Mathematical
Aspects of Quantum
Computing 2007
Kinki University Series on Quantum Computing
EditorinChief: Mikio Nakahara (Kinki University, Japan)
ISSN: 17937299
Published
Vol. 1 Mathematical Aspects of Quantum Computing 2007
edited by Mikio Nakahara, Robabeh Rahimi (Kinki Univ., Japan) &
Akira SaiToh (Osaka Univ., Japan)
ZhangJi  Math'l Aspects of Quan.pmd 3/11/2008, 2:05 PM 2
NE W J E RSE Y • L ONDON • SI NGAP ORE • BE I J I NG • SHANGHAI • HONG KONG • TAI P E I • CHE NNAI
World Scientifc
Kinki University Series on Quantum Computing – Vol. 1
Mathematical
Aspects of Quantum
Computing 2007
editors
Mikio Nakahara
Robabeh Rahimi
Kinki University, Japan
Akira SaiToh
Osaka University, Japan
British Library CataloguinginPublication Data
A catalogue record for this book is available from the British Library.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN13 9789812814470
ISBN10 9812814477
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
Copyright © 2008 by World Scientific Publishing Co. Pte. Ltd.
Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Printed in Singapore.
MATHEMATICAL ASPECTS OF QUANTUM COMPUTING 2007
Kinki University Series on Quantum Computing — Vol. 1
ZhangJi  Math'l Aspects of Quan.pmd 3/11/2008, 2:05 PM 1
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
v
PREFACE
This volume contains lecture notes and poster contributions presented
at the summer school “Mathematical Aspects of Quantum Computing”,
held from 27 to 29 August, 2007 at Kinki University in Osaka, Japan.
The aims of this summer school were to exchange and share ideas among
researchers working in various ﬁelds of quantum computing particularly
from a mathematical point of view, and to motivate students to tackle the
advanced subjects of quantum computing.
Invited speakers to the summer school were asked to prepare their lec
ture notes in a selfcontained way and therefore we expect that each contri
bution will be useful for students and researchers even with less background
to the topic.
This summer school was supported by “Open Research Center” Project
for Private Universities: matching fund subsidy from MEXT (Ministry of
Education, Culture, Sports, Science and Technology).
We yearn for continued outstanding success in this ﬁeld and wish to
expand summer schools and conferences further in the ﬁeld of quantum
computing based at Kinki University.
Finally, we would like to thank Ms Zhang Ji of World Scientiﬁc for her
excellent editorial work.
Osaka, January 2008
Mikio Nakahara
Robabeh Rahimi
Akira SaiToh
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
This page intentionally left blank This page intentionally left blank
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
vii
Summer School on
Mathematical Aspects of Quantum Computing
Kinki Univerisity, Osaka, Japan
27 – 29 August 2007
27 August
Mikio Nakahara (Kinki Univerisity, Japan)
Quantum Computing: An Overview
Kazuhiro Sakuma (Kinki University, Japan)
Braid Group and Topological Quantum Computing
28 August
Damian J. H. Markham (Tokyo University, Japan)
An Introduction to Entanglement Theory
Shogo Tanimura (Kyoto University, Japan)
Holonomic Quantum Computing and Its Optimization
29 August
Sahin Kaya Ozdemir (Osaka University, Japan)
Playing Games in Quantum Mechanical Settings: Features of Quan
tum Games
Manabu Hagiwara (National Institute of Advanced Industrial Science and
Technology, Japan)
Quantum Error Correction Codes
Poster Presentations
Vahideh Ebrahimi (Kinki University, Japan)
Controlled Teleportation of an Arbitrary Unknown TwoQubit En
tangled State
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
viii
Yukihiro Ota (Kinki University, Japan)
Notes on D¨ urCirac Classiﬁcation
Robabeh Rahimi (Kinki University, Japan)
BangBang Control of Entanglement in SpinBusBoson Model
Akira SaiToh (Osaka University, Japan)
Numerical Computation of TimeDependent Multipartite Nonclas
sical Correlation
Takuya Yamano (Ochanomizu University, Japan)
On Classical NoCloning Theorem under Liouville Dynamics and
Distances
M
a
r
c
h
1
1
,
2
0
0
8
8
:
5
9
W
S
P
C

P
r
o
c
e
e
d
i
n
g
s
T
r
i
m
S
i
z
e
:
9
i
n
x
6
i
n
M
a
t
h
e
m
a
t
i
c
a
l
A
s
p
e
c
t
s
i
x
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
x
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
xi
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
This page intentionally left blank This page intentionally left blank
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
xiii
LIST OF PARTICIPANTS
Aoki, Takashi Kinki University, Japan
Chinen, Koji Osaka City University, Japan
Ebrahimi Bakhtavar, Vahideh Kinki University, Japan
Hagiwara, Manabu National Institute of Advanced
Industrial Science and
Technology, Japan
Ikeda, Yasushi Kyoto University, Japan
Inoue, Kaiki Taro Kinki University, Japan
Kikuta, Toshiyuki Kinki University, Japan
Kobata, Kumi Kinki University, Japan
Kondo, Yasushi Kinki University, Japan
Markham, Damian James Harold Tokyo University, Japan
Minami, Kaori Kinki University, Japan
Nagaoka, Shoyu Kinki University, Japan
Nakagawa, Nobuo Kinki University, Japan
Nakahara, Mikio Kinki University, Japan
Ootsuka, Takayoshi Kinki University, Japan
Ota, Yukihiro Kinki University, Japan
Ozdemir, Sahin Kaya Osaka University, Japan
Rahimi Darabad, Robabeh Kinki University, Japan
SaiToh, Akira Osaka University, Japan
Sakuma, Kazuhiro Kinki University, Japan
Sasano, Hiroshi Kinki University, Japan
Segawa, Etsuo Yokohama National University, Japan
Sugita, Ayumu Osaka City University, Japan
Tanimura, Shogo Kyoto University, Japan
Tomita, Hiroyuki Kinki University, Japan
Unoki, Makoto Waseda University, Japan
Yamano, Takuya Ochanomizu University, Japan
Yamanouchi, Akiko ITOCHU TechnoSolutions Corpora
tions, Japan
Yoshida, Motoyuki Waseda University, Japan
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
This page intentionally left blank This page intentionally left blank
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
xv
CONTENTS
Preface v
Quantum Computing: An Overview 1
M. Nakahara
Braid Group and Topological Quantum Computing 55
T. Ootsuka, K. Sakuma
An Introduction to Entanglement Theory 91
D. J. H. Markham
Holonomic Quantum Computing and Its Optimization 115
S. Tanimura
Playing Games in Quantum Mechanical Settings: Features of
Quantum Games 139
S¸. K.
¨
Ozdemir, J. Shimamura, N. Imoto
Quantum ErrorCorrecting Codes 181
M. Hagiwara
—Poster Summaries—
Controled Teleportation of an Arbitrary Unknown TwoQubit
Entangled State 213
V. Ebrahimi, R. Rahimi, M. Nakahara
Notes on the D¨ ur–Cirac Classiﬁcation 215
Y. Ota, M. Yoshida, I. Ohba
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
xvi
BangBang Control of Entanglement in SpinBusBoson Model 217
R. Rahimi, A. SaiToh, M. Nakahara
Numerical Computation of TimeDependent Multipartite
Nonclassical Correlation 219
A. SaiToh, R. Rahimi, M. Nakahara, M. Kitagawa
On Classical NoCloning Theorem Under Liouville Dynamics
and Distances 221
T. Yamano, O. Iguchi
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
1
QUANTUM COMPUTING: AN OVERVIEW
MIKIO NAKAHARA
Department of Physics, Kinki University, HigashiOsaka 5778502, Japan
Email: nakahara@math.kindai.ac.jp
Elements of quantum computing and quantum infromation processing are in
troduced for nonspecialists. Subjects inclulde quantum physics, qubits, quan
tum gates, quantum algorithms, decoherece, quantum error correcting codes
and physical realizations. Presentations of these subjects are as pedagogical as
possible. Some sections are meant to be brief introductions to contributions by
other lecturers.
Keywords: Quantum Physics, Qubits, Quantum Gates, Quantum Algorithms.
1. Introduction
Quantum computing and quantum information processing are emerging dis
ciplines in which the principles of quantum physics are employed to store
and process information. We use the classical digital technology at almost
every moment in our lives: computers, mobile phones, mp3 players, just
to name a few. Even though quantum mechanics is used in the design of
devices such as LSI, the logic is purely classical. This means that an AND
circuit, for example, produces deﬁnitely 1 when the inputs are 1 and 1. One
of the most remarkable aspects of the principles of quantum physics is the
superposition principle by which a quantum system can take several diﬀer
ent states simultaneously. The input for a quantum computing device may
be a superposition of many possible inputs, and accordingly the output is
also a superposition of the corresponding output states. Another aspect of
quantum physics, which is far beyond the classical description, is entangle
ment. Given several objects in a classical world, they can be described by
specifying each object separately. Given a group of ﬁve people, for example,
this group can be described by specifying the height, color of eyes, personal
ity and so on of each constituent person. In a quantum world, however, only
a very small subset of all possible states can be described by such individual
speciﬁcations. In other words, most quantum states cannot be described by
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
2
such individual speciﬁcations, thereby being called “entangled”. Why and
how these two features give rise to the enormous computational power in
quantum computing and quantum information processing will be explained
in this contribution.
A part of this lecture note is based on our forthcoming book.
1
General
references are [2–4].
2. Quantum Physics
2.1. Notation and conventions
We will exclusively work with a ﬁnitedimensional complex vector space C
n
with an inner product , (Hilbert spaces). A vector in C
n
is called a
ket vector or a ket and is denoted as
x =
x
1
.
.
.
x
n
x
i
∈ C
while a vector in the dual space C
n∗
is called a bra vector or a bra and
denoted
α = (α
1
, . . . , α
n
) α
i
∈ C.
Index i sometimes runs from 0 to n−1. The inner product of x and α is
αx =
n
i=1
α
i
x
i
.
This inner product naturally introduces a correspondence
x = (x
1
, . . . , x
n
)
t
↔ x = (x
∗
1
, . . . , x
∗
n
),
by which an inner product of two vectors are deﬁned as xy =
n
i=1
x
∗
i
y
i
.
The inner product naturally deﬁnes the norm of a vector x as x =
xx.
Pauli matrices are generators of su(2) and denoted
σ
x
=
0 1
1 0
, σ
y
=
0 −i
i 0
, σ
z
=
1 0
0 −1
in the basis in which σ
z
is diagonalized. Symbols X = σ
x
, Y = −iσ
y
and
Z = σ
z
are also employed.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
3
Let A be an m×n matrix and B be a p ×q matrix. Then
A ⊗B =
a
11
B, a
12
B, . . . , a
1n
B
a
21
B, a
22
B, . . . , a
2n
B
. . .
a
m1
B, a
m2
B, . . . , a
mn
B
is an (mp) ×(nq) matrix called the tensor product of A and B.
2.2. Axioms of quantum mechanics
Quantum mechanics was discovered roughly a century ago.
5–10
In spite of
its long history, the interpretation of the wave function remains an open
question. Here we adopt the most popular one, called the Copenhagen in
terpretation.
A 1 A pure state in quantum mechanics is represented by a normalized vec
tor ψ in a Hilbert space H associated with the system. If two states
ψ
1
and ψ
2
are physical states of the system, their linear superpo
sition c
1
ψ
1
+ c
2
ψ
2
(c
k
∈ C), with
2
i=1
c
i

2
= 1, is also a possible
state of the same system (superposition principle).
A 2 For any physical quantity (observable) a, there exists a correspond
ing Hermitian operator A acting on H. When a measurement of a is
made, the outcome is one of the eigenvalues λ
j
of A. Let λ
1
and λ
2
be two eigenvalues of A: Aλ
i
= λ
i
λ
i
. Consider a superposition state
c
1
λ
1
+ c
2
λ
2
. If we measure a in this state, the state undergoes an
abrupt change (wave function collapse) to one of the eigenstates λ
i
corresponding to the observed eigenvalue λ
i
. Suppose we prepare many
copies of the state c
1
λ
1
+ c
2
λ
2
. The probability of collapsing to the
state λ
i
is given by c
i

2
(i = 1, 2). The complex coeﬃcient c
i
is called
the probability amplitude in this sense. It should be noted that a mea
surement produces one outcome λ
i
and the probability of obtaining
it is experimentally evaluated only after repeating measurements with
many copies of the same state. These statements are easily generalized
to superposition states of more than two states.
A 3 The time dependence of a state is governed by the Schr¨ odinger equation
i
∂ψ
∂t
= Hψ, (1)
where is a physical constant known as the Planck constant and H is a
Hermitian operator (matrix) corresponding to the energy of the system
and is called the Hamiltonian.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
4
Several comments are in order.
• In Axiom A 1, the phase of the vector may be chosen arbitrarily; ψ
in fact represents the “ray” {e
iα
ψ α ∈ R}. This is called the ray
representation. The overall phase is not observable and has no physical
meaning.
