system: if every single system could be assigned a pure state, a composite
system could be assigned a pair of pure states, one for each subsystem. A
correlated state should therefore be given by a probability distribution of
such pairs. A device which represented an arbitrary state of a composite
system as a mixture of uncorrelated pure product states might be called a
correlation resolver. It could be built given a classical teleportation line: when
one applies teleportation to one of the subsystems and applies conditions
on the classical measurement results of the intermediate stage, one obtains
precisely a representation of an arbitrary state in this form. But it is easy to
see that any state which can be so analyzed automatically satisfies all Bell-
type inequalities, and hence once again the experimental violations of Bell’s
inequalities show that such a correlation resolver cannot exist. Hence we
have here a second line of reasoning in favor of the no-teleportation theorem:
a teleportation device would allow classical correlation resolution, which is
shown to be impossible by the Bell experiments.
The distinction between resolvable states and their complement is one
of the starting points of entanglement theory, where the “resolvable” states
are called “separable”, or “classically correlated”, and all others are called
“entangled”. For a more detailed treatment and an up-to date overview, the
reader is referred to Chap. 5.
Without going into philosophical discussions about the foundations of
quantum mechanics, I should like to comment briefly on the individual-state
interpretation, which has suggested the two impossible machines discussed in
this subsection. First, this view is not at all uncommon, and it is quite possi-
ble to read some passages from the masters of the Copenhagen interpretation
as an endorsement of this view. Secondly, if we define a hidden-variable theory
as a theory in which individual systems are described by classical parame-
ters, whose distribution is responsible for the randomness seen in quantum
experiments, we have no choice but to call the individual-state interpreta-
tion a hidden-variable theory. The hidden variable in this theory is usually

denoted by ψ. And sure enough, as we have just pointed out, this theory
has all the difficulties with locality such a theory is known to have on gen-
eral grounds. Thirdly, avoiding an individual-state interpretation, and with
it some of its misleading intuitions, is easy enough. In practice this is done
anyhow, by concentrating on those aspects of the theory which have some
direct statistical meaning, and not on these involving hypothetical, and usu-
ally impossible devices. This common ground is the statistical interpretation
of quantum mechanics, in which states (pure or mixed) are the analogues
of classical probability distributions, and are not seen as a property of an
individual system, but of a specific way of preparing the system.
2.4 Possible Machines

2.4.1 Operations on Multiple Inputs
The no-teleportation theorem derived in the previous section says that there
is no way to measure a quantum state in such a way that the measuring
results suffice to reconstruct the state. At first sight this seems to deny that
the notion of “quantum states” has an operational meaning at all. But there is
no contradiction, and we shall resolve the apparent conflict in this subsection,
if only to sharpen the statement of the no-teleportation theorem.
Let us recall the operational definition of quantum states, according to
the statistical interpretation of quantum mechanics. A state is a description
of a way of preparing quantum systems, and in all its aspects it is related to
computing expectation values. We might also say that it is the assignment of
an expectation value to every observable of the system. So to the extent that
expectation values can be measured, it is possible to determine the state by
testing it on sufficiently many observables. What is crucial, however, is that
even the determination of a single expectation value is a statistical measure-
ment. Hence such a determination requires a repetition of the experiment
many times, using many systems prepared according to the same procedure.
In contrast, the above description of teleportation demands that it works
with a single quantum system as input, and that the measuring device does
not accumulate results from several input systems. Expressed in the current
jargon, teleportation is required to be a one-shot operation. Note that this
does not contradict our statistical criteria for the success of teleportation and
of other devices, which involve a statistics of independent “single shots”.
If we have available many identically prepared systems, many operations
which are otherwise impossible become easy. Let us begin with classical tele-
portation. Its multiinput analogue is the state estimation problem: how can
we design a measurement operating on samples of many (say, N ) systems
from the same preparing device, such that the measurement result in each
case is a collection of classical parameters forming a Hermitian matrix which
on average is close to the density matrix describing the initial preparation.

This is symbolized in Fig. 2.5 (with the box T omitted for the moment):
the box P at the end represents a repreparation of systems according to the
estimated density matrix. The overall output will then be a quantum system,
which can be directly compared with the inputs in statistical experiments. It
is clear that the state cannot be determined exactly from a sample with fi-
nite N , but the determination becomes arbitrarily good in the limit N → ∞.
Optimal estimation observables are known in the case when the inputs are
guaranteed to be pure [74], but in the case of general mixed states there are
no clear-cut theorems yet, partly owing to the fact that it is less clear what
“figure of merit” best describes the quality of such an estimator.
Estimate
T P
Fig. 2.5. Classical teleportation with multiple inputs, or state estimation
Given a good estimator we can, of course, proceed to good cloning by just

repeating the repreparation P as often as desired. The surprise here [75] is
that if only a fixed number M of outputs is required, it is possible to obtain
better clones with devices that stay entirely in the quantum world than by
going via classical estimation. Again, the problem of optimal cloning is fully
understood for pure states [76], but work has only just begun to understand
the mixed-state case.
Another operation which becomes accessible in this way is the universal
not operation, assigning to each pure qubit state the unique pure state or-
thogonal to it. Like time reversal, this is just a special case of an antiunitary
symmetry operation. In this case, a strategy using a classical estimation as
an intermediate step can be shown to be optimal [77]. In this sense “universal
not” is a harder task than “cloning”.
More generally, we can look at schemes such as those in Fig. 2.5, where T
represents any transformation of the density matrix data, whether or not this
transformation corresponds to a physically realizable transformation of quan-
tum states. A further interesting application is to the purification of states.
In this problem it is assumed that the input states were once pure, but were
later corrupted in some noisy environment (the same for all inputs). The task
is to reconstruct the original pure states. Usually, the noise corresponds to an
invertible linear transformation of the density matrices, but its inverse is not
a possible operation, because it transforms some density matrices to opera-
tors with negative eigenvalues. So the reversal of noise is not possible with a
one-shot device, but is easy to perform to high accuracy when many equally
prepared inputs are available. In the simplest case of a so-called depolarizing
channel, this problem is well understood [78]; it is also well understood in the
version requiring many outputs, as in the optimal-cloning problem [79].
2.4.2 Quantum Cryptography
It may seem impossible to find applications of impossible machines. But that

is not quite true: sometimes the impossibility of a certain task is precisely
what is called for in an application. A case in point is cryptography: here
one tries to make the deciphering of a code impossible. So if we could design
a code whose breaking would require one of the machines described in the
previous section, we could guarantee its security with the certainty of natural
law. This is precisely what quantum cryptography sets out to do. Because
only small quantum systems are involved it is one of the “easiest” applica-
tions of quantum information ideas, and was indeed the first to be realized
experimentally. For a detailed description we refer to Chap. 3. Here we just
describe in what sense it is the application of an impossible machine.
As always in cryptography, the basic situation is that two parties, Alice
and Bob, say, want to communicate without giving an “evil eavesdropper”,
conventionally named Eve, a chance to listen in. What classical eavesdrop-
pers do is to tap the transmission line, make a copy of what they hear for
later analysis, and otherwise let the signal pass undisturbed to the legitimate
receiver (Bob). But if the signal is quantum, the no-cloning theorem tells
us that faithful copying is impossible. So either Eve’s copy or Bob’s copy is
corrupted. In the first case Eve won’t learn anything, and hence there was
no eavesdropping anyway. In the second case Bob will know that something
may have gone wrong, and will tell Alice that they must discard that part
of the secret key which they were exchanging. Of course, intermediate situa-
tions are possible, and one has to show very carefully that there is an exact
trade-off between the amount of information Eve can get and the amount of
perturbation she must inflict on the channel.
2.4.3 Entanglement-Assisted Teleportation
This is arguably the first major discovery in the field of quantum informa-
tion. The no-cloning and no-teleportation theorems, although they had not
been formulated in such terms, would hardly have come as a surprise to peo-
ple working on the foundations of quantum mechanics in the 1960s, say. But
entanglement assistance was really an unexpected turn. It was first seen by
Bennett et al. [52], who also coined the term “teleportation”. It is gratifying
to see, though it is hardly a surprise on the same scale, that this prediction of
quantum mechanics has also been implemented experimentally. The experi-
ments are another interesting story, which will no doubt be told much better
Alice Bob
1 qubit 2 bit 1 qubit
M P
1 ebit
S
Fig. 2.6. Entanglement-assisted teleportation
in Chap. 3 by Weinfurter and Zeilinger, who represent one team in which a

major breakthrough in this regard was achieved.
The teleportation scheme is shown in Fig. 2.6. What makes it so surpris-
ing is that it combines two machines whose impossibility was discussed in the
previous section: omitting the distribution of the entangled state (the lower
half of Fig. 2.6), we get the impossible process of classical teleportation. On
the other hand, if we omit the classical channel, we get an attempt to transmit
information by means of correlations alone, i.e. a version of Bell’s Telephone.
Since the time dimension is not represented in this diagram, let us consider
the steps in the proper order. The first step is that Alice and Bob each receive
one half of an entangled system. The source can be a third party or can be
Bob’s lab. The last choice is maybe best for illustrative purposes, because it
makes clear that no information is flowing from Alice to Bob at this stage.
Alice is next given the quantum system whose state (which is unknown to
her) she is to teleport. Alice then makes a measurement on a system made
by combining the input and her half of the entangled system. She sends the
results via a classical channel to Bob, who uses them to adjust the settings
on his device, which then performs some unitary transformation on his half
of the entangled system. The resulting system is the output, and if every-
thing is chosen in the right way, these output systems are indeed statistically
indistinguishable from the outputs. To see just how the entangled state S,
the measurement M and the repreparation P have to be chosen requires the
mathematical framework of quantum theory. In the standard example one
teleports a state of one qubit, using up one maximally entangled two-qubit
system (in the current jargon, “1 ebit”) and sending two classical bits from
Alice to Bob. A general characterization of the teleportation schemes for
qubits and higher-dimensional systems is given below in Sect. 2.6.6.
2.4.4 Superdense Coding
It is easy to see and in fact is a commonplace occurrence that classical in-

formation can be transmitted on quantum channels. For example, one bit of
classical information can be coded in any two-level system, such as the polar-
ization degree of freedom of a photon. It is not entirely trivial to prove, but
hardly surprising that one cannot do better than “one bit per qubit”. Can we
beat this bound using the idea of entanglement assistance? It turns out that
one can. In fact, one can double the amount of classical information carried
by a quantum channel (“two bits per qubit”). Remarkably, the setups for do-
ing this are closely related to teleportation schemes, and in the simplest cases
Alice and Bob just have to swap their equipment for entanglement assisted
teleportation. This is explained in detail in Sect. 2.6.6.
2.4.5 Quantum Computation
Again, we shall be very brief on this subject, although it is certainly central

to the field. After all, it is partly the promise of a fantastic new class of
computers which has boosted the interest in quantum information in recent
years. But since in this book computation is covered in Chap. 4, we shall only
make a few remarks connecting this subject to the theme of possible versus
impossible machines.
So can quantum computers perform otherwise impossible tasks? Not re-
ally, because in principle we can solve the dynamical equations of quantum
mechanics on a classical computer and simulate all the results. Hence classi-
cally unsolvable problems such as the halting problem for Turing machines
and the word problem in group theory cannot be solved on quantum com-
puters either. But this argument only shows the possibility of emulating all
quantum computations on a classical computer, and omits the possibility
that the efficiency of this procedure may be terrible. The great promise of
quantum computation lies therefore in the reduction of running time, from
exponential to polynomial time in the case of Shor’s factorization algorithm
[80]. This reduction is comparable to replacing the task of counting all the
way up to a 137 digit number by just having to write it. No matter what the
constants are in the growth laws for the computing time (and they will prob-
ably not be very favorable for the quantum contestant), the polynomial time
is going to win if we are really interested in factoring very large numbers.
A word of caution is necessary here concerning the impossible/possible
distinction. While it is true that no polynomial-time classical factoring algo-
rithm is known, and this is what counts from a practical point of view, there
is no proof that no such algorithm exists. This is a typical state of affairs
in complexity theory, because the nonexistence of an algorithm is a state-
ment about the rather unwieldy set of all Turing machine programs. A proof
by inspecting all of them is obviously out, so it would have to be based on
some principle of “conservation of difficulties”, which rarely exists for real-

life problems. One problem in which this is possible is that of identifying
which (unique) element of a large list has a certain property (“needle in a
haystack”). In this case the obvious strategy of inspecting every element in
turn can be shown to be the optimal classical one, and has a running time
proportional to the length N of the list. √ But Grover’s quantum algorithm
[58] performs the task in the order of N steps, an amazing gain even if it
is not exponential. Hence there are problems for which quantum computers
are provably faster than any classical computer.
So what makes the reduction of running time work? This is not so easy to
answer, even after working through Shor’s algorithm and verifying the claim
of exponential speedup. Massive entanglement is used in the algorithm, so
this is certainly one important element. Then there is a technique known as
quantum parallelism, in which a quantum computation is run on a coherent
superposition of all possible classical inputs, and, in a sense, all values of a
function are computed simultaneously. A catchy paraphrase, attributed to
D. Deutsch, is to call this a computation in the parallel worlds of the many-
worlds interpretation.
But perhaps the best way to find out what powers quantum computation
is to to turn it around and to really try the classical emulation. The practi-
cal difficulty which then becomes apparent immediately is that the Hilbert
space dimensions grow extremely fast. For N qubits (two-level systems), one
has to operate in a Hilbert space of 2N dimensions. The corresponding space
of density matrices has 22N dimensions. For classical bits one has instead a
configuration space of 2N discrete points, and the analogue of the density ma-
trices, the probability densities, live in a merely 2N -dimensional space. Brute
force simulations of the whole system therefore tend to grind to a halt even on
fairly small systems. Feynman was the first to turn this around: maybe only
a quantum system can be used to simulate a quantum system, and maybe,
while we are at it, we can go beyond simulation and do some interesting
computations as well. So, putting it positively, in a quantum system we have
exponentially more dimensions to work with: there is lots of room in Hilbert
space. The added complexity of quantum versus classical correlations, i.e. the
phenomenon of entanglement, is also a consequence of this.
But it is not so easy to use those extra dimensions. For example, for
transmission of classical information an N -qubit system is no better than
a classical N -bit system. Only the entanglement assistance of superdense
coding brings out the additional dimensions. Similarly, quantum computers
do not speed up every computation, but are good only at specific tasks where
the extra dimensions can be brought into play.
2.4.6 Error Correction
Again, we shall only make a few remarks related to the possible/impossible

theme, and refer the reader to Chap. 4 for a deeper discussion. First of all,
error correction is absolutely crucial for the implementation of quantum com-

puters. Very early in the development of the subject the suspicion was raised
that exponential speedup was only possible if all component parts of the com-
puter were realized with exponentially high (and hence practically unattain-
able) precision.
In a classical computer the solution to this problem is digitization: every
bit is realized by a bistable circuit, and any deviation from the two wanted
states is restored by the circuit at the expense of some energy and with some
heat generation. This works separately for every bit, so in a sense every bit
has its own heat bath. But this strategy will not work for quantum comput-
ers: to begin with, there is now a continuum of pure states which would have
to be stabilized for every qubit, and, secondly, one heat bath per qubit would
quickly destroy entanglement and hence make the quantum computation im-
possible. There are many indications that entanglement is indeed more easily
destroyed by thermal noise and other sources of errors; this is summarily re-
ferred to as decoherence. For example, a Gaussian channel (this is a special
type of infinite-dimensional channel) has infinite capacity for classical infor-
mation, no matter how much noise we add. But its quantum capacity drops
to zero if we add more classical noise than that specified by the Heisenberg
uncertainty relations [81].
A standard technique for stabilizing classical information is redundancy:
just send a classical bit three times, and decide at the end by majority vote
which bit to take. It is easy to see that this reduces the probability of error
from order ε to order ε2 . But quantum mechanically, this procedure is for-
bidden by the no-cloning theorem: we simply cannot make three copies to
start the process.
Fortunately, quantum error correction is possible in spite of all these
doubts [82]. Like classical error correction, it also works by distributing the
quantum information over several parallel channels, but it does this in a much
more subtle way than copying. Using five parallel channels, one can obtain a
similar reduction of errors from order ε to order ε2 [63]. Much more has been
done, but many open questions remain, for which I refer once again to Chap.
4.
2.5 A Preview of the Quantum Theory of Information

Before we go on in the next section to turn some of the heuristic descriptions
of the previous sections into rigorous mathematical statements, I shall try
to give a flavor of the theory to be constructed, and of its motivations and
current state of development.
Theoretical physics contributes to the field of quantum information pro-
cessing in two distinct, though interrelated ways. One of these ways is the
construction of theoretical models of the systems which are being set up ex-
perimentally as candidates for quantum devices. Of course, any such system
will have very many degrees of freedom, of which only very few are singled
out as the “qubits” on which the quantum computation is performed. Hence
it is necessary to analyze to what degree and on what timescales it is justi-
fied to treat the qubit degrees of freedom separately, and with what errors
the desired quantum operation can be realized in the given system. These
questions are crucial for the realization of any quantum devices, and require
specialized in-depth knowledge of the appropriate theory, e.g. quantum op-
tics, solid-state theory or quantum chemistry (in the case of NMR quantum
computing). However, these problems are not what we want to look at in this
chapter. The other way in which theoretical physics contributes to the field of
quantum information processing is in the form of another kind of theoretical
work, which could be called the “abstract quantum theory of information”.
Recall the arguments in Sect. 2.2, where the possibility of translating be-
tween different carriers of (classical) information was taken as the justifica-
tion for looking at an abstracted version, the classical theory of information,
as founded by Shannon. While it is true that quantum information cannot
be translated into this framework, and is hence a new kind of information,
translation is often possible (at least in principle) between different carriers
of quantum information. Therefore, we can make a similar abstraction in the
quantum case. To this abstract theory all qubits are the same, whether they
are realized as polarizations of photons, nuclear spins, excited states of ions in
a trap, modes of a cavity electromagnetic field or whatever other realization
may be feasible. A large amount of work is currently being devoted to this
abstract branch of quantum information theory, so I shall list some of the
reasons for this effort.
• Abstract quantum theoretical reasoning is how it all started. In the early

papers of Feynman and Deutsch, and in the papers by Bennett and co-
workers, it is the structure of quantum theory itself which opens up all
these new possibilities. No hint from experiment and no particular the-
oretical difficulty in the description of concrete systems prompted this
development. Since the technical realizations are lagging behind so much,
the field will probably remain “theory driven” for some time to come.
• If we want to transfer ideas from the classical theory of information to
the quantum theory, we shall always get abstract statements. This works
quite well for importing good questions. Unfortunately, however, the an-
swers are most of the time not transferred so easily.
• The reason for this difficulty with importing classical results is that some
of the standard probabilistic techniques, such as conditioning, do not
work in quantum theory, or work only sporadically. This is the same
problem that the statistical mechanics of quantum many-particle sys-
tems faces in comparison with its classical sister. The cure can only be
the development of new, genuinely quantum techniques. Preferably these
should work in the widest (and hence most abstract) possible setting.
• One of the fascinating aspects of quantum information is that features

of quantum mechanics which were formerly seen only as paradoxical or
counterintuitive are now turned into an asset: these are precisely the
features one is trying to utilize now. But this means that naive intu-
itive reasoning tends to lead to wrong results. Until we know much more
about quantum information, we shall need rigorous guidance from a solid
conceptual and mathematical foundation of the theory.
• When we take as a selling point for, say, quantum cryptography that
secrets are protected “with the security of natural law”, the argument is
only as convincing as the proof that reduces this claim to first principles.
Clearly this requires abstract reasoning, because it must be independent
of the physical implementation of the device the eavesdropper uses. The
argument must also be completely rigorous in the mathematical sense.
• Because it does not care about the physical realization of its “qubits”,
the abstract quantum theory of information is applicable to a wide range
of seemingly very different systems. Consider, for example, some abstract
quantum gate like the “controlled not” (C-NOT). From the abstract the-
ory, we can hope to obtain relevant quality criteria, such as the minimal
fidelity with which this device has to be implemented for some algorithm
to work. So systems of quite different types can be checked according to
the same set of criteria, and a direct competition becomes possible (and
interesting) between different branches of experimental physics.
So what will be the basic concepts and features of the emerging quantum
theory of information? The information-theoretical perspective typically gen-
erates questions like
How can a given task of quantum information processing be performed
optimally with the given resources?
We have already seen a few typical tasks of quantum information process-
ing in the previous section and, of course, there are more. Typical resources
required for cryptography, quantum teleportation and dense coding are en-
tangled states, quantum channels and classical channels. In error correction
and computing tasks, the resources are the size of the quantum memory
and the number of quantum operations. Hence all these notions take on a
quantitative meaning.
For example, in entanglement-assisted teleportation the entangled pairs
are used up (one maximally entangled qubit pair is needed for every qubit
teleported). If we try to run this process with less than maximally entangled
states, we may still ask how many pairs from a given preparation device are
needed per qubit to teleport a message of many qubits, say, with an error less
than ε. This quantity is clearly a measure of entanglement. But other tasks
may lead to different quantitative measures of entanglement. Very often it is
possible to find inequalities between different measures of entanglement, and
establishing these inequalities is again a task of quantum information theory.
The direct definition of an entanglement measure based on teleportation,

of the quantum information capacity of a channel and of many similar quan-
tities requires an optimization with respect to all codings and decodings of
asymptotically long quantum messages, which is extremely hard to evaluate.
In the classical case, however, there is a simple formula for the capacity of a
noisy channel, called Shannon’s coding theorem, which allows us to compute
the capacity directly from the transition probabilities of a channel. Finding
quantum analogues of the coding theorem (and similar formulas for entan-
glement resources) is still one of the great challenges in quantum information
theory.
2.6 Elements of Quantum Information Theory
It is probably too early to write a definitive account of quantum information

theory – there are simply too many open questions. But the basic concepts
are clear enough, and it will be the task of the remainder of this chapter to
explain them, and use the precise definitions to state some of the interesting
open questions in the field. In the limited space available this cannot be done
in textbook style, with many examples and full proofs of all the things used
on the way (or even full references of them). So I shall try to emphasize the
main lines and to set up the basic definitions using as few primitive concepts
as possible. For example, the capacities of a channel for either classical or
quantum information will be defined on exactly the same pattern. This will
make it easier to establish the relations between these concepts.
The following sections begin with material which every physicist knows
from quantum mechanics courses, although maybe not in this form. We need
to go over this material, though, in order to establish the notation.
2.6.1 Systems and States
The systems occurring in the theory can be either quantum or classical, or

can be hybrids composed of a classical and a quantum part. Therefore, we
need a mathematical framework covering all these cases. A good choice is to
characterize each type of system by its algebra of observables. In this chapter
all algebras of observables will be taken to be finite-dimensional for simplicity.
Extensions to infinite dimensionality are mostly straightforward, though, and
in fact a strength of the algebraic approach to quantum theory is that it deals
not just with infinite-dimensional algebras, but also with systems of infinitely
many degrees of freedom, as in quantum field theory [83, 84] and statistical
mechanics [85].
The first main type of system consists of purely classical systems, whose
algebra of observables is commutative, and can hence be considered as a space
of complex-valued functions on a set X. Our assumption of finiteness requires
that X is a finite set, and the algebra of observables A will be C(X), the space
of all functions f : X → C. A single classical bit corresponds to the choice
X = {0, 1}. On the other hand, a purely quantum system is determined by the
choice A = B(H), the algebra of all bounded linear operators on the Hilbert
space H. The finiteness assumption requires that H has a finite dimension
d, so A is just the space Md of complex d × d matrices. A qubit is given by
A = M2 .
The basic statistical interpretation of the algebra of observables is the
same in the quantum and classical cases, and hinges on the cone of positive
elements in the algebra. Here Y is called positive (in symbols, Y ≥ 0) if it
can be written in the form Y = X ∗ X. Then Y ∈ Md is positive exactly
if it is given by a positive semidefinite matrix, and f ∈ C(X) is positive iff
f (x) ≥ 0 for all x. In any algebra of observables A, we shall denote by 1I ∈ A
the identity element.
A state Φ on A is a positive normalized linear functional on A. That is,
Φ : A → C is linear, with Φ(X ∗ X) ≥ 0 and Φ(1I) = 1. Each state describes
a way of preparing systems, in all the details that are relevant to subsequent
statistical measurements on the systems. The measurements are described by
assigning to each outcome from a device an effect F ∈ A, i.e. an element with
0 ≤ F ≤ 1I. The prediction of the theory for the probability of that outcome,
measured on systems prepared according to the state ρ, is then ρ(F ).
For explicit computations we shall often need to expand states and ele-
ments of A in a basis. The standard basis in C(X) consists of the functions
ex , x ∈ X, such that ex (y) = 1 for x = y and zero otherwise. Similarly,
if φµ ∈ H is an orthonormal basis of the Hilbert space of a quantum sys-
tem, we denote by eµν = |eµ eν | ∈ B(H) the corresponding “matrix units”.
Then a state p on the classical algebra C(X) is characterized by the numbers
px ≡ p(ex ), which form a probability distribution on X, i.e. p(x) ≥ 0 and

x p(x) = 1. Similarly, a quantum state ρ on B(H) is given by the numbers
ρµν ≡ ρ(eνµ ), which form the so-called density matrix. If we interpret them
as the expansion coefficients of an operator ρ = µν ρµν eµν , the density
operator of ρ, we can also write ρ(A) = tr( ρA).
A state is called pure if it is extremal in the convex set of all states, i.e.
if it cannot be written as a convex combination λρ + (1 − λ)ρ of other
states. These are the states which contain as little randomness as possible.
In the classical case, the only pure states are those concentrated on a single
point z ∈ X, i.e. pz = 1, or p(f ) = f (z). The pure states in the quantum
case are determined by “wave vectors” ψ ∈ H such that ρ(A) = ψ, Aψ,
and ρ = |ψψ|. Thus, in the simplest case of a classical bit, there are just
two extreme points, whereas in the case of a qubit the extreme points form
a sphere in three dimensions and are given by the expectations of the three
Fig. 2.7. State spaces as convex sets: left, one classical bit; right, one quantum bit
(qubit)
Pauli matrices:

1 1 + x3 x1 − ix2 1
ρ = = (1I + σ · x) ,
2 x1 + ix2 1 − x3 2
xk = ρ(σk ) . (2.5)
Then positivity requires |x| ≤ 1, with equality when ρ is pure. This is shown
2
in Fig. 2.7.
Thus, in addition to the north pole |1 and the south pole |0, which
roughly correspond to the extremal states of the classical bit, we have their
coherent superpositions corresponding to the wave vectors α|1 + β|0, where
α, β ∈ C, and |α|2 + |β|2 = 1. This additional freedom becomes even more
dramatic in higher-dimensional systems, and is crucial for the possibility of
entanglement.
Entanglement is a property of states of composite systems, so we must
introduce the notion of composition of systems. We shall define this in a
way which applies to classical and quantum systems alike. If A and B are
the algebras of observables of the subsystems, the algebra of observables of
the composition is defined to be the tensor product A ⊗ B. In the finite-
dimensional case, which is our main concern, this is defined as the space of
linear combinations of elements that can be written as A ⊗ B with A ∈ A
and B ∈ B, such that A ⊗ B is linear in A and linear in B. The algebraic
operations are defined by (A ⊗ B)∗ = A∗ ⊗ B ∗ , and (A1 ⊗ B1 )(A2 ⊗ B2 ) =
(A1 A2 ) ⊗ (B1 B2 ). Thus 1I = 1IA ⊗ 1IB . Since positivity is defined in terms of a
star operation (adjoint) and a product, these definitions also determine the
states and effects of the composite system.
Let us explore how this unifies the more common definitions in the clas-
sical and quantum cases. For two classical factors C(X) ⊗ C(Y ), a basis is
formed by the elements ex ⊗ ey , so the general element can be expanded as

f= f (x, y)ex ⊗ ey ,
x,y
and each element can be identified with a function on the Cartesian product
X × Y . Hence C(X) ⊗ C(Y ) ∼ = C(X × Y ). Similarly, in the purely quantum
case, we can expand in matrix units and obtain quantities with four indices:
(A ⊗ B)µν,µ ν = Aµµ Bνν . In a basis-free way, i.e. when A, B are considered
as operators on Hilbert spaces HA , HB , this is defined by the equation
(A ⊗ B)(φ ⊗ ψ) = (Aφ) ⊗ (Bψ) ,
where φ ∈ HA and ψ ∈ HB , and the tensor product of the Hilbert spaces is
formed in the usual way. Hence B(HA ) ⊗ B(HB ) ∼ = B(HA ⊗ HB ).
But the definition of a composition by a tensor product of algebras of ob-
servables also determines how a quantum–classical hybrid must be described.
Such systems occur frequently in quantum information theory, whenever a
combination of classical and quantum information is given. We shall approach
hybrids in two equivalent ways, which are also useful more generally. Suppose
we know only that the first subsystem is classical and make no assumptions
about the nature of the second, i.e. we want to characterize tensor prod-
ucts of the form C(X) ⊗ B. Then every element can be expanded in the
form B = x ex ⊗ Bx , where now Bx ∈ B. Clearly, the elements Bx de-
termine B, and hence we can identify the tensor product with the space
(sometimes denoted by C(X; B)) of B-valued functions on X with pointwise
algebraic operations. Similarly, suppose that we know only that B = Md is
the algebra
of d × d matrices. Then, expanding in matrix units, we find that
A = µν Aµν ⊗ eµν with Aµν ∈ A. That is, we can identify A ⊗ Md with
the space (sometimes denoted by Md (A)) of d × d matrices with entries from
A. By using the relation eµν eαβ = δνα eµβ , one can readily verify that the
product in A ⊗ Md indeed corresponds to the usual matrix multiplication
in Md (A), with due care given to the order of factors in products with ele-
ments from A, if A happens to be noncommutative. The adjoint is given by
(A∗ )µν = (Aνµ )∗ . Hence a hybrid algebra C(X) ⊗ Md can be viewed either
as the algebra of C(X)-valued d × d matrices or as the space of Md -valued
functions on X.
The physical interpretation of a composite system A⊗B in terms of states
and effects is straightforward. When F ∈ A and G ∈ B are effects, so is F ⊗G,
and this is interpreted as the joint measurement of F on the first subsystem
and of G on the second subsystem, where the “yes” outcome is taken as “both
effects give yes”. In particular, F ⊗1IB corresponds to measuring F on the first
system, completely ignoring the second. Thus, for any state ρ on A ⊗ B we
define the restriction ρA of ρ to A by ρA (A) = ρ(A⊗1IB ). In the classical case,
the probability density for ρA is obtained by integrating out the B variables.
In the quantum case, it corresponds to the partial trace of density matrices
with respect to HB . In general, it is not possible to reconstruct the state ρ
from the restrictions ρA and ρB , which is another way of saying that ρ also
describes correlations between the systems. However, given ρA and ρB , there
is always a state with these restrictions, namely the tensor product ρA ⊗ ρB ,
which corresponds to an independent preparation of the subsystems.
A fundamental difference between quantum and classical correlations lies

in the nature of pure states of composite systems. Classically the situation is
easy: a pure state of the composite system C(X) ⊗ C(Y ) ∼ = C(X × Y ) is just
a point (x, y) ∈ X × Y . Obviously, the restrictions of this state are the pure
states concentrated on x and y, respectively. More generally, whenever one of
the algebras in A ⊗ B is commutative, every pure state will restrict to pure
states on the subsystems. Not so in the purely quantum case. Here the pure
states are given by unit vectors Φ in the tensor product HA ⊗ HB , and unless
Φ happens to be of the special form φA ⊗ φB (and not a linear combination
of such vectors), the state will not be a product, and the restrictions will not
be pure. The following standard form of vectors in a tensor product, known
as the Schmidt decomposition, is used in entanglement theory every day, and
twice on Sundays.
Lemma 2.1. (1) (“Schmidt decomposition”) Let Φ ∈ HA ⊗ HB be a unit

vector, and let ρA denote
the density operator of its restriction to the first
factor. Then if ρA = µ λµ |eµ eµ | (with λµ > 0) is the spectral resolution,
we can find an orthonormal system eµ ∈ HB such that

Φ= λµ eµ ⊗ eµ .
µ
(2) (“Purification”) An arbitrary quantum state ρ on H can be extended to

a pure state on a larger system with a Hilbert space H ⊗ HB . Moreover, the
restricted density matrix ρB can be chosen to have no zero eigenvalues, and
with this additional condition the space HB and the extended pure state are
unique up to a unitary transformation.

Proof. (1) We may expand Φ as Φ = µ eµ ⊗ ψµ , with suitable vectors
Ψµ ∈ HB . The reduced density matrix is determined by

ρA F ) = Φ, (A ⊗ 1I)Φ =
tr( eµ , Aeν ψµ , ψν = λµ eµ , Aeµ .
µν µ
Since A is arbitrary (e.g. A = |eα eβ |), we may compare coefficients, and
obtain ψµ , ψν = λµ δµν . Hence eµ = λ−1/2 ψµ is the desired orthonormal
system.
(2) The existence of the purification is evident if one defines Φ as above,
with
the orthonormal system eµ chosen in an arbitrary way. Then ρB =

µ λµ |eµ eµ |, and the above computation shows that choosing the basis eµ
is the only freedom in this construction. But any two bases are linked by a
unitary transformation.
A nonproduct pure state is a basic example of an entangled state in the

sense of the following definition:
Definition 2.1. A state ρ on A ⊗ B is called separable (or “classically cor-

related”) if it can be written as

ρ= λµ ρA B
µ ⊗ ρµ ,
µ
with states ρA B
µ and ρµ on A and B, respectively, and weights λµ > 0. Other-
wise, ρ is called entangled.
Thus a classically correlated state may well contain nontrivial correlations.

In fact, if either A or B is classical, every state is classically correlated. What
the definition expresses is only that we may generate these correlations by a
purely classical mechanism. We can use a classical random generator, which
produces the result “µ” with probability λµ , together with two preparing
devices operating independently but receiving instructions from the random
generator: ρA µ is the state produced by the A device if it receives the input
“µ” from the random generator, and similarly for B. Then the overall state
prepared by this setup is ρ, and clearly the source of all correlations in this
state lies in the classical random generator.
For an extensive treatment of these concepts the reader is now referred
to Chap. 5. We shall turn instead to the second fundamental type of objects
in quantum information theory, the channels.
2.6.2 Channels
Any processing step of quantum information is represented by a “channel”.

This covers a great variety of operations, from preparation to time evolution,
measurement, and measurement with general state changes. Both the input
and the output of a channel may be an arbitrary combination of classical
and quantum information. The combination of different kinds of inputs or
outputs causes no special problems of formulation: it simply means that the
algebras of observables of the input and output systems of a channel must
be chosen as suitable tensor products.
The basic idea of the mathematical description of a channel is to charac-
terize the transformation T in terms of the way it modifies subsequent mea-
surements. Suppose the channel converts systems with algebra A into systems
with algebra B. Then, by applying first the channel, and then a yes/no mea-
surement F on the B-type output system, we have effectively measured an
effect on the A-type system, which will be denoted by T (F ). Hence a channel
is completely specified by a map T : B → A, and we shall say, for simplicity,
that this map is the channel. There is, of course an alternative way of viewing
a channel, namely as a map taking input states to output states, i.e. states
on A into states on B, which we we shall denote by T∗ . We shall say that
T describes the channel in the Heisenberg picture, whereas T∗ describes the
same channel in the Schrödinger picture. The descriptions are connected by

the equation
[T∗ (ρ)] (F ) = ρ[T (F )] (2.6)
where ρ is an arbitrary state on A, and F ∈ B is also arbitrary. The notation

on the left-hand side is sometimes a little clumsy; therefore we shall often
write T∗ (ρ) = ρ ◦ T , where “◦” denotes a composition of maps, in this case
from B to A to C. A composition of channels will then also be written in the
form S ◦T . This has the advantage that things are written from left to right in
the order in which they happen: first the preparation, then some channels, and
finally the yes/no measurement F . As a further simplification, we shall often
follow the convention of dropping the parentheses of the arguments of linear
operators (e.g. T (A) ≡ T A) and dropping the ◦ symbols, but reintroducing
any of these elements for punctuation whenever they help to make expressions
unambiguous or just more readable.
For many questions in quantum information theory it is crucial to have a
precise notion of the set of possible channels between two types of systems:
clearly, the distinction between “possible” and “impossible” in Sect. 2.3 is of
this kind, but the search for the “optimal device” performing a certain task
is also of this kind. There are two different approaches for defining the set of
maps T : B → A which should qualify as channels, and luckily they agree. The
first approach is axiomatic: one just lists the properties of T which are forced
on us by the statistical interpretation of the theory. The second approach is
constructive: one lists operations which can actually be performed according
to the conventional wisdom of quantum mechanics and defines the admissible
channels as those which can be assembled from these building blocks. The
equivalence between these approaches is one of the fundamental theorems in
this field, and is known as the Stinespring dilation theorem. We shall state
this theorem after describing both approaches and giving a formal definition
of “channels”.
Note that the left-hand side of (2.6) is linear in F , which reflects the
fact that a mixture of effects (“use effect F1 in 42% of the cases and F2
in the remaining cases”) directly becomes a mixture of the corresponding
probabilities. Therefore, the right-hand side also has to be linear in F , i.e.
T : B → A must be a linear operator, from the statistical interpretation of the
theory. Obviously, T also has to take positive operators F into a positive T (F )
(“T is positive”), and the trivial measurement has to remain trivial: T 1IB =
1IA (“T is unit preserving, or unital”). This is equivalent to T∗ being likewise
a positive linear operator, with the normalization condition tr T∗ (ρ) = tr ρ.
Finally, we would like to have an operation of “running two channels in
parallel”, i.e. we would like to define T ⊗ S : A1 ⊗ B1 → A2 ⊗ B2 for arbitrary
channels T : A1 → A2 and S : B1 → B2 . Since the identity In on an n-
level quantum system Mn is one of the channels we want to consider, we
must demand that T ⊗ In also takes positive elements to positive elements.
This “complete positivity” of T is a nontrivial requirement for maps between

quantum systems. If A or B is classical, any positive linear map from A to B
is automatically completely positive. For arbitrary completely positive maps,
the product T ⊗S is defined and is again completely positive, so just requiring
the ability to form a tensor product with the “innocent bystander” In suffices
to make all parallel channels well defined.
Definition 2.2. A channel converting systems with an algebra of observables

A to systems with an algebra of observables B is a completely positive, unit-
preserving linear operator T : B → A.
In the “constructive” approach one allows only maps which can be built
from the basic operations of (1) tensoring with a second system in a speci-
fied state, (2) unitary transformation and (3) reduction to a subsystem. Let
us describe these and some other basic channels more formally, if only to
show the richness of this concept. We leave the verification of the channel
properties, including complete positivity, to the reader.
• Expansion. This expands system A by system B in the state ρ , say. Thus

T∗ (ρ) = ρ ⊗ ρ , or, from (2.6), T : A ⊗ B → A with T (A ⊗ B) = ρ (B)A.
• Restriction. In the Heisenberg picture, the operation of discarding system

B from the composite system A⊗B is T : A → A⊗B, with T (A) = A⊗1IB .
As noted before, this corresponds to taking partial traces if B is quantum,
and to an integration over Y if B = C(Y ) is classical.
• Symmetry. By definition, the symmetries of a quantum system with

an algebra of observables A are the invertible channels, i.e. channels
T : A → A such that there is a channel S with ST = T S = IA . It turns
out that these are precisely the automorphisms of A, i.e. invertible linear
maps T : A → A such that T (AB) = T (A)T (B), and T (A∗ ) = T (A)∗ .
For a pure quantum system the symmetries are precisely the unitarily
implemented maps, i.e. the maps of the form T (A) = U AU ∗ , where U is
a unitary element of A. To readers familiar with Wigner’s theorem (e.g.
corollary 3.3 in [86]), another class of maps is conspicuously absent here,
namely positive maps of the form T (A) = ΘA∗ Θ∗ where Θ is antiunitary.
It is well known that owing to the positivity of energy, a time-reversal
symmetry can be implemented only by such an antiunitary transforma-
tion. But since such symmetries are not completely positive, they can only
be global symmetries, and can never occur as symmetries affecting only
a subsystem of the world.
• Observable. A measurement is simply a channel with a classical output

algebra, say B = C(X). Obviously, T : B → A is uniquely determined by
the collection of operators Fx = T (ex ) via T f = x f (x)Fx . The channel
property of T is equivalent to

Fx ∈ A , Fx ≥ 0 , Fx = 1IA .
x
Either the “resolution of the identity” {Fx } or the channel T will be

called an observable here. This differs in two ways from the usual text-
book definitions of this term: firstly, the outputs x ∈ X need not be
real numbers, and secondly, the operators Fx , whose expectations are the
probabilities for obtaining the output x, need not be projection opera-
tors. This is sometimes expressed by calling T a generalized observable or
a POVM (positive operator-valued measure). This is to distinguish them
from the old-style “nongeneralized” observables, which are called PVMs
(projection-valued measures) because Fx2 = Fx .
• Separable Channel. A classical teleportation scheme is a composition of

an observable and a preparation depending on a classical input, i.e. it is
of the form

T (A) = Fx ⊗ ρx (A) , (2.7)
x
where the Fx form an observable, and ρx is the reconstructed state when

the measurement result is x. Equivalently, we can say that T = RS,
where ‘input of S’ = ‘output of R’ is a classical system with an algebra
of observables C(X). The impossibility of classical teleportation, in this
language, is the statement that no separable channel can be equal to the
identity.
• Instrument. An observable describes only the statistics of the measuring

results, and contains no information about the state of the system after
the measurement. If we want such a more detailed description, we have to
count the quantum system after the measurement as one of the outputs.
Thus we obtain a composite output algebra C(X) ⊗ B, where X is the
set of classical outcomes of the measurement and B describes the output
systems, which can be of a different type in general from the input systems
with an algebra of observables A. The term “instrument” for such devices
was coined by Davies [86]. As in the case of observables, it is convenient
to expand in the basis {ex } of the classical algebra. Thus T : C(X) ⊗ B →
A can be considered
as a collection of maps Tx : B → A, such that
T (f ⊗ B) = x f (x)Tx (B). The conditions on Tx are

Tx : B → A completely positive, and Tx (1I) = 1I .
x
Note that an instrument has two kinds of “marginals”: we can ignore

the B output, which leads to the observable Fx = Tx (1IB ), or we can
ignore
the measurement results, which gives the overall state change T̄ =
x T x : B → A.
• Von Neumann Measurement. A von Neumann measurement is a special

case of an instrument, associated with
a family of orthogonal projections,
i.e. px ∈ A with p∗x py = δxy px and x px = 1I. These define an instrument
T : C(X) ⊗ A → A via Tx (A) = px Apx . What von Neumann actually
proposed [87] was the version of this with one-dimensional projections
px , so the general case is sometimes called an incomplete von Neumann
measurement or a Lüders measurement. The characteristic property of
such measurements is their repeatability: since Tx Ty = 0 for x = y, re-
peating the measurement a second time (or any number of times) will
always give the same output. For this reason the “projection postulate”,
which demanded that any decent measurement should be of this form,
dominated the theory of quantum measurement processes for a long time.
• Classical Input. Classical information may occur in the input of a device

family of maps Tx :
just as well as in the output. Again this leads to a
B → A such that T : B → C(X) ⊗ A, with T (B) = x ex ⊗ Tx (B). The
conditions on {Tx} are
Tx : B → A completely positive, and Tx (1I) = 1I .
Note that this looks very similar to the conditions for an instrument, but
the normalization is different. An interesting special case is a “prepara-
tor”, for which A = C is trivial. This prepares B states that depend in
an arbitrary way on the classical input x.
• Kraus Form. Consider quantum systems with Hilbert spaces HA and

HB , and let K : HA → HB be a bounded operator. Then the map
TK (B) = K ∗ BK from B(HB ) to B(HA ) is positive. Moreover, TK ⊗ In
can be written in the same form, with K replaced by K ⊗ 1In . Hence TK
is completely positive. It follows that maps of the form

T (B) = Kx∗ BKx , where Kx∗ Kx = 1I , (2.8)
x x
are channels. It is be a consequence of the Stinespring theorem that any

channel B(HB ) to B(HA ) can be written in this form, which we call the
Kraus form, following current usage. This refers to the book [88], which
is a still recommended early account of the notion of complete positivity
in physics.
• Ancilla Form. As stated above, every channel, defined abstractly as a

completely positive normalized map, can be constructed in terms of sim-
pler ones. A frequently used decomposition is shown in Fig. 2.8. The
Unitary
=
A
Fig. 2.8. Representation of an arbitrary channel as a unitary transformation on a

system extended by an ancilla
input system is coupled to an auxiliary system A, conventionally called

the “ancilla” (“maidservant”). Then a unitary transformation is carried
out, e.g. by letting the system evolve according to a tailor-made interac-
tion Hamiltonian, and finally the ancilla (or, more generally, a suitable
subsystem) is discarded.
The claim that every channel can be represented in the last two forms is
a direct consequence of the fundamental structural theorem for completely
positive maps, due to Stinespring [89]. We state it here in a version adapted
to pure quantum systems, containing no classical components.
Theorem 2.1. (Stinespring Theorem). Let T : Mn → Mm be a completely
positive linear map. Then there is a number + and an operator V : Cm →
Cn ⊗ C such that
T (X) = V ∗ (X ⊗ 1I )V , (2.9)
and the vectors of the form (X ⊗ 1I )V φ, where X ∈ Mn and φ ∈ Cm , span

Cn ⊗ C . This decomposition is unique up to a unitary transformation of C .
The ancilla form of a channel T is obtained by tensoring the Hilbert spaces
Cm and Cn ⊗ C with suitable tensor factors Ca and Cb , so that ma = n+b.
One picks pure states in ψa ∈ Ca and ψb ∈ Cb and looks for a unitary
extension of the map V
φ ⊗ ψa = (V φ) ⊗ ψb . There are many ways to do this,
and this is a weakness of the ancilla approach in practical computations: one
is always forced to specify an initial state ψa of the ancilla and many matrix
elements of the unitary interaction, which in the end drop out of all results.
As the uniqueness clause in the Stinespring theorem shows, it is the isometry
V which neatly captures the relevant part of the ancilla picture.
In order to obtain the Kraus form of a general positive map T from its
Stinespring representation, we choose vectors φx ∈ C such that

|χx χx | = 1I , (2.10)
x
and define Kraus operators Kx for T by φ, Kx ψ = φ ⊗ χx , V ψ (we leave

the straightforward verification of (2.8) to the reader). Of course, we can take
the χx as an orthonormal basis of C , but overcomplete systems of vectors
do just as well.
It turns out that all Kraus decompositions of a given completely posi-
tive operator are obtained in the way just described. This follows from the
following theorem, which solves the more general problem of finding all de-
compositions of a given completely positive operator into completely positive
summands. In terms of channels, this problem has the following interpreta-
tion: for an instrument {Tx }, the sum T̄ = x Tx describes the overall state
change, when the measurement results are ignored. So the reverse problem
is to find all measurements which are consistent with a given overall state
change (perturbation) of the system, or, in physical terms, all delayed-choice
measurements consistent with a given interaction between the system and its
environment. By analogy with results for states on abelian algebras (proba-
bility measures) and states on C* algebras, we call the following theorem the
Radon–Nikodym theorem. For a proof see [90].
Theorem 2.2. (Radon–Nikodym Theorem). Let Tx : Mn → Mm , x ∈ X

m n
maps, and let V : C → C ⊗ C be
be a family of completely positive
the Stinespring operator of T̄ = x Tx .Then there are uniquely determined
positive operators Fx ∈ M with x Fx = 1I such that
Tx (X) = V ∗ (X ⊗ Fx )V.
A simple but important special case is the case + = 1: then, since Cn ⊗

C ≡ Cn , we can just omit the tensor factor C . The Stinespring form is
then exactly that of a single term in the Kraus form with Kraus operator
K = V . The Radon–Nikodym part of the theorem then says that the only
decompositions of T̄ into completely positive summands are decompositions
into positive multiples of T̄ . Such maps are called “pure”. Since the identity
and, more generally, symmetries are of this type, we obtain the following
corollary:
Corollary 2.1. (“No information without perturbation”). Let T : C(X) ⊗
Mn → Mn be an instrument with a unitary global state change T̄ (A) =
T (1 ⊗ A) = U ∗ AU . Then there is a probability distribution px such that
Tx = px T̄ , and the probability ρ[Tx (1I)] ≡ px for obtaining the measurement
result x is independent of the input state ρ.
2.6.3 Duality between Channels and Bipartite States
There are many connections between the properties of states on bipartite

systems, and channels. For example, if Alice has locally created a state, and
wants to send one half to Bob, the properties of the channel available for that
transmission are crucial to the kind of distributed entangled state they can
create in this way. For example, if the channel is separable, the state will also
be separable.
Mathematically, the kind of relationship we shall describe here is very
reminiscent of the relationship between bilinear forms and linear operators:
an operator from an n-dimensional vector space to an m-dimensional vector
space is parametrized by an n × m matrix, just like a bilinear form with
arguments from an n-dimensional and an m-dimensional space. It is there-
fore hardly surprising that the matrix elements of a density operator on a
tensor product can be reorganized and reinterpreted as the matrix elements
of an operator between operator spaces. What is perhaps not so obvious,
however, is that the positivity conditions for states and for channels exactly
match up in this correspondence. This is the content of the following Lemma,
graphically represented in Fig. 2.9.
S = P
T
Fig. 2.9. The duality scheme of Lemma 2.2: an arbitrary preparation P is uniquely
represented as a preparation S of a pure state and the application of a channel T
to half of the system
Lemma 2.2. Let ρ be a density operator on H ⊗ K. Then there is a Hilbert

space H , a pure state σ on H ⊗ H and a channel T : B(K) → B(H ) such
that
ρ = σ ◦ (IH ⊗ T ) . (2.11)
Moreover, the restriction of σ to H can be chosen to be nonsingular, and in

this case the decomposition is unique in the sense that any other decomposi-
tion ρ = σ ◦ (IH ⊗ T ) is of the form σ = σ ◦ R and T = R−1 T , with a
unitarily implemented channel R.
It is clear that σ must be the purification of ρ, restricted to the first factor.