• Axiom A 2 may be formulated in a diﬀerent but equivalent way as
follows. Suppose we would like to measure an observable a. Let the
spectral decomposition of the corresponding operator A be
A =
i
λ
i
λ
i
λ
i
, where Aλ
i
= λ
i
λ
i
.
Then the expectation value A of a after measurements with respect
to many copies of ψ is
A = ψAψ. (2)
Let us expand ψ in terms of λ
i
as ψ =
i
c
i
λ
i
. According to
A 2, the probability of observing λ
i
upon measurement of a is c
i

2
and
therefore the expectation value after many measurements is
i
λ
i
c
i

2
.
If, conversely, Eq. (2) is employed, we will obtain the same result since
ψAψ =
i,j
c
∗
j
c
i
λ
j
Aλ
i
=
i,j
λ
i
c
∗
j
c
i
δ
ij
=
i
λ
i
c
i

2
.
This measurement is called the projective measurement. Any particular
outcome λ
i
will be found with the probability
c
i

2
= ψP
i
ψ, (3)
where P
i
= λ
i
λ
i
 is the projection operator and the state immediately
after the measurement is λ
i
or equivalently
P
i
ψ/
ψP
i
ψ. (4)
• The Schr¨ odinger equation (1) in Axiom A 3 is formally solved to yield
ψ(t) = e
−iHt/
ψ(0), (5)
if the Hamiltonian H is timeindependent, while
ψ(t) = T exp
−
i
t
0
H(t)dt
ψ(0) (6)
if H depends on t, where T is the timeordering operator. The state
at t > 0 is ψ(t) = U(t)ψ(0). The operator U(t) : ψ(0) → ψ(t),
called the timeevolution operator, is unitary. Unitarity of U(t) guar
antees that the norm of ψ(t) is conserved: ψ(0)U
†
(t)U(t)ψ(0) =
ψ(0)ψ(0) = 1 (∀t > 0).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
5
Two mutually commuting operators A and B have simultaneous eigen
states. If, in contrast, they do not commute, the measurement outcomes of
these operators on any state ψ satisfy the following uncertainty relations.
Let A = ψAψ and B = ψBψ be their respective expectation
values and ∆A =
(A − A)
2
and ∆B =
(B − B)
2
be respective
standard deviations. Then they satisfy
∆A∆B ≥
1
2
ψ[A, B]ψ. (7)
2.3. Simple example
Examples to clarify the axioms introduced in the previous subsection are
given. They are used to controll quantum states in physical realizations of a
quantum computer. A spin1/2 particle has two states, which we call spin
up state  ↑ and spindown state  ↓. It is common to assign components
 ↑ = (1, 0)
t
and  ↓ = (0, 1)
t
. They form a basis of a vector space C
2
.
Let us consider a timeindependent Hamiltonian
H = −
2
ωσ
x
(8)
acting on the spin Hilbert space C
2
. Suppose the system is in the eigenstate
of σ
z
with the eigenvalue +1 at time t = 0; ψ(0) =  ↑. The wave function
ψ(t) (t > 0) is then found from Eq. (5) as
ψ(t) = exp
i
ω
2
σ
x
t
ψ(0) =
cos ωt/2 i sin ωt/2
i sinωt/2 cos ωt/2
1
0
=
cos ωt/2
i sin ωt/2
= cos
ω
2
t ↑ +i sin
ω
2
t ↓. (9)
Suppose we measure σ
z
in ψ(t). The spin is found spinup with probability
P
↑
(t) = cos
2
(ωt/2) and spindown with probability P
↓
(t) = sin
2
(ωt/2).
Consider a more general Hamiltonian
H = −
2
ωˆ n · σ, (10)
where ˆ n is a unit vector in R
3
. The timeevolution operator is readily
obtained, by making use of a well known formula
e
iα(
ˆ
n·σ)
= cos αI +i(ˆ n · σ) sin α (11)
as
U(t) = exp(−iHt/) = cos ωt/2 I +i(ˆ n · σ) sin ωt/2. (12)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
6
Suppose the initial state is ψ(0) = (1, 0)
t
for example. Then we ﬁnd, at a
later time t > 0,
ψ(t) = U(t)ψ(0) =
cos(ωt/2) + in
z
sin(ωt/2)
i(n
x
+ in
y
) sin(ωt/2)
. (13)
2.4. Multipartite system, tensor product and entangled
state
So far, we have implictly assumed that the system is made of a single
component. Suppose a system is made of two components, one lives in a
Hilbert space H
1
and the other in H
2
. A system composed of two separate
components is called bipartite. The system as a whole lives in a Hilbert
space H = H
1
⊗ H
2
, whose general vector is written as
ψ =
i,j
c
ij
e
1,i
⊗ e
2,j
, (14)
where {e
a,i
} (a = 1, 2) is an orthonormal basis in H
a
and
i,j
c
ij

2
= 1.
A state ψ ∈ H written as a tensor product of two vectors as ψ =
ψ
1
⊗ψ
2
, (ψ
a
∈ H
a
) is called a separable state or a tensor product state.
A separable state admits a classical interpretation “The ﬁrst system is in
the state ψ
1
while the second system is in ψ
2
”. It is clear that the set
of separable state has dimension dimH
1
+dimH
2
. Note, however, that the
total space H has diﬀerent dimension than this: dimH = dimH
1
dimH
2
.
This number is considerably larger than the dimension of the sparable states
when dimH
a
(a = 1, 2) are large. What are the missing states then? Let
us consider a spin state
ψ =
1
√
2
( ↑ ⊗  ↑ +  ↓ ⊗  ↓) (15)
of two electrons. Suppose ψ may be decomposed as
ψ = (c
1
 ↑ + c
2
 ↓) ⊗ (d
1
 ↑ + d
2
 ↓)
= c
1
d
1
 ↑ ⊗  ↑ + c
1
d
2
 ↑ ⊗  ↓ + c
2
d
1
 ↓ ⊗  ↑ + c
2
d
2
 ↓ ⊗  ↓.
However this decomposition is not possible since we must have c
1
d
2
=
c
2
d
1
= 0, c
1
d
1
= c
2
d
2
= 1/
√
2 simultaneously and it is clear that the above
equations have no common solution, showing ψ is not separable.
Such nonseparable states are called entangled. Entangled states refuse
classical descriptions. Entanglement is used extensively as a powerful com
putational resource in the following.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
7
Suppose a bipartite state (14) is given. We are interested in when the
state is separable and when entangled. The criterion is given by the Schmidt
decomposition of ψ.
Theorem 2.1. Let H = H
1
⊗H
2
be the Hilbert space of a bipartite system.
Then a vector ψ ∈ H admits the Schmidt decomposition
ψ =
r
i=1
√
s
i
f
1,i
⊗f
2,i
, (16)
where s
i
> 0 are called the Schmidt coeﬃcients satisfying
i
s
i
= 1 and
{f
a,i
} is an orthonormal set of H
a
. The number r ∈ N is called the
Schmidt number of ψ.
It follows from the above theorem that a bipartite state ψ is separable
if and only if its Schmidt number r is 1. See
1
for the proof.
2.5. Mixed states and density matrices
It might happen in some cases that a quantum system under considertation
is in the state ψ
i
with a probability p
i
. In other words, we cannot say
deﬁnitely which state the system is in. Therefore some random nature comes
into the description of the system. Such a system is said to be in a mixed
state while a system whose vector is uniquely speciﬁed is in a pure state.
A pure state is a special case of a mixed state in which p
i
= 1 for some i
and p
j
= 0 (j = i).
A particular state ψ
i
∈ H appears with probability p
i
in an ensemble
of a mixed state, in which case the expectation value of the observable a
is ψ
i
Aψ
i
. The mean value of a averaged over the ensemble is then given
by
A =
N
i=1
p
i
ψ
i
Aψ
i
, (17)
where N is the number of available states. Let us introduce the density
matrix by
ρ =
N
i=1
p
i
ψ
i
ψ
i
. (18)
Then Eq. (17) is rewritten in a compact form as A = Tr(ρA).
Let A be a Hermitian matrix. A is called positivesemideﬁnite if
ψAψ ≥ 0 for any ψ ∈ H. It is easy to show all the eigenvalues of
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
8
a positivesemideﬁnite Hermitian matrix are nonnegative. Conversely, a
Hermitian matrix A whose every eigenvalue is nonnegative is positive
semideﬁnite.
Properties which a density matrix ρ satisﬁes are very much like axioms
for pure states.
A 1’ A physical state of a system, whose Hilbert space is H, is completely
speciﬁed by its associated density matrix ρ : H → H. A density matrix
is a positivesemideﬁnite Hermitian operator with tr ρ = 1, see remarks
below.
A 2’ The mean value of an observable a is given by
A = tr (ρA). (19)
A 3’ The temporal evolution of the density matrix follows the Liouvillevon
Neumann equation
i
d
dt
ρ = [H, ρ] (20)
where H is the system Hamiltonian, see remarks below.
Several remarks are in order.
• The density matrix (18) is Hermitian since p
i
∈ R. It is positive
semideﬁnite since ψρψ =
i
p
i
ψ
i
ψ
2
≥ 0.
• Each ψ
i
follows the Schr¨ odinger equation i
d
dt
ψ
i
= Hψ
i
in a closed
quantum system. Its Hermitian conjugate is −i
d
dt
ψ
i
 = ψ
i
H. We
prove the Liouvillevon Neumann equation from these equations as
i
d
dt
ρ = i
d
dt
i
p
i
ψ
i
ψ
i
 =
i
p
i
Hψ
i
ψ
i
−
i
p
i
ψ
i
ψ
i
H = [H, ρ].
We denote the set of all possible density matrices as S(H).
Example 2.1. A pure state ψ is a special case in which the corresponding
density matrix is ρ = ψψ. Therefore ρ is nothing but the projection
operator onto the state. Observe that A = tr ρA =
i
e
i
ψψAe
i
=
ψA
i
e
i
e
i
ψ = ψAψ, where {e
i
} is an orthonormal set.
Let us consider a beam of photons. We take a horizontally polarized
state e
1
=  ↔ and a vertically polarized state e
2
=  as orthonormal
basis vectors. If the photons are a totally uniform mixture of two polarized
states, the density matrix is given by
ρ =
1
2
e
1
e
1
 +
1
2
e
2
e
2
 =
1
2
1 0
0 1
=
1
2
I.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
9
This state is called a maximally mixed state.
If photons are in a pure state ψ = (e
1
+e
2
)/
√
2, the density matrix,
with {e
i
} as basis, is
ρ = ψψ =
1
2
1 1
1 1
.
We are interested in when ρ represents a pure state or a mixed state.
Theorem 2.2. A state ρ is pure if and only if tr ρ
2
= 1.
Proof: Since ρ is Hermitian, all its eigenvalues λ
i
(1 ≤ i ≤ dimH) are real
and the corresponding eigenvectors {λ
i
} are made orthonormal. Then
ρ
2
=
i,j
λ
i
λ
j
λ
i
λ
i
λ
j
λ
j
 =
i
λ
2
i
λ
i
λ
i
. Therefore tr ρ
2
=
i
λ
2
i
≤
λ
max
i
λ
i
= λ
max
≤ 1, where λ
max
is the largest eigenvalue of ρ. There
fore tr ρ
2
= 1 implies λ
max
= 1 and all the other eigenvalues are zero. The
converse is trivial.
We classify mixed states into three classes, similarly to the classiﬁcation
of pure states into separable states and entangled states. We use a bipartite
system in the deﬁnition but generalization to multipartitle systems should
be obvious.
Deﬁnition 2.1. A state ρ is called separable if it is written in the form
ρ =
i
p
i
ρ
1,i
⊗ ρ
2,i
, (21)
where 0 ≤ p
i
≤ 1 and
i
p
i
= 1. It is called inseparable, if ρ does not admit
the decomposition (21).
It is important to realize that only inseparable states have quantum cor
relations analogous to entangled pure states. It does not necessarily imply
that a separable state has no nonclassical correlation. It is pointed out that
useful nonclassical correlation exists in the subset of separable states.
11
In the next subsection, we discuss how to ﬁnd whether a given bipartite
density matrix is separable or inseparable.
2.6. Negativity
Let ρ be a bipartite state and deﬁne the partial transpose ρ
pt
of ρ with
respect to the second Hilbert space as
ρ
ij,kl
→ ρ
il,kj
, (22)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
10
where ρ
ij,kl
= (e
1,i
 ⊗e
2,j
) ρ (e
1,k
⊗e
2,l
). Here {e
1,k
} is the orthonor
mal basis of the ﬁrst system while {e
2,k
} of the second system. Suppose
ρ takes a separable form (21). Then the partial transpose yields
ρ
pt
=
i
p
i
ρ
1,i
⊗ ρ
t
2,i
. (23)
Note here that ρ
t
for any density matrix ρ is again a density matrix since
it is still positive semideﬁnite Hermitian with unit trace. Therefore the
partial transposed density matrix (23) is another density matrix. It was
conjectured by Peres
12
and subsequently proven by the Hordecki family
13
that positivity of the partially transposed density matrix is necessary and
suﬃcient condition for ρ to be separable in the cases of C
2
⊗C
2
systems and
C
2
⊗C
3
systems. Conversely, if the partial transpose of ρ of these systems
is not a density matrix, then ρ is inseparable. Instead of giving the proof,
we look at the following example.