√
Thus we may set σ = |Ψ Ψ |, where Ψ = k rk ek ⊗ ek ; rk > 0 are the
nonzero eigenvalues of the restriction of ρ to the first system, and ek is a basis
of H . Note that the ek are indeed unique up to a unitary transformation, so
we only have to show that for one choice of ek we obtain a unique T . From
the equation ρ = σ ◦ (IH ⊗ T ), we can then read off the matrix elements of

T:
−1/2 −1/2
ek , T (|eµ eν |) e = rk r ρ |ek ⊗ eµ e ⊗ eν | . (2.12)
We have to show that T , as defined by this equation, is completely positive

whenever ρ is positive. For fixed coefficients rk the map ρ → T is obviously
linear. Hence it suffices to prove complete positivity for ρ = |ϕϕ|. But in
−1/2
that case T = V ∗ AV , with eν , V e = r e ⊗ eν , ϕ, so T is indeed
completely positive. The normalization T (1I) = 1I follows from the choice of
rk , and the lemma is proved.
The main use of this lemma is to translate results about entangled states
to results about channels, and conversely. For this it is necessary to have
a translation table of properties. Some entries are easy: for example, ρ is a
product state iff T is depolarizing in the sense that T (A) = tr(ρ2 A) for some
density operator ρ2 , and ρ is separable in the sense of Definition 2.1 iff T is
separable (see (2.7)).
2.6.4 Channel Capacity
In the definition of channel capacity, we shall have to use a criterion for the
approximation of one channel by another. Since channels are maps between
normed spaces, one obvious choice would be to use the standard norm

S − T := sup S(A) − T (A) | A ≤ 1 . (2.13)
However, as in the case of positivity, there is a problem with this definition

when one considers tensor products: the norms T ⊗ In , where In is the
identity on Mn , may increase with n. This introduces complications when
one has to make estimates for parallel channels. Therefore we stabilize the
norm with respect to tensoring with “innocent bystanders”, and introduce,
for any linear map T between C* algebras, the norm
T cb := sup T ⊗ In , (2.14)

n
called the norm of complete boundedness, or “cb norm” for short. This name
derives from the observation that on infinite-dimensional C* algebras the
above supremum may be infinite even though each term in the supremum
is finite. By definition, a completely bounded map is one with T cb < ∞.
On a finite-dimensional C* algebra, every linear map is completely bounded:
for maps into Md we have T cb ≤ dT . (As a general reference on these
matters, I recommend the book [91].) One might conclude from this that
the distinction between these norms is irrelevant. However, since we shall
need estimates for large tensor products, every factor that increases with
dimension can make a decisive difference. This is the reason for employing
the cb norm in the definition of channel capacity. It will turn out, however,
that in the most important cases one has only to estimate differences from
the identity, and T − I and T − Icb can be estimated in terms of each
other with dimension-independent bounds.
The basis of the notion of channel capacity is a comparison between the
given channel T : A2 → A1 and an “ideal” channel S : B1 → B2 . The
comparison is effected by suitable encoding and decoding transformations
E : A1 → B1 and D : B2 → A2 so that the composed operator ET D : B2 →
B1 is a map which can be compared directly with the ideal channel S. Of
course, we are only interested in such a comparison in the case of optimal
encoding and decoding, i.e. in the quantity
∆(S, T ) = inf S − ET Dcb , (2.15)

E,D
where the infimum is over all channels (i.e. all unit-preserving completely
positive maps) E and D with appropriate domain and range. Since these
data are at least implicitly given together with S and T , there is no need
to specify them in the notation. S should be thought of as representing one
word of the kind of message to be sent, whereas T represents one invocation
of the channel. Channel capacity is defined as the number of S words per
invocation of the channel T which can be faithfully transmitted, with suitable
encoding and decoding for long messages. Here “messages of length n” are
represented by the tensor power S ⊗n , and “m invocations of the channel T ”
are represented by the tensor power T ⊗m .
Definition 2.3. Let S and T be channels. Then a number c ≥ 0 is called

an achievable rate for T with respect to S if, for any sequences nα , mα of
integers with mα → ∞ and lim supα (nα /mα ) < c, we have

lim ∆ S ⊗nα , T ⊗mα = 0.
α
The supremum of all achievable rates is called the capacity of T with respect
to S, and is denoted by C(S, T ).
Note that by definition, 0 is an achievable rate (no integer sequences with

asymptotically negative ratio exist), and hence C(S, T ) ≥ 0. If all c ≥ 0 are
achievable, then of course we write C(S, T ) = ∞. It may be cumbersome to
check all pairs of integer sequences with a given upper ratio when testing
c. However, owing to the monotonicity of ∆, it suffices to check only one
sequence, provided it is not too sparse: if there is any pair of sequences
nα , mα satisfying the conditions in the definition (including ∆ → 0) plus the
extra requirement that (mα /mα+1 ) → 1, then c is achievable.
The ideal channel for systems with an algebra of observables A is by
definition the identity map IA on A. For typographical convenience we shall
abbreviate “IA ” to “A” whenever it appears as an argument of ∆ or C. Using
this notation, we shall now summarize the capacities of ideal quantum and
classical channels. Of course, these are basic data for the whole theory:
C(Mk , Cn ) = 0 for k ≥ 2, (2.16)
log n
C(Ck , Cn ) = C(Mk , Mn ) = C(Mk , Cn ) = . (2.17)
log k
Here the first equation is the capacity version of the no-teleportation theorem:
it is impossible to transport any quantum information on a classical channel.
The second equation shows that for capacity purposes, Mn is indeed best
compared with Cn . In classical information theory one uses the one-bit system
C2 as the ideal reference channel. Similarly, we use the one-qubit channel as
the reference standard for quantum information , i.e. we define the classical
capacity Cc (T ) and the quantum capacity Cq (T ) of an arbitrary channel by
Cc (T ) = C(C2 , T ), (2.18)
Cq (T ) = C(M2 , T ) . (2.19)
Combining the results (2.17) with the “triangle inequality”, or two-step coding
inequality,
C(T1 , T3 ) ≥ C(T1 , T2 )C(T2 , T3 ), (2.20)
we see that this is really only a choice of units, i.e. for arbitrary channels
T we obtain C(Mn , T ) = (log 2/ log n)C(M2 , T ), and a similar equation for
classical capacities. Note that the term “qubit” refers to the reference system
M2 , but it is not advisable to use “qubit” as a special unit for quantum
information (rather than just “bit”): this would be like distinguishing between
the units “vertical meter” and “horizontal meter” and would create problems
in every equation in which the two capacities were directly compared. The
simplest relation of this kind is
Cq (T ) ≤ Cc (T ) , (2.21)
which follows from combining (2.20) with (2.17). Note that both definitions
apply to arbitrary channels T , whether the input and/or output are classical
or quantum or hybrids. In order for a channel to have a positive quantum
capacity, it is necessary that both the input and the output are quantum
systems. This is shown by combining (2.16) with the bottleneck inequality

C(S, T1 T2 ) ≤ min C(S, T1 ), C(S, T2 ) . (2.22)
Another application of the bottleneck inequality is to separable channels.

These are by definition the channels with a purely classical intermediate
stage, i.e. T = SR, where “output of S” = “input of R” is a classical system.
For such channels Cq (T ) = 0.
An important operation on channels is running two channels in parallel,

represented mathematically by the tensor product. The relevant inequality is
C(S, T1 ⊗ T2 ) ≥ C(S, T1 ) + C(S, T2 ) (2.23)
for the standard ideal channels, and when all systems involved are classical,
we even have equality. However, it is one of the big unsolved problems to
decide under what general circumstances this is true.
Comparison with the Classical Definition. Since the definition of clas-
sical capacity Cc (T ) also applies to the purely classical situation, we have to
verify that it is indeed equivalent to the standard definition in this case. To
that end, we have to evaluate the error quantity T − Icb for a classical-to-
classical channel. As noted, a classical channel T : C(Y ) → C(X) is given by
a transition probability matrix T (x → y). Since the cb norm coincides with
the ordinary norm in the classical case, we obtain

I − T cb = I − T = sup δxy − T (x → y) f (y)
x,f y

= 2 sup 1 − T (x → x) ,
x
where the supremum is over all f ∈ C(Y ) with |f (y)| ≤ 1 and is attained
where f is just the sign of the parenthesis in the second line, and we have
used the normalization of the transition probabilities. Hence, apart from an
irrelevant factor of two, T − Icb is just the maximal probability of error,
i.e. the largest probability for sending x and obtaining anything different.
This is precisely the quantity which is required to go to zero (after suitable
coding and decoding) in Shannon’s classical definition of the channel capacity
of discrete memoryless channels [92]. Hence the above definition agrees with
the classical one.
When considering the classical capacity Cc (T ) of a quantum channel,
it is natural to look at a coded channel ET D as a channel in its own right.
Since we are considering transmission of classical information, this is a purely
classical channel, and we can look at its classical capacity. Optimizing over
coding and decoding, we obtain the quantity
Cc,1 (T ) = sup Cc (ET D) . (2.24)

ET D classical
This is called the one-shot classical capacity, because it can be said to involve
only one invocation of the channel T . Of course, many uses of the channel
are implicit in the capacity on the right-hand side, but these are in some
sense harmless. In fact, every coding and decoding scheme for comparing
(ET D)⊗n with an ideal classical channel is also a coding/decoding for T ⊗n ,
but the codings/decodings that arise in this way from the coding ET D are
only those in which the coded input states and the measurements at the
outputs are not entangled. If we allow entanglement over blocks of a large

length +, we thus recover the full classical capacity:
1
Cc,1 (T ) ≤ Cc (T ) = sup Cc,1 (T ⊗ ) . (2.25)
+
It is not clear whether equality holds here. This is a fundamental question,

which can be paraphrased as follows: “Does entangled coding ever help in
sending classical information over quantum channels?” At the moment, all
partial results known to the author seem to suggest that this is not the case.
Comparison with other Error Criteria. Coming now to the quantum
capacity Cq (T ), we have to relate our definition to more current definitions.
One version, first stated by Bennett, is very similar to the one given above,
but differs slightly in the error quantity, which is required to go to zero.
Rather than T − Icb, he considers the lowest fidelity of the channel, defined
as

F (T ) = inf ψ, T |ψψ| ψ , (2.26)
ψ
where the supremum is over all unit vectors. Hence the achievable rates are
those for which F (ET ⊗nα D) → 1, where E, D map to a system of mα qubits,
and these integer sequences satisfy the same constraints as above. This def-
inition is equivalent to ours, because the error estimates are equivalent. In
fact, if we introduce the off-diagonal fidelity

F% (T ) = sup !eφ, T |φψ| ψ (2.27)
φ,ψ
for any channel T : Md → Md with d < ∞, we have the following system of

estimates:

T − I ≤ T − Icb ≤ 4 1 − F% (T ) ≤ 4 T − I, (2.28)

T − I ≤ 4 1 − F(T ) ≤ 4 1 − F% (T ) , (2.29)
which will be proved elsewhere. The main point is, though, that the dimension
does not appear in these estimates, so if one such quantity goes to zero, all
others do, and we can build an equivalent definition of capacity out of any
one of them.
Yet another definition of quantum capacity has been given in terms of
entropy quantities [93], and has also been shown to be equivalent [94].
2.6.5 Coding Theorems
The definition of channel capacity looks simple enough, but computing it on

the basis of this definition is in general a very hard task: it involves an opti-
mization over all coding and decoding channels in systems of asymptotically
many tensor factors. Hence it is crucial to obtain simpler expressions which

can be computed in a much more direct way from the matrix elements of the
given channel. Such results are called coding theorems, after the first theorem
of this type, established by Shannon.
In order to state this theorem, we need some entropy quantities. The von
Neumann entropy of a state with density matrix ρ is defined as

S(ρ) = −tr ρ log ρ , (2.30)
where the function of ρ on the right-hand side is evaluated in the functional

calculus, and 0 log 0 is defined to be zero. The logarithm is chosen here as the
logarithm to base 2, so the unit of entropy is a “bit”. The relative entropy of
a state ρ with respect to another, σ, is defined by

S(ρ, σ) = tr ρ(log ρ − log σ) . (2.31)
Both quantities are positive, and may be infinite on an infinite-dimensional

space. The von Neumann entropy is concave, whereas the relative entropy
is convex jointly in both arguments. For more precise definitions and many
further results, I recommend the book by Ohya and Petz [95].
The strongest coding theorem for quantum channels known so far is the
following expression for the one-shot classical capacity, proved by Holevo [96]:

Cc,1 (T ) = max S pi T∗ [ρi ] − pi S(T∗ [ρi ]) . (2.32)
i i
Whether or not this is equal to the classical capacity depends on whether the
conjectured equality in (2.25) holds or not. In any case, equality is known
to hold for channels with classical input, so Holevo’s coding theorem is a
genuine extension of Shannon’s.
No coding theorem has been proved yet for the quantum capacity. How-
ever, there is a fairly good candidate for the right-hand side, related to a
quantity called “coherent information” [97]. The formula is written most
compactly by relating it to an entanglement quantity via Lemma 2.2. For
any bipartite state ρ with restriction ρB to the second factor, let
ES (ρ) = S(ρB ) − S(ρ) . (2.33)
This is a measure of entanglement of sorts, because it is large when S(ρ)

is small, e.g. when ρ is pure, and ρB is very mixed when, for example, ρ is
maximally entangled. It can be negative, though (see [98] for a discussion).
We set
CS,1 (T ) = sup ES [σ ◦ (I ⊗ T )] , (2.34)

σ
where the supremum is over all bipartite pure states σ. Note that any measure
of entanglement can be turned into a capacity-like expression by this proce-
dure. Since this quantity is known not to be additive [99], the candidate for
the right-hand side of the quantum coding theorem is
1
CS (T ) = sup CS,1 (T ⊗ ) , (2.35)
+
in analogy to (2.25). So far there have been some good heuristic arguments
[100, 101] in favor of this candidate, but a full proof remains one of the main
challenges in the field.
An interesting upper bound on Cq (T ) can be written in terms of the
transpose operation Θ on the output system [81]: we have
Cq (T ) ≤ log2 ΘT cb . (2.36)
Hence, if ΘT happens to be completely positive (as for any channel with an

intermediate classical state), this map is a channel; hence, it has a cb norm
of 1, and Cq (T ) = 0. This criterion can also be used to show that whenever
there is sufficiently high noise in a channel, it will have a quantum capacity
of zero.
2.6.6 Teleportation and Dense Coding Schemes
In this section we shall show that entanglement-assisted teleportation and

dense coding, as described in Sects. 2.4.3 and 2.4.4, really work.
Rather than going through the now standard derivations in the basic ex-
amples involving qubits, we shall use the structure assembled so far to reverse
the question, i.e. we try to find the most general setup in which teleporta-
tion and dense coding work without errors. This will give some additional
insights, and possibly some welcome flexibility when it comes to realizing
these processes for systems larger than a qubit. The task as stated in this
form is somewhat beyond the scope of this chapter, mainly because there are
so many ways to waste resources, which do not necessarily have a compact
characterization. So, in order to obtain a readable result, we look only at the
“tight case” [102], in which resources are used, in a sense, optimally. By this
we mean that all Hilbert spaces involved have the same finite but arbitrary
dimension d (so we can take them all equal to H = Cd ), and the classical
channel distinguishes exactly |X| = d2 signals.
For both teleportation and dense coding, the beginning of each trans-
mission is the distribution of the parts of an entangled state ω between the
sender Alice and the receiver Bob. Only then is Alice given the message she
is supposed to send, which is a quantum state in the case of teleportation
and a classical value in the case of dense coding. She codes this in a suitable
way, and Bob reconstructs the original message by evaluating Alice’s signal
jointly with his entangled subsystem.
For dense coding, assume that x ∈ X is the message given to Alice.

She encodes it by transforming her entangled system with a channel Tx and
sending the resulting quantum system to Bob, who measures an observable
F jointly on Alice’s
his. The probability for obtaining y as a
particle and
result is then tr ω(Tx ⊗ I)(Fy ) , where the “⊗I” expresses the fact that no
transformation is applied to Bob’s particle, while Alice applies Tx to hers. If
everything works correctly, this expression has to be equal to 1 for x = y,
and 0 otherwise:

tr ω(Tx ⊗ I)(Fy ) = δxy . (2.37)
Let us take a similar look at teleportation. Here three quantum systems are
involved: the entangled pair in state ω, and the input system given to Alice,
in state ρ. Thus the overall initial state is ρ ⊗ ω. Alice measures an observable
F on the first two factors, obtaining a result x, which is sent to Bob. Bob
applies a transformation Tx to his particle, and makes a final measurement
of an observable A of his choice. Thus the probability of Alice measuring x
and of Bob obtaining a result “yes” on A is tr(ρ ⊗ ω)[Fx ⊗ Tx (A)]. Note that
the tensor symbols in this equation refer to different splittings of the system
(1 ⊗ 23 and 12 ⊗ 3, respectively). Teleportation is successful if the overall
probability of obtaining A, computed by summing over all possibilities x, is
the same as for an ideal channel, i.e.

tr(ρ ⊗ ω)[Fx ⊗ Tx (A)] = tr(ρA) . (2.38)
x∈X
Surprisingly, in the tight case one obtains exactly the same conditions on
ω, Tx , Fx for teleportation and for dense coding, i.e. a dense-coding scheme
can be turned into a teleportation scheme simply by letting Bob and Al-
ice swap their equipment. However, this symmetry depends crucially on the
tightness condition, because teleportation schemes with |X| > d2 signals are
trivial to achieve, but |X| > d2 is impossible for dense coding. Conversely,
dense coding through a d > d-dimensional channel is trivial to achieve, while
teleportation of states with d > d dimensions (with the same X) is impossi-
ble.
Let us now give a heuristic sketch of the arguments leading to the neces-
sary and sufficient conditions for (2.37) and (2.38) to hold. For full proofs we
refer to [102]. A crucial ingredient in the analysis of the teleportation equation
is the “no measurement without perturbation” principle from Lemma 2.1: the
left-hand side of (2.38) is indeed such a decomposition, so each term must be
equal to λx tr(ρA) for all ρ, A. But we can carry this even further: suppose
we decompose ω, Fx or Tx into a sum of (completely) positive terms. Then
each term in the resulting sum must also be proportional to tr(ρA). Hence
any components of ω, Tx or Fx satisfy a teleportation equation as well (up
to normalization). Similarly, the vanishing of the dense-coding equation for
x = y carries over to every positive summand in ω, Tx or Fx . Hence it is
plausible that we must first analyze the case where all ω, Fx , Tx are “pure”,
i.e. have no nontrivial decompositions as sums of (completely) positive terms:
ω = |ΩΩ|, (2.39)
Fx = |Φx Φx |, (2.40)
Tx (A) = Ux∗ AUx . (2.41)
The further analysis will show that in the pure case any two of these elements
determine the third via the teleportation or the dense-coding equation, so
that in fact all components of ω (and correspondingly Tx or Fx ) have to be
proportional. Hence each of these has to be pure in the first place. For the
present discussion, let us just assume purity in the form (2.39)–(2.41) from
now on. Note that normalization requires that each Ux is unitary.

The second normalization condition, x |Φx Φx | = x Fx = 1I, has
an interesting consequence in conjunction with the tightness condition: the
vectors Φx live in a d2 -dimensional space, and there are exactly d2 of them.
This implies that they are orthogonal: since each vector Φx satisfies Φx ≤ 1,
and d = tr(1
I) = x Φx , we must have Φx = 1 for all x. Hence, in the
2
sum 1 = x Φy , Fx Φy the term y = x is equal to 1, and hence the others

must be be zero.
Now consider the term with index x in the teleportation equation and set
ρ = |φ φ| and A = |ψψ |. Then the trace splits into two scalar products,
in which the variables φ, φ , ψ, ψ can be chosen independently, which leads
to an equation of the form
φ ⊗ Ω, Φx ⊗ (Ux∗ ψ) = λx φ, ψ , (2.42)

for all φ, ψ, and to coefficients which must satisfy x |λx |2 = 1. Note how
in this equation a scalar product between the vectors in the first and third
tensor factors is generated. This type of equation, which is clearly the core
of the teleportation process, may be solved in general:
Lemma 2.3. Let H, K be finite-dimensional Hilbert spaces, and let Ω1 ∈
K ⊗ H and Ω2 ∈ H ⊗ K be unit vectors such that, for all φ, ψ ∈ H,
φ ⊗ Ω1 , Ω2 ⊗ ψ = λφ, ψ . (2.43)
Then |λ| ≤ 1/ dim H, with equality iff Ω1 and Ω2 are maximally entangled

and equal up to the exchange of the tensor factors H and K.
√
For the proof, consider the Schmidt decomposition Ω1 = k wk fk ⊗ ek ,
and insert φ = en , ψ = em into (2.43) to find the matrix elements of Ω2 :
−1/2
en ⊗ fm , Ω2 = λ wm δnm .
−1
2 = |λ|
Clearly, Ω 2 2
m wm . This sum takes its smallest value under the
constraint m wm = Ω1 2 = 1 only at the point where all wm are equal.
This proves the lemma.

We apply this lemma to Ω1 = (1I ⊗ Ux )Ω and Ω = Φx . Then x |λx |2 ≤
d2 d−2 = 1, with equality only if all the vectors involved are maximally en-
tangled and are pairwise equal up to an exchange of factors:
Φx = (Ux ⊗ 1I)Ω , (2.44)

where we take Ω = d−1/2 k ek ⊗ ek by an appropriate choice of bases. If
Ω is maximally entangled, (2.44) sets up a one-to-one correspondence be-
tween unitary operators Ux and the vectors Φx , as independent elements in
the construction. The Φx have to be an orthonormal basis of maximally en-
tangled vectors, and there are no further constraints. In terms of the Ux ,
the orthogonality of the Φx translates into orthogonality with respect to the
Hilbert–Schmidt scalar product:
tr(Ux∗ Uy ) = d δxy . (2.45)
Again, there are no further constraints, so any collection of d2 unitaries sat-
isfying these equations leads to a teleportation scheme.
For the dense-coding case we obtain the same result, although by a differ-
ent route. Equation (2.44) follows easily if we write the teleportation equation
as |Ω, (Ux∗ ⊗ 1I)Φx |2 = δxy . The problem is to show that Ω has to be max-
imally entangled. Using the reduced density operator ω1 of ω, this becomes
tr(ω1 Ux∗ Uy ) = Ω, (Ux∗ Uy ⊗ 1I)Ω = Φx , Φy = δxy . (2.46)

We claim that this equation, for a positive operator ω1 and d2 unitaries Ux ,
implies that ω1 = d−1 1I. To see this, expandthe operator A = |φek |ω1−1 in
the basis Ux according to the formula A = x Ux tr(Ux∗ Aω1 ):

ek , Ux∗ φ Ux = |φek |ω1−1 .
x
Taking the matrix element φ| · |ek of this equation and summing over k, we
find

ek , Ux∗ φ φ, Ux ek = tr(Ux∗ |φφ|Ux ) = d2 φ2 = φ2 tr(ω1−1 ) .
x,k x

Hence tr(ω1−1 ) = d2 = −1
k rk , where rk are the eigenvalues of ω1 . Us-
ing
again the fact that the smallest value of this sum under −1 the constraint
k r k = 1 is attained only for constant rk , we find ω 1 = d 1I, and Ω is
indeed maximally entangled.
To summarize, we have the following theorem (again, for a detailed proof
see [102]):
Theorem 2.3. Given either a teleportation scheme or a dense-coding scheme,
which is tight in the sense that all Hilbert spaces are d-dimensional and
|X| = d2 classical signals are distinguished, then
• ω = |ΩΩ| is pure and maximally entangled,

• Fx = |Φx Φx |, where the Φx form an orthonormal basis of maximally
entangled vectors,
• Tx (A) = Ux∗ AUx , where the Ux are unitary and orthonormal in the sense
that tr(Ux∗ Uy ) = d δxy , and
• these objects are connected by the equation Φx = (Ux ⊗ 1I)Ω.
Given either the Φx or the Ux with the appropriate orthogonality properties,
and a maximally entangled vector Ω, the above conditions determine a dense
coding and a teleportation scheme.
In particular, we have shown that a teleportation scheme becomes a dense-

coding scheme and vice versa, when Alice and Bob swap their equipment.
However, this is only true in the tight case: for a larger quantum channel,
dense coding becomes easier but teleportation becomes more demanding.
Similarly, teleportation becomes easier with more allowed classical informa-
tion exchange, whereas dense coding of more than d2 signals is impossible.
In order to construct a scheme, it is best to start from the equation
tr(Ux∗ Uy ) = d δxy , i.e. to look for orthonormal bases in the space of operators
consisting of unitaries. For d = 2 the solution is essentially unique: U1 , . . . , U4
are the identity and the three Pauli matrices, which leads to the standard
examples. Group theory helps to construct examples of such bases for any
dimension d, but this construction by no means exhausts the possibilities.
A fairly general construction is given in [102]. It requires two combinatorial
structures known from classical design theory [103, 104]: a Latin square of
order d, i.e. a matrix in which each row and column is a permutation of
(1, ..., d), and d Hadamard matrices , i.e. unitary d×d matrices, in which each
entry has modulus d−1/2 . For neither Latin squares nor Hadamard matrices
does an exhaustive construction exist, so these are rich fields for hunting
and gathering new examples, or even infinite families of examples. Certainly,
this connection suggests that a full classification or exhaustive construction
of teleportation and dense-coding schemes cannot be expected. However, it
may still be a good project to look for schemes with additional desirable
features.
3 Quantum Communication
Harald Weinfurter and Anton Zeilinger
Quantum entanglement lies at the heart of the new field of quantum com-
munication and computation. For a long time, entanglement was seen just as
one of those fancy features which make quantum mechanics so counterintu-
itive. But recently, quantum information theory has shown the tremendous
importance of quantum correlations for the formulation of new methods of
information transfer and for algorithms exploiting the capabilities of quan-
tum computers. While the latter needs entanglement between a large number
of quantum systems, the basic quantum communication schemes rely only on
entanglement between the members of a pair of particles, directly pointing
to a possible realization of such schemes by means of correlated photon pairs
such as those produced by parametric down-conversion.
This chapter describes the first experimental realizations of quantum com-
munication schemes using entangled photon pairs. We show how to make com-
munication secure against eavesdropping using entanglement-based quantum
cryptography, how to increase the information capacity of a quantum chan-
nel by quantum dense coding and, finally, how to communicate quantum
information itself in the process of quantum teleportation.
3.1 Introduction
Quantum mechanics is probably the most successful physical theory of this
century. It provides powerful tools which form one of the cornerstones of
scientific progress, and which are indispensable for the understanding of om-
nipresent technical devices such as the transistor, semiconductor chips and
the laser. The most important areas where those devices are used are mod-
ern communication and information-processing technologies. But quantum
mechanics, until now, has only been used to construct these devices – quan-
tum effects are absolutely avoided in the representation and manipulation
of information. Rather than using single photons, one still uses strong light
pulses to send information along optical high-speed connections, and one re-
lies on electrical currents in semiconductor logic chips instead of applying
single electrons as signal carriers.
This caution surely is due to the fact that, at first glance, the inherent
stochastic character of quantum effects seems only to introduce unavoidable

3. Quantum Communication 59
noise and thus does not really recommend their use. Yet quantum informa-
tion theory shows us, in more and more examples, how one can profit from
the peculiar properties of quantum systems, and, when applied correctly,
how fundamental quantum effects can add to the power and features of clas-
sical information processing and transmission [12, 105, 106]. For example,
quantum computers will outperform conventional computers, and quantum
cryptography enables, for the first time, secure communication. While quan-
tum cryptography, in principle, can be performed even with single quantum
particles, all the other proposals utilize entanglement between two or more
particles, for example to enhance communication rates or to enable the tele-
portation of quantum states.
Entanglement between quantum systems is a pure quantum effect. It is
closely related to the superposition principle and describes correlations be-
tween quantum systems that are much stronger and richer than any classical
correlation could be. Originally this property was introduced by Einstein,
Podolsky and Rosen (EPR) [24], and also by Schrödinger [5] and Bohr [107]
in the discussion of the completeness of quantum mechanics and by von Neu-
mann [108] in his description of the measurement process. Entanglement also
provides a handle to distinguish various interpretations of quantum mechan-
ics via Bell’s theorem [72, 109] or the GHZ argument [37]. The development
of experimental techniques has enabled researchers to perform the recent
long-distance tests of entanglement [30], the first Bell experiment fulfilling
Einstein locality conditions [31] and the first GHZ experiment [38], which
all provided convincing demonstrations of the validity of standard quantum
mechanics.1
The field of quantum information is not concerned with the fundamental
issues. Instead, it builds on the validity of quantum mechanics and applies the
characteristic features of entangled systems to devise new, powerful schemes
for communication and computation. Entanglement between a large number
of quantum systems will enable very efficient computations. In particular,
the factorization algorithm of Shor [9] and the search algorithm of Grover
[58] (together with the increasing number of algorithms derived from one or
the other) show how entanglement and the associated interference between
entangled states can boost the power of quantum computers.
Quantum communication exploits entanglement between only two or three
particles. As will be seen in the following sections, the often counterintu-
itive features of such small entangled systems enable powerful communication
methods. After the very basic properties of pairs of entangled particles have
been introduced (Sect. 3.2), Sect. 3.3 gives an overview of the possibilities
of three important quantum communication schemes: entanglement-based
quantum cryptography enables secret key exchange and thus truly secure
communication [46]; using quantum dense coding, one can send classical in-
1
We are aware of the detection loophole [110], which will be closed whenever
technology allows.
60 Harald Weinfurter and Anton Zeilinger
formation more efficiently [51]; and, finally, with quantum teleportation one
can transfer quantum information, that is, the quantum state itself, from
one quantum system to another [52]. The tools for the experimental real-
ization of those quantum communication schemes are presented in Sect. 3.4.
In particular, we show how to produce polarization-entangled photon pairs
by parametric down-conversion [111] and how to observe these nonclassical
states by interferometric Bell-state analysis [112]. In Sect. 3.5 we describe the
first experimental realizations of basic quantum communication schemes. In
experiments performed during recent years at the University of Innsbruck,
we could realize entanglement-based quantum cryptography with randomly
switched analyzers and with the two users separated by more than 400 m
[113]; we demonstrated the possibility of transmitting 1.58 bits of classical
information by encoding trits on a single two-state photon [114]; and we
could transfer a qubit, in our case the polarization state, from one photon to
another by quantum teleportation [10, 11] and entanglement swapping [115].
3.2 Entanglement – Basic Features

For a long time, entanglement was seen merely as one of the counterintuitive
features of quantum mechanics, important only within the realm of the EPR
paradox. Only lately has the field of quantum information exploited these
features to obtain new types of information transmission and processing.
Recent literature now offers a thorough discussion of all the various properties
of entangled systems [37, 72, 116] (see also Chap. 5); in this review, we
concentrate on those features which form the foundation of the basic quantum
communication schemes.
At the heart of entanglement lies another fundamental feature of quantum
mechanics, the superposition principle. If we look at a classical, two-valued
system, for example a coin, we find it in either one of its two possible states,
that is, either head or tail. Its quantum mechanical counterpart, a two-state
quantum system, however,√can be found in any superposition of two possible
basis states, e.g. |Ψ = (1/ 2)(|0 + |1). Here we denote the two orthogonal
basis states by |0 and |1, respectively.2 This generic notation can stand
for any of the properties of various two-state systems, for example for the
ground state |g and excited state |e of an atom, or, as is the case in our
experiments, for the horizontal polarization |H and vertical polarization |V
of a photon.
In the classical world, we find two coins to be in the states of either
head/head, head/tail, tail/head or tail/tail, and we can identify these four
possibilities with the four quantum states
|01 |02 , |01 |12 , |11 |02 and |11 |12 ,
2
This notation should not be confused with the description of an electromagnetic
field (vacuum or single-photon state) in second quantization. Here we use only
the notions of first quantization to describe the properties of two-state systems.
describing two two-state quantum systems. As the superposition principle

holds for more than one quantum system, the two quantum particles are no
longer restricted to the four “classical” basis states above, but can be in any
superposition thereof, for example in the entangled state
1
|Ψ = √ (|01 |02 + |11 |12 ) . (3.1)
2
Of course, one is restricted neither to only two particles nor to such

maximally entangled states. During the last decade, enormous progress was
achieved in the theoretical studies of the quantum features of multiparti-
cle systems. One will observe even more stunning correlations between three
or more entangled particles [37, 117]; one can generalize to the observation
of interference and entanglement between multistate particles [118] and to
entanglement for mixed states. There also exists the possibility to purify en-
tanglement [119], and one can even find two-particle systems which are not
actually entangled, but are such that a local observer cannot distinguish them
from entangled states [120]. For the basic quantum communication schemes
and experiments, we can concentrate on the particular properties of max-
imally entangled two-particle systems. Considering two two-state particles,
we find a basis of four orthogonal, maximally entangled states, the so called
Bell-states basis:
1
|Ψ + 12 = √ (|01 |12 + |11 |02 ), (3.2)
2
− 1
|Ψ 12 = √ (|01 |12 − |11 |02 ), (3.3)
2
1
|Φ+ 12 = √ (|01 |02 + |11 |12 ), (3.4)
2
− 1
|Φ 12 = √ (|01 |02 − |11 |12 ) . (3.5)
2
The name “Bell states” was given to these states since they maximally
violate a Bell inequality [121]. This inequality was deduced in the context of
so-called local realistic theories (see Chap. 1), and gives a range of possible
results for certain statistical tests on identically prepared pairs of particles
[109]. Quantum mechanics predicts different results if the measurements are
performed on entangled pairs. If the two particles are not correlated, i.e. are
described by a product state, the quantum mechanical prediction is within
the range given by Bell’s inequality.
The remarkably nonclassical features of entangled pairs arise from the
fact that the two systems can no longer be seen as being independent but
now have to be seen as one combined system, where the observation of one of
the two will change the possible predictions of measurement results obtained
for the other [5, 107]. Formally, this mutual dependence is reflected by the
fact that the entangled state can no longer be factored into a product of two
states for the two subsystems separately.
If one looks only at one of the two particles, one finds it with equal proba-
bility in state |0 or in state |1. One has no information about the particular
outcome of a measurement to be performed. However, the observation of
one of the two particles determines the result of a measurement of the other
particle. This holds not only for a measurement in the basis |0/|1, but for
any arbitrary superposition, that is, for any arbitrary orientation of the mea-
surement apparatus. In particular, for the state |Ψ − we shall find the two
particles always in orthogonal states, no matter which measurement appa-
ratus is used. If, for the case of polarization-entangled photons, we observe
only one of the two photons, it appears to be completely unpolarized, and
any polarization direction is observed with equal probability. However, the
results for both photons are perfectly correlated. For example, this means
that photon 2 has vertical polarization if we found horizontal polarization
for photon 1, but also that photon 2 will be circularly polarized left if we
observed right circular polarization for photon 1.
Another important feature of the four Bell states is that a manipulation
of only one of the two particles suffices to transform from any Bell state to
any of the other three states. This is not possible for the basis formed by the
products. For example, to transform |01 |02 into |11 |12 one has to flip the
state of both particles.
These three features,
• different statistical results for measurements on entangled or unentangled
pairs
• perfect correlations between the observations of the two particles of a pair,
although the results of the measurements on the individual particles are
fully random
• the possibility to transform between the Bell states by manipulating only
one of the two particles,
are the ingredients of the fundamental quantum communication schemes de-
scribed here.
3.3 The Quantum Communication Schemes
Quantum communication methods utilize fundamental properties of quan-

tum mechanics to enhance the power and potential of today’s communica-
tion systems. The first step towards quantum information processing is the
generalization of classical digital encoding, which uses the bit values “0” and
“1”. Quantum information associates two distinguishable, orthogonal states
of a two-state system with these bit values. We thus directly translate the
two values of a classical bit to the two basis states |0 and |1.
As an extension to the situation for classical communication, the quantum

system can be in any superposition of the two basis states. To distinguish such
a quantum state and the information contained in it from a classical bit, such
a state is called a “qubit” [69]. The general state of a qubit is
|Ψ = a0 |0 + a1 |1 , (3.6)

2 2
where a0 and a1 are complex amplitudes (with |a0 | + |a1 | = 1).
A measurement of the qubit projects the state onto either |0 or |1 and
therefore cannot give the full quantum information about the state. Evi-
dently, if we want to communicate information, we have to restrict ourselves
to sending only basis states in order to avoid errors, and thus only one bit
of classical information can be sent with a single qubit. Therefore, the new
features do not seem to offer additional power. However, by provoking errors,
quantum cryptography [42, 122] enables one to check the security of quantum
key generation. The security of quantum cryptography relies on the fact that
an eavesdropper cannot unambiguously read the state of a single quantum
particle which is transferred from Alice to Bob.
When two-particle systems are used, entanglement adds many more fea-
tures to the capabilities of quantum communication systems compared with
classical systems. In recent years, several proposals have shown how to ex-
ploit the basic features of entangled states in new quantum communication
schemes. In the following we shall see how entangled pairs enable a new for-
mulation of quantum cryptography (Sect. 3.3.1), how we can surpass the limit
of transmitting only one bit per qubit (Sect. 3.3.2) and how entanglement
allows one to transfer quantum information from one particle to another in
the process of quantum teleportation (Sect. 3.3.3).
Suppose two parties, let us call them Alice and Bob, want to send each
other secret messages. There exists a cryptographic method, the one-time
pad scheme,3 which is secure against eavesdropping attacks – provided the
key used for encoding and decoding the message is perfectly random, is as
long as the original message and, most importantly, is secret and known
only to Alice and Bob. But how can they be sure that the key was securely
distributed to the two, and that no third person has knowledge about the key?
Quantum cryptography [42, 122] provides a means to ensure the security of
3
In the so-called “one-time pad” encryption (see Sect. 3.1), every character of the
message is encoded with a random key character. As shown by Shannon [44], the
cipher cannot be decoded without a knowledge of the key. The eavesdropping
is impossible as long as the key is securely exchanged between the sender and
receiver.
Fig. 3.1. Scheme for entanglement-based quantum cryptography [46]
the key distribution and thus enables, together with the one-time pad scheme,
absolutely secret communication.4
Let us first discuss how quantum cryptography can profit from the fas-
cinating properties of entangled systems to provide secure key exchange
[46, 123]. Suppose that Alice and Bob receive particles which are in entan-
gled pairs, from an EPR source (Fig. 3.1). Beforehand, Alice and Bob agreed
on some preferred basis, again called |0/|1, in which they start to perform
measurements. The possible results, +1 and −1, correspond to observation
of the state |1 or |0, respectively. Owing to the entanglement of the parti-
cles, the measurement results of Alice and Bob will be perfectly correlated
or, in a case where the source produces pairs in the |Ψ − state, perfectly an-
ticorrelated. For each instance where Alice obtained −1, she knows that Bob
observed +1, and if she obtained the result +1, she knows that Bob had −1.
Alice and Bob can translate the result −1 to the bit value 0 and the result
+1 to the bit value 1 and thereby establish a random key, ideal for encoding
messages. But how can they be sure that no eavesdropper has intercepted
the key exchange? There are two different techniques. The first scheme for
entanglement-based quantum cryptography [123] builds on the ideas of the
basic quantum cryptography protocol for single photons [42, 122]. In this
case, Alice and Bob randomly and independently vary their analysis direc-
tions between 0◦ , corresponding to the |0/|1 basis, and 45◦ , corresponding
to a second, noncommuting basis. They will observe perfect anticorrelations
of their measurements whenever they happen to have polarizers oriented par-
allel (Alice and Bob thus obtain identical keys, if one of them inverts all bits
of his/her key string). This can be viewed in the following way: as Alice makes
4
For descriptions of quantum cryptography schemes not relying on entanglement,
see [105, 106].
a measurement on photon A she projects photon B into the orthogonal state,