Example 2.2. Let us consider the Werner state
ρ =
1−p
4
0 0 0
0
1+p
4
−
p
2
0
0 −
p
2
1+p
4
0
0 0 0
1−p
4
, (24)
where 0 ≤ p ≤ 1. Here the basis vectors are arranged in the order
e
1,1
e
2,1
, e
1,1
e
2,2
, e
1,2
e
2,1
, e
1,2
e
2,2
.
Partial transpose of ρ yields
ρ
pt
=
1−p
4
0 0 −
p
2
0
1+p
4
0 0
0 0
1+p
4
0
−
p
2
0 0
1−p
4
.
ρ
pt
must have nonnegative eigenvalues to be a physically acceptable state.
The characteristic equation of ρ
pt
is
D(λ) = det(ρ
pt
− λI) =
λ −
p + 1
4
3
λ −
1 − 3p
4
= 0.
There are threefold degenerate eigenvalue λ = (1+p)/4 and nondegenerate
eigenvalue λ = (1 − 3p)/4. This shows that ρ
pt
is an unphysical state for
1/3 < p ≤ 1. If this is the case, ρ is inseparable.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
11
From the above observation, entangled states are characterized by non
vanishing negativity deﬁned as
N(ρ) ≡
1
2
(
i
λ
i
 −1). (25)
Note that negativity vanishes if and only if all the eigenvalues of ρ
pt
are
nonnegative. However there is a class of inseparable states which are not
characterized by negativity.
2.7. Partial trace and puriﬁcation
Let H = H
1
⊗H
2
be a Hilbert space of a bipartite system made of compo
nents 1 and 2 and let A be an arbitrary operator acting on H. The partial
trace of A over H
2
generates an operator acting on H
1
deﬁned as
A
1
= tr
2
A ≡
k
(I ⊗k)A(I ⊗k). (26)
We will be concerned with the partial trace of a density matrix in prac
tical applications. Let ρ = ψψ ∈ S(H) be a density matrix of a pure
state ψ. Suppose we are interested only in the ﬁrst system and have no
access to the second system. Then the partial trace allows us to “forget”
about the second system. In other words, the partial trace quantiﬁes our
ignarance on the second sytem.
To be concrete, consider a pure state
ψ =
1
√
2
(e
1
e
1
+e
2
e
2
),
where {e
i
} is an orthonormal basis of C
2
. The corresponding density ma
trix is
ρ =
1
2
1 0 0 1
0 0 0 0
0 0 0 0
1 0 0 1
,
where the basis vectors are ordered as {e
1
e
1
, e
1
e
2
, e
2
e
1
, e
2
e
2
}.
The partial trace of ρ is
ρ
1
= tr
2
ρ =
i=1,2
(I ⊗e
i
)ρ(I ⊗e
i
) =
1
2
1 0
0 1
. (27)
Note that a pure state ψ is mapped to a maximally mixed state ρ
1
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
12
We have seen above that the partial trace of a purestate density matrix
of a bipartite system over one of the constituent Hilbert spaces yields a
mixed state. How about the converse? Given a mixed state density matrix,
is it always possible to ﬁnd a pure state density matrix whose partial trace
over the extra Hilbert space yields the given density matrix? The answer
is yes and the process to ﬁnd the pure state is called the puriﬁcation. Let
ρ
1
=
k
p
k
ψ
k
ψ
k
 be a general density matrix of a system 1 with the
Hilbert space H
1
. Now let us introduce the second Hilbert space H
2
whose
dimension is the same as that of H
1
. Then formally introduce a normalized
vector
Ψ =
k
√
p
k
ψ
k
⊗ φ
k
, (28)
where {φ
k
} is an orthonormal basis of H
2
. We ﬁnd
tr
2
ΨΨ =
i,j,k
(I ⊗ φ
i
)
√
p
j
p
k
ψ
j
φ
j
ψ
k
φ
k

(I ⊗ φ
i
)
=
k
p
k
ψ
k
ψ
k
 = ρ
1
. (29)
It is always possible to purify a mixed state by tensoring an extra Hilbert
space of the same dimension as that of the original Hilbert space. Puriﬁca
tion is far from unique.
3. Qubits
A (Boolean) bit assumes two distinct values, 0 and 1, and it constitutes the
building block of the classical information theory. Quantum information
theory, on the other hand, is based on qubits.
3.1. One qubit
A qubit is a (unit) vector in the vector space C
2
, whose basis vectors are
denoted as
0 = (1, 0)
t
and 1 = (0, 1)
t
. (30)
What these vectors physically mean depends on the physical realization
employed for quantum information processing.
They might represent spin states of an electron, 0 =  ↑ and 1 =  ↓.
Electrons are replaced by nuclei with spin 1/2 in NMR (Nuclear Magnetic
Resonance).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
13
In some cases, 0 stands for a vertically polarized photon  while 1
represents a horizontally polarized photon  ↔. Alternatively they might
correspond to photons polarized in diﬀerent directions. For example, 0
may represent a polarization state 
↔
=
1
√
2
( + ↔) while 1 represents
a state 
↔
=
1
√
2
( −  ↔).
Truncated two states from many levels may be employed as a qubit. We
may assign 0 to the ground state and 1 to the ﬁrst excited state of an
atom or an ion.
In any case, we have to ﬁx a set of basis vectors when we carry out
quantum information processing. In the following, the basis is written in
an abstract form as {0, 1}, unless otherwise stated.
It is convenient to assume the vector 0 corresponds to the classical bit
0, while 1 to 1. Moreover a qubit may be in a superposition state:
ψ = a0 + b1, with a
2
+ b
2
= 1. (31)
If we measure ψ to see whether it is in 0 or 1, the outcome will be
0 (1) with the probability a
2
(b
2
) and the state immediately after the
measurement is 0 (1).
Although a qubit may take inﬁnitely many diﬀerent states, it should be
kept in mind that we can extract from it as the same amount of informa
tion as that of a classical bit. Information can be extracted only through
measurements. When we measure a qubit, the state vector ‘collapses’ to the
eigenvector that corresponds to the eigenvalue observed. Suppose a spin is
in the state a0 + b1. If we observe that the zcomponent of the spin is
+1/2, the system immediately after the measurement is in 0. This hap
pens with probability ψ00ψ = a
2
. The measurement outcome of a
qubit is always one of the eigenvalues, which we call abstractly 0 and 1.
3.2. Bloch sphere
It is useful, for many purposes, to express a state of a single qubit graphi
cally. Let us parameterize a onequbit pure state ψ with θ and φ as
ψ(θ, φ) = cos
θ
2
0 + e
iφ
sin
θ
2
1. (32)
The phase of ψ is ﬁxed in such a way that the coeﬃcient of 0 is real. It is
easy to verify that (ˆ n(θ, φ) · σ)ψ(θ, φ) = ψ(θ, φ), where σ = (σ
x
, σ
y
, σ
z
)
and ˆ n(θ, φ) is a real unit vector called the Bloch vector with components
ˆ n(θ, φ) = (sin θ cos φ, sin θ sin φ, cos θ)
t
. It is therefore natural to assign
ˆ n(θ, φ) to a state vector ψ(θ, φ) so that ψ(θ, φ) is expressed as a unit
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
14
vector ˆ n(θ, φ) on the surface of the unit sphere, called the Bloch sphere.
This correspondence is onetoone if the ranges of θ and φ are restricted to
0 ≤ θ ≤ π and 0 ≤ φ < 2π.
It is veriﬁed that state (32) satisﬁes
ψ(θ, φ)σψ(θ, φ) = ˆ n(θ, φ). (33)
A density matrix ρ of a qubit can be represented as a point on a unit
ball. Since ρ is a positive semideﬁnite Hermitian matrix with unit trace,
its most general form is
ρ =
1
2
I +
i=x,y,z
u
i
σ
i
, (34)
where u ∈ R
3
satisﬁes u ≤ 1. The reality follows from the Hermitic
ity requirement and tr ρ = 1 is obvious. The eigenvalues of ρ are λ
±
=
1
2
1 ±
u
/2 and therefore nonnegative. The eigenvalue λ
−
vanishes in
case u = 1, for which rank ρ = 1. Therefore the surface of the unit sphere
corresponds to pure states. The converse is also shown easily. In contrast,
all the points u inside a unit ball correspond to mixed states. The ball is
called the Bloch ball and the vector u is also called the Bloch vector.
It is easily veriﬁed that ρ given by Eq. (34) satisﬁes
σ = tr (ρσ) = u. (35)
3.3. Multiqubit systems and entangled states
Let us consider a group of many (n) qubits next. Such a system behaves
quite diﬀerently from a classical one and this diﬀerence gives a distinguish
ing aspect to quantum information theory. An nqubit system is often called
a (quantum) register in the context of quantum computing.
As an example, let us consider an nqubit register. Suppose we specify
the state of each qubit separately like a classical case. Each of the qubit is
then described by a 2d complex vector of the form a
i
0+b
i
1 and we need
2n complex numbers {a
i
, b
i
}
1≤i≤n
to specify the state. This corresponds the
a tensor product state (a
1
0 + b
1
1) ⊗ . . . ⊗ (a
n
0 + b
n
1) ∈ C
2n
. If the
system is treated in a fully quantummechanical way, however, a general
state vector of the register is represented as
ψ =
i
k
=0,1
a
i
1
i
2
...i
n
i
1
⊗ i
2
⊗ . . . ⊗ i
n
∈ C
2
n
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
15
Note that 2
n
2n for a large number n. The ratio 2
n
/2n is ∼ 10
298
for
n = 1000. Most quantum states in a Hilbert space with large n are entan
gled having no classical analogues. Entanglement is an extremely powerful
resource for quantum computation and quantum communication.
Let us consider a 2qubit system for deﬁniteness. The system has a
binary basis {00, 01, 10, 11}. More generally, a basis for a system of
n qubits may be {b
n−1
b
n−2
. . . b
0
}, where b
n−1
, b
n−2
, . . . , b
0
∈ {0, 1}. It is
also possible to express the basis in terms of the decimal system. We write
x, instead of b
n−1
b
n−2
. . . b
0
, where x = b
n−1
2
n−1
+b
n−2
2
n−2
+. . . +b
0
.
The basis for a 2qubit system may be written also as {0, 1, 2, 3} with
this decimal notation.
The set
{Φ
+
=
1
√
2
(00 +11), Φ
−
=
1
√
2
(00 −11),
Ψ
+
=
1
√
2
(01 +10), Ψ
−
=
1
√
2
(01 −10)}
(36)
is an orthonormal basis of a twoqubit system and is called the Bell basis.
Each vector is called the Bell state or the Bell vector. Note that all the Bell
states are entangled.
4. Quantum Gates, Quantum Circuit and Quantum
Computation
4.1. Introduction
Now that we have introduced qubits to store information, it is time to
consider operations acting on them. If they are simple, these operations
are called gates, or quantum gates, in analogy with those in classical logic
circuits. More complicated quantum circuits are composed of these simple
gates. A collection of quantum circuits for executing a complicated algo
rithm, a quantum algorithm, is a part of a quantum computation.
Deﬁnition 4.1. (Quantum Computation) A quantum computation is a
collection of the following three elements:
(1) A register or a set of registers,
(2) A unitary matrix u, which is taylored to execute a given quantum
algorithm and
(3) Measurements to extract information we need.
More formally, a quantum computation is the set {H, U, {M
m
}}, where
H = C
2
n
is the Hilbert space of an nqubit register, U ∈ U(2
n
) represents
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
16
a quantum algorithm and {M
m
} is the set of measurement operators. The
hardware (1) is called a quantum computer.
Suppose the register is set to a ﬁducial initial state, ψ
in
= 00 . . . 0
for example. A unitary matrix U
alg
is generated by an algorithm which
we want to execute. Operation of U
alg
on ψ
in
yields the output state
ψ
out
= U
alg
ψ
in
. Information is extracted from ψ
out
by appropriate
measurements.
4.2. Quantum gates
We have so far studied the change of a state upon measurements. When
measurements are not made, the time evolution of a state is described
by the Schr¨ odinger equation. The time evolution operator U is unitary:
UU
†
= U
†
U = I. We will be free from the Schr¨odinger equation in the
following and assume there always exist unitary matrices which we need.
One of the important conclusions derived from the unitarity of gates is
that the computational process is reversible.
4.2.1. Simple quantum gates
Examples of quantum gates which transform a onequbit state are given
below. We call them onequbit gates in the following. Linearity guarantees
that the action of a gate is completely speciﬁed if its action on the basis
{0, 1} is given. Consider the gate I whose action on the basis vectors is
I : 0 → 0, 1 → 1. The matrix expression of this gate is
I = 00 + 11 =
1 0
0 1
. (37)
Similarly we introduce X : 0 → 1, 1 → 0, Y : 0 → −1, 1 → 0
and Z : 0 → 0, 1 → −1 by
X = 10 + 01 =
0 1
1 0
= σ
x
, (38)
Y = 01 − 10 =
0 −1
1 0
= −iσ
y
, (39)
Z = 00 − 11 =
1 0
0 −1
= σ
z
. (40)
The transformation I is the identity transformation, while X is the negation
(NOT), Z the phase shift and Y = XZ the combination thereof.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
17
CNOT (controlledNOT) gate is a 2qubit gate, which plays an impor
tant role. The gate ﬂips the second qubit (the target qubit) when the ﬁrst
qubit (the control qubit) is 1, while leaving the second bit unchanged
when the ﬁrst bit is 0. Let {00, 01, 10, 11} be a basis for the 2qubit
system. We use the standard basis vectors with components
00 = (1, 0, 0, 0)
t
, 01 = (0, 1, 0, 0)
t
, 10 = (0, 0, 1, 0)
t
, 11 = (0, 0, 0, 1)
t
.
The action of CNOT gate, whose matrix expression will be written as
U
CNOT
, is U
CNOT
: 00 → 00, 01 → 01, 10 → 11, 11 → 10. It
has two equivalent expressions
U
CNOT
= 0000 + 0101 + 1110 + 1011
= 00 ⊗ I + 11 ⊗ X, (41)
having a matrix form
U
CNOT
=
1 0 0 0
0 1 0 0
0 0 0 1
0 0 1 0
. (42)
Let {i} be the basis vectors, where i ∈ {0, 1}. The action of CNOT on
the input state i, j is written as i, i ⊕j, where i ⊕j is an addition mod2.
A 1qubit gate whose unitary matrix is U is graphically depicted as
The left horizontal line is the input qubit while the right horizontal line is
the output qubit: time ﬂows from the left to the right.
A CNOT gate is expressed as
where • denotes the control bit, while
denotes the conditional negation.
There may be many control bits (see CCNOT gate below). More generally,
we consider a controlledU gate, V = 00 ⊗ I + 11 ⊗ U, in which the
target bit is acted on by a unitary transformation U only when the control
bit is 1. This gate is denoted graphically as
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
18
CCNOT (ControlledControlledNOT) gate has three inputs and the
third qubit ﬂips only when the ﬁrst two qubits are both in the state 1.
The explicit form of the CCNOT gate is
U
CCNOT
= (0000 +0101 +1010) ⊗I +1111 ⊗X. (43)
This gate is graphically expressed as
4.2.2. WalshHadamard transformation
The Hadamard gate or the Hadamard transformation H is an important
unitary transformation deﬁned by
U
H
: 0 →
1
√
2
(0 +1)
: 1 →
1
√
2
(0 −1).
(44)
The matrix representation of H is
U
H
=
1
√
2
(0 +1)0 +
1
√
2
(0 −1)1 =
1
√
2
1 1
1 −1
. (45)
A Hadamard gate is depicted as
There are numerous important applications of the Hadamard transfor
mation. All possible 2
n
states are generated when U
H
is applied on each
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
19
qubit of the state 00 . . . 0:
(U
H
⊗U
H
⊗. . . ⊗U
H
)00 . . . 0
=
1
√
2
(0 +1) ⊗
1
√
2
(0 +1) ⊗. . .
1
√
2
(0 +1)
=
1
√
2
n
2
n
−1
x=0
x. (46)
Therefore, we produce a superposition of all the states x with 0 ≤ x ≤
2
n
− 1 simultaneously. The transformation U
⊗n
H
is called the Walsh trans
formation, or WalshHadamard transformation and denoted as W
n
.
The quantum circuit
is used to generate Bell states from inputs 00, 01, 10 and 11.
4.2.3. SWAP gate and Fredkin gate
The SWAP gate acts on a tensor product state as
U
SWAP
ψ
1
, ψ
2
= ψ
2
, ψ
1
. (47)
The explict form of U
SWAP
is given by
U
SWAP
= 0000 +0110 +1001 +1111 =
1 0 0 0
0 0 1 0
0 1 0 0
0 0 0 1
. (48)
The SWAP gate is expressed as
Note that the SWAP gate is a special gate which maps an arbitrary tensor
product state to a tensor product state. In contrast, most 2qubit gates
map a tensor product state to an entangled state.
The controlledSWAP gate
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
20
is also called the Fredkin gate. It ﬂips the second (middle) and the third
(bottom) qubits only when the ﬁrst (top) qubit is in the state 1. Its explicit
form is U
Fredkin
= 00 ⊗I
4
+ 11 ⊗U
SWAP
.
4.3. Nocloning theorem
Theorem 4.1. (Wootters and Zurek
14
) An unknown quantum system can
not be cloned by unitary transformations.
Proof: Suppose there would exist a unitary transformation U that makes
a clone of a quantum system. Namely, suppose U acts, for any state ϕ, as
U : ϕ0 → ϕϕ. Let ϕ and φ be two states that are linearly independent.
Then we should have Uϕ0 = ϕϕ and Uφ0 = φφ by deﬁnition. Then
the action of U on ψ =
1
√
2
(ϕ + φ) yields
Uψ0 =
1
√
2
(Uϕ0 +Uφ0) =
1
√
2
(ϕϕ + φφ).
If U were a cloning transformation, we must also have
Uψ0 = ψψ =
1
2
(ϕϕ + ϕφ + φϕ + φφ),
which contradicts the previous result. Therefore, there does not exist a uni
tary cloning transformation.
Note however that the theorem does not apply if the states to be cloned
are limited to 0 and 1. For these cases, the copying operator U should
work as U : 00 → 00, : 10 → 11. We can assign arbitrary action of
U on a state whose second input is 1 since this case will never happen.
What we have to keep in mind is only that U be unitary. An example of
such U is
U = (0000 + 1110) + (0101 + 1011). (49)
where the ﬁrst set of operators renders U the cloning operator and the
second set is added just to make U unitary. We immediately notice that U
is nothing but the CNOT gate.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
21
Therefore, if the data under consideration is limited within 0 and 1,
we can copy the qubit states even in a quantum computer. This fact is used
to construct quantum error correcting codes.
4.4. Quantum teleportation
The purpose of quantum teleportation is to transmit an unknown quantum
state of a qubit using two classical bits in such a way that the recipient
reproduces the same state as the original qubit state. Note that the qubit
itself is not transported but the information required to reproduce the quan
tum state is transmitted. The original state is destroyed such that quantum
teleportation is not in contradiction with the nocloning theorem.
Fig. 1. In quantum teleportation, Alice sends Bob two classical bits so that Bob repro
duces a qubit state Alice initially had.
Alice: Alice has a qubit, whose state she does not know. She wishes to
send Bob the quantum state of this qubit through a classical communication
channel. Let φ = a0 + b1 be the state of the qubit. Both of them have
been given one of the qubits of the entangled pair
Φ
+
=
1
√
2
(00 +11)
in advance. They start with the state
φ ⊗Φ
+
=
1
√
2
(a000 + a011 + b100 + b111) , (50)
where Alice possesses the ﬁrst two qubits while Bob has the third. Alice
applies U
CNOT
⊗I followed by U
H
⊗I ⊗I to this state, which results in
(U
H
⊗I ⊗I)(U
CNOT
⊗I)(φ ⊗Φ
+
)
=
1
2
[00(a0 + b1) +01(a1 + b0)
+10(a0 −b1) +11(a1 −b0)]. (51)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
22
If Alice measures the 2 qubits in her hand, she will obtain one of the states
00, 01, 10 or 11 with equal probability 1/4. Bob’s qubit (one of the
EPR pair previously) collapses to a0+b1, a1+b0, a0−b1 or a1−
b0, respectively, depending on the result of Alice’s measurement. Alice
then sends Bob her result of the measurement using two classical bits.
Bob: After receiving two classical bits, Bob knows the state of the qubit
in his hand;
received bits Bob’s state decoding
00 a0 + b1 I
01 a1 + b0 X
10 a0 − b1 Z
11 a1 − b0 Y
(52)
Bob reconstructs the intial state φ by applying the decoding process shown
above. Suppose Alice sends Bob the classical bits 10, for example. Then Bob
applies Z on his qubit to reconstruct φ as Z : (a0−b1) → (a0+b1) =
φ.
Figure 2 shows the actual quantum circuit for quantum teleportation.
Fig. 2. Quantum circuit implementation of quantum teleportation.
4.5. Universal quantum gates
It can be shown that any classical logic gate can be constructed by using a
small set of gates, AND, NOT and XOR for example. Such a set of gates
is called the universal set of gates. It can be shown that the CCNOT gate
simulates these classical gates, and hence quantum circuits simulate any
classical circuits. The set of quantum gates is, however, much larger than
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
23
those classical gates. Thus we want to ﬁnd a universal set of quantum gates
from which any quantum circuits can be constructed.
It can be shown that
(1) the set of single qubit gates and
(2) CNOT gate
form a universal set of quantum circuits (universality theorem). The proof
is highly technical and is not given here.
1,2,16
We, instead, sketch the proof
in several lines.
It can be shown that any U ∈ U(n) is written as a product of N twolevel
unitary matrices, where N ≤ n(n − 1)/2 and a twolevel unitary matrix is
a unit matrix I
n
in which only four components V
aa
, V
ab
, V
ba
and V
bb
are
diﬀerent from I
n
. Moreover V = (V
ij
) is an element of U(2). An example
of a twolevel unitary matrix is
V =
α
∗
0 0 β
∗
0 1 0 0
0 0 1 0
−β 0 0 α
, (α
2
+β
2
= 1)
where a = 1 and b = 4.
Now we need to prove the universality theorem for twolevel unitary
matrices, which is certainly simpler than the general proof. By employing
CNOT gates and their generalizations, it is possible to move the elements
V
aa
, V
ab
, V
ba
and V
bb
so that they acts on a single qubit in the register. We
need to implement the controlledV gate whose target qubit is the one on
which V acts. Implementation of the controlledV gate requires generalized
CNOT gates and several U(2) gates.
1,2,16
4.6. Quantum parallelism and entanglement
Given an input x, a typical quantum computer “computes” f(x) as
U
f
: x0 → xf(x), (53)
where U
f
is a unitary matrix which implements the function f.
Suppose U
f
acts on an input which is a superposition of many x.
Since U
f
is a linear operator, it acts on all the constituent vectors of the
superposition simultaneously. The output is also a superposition of all the
results;
U
f
:
x
x0 →
x
xf(x). (54)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
24
This feature, called the quantum parallelism, gives quantum computer an
enormous power. A quantum computer is advantageous over a classical
counterpart in that it makes use of this quantum parallelism and also en
tanglement.
A unitary transformation acts on a superposition of all possible states
in most quantum algorithms. This superposition is prepared by the action
of the WalshHadamard transformation on an nqubit register in the initial
state 00 . . . 0 = 0 ⊗0 ⊗. . . ⊗0 resulting in
2
n
−1
x=0
x/
√
2
n
. This state
is a superposition of vectors encoding all the integers between 0 and 2
n
−1.
Then the linearlity of U
f
leads to
U
f
1
√
2
n
2
n
−1
x=0
x0
=
1
√
2
n
2
n
−1
x=0
U
f
x0 =
1
√
2
n
2
n
−1
x=0
xf(x). (55)
Note that the superposition is made of 2
n
= e
nln2
states, which makes
quantum computation exponentially faster than classical counterpart in a
certain kind of computation.
What about the limitation of a quantum computer
3
? Let us consider
the CCNOT gate for example. This gate ﬂips the third qubit if and only
if the ﬁrst and the second qubits are both in the state 1 while it leaves
the third qubit unchanged otherwise. Let us ﬁx the third input qubit to
0. The third output qubit state is x ∧ y, where x and y are the ﬁrst
and the second input qubits respectively. Suppose the input state of the
ﬁrst and the second qubits is a superposition of all possible states while the
third qubit is ﬁxed to 0. This can be achieved by the WalshHadamard
transformation as
U
H
0 ⊗U
H
0 ⊗0 =
1
√
2
(0 +1) ⊗
1
√
2
(0 +1) ⊗0
=
1
2
(000 +010 +100 +110). (56)
By operating CCNOT on this state, we obtain
U
CCNOT
(U
H
0 ⊗U
H
0 ⊗0) =
1
2
(000 +010 +100 +111). (57)
This output may be thought of as the truth table of AND: x, y, x∧y. It is
extremely important to note that the output is an entangled state and the
measurement projects the state to one line of the truth table, i.e., a single
term in the RHS of Eq. (57).
There is no advantage of quantum computation over classical one at this
stage. This is because only one result may be obtained by a single set of
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
25
measurements. What is worse, we cannot choose a speciﬁc vector x, y, x∧y
at our will! Thus any quantum algorithm should be programmed so that
the particular vector we want to observe should have larger probability to
be measured compared to other vectors. The programming strategies to
deal with this feature are
(1) to amplify the amplitude, and hence the probability, of the vector that
we want to observe. This strategy is employed in the Grover’s database
search algorithm.
(2) to ﬁnd a common property of all the f(x). This idea was employed in
the quantum Fourier transform to ﬁnd the order
a
of f in the Shor’s
factoring algorithm.