which is then analyzed by Bob. An eavesdropper, not knowing the actual ba-
sis, causes errors, since he/she cannot determine the quantum state without
information about the basis. Thus, Alice and Bob can find out, by communi-
cation over a classical, public channel, whether or not their key exchange has
been attacked by checking whether or not some of the key bits are different.
Of course, those test key bits cannot be used for secure communication and
have to be sacrificed.
The other technique uses the fragility of entanglement against measure-
ments. Any attack an eavesdropper might perform reduces the entanglement
and allows Alice and Bob to check the security of their quantum key ex-
change. As described in Sect. 3.2, measurements on entangled pairs obey
statistical correlations and will violate a Bell inequality. It can be shown that
the more knowledge the eavesdropper has gained when he/she intercepted
the key exchange, the less the inequality is violated. The amount by which
a Bell inequality is violated is thus an ideal measure of the security of the
key. Alice and Bob therefore measure the entangled particles not only in the
basis |0/|1, but also along some other directions, depending on the Bell
inequality used. A particularly simple form of Bell inequality, which is well
suited for experimental application, is the version deduced by Wigner [28],
which can be used as follows.
Alice chooses between two polarization measurements of photon A, either
along the axis α or along the axis β, and Bob chooses between measurements
along β and γ of photon B. We identify the direction β, which is common
to the two users, with our preferred basis |0/|1. A detected polarization
parallel to the analyzer axis corresponds to a +1 result, and a polarization
orthogonal to the analyzer axis corresponds to −1. If, heretically, one assumes
that every photon carries preassigned values determining the outcomes of the
measurements on each of the photon pairs, it follows that the probabilities
of obtaining +1 on both sides, p++ , must obey Wigner’s inequality:
p++ (αA , βB ) + p++ (βA , γB ) − p++ (αA , γB ) ≥ 0 . (3.7)
The quantum mechanical prediction pqm
for these probabilities with some
++
arbitrary analyzer settings ΘA (Alice) and ΘB (Bob) and measurement of the
Ψ − state is
1
pqm
++ (ΘA , ΘB ) = sin2 (ΘA − ΘB ) . (3.8)
2
The analyzer settings α = −30◦ , β = 0◦ and γ = 30◦ lead to a maximum
violation of Wigner’s inequality (3.7):
◦ ◦ ◦ ◦ ◦ ◦
pqm qm qm
++ (−30 , 0 ) + p++ (0 , 30 ) − p++ (−30 , 30 )
1 1 3 1
= + − =− , (3.9)
8 8 8 8
which is not greater than or equal to 0.
In order to implement quantum key distribution, Alice and Bob each vary
their analyzers randomly between two settings: Alice uses −30◦ , 0◦ , and Bob
uses 0◦ , 30◦ . Because Alice and Bob operate independently, four possible com-
binations of analyzer settings will occur, of which the three oblique settings
allow a test of Wigner’s inequality and the remaining combination of parallel
settings allows the generation of keys via the perfect anticorrelation (where,
again, either Alice or Bob has to invert all bits of the key to obtain identical
keys). If the measured probabilities violate Wigner’s inequality, the security
of the quantum channel is ascertained, and the keys generated can readily
be used. This scheme is an improvement on the Ekert scheme, which uses
the CHSH inequality. Since there are fewer settings on each side, the above
version is technically easier to implement and also uses the photon pairs more
efficiently for key generation.
Compared with standard attenuated-pulse quantum cryptography, such
systems are practically immune to any beam-splitter attack (or other attacks
that try to split pulses containing more than one photon) by a potential
eavesdropper. First of all, a photon pair source can be used as an (almost)
ideal source of single photons. If one of the photons is detected, the gate
time of the coincidence electronics (typically on the order of 1 ns) determines
the equivalent pulse duration in standard quantum cryptography. Since the
probability of generating one photon pair during such a short time is very
low, e.g., for the experiment described in Sect. 3.5.1, only about 6.8 × 10−4 ,
the probability of having two photons in the gate time is less than 3 × 10−7
and can be almost neglected. This has to be compared with a probability of
having two photons in a pulse of 0.005 for a typical quantum cryptography
realization using a mean of 0.1 photons per pulse.
However, the security against beam-splitting attacks can be further in-
creased when entanglement-based schemes are used. In this case, there is
only a correlation between two entangled pairs if they are simultaneously
generated during a time interval of the order of the coherence time of the
photons, i.e. during a time of typically 500 fs. This reduces the chances of
an eavesdropper learning the value of a key bit to about 6 × 10−14 and guar-
antees unprecedented security of the quantum key. Moreover, by utilizing
the peculiar properties of entangled photon pairs produced by parametric
down-conversion, one immediately profits from the inherent randomness of
quantum mechanical observations, which guarantees a truly random and non-
deterministic key.
3.3.2 Quantum Dense Coding
If one wants to send some information, one encodes the message with dis-
tinguishable symbols, writes them on some physical entity and finally, this is
transmitted to the receiver. To send one bit of information one uses, for exam-
ple, the binary values “0” and “1” as code symbols written on the information
Fig. 3.2. Scheme for the efficient transmission of classical information by quantum
dense coding [51] (BSM, Bell-state measurement; U, unitary transformation)
carrier. If one wants to send two bits of information, one consequently has to
perform the process twice; that means one has to send two such entities.
As mentioned above, in the case of quantum information one identifies
the two binary values with the two orthogonal basis states |0 and |1 of
the qubit. In order to send a classical message to Bob, Alice uses quantum
particles, all prepared in the same state by some source. Alice translates the
bit values of the message by either leaving the state of the qubit unchanged or
flipping it to the other, orthogonal state, and Bob, consequently, will observe
the particle in one or the other state. That means that Alice can encode one
bit of information in a single qubit. Obviously, she cannot do better, since in
order to avoid errors, the states arriving at Bob have to be distinguishable,
which is only guaranteed when orthogonal states are used. In this respect,
they do not gain anything by using qubits as compared with classical bits.
Also, if she wants to communicate two bits of information, Alice has to send
two qubits.
Bennett and Wiesner found a clever way to circumvent the classical limit
and showed how to increase the channel capacity by utilizing entangled par-
ticles [51]. Suppose the particle which Alice obtained from the source is en-
tangled with another particle, which was sent directly to Bob (Fig. 3.2). The
two particles are in one of the four Bell states, say |Ψ − . Alice now can use a
particular feature of the Bell basis, that manipulation of one of the two entan-
gled particles suffices to transform to any other of the four Bell states. Thus
she can perform one out of four possible transformations – that is, doing
nothing, shifting the phase by π, flipping the state, or flipping and phase-
shifting the state – to transform the two-particle state of their common pair
to another state. After Alice has sent the transformed two-state particle to
Bob, he can read the information by performing a combined measurement on
both particles. He makes a measurement in the Bell-state basis and can iden-
tify which of the four possible messages was sent by Alice. Thus it is possible
to encode two bits of classical information by manipulating and transmit-
ting a single two-state system. Entanglement enables one to communicate
information more efficiently than any classical system could do.
The preceding examples show how quantum information can be applied
for secure and efficient transmission of classical information. But can one
also transmit quantum information, that is, the state of a qubit? Obviously,
quantum mechanics places a number of obstacles in the way of this intention,
above all, the problem of measuring quantum states, which is utilized in
quantum cryptography as already described.
3.3.3 Quantum Teleportation
The Idea. It is an everyday task, in our classical world, for Alice to send
some information to Bob. Imagine a fax machine. Alice might have some
message, written on a sheet of paper. For the fax machine the actual written
information does not matter, in fact, it reduces to just a sequence of white
and black pixels. For the transmission, the machine scans the paper pixel by
pixel. It measures whether a pixel is white or black and sends this information
to Bob’s machine, which writes the state of each pixel onto another sheet of
paper. In classical physics, by definition, one can make the measurements with
arbitrary precision, and Bob’s sheet can thus become an ideal copy of Alice’s
original sheet of paper. If Alice’s pixels were made smaller and smaller, they
would, in reality, sooner or later be encoded on single molecules or atoms.
If we again confined ourselves to coding in only the basis states, we surely
could measure and transfer the binary value of even such pixels.
Now, imagine Alice not only has classical binary values encoded on her
system, but wants to send a quantum state, i.e. quantum information, to
Bob. She has a qubit encoded on some quantum system such as a molecule
or atom, and wishes, that a quantum system in Bob’s hands should represent
this qubit at the end of the transmission. Evidently, Alice cannot read the
quantum information, that is, measure the state of the quantum object with
arbitrary precision. All she would learn from her measurement would be that
the amplitude of the observed basis state was not zero. But this is not enough
information for Bob to reconstruct the qubit on his quantum particle.
Another limitation, which definitely seems to bring the quest for perfect
transfer of the quantum information to an end, is the no-cloning theorem
(see Sect. 3.1) [4]. According to this theorem, the state of a quantum system
cannot be copied onto another quantum system with arbitrary precision.
Thus, how could Bob’s quantum particle obtain the state of Alice’s particle?
In 1993 Bennett et al. found the solution to this problem [52]. In their
scheme, a chain of quantum correlations is established between the particle
Fig. 3.3. Scheme for teleporting a quantum state from one system to another one
[52]
carrying the initial quantum state and Bob’s particle. They dispense with
measuring the initial state; actually, they avoid gaining any knowledge about
this state at all!
To perform quantum teleportation, initially Alice and Bob share an en-
tangled pair of particles 2 and 3, which they have obtained from some source
of entangled particles, in, say, the state |Ψ − 2,3 (Fig. 3.3). As mentioned be-
fore, we cannot say anything about the state of particle 2 on its own. Nor do
we know the state of particle 3. In fact, these particles do not have a (pure)
state at all. But, whatever the results of measurements might be, we know
for sure that they are orthogonal to each other. Next, particle 1, which car-
ries the state to be sent to Bob, is given to Alice. She now measures particle
1 and 2 together, by projecting them onto the Bell-state basis. After pro-
jecting the two particles into an entangled state, she cannot infer anything
about the individual states of particles 1 and 2 anymore. However, she knows
about correlations between the two. Let us assume she has obtained the result
|Ψ − 1,2 . This tells her, that whatever the two states of particles 1 and 2 have
been, they have been orthogonal to each other. But from this, Alice already
knows that the state of particle 3 is equal to the state of particle 1 (up to a
possible overall phase shift). This follows because the state of particle 1 was
orthogonal to 2 and, owing to the preparation of particles 2 and 3, the state
of particle 2 was orthogonal to 3. All Alice has to do is to tell this to Bob,
to let him know that, in this particular case, the state of his particle 3 is the
same as that which particle 1 had initially.
Of course, since there are four orthogonal Bell states, there are four equally
probable outcomes for Alice’s Bell-state measurement. If Alice has obtained
another result, the state of Bob’s particle is again related to the initial state
of particle 1, up to a characteristic unitary transformation. This stems from

the fact that a unitary transformation of one of two entangled particles can
transform from any Bell state to any other.
Therefore, Alice has to send the result of her Bell-state measurement
(i.e. a number between 0 and 3, or, equivalently, two bits of information)
via a classical communication channel to Bob. He then can restore the initial
quantum state of particle 1 on his particle 3 by performing the correct unitary
transformation.
Formally, we describe the initial state of particle 1 by |χ1 = a|H1 +b|V 1 ,
and the state of the EPR pair 2 and 3 by |Ψ − 2,3 . Then the joint three-photon
system is in the product state
|Ψ 1,2,3 = |χ1 ⊗ |Ψ − 2,3 , (3.10)
which can be decomposed into

1
|Ψ 1,2,3 = [ −|Ψ − 1,2 (a|H3 + b|V 3 ) − |Ψ + 1,2 (a|H3 − b|V 3 ) (3.11)
2
+|Φ− 1,2 (a|V 3 + b|H3 ) + |Φ+ 1,2 (a|V 3 − b|H3 ) ] .
One easily sees that after observation of particles 1 and 2 in one of the
four Bell states, the corresponding unitary transformation enables Bob to
transfer the initial state of particle 1 to particle 3.
Some Remarks. The principle of quantum teleportation incorporates all
the characteristic features of entangled systems, and, in an astounding man-
ner, profits from the obstacles seemingly imposed by quantum mechanics. It
should be emphasized that quantum teleportation is well within the concepts
of conventional physics and quantum mechanics. Let us briefly discuss a few
not infrequently occurring misunderstandings.
First, the no-cloning theorem is not violated. The state of particle 1 can
only be restored on particle 3 if the measurement performed by Alice does not
give any information about the state! After Alice’s Bell-state measurement,
particle 1 is in a mixed state, which is absolutely uncorrelated with the initial
state of particle 1. Therefore, the particular quantum state which is teleported
can be attributed to only one particle at a time, never to two.
Secondly, there is no faster-than-light communication achieved in quan-
tum teleportation. Even if Alice knows, right after her measurement, that
Bob’s particle is already either in the correct state or in one of the three
other possible states, she has to send this information to Bob. The classical
information sent to Bob is transmitted, according to the theory of relativ-
ity, at the speed of light at maximum. Only after receiving the result and
after performing the correct unitary transformation can Bob restore the ini-
tial quantum state. If Bob does not know the result of Alice’s measurement,
his particle is in a mixed state, which is not correlated at all with the ini-
tial state. Thus quantum information, the qubit, cannot be transferred faster
than classical information.
Fig. 3.4. Scheme for entangling particles that have never interacted by the process
of entanglement swapping [125]
And, thirdly, there is also no transfer of matter or energy (other than that
required for the transmission of classical information). All that makes up a
particle are its properties, described by the quantum state. For example, the
state of a free neutron defines its momentum and its spin. If one transfers
the state onto another neutron, this particle obtains all the properties of the
first one; in fact, it becomes the initial particle. We leave it to the science
fiction writers to apply the scheme to bigger and bigger objects. Whether or
not this idea will help some Captain Kirk to get back to his space ship or
not cannot be answered here. Certainly, a lot of other problems need to be
solved as well.5
It is appropriate to point out some generalizations of the principle of
quantum teleportation. It is not necessary that the initial state which is to
be teleported is a pure state. In fact it can be any mixed state, or even the
undefined state of an entangled particle. This is best demonstrated by entan-
glement swapping [125]. Here, the particle to be teleported (1) is entangled
with yet another one (4) (Fig. 3.4). The state of particle 1 on its own is a
mixed state; however, it can be determined by the observation of particle 4.
Quantum teleportation allows us to transfer the state of particle 1 onto par-
ticle 3. Since quantum teleportation works for any arbitrary quantum state,
particle 3 thus becomes entangled with particle 4. Note, that particles 3 and 4
do not come from the same source, nor did they ever interact with each other.
Nevertheless, it is possible to entangle them by swapping the entanglement
in the process of quantum teleportation.
5
The “technical manuals of Star Trek” mention, as a necessary part of their trans-
porter, a “Heisenberg compensator” [124]. Quantum teleportation seems to pro-
vide a solution for this marvelous device. However, a lot more is necessary to
beam large objects.
Fig. 3.5. Remote state preparation of Bob’s particle 2, by a manipulation (M) of

particle 1
Quantum teleportation is not confined to transferring two-state quantum

systems. If Alice and Bob share an entangled pair of N-state particles, they
can teleport the state of an N-dimensional quantum system [126]. As before,
Alice performs projection onto the N2 -dimensional basis of entangled states
spanning the product space of particles 1 and 2. The result, one out of N2
equally probable results, has to be communicated to Bob, who then can again
restore the initial state of particle 1 by the corresponding unitary transfor-
mation of his particle 3. If Alice and Bob share a pair of particles that are
entangled in the original sense of EPR, that is, for continuous variables or
∞-dimensional states, they also can teleport properties such as the position
and momentum of particles or the phase and amplitude of electromagnetic
fields [11].
A considerable simplification of quantum teleportation, especially in terms
of experimental realization, transfers not the quantum state of a particle, but
rather the manipulation performed on the entangled particle which is given
to Alice [127] (Fig. 3.5). Again, one first distributes an entangled pair to Alice
and Bob. But before Alice gets hold of her particle 1 and can perform mea-
surements on it, the state of this particle is manipulated in another degree of
freedom. One cannot talk about a two-state system anymore. Rather, particle
1 now is described in a four-dimensional Hilbert space, spanned by the orig-
inal degree of freedom and the new one. Formally, however, this mimics the
two two-state particles given to Alice in the standard quantum teleportation
scheme. Consequently, a measurement in the four-dimensional Hilbert space
of particle 1, which perfectly erases the quantum information by mixing the
two degrees of freedom, is performed. This gives the necessary information for
Bob to perform the correct unitary transformation on the particle. That way,
the originally mixed state of particle 2 can be turned into a pure state which
depends on the manipulation initially performed on particle 1. Using such
a scheme, one can remotely prepare particle 3 in any pure quantum state.
Thus, it is not necessary to send two real numbers to Bob if one wants him to
have a certain, pure quantum state prepared on his particle. If he is provided
with one of a pair of entangled particles, Alice simply has to transmit two
bits of classical information to Bob.
3.4 The Experimental Prerequisites

Before turning to the fascinating applications of entangled systems, let us re-
view how to produce, how to manipulate and how to measure such quantum
systems experimentally. The last decade saw incredible progress in the ex-
perimental techniques for handling various quantum systems. However, there
are additional challenges when working with entangled systems, especially
the careful control of interactions and decoherence of the quantum systems.
In their seminal work Einstein, Podolsky and Rosen considered particles
which interacted with each other for a certain time and which thus became
entangled and thereafter exhibited the puzzling, nonclassical correlations.
The interaction needed to entangle a pair of particles is just the same as von
Neumann had in mind when describing the measurement process. Ideally, it
couples two quantum systems in such a way that, if the first system is in one
of a set of distinguishable (orthogonal) states, the second system will change
into a well-defined, corresponding state. Let us look at such a coupling for
the simplest case of two two-state systems. As before, the two basis states
are denoted as |0 and |1. The coupling is such that if system 1 is in state
|01 , system 2 will remain in its initial state, say |02 , whereas if system 1
is in state |11 , system 2 will flip to the orthogonal state, i.e. to |12 . The
nonclassical features arise if system 1 is in a superposition of its basis states.
Then, coupling it with the second system results in an entangled state:
|01 |02 −→ |01 |02 ,

|11 |02 −→ |11 |12 ,
1 1
√ (|01 + |11 ) |02 −→ √ (|01 |02 + |11 |12 ) . (3.12)
2 2
Although this basic principle of producing entangled states has been

known since the very beginning of quantum mechanics, until recently there
was no physical system where the necessary coupling could be realized. The
progress in cavity QED [36] and ion trap experiments [128] allowed the first
observation of entanglement between two atoms or two ions. These experi-
ments are of great importance for the further development of experimental
quantum computation. However, for quantum communication one needs to
transfer the entangled particles over reasonable distances. Thus photons (with
wavelengths in the visible or near infrared) are clearly a better choice. For
entangling photons via such a coupling, various methods have been proposed
and partially realized [129, 130, 131] but still need to be investigated more
thoroughly. Fortunately, the process of parametric down-conversion offers an
ideal source of entangled photon pairs without the need for strong coupling
(see Sect. 3.4.2).
To perform Bell-state analysis, one first has to transform the entangled
state into a product state. This is necessary since two particles can be an-
alyzed only if they are measured separately. Otherwise one would need to
entangle the two measurement apparatuses, each of which analyzes one of the
two particles – clearly an even more challenging task. In principle, a disentan-
gling transformation can be performed by reversing the entangling interaction
described above. However, as long as such couplings are not achievable, one
has to find replacements. In the following it is shown how two-particle inter-
ference can be employed for partial Bell-state analysis (see Sect. 3.4.3). Since
the manipulations and unitary transformations have to be performed on only
one quantum particle at a time, this does not create new obstacles. These
operations are often routine; in the case of light they have been routine for
two centuries.
3.4.1 Entangled Photon Pairs
Entanglement between photons cannot be generated by coupling them via

an interaction yet. However, there are several emission processes, such as
atomic cascade decays or parametric down-conversion, where, owing to the
conservation of energy and of linear or angular momentum, the properties of
two emitted photons become entangled.
Historically, entanglement between spatially separated quanta was first
observed in measurements of the polarization correlation between γ + γ − emis-
sions in positron annihilation [132], soon after Bohm’s proposal for observing
EPR phenomena in spin-1/2 systems. After Bell’s discovery that contradic-
tory predictions between quantum theories can actually be observed, a series
of measurements was performed, mostly with polarization-entangled photons
from a two-photon cascade emission from calcium [133]. In these experiments,
the two photons were in the visible spectrum, and thus could be manipulated
and controlled by standard optical techniques. Of course, this is a great ad-
vantage compared with the positron annihilation source; however, the two
photons are now no longer emitted in opposite directions, since the emitting
atom carries away some randomly determined momentum. This makes exper-
imental handling more difficult and also reduces the brightness of the source.
The process of parametric down-conversion now offers a new possibility for
efficiently generating entangled pairs of photons [111].
When light propagates through an optically nonlinear medium with second-
order nonlinearity χ(2) (only possible in noncentrosymmetric crystals), the
Fig. 3.6. The different relations between the emission directions for type I and type
II down-conversion
conversion of a light quantum from the incident pump field into a pair of
photons in the “idler” and “signal” modes can occur. In principle, this can
be seen as the inverse of the frequency-doubling process in nonlinear optics
[134].
As mentioned above, energy and momentum conservation can give rise
to entanglement in various degrees of freedom, such as position–momentum
and time–energy entanglement. However, the interaction time and volume
will determine the sharpness and quality of the correlations observed , which
are formally obtained by integration of the interaction Hamiltonian [135].
The interaction time is given by the coherence time τc of the UV pump light;
the volume is given by the extent and spatial distribution of the pump light
in the nonlinear crystal.
The relative orientations of the direction and polarization of the pump
beam, and the optic axis of the crystal determine the actual direction of
the emission of any given wavelength. We distinguish two possible alignment
types (Fig. 3.6): for type I down-conversion, the pump has, for example,
the extraordinary polarization and the idler and signal beams both have the
ordinary polarization. Different colors are emitted into cones centered on the
pump beam.
In type II down-conversion, the pump has the extraordinary polarization
and, in order to fulfill the momentum conservation condition inside the crystal
(phase-matching), the two down-converted photons have different, for most
directions orthogonal, polarizations, offering the possibility of a new source
of polarization-entangled photon pairs (Sect. 3.4.2).
One can distinguish two basic ways to observe entanglement. In the first
way, by selecting detection events one can chose a subensemble of possible
outcomes which exhibits the nonclassical features of entangled states.6 This
additional selection seems to contradict the spirit of EPR–Bell experiments;
however, it was shown recently, that, after a detailed analysis of all detection
events, the validity of local hidden-variable theories can be tested on the basis
6
For the observation of polarization entanglement, see [136]. For momentum entan-
glement, see the proposal [137] and the experimental results [138]. Time–energy
entanglement was proposed in [139]. Experiments are described in [140].
Fig. 3.7. Photons emerging from type II

down-conversion. The photons are always
emitted with the same wavelength but have
orthogonal polarizations. At the intersection
points their polarizations are undefined but
different, resulting in entanglement
of refined Bell inequalities [141]. Therefore, such sources can also be useful
for entanglement-based quantum cryptography [142].
In the second way, true entangled photon pairs can be generated. This is
essential for all the other quantum communication schemes, where one cannot
use the detection selection method. Several methods to obtain momentum-
entangled pairs [143] have been demonstrated experimentally [144], but are
extremely difficult to handle experimentally owing to the huge requirements
on the stability of the whole setup. Any phase change, i.e. a change in the path
lengths by as little as 10 nm, is devastating for the experiment. Also, the re-
cently developed source of time–energy-entangled photon pairs [145] partially
shares these problems and, to avoid detection selection, requires fast optical
switches. Fortunately, with polarization entanglement as produced by type
II parametric down-conversion, the stability requirements are considerably
more relaxed.
3.4.2 Polarization-Entangled Pairs from Type II

Down-Conversion
In type II down-conversion, the down-converted photons are emitted into

two cones, one with the ordinary polarization and the other with the ex-
traordinary polarization. Because of conservation of transverse momentum
the photons of each pair must lie on opposite sides of the pump beam. For
the proper alignment of the optic axis of the nonlinear crystal, the two cones
intersect along two lines (see Fig. 3.7) [111, 146]. Along the two directions
(“1” and “2”) where the cones overlap, the light can be essentially described
by an entangled state:

|Ψ = |H1 |V 2 + eiα |V 1 |H2 , (3.13)
where the relative phase α arises from the crystal birefringence, and an overall
phase shift is omitted. Using an additional birefringent phase shifter (or even
by slightly rotating the down-conversion crystal itself), the value of α can be
set as desired, e.g. to the value 0 or π. Thus, polarization-entangled states
are produced directly out of a single nonlinear crystal (beta barium borate,
BBO), with no need for extra beam splitters or mirrors and no requirement
to discard detected pairs.
Best of all, by using two extra birefringent elements, one can easily pro-
duce any of the four orthogonal Bell states. For example, when starting with
the state |Ψ + , a net phase shift of π and thus a transformation to the state
|Ψ − may be obtained by rotating a quarter-wave plate in one of the two
paths by 90◦ from the vertical to the horizontal direction. Similarly, a half-
wave plate in one path can be used to change a horizontal polarization to
vertical and to switch to the states |Φ± .
The birefringent nature of the down-conversion crystal complicates the
actual entangled state produced, since the ordinary and the extraordinary
photons have different velocities inside the crystal, and propagate along dif-
ferent directions even though they become parallel and, for short crystals,
collinear outside the crystal. The resulting longitudinal and transverse walk-
off between the two polarizations in the entangled state is maximal for pairs
created near the entrance face of the crystal, which consequently acquire the
greatest time delay and relative lateral displacement. Thus the two possible
emissions become, in principle, distinguishable by the order in which the de-
tectors would fire or by their spatial location, and no entanglement will be
observable. However, the photons are produced coherently along the entire
length of the crystal. One can thus completely compensate for the longitudi-
nal walk-off and partially for the transverse walk-off by using two additional
crystals, one in each path [147]. By verifying the correlations produced by
this source, one can observe strong violations of Bell’s inequalities (modulo
the typical auxiliary assumptions) within a short measurement time [31].
The experimental setup is shown in Fig. 3.8a. The 351.1 nm pump beam
(150 mW) is obtained from a single-mode argon ion laser, followed by a dis-
persion prism to remove unwanted laser fluorescence (not shown) [111]. Our
3 mm long BBO crystal was nominally cut such that θpm , the angle between
the optic axis and the pump beam, was 49.2◦, to allow collinear, degener-
ate operation when the pump beam is precisely orthogonal to the surface.
The optic axis was oriented in the vertical plane, and the entire crystal was
tilted (in the plane containing the optic axis, the surface normal and the
pump beam) by 0.72◦ , thus increasing the effective value of θpm inside the
crystal to 49.63◦ . The two cone overlap directions, selected by irises before
the detectors, were consequently separated by 6.0◦ . Each polarization ana-
lyzer consisted of two-channel polarizers (polarizing beam splitters) preceded
Fig. 3.8. (a) Experimental setup for the observation of entanglement produced by
a type II down-conversion source. The additional birefringent crystals are needed
to compensate for the birefringent walk-off effects from the first crystal. (b) Coin-
cidence fringes for the Bell states |Ψ + (•) and |Ψ + (◦) obtained when varying the
analyzer angle Θ1 , with Θ2 set to 45◦
by a rotatable half-wave plate. The detectors were cooled silicon avalanche

photodiodes operated in the Geiger mode. Coincidence rates C (θ1 , θ2 ) were
recorded as a function of the polarizer settings θ1 and θ2 .
In this experiment the transverse walk-off (0.3 mm) was small compared
with the coherent pump beam width (2 mm), so the associated labeling effect
was minimal. However, it was necessary to compensate for the longitudinal
walk-off, since the 3.0 mm BBO crystal produced a time delay which was
about the same as the coherence time of the detected photons (≈390 fs, de-
termined by interference filters with a width of 5 nm at 702 nm). We used
an additional BBO crystal (1.5 mm thick) as a compensator in each of the
paths, preceded by a half-wave plate to exchange the roles of the horizontal
and vertical polarizations.
Under such conditions, we now obtain routinely a coincidence fringe visi-
bility (as polarizer 2 is rotated, with polarizer 1 fixed at −45◦ ) of more than
97%, for irises with a size of 2 mm at a distance of 1.5 m from the crystal (Fig.
3.8b). The high quality of this source is crucial for the overall performance of
our experiments in quantum dense coding [114], quantum cryptography [113]
and tests of Bell’s inequalities [31]. For the later experiments, the photons
were coupled into single-mode fibers, to bridge long distances of the order
of 400 m. To achieve a high coupling, the pump beam should be slightly fo-
cused into the BBO crystal, to optimally match the microscope objectives
used. Since the compensation crystals partially compensate the transverse
walk-off, focusing down to 0.2 mm is not crucial. Visibilities of more than
98% have been obtained this way, with an overall collection and detection
efficiency of 10%.
Such a source has a number of distinct advantages. It seems to be rel-
atively insensitive to larger collection irises, an important feature in exper-
iments where high count rates are crucial. In addition, owing to its sim-
plicity, the source is much quicker to align than other down-conversion set-
ups and is remarkably stable. One of the reasons is that phase drifts are
not detrimental to a polarization-entangled state unless they are birefrin-
gent, i.e. polarization-dependent – this is a clear advantage over experiments
with momentum-entangled or energy–time-entangled photon pairs. Recently,
Kwiat and coworkers tested sandwiched type I crystals and achieved, for thin
crystals, a significantly higher relative yield of entangled photon pairs [148].
Also, utilizing cavities to enhance the pump field in the nonlinear crystal can
boost the output by a factor of 20 [149]. This gives hope that even more
efficient generation of entangled photon pairs will be obtained in the future.
3.4.3 Interferometric Bell-State Analysis
At the heart of Bell-state analysis of a pair of particles is the transformation

of an entangled state to an unentangled, product state. The necessary cou-
pling, however, has not been achieved for photons yet. But it turns out that
interference of two entangled particles, and thus the photon statistics behind
beam splitters depend on the entangled state that the pair is in [112, 150, 151].
The Principle. Let us discuss first the generic case of two interfering par-
ticles. If we have two otherwise indistinguishable particles in different beams
and overlap these two beams at a beam splitter, we ask ourselves, what is the
probability to find the two particles in different output beams of the beam
splitter (Fig. 3.9a). Alternatively we can ask, what is the probability that
two detectors, one in each output beam, detect one photon each.
If we performed this experiment with fermions, we would at first naively
expect the two fermions to arrive in different output beams. This is sug-
gested by the Pauli principle, which requires that the two particles cannot
be in the same quantum state, that is, they cannot exit in the same output
beam. Analogously, interference of bosons at a beam splitter will result in
the expectation of finding both bosons in one output beam. For a symmetric
50/50 beam splitter, it is fully random whether the two bosons will be de-
tected in the upper or the lower detector, but they will be always detected
by the same detector. However, it is important to realize that the statements
above are only correct if one disregards the internal degrees of freedom of the
interfering particles.
Ultimately, the reason for the different behaviors lies in the different sym-
metries of the wave functions describing bosonic and fermionic particles [150].
There are four different possibilities for how the two particles could propa-
gate from the input to the output beams of the beam splitter. We obtain
one particle in each output if both particles are reflected or both particles
are transmitted; we observe both particles at one detector if one particle
is transmitted and the other reflected, or vice versa. For the antisymmetric
states of fermions, the two possibilities of both particles being transmitted
and both being reflected interfere constructively, resulting in firing of each
Fig. 3.9. (a) Interference of two particles at a beam splitter. The observation of
coincident detection, i.e. detection of one particle at each of the two detectors, is
sensitive to the symmetry of the spatial component of the quantum state of the
combined system. (b) Bell-state analyzer for identifying the Bell states |Ψ + and
|Ψ − by observing different types of coincidences. The other two Bell states |Φ±
exhibit the same detection probabilities (both photons are detected by one detector)
for this setup and cannot be distinguished
of the two detectors. For the symmetric states of bosons, these two ampli-
tudes interfere destructively, giving no simultaneous detection in different
output beams [152]. For photons with identical polarizations, which means
for bosons, this interference effect has been known since the experiments by
Hong et al. [153],7 but up to now it has not been observed for fermions yet.
What kinds of interference effects of two photons at a beam splitter are to
be expected if we consider also the internal degree of freedom of the photons,
i.e. their polarization? In particular, if we interfere two polarization-entangled
photons at a beam splitter, the Bell state describes only the internal degree of
freedom. Inspection of the four Bell states shows that the state |Ψ − is anti-
symmetric, whereas the other three are symmetric. However, if two particles
interfere at a nonpolarizing beam splitter, what matters is only the spatial
part of the wave function. The symmetry of the wave function is determined
by the requirement that for two photons, the total state has to be symmetric
again. We therefore obtain, for the total state of two photons in the anti-
symmetric Bell state formed from two beams a and b at the beam splitter,
1
|Ψ = (|H1 |V 2 − |V 1 |H2 ) (|a1 |b2 − |b1 |a2 ) . (3.14)
2
7
For further experiments and theoretical generalizations, see [154].
This means that, for the state |Ψ − , we also have an antisymmetric spatial
part of the wave function and thus expect a different detection probability,
that is, different coincidences between the two detectors, compared with the
other three Bell states.
We therefore can discriminate the state |Ψ − from all the other states. It
is the only one which leads to coincidences between the two detectors in the
output beams of the beam splitter. Can we also identify the other Bell states?
If two photons are in the state |Ψ + , they will both propagate in the same
output beam but with orthogonal polarizations in the horizontal/vertical
(H/V) basis, whereas two photons in the state |Φ+ or in the state |Φ− ,
which also both leave the beam splitter in the same output arm, have the
same polarization in the H/V basis. Thus we can discriminate between the
state |Ψ + and the states |Φ± by a polarization analysis in the H/V basis
and by observing either coincidences between the outputs of a two-channel
polarizer or both photons again in only one output (Fig. 3.9b). Note that
reorientation of the polarization analysis allows one to separate any other of
these three states from the other two, but it is not possible to distinguish
between all of them simultaneously [155]. If the photons were entangled in
yet another degree of freedom, i.e. they were four-state systems rather than
regular qubits, one could also discriminate between the states |Φ+ and |Φ−
[156]. But up to now, no quantum communication scheme seems to have
profited from this fact.
Summarizing, we conclude that two-photon interference can be used to
identify two of the four Bell states, with the other two giving the same third
detection result. One thus cannot perform complete Bell-state analysis by
these interferometric means, but we can identify three different settings in
quantum dense coding and, for teleportation, even identification of only one
of the Bell states is sufficient to transfer any quantum state from one particle
to another, although then only in a quarter of the trials.
Bell-State Analysis of Independent Photons. The above description

of how to apply two-photon interference for Bell-state analysis can give only
some hints about the possible procedures. One intuitively feels that the neces-
sary joint detection of the two photons has to be “in coincidence”. But what
really are the experimental requirements for the two photons to interfere?
The coincidence conditions can be obtained using a more refined analysis
that takes the multimode nature of the states involved into account [157].
Interference occurs only if the contributing possibilities for finding one
photon in each output are indistinguishable. If the two photons come from
different sources or, as is the case in the experiments, from different down-
conversion emissions, there might be some timing information, in our case
detection of the second photon from each down-conversion, which might ren-
der the possibilities distinguishable.
For example, if we detect one photon behind the beam splitter at almost
the same time as one of the additional down-conversion photons, we can infer
the origin of the photon that is to interfere. However, if the time difference
between the detection events of the two interfering photons, that is, the over-
lap at the beam splitter, is much less than their coherence time, then the
detection of any other photon cannot give any additional information about
their origin. This ultra-coincidence condition requires the use of narrow filters
in order to make the coherence time as long as possible. However, even if we
consider using state-of-the-art interference filters that yield a coherence time
of about 3 ps, no detectors fast enough exist at present. And an even stronger
filtering by Fabry–Perot cavities (to achieve the necessary coherence time of
about 500 ps) results in prohibitively low count rates. Only a considerable
increase of the number of photon pairs emitted into a narrow wavelength
window may allow one to use this technique (e.g. with a subthreshold OPO
configuration as demonstrated in [158]).
The best choice, as it turns out, is not to try to detect the two photons
simultaneously, but rather to generate them with a time definition much
better than their coherence time. Consider two down-conversion processes
pumped by pulsed UV beams (using either two crystals or, as is the case
in our experiments, one crystal pumped by two passages of a UV beam).
Again we attempt to observe interference between two photons, one from each
down-conversion process. Then, without any narrow filters in the beams, the
tight time correlation of the photons coming from the same down-conversion
permits one again to associate simultaneously detected photons with each
other. This provides path information and hence prohibits interference.
We now insert filters before (or behind) the beam splitter. With stan-
dard filters, and thus also with high enough count rates, one easily achieves
coherence times on the order of 1 ps. And it is possible to pump the two down-
conversion processes with UV pulses with a duration shorter than 200 fs. Thus
it follows that the photons detected behind the beam splitter carry practi-
cally no information anymore on the detection times of their twin photons,
and, vice versa, detection of those latter photons does not give which-path
information, which would destroy the interference.
The “coincidence time” for registering the photons now can be very long;
it merely needs to be shorter than the repetition time of the UV pulses, which
is on the order of 10 ns for commercially available laser systems. One thus
can expect very good visibility of interference and very good precision of the
Bell-state analysis.
3.4.4 Manipulation and Detection of Single Photons
For polarization-entangled photons, the unitary transformations transform-

ing between the four Bell states can be performed with standard half-wave
and quarter-wave retardation plates.
As mentioned before, in order to have the maximum freedom in setting

any of the Bell states, one inserts one half-wave and one quarter-wave plate
into the beam. By precompensating the additional quarter-wave shifts with
the compensator plates of the EPR source, one obtains at the output of the
transformation plates the state |Ψ − if both optic axes are aligned along the
vertical direction. Rotation of only the quarter-wave plate to the horizontal
direction transforms this state to |Ψ + , and rotation of only the half-wave
plate by 45◦ gives |Φ− . Finally, rotating one plate by 90◦ and the other one
by 45◦ gives |Φ+ .
For initial experimental realizations of the ideas of quantum communica-
tion, such a static polarization manipulation is sufficient. However, for quan-
tum cryptography, and also for practical applications of other schemes, one
would like to be able to switch the unitary transformation rapidly to any po-
sition. This can be achieved by fast Pockels cells. Depending on the applied
voltage, these devices have different indices of refraction for two orthogonal
polarization components, and can be used in a similar way to the quartz
retardation plates [31].
Detection of the single photons has been performed using silicon avalanche
photodiodes operated in the Geiger mode. The diodes used have a detection
efficiency of about 40%. Owing to losses in the interference filters and other
optical components, the overall detection efficiency of a photon emitted from
the source was around 10% in cw experiments; for experiments using a pulsed
source, we achieved an efficiency of only about 4%. In many interference
experiments, a good definition of the transverse-mode structure of the beams
is necessary. An ideal solution for achieving high interference contrast is thus
to couple the output arms of a beam splitter into single-mode fibers and
connect these fibers to pigtailed avalanche photodiodes. The single-mode fiber
acts as a very good spatial filter for the transverse modes and couples the
light efficiently to the diodes.
3.5 Quantum Communication Experiments

In the first experiments [159], the researchers concentrated on the distribution

of pairs of entangled photons over large lengths of fibers, rather than on
including fast, random switching. In these indoor experiments, where the
optical fiber was wound on a fiber drum, one of the photons was chosen
to have a wavelength of λ = 1300 nm, and the other in the near infrared
for optimal detection efficiency (here the down-conversion was pumped by
a krypton ion laser at 460 nm). Time–energy entanglement was used, with
asymmetric interferometers at the observer stations and selection of true
coincidences. Such a scheme allows the correlated photons to have a wide
frequency distribution, and thus a relatively high intensity, since the visibility
Fig. 3.10. Setup for entanglement-based quantum cryptography. The polarization-

entangled photons are transmitted via optical fibers to Alice and Bob, who are
separated by 400 m, and both photons are analyzed, detected and registered in-
dependently. After a measurement run, the quantum keys are established by Alice
and Bob through classical communication over a standard computer network
of the interference effects depends on the monochromaticity of the pump laser

light.
In a more recent experiment, both photons were produced with a wave-
length around 1300 nm. Here, for the first time, laser diodes (λ = 650 nm)
were used for pumping the down-conversion, in contrast to the expensive
laser systems used in other experiments. This allowed the demonstration of
nonclassical correlations between two observers separated by more than 10
km in the Geneva area [30]. Standard optical telecommunication fibers con-
necting offices of the Swiss telecommunication company were used to send
the photons to two interferometers, where phase modulation served to set
the analysis parameters. The robustness of the source, together with the
high degree of quantum entanglement, opens new prospects for this secure
communication technique.8
The scheme of the first realistic demonstration of entanglement-based key
distribution is sketched in Fig 3.10 [113]. The source uses type II parametric
down-conversion in BBO, pumped with an argon ion laser working at a wave-
length of 351 nm and a power of 350 mW. The photons, with a wavelength
of 702 nm, are each coupled into 500 m long optical fibers and transmitted to
Alice and Bob, who are separated by 400 m. Alice and Bob both use a Wol-
laston polarizing beam splitter as a polarization analyzer. We shall associate
a detection of parallel polarizations, +1, with the key bit 1 and detection of
orthogonal polarizations, −1, with the key bit 0.
Electrooptic modulators in front of the analyzers rapidly switch (<15 ns)
the axis of the analyzer between two desired orientations, controlled by quan-
tum random signal generators. These quantum random number generators
are based on the quantum mechanical process of splitting a beam of photons
8
For recent experiments on entanglement-based cryptography, see [160].
and have a correlation time of less than 100 ns [161]. The photons are de-
tected by silicon avalanche photodiodes, and time interval analyzers on local
personal computers register all detection events as time stamps together with
the settings of the analyzers and the detection results.
Quantum key distribution is started by a single light pulse sent from the
source to Alice and Bob via a second optical fiber. After a run of about 5 s
duration has been completed, Alice and Bob compare their lists of detec-
tions to extract the coincidences. In order to record the detection events very
accurately, the time bases in Alice’s and Bob’s time interval analyzers are
controlled by two rubidium oscillators. Overall, the system has a measured
rate of total coincidences of ∼ 1700 per second, and a collection efficiency of
each photon path of 5%. All the necessary equipment for the source, Alice and
Bob have been proven to operate outside shielded laboratory environments
with a very high reliability.
For the realization of entanglement-based quantum cryptography using
the Wigner inequality, Alice switches the analyzer randomly between −30◦
and 0◦ , and Bob between 0◦ and +30◦ . After a run, Alice and Bob ex-
tract from the coincidences the probabilities p++ (0◦ , 30◦ ), p++ (−30◦ , 0◦ ), and
p++ (−30◦ , 30◦ ) for the corresponding analyzer settings. We obtain −0.112 ±
0.014 for the left-hand side of the Wigner inequality (3.7), which is in good
agreement with the predictions of quantum mechanics, and the coincidences
obtained at the parallel settings, (0◦ , 0◦ ), can be used as a quantum key. In a
typical run, Alice and Bob established 2162 bits of raw quantum key material
at a rate of 420 baud, and observed a quantum bit error rate (QBER) of 3.4%.
By biasing the frequencies of the analyzer combinations, the production rate
of the quantum keys can be increased to about 1700 baud without sacrificing
security.
To demonstrate the entanglement-based BB84 scheme, Alice’s and Bob’s
analyzers both switched independently and randomly between 0◦ and 45◦ .
After a measurement run, Alice and Bob extracted the coincidences measured
with parallel analyzers to generate the quantum key. In the experiment, Alice
and Bob collected 80 000 bits of quantum key at a rate of 850 baud and
observed a quantum bit error rate of 2.5%. To correct the remaining errors
and ensure the secrecy of the key, various classical error correction and privacy
amplification schemes have been developed. With a very fast and efficient
algorithm, a single iteration gives 49 984 bits with a significantly reduced
QBER of 0.40% [113].
3.5.2 Quantum Dense Coding
For the first realization of this quantum communication scheme, the experi-
ment consisted of three distinct parts (Fig. 3.11): the EPR source, generating
entangled photons in a well-defined state; Alice’s station, for encoding the
messages by a unitary transformation of her particle; and Bob’s Bell-state
analyzer, for reading the signal sent by Alice.
Fig. 3.11. Experimental setup for quantum dense coding. The two entangled pho-
tons created by type II down-conversion are distributed to Alice and Bob. Alice
sends her photon, after manipulation with birefringent plates, to Bob, who can read
the encoded information by interferometric Bell-state analysis. The path length de-
lay ∆ is varied to achieve optimal interference
The polarization-entangled photons, with a wavelength of λ = 702 nm,

were, similarly to the quantum cryptography experiment, produced by de-
generate noncollinear type II down-conversion in a nonlinear BBO crystal
along two distinct emission directions (carefully selected by 2 mm irises,
1.5 m away from the crystal). One beam was directed to Alice’s encoding
station, the other directly to Bob’s Bell-state analyzer. The settings were
such that we obtained the entangled state |Ψ + behind the compensation
crystals (not shown in the figure) and Alice’s manipulation unit when the
retardation plates were both set to the vertical direction after compensation
of birefringence in the BBO crystal.
The beam manipulated in Alice’s encoding station was combined with
the other beam in Bob’s Bell-state analyzer. Bob’s analyzer consisted of a
single beam splitter followed by two-channel polarizers in each of its outputs,
and proper coincidence analysis between four single-photon detectors. In the
alignment procedure, optical trombones were employed to equalize the path
lengths to well within the coherence length of the down-converted photons
(∆λ = 100 µm), in order to observe the two-photon interference.
To characterize the interference observable at Bob’s Bell-state analyzer,
we varied the path length difference ∆ of the two beams with the optical
trombone. If the path length difference is larger than the coherence length,
no interference occurs and one obtains classical statistics for the coincidence
count rates at the detectors. With optimal path length tuning, interference
enables one to read the encoded information.
Fig. 3.12. Coincidence rates CHV (•) and CHV (◦) as functions of the path length
difference ∆, when the states |Ψ + (left) or |Ψ − (right) are analyzed by Bob’s
interferometric Bell-state analyzer
Figure 3.12 shows the dependence of the coincidence rates CHV (•) and
CHV (◦) on the path length difference, when either the state |Ψ + (left) or
the state |Ψ − has been sent to the Bell-state analyzer (the rates CH V and
CH V display analogous behavior; we use the notation CAB for the coincidence
rate between the detectors DA and DB ). For perfect path length tuning, CHV
reaches its maximum for |Ψ + (left) and vanishes (apart from noise) for |Ψ −
(right). CHV displays the opposite dependence and clearly signifies |Ψ − . The
results of these measurements imply that if both photons are detected, we
can identify the state |Ψ + with a reliability of 95%, and 93% for the state
|Ψ − .
The performance of the dense-coding transmission is influenced not only
by the quality of the alignment procedure, but also by the quality of the
states sent by Alice. In order to evaluate the latter, the beam splitter was
translated out of the beams. Then an Einstein–Podolsky–Rosen–Bell-type
correlation measurement analyzed the degree of entanglement of the source,
as well as the quality of Alice’s transformations. The correlations were only 1–
2% higher than the visibilities with the beam splitter in place, which means
that the quality of this experiment is limited more by the quality of the
entanglement of the two beams than by that of the interference achieved.
When using Si avalanche diodes in the Geiger mode for single-photon de-
tection, a modification of the Bell-state analyzer is necessary, since then, for
the states |Φ± , one has to register the two photons leaving the Bell-state
analyzer via a coincidence detection. One possibility is to avoid interference
at all for these states by introducing polarization-dependent delays before
Bob’s beam splitter. Another approach is to split the incoming two-photon
state at an additional beam splitter and to detect it (with 50% likelihood)
by a coincidence count between detectors in each output (inset of Fig. 3.13).
For the purpose of a proof-of-principle demonstration, we put such a con-
Fig. 3.13. Coincidence rates as functions of the path length difference ∆. Because
of the nature of the Si avalanche photodiodes, the extension shown in the inset is
necessary for identifying two-photon states in one output
figuration in place of detector DH only. Figure 3.13 shows the increase of

the coincidence rate CHH () for zero path length difference, with the other
rates at the background level, when Alice sends the state |Φ− . Since we can
now distinguish the three different messages, the stage is set for the quantum
dense-coding transmission. Figure 3.14 shows the various coincidence rates
(normalized to the corresponding maximum rate of the transmitted state),
when the ASCII codes of “KM◦ ” (i.e. codes 75, 77, 179) were sent in only 15
trits instead of 24 classical bits.
From this measurement, one can also obtain a signal-to-noise ratio by
comparing the rates signifying the actual state with the sum of the two other
rates registered. The ratios for the transmission of the three states varied
Fig. 3.14. “1.58 bits per photon” quantum dense coding: the ASCII codes for the
letters “KM◦ ” (i.e. 75, 77, 179) are encoded in 15 trits instead of the 24 bits usually
necessary. The data for each type of encoded state are normalized to the maximum
coincidence rate for that state
owing to the different visibilities of the corresponding interferences and were

about 14% and 9%. The signal-to-noise ratio achieved results in an actual
channel capacity of 1.13 bits per transmitted (and detected) two-state photon
and thus clearly exceeds the channel capacity of 1 bit achievable with noise-
free classical communication.
3.5.3 Quantum Teleportation of Arbitrary Qubit States
In this experiment, polarization-entangled photons were produced again by