Now we consider the power of entanglement. Suppose we have an n
qubit register, whose Hilbert space is 2
n
dimensional. Since each qubit has
two basis states {0, 1}, there are 2n basis states, i.e., n 0’s and n 1’s,
involved to span this Hilbert space. Imagine that we have a single quantum
system, instead, which has the same Hilbert space. One might think that
the system may do the same quantum computation as the nqubit register
does. One possible problem is that one cannot “measure the kth digit”
leaving other digits unaﬀected. Even worse, consider how many diﬀerent
basis vectors are required for this system. This single system must have
an enormous number, 2
n
, of basis vectors! Multipartite implementation
of a quantum algorithm requires exponentially smaller number of basis
vectors than monopartite implementation since the former makes use of
entanglement as a computational resource.
5. Simple Quantum Algorithms
Let us introduce a few simple quantum algorithms which will be of help
to understand how quantum algorithms are diﬀerent from and superior to
classical algorithms.
5.1. Deutsch algorithm
The Deutsch algorithm is one of the ﬁrst quantum algorithms which showed
quantum algorithms may be more eﬃcient than their classical counterparts.
a
Let m, N ∈ N (m < N) be numbers coprime to each other. Then there exists P ∈ N
such that m
P
≡ 1 (mod N). The smallest such number P is called the period or the
order. It is easily seen that m
x+P
≡ m
x
(mod N), ∀x ∈ N.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
26
In spite of its simplicty, full usage of superposition principle and entangle
ment has been made here.
Let f : {0, 1} → {0, 1} be a binary function. Note that there are only
four possible f, namely
f
1
: 0 → 0, 1 → 0, f
2
: 0 → 1, 1 → 1,
f
3
: 0 → 0, 1 → 1, f
4
: 0 → 1, 1 → 0.
First two cases, f
1
and f
2
, are called constant,while the rest, f
3
and f
4
, are
balanced.If we only have classical resources, we need to evaluate f twice to
tell if f is constant or balanced. There is a quantum algorithm, in contrast,
with which it is possible to tell if f is constant or balanced with a single
evaluation of f, as was shown by Deutsch.
18
Let 0 and 1 correspond to classical bits 0 and 1, respectively, and
consider the state ψ
0
=
1
2
(00 − 01 + 10 − 11). We apply f on this
state in terms of the unitary operator U
f
: x, y → x, y ⊕ f(x), where ⊕
is an addition mod 2. To be explicit, we obtain
ψ
1
= U
f
ψ
0
=
1
2
(0, f(0) −0, ¬f(0) +1, f(1) −1, ¬f(1)),
where ¬ stands for negation. Therefore this operation is nothing but the
CNOT gate with the control bit f(x); the target bit y is ﬂipped if and
only if f(x) = 1 and left unchanged otherwise. Subsequently we apply the
Hadamard gate on the ﬁrst qubit to obtain
ψ
2
= U
H
ψ
1
=
1
2
√
2
[(0 +1)(f(0) −¬f(0)) + (0 −1)(f(1) −¬f(1))]
The wave function reduces to
ψ
2
=
1
√
2
0(f(0) −¬f(0)) (58)
in case f is constant, for which f(0) = f(1), and
ψ
2
=
1
√
2
1(f(0) −f(1)) (59)
if f is balanced, for which ¬f(0) = f(1). Therefore the measurement of
the ﬁrst qubit tells us whether f is constant or balanced.
Let us consider a quantum circuit which implements the Deutsch algo
rithm. We ﬁrst apply the WalshHadamard transformation W
2
= U
H
⊗U
H
on 01 to obtain ψ
0
. We need to introduce a conditional gate U
f
,
i.e., the controlledNOT gate with the control bit f(x), whose action is
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
27
U
f
: x, y → x, y ⊕f(x). Then the Hadamard gate is applied on the ﬁrst
qubit before it is measured. Figure 3 depicts this implementation.
Fig. 3. Implementation of the Deutsch algorithm.
In the quauntum circuit, we assume the gate U
f
is a black box for which
we do not ask the explicit implementation. We might think it is a kind of
subroutine. Such a black box is often called an oracle. The gate U
f
is called
the Deutsch oracle. Its implementation is given only after f is speciﬁed.
Then what is the merit of the Deutsch algorithm? Suppose your friend
gives you a unitary matrix U
f
and asks you to tell if f is constant or
balanced. Instead of applying 0 and 1 separately, you may contruct the
circuit in Fig. 3 with the given matrix U
f
and apply the circuit on the input
state 01. Then you can tell your friend whether f is constant or balanced
with a single use of U
f
.
5.2. DeutschJozsa algorithm
The Deutsch algorithm introduced in the previous section may be general
ized to the DeutschJozsa algorithm.
19
Let us ﬁrst deﬁne the DeutschJozsa
problem. Suppose there is a binary function
f : S
n
≡ {0, 1, . . . , 2
n
−1} → {0, 1}. (60)
We require f be either constantor balancedas before. When f is constant,
it takes a constant value 0 or 1 irrespetive of the input value x. When it is
balanaced the value f(x) for a half of x ∈ S
n
is 0 while it is 1 for the rest of
x. Although there are functions which are neither constant nor balanced, we
will not consider such cases here. Our task is to ﬁnd an algorithm which tells
if f is constant or balanced with the least possible number of evaluations
of f.
It is clear that we need at least 2
n−1
+ 1 steps, in the worst case with
classical manipulations, to make sure if f(x) is constant or balanced with
100 % conﬁdence. It will be shown below that the number of steps reduces
to a single step if we are allowed to use a quantum algorithm.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
28
The algorithm is divided into the following steps:
(1) Prepare an (n + 1)qubit register in the state ψ
0
= 0
⊗n
⊗1. First
n qubits work as input qubits while the (n + 1)st qubit serves as a
“scratch pad”. Such qubits, which are neither input qubits nor output
qubits, but work as a scratch pad to store temporary information are
called ancillas or ancillary qubits.
(2) Apply the WalshHadamard transforamtion to the register. Then we
have the state
ψ
1
= U
⊗n+1
H
ψ
0
=
1
√
2
n
(0 +1)
⊗n
⊗
1
√
2
(0 −1)
=
1
√
2
n
2
n
−1
x=0
x ⊗
1
√
2
(0 −1). (61)
(3) Apply the f(x)controlledNOT gate on the register, which ﬂips the
(n+1)st qubit if and only if f(x) = 1 for the input x. Therefore we need
a U
f
gate which evaluates f(x) and acts on the register as U
f
xc =
xc ⊕ f(x), where c is the onequbit state of the (n + 1)st qubit.
Observe that c is ﬂipped if and only if f(x) = 1 and left unchanged
otherwise. We then obtain a state
ψ
2
= U
f
ψ
1
=
1
√
2
n
2
n
−1
x=0
x
1
√
2
(f(x) −¬f(x))
=
1
√
2
n
x
(−1)
f(x)
x
1
√
2
(0 −1). (62)
Although the gate U
f
is applied once for all, it is applied to all the
nqubit states x simultaneously.
(4) The WalshHadamard transformation (46) is applied on the ﬁrst n
qubits next. We obtain
ψ
3
= (W
n
⊗I)ψ
2
=
1
√
2
n
2
n
−1
x=0
(−1)
f(x)
U
⊗n
H
x
1
√
2
(0 −1). (63)
It is instructive to write the action of the onequbit Hadamard gate as
U
H
x =
1
√
2
(0 + (−1)
x
1) =
1
√
2
y∈{0,1}
(−1)
xy
y,
where x ∈ {0, 1}, to ﬁnd the resulting state. The action of the Walsh
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
29
Hadamard transformation on x = x
n−1
. . . x
1
x
0
yields
W
n
x = (U
H
x
n−1
)(U
H
x
n−2
) . . . (U
H
x
0
)
=
1
√
2
n
y
n−1
,y
n−2
,...,y
0
∈{0,1}
(−1)
x
n−1
y
n−1
+x
n−2
y
n−2
+...+x
0
y
0
×y
n−1
y
n−2
. . . y
0
=
1
√
2
n
2
n
−1
y=0
(−1)
x·y
y, (64)
where x· y = x
n−1
y
n−1
⊕x
n−2
y
n−2
⊕. . . ⊕x
0
y
0
. Substituting this result
into Eq. (63), we obtain
ψ
3
=
1
2
n
2
n
−1
x,y=0
(−1)
f(x)
(−1)
x·y
y
1
√
2
(0 −1). (65)
(5) The ﬁrst n qubits are measured. Suppose f(x) is constant. Then ψ
3
is put in the form
ψ
3
=
1
2
n
x,y
(−1)
x·y
y
1
√
2
(0 −1)
up to an overall phase. Let us consider the summation
1
2
n
2
n
−1
x=0
(−1)
x·y
for a ﬁxed y ∈ S
n
. Clearly it vanishes since x · y is 0 for half of x and
1 for the other half of x unless y = 0. Therefore the summation yields
δ
y0
. Now the state reduces to ψ
3
= 0
⊗n 1
√
2
(0 − 1) and the mea
surement outcome of the ﬁrst n qubits is always 00 . . . 0. Suppose f(x)
is balanced next. The probability amplitude of y = 0 in ψ
3
is pro
portional to
2
n
−1
x=0
(−1)
f(x)
(−1)
x·y
=
2
n
−1
x=0
(−1)
f(x)
= 0. Therefore
the probability of obtaining measurement outcome 00 . . . 0 for the ﬁrst
n qubits vanishes. In conclusion, the function f is constant if we obtain
00 . . . 0 upon the meaurement of the ﬁrst n qubits in the state ψ
3
and
it is balanced otherwise.
6. Decoherence
A quantum system is always in interaction with its environment. This in
teraction inevitably alter the state of the quantum system, which causes
loss of information encoded in this system. The system under consideration
is not a closed system when interaction with outside world is in action.
We formulate the theory of open quantum system in this chapter by re
garding the combined system of the quantum system and its environment
as a closed system and subsequently trace out the environment degrees of
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
30
freedom. Let ρ
S
and ρ
E
be the initial density matrices of the system and
the environment, respectively. Even when the initial state is an uncorre
lated state ρ
S
⊗ρ
E
, the systemenvironment interaction entangles the total
system so that the total state develops to an inseparable entangled state
in general. Decoherence is a process in which environment causes various
changes in the quantum system, which manifests itself as undesirable noise.
6.1. Open quantum system
Let us start our exposition with some mathematical background materi
als.
1,2,24
We deal with general quantum states described by density matrices. We
are interested in a general evolution of a quantum system, which is described
by a powerful tool called a quantum operation. One of the simplest quantum
operations is a unitary time evolution of a closed system. Let ρ
S
be a density
matrix of a closed system at t = 0 and let U(t) be the time evolution
operator. Then the corresponding quantum map E is deﬁned as
E(ρ
S
) = U(t)ρ
S
U(t)
†
. (66)
One of our primary aims in this section is to generalize this map to cases
of open quantum systems.
6.1.1. Quantum operations and Kraus operators
Suppose a system of interest is coupled with its environment. We must
specify the details of the environment and the coupling between the system
and the environment to study the eﬀect of the environment on the behav
ior of the system. Let H
S
, H
E
and H
SE
be the system Hamiltonian, the
environment Hamiltonian and their interaction Hamiltonian, respectively.
We assume the systemenvironment interaction is weak enough so that this
separation into the system and its environment makes sense. To avoid con
fusion, we often call the system of interest the principal system. The total
Hamiltonian H
T
is then
H
T
= H
S
+ H
E
+ H
SE
. (67)
Correspondingly, we denote the system Hilbert space and the environment
Hilbert space as H
S
and H
E
, respectively, and the total Hilbert space as
H
T
= H
S
⊗ H
E
. The condition of weak systemenvironment interaction
may be lifted in some cases. Let us consider a qubit propagating through
a noisy quantum channel, for example. “Propagating” does not necessarily
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
31
mean propagating in space. The qubit may be spatially ﬁxed and subject
to timedependent noise. When the noise is localized in space and time, the
input and the output qubit states belong to a well deﬁned Hilbert space H
S
and the above separation of the Hamiltonian is perfectly acceptable even for
strongly interacting cases. We consider, in the following, how the principal
system state ρ
S
at t = 0 evolves in time in the presence of its environment.
A map which describes a general change of the state from ρ
S
to E(ρ
S
) is
called a quantum operation. We have already noted that the unitary time
evolution is an example of a quantum operation. Other quantum operations
include state change associated with measurement and state change due to
noise. The latter quantum map is our primary interest in this chapter.
The state of the total system is described by a density matrix ρ. Suppose
ρ is uncorrelated initally at time t = 0,
ρ(0) = ρ
S
⊗ ρ
E
, (68)
where ρ
S
(ρ
E
) is the initial density matrix of the principal system (envi
ronment). The total system is assumed to be closed and to evolve with a
unitary matrix U(t) as
ρ(t) = U(t)(ρ
S
⊗ ρ
E
)U(t)
†
. (69)
Note that the resulting state is not a tensor product state in general. We
are interested in extracting information on the state of the principal system
at some later time t > 0.
Even under these circumstances, however, we may still deﬁne the system
density matrix ρ
S
(t) by taking partial trace of ρ(t) over the environment
Hilbert space as
ρ
S
(t) = tr
E
[U(t)(ρ
S
⊗ ρ
E
)U(t)
†
]. (70)
We may forget about the environment by taking a trace over H
E
. This is
an example of a quantum operation, E(ρ
S
) = ρ
S
(t). Let {e
j
} be a basis of
the system Hilbert space while {ε
a
} be that of the environment Hilbert
space. We may take the basis of H
T
to be {e
j
⊗ε
a
}. The initial density
matrices may be written as ρ
S
=
j
p
j
e
j
e
j
, ρ
E
=
a
r
a
ε
a
ε
a
.
Action of the time evolution operator on a basis vector of H
T
is explicitly
written as
U(t)e
j
, ε
a
=
k,b
U
kb;ja
e
k
, ε
b
, (71)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
32
where e
j
, ε
a
= e
j
⊗ ε
a
for example. Using this expression, the density
matrix ρ(t) is written as
U(t)(ρ
S
⊗ ρ
E
)U(t)
†
=
j,a
p
j
r
a
U(t)e
j
, ε
a
e
j
, ε
a
U(t)
†
=
j,a,k,b,l,c
p
j
r
a
U
kb;ja
e
k
, ε
b
e
l
, ε
c
U
∗
lc;ja
. (72)
The partial trace over H
E
is carried out to yield
ρ
S
(t) = tr
E
[U(t)(ρ
S
⊗ ρ
E
)U(t)
†
] =
j,a,k,b,l
p
j
r
a
U
kb;ja
e
k
e
l
U
∗
lb;ja
=
j,a,b
p
j
k
√
r
a
U
kb;ja
e
k
l
√
r
a
e
l
U
∗
lb;ja
. (73)
To write down the quantum operation in a closed form, we assume the
initial environment state is a pure state, which we take, without loss of
generality, ρ
E
= ε
0
ε
0
. Even when ρ
E
is a mixed state, we may always
complement H
E
with a ﬁctitious Hilbert space to “purify” ρ
E
, see § 2.7.
With this assumption, ρ
S
(t) is written as
ρ
S
(t) = tr
E
[U(t)(ρ
S
⊗ ε
0
ε
0
)U(t)
†
]
=
a
(I ⊗ ε
a
)U(t)(ρ
S
⊗ ε
0
ε
0
)U(t)
†
(I ⊗ ε
a
)
=
a
(I ⊗ ε
a
)U(t)(I ⊗ ε
0
)ρ
S
(I ⊗ ε
0
)U(t)
†
(I ⊗ ε
a
).
We will drop I⊗ from I ⊗ε
a
 hereafter, whenever it does not cause confu
sion. Let us deﬁne the Kraus operator E
a
(t) : H
S
→ H
S
by
E
a
(t) = ε
a
U(t)ε
0
. (74)
Then we may write
E(ρ
S
) = ρ
S
(t) =
a
E
a
(t)ρ
S
E
a
(t)
†
. (75)
This is called the operatorsum representation (OSR) of a quantum opera
tion E. Note that {E
a
} satisﬁes the completeness relation
a
E
a
(t)
†
E
a
(t)
kl
=
a
ε
0
U(t)
†
ε
a
ε
a
U(t)ε
0
kl
= δ
kl
, (76)
where I is the unit matrix in H
S
. This is equivalent with the trace
preserving property of E as 1 = tr
S
ρ
S
(t) = tr
S
(E(ρ
S
)) = tr
S
a
E
†
a
E
a
ρ
S
for any ρ
S
∈ S(H
S
). Completeness relation and tracepreserving property
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
33
are satisﬁed since our total system is a closed system. A general quantum
map does not necessarily satisfy these properties.
25
At this stage, it turns out to be useful to relax the condition that U(t) be
a time evolution operator. Instead, we assume U be any operator including
an arbitrary unitary gate. Let us consider a twoqubit system on which the
CNOT gate acts. Suppose the principal system is the control qubit while
the environment is the target qubit. Then we ﬁnd
E
0
= (I ⊗ 0)U
CNOT
(I ⊗ 0) = P
0
, E
1
= (I ⊗ 1)U
CNOT
(I ⊗ 0) = P
1
,
where P
i
= ii, and consequently
E(ρ
S
) = P
0
ρ
S
P
0
+ P
1
ρ
S
P
1
= ρ
00
P
0
+ ρ
11
P
1
=
ρ
00
0
0 ρ
11
, (77)
where ρ
S
=
ρ
00
ρ
01
ρ
10
ρ
11
. Unitarity condition may be relaxed when mea
surements are included as quantum operations, for example.
Tracing out the extra degrees of freedom makes it impossible to invert
a quantum operation. Given an initial principal system state ρ
S
, there are
inﬁnitely many U that yield the same E(ρ
S
). Therefore even though it is
possible to compose two quantum operations, the set of quantum operations
is not a group but merely a semigroup.
b
6.1.2. Operatorsum representation and noisy quantum channel
Operatorsum representation (OSR) introduced in the previous subsection
seems to be rather abstract. Here we give an interpretation of OSR as a
noisy quantum channel. Suppose we have a set of unitary matrices {U
a
}
and a set of nonnegative real numbers {p
a
} such that
a
p
a
= 1. By
choosing U
a
randomly with probability p
a
and applying it to ρ
S
, we deﬁne
the expectation value of the resulting density matrix as
M(ρ
S
) =
a
p
a
U
a
ρ
S
U
†
a
, (78)
which we call a mixing process.
26
This occurs when a ﬂying qubit is sent
through a noisy quantum channel which transforms the density matrix by
U
a
with probability p
a
, for example. Note that no enviroment has been
introduced in the above deﬁnition, and hence no partial trace is involved.
b
A set S is called a semigroup if S is closed under a product satisfying associativity
(ab)c = a(bc). If S has a unit element e, such that ea = ae = a, ∀a ∈ S, it is called a
monoid.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
34
Now the correspondence between E(ρ
S
) and M(ρ
S
) should be clear. Let
us deﬁne E
a
≡
√
p
a
U
a
. Then Eq. (78) is rewritten as
M(ρ
S
) =
a
E
a
ρ
S
E
†
a
(79)
and the equivalence has been shown. Operators E
a
are identiﬁed with the
Kraus operators. The system transforms, under the action of U
a
, as
ρ
S
→ E
a
ρ
S
E
†
a
/tr
E
a
ρ
S
E
†
a
. (80)
Conversely, given a noisy quantum channel {U
a
, p
a
} we may intro
duce an “environment” with the Hilbert space H
E
as follows. Let H
E
=
Span(ε
a
) be a Hilbert space with the dimension equal to the number of
the unitary matrices {U
a
}, where {ε
a
} is an orthonormal basis. Deﬁne
formally the environment density matrix ρ
E
=
a
p
a
ε
a
ε
a
 and
U ≡
a
U
a
⊗ε
a
ε
a
 (81)
which acts on H
S
⊗ H
E
. It is easily veriﬁed from the orthonormality of
{ε
a
} that U is indeed a unitary matrix. Partial trace over H
E
then yields
E(ρ
S
) = tr
E
[U(ρ
S
⊗ρ
E
)U
†
]
=
a
(I ⊗ε
a
)
b
U
b
⊗ε
b
ε
b

ρ
S
⊗
c
p
c
ε
c
ε
c

×
d
U
d
⊗ε
d
ε
d

(I ⊗ε
a
)
=
a
p
a
U
a
ρ
S
U
†
a
= M(ρ
S
) (82)
showing that the mixing process is also decribed by a quantum operation
with a ﬁctitious environment.
6.1.3. Completely positive maps
All linear operators we have encountered so far map vectors to vectors.
A quantum operation maps a density matrix to another density matrix
linearly.
c
A linear operator of this kind is called a superoperator. Let Λ
be a superoperator acting on the system density matrices, Λ : S(H
S
) →
c
Of course, the space of density matrices is not a linear vector space. What is meant
hear is a linear operator, acting on the vector space of Hermitian matrices, also acts on
the space of density matrices and it maps a density matrix to another density matrix.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
35
S(H
S
). The operator Λ is easily extended to an operator acting on H
T
by
Λ
T
= Λ ⊗ I
E
, which acts on S(H
S
⊗ H
E
). Note, however, that Λ
T
is not
necessarily a map S(H
T
) → S(H
T
). It may happen that Λ
T
(ρ) is not a
density matrix any more. We have already encountered this situation when
we have introduced partial transpose operation in § 2.7. Let H
T
= H
1
⊗H
2
be a twoqubit Hilbert space, where H
k
is the kth qubit Hilbert space.
It is clear that the transpose operation Λ
t
: ρ
1
→ ρ
t
1
on a singlequbit
state ρ
1
preserves the density matrix properties. For a twoqubit density
matrix ρ
12
, however, this is not always the case. In fact, we have seen that
Λ
t
⊗ I : ρ
12
→ ρ
pt
12
deﬁned by Eq. (22) maps a density matrix to a matrix
which is not a density matrix when ρ
12
is inseparable.
A map Λ which maps a positive operator acting on H
S
to another
positive operator on H
S
is said to be positive. Moreover, it is called a
completely positive map (CP map), if its extension Λ
T
= Λ⊗I
n
remains a
positive operator for an arbitrary n ∈ N.
Theorem 6.1. A linear map Λ is CP if and only if there exists a set of
operators {E
a
} such that Λ(ρ
S
) can be written as
Λ(ρ
S
) =
a
E
a
ρ
S
E
†
a
. (83)
We require not only that Λ be CP but also Λ(ρ) be a density matrix:
tr Λ(ρ
S
) = tr
a
E
a
ρE
†
a
= tr
a
E
†
a
E
a
ρ
= 1. (84)
This condition is satisﬁed for any ρ if and only if
a
E
†
a
E
a
= I
S
. (85)
Therefore, any quantum operation obtained by tracing out the environment
degrees of freedom is CP and preserves trace.
6.2. Measurements as quantum operations
We have already seen that a unitary evoluation ρ
S
→ Uρ
S
U
†
and a mixing
process ρ
S
→
i
p
i
U
i
ρ
S
U
†
i
are quantum operations. We will see further
examples of quantum operations in this section and the next. This section
deals with measurements as quantum operations.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
36
6.2.1. Projective measurements
Suppose we measure an observable A =
i
λ
i
P
i
, where P
i
= λ
i
λ
i
 is the
projection operator corresponding to the eigenvector λ
i
. We have seen in
Chapter 2 that the probability of observing λ
i
upon a measurement of A
in a state ρ is
p(i) = λ
i
ρλ
i
= tr (P
i
ρ) (86)
and the state changes as ρ → P
i
ρP
i
/p(i). This process happens with a prob
ability p(i). Thus we may regard the measurement process as a quantum
operation
ρ
S
→
i
p(i)
P
i
ρ
S
P
i
p(i)
=
i
P
i
ρ
S
P
i
, (87)
where the set {P
i
} satisifes the completeness relation
i
P
i
P
†
i
= I.
The projective measurement is a special case of a quantum operation in
which the Kraus operators are E
i
= P
i
.
6.2.2. POVM
We have been concerned with projective measurements so far. However, it
should be noted that they are not unique type of measurements. Here we
will deal with the most general framework of measurement and show that
it is a quantum operation.
Suppose a system and an environment, prepared initially in a product
state ψe
0
, are acted by a unitary operator U, which applies an operator
M
i
on the system and, at the same time, put the environment to e
i
for
various i. It is written explicitly as
Ψ = Uψe
0
=
i
M
i
ψe
i
. (88)
The system and its environment are correlated in this way. This state must
satisfy the normalization condition since U is unitary; ψe
0
U
†
Uψe
0
=
i,j
ψe
i
M
†
i
M
j
⊗Iψe
j
= ψ
i
M
†
i
M
i
ψ = 1. Since ψ is arbitrary,
we must have
i
M
†
i
M
i
= I
S
, (89)
where I
S
is the unit matrix acting on the system Hilbert space H
S
. Opera
tors {M
†
i
M
i
} are said to form a POVM (positive operatorvalued measure).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
37
Suppose we measure the environment with a measurement operator
O = I
S
⊗
i
λ
i
e
i
e
i
 =
i
λ
i
(I
S
⊗ e
i
e
i
) .
We obtain a measurement outcome λ
k
with a probability
p(k) = Ψ(I
S
⊗ e
k
e
k
)Ψ
=
i,j
ψe
i
M
†
i
(I
S
⊗ e
k
e
k
)M
j
ψe
j
= ψM
†
k
M
k
ψ, (90)
where Ψ = Uψe
0
. The combined system immediately after the mea
surement is
1
p(k)
(I
S
⊗ e
k
e
k
)Uψe
0
=
1
p(k)
(I
S
⊗ e
k
e
k
)
i
M
i
ψe
i
=
1
p(k)
M
k
ψe
k
. (91)
Let ρ
S
=
i
p
i
ψ
i
ψ
i
 be an arbitrary density matrix of the principal
system. It follows from the above observation for a pure state ψψ that
the reduced density matrix immediately after the measurement is
k
p(k)
M
k
ρ
S
M
†
k
p(k)
=
k
M
k
ρ
S
M
†
k
. (92)
This shows that POVM measurement is a quantum operation in which the
Kraus operators are given by the generalized measurement operators {M
i
}.