type II down-conversion in a nonlinear BBO crystal (see Fig. 3.15), but here
the UV beam was pulsed to obtain a high time definition of the creation of
the pairs (the pulses had a duration of about 200 fs and λ = 394 nm). The
entangled pair of photons 2 and 3 is produced in the first passage of the
UV pulse through the nonlinear crystal, and the pair 1 and 4 after the pulse
has been reflected at a mirror back through the crystal. Mirrors and beam
splitters (BS) are used to steer and to overlap the light beams. Polarizers
(Pol) and polarizing beam splitters (PBS), together with birefringent retar-
dation plates (λ/2), prepare and analyze the polarization of the photons. All
single-photon detectors indicated in the figure (silicon avalanche photodiodes
operated in the Geiger mode) are equipped with narrow-band interference fil-
ters; the detectors of Alice’s Bell-state analyzer are equipped with additional
single-mode fiber couplers for spatial filtering.
For the first demonstration of quantum teleportation [10, 11], we prepared
particle 1 in various nonorthogonal polarization states using a polarizer and a
quarter-wave plate (not shown). Behind Bob’s “receiver”, polarization analy-
sis was performed to prove the dependence of the polarization of photon 3 on
the polarization of photon 1. (In this case we used the registration of photon
4 only to define the time of appearance of photon 1.)
The first task now is to prove that no information about the state of
photon 1 is revealed during the Bell-state measurement of Alice. Figure 3.16
shows the coincidence rate between detectors f1 and f2 when the overlap
of photons 1 and 2 at the beam splitter was varied (for this, we changed
the position of the mirror reflecting the pump beam back into the crystal).
The characteristic interference effect, a reduction of the coincidence rate,
occurs only around zero delay. Outside this region, which is on the order of
the coherence length of the detected photons, no reduction occurs, and the
two photons are detected in coincidence with 50% probability. Within the
statistics, there is no difference between the two data sets, although particle 1
was prepared in two mutually orthogonal states (+45◦ and −45◦). Obviously,
Alice has no means to determine which of the two states particle 1 was in
after the projection into the Bell-state basis.
Figure 3.17 shows the polarization of photon 3 after the teleportation
protocol has been performed, again as the delay between photons 1 and 2 is
varied. When interference occurs at the beam splitter, i.e. around zero delay,
the polarization of photon 3 is given by the settings for photon 1. The two
Fig. 3.15. Experimental setup for quantum teleportation. A UV pulse passing

through a nonlinear crystal creates an entangled pair of photons 2 and 3 in the state
|Ψ − , which is distributed to Alice and Bob. During its second passage through the
crystal, after retroflection the UV pulse creates another pair of photons, of which
one is prepared in the initial state to be teleported (photon 1), and the other one
(4) serves as a trigger indicating that a photon to be teleported is on its way. Alice
then looks for coincidences behind her beam splitter, where the initial photon and
one of the ancillaries are superposed. Bob, after receiving the classical information
that Alice has obtained a coincidence count identifying the |Ψ − Bell state, knows
that his photon 3 is in the initial state of photon 1, which then can be verified using
polarization analysis
Fig. 3.16. Coincidence rate between the two detectors of Alice’s Bell-state analyzer
as a function of the delay between the two photons 1 and 2. The data for the +45◦
and −45◦ polarizations of photon 1 are equal within the statistics, which shows
that no information about the state of photon 1 is revealed to Alice
graphs show the results obtained when the initial polarization of photon 1
was set either to 45◦ or to vertical polarization and then the polarization of
photon 3 along the corresponding direction was analyzed. The reduction in
the polarization to about 65% is due to the limited degree of entanglement
between photons 2 and 3 (85%), and to the reduced contrast of the interfer-
ence at the beam splitter as a consequence of the relatively short coherence
time of the detected photons. Of course, better beam definition by narrow
pinholes and more stringent filtering could improve this value. However, this
would cause further, unacceptable loss in the fourfold coincidence rates. Each
of the polarization data points shown was obtained from about 100 four-fold
coincidence counts in 4000 s.
Fig. 3.17. Polarization of photon 3 after teleportation, compared with the po-
larization initially prepared on photon 1. The analyzer testing the quality of the
teleportation performed by Alice and Bob was oriented parallel to the initial po-
larization
These measurements and also runs with the initial polarization along
other directions demonstrate the ability to teleport the polarization of any
pure state. Of course, since the directions used are mutually nonorthogonal,
one can infer that the scheme works for any arbitrary quantum state. How-
ever, there is a much more direct way to experimentally demonstrate the full
power of quantum teleportation.
One way to demonstrate that any arbitrary quantum state can be trans-
ferred is to use the fact that we can also obtain entanglement between photons
1 and 4 (Fig. 3.18). After the polarizer was removed from arm 1 and put into
arm 4, the state of 1 was not defined anymore, but still could be teleported
to photon 3; this was demonstrated by showing that now the entanglement
had been swapped to photons 3 and 4.
The state of photon 1 (Fig. 3.18), which is part of an entangled pair (pho-
tons 1 and 4), is fully undetermined and is formally described by a mixed
state. If one can teleport this state to another photon, i.e. to Bob’s photon 3,
we expect to find this photon in a mixed state, that means it is unpolarized.
Fig. 3.18. Experimental setup that demonstrates teleportation of arbitrary quan-

tum states: by teleporting the as yet undefined state of photon 1 to photon 3, one
is able to swap the entanglement, initially between particles 1 and 4 and between
particles 2 and 3, to the newly entangled pair of particles 3 and 4 by projecting 1
and 2 into an entangled pair
Now, since Bob’s photon was originally also part of an entangled pair (pho-
tons 2 and 3), it was unpolarized anyway. One might conclude that here we
did not achieve anything. However, if one determines not only the polariza-
tion of photon 3 but the correlations between photons 3 and 4, one finds that
now these two photons, which have been produced independently by different
processes, are entangled [115].
Figure 3.19 verifies the entanglement between photons 3 and 4, condi-
tioned on coincidence detection of photons 2 and 3. Varying the angle Θ of
the polarizer in arm 4 causes a sinusoidal variation of the count rate, here
with the analyzer of photon 3 set to ±45◦ . This shows that we did not tele-
port just a mixed state, but actually the as yet undetermined state of the
entangled photon.
These experiments present the first demonstration of quantum teleporta-
tion, that is, the transfer of a qubit from one two-state particle to another.
In the meantime, further steps have been achieved, in particular the remote
preparation of the state of Bob’s photon (sometimes also called “telepor-
tation”) [10] and, especially important, the teleportation of the state of an
electro-magnetic field [11]. The latter is the first example of teleportation of
continuous variables based on the original EPR entanglement. The first ex-
periment demonstrated the feasibility of transfer of fluctuations of a coherent
state from one light beam to another. Although the experiment was limited
to a narrow bandwidth of 100 kHz, this was only a technical limitation due
to the detection electronics, the modulators and the bandwidth of the source
Fig. 3.19. Verification of the entanglement between photons 3 and 4. The sinusoidal
dependence of the fourfold coincidence rate on the orientation Θ of the polarizer
in arm 4 for ±45◦ polarization analysis of photon 3 demonstrates the possibility to
teleport any arbitrary quantum state
of EPR-entangled light beams. In principle, it soon should be possible also to

transfer nonclassical states of light, such as squeezed light or number states.
3.6 Outlook
Quantum communication with entangled photons has shown its power and
its fascinating features. Our experiments, where realistic entanglement-based
quantum cryptography has been performed, where the capacity of commu-
nication channels has been increased beyond classical limits and where the
polarization state of a photon has been transferred to another one by means
of quantum teleportation, are only the first steps towards the exploitation of
new resources for communication and information processing.
Quantum communication can offer a wealth of further possibilities, es-
pecially when combined with simple quantum logic circuitry. Quantum com-
puters have to operate on large numbers of qubits to really demonstrate their
power. But quantum communication schemes already profit from combining
only a few qubits and entangled systems. Quantum logic operations with sev-
eral particles are already useful in examples of the quantum coding theorem
[69], but have shown their importance particularly in the proposals for entan-
glement purification [119]. Any realistic transmission of quantum states will
suffer from noise and decoherence along the line. If one wants to distribute
entangled pairs of particles to, say, Alice and Bob, the entanglement between
the received particles will be considerably degraded, which would prevent
successful quantum teleportation, for example. If Alice and Bob now com-
bine the particles of several such noisy pairs on each side by quantum logic
operations, they can improve the quality of entanglement by the proposed

“distillation” process.
These ideas are closely related to quantum error correction for quantum
computers and were recently implemented in a proposal for efficient distri-
bution of entanglement via so-called quantum repeaters [162]. One day, such
systems might form the core of quantum networks [163] allowing quantum
communication and even computation over large distances. Of course, one
should always keep in mind the obstacles put in the way by the decoherence
of quantum states [164]. However, quantum communication schemes should
be significantly more stable owing to the much lower number of quantum
systems involved.
Once entangled particles have been distributed, various quantum com-
munication protocols could be implemented. Besides those described in the
preceding sections, there are some recent proposals that give a new twist
to quantum information processing. Quantum gambling [165] and quantum
games [166], e.g. a “quantized” version of the prisoner dilemma, bring the
field of game theory to the quantum world and demonstrate new strategies
in well-known classical games. But the new ideas and thoughts might also
be quite useful for other types of communication problems. For example, the
quantum version of “Chinese whispers” [167] can be also seen as a special
type of error correction scheme. Errors in the classical communication, the
whispering, can be more efficiently corrected if the sender and receiver have
been provided with entangled pairs of particles.
New possibilities arise if entangled triples of particles are used. For cer-
tain tasks, the communication between three or more parties becomes less
complex, and thus more efficient, if the parties share the entanglement ini-
tially [168], and also schemes for quantum cloning [169] of the state of a qubit
become feasible with entangled triples.
Now that significant improvements of down-conversion sources [145, 148,
149] and the first observation of three-particle entanglement [38, 170] have
been achieved, the realization with entangled triples of those schemes that
have previously used entangled pairs is within the reach of future experiments.
For realizing entanglement purification and similar schemes, the experiments
immediately become much more complex. It first has to be seen what methods
can be used to perform quantum logic operations with photons, and also what
types of photon sources should be used then. However, the combined progress
in the form of improving experimental techniques and of better understanding
of the principles of quantum information theory makes the more complicated
schemes feasible. Most likely, there will also be novel schemes for quantum
communication using higher numbers of qubits and/or even more complex
types of entanglement.
Quantum cryptography was the first quantum communication method to
literally leave the shielded environment of quantum physics laboratories and
to become a promising candidate for commercial exploitation. We expect
that the future will show an enormous potential for and benefit from the
use of other quantum communication methods, such as the distribution of
entanglement over large distances and the transfer of quantum information
in the process of quantum teleportation.
4 Quantum Algorithms: Applicable Algebra
and Quantum Physics
Thomas Beth and Martin Rötteler
4.1 Introduction
Classical computer science relies on the concept of Turing machines as a

unifying model of universal computation. According to the modern Church–
Turing Thesis, this concept is interpreted in the form that every physically
reasonable model of computation can be efficiently simulated on a proba-
bilistic Turing machine. Recently this understanding, which was taken for
granted for a long time, has required a severe reorientation because of the
emergence of new computers that do not rely on classical physics but, rather,
use effects predicted by quantum mechanics.
It has been realized that, by using the principles of quantum mechanics,
there are problems for which a putative quantum computer could outperform
any classical computer. Quantum algorithms benefit from the application of
the superposition principle to the internal states of the quantum computer,
which are considered to be states in a (finite-dimensional) Hilbert space. As
a result, these algorithms lead to a new theory of computation and might be
of central importance to physics and computer science. Striking examples of
quantum algorithms are Shor’s factoring algorithm, Grover’s search algorithm
and algorithms for quantum error-correcting codes, all of which will be part
of this contribution.
We shall introduce the complexity model of quantum gates, which are
most familiar to researchers in the field of quantum computing, and shall give
many examples of the usefulness and conciseness of this formalism. Quantum
circuits provide a computational model equivalent to quantum Turing ma-
chines. This means that, very much like the situation in classical computing,
there are several ways of describing computations by appropriate theoretical
models.
Amongst such quantum circuits, quantum signal transforms form basic
primitives in the treatment of controlled quantum systems. A surprising and
important result, in view of the algorithms of Shor, is the fact that it is
possible to compute a Fourier transform (of size 2n ) on a quantum computer
by means of a quantum circuit which requires only O(n2 ) basic operations.
This is a substantial speedup compared to the classical case, where the fast
Fourier transform [171] yields an algorithm that requires O(n2n ) arithmetic

4 Quantum Algorithms: Applicable Algebra and Quantum Physics 97
operations. Applications of such Fourier transforms to finite abelian groups

arise in the algorithms of Simon and Shor. We shall present these algorithms
and the underlying principle leading to their surprisingly fast solution on a
quantum computer.
As already mentioned, one of the basic results is that in the complexity
model of quantum circuits, the Fourier transform can be realized with an
exponential speedup compared with the classical case. However, in the quan-
tum regime the only way to extract information from a system is to make
measurements and thereby project out nearly all aspects of the whole system.
Thus, the art and science of designing quantum algorithms lies in the ability
to obtain enough information from measurements, i.e. to choose the right
bases from which relevant information can be read off. On the basis of the
example of the Fourier observable, which represents the most important case
of such a base change, we explain the underlying principle by means of the
so-called hidden-subgroup algorithms and present an analysis of sampling in
the Fourier basis with respect to the appropriate groups.
We then show how recent results in the theory of signal processing (for a
classical computer) can be applied to obtain fast quantum algorithms for var-
ious discrete signal transforms, including Fourier transforms for nonabelian
groups. Finally, we give a brief introduction to the theory of (quantum) error-
correcting codes and their algorithmic implementation.
4.2 Architectures and Machine Models
The definition of an architecture and a machine model, on which the compu-

tations are considered to be carried out, is indispensable if one is to have a
common computational model for which algorithms can be devised.
Each reasonable model of computation should give us the possibility of
performing arbitrary operations, up to a desired accuracy, on the system by
execution of elementary operations. By counting the elementary operations
necessary to complete a given task, we arrive at complexity models. Finally,
if different approaches defining universal computational models are possible,
it is desirable to show the equivalence of these models, in the sense that they
can simulate each other with a slowdown that is polynomial in the size of the
input. In the case of quantum computing, we give two models for universal
quantum computation, namely quantum networks in Sect. 4.2.1 and quantum
Turing machines in Sect. 4.2.6. We shall put more emphasis on gates and
networks, relying on the result that these two models are equivalent in the
sense described.
One remark concerning the architecture is in order: we restrict ourselves
to the case of operational spaces with a dimension that is a power of two,
which are called qubit architectures. These systems incorporate the features
necessary to do quantum computing, i.e. superposition of an exponentially
98 Thomas Beth and Martin Rötteler
.. ..
. .
... ... • j
U (i) = CNOT(i,j) = ... ...
U i ❢
.. .. i
. . .. ..
. .
Fig. 4.1. Elementary quantum gates
growing number of states, interference between computational paths and en-

tanglement between quantum registers.
It is possible to perform an embedding of an arbitrary finite-dimensional
operational space into a qubit architecture (see Sect. 4.2.3); however, this
reduction involves a suitable encoding of the states of the system into the
basis states of the qubit architecture, and hence genuine properties of the
system might be lost by this procedure.
4.2.1 Quantum Networks
The state of a quantum computer is given by a normalized vector in a Hilbert

space H2n of dimension 2n , which is endowed with a natural tensor structure
H2n ∼= C 2 ⊗. . .⊗C 2 (n factors). The standard basis for this Hilbert space is the
set {|x : x ∈ Zn2 } of binary strings of length n. Restricting the computational
space to Hilbert spaces of this particular form is motivated by the idea of a
quantum register consisting of n quantum bits. A quantum bit, also called a
qubit, is a state corresponding to one tensor component of H2n and has the
form
|ϕ = α|0 + β|1 , α, β ∈ C , |α|2 + |β|2 = 1 .
The possible operations that this computer can perform are the elements
of the unitary group U(2n ). To study the complexity of performing unitary
operations on n-qubit quantum systems, we introduce the following two types
of computational primitives: local unitary operations on a qubit i are matrices
of the form U (i) = 12i−1 ⊗ U ⊗ 12n−i , where U is an element of the unitary
group U(2) of 2 × 2 matrices and 1N denotes the identity matrix of size
N . Furthermore, we need operations which affect two qubits at a time, the
most prominent of which is a so-called controlled NOT gate (also called a
measurement gate) between the qubits j (control) and i (target), denoted by
CNOT(i,j) . On the basis vectors |xn , . . . , x1 of H2n , the operation CNOT(i,j)
is defined by
|xn , . . . , xi+1 , xi , xi−1 , . . . , x1 → |xn , . . . , xi+1 , xi ⊕ xj , xi−1 , . . . , x1 ,
where the addition ⊕ is performed in Z2 .

In graphical notation, using quantum wires, these transformations are

written as shown in Fig. 4.1. Lines correspond to qubits, unaffected qubits
are omitted and a dot • sitting on a wire denotes a control bit. Note that
we draw the qubits according to their significance, starting with the most
significant qubit on top. Quantum circuits are always read from left to right.
The two types of gates shown in Fig. 4.1 suffice to generate all unitary
transformations, i.e. they form a universal set of gates. This is the content of
the following theorem [172].
Theorem 4.1. G1 := {U (i) , CNOT(i,j) | U ∈ U(2), i, j ∈ {1, . . . , n}, i = j}

is a generating set for the unitary group U(2n ).
This means that for each U ∈ U(2n ) there is a word w1 w2 . . . wk (where

wi ∈ G1 for i = 1, . . . , k is an elementary gate) such that U factorizes as
U = w1 w2 . . . wk . On the basis of theorem 4.1, we now define a complexity
measure for unitary operations on qubit architectures.
Definition 4.1. Let U ∈ U(2n ) be a given unitary transformation. Then

κ(U ) is defined as the minimal number k of operations in G1 necessary to
write U = w1 w2 . . . wk as a sequence of elementary gates.
For the complexity measure κ, the following holds: κ(A ⊗ B) ≤ κ(A) +

κ(B) for all A ∈ U(2n1 ) and B ∈ U(2n2 ), because tensor products are free
of cost in a computational model based on quantum mechanical principles.
Also, by concatenation of operations, we obtain κ(A · B) ≤ κ(A) + κ(B) for
A, B ∈ U(2n ). Note that whereas in the usual linear complexity measure Lc
[173, 174] permutation matrices are free (i.e. Lc (π) = 0, for all π ∈ Sn ),
we have to take them into account when using the complexity measure κ.
Instead of the universal set of gates G1 we can, alternatively, use the set
G2 := {U (i,j) : U ∈ U(4), i, j ∈ {1, . . . , n}, i = j} of all two-bit gates,
changing the value of κ by only a constant.
Whereas the complexity measure κ is used in cases where we want to
implement a given unitary operation U exactly in terms of the generating
sets G1 and G2 , it is also expedient to consider unitary approximations by
quantum networks. By this we mean a sequence of operations w1 , . . . , wn
which approximates U up to a given , i.e. such that U − ni=1 wi < , where
· denotes the spectral norm.1 We denote the corresponding complexity
measure by κ .
Remark 4.1. The following facts concerning approximation by elementary

gates are known:
• There are two-bit gates which are universal [175, 176], i.e. there exists a
unitary transformation A ∈ U(4) with respect to which it is possible to
approximate any given U up to > 0 by a sequence of applications of A
1
Recall that the spectral norm of a matrix A ∈ C n×n is given by maxλ∈Spec(A) |λ|.
k
to two tensor components of H2n only: U − i=1 A(mi ,ni ) < . Even
though the much stronger statement of universality of a generic two-bit
gate is known to be true, it is hard to prove the universality for a given
two-bit gate [176, 177].
• Small generating sets are known; for instance, we can choose
1
1 0 √

1 1
G3 := √ , , CNOT(k,l) , k = l ,
2 1 −1 0 (1 + i) 2
where the first two operations generate a dense subgroup in U(2). The
Hadamard transformation, which is part of this generating set, is denoted
by
1 1 1
H2 := √
2 1 −1
and is an example of a Fourier transformation on the abelian group Z2
(see Sect. 4.4).
• Knill has obtained a general upper bound O(n4n ) for the approximation
of unitary matrices using a counting argument [178, 179].
• We cite the following approximation result from Sect. 4.2 of [180]: Fix
a number n of qubits and suppose that X1 , . . . , Xr = SU(2n ), i.e.
X1 , . . . , Xr generate a dense subgroup in the special unitary group SU(2n ).
Then it is possible to approximate a given matrix U ∈ SU (2n ) with
given accuracy > 0 by a product of length O{poly[log(1/)]}, where the
factors belong to the set {X1 , . . . , Xr , X1−1 , . . . , Xr−1 }. Furthermore, this
approximation is constructive and efficient, since there is an algorithm
with running time O{poly[log(1/)]} which computes the approximating
product. However, we remind the reader that this holds only for a fixed
value of n; the constant hidden in the O-calculus grows exponentially
with n (see Theorem 4.8 of [180]).
From now on, we put the main emphasis on the model for realizing uni-
tary transformations exactly and on the associated complexity measure κ.
In general, only exponential upper bounds for the minimal length occuring
in factorizations are known. However, there are many interesting classes of
unitary matrices in U(2n ) that lead to only a polylogarithmic word length,
which means that the length of a minimal factorization grows asymptotically
like O[p(n)], where p is a polynomial.
In the following we give some examples of transformations, their factor-
ization into elementary gates and their graphical representation in terms of
quantum gate arrays. The operations considered in these examples admit
short factorizations and will be useful in the subsequent parts of this chap-
ter.
Example 4.1 (Permutation of Qubits). The symmetric group Sn is embedded

in U(2n ) by the natural operation of Sn on the tensor components (qubits).
Let τ ∈ Sn and let Πτ be the corresponding permutation matrix on 2n

points. Then κ(Πτ ) = O(n): to prove this, we first note that each element
σ ∈ Sn can be written as a product σ = τ1 · τ2 of two involutions τ1 and
τ2 ∈ Sn , i.e. τ12 = τ22 = id. To see that it is always possible to find a suitable
τ1 and τ2 , we can, when considering the decomposition of σ into disjoint
cycles, restrict ourselves to the case of an n-cycle. Now the decomposition
follows immediately from the fact that there is a dihedral group of size 2n
containing σ as the canonical n-cycle and that this rotation is the product of
two reflections.
The unitary transformation corresponding to Πτ , where τ ∈ Sn is an
involution, can be realized by swappings of quantum wires, which, in turn,
can be performed efficiently and in parallel. To swap two quantum wires we
can use the well-known identity Π(1,2) = CNOT(1,2) · CNOT(2,1) · CNOT(1,2) ,
yielding a circuit of depth three. Writing an arbitrary permutation Πτ of the
qubits as a product of two involutions, we therefore obtain a realization by a
circuit of depth six at most (see also [181]).
As an example, the permutation (1, 3, 2) of the qubits (which corresponds
to the permutation (1, 4, 2)(3, 5, 6) on the register) is factored as (1, 3, 2) =
(1, 2)(2, 3) (see Fig. 4.2).
• ❣ •
❅✁✁
❅ = ❣ ❣ ❣
❅✁ • • •
✁❅ ❣ ❣
•
Fig. 4.2. Factorization (1, 3, 2) = (1, 2)(2, 3)
Example 4.2 (Controlled Operations). Following [172], we introduce a special

class of quantum gates with multiple control qubits, yielding a natural gen-
eralization of the controlled NOT gate. This class of gates is given by the
transformations Λk (U ), where U is a unitary transformation in U(2l ). The
gate Λk (U ) is a transformation acting on k + l qubits, where the k most
significant bits serve as control bits and the l least significant bits are target
bits: the operation U is applied to the l target bits if and only if all k control
bits are equal to 1. Denoting by M := 2l (2k −1) the number of basis vectors
on which Λk (U ) acts trivially, the corresponding unitary matrix is given by
1M ⊕ U , where we have used ⊕ to denote a direct sum of matrices.
To provide further examples of the graphical notation for quantum cir-
cuits, we give in Fig. 4.3 a Λ1 (U ) gate for U ∈ U(2n ) with a normal control
qubit, a gate Λ1 (U ) with an inverted control qubit, and the matrices repre-
sented. Lemmas 7.2 and 7.5 of [172] show that for U ∈ U(2), the gate Λk (U )
can be realized with gate complexity O(n), for k < n−1. If there are auxiliary
qubits (so-called ancillae) available, a gate Λn−1 (U ) can also be computed
using O(n) operations from G1 ; otherwise, we have κ[Λn−1 (U )] = O(n2 ).
r n+1 ❜ n+1
Λ1 (U ) n Λ1 (U ) n
= 12n ⊕ U = ... U ... ..
. = U ⊕ 12n = ... U ... ..
.
1 1
Fig. 4.3. Controlled gates with (left) normal and (right) inverted control bit. Here
⊕ is used to denote the block-direct sum of matrices
We remark that, if U ∈ U(2n ) can be realized in p elementary operations

then, Λ1 (U ) ∈ U(2n+1 ) can be realized in c×p basic operations, where c ∈ N is
a constant that does not depend on U . To see this, we first assume, without
loss of generality, that U is decomposed into elementary gates. Therefore,
we have to show that a doubly controlled NOT (also called a Toffoli gate,
see Sect. 4.2.3) and a singly controlled U ∈ U(2) gate can be realized with a
constant increase of length. It is possible to obtain the bound c ≤ 17 according
to the following decompositions [172]: for each unitary transformation U ∈
U(2) we can write Λ1 (U ) = A(1) · CNOT(1,2) · B (1) · CNOT(1,2) · C (1) with
suitably chosen A, B, C ∈ U(2), i.e. we need at most five elementary gates
for the realization of Λ1 (U ). To bound the number of operations necessary
to realize τ := Λ2 (σx ) with respect to the set G1 , we choose a square root R
of σx , i.e. R2 = σx , and use the identity
τ = [12 ⊗ Λ1 (R)] · CNOT(2,3) · [12 ⊗ Λ1 (R† )] · CNOT(2,3) · Λ1 (12 ⊗ R).
This shows that we need at most 5 + 1 + 5 + 1 + 5 = 17 elementary gates to

realize τ .
Example 4.3 (Cyclic Shift). Let Pn ∈ S2n be the cyclic shift acting on the
states of the quantum register as x → x + 1 mod 2n . The corresponding
permutation matrix is the 2n -cycle (0, 1, . . . , 2n −1). The unitary matrix Pn
can be realized in a polylogarithmic number of operations; see Fig. 4.4 for a
realization using Boolean gates only.
❣ ···
r ❣ ···
r r ❣ ···
..
.
r r r ··· ❣
r r r ··· r ❣
Fig. 4.4. Realizing a cyclic shift on a quantum register

Other, non-Boolean factorizations are also possible: using a basic fact

about group circulants (see also Sect. 4.5.1) and anticipating the fact that
the discrete Fourier transform DFT2n can be performed in O(n2 ) operations
(which is shown in Sect. 4.4.1), we can use the identity

DFT−1 i n
2n · Pn · DFT2n = diag(ω2n : i = 0, . . . , 2 − 1)
n−1
= diag(1, ω22n ) ⊗ · · · ⊗ diag(1, ω2n )
to obtain κ(Pn ) = O(n ). 2
4.2.2 Boolean Functions and the Ring Normal Form
Boolean functions are important primitives used throughout classical infor-

matics. Denoting the finite field with q elements2 by GF (q), we obtain the
Boolean numbers as the special case q = 2. A multivariate Boolean function
f : GF (2)n → GF (2) can be represented in various ways. Besides the truth
table, which is a common but uneconomic way to represent f as the sequence
of its values f (0 . . . 0), . . . , f (1 . . . 1) for all binary strings of length n, promi-
nent examples of normal forms are the conjunctive and disjunctive normal
forms [183] which originate from predicate logic and are used in transistor
circuitry.
For quantum computational purposes, another way of representing f offers
itself, namely the ring normal form (RNF), defined as the (unique) expansion
of f as a polynomial in the ring Rn of Boolean functions of n variables. This
ring is defined by Rn := GF (2)[X1 , . . . , Xn ]/(X12 − X1 , . . . , Xn2 − Xn ) [184].
Multiplication and addition in Rn are the usual multiplication and addition of
polynomials modulo the relations given by the ideal (X12 − X1 , . . . , Xn2 − Xn ),
and addition is usually denoted as “⊕”. Therefore f is represented as
n

f (X1 , . . . , Xn ) := au Xiui , (4.1)
u=(u1 ,...,un )∈{0,1}n i=1
with coefficients au ∈ GF (2).

Example 4.4. The logical complement is given by NOT(X) = 1 ⊕ X. The
n of the AND function on n variables X1 , . . . , Xn is AND(X1 , . . . , Xn ) =
RNF
i=1 Xi . The RNF of thePARITY function of n variables is given by
n
PARITY(X1 , . . . , Xn ) = i=1 Xi . Finally, the RNF of the OR function
on n variables is given by
n

OR(X1 , . . . , Xn ) = 1 ⊕ (1 ⊕ Xi ) = m(X1 , . . . , Xn ) ,
i=1 m
2
Necessarily, we have q = pn , where p is a prime and n ≥ 1 [182]. Finite fields are
also called Galois fields after Évariste Galois (1811–1832).
where the last sum runs over all nonconstant multilinear monomials m in the
ring Rn .
We can implement a Boolean function given in the RNF shown in (4.1)

with the use of the so-called Toffoli gate τ . The action of τ on the basis states
of the Hilbert space H8 is given by τ : |x|y|z → |x|y|z ⊕ x · y [185].
Horner’s rule for multivariate polynomials yields a method for implement-
ing the function f given in (4.1). To achieve this, we write f (X1 , . . . , Xn ) =
f1 (X1 , . . . , Xn−1 ) · Xn ⊕ f2 (X1 , . . . , Xn−1 ) and observe that this function can
be computed using one Toffoli gate, assuming that f1 and f2 have already
been computed. Therefore, we obtain a recursive factorization for f , which
in general will make use of auxiliary qubits.
4.2.3 Embedded Transforms
This section deals with the issue of embedding a given transform A into a
unitary matrix of larger size. We start by considering the problem of realizing
a given matrix A as a submatrix of a unitary matrix of larger size. The
following theorem (see also [186]) shows that the only condition A has to
fulfill in order to allow an embedding involving one additional qubit is to be
of bounded norm, i.e. A ≤ 1, with respect to the spectral norm.
Theorem 4.2. Let A ∈ C n×n be a given matrix of norm A ≤ 1. Then

A (1n − AA† )1/2
UA := (4.2)
(1n − A† A)1/2 −A†
yields a unitary matrix UA ∈ U(2n) which contains A as the n × n submatrix

in the upper left corner.
Proof. Observe that the n × 2n matrix U1 := (A, (1n − A† A)1/2 )t has the
property U1† · U1 = 1n . Analogously, for the matrix U2 := [A, (1n − AA† )1/2 ],
the identity U2 · U2† = 1n holds. An easy computation shows that (4.2) is
indeed unitary.
Since each matrix in C n×n can be renormalized by multiplication with a
suitable scalar to fulfill the requirement of a bounded norm, we can realize
all operations up to a scalar prefactor by unitary embeddings. The embed-
ding (4.2) is by no means unique. However, it is possible to parametrize all
embeddings by (1n ⊕ V1 ) · UA · (1n ⊕ V2 ), where V1 , V2 ∈ U(n) are arbitrary
unitary transforms.
We are naturally led to a different kind of embedding if the given trans-
formation is unitary and we want to realize it on a qubit architecture, i.e. if
we restrict ourselves to matrices whose size is a power of 2. Then, a given
unitary matrix U ∈ U(N ) can be embedded into a unitary matrix in U(2n ) by
choosing n = log N and padding U with an identity matrix 12n −N of size
2n − N . This is noncanonical since we have degrees of freedom in the choice

of the subspace on which this newly formed matrix acts as the identity.
In general, it is a difficult problem to find the optimal embedding for a
given transform. An example in which it is not natural to go to the next
power of 2 is given by U = U1 ⊗ U2 ∈ U(15), where U1 ∈ U(3) and U2 ∈
U(5). We then have the possibilities U ⊕ 11 ∈ U(24 ) and the embedding
(U1 ⊕ 11 ) ⊗ (U2 ⊕ 13 ) ∈ U(25 ), which respects the tensor decomposition of U .
A third type of embedding occurs in the context of quantum and reversible
computing, where a general method is required to make a given map f :
X → Y bijective (here X and Y are finite sets). If we consider the map
f˜ : X × Y → X × Y which maps (x, y0 ) → (x, f (x)), where y0 is a fixed
element in the codomain Y of f , then this map is obviously injective when
restricted to the fibre X × {y0 }. Observe now that it is always possible to
extend f˜|X×{y0 } to a unitary operation on the Hilbert space HX,Y spanned
by the basis consisting of {|x|y : x ∈ X, y ∈ Y }. Hence, it is always possible
to construct a unitary operation Vf : HX,Y → HX,Y which has the property
|x|0 → |x|f (x) , for all x ∈ X , (4.3)
i.e. Vf implements the graph Γf = {(x, f (x)) : x ∈ X} of f in the Hilbert

space HX,Y . Here, we have identified the special element y0 with the basis
vector |0 ∈ HY .
Example 4.5. Let f : GF (2)2 → GF (2) be the AND function, i.e. let f = x·y
be the RNF of f . Note that f˜ can be chosen to be the function (x, y, z) →
(x, y, z ⊕ f (x, y)) since the codomain is endowed with a group structure.
Overall, we obtain the function table of f˜ given in Fig. 4.5. The variables
with a prime correspond to the values after the transformation has been
performed.
x y z x y z x y z x y z |x • |x
0 0 0 0 0 0 1 0 0 1 0 0
0 0 1 0 0 1 1 0 1 1 0 1 |y • |y
0 1 0 0 1 0 1 1 0 1 1 1
|z ❢ |z ⊕ x · y
0 1 1 0 1 1 1 1 1 1 1 0
Fig. 4.5. Truth table for the Toffoli gate and the corresponding quantum circuit
We recognize f˜ as the unitary operation τ : |x|y|z → |x|y|z ⊕ x · y on

the Hilbert space H8 , which is the Toffoli gate [185].
The method described in Example 4.5 is quite general, as the following

theorem shows (for a proof see [187]).
Theorem 4.3. Suppose f : {0, 1}n → {0, 1}m is a Boolean function which
can be computed using c operations from the universal set {AND, NOT}
of classical gates. Then f˜ : {0, 1}n+m → {0, 1}n+m, defined by (x, y) →
(x, y ⊕ f (x)), is a reversible Boolean function which can be computed by a
circuit of length 2c + m built up from the set {CNOT, τ } of reversible gates.
Even though the construction described in Theorem 4.3 works for arbi-
trary f : {0, 1}n → {0, 1}m, in general only rf = log maxy∈{0,1}m |f −1 (y)|
additional bits are necessary to define a reversible Boolean function frev :
{0, 1}n+rf → {0, 1}n+rf with the property frev |{0,1}n = f . The reason is that
by using the additional rf bits, the preimages of f can be separated via a
suitable binary encoding. However, the complexity of a Boolean circuit of a
realization of frev constructed in such a way is such that the circuit cannot
be controlled as easily as for the function defined in Theorem 4.3.
4.2.4 Permutations
We have already mentioned in Sect. 4.2.1 that on a quantum computer per-

mutations of the basis states have to be taken into account when considering
the complexity: in general, for the cost κ(π), where π is a permutation matrix
in U(2n ) and κ is the complexity measure introduced in Sect. 4.2.1, nothing
better is known than an exponential upper bound of O(n4n ).
Nevertheless, there are quite a few classes of permutations admitting a
better, even polylogarithmic word length, as the examples of permutations of
quantum wires and of the cyclic shift Pn : x → x + 1 mod 2n on a quantum
register have shown (see Examples 4.1 and 4.3).
In what follows we consider a further class of permutations of a quantum
register that admits efficient realizations, which operate by linear transfor-
mations on the names of the kets. Recall that the basis states can be iden-
tified with the binary words of length n and hence, with the elements of
GF (2)n , the n-dimensional vector space over the finite field GF (2) of two el-
ements. Denoting the group of invertible linear transformations of GF (2)n by
GL(n, GF (2)), we see that each transform A ∈ GL(n, GF (2)) corresponds to
a permutation of the binary words of length n and, hence, to a permutation
matrix ΦA of size 2n × 2n .
It turns out that these permutations are efficiently realizable on a quan-
tum computer (see Sect. 4 of [188]). First we need the following lemma.
Lemma 4.1. Let K be a field and let A ∈ GL(n, K) be an invertible matrix

with entries in K. Then there exist a permutation matrix P , a lower trian-
gular matrix L and an upper triangular matrix U such that A = P · L · U .
In numerical mathematics this decomposition is also known as the “LU

decomposition” (see, e.g., Sect. 3.2. of [189]). The statement is a consequence
of Gauss’s algorithm. We are now ready to prove the following theorem.
Theorem 4.4. Given A ∈ GL(n, GF (2)), ΦA can be realized in O(n2 ) ele-

mentary operations on a quantum computer.
Proof. First, decompose A according to Lemma 4.1 into A = P · L · U and
observe that the permutation matrix P is a permutation of the quantum
wires, and hence κ(P ) = O(n) (see Example 4.1). The matrices L and U can
be realized using CNOT gates only. Without loss of generality, we consider the
factorization of L: proceeding along the diagonals of these matrices, we find
all diagonal entries to be 1 (otherwise the matrices would not be invertible).
Therefore ΦL maps the basis vector |ei , where ei = (0 . . . 1 . . . 0) is the ith
basis vector in the standard basis of GF (2)n , to the sum j≥i αj ej , where
the vector (αi )i=1,...,n is the ith column of A. Application of the sequence

j≥i CNOT
(i,j)
, where the product runs over all j = 0, has the same effect
on the basis vector ei .
Proceeding column by column in L yields a factorization into O(n2 ) el-
ementary gates. Combining the factorizations for P , L and U , we obtain
κ(A) = O(n2 ).
As an example, we take a look at the matrix

10
A= ∈ GL(2, GF (2)) .
11
To see what the corresponding ΦA looks like, we compute the effect of A on
the basis vectors:

0 0 0 0 1 1 1 1
→ , → , → , → ,
0 0 1 1 0 1 1 0
i.e. ΦA = CNOT(2,1) in accordance with Theorem 4.4, since A is already

lower triangular.
As further examples of permutations arising as unitary transforms on a
quantum computer, we mention gates for modular arithmetic [190, 191]. More
specifically, we consider the following operation, acting on kets which have
been endowed with the group structure of ZN := Z/N Z, i.e. {|x : x ∈ ZN }
is a basis of this operational space H. Then
Υa : |x|0 → |x|a · x mod N ,
where a ∈ Z× N is an element of the multiplicative group of units in ZN ;
this can be extended to a permutation of the whole space H ⊗ H using the
methods of Sect. 4.2.3. Using a number of ancilla qubits which is polynomial
in log[dim(H)], it is possible to realize Υa , as well as other basic primitives
known from classical circuit design [183], such as
• adders modulo N : |x|y → |x|x + y mod N
• modular exponentiation: |x|0 → |x|ax mod N ,
in polylogarithmic time on a quantum computer [190, 191].
4.2.5 Preparing Quantum States
If we are interested in preparing particular quantum states by means of an ef-

fective procedure, in most cases it is straightforward to write down a quantum
circuit which yields the desired state when applied to the ground state |0.
For instance, by means of the quantum circuit given in Fig. 4.6, a Schrödinger
cat state |Ψn on n qubits can be prepared using n+1 elementary gates. These
are the states
1 1
|Ψn := √ |0 . . . 0 + √ |1 . . . 1 ,
2
2

n zeros n ones
and we remind the reader that |Ψ2 is locally equivalent to a so-called EPR
state [24] and |Ψ3 is a so-called GHZ state [37].
|0 H2 r r ··· r
❤ ···
..
.
❤ ···
.. .. .. ..
. . . .
|0 ··· ❤
Fig. 4.6. Quantum circuit that prepares a cat state
We remark that there is an algorithm to prepare an arbitrary quantum

state |ϕ starting from the ground state |0, i.e. to construct a quantum
circuit Uϕ yielding Uϕ |0 = |ϕ.

Algorithm 1 Let |ϕ = x∈Zn αx |x be a quantum state which we would
2
like to prepare. Do the following in a recursive way. Write
|ϕ = a|0|ϕ0 + b|1|ϕ1 ,
where a and b are complex numbers fulfilling |a|2 + |b|2 = 1. The states

|ϕ0 := α(0)
x |x and |ϕ1 := α(1)
x |x
x∈Zn−1
2 x∈Zn−1
2
appearing on the right-hand side of the above equation can be prepared by in-
duction hypotheses using the circuits U1 and U2 . Let A be the local transform

a −b
A := ⊗ 12n−1 .
b a
Then |ϕ can be prepared by application of the circuit A · Λ1 (U1 ) · Λ1 (U2 ) to

the ground state |0.
In general, the quantum circuit Uϕ for preparing a state |ϕ ∈ H2n generated
by this algorithm has a complexity κ(Uϕ ) = O(2n ), which is linear in the
dimension of the Hilbert space but exponential in the number of qubits.
However, as the example of cat states previously mentioned shows, there are
states which admit much more efficient preparation sequences. In such a set
of states, we also find the so-called symmetric states [192]
1
√ (|00 . . . 0 + |10 . . . 0 + |01 . . . 0 + . . . + |00 . . . 1) ,
n+1
i.e. the union of the orbits of |00 . . . 0 and |10 . . . 0 under the cyclic group
acting on the qubits. As shown in Sect. 4 of [192], these states can be prepared
using O(n) operations and a quadratic overhead of ancilla qubits. √ ν
Finally, we give circuits for preparation of the states |ψν := (1/ ν) i=1 |i
for ν = 1, . . . , 2n , which represent equal amplitudes over the first ν basis
states of H2n . The states |ψν can be efficiently prepared from the ground
state |0 by the following procedure (using the principle of binary search
[193]), which is described in Sect. 4 of [187].
Since |ψ2n can easily be prepared by application of the Hadamard trans-
formation H2⊗n , we can assume ν < 2n without loss of generality. We now
choose k ∈ N such that 2k ≤ ν < 2k+1 and apply the transformation
√ √
1 2 k − ν − 2k
U := √ √ √
ν ν − 2k 2k
to the first bit of the ground state |0. Next we achieve equal superposition on
the first 2k basis states |0 . . . 0, . . . , |0 . . . 01 . . . 1 by application of an (n−k)-
fold controlled Λ1 (H2⊗k ) operation, which can be implemented using O(n2 )
operations. Finally, we apply the preparation circuit for the state |ψν−2k
(which has been constructed by induction), conditioned on the (k + 1)th
bit. Overall, we obtain a complexity for the preparation of |ψν of O(n3 )
operations.
4.2.6 Quantum Turing Machines
Quantum circuits provide a natural framework to specify unitary transforma-

tions on finite-dimensional Hilbert spaces and give rise to complexity models
when factorizations into elementary gates, e.g. with respect to the universal
sets G1 or G2 , are taken into account.
Besides the formalism of quantum networks, there are other ways of de-
scribing computations performed by quantum mechanical systems. In the
following we briefly review the model of quantum Turing machines (QTMs)
defined by Deutsch [7]. We remind the reader that Turing machines [194]
provide a unified model for classical deterministic and probabilistic compu-
tation (see, e.g., [195]). The importance of Turing machines as a unifying
r q0
✟
✟ ❍❍
✟ ✟ ❍❍
q1
r✟✟ ❍❍r q2
❅ ❅
❅
· ❅ ··
·· ❅r qf1
r ·
❅ ❅
··
r qf2 ·
Fig. 4.7. Configurations of a Turing machine

concept for classical computing manifests itself in the Church–Turing thesis
[196, 197], which, in its modern form, claims that every physically reasonable
model of computation can be efficiently simulated on a probabilistic Turing
machine.
Definition 4.2. A deterministic Turing machine T is defined by the data

(Q, Σ,q0 , F, t), where Q is a finite set of states, Σ a set of symbols, q0 ∈ Q
a distinguished initial state, F ⊆ Q the set of final states, and t : Q × Σ −→
Q × Σ × {←, ↓, →} the transition function. The admissible actions of T are
movements of a read–write head, which in one computational step can move
to the left, stay where it is or move to the right (we have denoted these actions
by {←, ↓, →}). Along with the Turing machine T comes a infinite tape of cells
(the cells are in bijection with Z) which can take symbols from Σ.
A configuration of a Turing machine T is therefore given by a triplet

(v, p, q) ∈ (Σ Z , Z, Q), consisting of the state v of the tape, the position p
of the head and the internal state q. We obtain a tree of configurations by
considering two configurations c1 and c2 to be adjacent if and only if c2 is
obtained from c1 by an elementary move, i.e. scanning a symbol from the
tape, changing the internal state, writing back a symbol to the tape and
moving the head. The initial state is the root of this tree, whereas the final
states constitute its leaves (see Fig. 4.7).
A probabilistic Turing machine differs from a deterministic one only in the
nature of the transition function t, which is then a mapping
t : Q × Σ × Q × Σ × {←, ↓, →} −→ [0, 1] ,
which assigns probabilities from the real interval [0, 1] to the possible actions
of T . A normalization condition, which guarantees the well-formedness of a
probabilistic Turing machine, is that for all configurations the sum of the
probabilities of all successors is 1. Therefore, the admissible state transitions
of a probabilistic Turing machine T can be described by a stochastic matrix
ST ∈ [0, 1]Z×Z, where stochasticity means that the rows of ST add up to 1
and the successor csucc of c is obtained by csucc = ST · c. Note that a deter-
ministic Turing machine is a special case of a probabilistic Turing machine
with a “subpermutation” matrix ST , i.e. ST can be obtained from a suitable

permutation matrix by deleting some of its rows.
One more variation of this idea is needed to finally arrive at the concept
of a quantum Turing machine: we have to require that the transition function
t is a mapping with a normalization condition
t : Q × Σ × Q × Σ × {←, ↓, →} −→ C (4.4)
from a configuration to possibly many successors, each of which is given a

complex amplitude. Here the normalization constraint says that the matrix
UT describing the dynamics of T on the state space is unitary.
Observe that there is one counterintuitive fact implied by this definition:
as t assigns complex amplitudes to Q × Σ × Q × Σ × {←, ↓, →} (according to
(4.4)), one can interpret a configuration of T as being in a superposition of
(i) tape symbols in each individual cell, (ii) states of the finite-state machine
supported by Q and (iii) positions of the head. The last point in particular
might look uncomfortable at first sight, but we remind the reader of the fact
that in classical probabilistic computation each individual configuration is
assigned a probability, so one can think of a probabilistic computation as
traversing an exponentially large configuration space! The main difference
of the QTM model is that because of negative amplitudes, computational
paths in this configuration space can cancel each other out, i.e. the effects of
interference can force the Turing machine into certain paths which ultimately
may lead to the desired solution of the computational task.
Now that the computational model of a QTM has been defined, the ques-
tion arises as to what can be computed on a QTM, compared with a classical
deterministic or probabilistic Turing machine. An important result in this
context is that everything which can be computed classically in polynomial
time can also be computed on a QTM because of the following theorem, which
relies on some results of Bennett for reversible Turing machines [55, 198, 199]
and was adapted to the QTM setting in [200]. As usual, we denote by L∗ the
language {0, 1}∗ consisting of all binary strings and denote by |x| the length
of the word x ∈ L∗ .
Theorem 4.5. Let f : L∗ → L∗ be a polynomial-time computable function

such that |f (x)| depends only on |x|. Then there is a polynomial-time QTM
T that computes |x|0 → |x|f (x). The running time of T depends only on
|x|.
Proof. The basic idea is to replace each elementary step in the computation
of f by a reversible operation (using theorem 4.3), keeping in mind that an-
cilla qubits are needed to make the computation reversible (see Sect. 4.2.3).
We now adjoin additional qubits to the system, which are initialized in the
ground state |0, and apply a controlled NOT operation using the compu-
tational qubits holding the result f (x). Of course, after the application of
this operation the state is highly entangled between the computational reg-
ister and the additional register holding the result. Next, we run the whole
computation that was done to compute f , backwards, reversibly, on the com-
putation register to get rid of the garbage which might destroy the coherence,
and end up with the state |x, 0, f (x), where the |0 refers to the ancilla bits
used in the first step of this procedure.
Remark 4.2.
• The class of quantum Turing machines allows the definition and study
of the important complexity class BQP [200],3 as well as the relation of
BQP to other classes known from classical complexity theory.
• There are programming primitives for QTMs, such as composition, loops
and branching [200], as in the classical case. However, a problem arises
in realizing while-loops, since the predicate which decides whether the
loop terminates can be in a superposition of true and false, depending on
the computation path. Therefore all computations have to be arranged
in such a way that this predicate is never in a superposed state, i.e. the
state of the predicate has to be classical. As a consequence, we obtain
the result that a quantum Turing machine can only perform loops with
a prescribed number of iterations, which in turn can be determined by a
classical Turing machine.
• An important issue is whether QTMs constitute an analog or discrete
model of computation. One might be tempted to think of the possibil-
ity of encoding an arbitrary amount of information into the transition
amplitudes of t, i.e. of producing a machine model which could benefit
from computing with complex numbers to arbitrary precision (for the
strange effects of such models see, e. g., [201]). However, see [200, 202] for
a proof of the fact that it is sufficient to take transition amplitudes from
the finite set {±3/5, ±4/5, ±1, 0} in order to approximate a given QTM
to arbitrary precision. The reason for this is that the Pythagorean-triple
transformation

1 3 −4
∈ SO(2)
5 4 3
has eigenvalues of the form e2πiν with ν ∈

/ Q and, therefore, generates a
dense subgroup in SO(2). The statement then follows from the fact that
the full unitary group on H2n can be parametrized by SO(2) matrices
applied to arbitrary basis states and phase rotations [172, 203].
• Yao has shown [204] that the computational models of QTM and uniform
families of quantum gates (see Sect. 4.2.1) are polynomially equivalent,
i.e. each model can simulate the other with polynomial time overhead.
3
BQP stands for “bounded-error quantum polynomial time”.
4.3 Using Entanglement for Computation:

A First Quantum Algorithm
Entanglement between registers holding quantum states lies at the heart of

the quantum algorithm which we shall describe in this section. The formula-
tion of a problem on which a quantum computer will exceed the performance
of any classical (probabilistic) computer may appear artificial. However, this
was one of the first examples of problems on which a quantum computer could
provably outperform any classical computer, with an exponential speedup.
Because of its clarity and methodology, we present the quantum algo-
rithm of Simon, in which many of the basic principles of quantum computing,
namely the superposition principle, computing with preimages and the use
of the Fourier transform, become apparent. We briefly remind the reader of
the problem and mention that we are considering here a slightly generalized
situation compared with the original setup (see also [57, 205, 206]).
Quantum algorithms relying on the same principles have been given in [56,
200]. As in the case of Simon’s problem described below, these algorithms rely
on the Fourier transform for a suitably chosen abelian group. In both cases
it has been shown that these quantum algorithms provide a superpolynomial
gap over any classical probabilistic computer in the number of operations
necessary to solve the corresponding problems.
In the following, we denote by Zn2 the elementary abelian 2-group of order
n
2 , the elements of which we think of as being identified with binary strings
of length n, and denote addition in Zn2 by ⊕.
Definition 4.3 (Simon’s Problem). Let f : Zn2 → Zn2 be a function given

as a black-box quantum circuit, i.e. f can be evaluated on superpositions of
states and is realized by a unitary transform Vf specified by
|x|0 → |x|f (x), for all x ∈ Zn2 ,
as described in Sect. 4.2.3 (see, in particular, (4.3)). In addition, it is specified

that there is a subgroup U ⊆ Zn2 (the “hidden” subgroup) such that f takes a
constant value on each of the cosets g ⊕ U for g ∈ Zn2 and, furthermore, f
takes different values on different cosets. The problem is to find generators
for U .
We can now formulate a quantum algorithm which solves Simon’s problem

in a polynomial number of operations on a quantum computer. This algorithm
uses O(n) evaluations of the black-box quantum circuit f , and the classical
postcomputation, which is essentially linear algebra over GF (2), and also
takes a number of operations which is polynomial in n.
Algorithm 2 This algorithm needs two quantum registers of length n, hold-

ing elements of the domain and codomain of f , and consists of the following
steps.
1. Prepare the ground state
|ϕ1 = |0 . . . 0 ⊗ |0 . . . 0
in both quantum registers.