The projective measurement is a special class of POVM, in which {M
i
} are
the projective operators.
6.3. Examples
Now we examine several important examples which have relevance in quan
tum information theory. Decoherence appears as an error in quantum infor
mation processing. The next chapter is devoted to strategies to ﬁght against
errors introduced in this section.
6.3.1. Bitﬂip channel
Consider a closed twoqubit system with a Hilbert space C
2
⊗C
2
. We call
the ﬁrst qubit the “(principal) system” while the second qubit the “envi
ronment”. A bitﬂip channel is deﬁned by a quantum operation
E(ρ
S
) = (1 − p)ρ
S
+ pσ
x
ρ
S
σ
x
, 0 ≤ p ≤ 1. (93)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
38
The input ρ
S
is bitﬂipped with a probability p while it remains in its input
state with a probability 1 − p. The Kraus operators are read oﬀ as
E
0
=
1 − pI, E
1
=
√
pσ
x
. (94)
Fig. 4. Quantum circuit modelling a bitﬂip channel. The gate is the inverted CNOT
gate I ⊗ 00 + σ
x
⊗ 11.
The circuit depicted in Fig. 4 models the bitﬂip channel provided that
the second qubit is in a mixed state (1 − p)00 + p11. The circuit is
nothing but the inverted CNOT gate V = I ⊗ 00 + σ
x
⊗ 11. The
output of this circuit is
V (ρ
S
⊗ [(1 − p)00 + p11]) V
†
= (1 − p)ρ
S
⊗ 00 + pσ
x
ρ
S
σ
x
11, (95)
from which we obtain
E(ρ
S
) = (1 − p)ρ
S
+ pσ
x
ρ
S
σ
x
(96)
after tracing over the environment Hilbert space.
The choice of the second qubit input state is far from unique and so is
the choice of the circuit. Suppose the initial state of the environment is a
pure state ψ
E
=
√
1 − p0 +
√
p1, for example. Then the output of the
circuit in Fig. 4 is
E(ρ
S
) = tr
E
[V ρ
S
⊗ ψ
E
ψ
E
V
†
] = (1 − p)ρ
S
+ pσ
x
ρ
S
σ
x
, (97)
producing the same result as before.
Let us see what transformation this quantum operation brings about in
ρ
S
. We parametrize ρ
S
using the Bloch vector as
ρ
S
=
1
2
I +
k=x,y,z
c
k
σ
k
, (c
k
∈ R) (98)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
39
where
k
c
2
k
≤ 1. We obtain
E(ρ
S
) = (1 − p)ρ
S
+ pσ
x
ρ
S
σ
x
=
1 − p
2
(I + c
x
σ
x
+ c
y
σ
y
+ c
z
σ
z
) +
p
2
(I + c
x
σ
x
− c
y
σ
y
− c
z
σ
z
)
=
1
2
1 + (1 − 2p)c
z
c
x
− i(1 − 2p)c
y
c
x
+ i(1 − 2p)c
y
1 − (1 − 2p)c
z
. (99)
Observe that the radius of the Bloch sphere is reduced along the y and the
zaxes so that the radius in these directions is 1−2p. Equation (99) shows
that the quantum operation has produced a mixture of the Bloch vector
states (c
x
, c
y
, c
z
) and (c
x
, −c
y
, −c
z
) with weights 1 − p and p respectively.
Figure 5 (a) shows the Bloch sphere which represents the input qubit states.
The Bloch sphere shrinks along the y and zaxes, which results in the
ellipsoid shown in Fig. 5 (b).
1
0.5
0
0.5
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
Fig. 5. Bloch sphere of the input state ρ
S
(a) and output states of (b) bitﬂip channel
and (c) phaseﬂip channel. The probability p = 0.2 is common to both channels.
6.3.2. Phaseﬂip channel
Consider again a closed twoqubit system with the “(principal) system” and
its “environment”.
The phaseﬂip channel is deﬁned by a quantum operation
E(ρ
S
) = (1 − p)ρ
S
+ pσ
z
ρ
S
σ
z
, 0 ≤ p ≤ 1. (100)
The input ρ
S
is phaseﬂipped (0 → 0 and 1 → −1) with a proba
bility p while it remains in its input state with a probability 1 − p. The
corresponding Kraus operators are
E
0
=
1 − pI, E
1
=
√
pσ
z
. (101)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
40
Fig. 6. Quantum circuit modelling a phaseﬂip channel. The gate is the inverted
controlledσ
z
gate.
A quantum circuit which models the phaseﬂip channel is shown in
Fig. 6. Let ρ
S
be the ﬁrst qubit input state while (1 − p)00 + p11 be
the second qubit input state. The circuit is the inverted controlledσ
z
gate
V = I ⊗ 00 + σ
z
⊗ 11.
The output of this circuit is
V (ρ
S
⊗ [(1 − p)00 + p11]) V
†
= (1 − p)ρ
S
⊗ 00 + pσ
z
ρ
S
σ
z
⊗ 11, (102)
from which we obtain
E(ρ
S
) = (1 − p)ρ
S
+ pσ
z
ρ
S
σ
z
. (103)
The second qubit input state may be a pure state
ψ
E
=
1 − p0 +
√
p1, (104)
for example. Then we ﬁnd
E(ρ
S
) = tr
E
[V ρ
S
⊗ ψ
E
ψ
E
V
†
] = E
0
ρ
S
E
†
0
+ E
1
ρ
S
E
†
1
, (105)
where the Kraus operators are
E
0
= 0V ψ
E
=
1 − pI, E
1
= 1V ψ
E
=
√
pσ
z
. (106)
Let us work out the transformation this quantum operation brings about
to ρ
S
. We parametrize ρ
S
using the Bloch vector as before. We obtain
E(ρ
S
) = (1 − p)ρ
S
+ pσ
z
ρ
S
σ
z
=
1 − p
2
(I + c
x
σ
x
+ c
y
σ
y
+ c
z
σ
z
) +
p
2
(I − c
x
σ
x
− c
y
σ
y
+ c
z
σ
z
)
=
1
2
1 + c
z
(1 − 2p)(−c
x
− ic
y
)
(1 − 2p)(c
x
+ ic
y
) 1 − c
z
. (107)
Observe that the oﬀdiagonal components decay while the diagonal com
ponents remain the same. Equation (107) shows that the quantum oper
ation has produced a mixture of the Bloch vector states (c
x
, c
y
, c
z
) and
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
41
(−c
x
, −c
y
, c
z
) with weights 1 −p and p respectively. The initial state has a
deﬁnite phase φ = tan
−1
(c
y
/c
x
) in the oﬀdiagonal components. The phase
after the quantum operaition is applied is a mixture of states with φ and
φ+π. This process is called the phase relaxation process, or the T
2
process
in the context of NMR. The radius of the Bloch sphere is reduced along
the x and the yaxes as 1 → 1 − 2p. Figure 5 (c) shows the eﬀect of the
phaseﬂip channel on the Bloch sphere for p = 0.2.
Other examples will be found in [1,2].
7. Quantum Error Correcting Codes
7.1. Introduction
It has been shown in the previous chapter that interactions between a quan
tum system with environment cause undesirable changes in the state of the
quantum system. In the case of qubits, they appear as bitﬂip and phase
ﬂip errors, for example. To reduce such errors, we must implement some
sort of error correcting mechanism in the algorithm.
Before we introduce quantum error correcting codes, we have a brief look
at the simplest version of error correcting code in classical bits. Suppose
we transmit a serise of 0’s and 1’s through a noisy classical channel. Each
bit is assumed to ﬂip independently with a probability p. Thus a bit 0 sent
through the channel will be received as 0 with probability 1 − p and as 1
with probability p. To reduce channel errors, we may invoke to majority
vote. Namely, we encode logical 0 by 000 and 1 by 111, for example. When
000 is sent through this channel, it will be received as 000 with probability
(1−p)
3
, as 100, 010 or 001 with probability 3p(1−p)
2
, as 011, 101 or 110 with
probability 3p
2
(1 −p) and ﬁnally as 111 with probability p
3
. By taking the
majority vote, we correctly reproduce the desired result 0 with probability
p
0
= (1 − p)
3
+ 3p(1 − p)
2
= (1 − p)
2
(1 + 2p) while fails with probability
p
1
= 3p
2
(1 −p) +p
3
= (3 −2p)p
2
. We obtain p
0
p
1
for suﬃciently small
p ≥ 0. In fact, we ﬁnd p
0
= 0.972 and p
1
= 0.028 for p = 0.1. The success
probability p
0
increases as p approaches to 0, or alternatively, if we use
more bits to encode 0 or 1.
This method cannot be applicable to qubits, however, due to nocloning
theorem. We have to somehow think out the way to overcome this theorem.
7.2. Threequbit bitﬂip code: the simplest example
It is instructive to introduce a simple example of quantum error correcting
codes (QECC). We closely follow Steane
30
here.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
42
7.2.1. Bitﬂip QECC
Suppose Alice wants to send a qubit or a series of qubits to Bob through a
noisy quantum channel. Let ψ = a0+b1 be the state she wants to send.
If she is to transmit a serise of qubits, she sends them one by one and the
following argument applies to each of the qubits. Let p be the probability
with which a qubit is ﬂipped and we assume there are no other types of
errors in the channel. In other words, the operator X is applied to the qubit
with probability p and consequently the state is mapped to
ψ → ψ
= Xψ = a1 + b0. (108)
We have already seen in the previous section that this channel is described
by a quantum operation (93).
7.2.2. Encoding
To reduce the error probability, we want to mimic somehow the classical
counterpart without using a clone machine. Let us recall that the action of
a CNOT gate is CNOT : j0 → jj, j ∈ {0, 1} and therefore it duplicates
the control bit j ∈ {0, 1} when the target bit is initially set to 0. We use
this fact to triplicate the basis vectors as
ψ00 = (a0 + b1)00 → ψ
E
= a000 + b111, (109)
where ψ
E
denotes the encoded state. The state ψ
E
is called the logical
qubit while each constituent qubit is called the physical qubit. We borrow
terminologies from classical error correcting code (ECC) and call the set
C = {a000 + b111a, b ∈ C, a
2
+ b
2
= 1} (110)
the code and each member of C a codeword. It is important to note that the
state ψ is not triplicated but only the basis vectors are triplicated. This
redundancy makes it possible to detect errors in ψ
E
and correct them as
we see below.
A quantum circuit which implements the encoding (109) is easily found
from our experience in CNOT gate. Let us consider the circuit shown in
Fig. 7 (a) whose input state is ψ00. It is immediately found that the
output of this circuit is ψ
E
= a000 + b111 as promised.
7.2.3. Transmission
Now the state ψ
E
is sent through a quantum channel which introduces
bitﬂip error with a rate p for each qubit independently. We assume p is
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
43
Fig. 7. Quantum circuits to (a) encode, (b) detect bitﬂip error syndrome, (c) make
correction to a relevant qubit and (d) decode. The gate N
X
stands for the bitﬂip noise.
suﬃciently small so that not many errors occur during qubit transmission.
The received state depends on in which physical qubit(s) the bitﬂip error
occurred. Table 1 lists possible received states and the probabilities with
which these states are received.
Table 1. State Bob receives and the
probability which this may happen.
State Bob receives Probability
a000 + b111 (1 − p)
3
a100 + b011 p(1 − p)
2
a010 + b101 p(1 − p)
2
a001 + b110 p(1 − p)
2
a110 + b001 p
2
(1 − p)
a101 + b010 p
2
(1 − p)
a011 + b100 p
2
(1 − p)
a111 + b000 p
3
7.2.4. Error syndrome dectection and correction
Now Bob has to extract from the received state which error occurred during
qubits transmission. For this purpose, Bob prepares two ancillary qubits in
the state 00 as depicted in Fig. 7 (b) and apply four CNOT operations
whose control bits are the encoded qubits while the target qubits are Bob’s
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
44
two ancillary qubits. Let x
1
x
2
x
3
be a basis vectors Bob has received and
let A (B) be the output state of the ﬁrst (second) ancilla qubit. It is seen
from Fig. 7 (b) that A = x
1
⊕x
2
and B = x
1
⊕x
3
. Let a100+b011 be the
received logical qubit for example. Note that the ﬁrst qubit state in both
of the basis vectors is diﬀerent from the second and the third qubit states.
These diﬀerence are detected by the pairs of CNOT gates in Fig. 7 (b). The
error extracting sequence transforms the anicillary qubits as
(a100 + b011)00 → a10011 + b01111 = (a100 + b011)11.
Both of the ancillary qubits are ﬂipped since x
1
⊕ x
2
= x
1
⊕ x
3
= 1 for
both 100 and 011. It is important to realize that (i) the syndrome is
independent of a and b and (ii) the received state a100 + b011 remains
intact; we have detected an error without measuring the received state!
These features are common to all QECC.
We list the result of other cases in Table 2. Note that among eight
Table 2. States after error extraction is made and the
probabilities with which these states are produced.