2. Achieve equal amplitude distribution in the first register, for instance by
an application of a Hadamard transformation to each qubit:
1
|ϕ2 = √ |x ⊗ |0 . . . 0 .
2n x∈Zn
2
3. Apply Vf to compute f in superposition. We obtain
1
|ϕ3 = √ |x|f (x) .
2n x∈Zn
2
4. Measure the second register to obtain some value z in the range of f .

Owing to the condition on f specified, the first register now holds a coset
g0 ⊕ U of the hidden subgroup U , namely the set of elements equal to
z = f (g0 ):
1 1
|ϕ4 = |x|z = |x|z .
|U | f (x)=z |U | x∈g0 ⊕U
⊗n
5. Application of the Hadamard transformation H2 to y·g the first register
transforms the coset into the superposition y∈U ⊥ (−1) 0 |y. The sup-
ported vectors of this superposition are the
elements of U ⊥ , which is the
⊥ n n
group defined by U := {y ∈ Z2 : x · y = i=1 xi yi = 0}, i.e. the orthog-
onal complement of U with respect to the scalar product in Zn2 (see also
Sect. 4.4.2).
6. Now measure the first register. We draw from the set of irreducible rep-
resentations of Zn2 having U in the kernel, i.e. we obtain an equal distri-
bution over the elements of U ⊥ .
7. By iterating steps 1-6, we produce elements of Zn2 which generate the
group U ⊥ with high probability. After performing this experiment an ex-
pected number of n times, we generate with probability greater than 1−2−n
the group U ⊥ .
8. By solving linear equations over GF (2), it is easy to find generators for
(U ⊥ )⊥ = U .
Therefore we obtain generators for U by computing the kernel of a matrix

over GF (2) in time O(n3 ).
Analysis of Algorithm 2
• We first address the measurement in step 4. If we do not perform this
measurement, we are left with the state
1
|ϕ4 = √ |σ ⊕ x|f (σ) ,
n
2 σ∈Zn /U x∈U
2
and if we continue with step 5 we shall obtain the state

1
|ϕ5 = √ (−1)σ·y |y|f (σ) .
2n σ∈Zn /U y∈U ⊥
2
Therefore, sampling of the first register as in step 6 will yield an equal

distribution over U ⊥ and we can go on as in step 7. Hence we can omit
step 4.
• The reason for the application of the transformation H2⊗n and the ap-
pearance of the group U ⊥ will be clarified in the following sections. As it
turns out, H2⊗n is an instance of a Fourier transform for an abelian group
and ⊥ is an antiisomorphism of the lattice of subgroups of Zn2 .
• For the linear-algebra part in step 8, we refer the reader to standard texts
such as [207]. Gauss’s algorithm for computing the kernel of an n × n ma-
trix takes O(n3 ) arithmetic operations over the finite field Z2 = GF (2).
Overall, we obtain the following cost: O(n) applications of Vf , O(n2 ) el-
ementary quantum operations (which are all Hadamard operations H2 ),
O(n2 ) measurements of individual qubits, and O(n3 ) classical operations
(arithmetic in GF (2)).
4.4 Quantum Fourier Transforms: the Abelian Case

In this section we recall the definition and basic properties of the discrete
Fourier transform (DFT) and give examples of its use in quantum computing.
In most standard texts on signal processing (e.g. [208]), the DFTN of a
periodic signal given by a function f : ZN → C (where ZN is the cyclic group
of order N ) is defined as the function F given by

F (ν) := e2πiνt/N f (t) .
t∈ZN
If, equivalently, we adopt the point of view that the signal f and the Fourier
transform are vectors in C N , we see that performing the DFTN is a matrix
vector multiplication of f with the unitary matrix
1
DFTN := √ · ω i·j i,j=0,...,N −1 ,
N
where ω = e2πi/N denotes a primitive N th root of unity.
From an algebraic point of view, the DFTN gives an isomorphism Φ of

the group algebra CZN ,
Φ : CZN −→ C N ,
onto the direct sum of the irreducible matrix representations of ZN (where

multiplication is performed pointwise). This means that DFTN decomposes
the regular representation of ZN into its irreducible constituents. It is known
that this property allows the derivation of a fast convolution algorithm [171]
in a canonical way and can be generalized to more general group circulants.
Viewing the DFT as a decomposition matrix for the regular representation
of a group leads to the generalization of Fourier transforms to arbitrary finite
groups (cf. [171] and Sect. 4.6.1).
The fact that DFTN can be computed in O(N log N ) arithmetic opera-
tions (counting additions and multiplications) is very important for applica-
tions in classical signal processing. This possibility to perform a fast Fourier
transform justifies the heavy use of the DFTN in today’s computing technol-
ogy. In the next section we shall show that the O(N log N ) bound, which is
sharp in the arithmetic complexity model (see Chap. 4 and 5 of [174]) can be
improved with a quantum computer to O[(log N )2 ] operations.
4.4.1 Factorization of DFTN
From now on we restrict ourselves to cases of DFTN where N = 2n is a power

of 2, since these transforms naturally fit the tensor structure imposed by the
qubits.
The efficient implementation of the Fourier transform on a quantum com-
puter starts from the well-known Cooley–Tukey decomposition [209]: after the
row permutation Πτ , where τ = (1, . . . , n) is the cyclic shift on the qubits, is
performed, the DFT2n has the following block structure [171]:
 
DFT2n−1 DFT2n−1
Πτ DFT2n =  
DFT2n−1 Wn −DFT2n−1 Wn
= (12 ⊗ DFT2n−1 ) · Tn · (DFT2 ⊗ 12n−1 ) .
Here we denote by
 
1
 ω2n 
 
 ω22n 
Tn := 12n−1 ⊕ Wn , Wn :=  
 .. 
 . 
n−1
ω22n
the matrix of twiddle factors [171]. Taking into account the fact that Wn has
the tensor decomposition
1
n−2
Wn = i ,
ω22n
i=0
we see that Tn can be implemented by n−1 gates having one control wire each.
These can be factored into the elementary gates G1 with constant overhead.
Because tensor products are free in our computational model, by recursion
we arrive at an upper bound of O(n2 ) for the number of elementary operations
necessary to compute the discrete Fourier transform on a quantum computer
(this operation will be referred to as “QFT”).
In Fig. 4.8, the derived decomposition into quantum gates is displayed
using the graphical notation introduced in Sect. 4.2.1. The gates labeled by
Dk in this circuit are the diagonal phase shifts diag(1, e2πi/k ) and, in ad-
dition, we have used the abbreviation N = 2n . We observe that the per-
mutations Πn , which arose in the Cooley–Tukey formula, have all been col-
lected together, yielding the so-called bit reversal, which is the permutation
of the quantum wires (1, n) (2, n−1) . . . (n/2, n/2 + 1) when n is even and
(1, n) (2, n−1) . . . ((n − 1)/2, (n + 3)/2) when n is odd.
··· • • ··· • H2
❆ ✁✁
❆ ✁ ··· ··· D4
❅❆ ✁
.. ❅❆✁ .. · · ..
. . ·· ·· .
✁ ❆❅❅ • ··· DN ···
✁ ❆ H2 2
✁ ❆❆ ··· ···
H2 D4 DN
Fig. 4.8. Quantum circuit that computes a Fourier transform QFT2n
If we intend to use the Fourier transform as a sampling device, i.e. the

application of QFT2n to a state |ψ followed directly by a measurement in
the standard basis, we can use the structure of this quantum circuit that
computes a Fourier transform QFT2n to avoid nearly all quantum gates [210].
In order to give circuits for the QFT for an arbitrary abelian group, we
need to describe the structure of these groups and their representations first,
which is done in the next section.
4.4.2 Abelian Groups and Duality Theorems
Let A be a finite abelian group. Then A splits into a direct product of its
p-components: A ∼ = Ap1 × . . . × Apn , where pi , i = 1, . . . , n are the prime
divisors of the order |A| of A (see [211], Part I, paragraph 8).
Furthermore, each Api , which consists of all elements annihilated by a

power of pi , has the form Api ∼
= Zpri,1 × . . . × Zpri,ki (see [211], Part I, para-
graph 8). Both statements follow immediately from the structure theorem for
finitely generated modules over principle ideal rings.
We remark that a Fourier transform for a finite abelian group can easily be
constructed from this knowledge: given two groups G1 and G2 , the irreducible
representations of their direct product G1 × G2 can be obtained from those
of G1 and G2 as follows.
Theorem 4.6. The irreducible representations of G = G1 × G2 are given by
Irr(G) = {φ1 ⊗ φ2 : φ1 ∈ Irr(G1 ), φ2 ∈ Irr(G2 )} .
(For a proof see [212].) If we therefore encode the elements of A according to

the direct-product decomposition as above, the matrix
ki
n

DFTA = DFTpi ri,j
i=1 j=1
is a decomposition matrix for the regular representation of A.
Corollary 4.1. Let A be a finite abelian group of order 2n . Then a Fourier

transform for A can be computed in O(n2 ) elementary operations.
Proof. The decomposition of DFT2n has already been considered and an im-
plementation in O(n2 ) many operations has been derived from the Cooley–
Tukey formula in Sect. 4.4.1. Since tensor products are free in our computa-
tional model, we can conclude that a direct factor Z2n is already the worst
case for an implementation.
Example 4.6. The Fourier transform for the elementary abelian 2-group Zn2
is given by the tensor product DFT2 ⊗ · · · ⊗ DFT2 of the Fourier transform
for the cyclic factors and hence coincides with the Hadamard matrix H2⊗n
used in Algorithm 2.
The Dual Group. Given a finite abelian group A, we can consider Hom(A,C ∗),
i.e. the group of characters of A (NB: in the nonabelian case a character is
generalized to the traces of the representing matrices and hence is not a ho-
momorphism anymore). The following theorem says that A is isomorphic to
its group of characters.
Theorem 4.7. For a finite abelian group A, we have A ∼

=φ Hom(A, C ∗ ).
We shall make the isomorphism φ explicit. Choose ωp1 , . . . , ωpn , where

the ωi are primitive pi th roots of unity in C ∗ . Then φ is defined by the
assignment φ(ei ) := ωpi on the elements ei := (0, . . . , 1, . . . , 0), where the 1
is in the ith position. The isomorphism φ is not canonical, since a different

choice of primitive roots of unity will yield a different isomorphism.
Next we observe that there is a pairing β (i.e. a bilinear map) between A
and Hom(A, C ∗ ) via β : (a, f ) → f (a) ∈ C ∗ . We suppose that a is orthogonal
to f (in symbols, a ⊥ f ) if β(a, f ) = 1. For a given subgroup U ⊆ A we can
now define (see also Algorithm 2) the orthogonal complement U ⊥ := {y ∈
A : β(y, x) = 1, ∀x ∈ U }. The following duality theorem holds (see [184]):
Theorem 4.8. For all subgroups U, U ⊆ A, the following identities hold:
(a) Self-duality of ⊥:
(U ⊥ )⊥ = U .
(b) Complementarity:
⊥
(U ∩ U )⊥ = U ⊥ , U .
(c) The mapping ⊥ is an inclusion-reversing antiisomorphism on the lattice
of subgroups of A (i.e. ⊥ is a Galois correspondence).
4.4.3 Sampling of Fourier Coefficients

In this section we address the problem of gaining information from the Fourier
coefficients of a special class of states. More precisely, we consider the Fourier
transforms of the (normalized) characteristic function
1
|χc+U := |x (4.5)
|U | x∈c+U
of a coset c + U of a subgroup U ⊆ A of an abelian group A. The question of

what conclusions about U can be drawn from measuring the Fourier trans-
formed state (4.5) is of special interest, as we have already seen in Sect. 4.3.
The following theorem shows that for an abelian group, the Fourier transform
maps subgroups to their duals.

Theorem 4.9. Let DFTA = (1/ |A|) x,y∈A β(x, y)|y x| be the Fourier
transform for the abelian group A. Then for each subgroup U ⊆ A, we have
1 1
DFTA |x = |y .
|U | x∈U |U ⊥ | y∈U ⊥
Proof. Since
 
1 1 
DFTA |x = β(x, y)|y x| |x
|U | x∈U |U | x,y∈A x∈U
1
= β(x, y)|y ,
|A||U | y∈A x∈U

it suffices to show that x∈U β(x, y) = 0 for y ∈ / U ⊥ , but this statement
follows
from the fact that the existence of x0 with β(x0 , y) = 0 implies that
β(x, y) = β(x + x , y) = β(x , y) x∈U β(x, y). The other case,
x∈U x∈U
⊥
0 0
x∈U β(x, y) = |U | for y ∈ U , is obvious.

Hence, measuring the Fourier spectrum of |χU yields an equal distribu-
tion on the elements of the dual group U ⊥ . Also, in the case of the character-
istic function of a coset c + U instead of U , we obtain the same probability
distribution on U ⊥ since the Fourier transform diagonalizes the group ac-
tion completely in the abelian case, i.e. the translation by c corresponds to a
pointwise multiplication by phases in the Fourier basis:
1 1
DFTA |x = ϕc,y |y ,
|U | x∈c+U |U ⊥ | y∈U ⊥
where ϕc,y ∈ U(1) are phase factors which depend on c and y but are al-
ways eth roots of unity, where e denotes the exponent of A. Since making
measurements involves taking the squares of the amplitudes, we obtain an
equal distribution over U ⊥ . The states |χc+U , which in general will be highly
entangled, make the principle of interference and its use in quantum algo-
rithms apparent: only those Fourier coefficients remain which correspond to
the elements of U ⊥ (constructive interference), whereas the amplitudes of all
other elements vanish (destructive interference).
4.4.4 Schur’s Lemma and its Applications

in Quantum Computing
In this section we explore the underlying reason behind Theorem 4.9, namely
Schur’s lemma. We present further applications of this powerful tool from
representation theory. For a proof of Schur’s lemma we refer the reader to
Sect. 2 of [213].
Lemma 4.2 (Schur’s lemma). Let ρ1 : G → GLC (V ) and ρ2 : G →
GLC (W ) be irreducible complex representations of a group G. Suppose that
the element A ∈ End(V, W ) has the property
ρ1 (g) · A = A · ρ2 (g), ∀g ∈ G ,
i.e. A commutes with all pairs of images ρ1 (g), ρ2 (g). Then exactly one of the
following cases holds:
∼ ρ2 . In this case A = 0End(V,W ) .
(i) ρ1 =
(ii) ρ1 ∼= ρ2 . In this case A is a homothety, i.e. A = λ · 1End(V,W ) with an
element λ ∈ C.
We give two applications of Schur’s lemma which make its importance in
the context of Fourier analysis apparent. First, we present a reformulation
of Theorem 4.9 in representation-theoretical terms and state then a theorem
which is useful in the sampling of functions having a hidden normal subgroup.
Theorem 4.10. Let G be a finite abelian group, Hom(G, C ∗ ) the dual group,
and β : G × Hom(G, C ∗ ) → C ∗ the canonical pairing defined by β(g, ϕ) :=
ϕ(g). Then for each subgroup U ⊆ G, the following holds (normalization
omitted):

DFTG |u = |ϕ .
u∈U ϕ∈Hom(G,C ∗ )
U⊆Ker(ϕ)
Moreover, we obtain the following identity for the cosets u0 + U :

DFTG |u0 + u = β(u0 , ϕ) · |ϕ .
u∈U ϕ∈Hom(G,C ∗ )
U⊆Ker(ϕ)
|G|
Proof. The mapping DFTG : CG → i=1 C is given by evaluation of elements
of G for the irreducible representations {ϕ1 , . . . , ϕs } of G, which are all one-
dimensional (i.e. s = |G|) and hence are characters, since G was assumed to
be abelian. Therefore, the coefficient for the irreducible representation ϕi is
computed from

ϕi (u0 + u) = ϕi (u0 ) · ϕi (u)
u∈U u∈U

= ϕi (u0 ) · ϕi (u) .
u∈U
Considering the restricted

characters ϕi ↓ U , we use Schur’s lemma (Lemma
4.2) to deduce that u∈U ϕi (u) = 0 iff U ⊆ Ker(ϕi ).
Theorem 4.11. Let G be an arbitrary finite group and N G a normal

subgroup of G. Then for each irreducible representation ρ of G of degree d,
exactly one of the following cases holds:
1 1
ρ(n) = λ · 1d×d , or ρ(n) = 0d×d ,
|N | |N |
n∈N n∈N
where the first case applies iff N is contained in the kernel of ρ.

Proof. Let A := (1/|N |) n∈N ρ(n) denote the equal distribution over all
images of N under ρ. Then, from the assumption of normality of N , we
obtain
ρ(g)−1 · A · ρ(g) = A ,
i.e. A commutes with the irreducible representation ρ. Using Schur’s lemma,

we conclude that A = λ · 1d×d . If N ⊆ Ker(ρ) we find A = 1d×d. On the
other hand, if N ⊆ Ker(ρ), we conclude that A is the zero matrix 0d×d ,
because otherwise an element n0 ∈/ N equal to ρ(n0 ) = 1d×d would lead to

the contradiction

ρ(n0 ) · ρ(n) = ρ(n0 · n) = ρ(n) ,
n∈N n∈N n∈N
since from this we could conclude ρ(n0 ) = 1d×d , contrary to the assumption
n0 ∈
/ Ker(ρ).
4.5 Exploring Quantum Algorithms

4.5.1 Grover’s Algorithm
We give an outline of Grover’s algorithm for searching an unordered list and
present an optical implementation using Fourier lenses.
The search algorithm allows one to find elements in a list of N items
fulfilling a given predicate in time O(N 1/2 ). We assume that the predicate f
is given by a quantum circuit Vf and, as usual in this setting, we count the
invocations of Vf (the “oracle”). It is straightforward to construct from Vf an
operator Sf which flips the amplitudes of the states that fulfill the predicate,
i.e. Sf : |x → (−1)f (x) |x for all basis states |x. The Grover algorithm relies
on an averaging method called inversion about average, which is described in
the following.
Consider the matrix
 
−1 + 22n 2
n . . . 2
n
 2 2 
 
 2
2n −1 + 2
2n . . . 2
2n 
Dn :=    . (4.6)
.. .. .. . 
 . . . .. 
 
2
2n
2
2n . . . −1 + 22n
Dn is a circulant matrix [214], i.e. Dn = circG (−1+(2/2n), (2/2n ), . . . , (2/2n ))

for any choice of a finite group G. To see this, we recall the definition of a
general group circulant,
circG (v) := (vgi −1 ·gj )1≤i,j≤|G|
for a fixed ordering G = {g1 , . . . , g|G| } of the elements of G and for a vector v
which is labeled by G. To implement (4.6) on a quantum computer we apply
the circulant for the group Zn2 , making use of the following theorem.
Theorem 4.12. The Fourier transform DFTA for a finite abelian group A
implements a bijection between the set of A-circulant matrices and the set of
diagonal matrices over C. Explicitly, each circulant C is of the form
C = DFT−1
A · diag(d1 , . . . , dn ) · DFTA
and the vector d = (d1 , . . . , dn ) of diagonal entries is given by d = DFTA · c,

where c is the first row of C.
✻ ✻ ✻
a) ✲ b) ✲ c) ✲
Fig. 4.9. (a) Equal distribution, (b) flip solutions, (c) inversion about average
Grover’s original idea was to use correlations to amplify the amplitudes

of the states fulfilling the predicate, i.e. to correlate – starting from an ini-
tial distribution, which is chosen to be the equal distribution P (X = i) =
1/N, i = 1, . . . , N – the vector (−1, 1, . . . , 1) with the probability distribution
obtained by flipping the signs of the states fulfilling the predicate.
Starting from the equal distribution H2⊗n |0, the amplification process in
Grover’s
√ algorithm consists of an iterated application of the operator −Dn Sf
O( 2n ) times [215]. In Fig. 4.9, the steps of this procedure are illustrated in
a qualitative way.
We present a realization of the Grover algorithm with a diffractive optical
system. This is not a quantum mechanical realization in the sense of quan-
tum computing, since such a system does not support a qubit architecture.
However, the transition matrices are unitary and hence we can consider sim-
ulations of quantum algorithms via optical devices. These optical setups scale
linearly with the dimension of the computational Hilbert space rather than
with the logarithmic growth of a quantum register; nevertheless they have
some remarkable properties, the best known of which is the ability to perform
a Fourier transform by a simple application of a cylindrical lens [216]. This
resembles the classical Fourier transform (corresponding to DFTZ2n rather
than DFTZn2 ), which is correct because of the comments preceding Theorem
4.12.
Observe that by starting from this operation, we can easily perform cor-
relations since this corresponds to a multiplication with a circulant matrix.
Every circulant matrix can be realized optically using a so-called 4f setup,
which corresponds to a factorization of a circulant matrix C into diagonal ma-
trices and Fourier transforms (following Theorem 4.12): C = DFT−1 ·D·DFT,
where D is a given diagonal matrix (see also Fig. 4.10).
incoming diffractive element outgoing

wavefront G=Fg wavefront
h h«G
f f f f
Fig. 4.10. Optical 4f setup that computes a convolution of h and g

Remark 4.3. The question of what linear transforms can be performed op-
tically is equivalent to the question of what matrices can be factored into
diagonal matrices and Fourier transforms, which correspond to diagonal ma-
trices and circulant matrices. It has been shown in [217] that for fields
K ∈ {GF (3), GF (5)}, every square matrix M with entries in K can be
written as a product of circulant and diagonal matrices with entries in K.
Furthermore, if M is unitary the circulant and diagonal factors can also be
chosen to be unitary [217].
Remark 4.4. There have been several generalizations and modifications of

the original Grover algorithm.
• The case of an unknown number of solutions fulfilling the given predicate

is analyzed in [218].
• The issue of arbitrary initial distributions (instead of an equal distribu-
tion) is considered in [219].
• In [215] it is shown that instead of the diffusion operators Dn , almost
any (except for a set of matrices of measure zero) unitary matrix can be
used to perform the Grover algorithm.
• There is a quantum algorithm for the so-called collision problem which
needs O( 3 N/r) evaluations of a given r-to-one function f to find a pair
of values which are mapped to the same element [220].
• The Grover algorithm has been shown to be optimal, √ i.e. the problem
of finding an element in an unordered list takes Θ( N ) operations on a
quantum computer [200].
• Some other problems have been solved by a subroutine call to the Grover
algorithm, e.g. a problem in communication complexity: the problem of
deciding whether A, B ⊆ {1, . . . , N } are disjoint or not. Using the Grover
algorithm, it can be shown that it is sufficient for the√two parties, one
holding A and the other holding B, to communicate O( N log N ) qubits
to solve this problem [221].
4.5.2 Shor’s Algorithm
In this section we briefly review Shor’s factorization algorithm and show how
the Fourier transform comes into play. It is known that factoring a number N
is easy under the assumption that it is easy to determine the (multiplicative)
order of an arbitrary element in (ZN )× . For a proof of this, we refer the
reader to [222] and to Shor’s original paper [9].
Once this reduction has been done, the following observation is the cru-
cial step for the quantum algorithm. Let y be randomly chosen and let
gcd(y, N ) = 1. To determine the multiplicative order r of y mod N , con-
sider the function
fy (x) := y x mod N .
Clearly fy (x + r) = fy (x), i.e. fy is a periodic function with period r. The

quantum algorithm to determine this period is as follows:
Algorithm 3 Let N be given; determine M = 2m such that N 2 ≤ M ≤ 2N 2 .

This number M will be the length of the Fourier transform to be performed
in the following.
1. Randomly choose y with gcd(y, N ) = 1.
2. Prepare the state |0 ⊗ |0 in two registers of lengths m and log2 N .
3. Application of the Hadamard transform H2⊗m to the left part of the reg-
ister results in a superposition of all possible inputs
M−1
1
√ |x ⊗ |0 .
M x=0
4. We construct a unitary operation which computes the (partial function)

|x|0 → |x|y x mod N following Sect. 4.2.3 and 4.2.4. Calculation of
fy (x) = y x mod N yields for this superposition (normalization omitted)
M−1

|x ⊗ |y x mod N .
x=0
5. Measuring the right part of the register gives a certain value z0 . The
remaining state is the superposition of all x satisfying fy (x) = z0 :
s−1
M
|x|z0 = |x0 + kr|z0 , where y x0 = z0 and s = .
y x =z0
r
k=0
6. Performing a QFTM on the left part of the register leads to

M−1
s−1
e2πi(x0 +kr)l/M |l|z0 .
l=0 k=0
7. Finally, a measurement of the left part of the register gives a value l0 .
Application of this algorithm produces data from which the period r can be
extracted after classical postprocessing involving Diophantine approximation
(see Sect. 4.5.2).
A thorough analysis of this algorithm must take into account the overhead
for the calculation of the function fy : x → y x mod N . However, after this
function has been realized as a quantum network once (which can be obtained
from a classical network for this function in polynomial time), the superpo-
sition principle applies since all inputs can be processed by one application
of fy .
The role played by Fourier transforms in this algorithm is twofold. In

step 3 it is used to generate a superposition of all inputs from the ground
state |0. This could have been done by any unitary transformation having
an all-one vector in the first column. In step 6 we use the QFTM to extract
the information about the period r which was hidden in the graph of the
function fy .
Remark 4.5. It should be noted that for small numbers, as in the example of
Fig. 4.11 and 4.12, an optical setup using Fourier lenses (cf. Fig. 4.10) could
implement Shor’s algorithm. This very example was initially simulated with
r
the DigiOpt system [223].
Fig. 4.11. Function graph of f (x) = 2x mod 187
Fig. 4.12. The function shown in Fig. 4.11. transformed by a DFT of length 1024
An important question is the behavior of the states obtained after step 6 if

the length M of the Fourier transform QFTM we are using does not coincide
with the period r of the function which is transformed. Note that this period
r is exactly what has to be determined, and therefore the length N must
be chosen appropriately in order to gain enough information about r from
sampling. As it turns out, the choice M = 2m according to the condition N 2 ≤
M ≤ 2N 2 yields peaks in the Fourier domain which are sharply concentrated
around the values lM/r [9].
In Fig. 4.11 and 4.12, this effect is illustrated for N = 187 and M = 1024.
The choice of M does not fulfill the condition N 2 ≤ M ≤ 2N 2 . Nevertheless,
the characteristic peaks in the Fourier spectrum, which are sharply peaked
around multiples of the inverse of the order, are apparent.
Let us now consider this state, which has been oversampled by trans-
forming it with a Fourier transform of length M followed
s−1 by a measurement,
a little more closely. Constructive interference in k=0 e2πi(x0 +kr)l/M |l will
occur for those basis states |l for which lr is close to M . The probability of
measuring a specific l in this sum can be bounded by
1 2 1 2
1 1 − eπirs/M |2
s−1 s−1
2πikrl/M πir/M
√ e ≥ √ e =
sM sM sM |1 − eπir/M |2
k=0 k=0
1 | sin[πrs/(2M )]|2 4
= ≥ 2 .
sM | sin[πr/(2M )]|2 π r
The fractions l/M that we obtain from sampling fulfill the condition
l p
1
− ≤ ,
M r 2M
for some integer p. Because of the choice M ≥ N 2 we obtain the result that
l/M can be approximated efficiently by continued fractions, as described in
the following section. This classical postprocessing completes the description
of Shor’s algorithm for finding the order r of y. To factor N we compute the
least common divisors of (y r/2 + 1, N ) and (y r/2 − 1, N ), obtaining nontrivial
factors of N if r is even and y r/2 ≡ ±1 mod N .
A similar method can be applied to the discrete logarithm problem [9].
We mention that both the factoring and the discrete logarithm problem can
be readily recognized as hidden-subgroup problems (see also Sect. 5): in the
case of factoring, this corresponds to the group U generated by y and we are
interested in the index [Z : U ], which equals the multiplicative order of y. The
discrete logarithm problem for GF (q)× can be considered a hidden-subgroup
problem for the group G = Z × Z and the function f : G → GF (q)× given
by f (x, y) → ζ x α−y , where ζ is the primitive element and α is the element
for which we want to compute the logarithm.
The main difference from the situation in Simon’s algorithm (see Sect.
4.3) is that in these cases we cannot apply the Fourier transform for the
parent group G, since we do not know its order a priori. Rather, we have to
compute larger Fourier transforms; preferably, the length is chosen to be a
power of 2, to oversample. We then obtain the information from the sampled
Fourier spectrum by classical postprocessing.
The basic features of the method of Fourier sampling which we invoked in
Shor’s factoring algorithm are also incorporated in Kitaev’s algorithm for the
abelian stabilizer problem [187]. At the very heart of Kitaev’s approach is a
method to measure the eigenvalues of a unitary operator U , supposing that
the corresponding eigenvectors can be prepared. This estimation procedure
i
becomes efficient if, besides U , the powers U 2 , i = 0, 1, . . . , can also be
implemented efficiently [180, 187]. We must also mention that the method of
Diophantine approximation is crucial for the phase estimation.
Diophantine Approximation. In the following we briefly review some
properties of the continued-fraction expansion of a real number. An impor-
tant property a number can have is to be one of the convergents of such an
expansion. More specifically, the following theorem holds, which is important

in the sampling part of Shor’s algorithm for recovering the exact eigenvalues
of an operator from the data sampled.
Theorem 4.13. For each x ∈ R and each fraction p/q fulfilling

x − p < 1 ,
q 2 q2
the following holds: p/q is a convergent of the continued-fraction expansion

of the real number x.
A proof of this theorem can be found in standard texts on elementary

number theory (e.g. [224]). There exists a simple algorithm for producing an
expansion of a rational number into a continued fraction:
Algorithm 4 Let x ∈ Q. Define a0 := &x', x1 := 1/(x−a0 ), a1 := &x1 ', x2 :=
1/(x1 − a1 ) and so on, until we obtain xi = ai for the first time. Then we
can write x as
1
x = a0 + .
1
a1 +
1
a2 +
···+ 1
an
We mention in addition that Algorithm 4 not only yields optimal approx-

imations (if applied to elements x ∈ R) but is also very efficient, since in
principle just a Euclidean algorithm is performed.
4.5.3 Taxonomy of Quantum Algorithms
The quantum algorithms which have been discovered by now fall into two
categories, the principles of which we shall describe in the following.
Entanglement-Driven Algorithms. Suppose we are given a function f :
X → Y from a (finite) domain X to a codomain Y . This function does not
have to be injective; however, for a quantum computer to be able to perform
f with respect to a suitably encoded X and Y , the function f has to be
embedded into a unitary matrix Vf (cf. Sect. 4.2.3). We then can compute
simultaneously the images of all inputs x ∈ X using |x|0 → |x|f (x) for all
x ∈ X by preparing an equal superposition x∈X |x in the X register and
the ground state |0 in the Y register first, and then applying the quantum
circuit Vf to obtain x∈X |x|f (x). This entangled state can then be written
as

|x |y ,
y∈Im(f ) x:f (x)=y
i.e. we obtain a separation of the preimages of f . Measuring the second reg-

ister leaves us with one of these preimages.
Pars pro toto, we mention the algorithms of Shor, which have been de-
scribed in Sect. 4.5.2. In this case the function f is given by f (x) = ax mod N ,
where N is the number to be factored and a is a random element in ZN .
Superposition-Driven Algorithms. The principle of this class of algo-
rithms is to amplify to a certain extent the amplitudes of a set of “good”
states, e.g. states which are specified by a predicate given in the form of a
quantum circuit, and on the other hand to shrink the amplitudes of the “bad”
states.
Pars pro toto, we mention the Grover algorithm for searching an un-
ordered list. We gave an outline of this algorithm in Sect. 4.5.1 and presented
an optical implementation using Fourier lenses.
4.6 Quantum Signal Transforms
Abelian Fourier transforms have been used extensively in the algorithms in

the preceding sections. We now consider further classes of unitary transforma-
tions which admit efficient realizations in a computational model of quantum
circuits. In Sect. 4.6.1 we introduce generalized Fourier transforms which yield
unitary transformations parametrized by (nonabelian) finite groups. Classi-
cally, advanced methods have been developed for the study of fast Fourier
transforms for solvable groups [171, 174, 225, 226]. While it is an open ques-
tion whether efficient quantum Fourier transforms exist for all finite solvable
groups, it is possible to give efficient circuits for special classes (see Sect.
4.6.1). In Sect. 4.6.2 we consider a class of real orthogonal transformations
useful in classical signal processing, for which quantum circuits of polyloga-
rithmic size exist.
4.6.1 Quantum Fourier Transforms: the General Case
In Sect. 4.4 we have encountered the special case of the discrete Fourier trans-
form for abelian groups. This concept can be generalized to arbitrary finite
groups, which leads to an interesting and well-studied topic for classical com-
puters. We refer to [171, 174, 225, 226] as representatives of a vast number of
publications. The reader not familiar with the standard notations concerning
group representations is referred to these publications and to standard ref-
erences such as [213, 227]. Following [228], we briefly present the terms and
notations from representation theory which we are going to use, and recall
the definition of Fourier transforms.
The Wedderburn Decomposition. Let φ be a regular representation of

the finite group G. Then a Fourier transform for G is any matrix A that
decomposes φ into irreducible representations with the additional property
that equivalent irreducibles in the corresponding decomposition are equal.
Regular representations are not unique, since they depend on the ordering
of the elements of G. Also, note that this definition says nothing about the
choice of the irreducible representations of G, which in the nonabelian case
are not unique.
On the level of algebras, the matrix A is a constructive realization of the
algebra isomorphism

CG ∼= Mdi (C) (4.7)
i
of C-algebras, where Mdi (C) denotes the full matrix ring of di × di matri-
ces with coefficients in C. This decomposition is also known as Wedderburn
decomposition of the group algebra CG.
We remind the reader of the fact that the decomposition in (4.7) is quite
familiar from the theory of error-avoiding quantum codes and noiseless sub-
systems [229, 230, 231, 232].
As an example of a Fourier transform, let G = Zn = x | xn = 1
be the cyclic group of order n with regular representation φ = 1E ↑T G,
T = (x0 , x1 , . . . , xn−1 ), and let ωn be a primitive nth root of unity.4 Now
√
φA = n−1 i ij
i=0 ρi , where ρi : x → ωn and A = DFTn = (1/ n)[ωn | i, j =
0 . . . n − 1] is the (unitary) discrete Fourier transform well known from signal
processing.
If A is a Fourier transform for the group G, then any fast algorithm
for the multiplication with A is called a fast Fourier transform for G. Of
course, the term fast depends on the complexity model chosen. Since we
are primarily interested in the realization of a fast Fourier transform on a
quantum computer (QFT), we first have to use the complexity measure κ, as
derived in Sect. 4.2.1.
Classically, a fast Fourier transform is given by a factorization of the
decomposition matrix A into a product of sparse matrices5 [171, 174, 225,
233]. For a solvable group G, this factorization can be obtained recursively
using the following idea. First, a normal subgroup of prime index (G : N ) = p
is chosen. Using transitivity of induction, φ = 1E ↑ G is written as (1E ↑ N ) ↑
G (note that we have the freedom to choose the transversals appropriately).
Then 1E ↑ N , which again is a regular representation, is decomposed (by
4
The induction of a representation φ of a subgroup H ≤ G with transversal
T = (t1 , . . . , tk ) is defined by (φ ↑T G)(g) := [φ̇(ti gt−1
j ) | i, j = 1 . . . n], where
φ̇(x) := φ(x) for x ∈ H or else is the zero matrix of the appropriate size.
5
Note that in general, sparseness of a matrix does not imply low computational
complexity with respect to the complexity measure κ.
recursion), yielding a Fourier transform B for N . In the last step, A is derived

from B using a recursion formula.
A Decomposition Algorithm. Following [228], we explain this procedure
in more detail by first presenting two essential theorems (without proof)
and then stating the actual algorithm for deriving fast Fourier transforms
for solvable groups. The special tensor structure of the recursion formula
mentioned above will allow us to use this algorithm as a starting point to
obtain fast quantum Fourier transforms in the case where G is a 2-group
(i. e. |G| is a power of 2).
First we need Clifford’s theorem, which explains the relationship between
the irreducible representations of G and those of a normal subgroup N of G.
Recall that G acts on the representations of N via inner conjugation: given
a representation ρ of N and t ∈ G we define ρt : n → ρ(tnt−1 ) for n ∈ N .
Theorem 4.14 (Clifford’s Theorem). Let N G be a normal subgroup
of prime index p with (cyclic) transversal T = (t0 , t1 , . . . , t(p−1) ), and denote
by λi : t → ωpi , i = 0, . . . , p − 1, the p irreducible representations of G arising
from G/N . Assume ρ is an irreducible representation of N . Then exactly one
of the two following cases applies:
1. ρ ∼= ρt and ρ has p pairwise inequivalent extensions to G. If ρ is one of
them, then all are given by λi · ρ, i = 0, . . . , p − 1. p−1 i
2. ρ ∼
ρt and ρ ↑T G is irreducible. Furthermore, (ρ ↑T G) ↓ N = i=0 ρt
=
and
D⊗1d
(λi · (ρ ↑T G)) = ρ ↑T G , D = diag(1, ωp , . . . , ωp(p−1) )i .
The following theorem provides the recursion formula and was used earlier
by Beth [171] to obtain fast Fourier transforms based on the tensor product
as a parallel-processing model.
Theorem 4.15. Let N G be a normal subgroup of prime index p having
a transversal T = (t0 , t1 , . . . , t(p−1) ), and let φ be a representation of degree
d of N . Suppose that A is a matrix decomposing φ into irreducibles, i.e.
φA = ρ = ρ1 ⊕ . . . ⊕ ρk , and that ρ is an extension of ρ to G. Then
p−1

(φ ↑T G)B = λi · ρ ,
i=0
where λi : t → ωpi , i = 0, . . . , p − 1, are the p irreducible representations of

G arising from the factor group G/N ,
p−1

B = (1p ⊗ A) · D · (DFTp ⊗ 1d ) , and D= ρ(t)i .
i=0
If, in particular, ρ is a direct sum of irreducibles, then B is a decomposition

matrix of φ ↑T G.
In the case of a cyclic group G the formula yields exactly the well-known
Cooley–Tukey decomposition (see also Sect. 4.4.1 and [209]), in which D is
usually called the twiddle matrix.
Assume that N G is a normal subgroup of prime index p with Fourier
m
transform A and decomposition φA = ρ = i=1 ρi . We can reorder the ρi
such that the first, say k, ρi have an extension ρi to G and the other ρi occur
(p−1)
as sequences ρi ⊕ ρti ⊕ . . . ⊕ ρti of inner conjugates (cf. Theorem 4.14; note
j
t
that the irreducibles ρi , ρi have the same multiplicity since φ is regular). In
the first case the extension may be calculated by Minkwitz’s formula [234];
in the latter case each sequence can be extended by ρi ↑T G (Theorem 4.14,
case 2). We do not state Minkwitz’s formula here, since we shall not need it
in the special cases treated later on. Altogether, we obtain an extension ρ of ρ
and can apply Theorem
p 4.15. The remaining task is to ensure that equivalent
irreducibles in i=1 λi · ρ are equal. For summands of ρ of the form ρi we
have the result that λj · ρi and ρi are inequivalent, and hence there is nothing
to do. For summands of ρ of the form ρi ↑T G, we conjugate λj · (ρi ↑T G)
onto ρi ↑T G using Theorem 4.14, case 2.
Now we are ready to formulate the recursive algorithm for constructing a
fast Fourier transform for a solvable group G due to Püschel et al. [228].
Algorithm 5 Let N G be a normal subgroup of prime index p with transver-
sal T = (t0 , t1 , . . . , t(p−1) ). Suppose that φ is a regular representation of N
with (fast) Fourier transform A, i.e. φA = ρ1 ⊕ . . . ⊕ ρk , fulfilling ρi ∼ = ρj ⇒
ρi = ρj . A Fourier transform B of G with respect to the regular representation
φ ↑T G can be obtained as follows.
1. Determine a permutation matrix P that rearranges the ρi , i = 1, . . . , k,
such that the extensible ρi (i.e. those satisfying ρi = ρti ) come first,
followed by the other representations ordered into sequences of length p
(p−1)
equivalent to ρi , ρti , . . . , ρti . (Note that these sequences need to be equal
(p−1)
t t
to ρi , ρi , . . . , ρi , which is established in the next step.)
2. Calculate a matrix M which is the identity on the extensibles and conju-
(p−1)
gates the sequences of length p to make them equal to ρi , ρti , . . . , ρti .
3. Note that A · P · M is a decomposition matrix for φ, too, and let ρ =
φA·P ·M . Extend ρ to G summand-wise. For the extensible summands use
(p−1)
Minkwitz’s formula; the sequences ρi , ρti , . . . , ρti can be extended by
ρi ↑T G. p−1

4. Evaluate ρ at t and build D = ρ(t)i .
i=0
5. Construct a block-diagonal matrix C with Theorem 4.14, case 2, conju-
p−1
gating i=0 λi · ρ such that equivalent irreducibles are equal. C is the
identity on the extended summands.
Result:
B = (1p ⊗ A · P · M ) · D · (DFTp ⊗ 1|N | ) · C (4.8)
is a fast Fourier transform for G.
DFT2 n
n–1
D C
... A ... P ... M ... ... ... ..
.
1
Fig. 4.13. Coarse quantum circuit visualizing Algorithm 5 for 2-groups
It is obviously possible to construct fast Fourier transforms on a classical

computer for any solvable group by recursive use of this algorithm.
Since we are restricting ourselves to the case of a quantum computer
consisting of qubits, i.e. two-level systems, we apply Algorithm 5 to obtain
QFTs for 2-groups (i.e. |G| = 2n and thus p = 2). In this case the two tensor
products occuring in (4.8) fit very well and yield a coarse factorization, as
shown in Fig. 4.13. The lines in the figure correspond to the qubits as in Sect.
4.2.1, and a box ranging over more than one line denotes a matrix admitting
no a priori factorization into a tensor product.
The remaining problem is the realization of the matrices A, P, M, D, C in
terms of elementary building blocks as presented in Sect. 4.2.1. At present,
the realization of these matrices remains a creative process to be performed
for given (classes of) finite groups.
In [228] Algorithm 5 is applied to a class of nonabelian 2-groups, namely
the 2-groups which contain a cyclic normal subgroup of index 2. These have
been classified (see [235], Sect. 14.9, pp. 90–91), and for n ≥ 3 there are
exactly the following four isomorphism types:
n
1. the dihedral group D2n+1 = x, y | x2 = y 2 = 1, xy = x−1
n
2. the quaternion group Q2n+1 = x, y | x2 = y 4 = 1, xy = x−1
n n−1
3. the group QP2n+1 = x, y | x2 = y 2 = 1, xy = x2 +1
n n−1
4. the quasi-dihedral group QD2n+1 = x, y | x2 = y 2 = 1, xy = x2 −1 .
Observe that the extensions 1, 3 and 4 of the cyclic subgroup Z2n = x split,
i.e. the groups have the structure of a semidirect product of Z2n with Z2 .
The three isomorphism types correspond to the three different embeddings
of Z2 = y into (Z2n )× ∼= Z2 × Z2n−2 .
In [228] quantum circuits with polylogarithmic gate complexity are given
for the Fourier transforms for each of these groups. See also [236, 237] for
quantum Fourier transforms for nonabelian groups.
An Example: Wreath Products. In this section we recall the definition

of wreath products in general (see also [235, 238]) and, as an example, give
efficient quantum Fourier transforms for a certain family of wreath products.
Definition 4.4. Let G be a group and H ⊆ Sn be a subgroup of the symmet-

ric group on n letters. The wreath product G * H of G with H is the set
{(ϕ, h) : h ∈ H, ϕ : [1, . . . , n] → G}
equipped with the multiplication
(ϕ1 , h1 ) · (ϕ2 , h2 ) := (ψ, h1 h2 )i ,
where ψ is the mapping given by i → ϕ1 (ih2 )ϕ2 (i) for i ∈ [1, . . . , n].
In other words, the wreath product is isomorphic to a semidirect product
of the so-called base group N := G×. . .×G, which is the n-fold direct product
of (independent) copies of G with H, in symbols G * H = N H, where H
operates via permutation of the direct factors of N . So we can think of the
elements as n-tuples of elements from G together with a permutation τ , and
multiplication is done componentwise after a suitable permutation of the first
n factors:
(g1 , . . . , gn ; τ ) · (g1 , . . . , gn ; τ ) = (gτ (1) g1 , . . . , gτ (n) gn ; τ τ ) .
In this section we show how to compute a Fourier transform for certain

wreath products on a quantum computer. We show how the general recursive
method to obtain fast Fourier transforms on a quantum computer described
in [228] can be applied directly in the case of wreath products A * Z2 , where
A is an arbitrary abelian 2-group. The recursion of the algorithm follows the
chain
A * Z2 A × A E ,
where the second composition factor is the base group. We first want to
determine the irreducible representations of G := A * Z2 . Let G∗ be the
base group of G, i. e. G∗ = A × A. G∗ is a normal subgroup of G of index
2. Denoting by A = {χ1 , . . . , χk } the set of irreducible representations of A,
recall that the irreducible representations of G∗ are given by the set {χi ⊗χj :
i, j = 1, . . . , k} of pairwise tensor products (e.g. Sect. 5.6 of [212]).
Since G∗ G, the group G operates on the representations of G∗ via inner
conjugation. Because G is a semidirect product of G∗ with Z2 , we can write
each element g ∈ G as g = (a1 , a2 ; τ ) with a1 , a2 ∈ A and we conclude
(χ1 ⊗ χ2 )g = (χa1 1 ⊗ χa2 2 )τ = (χ1 ⊗ χ2 )τ ,
i.e. only the factor group G/G∗ = Z2 operates via permutation of the tensor
factors. The operation of τ is to map χ1 ⊗ χ2 → χ2 ⊗ χ1 .
Therefore, it is easy to determine the inertia groups (see [171, 235] for
definitions) Tρ of a representation ρ of G∗ . We have to consider two cases:
(a) ρ = χi ⊗ χi . Then Tρ = G, since permutation of the factors leaves ρ
invariant.
(b) ρ = χi ⊗ χj , i = j. Here we have Tρ = G∗ .

The irreducible representations of G∗ fulfilling (a) extend to representa-
tions of G, whereas the induction of a representation fulfilling (b) is irre-
ducible. In this case the restriction of the induced representation to G∗ is, by
Clifford theory, equal to the direct sum χ1 ⊗ χ2 ⊕ χ2 ⊗ χ1 .
Example 4.7. We consider the special case Wn := Zn2 * Z2 , for which the
quantum Fourier transforms have an especially appealing form.
Applying the design principles for Fourier transforms described in this
section (see also [171, 174, 226, 228, 249]), we obtain the circuits for DFTWn
in a straightforward way. Once we have studied the extension/induction be-
havior of the irreducible representations of G∗ , the recursive formula

(12 ⊗ DFTG∗ ) · Φ(t) · (DFTZ2 ⊗ 1|A|2 ) (4.9)
t∈T
provides a Fourier transform for G. Here Φ(t) denotes the extension (as a
whole) of the regular representation of G∗ to a representation of G [171, 226,
228]. In the case of Wn , the transform DFTG∗ is the Fourier transform for
Z2n
2 and therefore a tensor product of 2n Hadamard matrices.
a r H2
yn H2 r ❤ r ···
.. .. .. ..
. . . .
y1 H2 ··· r ❤ r
xn H2 ❤ r ❤···
.. .. .. ..
. . . .
x1 H2 ··· ❤ r ❤
Fig. 4.14. The Fourier transform for the wreath product Zn

2 Z2
The circuits for the case of Wn are shown in Fig. 4.14. Obviously, the
complexity cost of this circuit is linear in the number of qubits,since the
conditional gate representing the evaluation at the transversal t∈T Φ(t)
can be realized with 3n Toffoli gates.
Nonabelian Hidden Subgroups. We adopt the definition of the hidden-

subgroup problem given in [239]. The history of the hidden subgroup problem
parallels the history of quantum computing, since the algorithms of Simon [57]
and Shor [9] can be formulated in the language of hidden subgroups (see, e.g.,
[240] for this reduction) for certain abelian groups. In [206] exact quantum
algorithms (with a running time polynomial in the number of evaluations of

the given black-box function and the classical postprocessing) are given for
the hidden-subgroup problem in the abelian case. For the general case we use
the following definition, which should be compared with Simon’s problem in
Sect. 4.3.
Definition 4.5 (The Hidden Subgroup Problem). Let G be a finite
group and f : G → R a mapping from G to an arbitrary domain R fulfilling
the following conditions:
(a) The function f is given as a quantum circuit, i.e. f can be evaluated in
superpositions.
(b) There exists a subgroup U ⊆ G such that f takes a constant value on
each of the cosets gU for g ∈ G.
(c) Furthermore, f takes different values on different cosets.
The problem is to find generators for U .
We have already seen that Simon’s algorithm and Shor’s algorithms for the
discrete logarithm and factoring can be seen as instances of abelian hidden-
subgroup problems.
The hidden-subgroup problem for nonabelian groups provides a natural
generalization of these quantum algorithms. Interesting problems can be for-
mulated as hidden-subgroup problems for nonabelian groups, e.g. the graph
isomorphism problem, which is equivalent to the problem of deciding whether
a graph has a nontrivial automorphism group [241]. In this case G is the sym-
metric group Sn acting on a given graph Γ with n vertices. To reduce the
graph automorphism problem to a hidden-subgroup problem for Sn , we en-
code Γ into a binary string in R := {0, 1}∗ and define f to be the mapping
which assigns to a given permutation σ ∈ Sn the graph Γ σ . Progress in the
direction of the graph isomorphism problem has been made in [242], but an
efficient quantum algorithm solving the hidden-subgroup problem for Sn or
the graph isomorphism problem is still not known.
In [239], the case of hidden subgroups of dihedral groups is addressed.
The authors show that it is possible to solve the hidden-subgroup problem
using only polynomially many queries to the black-box function f . However,
the classical postprocessing takes exponential time in order to solve a non-
linear optimization problem. In [243] the hidden-subgroup problem for the
wreath products Zn2 * Z2 is solved on a quantum computer, using polyno-
mially many queries to f and efficient classical postprocessing which takes
O(n3 ) steps. The fast quantum Fourier transforms for these groups, which
have been derived in Sect. 5, have been used in this solution.
4.6.2 The Discrete Cosine Transform

In this section we address the problem of computing further signal transforms
on a quantum computer. More specifically, we give a realization of the discrete
cosine transforms (DCTs) of type II on a quantum computer, which is based

on a well-known reduction to the computation of a discrete Fourier transform
(DFT) of double length.
The DCT has numerous applications in classical signal processing, and
hence, this transform might be useful in quantum computing and quantum
state engineering as well. For efficient quantum signal transforms, and espe-
cially for wavelet transforms on quantum computers, see also [236, 244, 245,
246].
The discrete cosine transforms come in different flavors, varying slightly
in their definitions. In this paper we restrict ourselves to the discrete cosine
transform of type II (DCTII ) as defined below.
Recall that the DCTII is the N × N matrix defined by (see [247], p. 11)
1/2
2 i(j + 1/2)π
DCTII := ki cos ,
N N i,j=0,...,N −1
√
where ki = 1 for i = 1, . . . , N − 1 and k0 = 1/ 2.
In [247] the DCTII is shown to be a real and orthogonal (and hence uni-
tary) transformation. Closely related is the discrete cosine transform DCTIII ,
which is defined as the transposed matrix (and hence the inverse) of DCTII .
Therefore, each efficient quantum circuit for the DCTIII yields one for the
DCTII and vice versa, since the inverse transform is obtained by reading the
circuit backwards (where each elementary gate is conjugated and transposed).
Concerning the applications in classical signal processing, we should men-
tion the well-known fact that the DCTs (of all families) are asymptotically
equivalent to the Karhunen–Loève transform for signals produced by a first-
order Markov process. The DCTII is also used in the JPEG image compres-
sion standard [247].
For a given (normalized) vector |x = (x(0), . . . , x(N − 1))t , we want to
compute the matrix vector product DCTII · |x on a quantum computer effi-
ciently. Since DCTII is a unitary matrix, it has a factorization into elementary
quantum gates, and we seek a factorization of polylogarithmic length. Note
that we restrict ourselves to the case N = 2n since matrices of this size fit
naturally into the tensor product structure of the Hilbert space imposed by
the qubits.
Theorem 4.16. The discrete cosine transform DCTII (2n ) of length 2n can
be computed in O(n2 ) steps using one auxiliary qubit.
Proof. The main idea (following Chap. 4 of [247]) is to reduce the computation
of DCTII to a computation of a DFT of double length.
Instead of the input vector |x of length 2n , we consider |y of length
2n+1 , defined by
√
x(i)/ 2 √ , i = 0, . . . , 2n −1
y(i) := .
x(2n+1 − i − 1)/ 2 , i = 2n , . . . , 2n+1 −1
The input state of the quantum computer considered here, which has a
register of length n carrying |x and an extra qubit (which is initialized in
the ground state), is |ϕin = |0|x (written explicitly, this is x(0)|00 . . . 0 +
· · · + x(2n )|01 . . . 1), which is transformed by the circuit given in Fig. 4.15
to yield |y.
|0 H2 s
.. .. ..
|x . . πrev .
Fig. 4.15. Circuit that prepares the vector |y
As usual, H2 is the Hadamard transformation and πrev is the permutation

obtained by performing a σx on each √ wire.
Application of DFT2n+1 := (1/ 2n+1 )[ω2i·jn+1 ]i,j=0,...,2n+1 −1 , where ω2n+1
denotes a primitive 2n+1 th root of unity, to the vector |y yields the compo-
nents
n
2 −1
1 m·(2n −i−1)
(DFT2n+1 · y)(m) = √ (ω2m·i
n+1 + ω n+1
2 )x(i) ,
2n+2 i=0
which holds for m = 0, . . . , 2n+1 − 1.

Note that multiplication with DFT2n+1 can be performed in O(n2 ) steps
on a quantum computer [9, 248], which is quite contrary to the classical FFT
algorithm, which requires O(n2n ) arithmetic operations. However, the tensor
product recursion formula [171, 249] is well suited for direct translation into
efficient quantum circuits [203].
m/2
Multiplication of the vector component y(m) by the phase factor ω2n+1
yields
m/2 ! m(2n+1 −i−1) " m(i+1/2) m(2n+1 −i−1/2)
ω2n+1 ω2mi
n+1 + ω n+1
2 = ω2n+1 + ω2n+1 ,
which means that this expression is equal to 2 cos [m(i + 1/2)π/2n ]. The
multiplication with these (relative) phase factors corresponds to a diagonal
m/2
matrix T = 12 ⊗ diag(ω2n+1 , m = 0, . . . , 2n −1), which can be implemented
by a tensor product T = 12 ⊗ Tn ⊗ · · · ⊗ T1 of local operations, where Ti =
i−1
diag(1, ω22n+2 ) for i = 1, . . . , n.
Looking at the state obtained so far, we see that |ϕin has been mapped
to |ϕ , the lower 2n components of which are given by
n
2 −1
1
ϕ (m) = √ cos [m(i + 1/2)π/2n]x(i) ,
2n i=0
where m = 0, . . . , 2n+1 − 1. Using an elementary property of the cosine, we

see that the following holds for the components of this vector:
ϕ (m) = −ϕ (2n+1 − m), for m = 1, . . . , 2n − 1 ,

n
and, furthermore, ϕ (0) = 2i=0−1 x(i) and ϕ (2n ) = 0. Hence we are nearly
finished, since, except for cleaning up the help register, the x register has
been transformed according to DCTII .
Cleaning up the help register can be accomplished by the matrix
01 1 01 1
BB √1
2.
...
- √12
CC B
B
1
C
C
BB ..
CC B
√1 - √12
C
√1 - √12
B C
2
BB 2
CC = π −1
B
B
√1
2
√1
2 C
C
π,
B@ CA
1 1 ...
√1
2
√
2 @ √1 - √12 A
... ...
√1
2
√1
√1 √1 2 2
2 2
where the permutation matrix π arranges the columns 0, . . . , 2n+1 in such a

way that (i, 2n+1 −i+1) stand next to each other for i = 1, . . . , 2n −1. This can
be achieved by a quantum circuit, used also in [228], where this permutation
appeared in the reordering of irreducible representations according to the
action of a dihedral group. The circuit implementing π is given in Fig. 4.16
and can be performed in O(n2 ) operations. Here Pn is the cyclic shift x → x+
1 mod 2n on the basis states, which can be implemented in O(n2 ) operations
(see [228], Sect. 3).
❡
❉ ❉ ❡
❉ ... x → −x ... ❉ Pn ...
.
.. = . ..
... ... ❡
❉ ❉ ❡
❉ • ❉ • •
Fig. 4.16. Grouping the matrix entries pairwise via π
The matrix obtained after conjugation with π is a tensor product of the

form 12n ⊗ D1 , where
1 1 −1
D1 := √ ,
2 1 1
up to a multiplication with the block diagonal matrix

! 1 1 1 "
diag √ , 12n+1 −2 ,
2 −1 1
which in turn can be implemented by an (n−1)-fold controlled operation, i.e.

in O(n2 ) elementary gates [172].
Overall, we obtain the circuit for the implementation of a DCTII in O(n2 )

elementary gates shown in Fig. 4.17 (we have set D2 := D1−1 ).
We close this section with the remark that it is possible to derive other
decompositions of DCTII into a product of elementary quantum gates which
have the advantage of being in-place, i.e. the overhead of one qubit that the
realization given in the previous section needs can be saved [250].
This factorization of DCTII (2n ) of length 2n can also be computed in
O(n2 ) elementary operations and does not make use of auxiliary registers.
|0 H2 r ❢
❉
Tn ❉
..
❉.
DFT2n+1 . ..
|x ... ... πrev ... ❢
❉ ❢
T2 ❉
T1 ❉ •
❜ ❢ |0
❅
❅
..
.
Pn Pn−1 .. ..
. DCTII |x
❜ ❢ .
❜ ❢ ❅
❅
❅
❅
• D1 D2 • •
Fig. 4.17. Complete quantum circuit for DCTII using one auxiliary qubit
4.7 Quantum Error-Correcting Codes
4.7.1 Introduction
The class of quantum error-correcting codes (QECCs) provides a master ex-

ample that represents multiple aspects of the features of quantum algorithms.
They are quantum algorithms per se, showing the applicability of the algorith-
mic concepts described so far to an intrinsically important area of quantum
computation and quantum information theory, namely, the stabilization of
quantum states.
This is achieved by methods which are based on signal-processing tech-
niques at two levels:
• the Hilbert space of the quantum system itself

• the discrete vector space of the underlying combinatorial configurations.
The latter structure of finite-geometric spaces gives, furthermore, a deep

insight into the features of entanglement, as the error-correcting codes known
from classical coding theory share a particular feature, namely that of gen-
erating sets of codewords with maximal minimum distance by constructing
suitable geometric configurations. On the other hand, these configurations
nicely display the features of maximally entangled quantum states.
4.7.2 Background
We follow the presentation in [104], Chap. 13 for the background on the

classical theory. A classical error-correcting code (ECC) consists of a set of
binary words
x = (x0 , . . . , xn−1 )
of length n, where each “bit” xi ∈ {0, 1} can take a value xi ∈ GF (2) in the
finite field of two elements. With this elementary notion, the codewords can be
treated as vectors of the n-dimensional GF (2) vector space GF (2)n , where
the set of codewords is usually assumed to form a k-dimensional subspace
C ≤ GF (2)n . This can be obtained canonically as the range im(G) of a GF (2)
linear mapping G : GF (2)k → GF (2)n of the so-called encoder matrix, which
maps k-bit messages onto n-bit codewords, adding a redundancy of r = n − k
bits in the r so-called parity check bits.
The characteristic parameters of such a linear code C are the rate R = k/n
and the minimum Hamming weight
wH := min{wgtH (c) : c ∈ C, c = 0} .
The Hamming weight for a vector x = (x0 , . . . , xn−1 ) ∈ GF (2)n is defined

by wgtH (x) := |supp(x)|, where supp(x) := {i ∈ [0, . . . , n − 1] : xi = 0}.
It may be noted that the Hamming distance dH (u, v) := #{i ∈ [0, . . . , n−
1] : ui = vi }, which measures the number of bit flips necessary to change v
into u, is given by dH (u, v) = wgtH (u− v). Thus, since the code C is assumed
to be a linear subspace of GF (2)n , the equality for the minimum distance
dH (C) = minu=v dH (u, v) = wH is easily derived.
In constructing error-correcting codes, besides solving the parametric op-
timization problem of maximizing the rate of C and its minimum weight, it
is a challenge to provide an efficient decoding algorithm at the same time. A
decoder for C can in principle be built as follows:
Lemma 4.3. The dual code C ⊥ of C, which is the (n − k)-dimensional sub-

space of GF (2)n orthogonal to C with respect to the canonical GF (2) bilinear
pairing (see also Sect. 4.4.2), is generated by any matrix H : GF (2)n−k →
GF (2)n with im(H) = C ⊥ .
Obviously C = Ker(H t ), the kernel of the transpose of H, and G · H t = 0.

For the sake of simplicity we assume that each codeword c ∈ C is trans-
mitted through a binary symmetric channel (BSC) (Fig. 4.18), which is the
master model of a discrete memoryless channel [251]. A BSC is assumed to
add the error vector e ∈ GF (2)n independently of the codeword so that a
noisy vector u = c + e is received with probability q n−wgt(e) · pwgt(e) , where
q = 1 − p.
q = 1 − p✲
0 ◗ 0
✸
✑
◗ ✑
◗✑
p ✑◗ p
✑ ◗
1 ✑ q = 1 − p◗
s
✲ 1
Fig. 4.18. Binary symmetric channel with parameter p
From the syndrome
s := u · H t = c · H t + e · H t = e · H t
we see that the error pattern determines an affine subspace of GF (2)n , namely
a coset of C in GF (2)n . In order to allow error correction, the syndrome has
to be linked to the unknown error vector e uniquely, usually according to
the maximum-likelihood decoding principle. Obviously, for the BSC this can
be achieved by finding the unique codeword c that minimizes the Hamming
distance to u. For combinatorial reasons this is possible if
w − 1
H
wgt(e) ≤ .
2
Thus a code with minimum Hamming weight wH = 2t + 1 is said to correct
t errors per codeword.
4.7.3 A Classic Code
An example which today can be considered a classic in the field of science and
technology is the application and design of (first-order) Reed–Müller codes
RM (1, m). This is not merely because they were a decisive piece of discrete
mathematics in producing the first pictures from the surface of Mars in the
early 1970s after the landing of Mariner 9 on 19 January 1972.
The description of their geometric construction and the method of er-
ror detection/correction of the corresponding wave functions, which we have
taken from [104], Chap. 13, is quite similar to the concept of quantum error-
correcting codes. In this section we describe Reed–Müller codes in a natural
geometrical setting (cf. [103, 104, 252]).
m
In the first-order Reed–Müller code RM (1, m), the codewords f ∈ GF (2)2
are GF (2)-linear combinations of class functions of the index-2 subgroups
H < Zm 2 and their cosets. From this notion, a natural transform into the
orthonormal basis of Walsh–Hadamard functions [253, 254] is given by the
characters of Zm
2 , i.e. the Hadamard transformation H2m (see also Sect. 4.2.1
and [255]). By these means, the GF (2) vector
m
f = (f (u))u∈Zm ∈ GF (2)2
2
is converted into the real (row) vector

F = (−1)f (u) m u∈Z2
by replacing an entry 0 in f by an entry 1 in F and an entry 1 in f by

an entry −1 in F . Modulation into a wavefunction F (t) is then achieved by
transmitting the step function in the interval [0, 2m − 1] defined by
m
2 −1
F (t) = FBinary(i) 1[i,i+1] (t) .
i=0
Note that for the first Hadamard coefficient Fˆ0 of F (t), which is given by
$ 2m
#
F0 = F (t) dt ,
0
the following identity holds:
F#0 = 2m − 2 wgt(f ) .
This provides a beautiful maximum-likelihood decoding device similar to that
needed for quantum codes.
We shall illustrate this with the example of the first-order Reed–Müller
code RM (1, 3).
The codewords of the first-order Reed–Müller code RM (1, m) with m = 3,
of length 8, can be regarded as incidence vectors of special subsets of points of
AG(3, 2) (see Fig. 4.19). The following table of Boolean functions (see Sect.
4.2.2) fi (x3 , x2 , x1 ) = 1 ⊕ xi of incidence vectors,
1 = (1, 1, 1, 1, 1, 1, 1, 1) corresponding to the whole space ,

f1 = (1, 0, 1, 0, 1, 0, 1, 0) corresponding to the 2-subspace (∗, ∗, 0) ,
(4.10)
f2 = (1, 1, 0, 0, 1, 1, 0, 0) corresponding to the 2-subspace (∗, 0, ∗) ,
f3 = (1, 1, 1, 1, 0, 0, 0, 0) corresponding to the 2-subspace (0, ∗, ∗) ,
thus defines an (m + 1) × 2m generator matrix G. Its range is the following
set of 16 codewords:
(0, 0, 0, 0, 0, 0, 0, 0) (1, 1, 1, 1, 1, 1, 1, 1) (0, 0, 0, 0, 1, 1, 1, 1) (1, 1, 1, 1, 0, 0, 0, 0)
(0, 0, 1, 1, 0, 0, 1, 1) (1, 1, 0, 0, 1, 1, 0, 0) (0, 1, 0, 1, 0, 1, 0, 1) (1, 0, 1, 0, 1, 0, 1, 0)
(0, 0, 1, 1, 1, 1, 0, 0) (1, 1, 0, 0, 0, 0, 1, 1) (0, 1, 0, 1, 1, 0, 1, 0) (1, 0, 1, 0, 0, 1, 0, 1)
(0, 1, 1, 0, 0, 1, 1, 0) (1, 0, 0, 1, 1, 0, 0, 1) (0, 1, 1, 0, 1, 0, 0, 1) (1, 0, 0, 1, 0, 1, 1, 0) .
P3 = 011
P5 = 101
P6 = 110
P 0 = 000
Fig. 4.19. Reed–Müller code RM (1, 3) interpreted as a 2-flat in AG(3, 2)
Thus the subset S = {P0 , P3 , P5 , P6 } determines the incidence vector

(1, 0, 0, 1, 0, 1, 1, 0), which is the codeword f 1 + f 2 + f 3 of RM (1, 3). The
geometric interpretation of this codeword as an (m − 1)-flat in AG(3, 2) is
shown in Fig. 4.19.
With respect to G, the signal function F transmitted (see Fig. 4.20) thus
corresponds to the (m + 1)-bit input vector (0, 1, 1, 1).
F(t)
1
t
-1
Fig. 4.20. Transmitted signal corresponding to the codeword f 0 + f 1 + f 2 =
(1, 0, 0, 1, 0, 1, 1, 0) with and without noise
In the more general case of Reed–Müller codes RM (1, m) and their modu-
lated signal functions, noise e(t) is added to the signal function by the channel
during transmission, so that the receiver will detect only a signal
ψ(t) = F (t) + e(t) .
The behavior of the RM (1, m) demodulator–decoder device can be sket-

ched on the basis of the underlying geometry as follows.
Transformation of the received signal ψ(t) into the orthonormal basis of
Walsh–Hadamard functions Wu (t) of order m is achieved by computing the
2m scalar products
#
ψ(u) = Wu | ψ ,
where, for u ∈ GF (2)m , Wu (t) is given by

m

2
Wu (t) = (−1)u·Binary(j−1) · 1[j−1,j] (t) ,
j=1
corresponding to the uth row of the Hadamard matrix H2n .

As the received signal function is represented by a sampling vector ψ =
F + e, after the Walsh–Hadamard transformation one obtains ψ # = ψ ·H2m =
F H2m +eH2m . We estimate a maximum-likelihood signal function as follows.
From F # = F · H2m , where

F# (u) = (−1)u·v F (v) = (−1)u·v+f (v) mod2 ,
v v
we obtain the identity
|2m − F#(u)| = 2 wgt(u⊥ ⊕ f ) ,
since (u·v)v∈GF (2)m is the incidence vector u⊥ of the hyperspace of AG(2, m)

orthogonal to the vector u ∈ GF (2)m . As the RM (1, m) codes consist of all
incidence vectors of such hyperspaces or their cosets (i.e. the complement),
a minimum distance decoder is realized as shown in Fig. 4.21.6
φ̂
Signal Sampling ψ ✲ Maximiser ✲ Decon- Message
✲ at rate ✲ H2m De- ✲
|ψ(t) verter coder M
1/2m ✲
φˆ0
Fig. 4.21. Minimum-distance decoder for Reed–Müller codes (cf. [104], Chap. 13)
Here the maximizer computes
x ∈ GF (2)m such that |ψ(x)| = max |ψ(u)| .

u∈GF (2)m
With the overall ± parity given by the sign (ψ#0 ) = (−1)ε(ψ) , the deconverter
produces a most likely codeword f = ε(ψ) · 1 + x⊥ . If the enumeration of
the generating hyperspaces ui (cf. (4.10)) is chosen suitably, the decoder
therefore reproduces the (m + 1)-bit message n = (ε, x1 , . . . , xm ), which is a
maximum-likelihood estimator of the word originally encoded into f = M ·G.
The reader is urged to compare this decoding algorithm with the quantum
decoding algorithm described in Sect. 4.7.4.
6
With the fast Hadamard transform algorithm (see Sect. 4.2.1), this decoder re-
quires O(m2m ) computational steps. Note that this is one of the earliest commu-
nication applications of generalized FFT algorithms (cf. [171]).
4.7.4 Quantum Channels and Codes
One of the basic features described so far in this contribution has been the use
of quantum mechanical properties (entanglement, superposition) to speed up
the solution of classical problems. Most notably, this basic idea of quantum
computing has been present in the complex of hidden-subgroup problems
(see Sect. 4.3, 4.5.2 and 5), where the problem was to identity an unknown
subgroup of a given group out of exponentially many candidates, given a su-
perposition over a coset of the unknown subgroup. In this section we shall take
the dual point of view: we construct states which are simultaneous eigenstates
of a suitably chosen subgroup of a fixed error group and have the additional
property that an element of an unknown coset of this group – this models an
error which happens to the states – can be identified and also corrected.
To construct these states, we rely on the powerful theory of classical ECC
[184] introduced in the preceding sections. In what follows, we shall introduce
the class of binary QECCs, the so-called CSS codes, referring to the elaborate
article [255] by Beth and Grassl. These codes were independently discovered
by Calderbank and Shor [82] and Steane [62].
Quantum Channels. Before describing the construction of these codes,
we shall loosely describe the similarities and differences between a classical
binary symmetric channel (BSC) (see Sect. 4.7.2, Fig. 4.18) and a quantum
channel (QC).
Much as in the idealized case of a BSC, where vectors c ∈ GF (2)n are
n
transmitted, we shall consider the QC to be a carrier of kets |ψ ∈ H2n = C 2
n
spanned by the basis kets |x, where x ∈ GF (2) . In this system, so-called
error operations, generated by local errors as bit-flip errors, and sign-flip
errors can occur in superposition. Similarly to the case of a BSC, where the
error group is isomorphic to (GF (2)n , ⊕) = (e1 , . . . , en ) : ei ∈ {0, 1}, i =
1, . . . , n, in QC the error group
E = e1 ⊗ . . . ⊗ en : ei ∈ {id, σx , σz }, i = 1, . . . , n} ,
generated by the local bit flips and phase flips, represents all possible error
operators in the quantum channel by the transition diagram described below.
Initially, the input wavefunction |ψ will interact with the environment
via |ψ → |ψ| through a modulator; within the channel, this waveform
evolves under the error group according to the channel characteristics,
channel -
|ψ| ✲ (γ|ψ)|γ , (4.11)
error γ∈E
whereas, in the “environment”, ancillae tacitly contain probability amplitudes

for the occurrence of group elements.
Much as in the BSC, where only errors e ∈ GF (2)n with a given maximal
weight wgt(e) ≤ t are allowed or assumed, in the QC the sum of the received
ket

|µ = γ|ψ|γ (4.12)
γ∈E
wgt(γ)≤t
will range only over those group elements γ which are products of at most
t local (single-bit) errors. Much in accordance with the classical case, for
γ = e1 ⊗ . . . ⊗ en ∈ E we define wgt(γ) to be the number of occurrences of
σx and σz needed to generate γ.
In order to protect quantum states against errors of this kind in a quan-
tum channel, a so-called Pauli channel, a quantum error-correcting code must
be constructed to map original quantum states into certain “protected” or-
thogonal subspaces, so that errors of the type of (4.12) cannot in practice
damage the original state seriously. For this purpose, the theory of classical
codes for the BSC can be successfully applied, as we describe below.
Quantum Codes. The basic principle is to consider encoded quantum states
which are superpositions of basis vectors belonging to classical codes, e.g.
1
|C := |c .
|C| c∈C
First we note a surprising fact expressing an old result of error-correcting

codes directly in the language of quantum theory.
Lemma 4.4 (MacWilliams). Let C < GF (2)n be an error-correcting code

with its dual as usual. Let A be the elementary abelian group (GF (2)n , ⊕).
Then DFTA is the Hadamard transformation (see Sect. 4.4.2) H2n := H2 ⊗
. . . ⊗ H2 (n factors). For each error vector e ∈ GF (2)n ,

1 1
H2n |C = H2n |c ⊕ e = (−1)c·e |c . (4.13)
|C| c∈C |C ⊥ | c∈C ⊥
Proof. This result is due to the identity

1 1
DFTA |c ⊕ e = β(c, e)|c , (4.14)
|U | c∈U |U ⊥ | c∈U ⊥
which holds for all subgroups U of an abelian group A (see Sect. 4.4.2 and
4.4.3).

The lemma says that any bit-flip error applied to the state |C will give
a state whose support C ⊥ is translation invariant, the shift being expressed
only in the phases of the elements of C ⊥ . Dually, a phase-flip error in |C will
occur as a bit-flip error in |C ⊥ .
From this, we deduce the following basic coding principle ([255], p. 462):
given a classical binary linear (n, k) code C of length n and dimension k, the
states of the related quantum code are given by
1
|ψwi = (−1)c·wi |c , (4.15)
|C| c∈C
with suitable w i ∈ GF (2)n encoding the ith basis ket of the initial state
to be protected. In addition to the choice of the classical code C, for the
construction of a binary quantum code a subset W ⊆ GF (2)n /C ⊥ has to be
given to define the vectors wi .
Example 4.8. Let C := {(0, 0, 0), (1, 1, 1)} ⊆ GF (2)3 be the dual of the Reed–
Müller code R = RM (1, 3) shown in Fig. 4.19. Here W := GF (2)3 /C ⊥ pro-
vides an appropriate choice,
w 0 = (0, 0, 0), w1 = (1, 1, 1) ,
for the following encoding:
|0 → |ψw0 = |0, 0, 0 + |1, 1, 1 ,

(4.16)
|1 → |ψw1 = |0, 0, 0 − |1, 1, 1 .
Note that the state |0, 0, 0 + |1, 1, 1 is the maximally entangled GHZ state.
We remark that the GHZ state is, up to local unitary transformations, the
unique maximally entangled state [256]. The “protected” subspace of code
vectors is, by definition,
H0 = {|ψ = α|ψw0 β|ψw 1 | |α|2 + |β|2 = 1} .
Since C is a one-error-correcting binary code, the quantum code is endowed

with this property with respect to single bit-flip errors, i.e. it is protected
against the error operators
ε3 = id ⊗ id ⊗ σx , ε2 = id ⊗ σx ⊗ id, ε1 = σx ⊗ id ⊗ id ∈ U(8) .
Obviously, the subspaces H0 and Hi = εi H0 (i = 1, 2, 3) are mutually or-

thogonal, so that
3
3
H2 = Hi
i=0
is the direct sum of the four orthogonal spaces
H0 = |0, 0, 0 + |1, 1, 1, |0, 0, 0 − |1, 1, 1 ,

H1 = |1, 0, 0 + |0, 1, 1, |1, 0, 0 − |0, 1, 1 ,
H2 = |0, 1, 0 + |1, 0, 1, |0, 1, 0 − |1, 0, 1 ,
H3 = |0, 0, 1 + |1, 1, 0, |0, 0, 1 − |1, 1, 0 .
Thus, for any linear combination E of these single bit-flip errors i , the en-
coded state |ψ = α|ψw0 + β|ψw1 is represented by

3
E|ψ = λi (αi |ψw0 + βi |ψw1 ) ,
i=0
as a direct sum of four orthogonal vectors, each being “proportional” to |ψ.

So, by a measurement, i.e. a random projection onto any of the spaces Hi , or
by applying a conditional gate U = diag(†i : i = 0, . . . , 3), the original state
can be reconstituted, thus correcting up to one bit-flip error.
But note that if, instead of C, the code R = C ⊥ had been selected, this
code construction could not have been successful, as R = RM (1, 3) can
detect one error but not correct it. It can be seen (see [255], p. 463) that the
corresponding quantum code inherits this property of detecting one phase-
error but not being capable of correcting it.
Motivated by and starting from this example and the properties and prob-
lems derived from it, we now quote the following theorem, which provides a
method to obtain quantum codes from classical codes.
Theorem 4.17 (CSS Codes). Suppose C1 , C2 ⊆ GF (2)n are classical bi-
nary codes with parameters (n1 , k1 , d1 ) and (n2 , k2 , d2 ), respectively, fulfilling
the additional requirement C1⊥ ⊆ C2 . Let W := {wi : i = 1, . . . , [C2 : C1⊥ ]} be
a system of representatives of the cosets C2⊥ /C1 . Then the set of states
1
|φwi := (−1)c·wi |c
|C1 | c∈C1
forms a quantum code Q which can correct (d1 −1)/2 bit errors and (d2 −1)/2
phase errors.
In practice, we have the following corollary in the case of weakly self-dual
codes C ⊆ C ⊥ , which were shown to be as good asymptotically in [82]:
Theorem 4.18. Let C be a weakly self-dual binary code with dual distance d.
Then the corresponding quantum code is capable of correcting up to (d − 1)/2
errors.
The construction in this theorem can be made more general, as Beth and
Grassl have shown in [255]:
Theorem 4.19. Let C ⊥ be a weakly self-dual binary code. If for
M0 := {w ∈ GF (2)n /C ⊥ | dH (C ⊥ , C ⊥ + w) ≤ t}
the following condition,
∀w i , w j : i = j ⇒ M0 ∩ (M0 + (w i − wj )) = ∅ , (4.17)
is satisfied, the quantum code can correct t errors, i.e. any error operator
ε ∈ E where the total number of positions exposed to bit flips or sign flips is
at most t.
This leads to the following decoding algorithm.
Algorithm 6 (Quantum Decoding Algorithm) Let |φ be encoded by a

QECC according to Theorems 4.17–4.19. The received vector E|φ will be
decoded and reconstructed by the following steps:
• Perform a measurement to determine the bit-flip errors, i.e. project onto
the code space HC or one of its orthogonal images HC+e under a bit-flip
error e.
• Correct this bit-flip error (“subtract” e) by applying the corresponding
tensor product of σx operators.
• Perform a Hadamard transformation.
• Perform a measurement to determine the sign-flip errors which corre-
sponds to a bit-flip error in the actual bases.
• Correct this error.
• Reencode the final state.
This decoding algorithm is easily understood from the point of view of

binary codes, which have been designed as general constructions for the BSC
(see Sect. 4.7.2). The reader should also observe the stunning analogy between
this quantum decoder and the Green-machine decoder described in Sect.
4.7.3, Fig. 4.21.
4.8 Conclusions
We have presented an introduction to a computational model of quantum

computers from a computer science point of view. Discrete Fourier transforms
have been introduced as important subroutines used in several quantum al-
gorithms. Throughout, unitary transformations which can be implemented
in terms of elementary gates using a quantum circuit of polylogarithmic size
have been of special interest; they yield an exponential speedup compared
with the classical situation in many cases.
The quantum algorithms presented here exploit the fundamental princi-
ples of interference, superposition and entanglement that quantum physics
offers. We have explored these principles in various algorithms, ranging from
Shor’s algorithms to algorithms for quantum error-correcting codes.
5 Mixed-State Entanglement and Quantum
Communication
Michal Horodecki, Pawel Horodecki and Ryszard Horodecki
5.1 Introduction
Quantum entanglement is one of the most striking features of the quantum
formalism [26]. It can be expressed as follows: If two systems interacted in
the past it is, in general, not possible to assign a single state vector to either
of the two subsystems [257]. This is what is sometimes called the principle of
nonseparability. A common example of an entangled state is the singlet state
[258],
1
ψ− = √ (|01 − |10) . (5.1)
2
One can see that it cannot be represented as a product of individual vectors
describing states of subsystems. Historically, entanglement was first recog-
nized by Einstein, Podolsky and Rosen (EPR) [24] and by Schrödinger [5].1
In their famous paper, EPR suggested a description of the world (called “local
realism”) which assigns an independent and objective reality to the physi-
cal properties of the well-separated subsystems of a compound system. Then
EPR applied the criterion of local realism to predictions associated with an
entangled state to conclude that quantum mechanics was incomplete. The
EPR criticism was the source of many discussions concerning fundamental
differences between the quantum and classical descriptions of nature.
The most significant progress toward the resolution of the EPR problem
was made by Bell [25], who proved that local realism implies constraints on
the predictions of spin correlations in the form of inequalities (called Bell’s
inequalities) which can be violated by quantum mechanical predictions for
a system in the state (5.1). The latter feature of quantum mechanics, usu-
ally called nonlocality, is one of the most evident manifestations of quantum
entanglement.
Information-theoretical aspects of entanglement were first considered by
Schrödinger, who wrote [5], in the context of the EPR problem,
“Thus one disposes provisionally (until the entanglement is resolved by actual obser-
vation) of only a common description of the two in that space of higher dimension.
1
In fact, entangled quantum states had been used in investigations of the proper-
ties of atomic and molecular systems [259].

152 Micha&l Horodecki, Pawe&l Horodecki and Ryszard Horodecki
This is the reason that knowledge of the individual systems can decline to the scant-
iest, even to zero, while that of the combined system remains continually maximal.
Best possible knowledge of a whole does not include best possible knowledge of its
parts – and that is what keeps coming back to haunt us”.
In this way Schrödinger recognized a profoundly nonclassical relation between
the information which an entangled state gives us about the whole system
and the information which it gives us about the subsystems.
The recent development of quantum information theory has shown that
entanglement can have important practical applications (see, e.g., [1, 2, 3,
260]). In particular, it turns out that entanglement can be used as a resource
for communication of quantum states in an astonishing process called quan-
tum teleportation [261].2 In the latter, a quantum state is transmitted by use
of a pair of particles in a singlet state (5.1) shared by the sender and receiver
(usually referred to as Alice and Bob), and two bits of classical communica-
tion. However, in real conditions, owing to interaction with the environment,
called decoherence, we encounter mixed states rather than pure ones. These
mixed states can still possess some residual entanglement. More specifically,
a mixed state is considered to be entangled if it is not a mixture of prod-
uct states [263]. In mixed states the quantum correlations are weakened, and
hence the manifestations of mixed-state entanglement can be very subtle
[263, 264, 265]. Nevertheless, it appears that it can be used as a resource for
quantum communication. Such a possibility is due to the discovery of distil-
lation of entanglement [266]: by manipulation of noisy pairs, involving local
operations and classical communication, Alice and Bob can obtain singlet
pairs and apply teleportation. This procedure provides a powerful protection
of the quantum data transmission against the environment.
Consequently, the fundamental problem was to investigate the structure
of mixed-state entanglement, especially in the context of quantum commu-
nication. These investigations have led to discovery of discontinuity in the
structure of mixed-state entanglement. It appears that there are at least two
qualitatively different types of entanglement [71]: free, which is useful for
quantum communication, and bound, which is a nondistillable, very weak
and mysterious type of entanglement.
The present contribution is divided into two main parts. In the first part
we report results of an investigation of the mathematical structure of entan-
glement. The main question is: given a mixed state, is it entangled or not? We
present powerful tools that allow us to obtain the answer in many interesting
cases. A crucial role is played here by the connection between entanglement
and the theory of positive maps [267]. In contrast to completely positive maps
[88], positive maps have not been applied in physics so far. The second part
is devoted to the application of the entanglement of mixed states to quantum
communication. Now, the leading question is: given an entangled state, can
it be distilled? The mathematical tools worked out in the first part allow us
2
For experimental realizations, see [262].
Mixed-State Entanglement and Quantum Communication 153
to answer the question. Surprisingly, the answer does not simplify the picture
but, rather, reveals a new horizon including the basic question: what is the
role of bound entanglement in nature?
Since entanglement is a basic ingredient of quantum information theory,
the scope of application of the research presented here goes far beyond the
quantum communication problem. The insight into the structure of entangle-
ment of mixed states can be helpful in many subfields of quantum information
theory, including quantum computing, quantum cryptography, etc.
Finally, it must be emphasized that our approach will be basically qual-
itative. Thus we shall not review here the beautiful work performed in the
domain of quantifying entanglement [66, 268, 269, 270, 271] (we shall only
touch on this subject in the second part). Owing to the limited space for
the present contribution, we shall also restrict our considerations to the en-
tanglement of bipartite systems, even though a number of results have been
recently obtained for multipartite systems (see, e.g., [272, 273]).
5.2 Entanglement of Mixed States: Characterization
We shall deal with states on the finite-dimensional Hilbert space HAB =

HA ⊗ HB . We shall call the system described by the Hilbert space HAB the
n ⊗ m system, where n and m are the dimensions of the spaces HA and HB ,
respectively. An operator acting on H is a state if Tr = 1 and if it is a
positive operator, i.e.
Tr P ≥ 0 (5.2)
for any projectors P (equivalently, positivity of an operator means that it is

Hermitian and has nonnegative eigenvalues).
A state acting on the Hilbert space HAB is called separable3 if it can be
approximated in the trace norm by states of the form
k

= pi i ⊗ ˜i , (5.3)
i=1
where i and ˜i are states on HA and HB , respectively. In finite dimensions

one can use a simpler definition [274] (see also [269]): is separable if it is of
the form (5.3) for some k (one can always find a k ≤ dim HAB 2
). Note that
the property of being entangled or not does not change if one subjects the
state to a product unitary transformation → = U1 ⊗ U2 U1† ⊗ U2† . We
call the states and equivalent.
3
The definition of separable states presented here is due to Werner [263], who
called them classically correlated states.
We shall further need the following maximally entangled pure state of the
d ⊗ d system:
d
d 1
ψ+ = √ |i ⊗ |i . (5.4)
d i=1
We shall denote the corresponding projector by P+d (the superscript d will

usually be omitted). Then, for any state , the quantity F =
ψ+ ||ψ+ is
called the singlet fraction.4 In general, by maximally entangled states we shall
mean vectors ψ that are equivalent to ψ+ :
ψ = U1 ⊗ U2 ψ+ ,
where U1 , U2 are unitary transformations. The most common two-qubit max-

imally entangled state is the singlet state (5.1). One can define the fully
entangled fraction of a state of the d ⊗ d system by
F () = max
ψ||ψ , (5.5)
ψ
where the maximum is taken over all maximally entangled vectors of the d⊗d
system.
5.2.1 Pure States
If is a pure state, i.e. = |ψ

ψ|, then it is easy to check if it is entangled
or not. Indeed, the above definition implies that it is separable if and only
if ψ = ψA ⊗ ψB , i.e. if either of its reduced density matrices is a pure state.
Thus it suffices to find eigenvalues of either of the reductions. Equivalently,
one can refer to the Schmidt decomposition [275] of the state. As one knows,
for any pure state ψ there exist bases {eA B
i }, {ei } in the spaces HA and HB
such that
k

ψ= ai |eA B
i ⊗ |ei , k ≤ dim HAB , (5.6)
i=1
where the positive coefficients ai are called Schmidt coefficients. The state
is entangled if at least two coefficients do not vanish. One finds that the
positive eigenvalues of either of the reductions are equal to the squares of
the Schmidt coefficients. In the next section we shall introduce a series of
necessary conditions for the separability for mixed states. It turns out that
all of them are equivalent to separability in the case of pure states [276, 277].
4
In fact, the state ψ+ used in the definition of the singlet fraction is a local
transformation of the true singlet state. Nevertheless, we shall keep the name
“singlet fraction”, while using the state ψ+ , which is more convenient for technical
reasons.
5.2.2 Some Necessary Conditions for Separability

of Mixed States
A condition that is satisfied by separable states will be called a separability

criterion. If a separability criterion is violated by the state, the state must
be entangled. It is important to have strong separability criteria, i.e. those
that are violated by the largest number of states if possible.
Since violation of the Bell inequalities is a manifestation of quantum en-
tanglement, a natural separability criterion is constituted by the Bell inequal-
ities. In [263] Werner first pointed out that separable states must satisfy all
possible Bell inequalities.5 The common Bell inequalities derived by Clauser,
Horne, Shimony and Holt (CHSH) are given by [27]
Tr B ≤ 2 , (5.7)
where the Bell–CHSH observable B is given by
B = âσ ⊗ (b̂ + b̂ )σ + â σ ⊗ (b̂ − b̂ )σ . (5.8)

3
Here â, â , b̂, b̂ are arbitrary unit vectors in R , âσ = i=1 ai σi , and σi are
3
the Pauli matrices. For any given set of vectors we have a different inequality.
In [278] we derived the condition for a two-qubit6 state that was equivalent
to satisfying all the inequalities jointly. This condition has the following form:
M () ≤ 1 , (5.9)
where M is constructed in the following way. One considers the 3 × 3 real

matrix T with entries Tij ≡ Tr σi ⊗ σj . Then M is equal to the sum of the
two greater eigenvalues of the matrix T † T . This condition characterizes states
violating the most common, and so far the strongest, Bell inequality for two
qubits (see [279] in this context). While it is interesting from the point of
view of nonlocality, it appears to be not a very strong separability criterion.
Indeed, there exists [263] a large class of entangled states that satisfy all
standard Bell inequalities.7
Another approach originated from the observation by Schrödinger [5] that
an entangled state gives us more information about the total system than
about subsystems. This gave rise to a series of entropic inequalities of the
form [277, 280]
S(A ) ≤ S(), S(B ) ≤ S() , (5.10)

5
In [263] Werner also provided a very useful criterion based on the so-called flip
operator (see Sect. 5.2.4).
6
A qubit is the elementary unit of quantum information and denotes a two-level
quantum system (i.e. a 2 ⊗ 2 system) [69].
7
See [264, 265, 283] in the context of more sophisticated nonlocality criteria.
where A = TrB , and similarly for B . The above inequalities were proved
[277, 280, 281, 282] to be satisfied by separable states for four different en-
tropies that are particular cases of the Renýi quantum entropies
Sα = (1 − α)−1 log Tr α :
S0 = log R() ,
S1 = −Tr log ,
S2 = − log Tr 2 ,
S∞ = − log |||| , (5.11)
where R() denotes the rank of the state (the number of nonvanishing
eigenvalues). The above inequalities are useful tools in many cases (as we
shall see in Sect. 5.3.5, one of them allows us to obtain a bound on the
possible rank of the bound entangled states); still, however, they are not very
strong criteria.
A different approach, presented in [66], is based on local manipulations of
entanglement (this approach was anticipated in [265]). The main idea is the
following: a given state is entangled because parties sharing many systems
(pairs of particles) in this state can produce a smaller number of pairs in a
highly entangled state (of easily “detectable” entanglement) by local oper-
ations and classical communication (LOCC). This approach initiated a new
field in quantum information theory: manipulating entanglement. The sec-
ond part of this contribution will be devoted to this field. It also initiated the
subject of the quantification of entanglement. Still, however, the seemingly
simple qualitative question of whether a given state is entangled or not was
not solved.
A breakthrough was achieved by Peres [284], who derived a surprisingly
simple but very strong criterion. He noted that a separable state remains a
positive operator if subjected to partial transposition (PT). We will call this
the positive partial transposition (PPT) criterion.
To define partial transposition, we shall use the matrix elements of a state
in some product basis:
mµ,nν =
m| ⊗
µ| |n ⊗ |ν , (5.12)
where the kets with Latin and Greek letters form an orthonormal basis in the
Hilbert space describing the first and the second system, respectively. Hence
the partial transposition of is defined as
T
mµ,nν ≡ mν,nµ .
B
(5.13)
The form of the operator TB depends on the choice of basis, but its eigenval-
ues do not. We shall say that a state “is PPT” if TB ≥ 0; otherwise we shall
say that the state “is NPT”. The partial transposition is easy to perform in
matrix notation. The state of the m ⊗ n system can be written as
 
A11 ... A1m
=  ... ... ...  , (5.14)
Am1 ... Amm
n
with n × n matrices Aij acting on the second (C ) space. These matrices are
defined by their matrix elements {Aij }µν ≡ iν,jµ . Then the partial transpo-
sition can be realized simply by transposition (denoted by T) of all of these
matrices, namely
 T 
A11 ... AT
1m
TB =  ... ... ...  . (5.15)
ATm1 ... Amm
T
Now, for any separable state , the operator TB must have still non-
negative eigenvalues [284]. Indeed, consider a partially transposed separable
state

TB = pi i ⊗ (˜
i )T . (5.16)
i
Since the state ˜i remains positive under transposition, so does the total
state.
Note that what distinguishes the Peres criterion from the earlier ones is
that it is structural. In other words, it does not say that some scalar function
of a state satisfies some inequality, but it imposes constraints on the structure
of the operator resulting from PT. Thus the criterion amounts to satisfying
many inequalities at the same time. In the next section we shall see that there
is also another crucial feature of the criterion: it involves a transposition that
is a positive map but is not a completely positive one. This feature, abstracted
from the Peres criterion, allows us to find an intimate connection between
entanglement and the theory of positive maps.
Finally, it should be mentioned that necessary conditions for separability
have been recently obtained in the infinite-dimensional case [285, 286]. In
particular, the Peres criterion was expressed in terms of the Wigner repre-
sentation and applied to Gaussian wave packets [286].
5.2.3 Entanglement and Theory of Positive Maps
To describe the very fruitful connection between entanglement and positive

maps we shall need mathematical notions such as positive operators, posi-
tive maps and completely positive maps. In the following section we establish
these notions. In the subsequent sections we use them to develop the charac-
terization of the set of separable states.
Positive and Completely Positive Maps. We start with the following no-
tation. By AA and AB we shall denote the set of operators acting on HA and
HB , respectively. Recall that the set A of operators acting on some Hilbert
space H constitutes a Hilbert space itself (a so-called Hilbert–Schmidt space)
with a scalar product
A, B = Tr A† B. One can consider an orthonormal
H
basis of operators in this space given by {|i
j|}dimi,j=1 , where |i is a basis
in the space H. Since we are dealing with a finite dimension, A is in fact a
space of matrices. Hence we shall sometimes denote it by Md , where d is the

dimension of H.
The space of the linear maps from AA to AB is denoted by L(AA , AB ).
We say that a map Λ ∈ L(AA , AB ) is positive if it maps positive operators in
AA into the set of positive operators, i.e. if A ≥ 0 implies Λ(A) ≥ 0. Finally,
we need the definition of a completely positive (CP) map. We say [88] that
a map Λ ∈ L(AA , AB ) is completely positive if the induced map
Λn = Λ ⊗ In : AA ⊗ Mn → AB ⊗ Mn (5.17)
is positive for all n; here In is the identity map on the space Mn .8 Thus
the tensor product of a CP map and the identity maps positive operators
into positive ones. An example of a CP map is → W W † , where W is an
arbitrary operator. As a matter of fact, the general form of a CP map is

Λ() = Wi Wi† . (5.18)
i
CP maps that do not increase the trace (Tr Λ() ≤ Tr ) correspond to

the most general physical operations allowed by quantum mechanics [88].
If Tr Λ() = Tr for any (we say the map is trace-preserving), then the
operation can be performed with probability 1; otherwise, it can be performed
with probability p = Tr Λ().
It is remarkable that there are positive maps that are not CP: an example
is just the transposition mentioned in the previous section. Indeed, if is
positive, then so is T , because
Tr T P = Tr P T ≥ 0 (5.19)
and P T is still some projector. We have used here the fact that Tr AT = Tr A.
On the other hand, I ⊗ T is no longer positive. One can easily check this,
showing that (I ⊗ T )P+ ≡ P+TB is not a positive operator.
A positive map is called decomposable [287] if it can be represented in the
form
Λ = Λ1CP + Λ2CP ◦ T , (5.20)
where ΛiCP are some CP maps. For low-dimensional systems (Λ : M2 → M2 or

Λ : M3 → M2 ) the set of positive maps can be easily characterized. Namely,
it has been shown [288, 289] that all the positive maps are decomposable in
this case. If, instead, at least one of the spaces is Mn with n ≥ 4, there exist
nondecomposable positive maps [287, 289] (see the example in Sect. 5.2.4).
No full characterization of positive maps has been worked out so far in this
case.
8
Of course, a completely positive map is also a positive one.
Characterization of Separable States via Positive Maps. The fact

that complete positivity is not equivalent to positivity is crucial for the prob-
lem of entanglement that we are discussing here. Indeed, trivially, the product
states are mapped into positive operators by the tensor product of a posi-
tive map and an identity: (Λ ⊗ I)( ⊗ ˜) = (Λ) ⊗ ˜ ≥ 0. Of course, the
same holds for separable states. Then the main idea is that this property of
separable states is essential, i.e., roughly speaking, if a state is entangled,
then there exists a positive map Λ such that (Λ ⊗ I) is not positive. This
means that one can seek the entangled states by means of the positive maps.
Now the point is that not all of the positive maps can help us to determine
whether a given state is entangled. In fact, the completely positive maps do
not “feel” entanglement. Thus the problem of characterization of the set of
the separable states reduces to the following: one should extract from the set
of all positive maps some essential ones. As we shall see later, this is possible
in some cases. Namely, it appears that for the 2 ⊗ 2 and 2 ⊗ 3 systems the
transposition is the only such map. For higher-dimensional systems, apart
from transposition, nondecomposable maps will also be relevant.
Consider the following lemma [267], which will lead us to the basic theo-
rem relating entanglement and positive maps.
Lemma 5.1. A state ∈ AA ⊗ AB is separable if and only if
Tr(A) ≥ 0 (5.21)
for any operator A satisfying Tr(AP ⊗ Q) ≥ 0, for all pure states P and Q
acting on HA and HB , respectively.
Remark. Note that operator A, which is positive on product states (i.e. it
satisfies TrA P ⊗ Q ≥ 0), is automatically Hermitian.
The lemma is a reflection of the fact that in real Euclidean space, a convex
set and a point lying outside it can always be separated by a hyperplane.9
Here, the convex set is the set of separable states, while the point is the
entangled state. The hyperplane is determined by the operator A. Though
this operator is not positive its restriction to product states is still positive.
Thus, this operator has been called the “entanglement witness” [290], as it
indicates the entanglement of some state (the first entanglement witness was
provided in [263]; see Sect. 5.2.4). Now, to pass to positive maps, we shall use
the isomorphism between entanglement witnesses and positive non-CP maps
[291]. Note that, if we have any linear operator A ∈ AA ⊗ AB , we can define
a map

k| Λ(|i
j|) |l =
i| ⊗
k| A |j ⊗ |l , (5.22)
9
For infinite dimensions one must invoke the Hahn–Banach theorem, whose geo-
metric form is a generalization of this fact.
which can be rephrased as follows:

1
A = (I ⊗ Λ) P+d , (5.23)
d
where d = dim HA . Conversely, given a map, the above formula allows one
to obtain a corresponding operator. It turns out that this formula also gives
a one-to-one correspondence between entanglement witnesses and positive
non-CP maps [291]. By applying this fact, one can prove [267] the following
theorem:
Theorem 5.1. Let act on the Hilbert space HA ⊗ HB . Then is separable

if and only if for any positive map Λ : AB → AA the operator (I ⊗ Λ) is
positive.
As we mentioned, the relevant positive maps here are the ones that are not
completely positive. Indeed, for the CP map Λ we have (I ⊗ Λ) ≥ 0 for any
state , and hence CP maps are of no use here. The above theorem presents, to
our knowledge, the first application of the theory of positive maps in physics.
So far, only completely positive maps have been of interest to physicists.
As we shall see, the theorem has proved fruitful both for mathematics (the
theory of positive maps) and for physics (the theory of entanglement).
Operational Characterization of Entanglement
in Low Dimensions (2 ⊗ 2 and 2 ⊗ 3 Systems). The first conclusion
derived from the theorem is an operational characterization of the separable
states in low dimensions (2 ⊗ 2 and 2 ⊗ 3). This follows from the previously
mentioned result that positive maps in low dimensions are decomposable.
Then the condition (I ⊗ Λ) ≥ 0 reads (I ⊗ ΛCP 1 ) + (I ⊗ Λ2 )
CP TB
. Now,
since is positive and Λ1 is CP, the first term is always positive. If TB
CP
is positive, then the second term is also positive, and hence their sum is
a positive operator. Thus, to check whether for all positive maps we have
(I ⊗ Λ) ≥ 0, it suffices to check only transposition. One obtains the following
[267] (see also [292]):
Theorem 5.2. A state of a 2 ⊗ 2 or 2 ⊗ 3 system is separable if and only
if its partial transposition is a positive operator.
Remark. Equivalently, one can use the partial transposition with respect to
the first space.
The above theorem is an important result, as it allows one to determine
unambiguously whether a given quantum state of a 2⊗2 or a 2⊗3 system can
be written as mixture of product states or not. The necessary and sufficient
condition for separability here is surprisingly simple; hence it has found many
applications. In particular, it has been applied in the context of broadcasting
entanglement [293], quantum information flow in quantum copying networks
[294], disentangling machines [295], imperfect two-qubit gates [296], analysis
of the volume of the set of entangled states [297, 298], decomposition of sepa-
rable states into minimal ensembles or pseudo-ensembles [299], entanglement
splitting [300] and analysis of entanglement measures [270, 301, 302].
In Sect. 5.3.2 we describe the first application [303]: by use of this theorem
we show that any entangled two-qubit system can be distilled, and hence is
useful for quantum communication.
Higher Dimensions – Entangled States
with Positive Partial Transposition. Since the Størmer–Woronowicz char-
acterization of positive maps applies only to low dimensions, it follows that
for higher dimensions partial transposition will not constitute a necessary and
sufficient condition for separability. Thus there exist states that are entangled
but are PPT (see Fig. 5.1). The first explicit examples of an entangled but
PPT state were provided in [274]. Later on, it became apparent, that the
mathematical literature concerning nondecomposable maps contains exam-
ples of matrices that can be treated as prototypes of PPT entangled states
[292, 304].
We shall now describe the method of obtaining such states presented
in [274], as it has proved to be a fruitful direction in searching for PPT
entangled states. Section 5.3 will describe the motivation for undertaking
the very tedious task of this search – the states represent a curious type of
entanglement, namely bound entanglement.
To find the desired examples we must find an entangled PPT state. Of
course, we cannot use the strongest tool so far described, i.e. the PPT crite-
rion, because we are actually trying to find a state that is PPT. So we must
derive a criterion that would be stronger in some cases. It appears that the
range10 of the state can tell us much about its entanglement in some cases.
This is contained in the following theorem, derived in [274] on the basis of
the analogous condition for positive maps considered in [289].
Theorem 5.3. (Range Criterion). If a state acting on the space HAB

is separable, then there exists a family of product vectors ψi ⊗ φi such that
(a) they span the range of
(b) the vectors {ψi ⊗ φ∗i }ki=1 span the range of TB (where ∗ denotes complex
conjugation in the basis in which partial transposition was performed).
In particular, any of the vectors ψi ⊗ φ∗i belongs to the range of .
Now, in [274] there were presented two examples of PPT states violating
the above criterion. We shall present the example for a 2 ⊗ 4 case.11 The
10
The range of an operator A acting on the Hilbert space H is given by R(A) =
{A(ψ) : ψ ∈ H}. If A is a Hermitian operator, then the range is equivalent to the
support, i.e. the space spanned by its eigenvectors with nonzero eigenvalues.
11
This is based on an example concerning positive maps [289].
(a)
PPT NPT
✬ ✩
separable entangled
states states
✫ ✪
(b)
PPT NPT
✬ ✩
entangled states
✬ ✩
separable
states
✫ ✪
✫ ✪
Fig. 5.1. Structure of entanglement of mixed states for 2 ⊗ 2 and 2 ⊗ 3 systems (a)
and for higher dimensions (b)
matrix is written in the standard product basis {|ij}:

 
b000 0 b0 0
 
0 b 0 0 0 0 b 0 
 
0 0 b 0 0 0 0 b 
 
1  0 0 0 b 0 0 0 √ 0 

b =  2  , (5.24)
7b + 1  0 0 0 0 1+b 0 0 1−b 
 2 2 
b 0 0 0 0 b 0 0 
 
 
0 b 0 0 √ 0 0 b 0 
2
0 0 b 0 1−b
2 0 0 1+b
2
where 0 < b < 1. Now, by performing PT as defined by (5.15), we can

check that Tb
B
remains a positive operator. By a tedious calculation, one
can check that none of the product vectors belonging to the range of b , if
they are partially conjugated (as stated in the theorem), belongs to the range
of TB
b . Thus the condition stated in the theorem is drastically violated, and
hence the state is entangled. As we shall see further, the entanglement is
masked so subtly that it cannot be distilled at all!
Range Criterion and Positive Nondecomposable Maps. The separa-
bility criterion given by the above theorem has been fruitfully applied in the
search for PPT entangled states [273, 305, 306]. Theorem 5.3 was applied
in [307], where a technique of subtraction of product vectors from the range
of the state was used to obtain the best separable approximation (BSA) of
the state. As a tool, the authors considered subspaces containing no product
vectors. Note that the (normalized) projector onto such a subspace must be
entangled, as condition (a) of the theorem is not satisfied. This approach
was successfully applied in [273] (see also [290, 305]) and, in connection with
the seemingly completely different concept of unextendible product bases, pro-
duced an elegant, and so far the most transparent, method of construction
of PPT entangled states.
To describe the construction,12 one needs the following definition [273]:
Definition 5.1. A set of product orthogonal vectors in HAB that

(a) has fewer elements than the dimension of the space
(b) is such that there does not exist any product vector orthogonal to all of
them
is called an unextendible product basis (UPB).
Here we recall an example of such basis in the 3 ⊗ 3 system:

1 1
|v0 = √ |0|0 − 1 , |v2 = √ |2|1 − 2 ,
2 2
1 1
|v1 = √ |0 − 1|2 , |v3 = √ |1 − 2|0 ,
2 2
1
|v4 = |0 + 1 + 2|0 + 1 + 2 . (5.25)
3
Of course, the above five vectors are orthogonal to each other. However, each
of the two subsets {|v0 , |v1 , |v4 } and {|v2 , |v3 , |v4 } spans the full three-
dimensional space. This prevents the existence of a sixth product vector that
would be orthogonal to all five of them. How do we connect this with the
problem we are dealing with in this section? The answer is: via the subspace
complementary to the one spanned by these vectors. Indeed, suppose that
12
The construction applies to the multipartite case [273]; in the present review we
consider only bipartite systems.
{wi = φi ⊗ ψi }ki=1 is a UPB. For a d ⊗ d system, consider the projector P =

k
i=1 |wi
wi | onto the subspace H spanned by the vectors wi (dim H = k).
Now, consider the state uniformly distributed on its orthogonal complement
H⊥ (dim H⊥ = d2 − k),
1
= (I − P ) . (5.26)
d2 −k
The range of the state (H⊥ ) contains no product vectors: otherwise one would
be able to extend the product basis {wi }. Then, by Theorem 5.3, the state
must be entangled. Let us now calculate TB . Since wi = φi ⊗ ψi , then
∗
(|wi
wi |)TB = |w̃i
w̃i |, where w̃i = φi ⊗ ψ
i . The vectors w̃i are orthogonal
to each other, so that the operator P TB
= i |w̃i
w̃i | is a projector. Conse-
quently, the operator (I − P )TB = I − P TB is also a projector, and hence it
is positive. We conclude that is PPT.
A different way of obtaining examples of PPT entangled states can be
inferred from the papers devoted to the search for nondecomposable positive
maps in the mathematical literature [292, 304] (see Sect. 5.2.3). A way to
find a nondecomposable map is the following. One constructs some map Λ
and proves somehow that it is positive. Thus one can guess some (possibly
unnormalized) state that is PPT. Now, if (I ⊗ Λ) is not positive, then
Λ cannot be decomposable, as shown in the discussion preceding Theorem
5.2. At the same time, the state must be entangled. In Sect. 5.2.4 we present
an example of a PPT entangled state (based on [288]) found in this way.
Thanks to its symmetric form, the state allowed researchers to reveal the
first quantum effect produced by bound entanglement (see Sect. 24).
Thus a possible direction for exploring the “PPT region” of entanglement
is to develop the description of nondecomposable maps. However, it appears
that there can be also a “back-reaction”: exploration of the PPT region may
allow us to obtain new results on nondecomposable maps. It turns out that
the UPB method described above allows for easy construction of new nonde-
composable maps [308]. We direct the interested reader to the original article,
as well as [290]. We note only that to find a nondecomposable map, one needs
only to construct some UPB. Then the procedure is automatic, like the pro-
cedure described above. To our knowledge, this is the first systematic way of
finding nondecomposable maps.
5.2.4 Examples
We present here a couple of examples, illustrating the results contained in
previous sections. In particular, we introduce two families of states that play
important roles in the problem of distillation of entanglement.
Reduction Criterion for Separability. As mentioned in Sect. 8, if Λ is a
positive map, then for separable states we have
(I ⊗ Λ) ≥ 0 . (5.27)
If the map is not CP, then this condition is not trivial, i.e. for some states
(I ⊗ Λ) is not positive. Consider the map given by Λ(A) = (Tr A)I − A.
The eigenvalues of the resulting operator Λ(A) are given by λi = Tr A − ai ,
where aiare the eigenvalues of A. If A is positive, then ai ≥ 0. Now, since
Tr A = i ai , then λi are also nonnegative. Thus the map is positive. Now,
the formula (5.27) and the dual formula (Λ⊗ I) ≥ 0 applied to this particular
map imply that separable states must satisfy the following inequalities:
I ⊗ B − ≥ 0 , A ⊗ I − ≥ 0 . (5.28)
The two conditions, taken jointly, are called the reduction criterion [281, 309].
One can check that it implies the entropic inequalities (hence it is better in
“detecting” entanglement). From the reduction criterion, it follows that states
of a d ⊗ d system with F () > 1/d must be entangled (this was originally
argued in [66]). Indeed, from the above inequalities, it follows that for a
separable state σ and a maximally entangled state ψme one has
ψme |σA ⊗
I − σ|ψme ≥ 0. Since the reduced density matrix ψAme of the state ψme is
proportional to the identity, we obtain
ψme |σA ⊗ I|ψme = Tr(ψ A σA ) =
me
1/d. Hence we obtain F ≤ 1/d. We conclude that the latter condition is a

separability criterion.
Let us note finally [281], that for 2 ⊗ 2 and 2 ⊗ 3 systems the reduc-
tion criterion is equivalent to the PPT criterion, and hence is equivalent to
separability.
Strong Separability Criteria from an Entanglement Witness. Con-
sider the unitary flip operator V on a d⊗d system defined by V ψ ⊗φ = φ⊗ψ.
Note that it can be written as V = PS − PA , where PS and PA are projectors
onto the symmetric and antisymmetric subspaces, respectively, of the total
space. Hence V is a dichotomic observable (with eigenvalues ±1). One can
check that Tr V A ⊗ B = Tr AB for any operators A, B. Then V is an en-
tanglement witness, so that Tr V ≥ 0 is a separability criterion [263]. Now,
let us find the corresponding positive map via the formula (5.23). One easily
finds that it is a transposition (up to an irrelevant factor). Remarkably, in
this way, given an entanglement witness, one can find the corresponding map,
so as to obtain the much stronger criterion given by (5.27).
Werner States. In [263] Werner considered states that do not change if
both subsystems are subjected to the same unitary transformation:
= U ⊗ U U† ⊗ U† for any unitary U . (5.29)
He showed that such states (called Werner states) must be of the following
form:
1
W (d) = (I + βV ) , −1 ≤ β ≤ 1 , (5.30)
d2 − βd
where V is the flip operator defined above. Another form for W is [310]
PA PS
W (d) = p + (1 − p) , 0≤p≤1, (5.31)
NA NS
where NA = (d2 − d)/2 and NS = (d2 + d)/2 are the dimensions of the
antisymmetric and symmetric subspaces, respectively. It was shown [263]
that W is entangled if and only if Tr V W < 0. Equivalent conditions are
β < −1/d, p > 0 or is NPT. Thus W is separable if and only if it is PPT.
For d = 2 (the two-qubit case) the state can be written as (see [264])
I 1
W (2) = p|ψ−
ψ− | + (1 − p) , − ≤p≤1. (5.32)
4 3
Note that any state , if subjected to a random transformation of the
form U ⊗ U (we call such an operation U ⊗ U twirling), becomes a Werner
state:

dU U ⊗ U U † ⊗ U † = W . (5.33)
Moreover, Tr V = Tr W V (i.e. Tr V is invariant under U ⊗ U twirling).

Isotropic State. If we apply a local unitary transformation to the state
(5.32), changing ψ− into ψ+ , we can generalize its form to higher dimensions
as follows [281]:
I 1
(p, d) = pP+ + (1 − p) , where − ≤p≤1. (5.34)
d2 d2 − 1
The state will be called “isotropic” [311] here.13 For p > 0 it is interpreted as
a mixture of a maximally entangled state P+ with a completely chaotic noise
represented by I/d. It was shown that it is the only state invariant under
U ⊗ U ∗ transformations.14 If we use the singlet fraction F = Tr P+ as a
parameter, we obtain

d2 I 1
(F, d) = 2 (1 − F ) 2 + (F − 2 )P+ , 0 ≤ F ≤ 1 . (5.35)
d −1 d d
The two parameters are related via p = (d2 F − 1)/(d2 − 1). The state is
entangled if and only if F > 1/d or, equivalently, if it is NPT. Similarly to
the case for Werner states, a state subjected to U ⊗ U ∗ twirling (random
U ⊗ U ∗ operations) becomes isotropic, and the parameter F () is invariant
under this operation.
13
In [281] it was called a “noisy singlet”.
14
The star denotes complex conjugation.
A Two-Qubit State. Consider the following two-qubit state:

= p|ψ−
ψ− | + (1 − p)|00
00| . (5.36)
√
From (5.9) we obtain the result that for p ≤ 1/ 2, the CHSH–Bell inequal-
ities are satisfied. A little bit stronger is the criterion involving the fully
entangled fraction; we have F ≤ 1/2 for p ≤ 1/2. The entropic inequalities,
apart from the one involving S0 , are equivalent to each other for this state
and give again p ≤ 1/2. Thus they reveal entanglement for p > 1/2. By
applying the partial transposition one can convince oneself that the state is
entangled for all p > 0 (for p = 0 it is manifestly separable).
Entangled PPT State Via Nondecomposable Positive Map. Con-
sider the following state (constructed on the basis of Størmer matrices [288])
of a 3 ⊗ 3 system:
2 α 5−α
σα = |ψ+
ψ+ | + σ+ + σ− 2≤α≤5 , (5.37)
7 7 7
where
1
σ+ = (|0|1
0|
1| + |1|2
1|
2| + |2|0
2|
0|) ,
3
1
σ− = (|1|0
1|
0| + |2|1
2|
1| + |0|2
0|
2|) . (5.38)
3
Using the formulas (5.15), one easily finds that for α ≤ 4 the state is PPT.
Consider now the following map [287]:
     
a11 a12 a13 a11 −a12 −a13 a33 0 0
Λ  a21 a22 a23  =  −a21 a22 −a23  +  0 a11 0  . (5.39)
a31 a32 a33 −a31 −a32 a33 0 0 a22
This map has been shown to be positive [287]. Now, one can calculate the
operator (I ⊗ Λ) and find that one of its eigenvalues is negative for α > 3
(explicitly, λ− = (3 − α)/2). This implies that
• the state is entangled (for a separable state we would have (I ⊗ Λ) ≥ 0)
• the map is nondecomposable (for a decomposable map and PPT state
we also would have (I ⊗ Λ) ≥ 0).
For 2 ≤ α ≤ 3 it is separable, as it can be written as a mixture of other
separable states σα = 6/71 + (α − 2)/7σ+ + (3 − α)/7σ− , where 1 =
(|ψ+
ψ+ | + σ+ + σ− )/3. The latter state can be written as an integral over
product states:

1 2π dθ
1 = |ψ(θ)
ψ(θ)| ⊗ |ψ(−θ)
ψ(−θ)| ,
8 0 2π
√
where |ψ(θ) = 1/ 3(|0 + eiθ |1 + e−2iθ |2) (there exists also a finite decom-
position exploiting phases of roots of unity [312]).
5.2.5 Volumes of Entangled and Separable States

The question of the volume of the set of separable or entangled states in
the set of all states, raised in [297], is important for several reasons. First,
one might be interested in the following basic question: is the world more
classical or more quantum? Second, the size of the volume would reflect a
consideration that is important for the numerical analysis of entanglement,
that of to what extent separable or entangled states are typical. Later it
became apparent that considerations of the volume of separable states led to
important results concerning the question of the relevance of entanglement
in quantum computing [313].
We shall mainly consider a qualitative question: is the volume of sepa-
rable (Vs ), entangled (Ve ) or PPT entangled (Vpe ) states nonzero? All these
problems can be solved by the same method [297]: one picks a suitable state
from either of the sets and tries to show that some (perhaps small) ball round
the state is still contained in the set.
For separable states one takes the ball round the maximally mixed state:
one needs a number p0 such that for any state ˜ the state
= pI/N + (1 − p)˜
(5.40)
is separable for all p ≤ p0 (here N is the dimension of the total system).
In [297] it was shown that, in the general case of multipartite systems of
any finite dimension, such a p0 exists. Note that in fact we have obtained
a sufficient condition for separability: if the eigenvalues of a given state do
not differ too much from the uniform spectrum of the maximally mixed state,
then the state must be separable. One would like to have some concrete values
of p0 that satisfy the condition (the larger p0 is, the stronger the condition).
Consider, for example, the 2 ⊗ 2 system. Here one can provide the largest
possible p0 , as there exists a necessary and sufficient condition for separability
(the PPT criterion). Consider the eigenvalues λi of the partial transposition
of the state (5.40). They are of the form λi = (1 − p)/N + pλ̃i , where λ̃i
are the eigenvalues of ˜TB (in our case N = 4). One easily can see (on the
basis of the Schmidt decomposition) that a partial transposition of a pure
state cannot have eigenvalues smaller than −1/2. Hence the same is true for
mixed states. In conclusion, we obtain the result that if (1 − p)/N − p/2 ≥ 0
then the eigenvalues λi are nonnegative for arbitrary . ˜ Thus for the 2 ⊗ 2
system one can take p0 = 1/3 to obtain a sufficient condition for separability.
Concrete values of p0 for the case of n-partite systems of dimension d were
obtained in [270]:
1
p0 = . (5.41)
(1 + 2/d)n−1
These considerations proved to be crucial for the analysis of the experimental
implementation of quantum algorithms in high-temperature systems via nu-
clear magnetic resonance (NMR) methods. This is because the generic state
used in this approach is the maximally mixed state with a small admixture of
some pure entangled state. In [313] the sufficient conditions of the above kind
were further developed, and it was concluded that in all of the NMR quantum
computing experiments performed so far, the admixture of the pure state was
too small. Thus the total state used in these experiments was separable: it
satisfied a condition sufficient for separability. This raised an interesting dis-
cussion as to what extent entanglement is necessary for quantum computing
[314, 315] (see also [316]). Even though there is still no general answer, it was
shown [314] that the Shor algorithm [9] requires entanglement.
Let us now turn back to the question of the volumes of Ve and Vpe . If one
takes ψ+ of a d ⊗ d system for simplicity, it is easy to see that a not very
large admixture of any state will ensure F > 1/d. Thus any state belonging
to the neighborhood must be entangled. Showing that the volume of PPT
entangled states is nonzero is a bit more involved [297].
In conclusion, all three types of states are not atypical in the set of all
states of a given system. However, it appears that the ratio of the volume of
the set of PPT states VPPT (and hence also separable states) to the volume
of the total set of states goes down exponentially with the dimension of the
system (see Fig. 5.2). This result was obtained numerically [297] and still
PN
0.6
0.4
0.2
0.0
4 8 12 16 20 24
N
Fig. 5.2. The ratio PN = VPPT /V of the volume of PPT states to the volume of
the set of all states versus the dimension N of the total system. Different symbols
distinguish different sizes of one subsystem (k = 2 (), k = 3 ()). (This figure is
reproduced from Phys. Rev. A 58, 883 (1998) by permission of the authors)
awaits analytical proof.15 However, it is compatible with the rigorous result

in [318] that in the infinite-dimensional case the set of separable states is
nowhere dense in the total set of states. Then the generic infinite-dimensional
state is entangled.
5.3 Mixed-State Entanglement as a Resource

for Quantum Communication
As one knows, if two distant observers (one usually calls them Alice and
Bob) share a pair of particles in a singlet state ψ− then they can send a
quantum state to one another by the use of only two additional classical bits.
This is called quantum teleportation [261].16 If the classical communication
is free of charge (since it is much cheaper than communication of quantum
bits), one can say that a singlet pair is a resource equivalent to sending one
qubit. In the following, it will be shown that mixed-state entanglement can
also be a resource for quantum communication. Quantum communication
via mixed entangled states will require, apart from teleportation, an action
called distillation. It will be also shown that there exists a peculiar type of
entanglement (bound entanglement) that is a surprisingly weak resource.
5.3.1 Distillation of Entanglement:

Counterfactual Error Correction
Now we will attempt to describe the ingenious concept of distillation of en-

tanglement introduced in [266] and developed in [66, 319] (see also [320]).
To this end, let us first briefly describe the idea of classical and quantum
communication via a noisy channel. As is known [321], the central idea of
classical information theory, pioneered by Shannon, is that one can send in-
formation reliably and with a nonzero rate via a noisy information channel.
This is achieved by coding: the k input bits of information are encoded into a
larger number of n bits. Such a package is sent down the noisy channel. Then
the receiver performs a decoding transformation, recovering the k input bits
with asymptotically (in the limit of large n and k) perfect fidelity. Moreover,
the asymptotic rate of information transmission k/n is nonzero.
In quantum domain, one would like to communicate quantum states in-
stead of classical messages. It appears that an analogous scheme can be ap-
plied here [62, 322]. The k input qubits of quantum information are supple-
mented with additional qubits in some standard initial state, and the total
system of n qubits is subjected to some quantum transformation. Now the
package can be sent down the channel. After the decoding operation, the
15
The result could depend on the measure of the volume chosen [317]. In [298] two
different measures were compared and produced similar results.
16
It is called “entanglement-assisted teleportation” in Chap. 2 of this book.
state of k qubits is recovered with asymptotically perfect fidelity [69] (now it

is quantum fidelity – characterizing how close the output state is to the input
state), regardless of the particular form of the state. The discovery of the
above possibility (called quantum error correction; we shall call it here direct
error correction) initiated, in particular, extensive studies of quantum error-
correcting codes (see [255] and references therein), as well as studies of the
capacities of quantum channels (see [94] and references therein).17 A common
example of a quantum channel is the one-qubit quantum depolarizing chan-
nel: here an input state is undisturbed with probability p and subjected to
a random unitary transformation with probability 1 − p. It can be described
by the following completely positive map:
I
→ Λ() = p + (1 − p) , (5.42)
2
where I/2 is the maximally mixed state of one qubit. This channel has been
thoroughly investigated [66, 266, 323, 324]. What is important here is that it
has been shown [324] that for p ≤ 2/3 the above method of error correction
does not work. In the classical domain, it would mean that the channel was
useless. Here, surprisingly, there is a trick that allows one to beat this limit,
even down to p = 1/3! The scheme that realizes this fact is quite mysterious.
In direct error correction we deal directly with the systems carrying the infor-
mation to be protected. Now, it appears that by using entanglement, one can
remove the results of the action of noise without even having the information
to be sent. Therefore, it can be called counterfactual error correction.
How does this work? The idea itself is not complicated. Instead of sending
the qubits of information, Alice (the sender) sends Bob particles from entan-
gled pairs (in the state ψ− ), keeping one particle from each pair. The pairs
are disturbed by the action of the channel, so that their state turns into a
mixture18 that still possesses some residual entanglement. Now, it turns out
that by local quantum operations (including collective actions over all mem-
bers of the pairs in each lab) and classical communication (local operations
and classical communication, LOCC) between Alice and Bob, Alice and Bob
are able to obtain a smaller number of pairs in a nearly maximally entan-
gled state ψ− (see Fig. 5.3). Such a procedure, proposed in [266], is called
distillation. As in the case of direct error correction, one can achieve a finite
asymptotic rate k/n for the distilled pairs per input pair, and the fidelity,
which now denotes the similarity of the distilled pairs to a product of sin-
glet pairs, is asymptotically perfect. Now, the distilled pairs can be used for
teleportation of quantum information. The maximal possible rate achievable
within the above framework is called the entanglement of distillation of the
17
The capacity Q of a quantum channel is the greatest ratio k/n for reliable trans-
mission down the given channel.
18
If the channel is memoryless, the mixture factorizes into states of individual
pairs.
state , and is denoted by D(). Thus, if Alice and Bob share n pairs each
in state , they can faithfully teleport k = nD() qubits.
As we have mentioned, the error correction stage and the transmission
stage are separated in time here; the error correction can be performed even
before the information to be sent was produced. Using the terminology of
[325], one can say that Alice and Bob operate on potentialities (an entangled
pair represents a potential communication) and correct the potential error,
so that when the actual information is coming, it can be teleported without
any additional action.
The above scheme is not only mysterious. It is also much more powerful
than the direct method. In the next section we describe a distillation protocol
that allows one to send quantum information reliably via a channel with
p > 1/3. A general question is: where are the limits of distillation? As we
have seen, the basic action refers to mixed bipartite states, so that instead of
talking about channels, we can concentrate on bipartite states. The question
can be formulated as follows: which states can be distilled by the most
general LOCC actions? Here, by saying that a state can be distilled, we
mean that Alice and Bob can obtain singlets from the initial state ⊗n of n
pairs (thus we shall work with memoryless channels).
One can easily see that separable states cannot be distilled: they contain
no entanglement, so it is impossible to convert them into entangled states by
LOCC operations. Then the final form of our question is: can all entangled
states be distilled? Before the answer to this question was provided, the
default was “yes”, and the problem was how to prove it. Now, we know that
the answer is “no”, so that the structure of the entanglement of bipartite
states is much more puzzling than one might have suspected.
Finally, we should mention that for pure states the problem of conversion
into singlet pairs has been solved. Here there is no surprise: all entangled pure
states can be distilled [326] (see also [327]). What is especially important
is that this distillation can be performed reversibly: from the singlet pairs
obtained, we can recover (asymptotically) the same number of input pairs
[326]. As we shall see, this is not the case for mixed states.
8 • • Alice and Bob

>
> operations 9
n pairs,
>
< •• • •\/\/\/\/\/\/\/\/• >
= k nearly
• =⇒ •\/\/\/\/\/\/\/\/•
each in
>
> . . > singlet
; pairs
state
>
: .• • •\/\/\/\/\/\/\/\/•
Fig. 5.3. Distillation of mixed-state entanglement

5.3.2 Distillation of Two-Qubit States
In this section, we shall describe what was historically the first distillation
protocol for two-qubit states, devised by Bennett, Brassard, Popescu, Schu-
macher, Smolin and Wootters (BBPSSW) [266]. Then we shall show that a
more general protocol can distill any entangled two-qubit state.19
BBPSSW Distillation Protocol. The BBPSSW distillation protocol still
remains the most transparent example of distillation. It works for two-qubit
states with a fully entangled fraction satisfying F > 1/2. Such states are
equivalent to those with F > 1/2, so that we can restrict ourselves to the
latter states. Hence we assume that Alice and Bob initially share a huge
number of pairs, each in the same state with F > 1/2, so that the total
state is ⊗n . Now they aim to obtain a smaller number of pairs with a higher
singlet fraction F . To this end they iterate the following steps:
1. They take two pairs and apply U ⊗ U ∗ twirling to each of them, i.e.
a random unitary transformation of the form U ⊗ U ∗ (Alice picks at
random a transformation U , applies it, and communicates to Bob which
transformation she has chosen; then he applies U ∗ to his particle). Thus
one has a transformation from two copies of to two copies of the isotropic
state F with an unchanged F :
⊗ → F ⊗ F . (5.43)
2. Each party performs the unitary transformation XOR20 on his/her mem-

bers of the pairs (see Fig. 5.4). The transformation is given by
UXOR |a|b = |a|(a + b) mod 2 (5.44)
(the first qubit is called the source, the second qubit the target). They
obtain some complicated state ˜ of two pairs.
3. The pair of target qubits is measured locally in the basis |0, |1 and it is
discarded. If the results agree (success), the source pair is kept and has a
greater singlet fraction. Otherwise (failure), the source pair is discarded
too.
If the results in step 3 agree, the final state of the source pair kept can be
calculated from the formula
= TrHt (Pt ⊗ Is ˜Pt ⊗ Is ) , (5.45)

19
In this contribution we restrict ourselves to distillation by means of perfect op-
erations. The more realistic case where there are imperfections in the quantum
operations performed by Alice and Bob is considered in [328].
20
The quantum XOR gate is the most common quantum two-qubit gate and was
introduced in [329].
source pair
• •
Alice XOR XOR Bob
• •
target pair
Fig. 5.4. Bilateral quantum XOR operation
where the partial trace is performed over the Hilbert space H(t) of the tar-
get pair, Is is the identity on the space of the source pair (because it was
not measured), and Pt = |00
00| + |11
11| acts on target pair space and
corresponds to the case “results agree”.
Subsequently, one can calculate the singlet fraction of the surviving pair
as a function of the singlet fraction of the two initial pairs, obtaining
F 2 + (1/9)(1 − F )2
F (F ) = . (5.46)
F 2 + (2/3)F (1 − F ) + (5/9)(1 − F )2
Since the function F (F ) is continuous, F (F ) > F for F > 1/2 and F (1) =
1, we obtain the result that by iterating the procedure, Alice and Bob can
obtain a state with arbitrarily high F . Of course, the larger F is required
to be, the more pairs must be sacrificed, and the less the probability p of
success is. Thus if Alice and Bob start with some Fin and would like to
end up with some higher Fout , the number of final pairs will be on average
k = np/2l , where l and p depend on Fin and Fout , and denote the number
of iterations of the function F (F ) required to reach Fout starting from Fin ,
and the probability of a string of l successful operations, respectively.
The above method allows one to obtain an arbitrarily high F , but the
asymptotic rate is zero. However, if F is high enough so that 1 − S > 0, where
S is the von Neumann entropy of the state , then there exists a protocol
(called hashing) that gives a nonzero rate [66]. We shall not describe this
protocol here, but we note that for any state with F > 1/2 Alice and Bob
can start by using the recurrence method to obtain 1 − S > 0, and then
apply the hashing protocol. This gives a nonzero rate for any state with
F > 1/2. This means that quantum information can be transmitted via a
depolarizing channel (5.42) only if p > 1/3. Indeed, one can check that if
Alice send one of the particles from a pair in a state ψ+ via the channel
to Bob, then the final state shared by them will be an isotropic one with
F > 1/2. By repeating this process, Alice and Bob can obtain many such
pairs. Then distillation will allow them to use the pairs for asymptotically
faithful quantum communication.
All Entangled Two-Qubit States are Distillable. As was mentioned in
Sect. 5.2.4, there exist entangled two-qubit states with F < 1/2, so that no
product unitary transformation can produce F > 1/2. Thus the BBPSSW
protocol cannot be applied to all entangled two-qubit states. We shall show
below that, nevertheless, all such states are distillable [303]. It was possible
to solve the problem mainly because of the characterization of the entangled
states as discussed in Sect. 5.2.3.
Since we are not interested in the value of the asymptotic rate, it suf-
fices to show that by starting with pairs in an entangled state, Alice and
Bob are able to obtain a fraction of them in a new state with F > 1/2 (and
then the BBPSSW protocol will do the job). Our main tool will be the so-
called filtering operation [326, 330], which involves generalized measurement
performed by one of the parties (say, Alice) on individual pairs. This mea-
surement consists of two outcomes {1, 2}, associated with operators W1 and
W2 satisfying
W1† W1 + W2† W2 = IA (5.47)
(IA and IB denote identities on Alice’s and Bob’s systems, respectively).

After such a measurement, the state becomes
1
→ Wi ⊗ IB Wi† ⊗ IB , i = 1, 2 (5.48)
pi
with probability pi = Tr(Wi Wi† ). The condition (5.47) ensures p1 + p2 = 1.

Now Alice will be interested only in one outcome (say, 1). If this outcome
occurs, Alice and Bob keep the pair; otherwise they discard it (this requires
communication from Alice to Bob). Then we are only interested in the op-
erator W1 . If its norm does not exceed 1, one can always find a suitable
W2 such that the condition (5.47) is satisfied. Now, since neither the form
of the final state (5.48) nor the fact whether p1 is zero or not depends on
the positive factor multiplying W1 , we are free to consider completely ar-
bitrary filtering operators W1 . In conclusion, for any entangled state we
must find an operator W such that the state resulting from the filtering
W ⊗ I W † ⊗ I/Tr(W ⊗ I W † ⊗ I) has F > 1/2. Consider then an arbitrary
two-qubit entangled state . From Theorem 5.2, we know that TB is not a
positive operator, and hence there exists a vector ψ for which

ψ|TB |ψ < 0 . (5.49)
Now let us note that any vector φ of a d ⊗ d system can be written as

φ = Aφ ⊗ Iψ+ , where Aφ is some operator. Indeed, write φ in a product
d
basis: φ = i,j=1 aij |i ⊗ |j. Then the matrix elements of the operator Aφ
√
are given by
i|Aφ |j = daij (in our case, d = 2). Therefore the formula
(5.49) can be rewritten in the form

Tr (A†ψ ⊗ I Aψ ⊗ I)TB P+ < 0 . (5.50)
Using the identity Tr ATB B = Tr AB TB , valid for any operators A, B, and

the fact that P+TB = 1/dV (where V is the flip operator, see Sect. 5.2.4), we
obtain

Tr (A†ψ ⊗ I Aψ ⊗ I) V < 0 . (5.51)
We conclude that A†ψ ⊗ I Aψ ⊗ I cannot be equal to the null operator, and

hence we can consider the following state:
A†ψ ⊗ I Aψ ⊗ I
˜ = .
Tr(A†ψ ⊗ I Aψ ⊗ I)
Now it is clear that the role of the filter W will be played by the operator A†ψ .
√
We shall show that
ψ− |˜ |ψ− > 1/2, where ψ− = (|01 − |10)/ 2. Then a
suitable unitary transformation by Alice can convert ˜ into a state with
F > 1/2.
From the inequality (5.51), we obtain
Tr ˜V < 0 . (5.52)
If we use the product basis |1 = |00, |2 = |01, |3 = |10, |4 = |11, the
inequality (5.52) can be written as
˜11 + ˜44 + ˜23 + ˜32 < 0 . (5.53)

The above inequality, together with the trace condition Tr ˜ = i ˜ii = 1,
gives
1 1

ψ− |˜
|ψ− = 22 + ˜33 − ˜23 − ˜32 ) > .
(˜ (5.54)
2 2
To summarize, given a large supply of pairs, each in an entangled state
, Alice and Bob can distill maximally entangled pairs in the following way.
First Alice applies a filtering determined by the operator W = A†ψ described
above. Then Alice and Bob obtain, on average, a supply of np surviving pairs
in the state ˜ (here p = Tr W ⊗ I W † ⊗ I is the probability that the outcome
of Alice’s measurement will be the one associated with the operator W † ). Now
Alice applies an operation iσy to obtain a state with F > 1/2. Then they
can use the BBPSSW protocol to distill maximally entangled pairs that are
useful for quantum communication. Note that we have assumed that Alice
and Bob know the initial state of the pairs. It can be shown that, if they√do
not know, they still can do the job (in the two-qubit case) by sacrificing n
pairs to estimate the state (P. Horodecki, unpublished).
The above protocol can easily be shown to work in the 2 ⊗ 3 case. The
protocol can be also fruitfully applied for the system 2 ⊗ n if the state is NPT
[331].
5.3.3 Examples
Consider the state (5.36) from Sect. 5.2.4, = p|ψ−

ψ− | + (1 − p)|00
00|.
It is entangled for all p > 0. In matrix notation we have
   
1−p 0 0 0 1 − p 0 0 − p2
 0 p p   0 p 0 0 
= 2 −2 0  
 0 −p p 0 , =  0 0 p 0  .
TB 2 (5.55)
2 2 2
0 0 0 0 − p2 0 0 0

The negative eigenvalue of TB is λ− = 1/2 1 − p − (1 − p)2 + p2 and
the corresponding (unnormalized) eigenvector ψ = λ− |00 − (p/2)|11, and
hence we can take the filter to be of the form W = diag[λ− , −p/2]. The new
state ˜ resulting from filtering is of the form
 2 
λ− (1 − p) 0 0 0
3 2
1  0 p p 
4 λ− 0  ,
˜ = 8 (5.56)
N 
2
p p
4 λ− 2 λ− 0
2
0
0 0 0 0
where N = λ2− (1 − p) + p2 /8 + λ2− p/2. Now the overlap with ψ− , given by
ψ− |˜
|ψ− = (p3 /8 + λ2 p/2 − λp2 /2)/N , is greater than 1/2 only if p > 0.
The new state can be distilled by the BBPSSW protocol.
Below we shall prove that some states of higher-dimensional systems are
distillable. We shall do this by showing that some LOCC operation can con-
vert them (possibly with some probability) into an entangled two-qubit state.
Distillation of Isotropic State for d ⊗ d System. For F > 1/d, an
isotropic state can be distilled [281, 313]. If both Alice and Bob apply the
projector P = |0
0| + |1
1|, where |0, |1 are vectors from the local basis,
then the isotropic state will be converted into a two-qubit isotropic state.
(Note that the projectors play the role of filters; also, the filtering is successful
if both Alice and Bob obtain outcomes corresponding to P .) Now, if the initial
state satisfied F > 1/d then the final state, as a two-qubit state, will have
F > 1/2. Thus it is entangled and hence can be distilled.
Distillation and Reduction Criterion. Any state of a d ⊗ d system
that violates the reduction criterion (see Sect. 5.2.4) can be distilled [281].
Indeed, take a vector ψ for which
ψ|A ⊗ I − |ψ < 0. It is easy to see that
by applying the filter W given by ψ = W ⊗ Iψ+ , one obtains a state with
F > 1/d. Now, the random U ⊗ U ∗ transformations will convert it into an
isotropic state with the same F . As shown above, the latter state is distillable.
5.3.4 Bound Entanglement
In the light of the result for two qubits, it was naturally expected that any
entangled state could be distilled. It was a great surprise when it became
apparent that it was not the case. In [71] it was shown that there exist
entangled states that cannot be distilled. The following theorem provides a
necessary and sufficient condition for the distillability of a mixed state [71].
Theorem 5.4. A state is distillable if and only if, for some two-dimen-
sional projectors P, Q and for some number n, the state P ⊗ Q⊗n P ⊗ Q is
entangled.
Remarks. (1) Note that the state P ⊗ Q⊗n P ⊗ Q is effectively a two-qubit
one as its support is contained in the C ⊗ C subspace determined by the
2 2
projectors P, Q. This means that the distillable entanglement is a two-qubit

entanglement. (2) One can see that the theorem is compatible with the fact
[326] that any pure state can be distilled.
As a consequence of this theorem, we obtain the following theorem [71]:
Theorem 5.5. A PPT state cannot be distilled.
Proof. We shall give here a proof independent of Theorem 5.4. As a matter

of fact, we shall show that (i) the set of PPT states is invariant under LOCC
operations [71] and (ii) it is bounded away from the maximally entangled
state [311, 332]. Then, since (⊗n )TB = (TB )⊗n , we obtain the proof of the
theorem. To prove (i), note that any LOCC operation can be written as [268]
1
→ = Ai ⊗ Bi A†i ⊗ Bi† , (5.57)
p i
where p is a normalization constant interpreted

as the probability of real-
ization of the operation, and the map → i Ai ⊗ Bi A†i ⊗ Bi† does not
increase the trace (this ensures p ≤ 1). Suppose now that is PPT, i.e.
TB ≥ 0, and examine partial transposition of the state . We shall use the
following property of partial transposition:
(A ⊗ BC ⊗ D)TB = A ⊗ DT TB C ⊗ B T (5.58)
for any operators A, B, C, D and . Then we obtain

( )TB = Ai ⊗ (Bi† )T TB Ai ⊗ (Bi )T . (5.59)
i
Thus ( )TB is a result of the action of some completely positive map on an
operator TB that by assumption is positive. Then also the operator ( )TB
must be positive. Thus a LOCC map does not move outside the set of PPT
states.
To prove (ii), let us now show that PPT states can never have a high
singlet fraction F . Consider a PPT state of a d ⊗ d system. We obtain
Tr P+ = Tr TB P+TB . (5.60)

Now, it is easy to check that P+ = 1/dV , where V is the flip operator

described in Sect. 5.2.4. Note that V is Hermitian and has eigenvalues ±1.
Since is PPT then ˜ = TB is a legitimate state, and the above expression
can be rewritten in terms of the mean value of the observable V as
1
Tr P+ = Tr ˜V . (5.61)
d
The mean value of a dichotomic observable cannot exceed 1, so that we obtain
1
F () ≤ . (5.62)
d
Thus the maximal possible singlet fraction that can be attained by PPT states
is the one that can be obtained without any prior entanglement between the
parties. Indeed, a product state |00 has a singlet fraction 1/d (if it belongs
to the Hilbert space C d ⊗ C d ). Consequently, for however large an amount
of PPT pairs, even a single two-qubit pair with F > 1/2 cannot be obtained
by LOCC actions.
Now, one can appreciate the results presented in the first part of this con-
tribution. From Sect. 5.2.3, we know that there exist entangled states that
are PPT. So far, the question of whether there exist entangled states that
are PPT has been merely a technical one. At this point, since the above the-
orem implies that PPT states are nondistillable, we can draw a remarkable
conclusion: there exist nondistillable entangled states. Since, in the process of
distillation, no entanglement can be liberated to the useful singlet form, they
have been called bound entangled (BE) states. Thus there exist at least two
qualitatively different types of entanglement: apart from the free entanglement
that can be distilled, there is a bound one that cannot be distilled and seems
to be completely useless for quantum communication. This discontinuity of
the structure of the entanglement of mixed states was considered to be pos-
sible for multipartite systems, but it was completely surprising for bipartite
systems. It should be emphasized here that the BE states are not atypical in
the set of all possible states: as we have mentioned in Sect. 5.2.5, the volume
of the PPT entangled states is nonzero. One of the main consequences of the
existence of BE states is that it reveals a transparent form of irreversibility
in entanglement processing. If Alice and Bob share pairs in a pure state, then
to produce a BE state they need some prior entanglement.21 However, once
they have produced the BE states, they are not able to recover the pure en-
tanglement from them. It is entirely lost. This is a qualitative irreversibility,
which is probably a source of the quantitative irreversibility [66, 266] that is
due to the fact that we need more pure entanglement to produce some mixed
states than we can then distill back from them [269, 334].22
21
This was rigorously proved recently in [333].
22
In fact, the existence of both kinds of irreversibility has not been rigorously
proved so far (see [335]). The proof of quantitative irreversibility in [336] turned
To analyse the phenomenon of bound entanglement, one needs as many

examples of BE states as possible. Hence there is a very exciting physical
motivation for the search for PPT entangled states. In Sect. 9 we discussed
different methods of searching . As a result we obtained a couple of exam-
ples of BE states via the separability criterion given by Theorem 5.3, from
the mathematical literature on nondecomposable maps and via unextendible
product bases.
The examples produced via UPBs are extremely interesting from the phys-
ical point of view. This is because a UPB is not only a mathematical object:
as shown in [273], it produces a very curious physical effect [120] called “non-
locality without entanglement”. Namely, suppose that Alice and Bob share
a pair in one of the states from the UPB, but they do not know which
state this is. It appears that by LOCC operations (with finite resources),
they are not able to read the identity of the state. However, if the particles
were together, then, since the states are orthogonal, they could be perfectly
distinguished from each other. Thus we have a highly nonclassical effect pro-
duced by an ensemble of separable states. On the other hand, the BE state
associated with the given UPB (the uniform state on the complementary
subspace, see (5.26)) presents opposite features: it is entangled but, since its
entanglement is bound, it ceases to behave in a quantum manner. Moreover,
in both situations we have a kind of irreversibility. As was mentioned, BE
states are a reflection of the formation–distillation irreversibility: to create
them by LOCC from singlet pairs, Alice and Bob need a nonzero amount of
the latter. However, once they are created, there is no way to distill singlets
out of them. On the other hand, a UPB exhibits preparation–measurement
irreversibility: any of the states belonging to the UPB can be prepared by
LOCC operations, but once Alice and Bob forget the identity of the state,
they cannot recover it by LOCC. This surprising connection between some
BE states and bases that are not distinguishable by LOCC implies many in-
teresting questions concerning the future unification of our knowledge about
the nature of quantum information.
Finally, we shall mention a result concerning the rank of the BE state. In
numerical analysis of BE states (especially their tensor products), it is very
convenient to have examples with low rank. However, in [282] the following
bound on the rank of the BE state was derived:
R() ≥ max{R(A ), R(B )} . (5.63)
(Recall that R() denotes the rank of .) Note that the above inequality
is nothing but the entropic inequality (5.10) with the entropy (5.11). Thus
it appears that the latter inequality is a necessary condition not only for
separability, but also for nondistillability. The proof is based on the fact [281]
out to be invalid: it was based on a theorem [311] on the additivity of the relative
entropy of Werner states. However, an explicit counterexample to this theorem
was provided in [337].
that any state violating a reduction criterion (see Sects. 5.2.4 and 5.3.3) can
be distilled. It can be shown that, if a state violates the above equation, then
it must also violate a reduction criterion, and hence can be distilled. Then
it follows that there does not exist any BE state of rank two [282]. Indeed,
if such a state existed, then its local ranks could not exceed two. Hence the
total state would be effectively a two-qubit state. However, from Sect. 19 we
know that two-qubit bound entangled states do not exist.
5.3.5 Do There Exist Bound Entangled NPT States?
So far we have considered BE states that arise from Theorem 5.5, which says
that the NPT condition is necessary for distillability. As mentioned in Sect.
19, for 2 ⊗ n systems all NPT states can be distilled [331], and hence the
condition is also sufficient in this case. However, it is not known whether
it is sufficient in general. A necessary and sufficient condition is given by
Theorem 5.4. To find if this condition is equivalent to the PPT one, it must
be determined whether there exists an NPT state such that, for any number
of copies n, the state ⊗n will not have an entangled two-qubit “substate”
(i.e. the state P ⊗ Q⊗n P ⊗ Q). In [281] it was pointed out that one can
reduce the problem by means of the following observation.
Proposition 5.1. The following statements are equivalent:
1. Any NPT state is distillable.
2. Any entangled Werner state (5.30) is distillable.
Proof. The proof of the implication (1) ⇒ (2) is immediate, as Werner states
are entangled if and only if they are NPT. If we can distill any NPT state,
then also Werner entangled states are distillable. To obtain (2) ⇒ (1), note
that the reasoning of Sect. 19, from (5.49) to (5.52), is insensitive to the
dimension d of the problem. Consequently, from any NPT state, a suitable
filtering produces a state ˜ satisfying Tr ˜V < 0. As mentioned in Sect. 5.2.4,
the parameter Tr V is invariant under U ⊗ U twirling, so that by applying
the latter (which is an LOCC operation), Alice and Bob obtain a Werner
state W satisfying Tr W V < 0. Thus any NPT state can be converted by
means of LOCC operations into an entangled Werner state, which completes
the proof.
The above proposition implies that to determine whether there exist NPT
bound entangled states, one can restrict oneself to the family of Werner states,
which is a one-parameter family of very high symmetry. Even after such a
reduction of the problem, the latter remains extremely difficult. In [310, 338]
the authors examine the nth tensor power of Werner states (in [338] a larger,
two-parameter family is considered). The results, though not conclusive yet,
strongly suggest that there exist NPT bound entangled states (see Fig. 5.5).
Thus it is likely that the characterization of distillable states is not as sim-
ple as reduction to a NPT condition. The possible existence of NPT bound
(a)
PPT NPT
✬ ✩
free
separable
entangled
states
states
✫ ✪
(b)
PPT NPT
✬ ✩
bound entangled states

✬ ✩✬ ✩
separable free
states entangled
states
✫ ✪
✫ ✪
✫ ✪
Fig. 5.5. Entanglement and distillability of mixed states for 2 ⊗ 2 and 2 ⊗ 3 systems
(a) and for higher dimensions (b). The area filled with diagonal lines denotes the
hypothetical set of bound entangled NPT states
entanglement would make the total picture much more obscure (and hence
much more interesting). Among others, there would arise the following ques-
tion: for two distinct BE states 1 and 2 , is the state 1 ⊗ 2 also BE? (If
BE was equivalent to PPT, this question would have an immediate answer
“yes”, because the PPT property is additive, i.e. if two states are PPT, then
so is their tensor product [284]). Recently, a negative answer to this question
was obtained in [339] in the case of a multipartite system. For bipartite states
the answer is still unknown.
5.3.6 Example
Consider the family of states (5.37) considered in Sect. 5.2.4. One obtains the
following classification: is
• separable for 2 ≤ α ≤ 3
• bound entangled (BE) for 3 < α ≤ 4
• free entangled (FE) for 4 < α ≤ 5 .
The separability was shown in Sect. (5.2.4). It was also shown there that, for
3 < α ≤ 4, the state is entangled and PPT. In this case we conclude that it is
BE. For α > 4, Alice and Bob can apply local projectors P = |0
0| + |1
1|,
obtaining an entangled two-qubit state. Hence the initial state is FE in this
region of α.
5.3.7 Some Consequences of the Existence

of Bound Entanglement
A basic question that arises in the context of bound entanglement is: what
is its role in quantum information theory? We shall show in the following
sections that even though it is indeed a very poor type of entanglement, it
can produce a nonclassical effect, enhancing quantum communication via a
subtle activation-like process [340]. This will lead us to a new paradigm of
entanglement processing that extends the “LOCC paradigm”. Moreover, the
existence of bound entanglement means that there exist stronger limits on
the distillation rate than were expected before. We shall report these and
other consequences in the next few subsections.
Bound Entanglement and Teleportation. By definition, BE states can-
not be distilled, and hence it is impossible to obtain faithful teleportation via
such states. However, it might be the case that the transmission fidelity of im-
perfect teleportation might still be better than that achievable with a purely
classical channel, i.e. without sharing any entanglement (this is a way of re-
vealing a manifestation of quantum features of some mixed states [264]).23
Initial searches produced a negative result [314]. Here we present more gen-
eral results, according to which the most general teleportation scheme cannot
produce better than classical fidelity if Alice and Bob share BE states.
23
For a detailed study of the standard teleportation scheme via a mixed two-qubit
state, see [341]. The optimal one-way teleportation scheme via pure states was
obtained in [342].
General Teleportation Scheme. Teleportation, as originally devised [261],

is a way of transmitting a quantum state by use of a classical channel and a bi-
partite entangled state (pure singlet state) shared by Alice and Bob. The most
general scheme of teleportation would then be of the following form [343].
There are three systems: that of the input particle, the state of which is to
be teleported (we ascribe to this system the Hilbert space HA ), and two sys-
tems that are in the entangled state AB (with Hilbert space H = HA ⊗ HB ).
For simplicity we assume that dim HA = dim HA = dim HB = d. The initial
state is
|ψA
ψA | ⊗ AB ,
where ψA is the state to be teleported (unknown to Alice and Bob). Now Al-
ice and Bob perform some trace-preserving LOCC operation (trace-preserving,
because teleportation is an operation that must be performed with probabil-
ity 1). The form of the operation depends on the state AB that is known to
Alice and Bob, but is independent of the input state ψA because that state
is unknown. Now the total system is in a new, perhaps very complicated
state A AB . The transmitted state is given by TrA A (A AB ). The overall
transmission stages are the following:
ψA → |ψA
ψA | ⊗ AB → Λ(|ψA
ψA | ⊗ AB ) → TrA A A AB = B .
The transmission fidelity is now defined by
f =
ψA |B |ψA ,
where the average is taken over a uniform distribution of the input states
ψA .24 In the original teleportation scheme (where AB is a maximally en-
tangled state), the state B is exactly equal to the input state, so that f = 1.
If Alice and Bob share a pair in a separable state (or, equivalently, share no
pair), then the best one can do is the following: Alice measures the state and
sends the results to Bob [264]. Since it is impossible to find the form of the
state when one has only a single system in that state [345] (it would con-
tradict the no-cloning theorem [346] (see Sect. 1)), the performance of such
a process will be very poor. One can check that the best possible fidelity is
f = 2/(d + 1). If the shared pair is entangled but is not a pure maximally
entangled state, we shall obtain some intermediate value of f .
Optimal Teleportation. Having defined the general teleportation scheme,
one can ask about the maximal fidelity that can be achieved for a given state
24
Note that the fidelity so defined is not a unique criterion of the performance of
teleportation. For example, one can consider a restricted input: Alice receives one
of two nonorthogonal vectors with some probabilities [344]. Then the formula for
the fidelity would be different. In general, the fidelity is determined by a chosen
distribution over input states.
AB within the scheme. Thus, for a given AB we must maximize f over all
possible trace-preserving LOCC operations. The problem is, in general, ex-
tremely difficult. However, the high symmetry of the chosen fidelity function
allows one to reduce it in the following way. It has been shown [343] that the
best Alice and Bob can do is the following. They first perform some LOCC
action that aims at increasing F (AB ) as much as possible. Then they per-
form the standard teleportation scheme, via the new state AB (just as if it
were the state P+ ). The fidelity obtained is given by
Fmax d + 1
fmax = , (5.64)
d+1
where Fmax = F (AB ) is the maximal F that can be obtained by trace-
preserving LOCC actions if the initial state is AB .
Teleportation Via Bound Entangled States. According to (5.64), to
check the performance of teleportation via BE states of a d ⊗ d system, we
should find the maximal F attainable from BE states via trace-preserving
LOCC actions. As was argued in Sect. 5.3.4, a BE state subjected to any
LOCC operation remains BE. Moreover, the singlet fraction F of a BE state
of a d⊗d system satisfies F ≤ 1/d (because states with F > 1/d are distillable,
as shown in Sect. 5.3.3). We conclude that, if the initial state is BE, then
the highest F achievable by any (not only trace-preserving) LOCC actions
is F = 1/d. However, as we have argued, this gives a fidelity f = 2/(d + 1),
which can be achieved without entanglement. Thus the BE states behave
here like separable states – their entanglement does not manifest itself.
Activation of Bound Entanglement. Here we shall show that bound
entanglement can produce a nonclassical effect, even though the effect is a
very subtle one. This effect is the so-called activation of bound entanglement
[340]. The underlying concept originates from a formal entanglement–energy
analogy developed in [71, 269, 325, 335, 347]. One can imagine that the bound
entanglement is like the energy of a system confined in a shallow potential
well. Then, as in the process of chemical activation, if we add a small amount
of extra energy to the system, its energy can be liberated.
In our case, the role of the system is played by a huge amount of bound
entangled pairs, while that of the extra energy is played by a single pair
that is free entangled. More specifically, we shall show that a process called
conclusive teleportation [348] can be performed with arbitrarily high fidelity
if Alice and Bob can perform joint operations over the BE pairs and the
FE pair. We shall argue that it is impossible if either of the two elements is
lacking.
Conclusive Teleportation. Suppose that Alice and Bob have a pair in a
state for which the optimal teleportation fidelity is f0 . Suppose, further, that
the fidelity is too poor for some of Alice and Bob’s purposes. What they can
do to change the situation is to perform a so-called conclusive teleportation.
Namely, they can perform some LOCC operation with two final outcomes 0
and 1. If they obtain the outcome 0, they fail and decide to discard the pair.
If the outcome is 1 they perform teleportation, and the fidelity is now better
than the initial f0 . Of course, the price they must pay is that the probability
of success (outcome 1) may be small. The scheme is illustrated in Fig. 5.6.
8
>
>
>
> Alice • • Bob
>
>
>
>
>
>
preparation
>
< LOCC
of a strongly
>
>
❄
entangled pair
>
> ✲ (failure)
>
>
>
> 1−p
>
>
>
: (success) p
❄
( •\/\/\/\/\/\/\/\/•
teleportation φ ◦ −→ + −→ ◦ φ
LOCC
Fig. 5.6. Conclusive teleportation. Starting with a weakly entangled pair, Alice
and Bob prepare with probability p a strongly entangled pair and then perform
teleportation
A simple example is the following. Suppose that Alice and Bob share a
pair in a pure state ψ = a|00 + b|11 which is nearly a product state (e.g. a
is close to 1). Then the standard teleportation scheme provides a rather poor
fidelity f = 2(1 + ab)/3 [341, 349]. However, Alice can subject her particle to
a filtering procedure [326, 330] described by the operation
Λ = W (·)W † + V (·)V † , (5.65)
where W = diag(b, a), V = diag(a, b). Here the outcome 1 (success) cor-
responds to the operator W . Indeed, if this outcome is obtained, the state
collapses to the singlet state
W ⊗ Iψ 1
ψ̃ = = √ (|00 + |11) . (5.66)
||W ⊗ Iψ|| 2
Then, in this case, perfect teleportation can be performed. Thus, if Alice
and Bob teleported directly via the initial state, they would obtain a very
poor performance. Now they have a small but nonzero chance of performing
perfect teleportation.
• •
LOCC
❄
✲ (failure)
1−p
(success) p
❄
•\/\/\/\/\/\/\/\/•
Fig. 5.7. Conclusive increase of the singlet fraction. Alice and Bob obtain, with
a probability p of success, a state with a higher singlet fraction than that of the
initial state
Similarly to the usual form of teleportation, conclusive teleportation can

be reduced to conclusively increasing F (illustrated in Fig. 5.7), followed by
the original teleportation protocol. If in the first stage Alice and Bob obtain
a state with some F , then the second stage will produce the corresponding
fidelity f = (F d + 1)/(d + 1). Thus we can restrict our consideration to con-
clusively increasing the singlet fraction. The latter process was developed in
[343, 350]. An interesting peculiarity of conclusively increasing the singlet
fraction is that sometimes it is impossible to obtain F = 1, but it is still
possible to make F arbitrarily close to 1. However, if F → 1, then the prob-
ability of success tends to 0, so that, indeed, it is impossible to reach F = 1
[343].
Activation Protocol. Suppose that Alice and Bob share a single pair of
spin-1 particles in the following free entangled mixed state:
free = (F ) ≡ F |ψ+

ψ+ | + (1 − F )σ+ , 0 < F < 1 , (5.67)
where σ± are separable states given by (5.38). It is easy to see that the
state (5.67) is free entangled. Namely, after action of the local projections
(|0
0| + |1
1|) ⊗ (|0
0| + |1
1|), we obtain an entangled 2 ⊗ 2 state (its
entanglement can be revealed by calculating partial transposition). Thus, ac-

cording to Theorem 5.4, the state (5.67) is FE. By complicated considerations
one can show [343] that there is a threshold F0 < 1 that cannot be exceeded
in the process of conclusively increasing the singlet fraction. In other words,
Alice and Bob have no chance of obtaining a state with F ( ) > F0 (we
do not know the value F0 , we only know that such a number exists).
Suppose now that Alice and Bob share, in addition, a very large number
of pairs in the following BE state (the one considered in Sect. (5.3.6)):
2 α 5−α
σα = |ψ+
ψ+ | + σ+ + σ− . (5.68)
7 7 7
As stated in Sect. 5.3.6, for 3 < α ≤ 4 the state is BE. As we know, there is
no chance of obtaining even a pair with F > 1/3 from BE pairs of a 3 ⊗ 3
system. Now, it turns out that, if Alice and Bob have both an FE pair and
the BE pairs, they can apply a simple protocol to obtain an F arbitrarily
close to 1. Thus, owing to the connection between conclusive increasing of
the singlet fraction and conclusive teleportation, the fidelity of the latter can
be arbitrarily close to unity only if both an FE pair and BE pairs are shared.
The protocol [340] is similar to the recurrence distillation protocol de-
scribed in Sect. 19. It is an iteration of the following two steps:
(i) Alice and Bob take the free entangled pair, in the state free (F ), and one
of the pairs, which is in the state σα . They perform the bilateral XOR
operation UBXOR ≡ UXOR ⊗ UXOR , each of them treating the member
of the free entangled pair as a source and the member of the bound
entangled pair as a target.25
(ii) Alice and Bob measure the members of the source pair in the basis
|0, |1, |2. Then they compare their results via classical communication.
If the compared results differ from one another, they have to discard both
pairs, and then the trial of the improvement of F fails. If the results agree,
then the trial succeeds and they discard only the target pair, coming back
with (as we shall see) an improved source pair to the first step (i).
After some algebra, one can see that the success in the step (ii) occurs with
a nonzero probability
2F + (1 − F )(5 − α)
PF →F = , (5.70)
7
25
Here we need the quantum XOR gate not for two qubits, as in Sect. 19, but for
two qutrits (three-level systems). A general XOR operation for a d ⊗ d system,
which was used in in [281, 351], is defined as
UXOR |a|b = |a|(b + a)mod d , (5.69)
where the initial states |a and |b correspond to the source and target states,
respectively.
Fig. 5.8. Liberation of bound entanglement. The singlet fraction of the FE state is
plotted versus the number of successful iterations of (i) and (ii), and the parameter
α of the state α of the BE pairs used. The initial singlet fraction of the FE pair is
taken as Fin = 0.3 (This figure is reproduced from Phys. Rev. Lett. 82, 1056 (1999)
by permission of the authors)
leading to the transformation (F ) → (F ), where the improved fidelity is
2F
F (F ) = . (5.71)
2F + (1 − F )(5 − α)
If α > 3, then the above continuous function of F exceeds the value of F

on the whole region (0, 1). Thus the successful repetition of the steps (i) and
(ii) produces a sequence of source fidelities Fn → 1. In Fig. 5.8 we have plotted
the value of F obtained versus the number of iterations of the protocol and
the parameter α. For α ≤ 3 the singlet fraction goes down: separable states
cannot help to increase it. We can see the dramatic qualitative change at the
“critical”26 point that occurs at the borderline between separable states and
bound entangled ones (α = 3). On the other hand, it is surprising that there
26
The term “critical” that we have used here reflects the rapid character of the
change (see [307] for a similar “phase transition” between separable and FE
states). On the other hand, the present development of thermodynamic analogies
in entanglement processing [71, 268, 325, 335, 347] allows us to hope that in
future one will be able to build a synthetic theory of entanglement based on
thermodynamic analogies: then the “critical” point would become truly critical.
is no qualitative difference between the behavior of BE states (3 < α ≤ 4)

and FE states (4 < α ≤ 5). Here the change is only quantitative, while the
shape of the corresponding curves is basically the same. To our knowledge,
this is the only effect we know about where bound entanglement manifests its
quantumness.27 Since the effect is very subtle, one must conclude that bound
entanglement is essentially different from free entanglement and is enormously
weak. For recent results on the activation effect in the multiparticle cases see
[353].
Entanglement-Enhanced LOCC Operations. The activation effect sug-
gests that we should extend the paradigm of LOCC operations by including
quantum communication (under suitable control). Then we obtain entangle-
ment-enhanced LOCC (LOCC + EE) operations (see [354]). For example, if
we allow LOCC operations and an arbitrary amount of shared bound entan-
glement, we obtain the LOCC + BE paradigm. One can now ask about the
entanglement of formation28 and distillation in this regime. Since BE states
contain entanglement, even though it is very weak, then an infinite amount
of bound entanglement could make DLOCC+BE much larger than the usual
DLOCC : one might expect DLOCC+BE to be the maximal possible, indepen-
dently of the input state [355] (e.g. for two-qubit pairs, we would have
DLOCC+BE = 1 for any state). In [305, 356] it was shown that this is im-
possible. The argument of [305] is as follows. First, the authors recall that
F
DLOCC ≤ ELOCC [66]. Otherwise, it would be possible to increase entangle-
ment by means of LOCC actions. Indeed, suppose that for some state we
have DLOCC () > ELOCC F
(). Then Alice and Bob could take n two-qubit
pairs in a singlet state and produce n/ELOCCF
pairs of the state . Then they
could distill n(DLOCC /ELOCC ) singlets, which would be greater number than
F
n. A similar argument is applied to LOCC+ BE actions: the authors show

that it is impossible to increase the number of singlet pairs by LOCC + BE
actions, and conclude that DLOCC+BE ≤ ELOCC+BE F
. On the other hand, ob-
viously we have ELOCC+BE ≤ ELOCC . Combining the inequalities, we obtain
F F
the result that DLOCC+BE is bounded by the usual entanglement of formation

ELOCC
F
, which is maximal only for singlet-type states. A different argument
in [356] is based on the results of Rains [311] on a bound for the distillation
of entanglement (see Sect. 28). Thus, even if an infinite amount of BE pairs
is employed, LOCC + BE operations are not enormously powerful. However,
it is still possible that they are better than LOCC operations themselves, i.e.
we can conjecture that DLOCC+BE () > DLOCC () for some states .
Bounds for Entanglement of Distillation. Bound entanglement is an
achievement in qualitative description; however, as we could see in the previ-
ous section, it also has an impact on the quantitative approach. Here we shall
27
In the multipartite case, two other effects have recently been found [339, 352].
28 F
The entanglement of formation ELOCC () of a state is the number of input
singlet pairs per output pair needed to produce the state by LOCC operations
[66].
see that it has helped to obtain a strong upper bound for the entanglement
of distillation D (recall that the latter has the meaning of the capacity of a
noisy teleportation channel constituted by bipartite mixed states, and hence
is a central parameter of quantum communication theory).
The first upper bound for D was the entanglement of formation [66],
calculated explicitly for two-qubit states [301]. However, a stronger bound
has been provided in [269] (see also [334]). It is given by the following measure
of entanglement [268, 269] based on the relative entropy:
EVP () = inf S(|σ) , (5.72)

σ
where the infimum is taken over all separable states σ. The relative entropy
is defined by
S(|σ) = Tr log − Tr log σ .
Vedral and Plenio provided a complicated argument [269] showing that EVP is
an upper bound for D(), under the additional assumption that it is additive.
Even though we still do not know if it is indeed additive, Rains showed [311]
that it is a bound for D even without this assumption. He also obtained a
stronger bound by use of BE states (more precisely, PPT states). It appears
that, if the infimum in (5.72) is taken over PPT states (which are bound
entangled), the new measure ER is a bound for the distillable entanglement,
too. However, since the set of PPT states is strictly greater than the set
of separable states, the bound is stronger. For example, the entangled PPT
states have zero distillable entanglement. Since they are not separable, EVP
does not vanish for them, and hence the evaluation of D by means of EVP is
too rough. The Rains measure vanishes for these states.
We will not provide here the original proof of the Rains result. Instead
we demonstrate a general theorem on bounds for distillable entanglement
obtained in [357], which allows a major simplification of the proof of the
result.
Theorem 5.6. Any function B satisfying the conditions (a)–(c) below is an

upper bound for the entanglement of distillation:
(a) Weak monotonicity: B() ≥ B[Λ()], where Λ is a superoperator realizable
by means of LOCC operations.
(b) Partial subadditivity: B(⊗n ) ≤ nB().
(c) Continuity for an isotropic state (F, d): suppose that we have a sequence
of isotropic states (Fd , d) (see Sect. 5.2.4, (5.34)) such that Fd → 1 if d →
∞. Then we require
1
lim B[(Fd , d)] → 1 . (5.73)
d→∞ log d
Remark. If, instead of LOCC operations, we take another class C of oper-

ations including classical communication in at least one direction (e.g. the
LOCC + BE operations mentioned previously), the proof, mutatis mutandis,
also applies. (The condition (a) then involves the class C)
Proof. The main idea of the proof is to exploit the monotonicity condition.
We shall show that if D were greater than B then, during the distillation pro-
tocol, the function B would have to increase. But this cannot be so, because
distillation is a LOCC action, and hence B would violate the assumption (a).
By subadditivity, we have
1
B() ≥ B(⊗n ) . (5.74)
n
Distillation of n pairs aims at obtaining k pairs, each in nearly a singlet state.
The asymptotic rate is lim k/n. It was shown [311] that one can equally
well think of the final d ⊗ d system as being in a state close to P+d . The
asymptotic rate is now lim(log d)/n. Then the only relevant parameters of
the final state out are the dimension d and the fidelity F (out ). Thus the
distillation protocol can be followed by U ⊗U ∗ twirling, producing an isotropic
final state (d, F ) (see Sect. 5.2.4). By condition (a), distillation does not
increase B, and hence
1 1
B(⊗n ) ≥ B[(Fdn , dn )] . (5.75)
n n
Now, in the distillation process F → 1, and if we consider an optimal protocol,
then (log d)/n → D. Hence, by condition (c), the right-hand side of the
inequality tends to D(). Thus we obtain the result that B() ≥ D().
We should check, whether the Vedral–Plenio and Rains measures satisfy
the assumptions of the theorem. Subadditivity and weak monotonicity are
immediate consequence of the properties of the relative entropy used in the
definition of ER (subadditivity was proved in [268], and weak monotonicity
in [269]). The calculation of ER for an isotropic state is a little bit more
involved, but by using the high symmetry of the state, it was found to be
[311] EVP [(F, d)] = ER [(F, d)] = log d+F log F +(1−F ) log[(1−F )/(d−1)].
Evaluating now this expression for large d, we easily find that the condition
(c) is satisfied. The argument applies without any change to the Rains bound.
Finally, let us note that the Rains entanglement measure attributes no
entanglement to some entangled states (the PPT entangled ones). Normally
we would require that a natural postulate for an entanglement measure would
be that the entanglement measure should vanish if and only if the state is
separable. However, then we would have to remove distillable entanglement
from the set of measures. Indeed, the distillable entanglement vanishes for
some manifestly entangled states – bound entangled ones. Now the problem
is: should we keep the postulate, or keep D as a good measure?
It is reasonable to keep D as a good measure, as it has a direct physical

sense: it describes entanglement as a resource for quantum communication.
If it is not a measure, then we must conclude that we are not interested in
measures. Consequently, we adopt as a main “postulate” for an entangle-
ment measure the following statement: “Distillable entanglement is a good
measure”. So we must abandon the postulate. The apparent paradox can be
removed by realizing that we have different types of entanglement. Then a
given state, even though it is entangled, may not contain some particular type
of entanglement, and the measure that quantifies that type will attribute no
entanglement to the state.
5.4 Concluding Remarks
In contrast to the case of pure states, the problem of mixed-state entangle-

ment is “nondegenerate” in the sense that the various scalar and structural
separability criteria are not equivalent. There is a fundamental connection
between entanglement and positive maps, represented by Theorem 5.1. How-
ever, there is still a problem of turning it into an operational criterion for
higher-dimensional systems. Recently [358, 359] the question was reduced to
a problem of investigation of the so-called “edge” PPT entangled states, as
well as of the positive maps and entanglement witnesses detecting their en-
tanglement. Some operational criteria for low-rank density matrices (and also
for the multiparticle case) have been worked out in [360].
It is remarkable that the structure of entanglement reveals a discontinuity.
There are two qualitatively different types of entanglement: distillable, “free”
entanglement, and “bound” entanglement, which cannot be distilled. All the
two-qubit entangled states are free entangled. Moreover, a free entangled state
in any dimension must have some features of two-qubit entanglement. Bound
entanglement is practically useless for quantum communication. However, it
is not a marginal phenomenon, as the volume of the set of BE states in the
set of all states for finite dimension is nonzero.
The discovery of activation of bipartite bound entanglement suggested
[340] the nonadditivity of the corresponding quantum communication chan-
nels,29 in the sense that the distillable entanglement D(BE ⊗ EF ) could
exceed D(FE ) for some free entangled state FE and bound entangled state
BE . Quite recently it has been shown [339] that, in the multipartite case, two
different bound entangled states, if tensored together, can make a distillable
state: D(1BE ⊗ 2BE ) > D(1BE ) + D(2BE ) = 0. This new nonclassical effect
was called superactivation. On the other hand, in [352] it was shown that
the four-party “unlockable” bound entangled states [361] can be used for re-
mote concentration of quantum information. It is intriguing that for bipartite
29
This could be then reformulated in terms of the so-called binding entanglement
channels [305, 355].
systems, with the exception of the activation effect, bound entanglement is

permanently passive. In general, there may be a qualitative difference between
bipartite bound entanglement and the multipartite form. Still, in the light of
recent results [362], it is quite possible that bipartite bound entanglement is
also nonadditive. The very recent investigations of bound entanglement for
continuous variables [363, 364] also raise analogous questions in this latter
domain.
As we have seen, there is a basic connection between bound entanglement
and irreversibility. As a consequence, it would be interesting to investigate
some dynamical features of BE. It cannot be excluded that some systems in-
volving BE states may reveal a nonstandard (nonexponential) decay of entan-
glement. In general, it seems that the role of bound entanglement in quantum
communication will be negative: in fact, the existence of BE constitutes a fun-
damental restriction on entanglement processing. One can speculate that it is
the ultimate restriction in the context of distillation, i.e. that it may allow one
to determine the value of the distillable entanglement. Hence it seems impor-
tant to develop an approach combining BE and the entanglement measures
involving relative entropy. It also seems reasonable to conjecture that, in the
case of general distillation processes involving the conversion of mixed states
→ [66], the bound entanglement EB never decreases30 (i.e. ∆EB ≥ 0) in
optimal processes.
The irreversibility inherently connected with distillation encourages us to
develop some natural formal analogies between mixed-state entanglement
processing and phenomenological thermodynamics. The construction of a
“thermodynamics of entanglement” (cf. [325, 335, 347, 365]) would be es-
sential for a synthetic understanding of entanglement processing. Of course,
progress in the above directions would require the development of various
techniques of searching for bound entangled states.
One of the challenges of mixed-state entanglement theory is to determine
which states are useful for quantum communication with given additional
resources. In particular, we still do not know (i) which states are distillable
under LOCC (i.e. which states are free entangled), and (ii) which states are
distillable under one-way classical communication and local operations.
A promising direction for mixed-state entanglement theory is its appli-
cation to the theory of quantum channel capacity, pioneered in [66]. In par-
ticular, the methods leading to upper bounds for distillable entanglement
described in Sect. 28 allow one to obtain upper bounds for quantum channel
capacities [101] (one of them was obtained earlier [97]). It has been shown
[101] that the following hypothetical inequality
D1 () ≥ S(B ) − S() , (5.76)

30
The bound entanglement can be quantified [71] as the difference between the
entanglement of formation and the entanglement of distillation (defined within
the original distillation scheme): EB = EF − D.
where D1 () is the one-way distillable entanglement,31 would imply equality

between the capacity of a quantum channel and the maximal rate of coher-
ent information [366]. The latter equality would be nothing but a quantum
Shannon theorem, with coherent information being the counterpart of mu-
tual information. All the results obtained so far in the domain of quantifying
entanglement indicate that the inequality is true. However, a proof of the
inequality has not been found so far.
Finally, one would like to have a clear connection between entanglement
and its basic manifestation – nonlocality. One can assume that free entangled
states exhibit nonlocality via a distillation process [265, 266]. However, the
question concerning the possible nonlocality of BE states remains open (see
[367, 368, 369]).
To answer the above and many other questions, one must develop the
mathematical description of the structure of mixed-state entanglement. In
this context, it would be especially important to push forward the mathe-
matics of positive maps. One hopes that the exciting physics connected with
mixed-state entanglement that we have presented in this contribution will
stimulate progress in this domain.
31
Classical messages can be sent only from Alice to Bob during distillation.

(Springer Tracts in Modern Physics 173) Gernot Alber, Thomas Beth, Michał Horodecki, Paweł Horodecki, Ryszard Horodecki, Martin Rötteler, Harald Weinfurter, Reinhard Werner, Anton Zeilinger (auth.) - .pdf

Uploaded by

Copyright:

Available Formats

(Springer Tracts in Modern Physics 173) Gernot Alber, Thomas Beth, Michał Horodecki, Paweł Horodecki, Ryszard Horodecki, Martin Rötteler, Harald Weinfurter, Reinhard Werner, Anton Zeilinger (auth.) - .pdf

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Springer Tracts in Modern Physics 173) Gernot Alber, Thomas Beth, Michał Horodecki, Paweł Horodecki, Ryszard Horodecki, Martin Rötteler, Harald Weinfurter, Reinhard Werner, Anton Zeilinger (auth.) - .pdf

Uploaded by

Copyright:

Available Formats

1 From the Foundations of Quantum Theory

to Quantum Technology – an Introduction

Nowadays, the new technological prospects of processing quantum informa-

G. Alber, T. Beth, M. Horodecki, P. Horodecki, R. Horodecki, M. Rötteler, H. Weinfurter,

mation processing reﬂect a profound change in the general attitude towards

1.1 Early Developments

Many of the current methods and developments in the processing of quantum

1.1.1 Entanglement and Local, Realistic Theories

In order to clarify the characteristic diﬀerences between quantum mechan-

such as a photon pair, is produced by a source s. Polarization properties of

It should be mentioned that this relation is counterfactual [26] in the sense

which is a variant of Bell’s inequality and which is due to Clauser, Horne,

âi b̂j QM = ψ|âi b̂j |ψ = −αi · β j . (1.5)

So far, numerous experiments testing and supporting violations of Bell’s

âi = αi · σ A , b̂j = β j · σ B and ĉk = γ k · σ C . The GHZ state of (1.7) fulﬁlls

1.1.2 Characteristic Quantum Eﬀects for Practical Purposes

According to a suggestion of Feynman [40], quantum systems are not only of

|1|ϕ|a → |1|1|a1  , (1.11)

A(lice)’s direction i 1 2 1 1 2 1 2 2 1 2 ···

1.1.3 Quantum Algorithms

Feynman’s suggestion also indicates interesting links between quantum phy-

Quantum mechanically, the situation is diﬀerent. The 2n possible binary

| ψ1 > |ψ2 >

is applied to all of the ﬁrst n qubits. We denote by H (i) the application

If in step 3 the quantum system is found in state |ψ1 , f is constant, otherwise

1.2 Quantum Physics and Information Processing

Quantum information and quantum computers have received a lot of public

2.2 What is Quantum Information?

Quantum information is that kind of information which is carried

So a “transmitter” of quantum information is nothing but a device preparing

of “information” used in everyday language, namely the kind available at an

2.3 Impossible Machines

The usefulness of considering impossible machines is well known from ther-

machine in this hierarchy, we are certain that teleportation is likewise im-

2.3.1 The Quantum Copier

This is the machine referred to in the well-known paper of Wootters and

2.3.2 Joint Measurement

Fig. 2.3. Obtaining joint measurements from a copier

2.3.3 Bell’s Telephone

performing superluminal communication using only correlations of the type

Fig. 2.4. Building Bell’s Telephone from a joint measurement

task forbidden by causality, and hence is impossible in general. This is the

β = C(A1 , B1 ) + C(A1 , B2 ) + C(A2 , B1 ) − C(A2 , B2 ) (2.1)

carries special signiﬁcance, as we shall see below. Because the inequality β ≤ 2

the probabilities pi (ai , b1 , b2 ) = P(ai , (b1 , b2 ) | Ai , B1 &B2 ). The condition

case in√experiments conducted to determine β (e.g. [73]), which give roughly

2.3.4 Entanglement, Mixed-State Analyzers

Violations of Bell’s inequalities can also be seen to prove the existence of a

tion a hidden-variable theory. The hidden variable in this theory is usually

2.4 Possible Machines

on average is close to the density matrix describing the initial preparation.

Fig. 2.5. Classical teleportation with multiple inputs, or state estimation

Given a good estimator we can, of course, proceed to good cloning by just

2.4.2 Quantum Cryptography

It may seem impossible to ﬁnd applications of impossible machines. But that

2.4.3 Entanglement-Assisted Teleportation

in Chap. 3 by Weinfurter and Zeilinger, who represent one team in which a

2.4.4 Superdense Coding

It is easy to see and in fact is a commonplace occurrence that classical in-

2.4.5 Quantum Computation

âi b̂j QM = ψ|âi b̂j |ψ = −αi · β j . (1.5)

|1|ϕ|a → |1|1|a1 , (1.11)

If in step 3 the quantum system is found in state |ψ1 , f is constant, otherwise

T (X) = V ∗ (X ⊗ 1I )V , (2.9)

and the vectors of the form (X ⊗ 1I )V φ, where X ∈ Mn and φ ∈ Cm , span

and deﬁne Kraus operators Kx for T by φ, Kx ψ = φ ⊗ χx , V ψ (we leave

It is clear that σ must be the puriﬁcation of ρ, restricted to the ﬁrst factor.

T cb := sup T ⊗ In , (2.14)

∆(S, T ) = inf S − ET Dcb , (2.15)