State after error syndrome extraction Probability
(a000 + b111)00 (1 − p)
3
(a100 + b011)11 p(1 − p)
2
(a010 + b101)10 p(1 − p)
2
(a001 + b110)01 p(1 − p)
2
(a110 + b001)01 p
2
(1 − p)
(a101 + b010)10 p
2
(1 − p)
(a011 + b100)11 p
2
(1 − p)
(a111 + b000)00 p
3
possible states, there are exactly two states with the same ancilla state. Does
it mean this error extraction scheme does not work? Now let us compare the
probabilities associated with the same ancillary state. When the ancillary
state is 10, for example, there are two possible reseived states a010 +
b101 and a101+b010. Note that the former is received with probability
p(1 − p)
2
while that latter with p
2
(1 − p). Therefore the latter probability
is negligible compared to the former for suﬃciently small p.
It is instructive to visualize what errors do to the encoded basis vectors.
Consider a cube with the unit length. The vertices of the cube have coor
dinates (i, j, k) where i, j, k ∈ {0, 1}. We assign a vector ijk to the vertex
(i, j, k), under which the vectors 000 and 111 correspond to diagonally
separated vertices. An action of X
i
, the operator X = σ
x
acting on the
ith qubit, sends these basis vectors to the nearest neighbor vetices, which
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
45
diﬀer from the correct basis vectors in the ith position. The intersection
of the sets of vectors obtained by a single action of X
i
on 000 and 111
is an empty set. Therefore an action of a single error operator X can be
corrected with no ambiguity.
Now Bob measures his ancillary qubits and obtains two bits of classical
information. The set of two bits is called the (error) syndrome and it tells
Bob in which physical qubit the error occurred during transmission. Bob
applies correcting procedure to the received state according to the error
sydrome he has obtained. Ignoring extra error states with small probabili
ties, we immediately ﬁnd that the following action must be taken:
error syndrome correction to be made
00 identity operation (nothing is required)
01 apply σ
x
to the third qubit
10 apply σ
x
to the second qubit
11 apply σ
x
to the ﬁrst qubit
Suppose the syndrome is 01, for example. The state Bob received is likely to
be a001 +b110. Bob recovers the initial state Alice has sent by applying
I ⊗ I ⊗ σ
x
on the received state:
(I ⊗ I ⊗ σ
x
)(a001 + b110) = a000 + b111.
If Bob receives the state a110 + b001, unfortunately, he will obtain
(I ⊗ I ⊗ σ
x
)(a110 + b001) = a111 + b000.
In fact, for any error sydrome, Bob obtaines either a000+b111 or a111+
b000. The latter case occurs if and only if more than one qubit are ﬂipped,
and hence it is less likely to happen for suﬃciently small error rate p. The
probability with which multiple error occurs is found from Table 1 as
P(error) = 3p
2
(1 − p) + p
3
= 3p
2
− 2p
3
. (111)
This error rate is less than p if p < 1/2. In contrast, success probability has
been enhanced from 1 − p to 1 − P(error) = 1 − 3p
2
+ 2p
3
. Let p = 0.1,
for example. Then the error rate is lowered to P(error) = 0.028, while the
success probability is enhanced from 0.9 to 0.972.
7.2.5. Decoding
Now that Bob has corrected an error, what is left for him is to decode
the encoded state. This is nothing but the inverse transformation of the
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
46
encoding (109). It can be seen from Fig. 7 (d) that
CNOT
12
CNOT
13
(a000 + b111) = a000 + b100 = (a0 + b1)00.
(112)
7.2.6. Miracle of entanglement
This example, albeit simple, contains almost all fundamental ingredients of
QECC. We prepare some redundant qubits which somehow “triplicate” the
original qubit state to be sent without violating nocloning theorem. Then
the encoded qubits are sent through a noisy channel, which causes a bitﬂip
in at most one of the qubits. The received state, which may be subject to an
error, is then entangled with ancillary qubits, whose state reﬂects the error
which occurred during the state transmission. This results in an entangled
state
k
A bitﬂip error in the kth qubit ⊗ corresponding error syndrome.
(113)
The wave function, upon the measurement of the ancillary qubits, collapses
to a state with a bitﬂip error corresponding to the observed error syndrome.
In a sense, syndrome measurement singles out a particular error state which
produces the observed syndrome.
Once syndrome is found, it is an easy task to transform the received state
back to the original state. Note that everything is done without knowing
what the origial state is.
7.2.7. Continuous rotations
We have considered noise X so far. Suppose noise in the channel is charac
terized by a contiuous paramter α as
U
α
= e
iαX
= cos αI + iX sin α, (114)
which maps a state ψ to
U
α
ψ = cos αψ + i sinαXψ. (115)
Suppose U
α
acts on the ﬁrst qubit, for example. Bob then receives
(U
α
⊗ I ⊗ I)(a000 + b111)
= cos α(a000 + b111) + i sinα(a100 + b011).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
47
The output of the error syndrome detection circuit, before the syndrome
measurement is made, is an entangled state
cos α(a000 + b111)00 + i sinα(a100 + b111)11, (116)
see Table 2. Measurement of the error syndrome yields either 00 or 11.
In the former case the state collapses to ψ = a000 + b111 and this
happens with a probability cos
2
α. In the latter case, on the other hand,
the received state collapses to Xψ = a100 + b011 and this happens
with a probability sin
2
α. Bob applies I (X) to the ﬁrst qubit to correct the
error when the syndrome readout is 00 (11).
It is clear that error U
α
may act on the second or the third qubit.
Continuous rotation U
α
for any α may be corrected in this way. In general,
linearity of a quantum circuit guarantees that any QECC, which corrects
the bitﬂip error X, corrects continuous error U
α
.
8. DiVincenzo Criteria
We have learned so far that information may be encoded and processed in
a quantummechanical way. This new discipline called quantum informa
tion processing (QIP) is expected to solve a certain class of problems that
current digital computers cannot solve in a practical time scale. Although
a small scale quantum information processor is already available commer
cially, physical realization of large scale quantum information processors is
still beyond the scope of our currently available technology.
A quantum computer should have at least 10
2
∼ 10
3
qubits to be able
to execute algorithms that are more eﬃcient than their classical counter
parts. DiVincenzo proposed necessary conditions, socalled the DiVincenzo
criteria that any physical system has to fulﬁll to be a candidate for a viable
quantum computer.
38
In the next section, we outline these conditions as
well as two additional criteria for networkability.
8.1. DiVincenzo criteria
In his inﬂuential article,
38
DiVincenzo proposed ﬁve criteria that any phys
ical system must satisfy to be a viable quantum computer. We summarize
the relevant parts of these criteria in this section.
(1) A scalable physical system with well characterized qubits.
To begin with, we need a quantum register made of many qubits to store
information. Recall that a classical computer also requires memory to
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
48
store information. The simplest way to realize a qubit physically is
to use a twolevel quantum system. For example, an electron, a spin
1/2 nucleus or two mutually orthogonal polarization states (horizontal
and vertical, for example) of a single photon can be a qubit. We may
also employ a twodimensional subspace, such as the ground state and
the ﬁrst excited state, of a multidimensional Hilbert space, such as
atomic energy levels. In any case, the two states are identiﬁed as the
basis vectors, 0 and 1, of the Hilbert space so that a general single
qubit state takes the form ψ = α0 + β1, α
2
+ β
2
= 1. A multi
qubit state is expanded in terms of the tensor products of these basis
vectors. Each qubit must be separately addressable. Moreover it should
be scalable up to a large number of qubits. The twodimensional vector
space of a qubit may be extended to threedimensional (qutrit) or, more
generally, ddimensional (qudit).
A system may be made of several diﬀerent kinds of qubits. Qubits
in an ion trap quantum computer, for instance, may be deﬁned as:
(1) hyperﬁne/Zeeman sublevels in the electronic ground state of ions
(2) a ground state and an excited state of a weakly allowed optical
transition and (3) normal mode of ion oscillation. A similar scenario is
also proposed for Josephson junction qubits, in which two ﬂux qubits
are coupled through a quantized LC circuit. Simultaneous usage of
several types of qubits may be the most promising way to achieving a
viable quantum computer.
(2) The ability to initialize the state of the qubits to a simple ﬁducial state,
such as 00 . . . 0.
Suppose you are not able to reset your (classical) computer. Then you
will never trust the output of some computation even though processing
is done correctly. Therefore initialization is an important part of both
quantum and classical information processors.
In many realizations, initialization may be done simply by cooling to
bring the system into its ground state. Let ∆E be the diﬀerence between
energies of the ﬁrst excited state and the ground state. The system is
in the ground state with a good precision at low temperatures satis
fying k
B
T ∆E. Alternatively, we may use projective measurement
to project the system onto a desired state. In some cases, we observe
the system to be in an undesired state upon such measurement. Then
we may transform the system to the desired ﬁducial state by applying
appropriate gates.
For some realizations, such as liquid state NMR, however, it is im
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
49
possible to cool the system down to extremely low temperatures. In
those cases, we are forced to use a thermally populated state as an ini
tial state. This seemingly diﬃcult problem may be amended by several
methods if some computational resources are sacriﬁced. We then obtain
an “eﬀective” pure state, socalled the pseudopure state, which works
as an initial state for most purposes.
Continuous fresh supply of qubits in a speciﬁed state, such as 0, is
also an important requirement for successful quantum error correction.
as we have seen in Section 7.
(3) Long decoherence times, much longer than the gate operation time.
Hardware of a classical computer lasts long, for on the order of 10 years.
Things are totally diﬀerent for a quantum computer, which is fragile
against external disturbance called decoherence, see Section 6.
Decoherence is probably the hardest obstacle to building a viable quan
tum computer. Decoherence means many aspects of quantum state
degradation due to interactions of the system with the environment
and sets the maximum time available for quantum computation. De
coherence time itself is not very important. What matters is the ratio
“decoherence time/gate operation time”. For some realizations, deco
herence time may be as short as ∼ µs. This is not necessarily a big
problem provided that the gate operation time, determined by the Rabi
oscillation period and the qubitcoupling strength, for example, is much
shorter than the decoherence time. If the typical gate operation time
is ∼ ps, say, the system may execute 10
12−6
= 10
6
gate operations
before the quantum state decays. We quote the number ∼ 10
5
of gates
required to factor 21 into 3 and 7 by using Shor’s algorithm.
40
There are several ways to eﬀectively prolong decoherence time. A
closedloop control method incorporates QECC, while an openloop
control method incorporates noiseless subsystem
41
and decoherence free
subspace (DFS).
42
(4) A “universal” set of quantum gates.
Suppose you have a classical computer with a big memory. Now you
have to manipulate the data encoded in the memory by applying various
logic gates. You must be able to apply arbitrary logic operations on the
memory bits to carry out useful information processing. It is known that
the NAND gate is universal, i.e., any logic gates may be implemented
with NAND gates.
Let H(γ(t)) be the Hamiltonian of an nqubit system under con
sideration, where γ(t) collectively denotes the control parameters in
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
50
the Hamiltonian. The timedevelopment operator of the system is
U[γ(t)] = T exp
−
i
T
H(γ(t))dt
∈ U(2
n
), where T is the time
ordering operator. Our task is to ﬁnd the set of control parameters
γ(t), which implements the desired gate U
gate
as U[γ(t)] = U
gate
. Al
though this “inverse problem” seems to be diﬃcult to solve, a theorem
by Barenco et al. guarantees that any U(2
n
) gate may be decomposed
into singlequbit gates ∈ U(2) and CNOT gates.
16
Therefore it suﬃces
to ﬁnd the control sequences to implement U(2) gates and a CNOT gate
to construct an arbitrary gate. Naturally, implementation of a CNOT
gate in any realization is considered to be a milestone in this respect.
Note, however, that any twoqubit gates, which are neither a tensor
product of two onequbit gates nor a SWAP gate, work as a component
of a universal set of gates.
43
(5) A qubitspeciﬁc measurement capability.
The result of classical computation must be displayed on a screen or
printed on a sheet of paper to readout the result. Although the read
out process in a classical computer is regarded as too trivial a part of
computation, it is a vital part in quantum computing.
The state at the end of an execution of quantum algorithm must be
measured to extract the result of the computation. The measurement
process depends heavily on the physical system under consideration. For
most realizations, projective measurements are the primary method to
extract the outcome of a computation. In liquid state NMR, in con
trast, a projective measurement is impossible, and we have to resort to
ensemble averaged measurements.
Measurement in general has no 100% eﬃciency due to decoherence, gate
operation error and many more reasons. If this is the case, we have to
repeat the same computation many times to achieve reasonably high
reliability.
Moreover, we should be able to send and store quantum information
to construct a quantum data processing network. This “networkability”
requires following two additional criteria to be satisﬁed.
(6) The ability to interconvert stationary and ﬂying qubits.
Some realizations are excellent in storing quantum information while
long distant transmission of quantum information might require diﬀer
ent physical resources. It may happen that some system has a Hamil
tonian which is easily controllable and is advantageous in executing
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
51
quantum algorithms. Compare this with a current digital computer, in
which the CPU and the system memory are made of semiconductors
while a hard disk drive is used as a mass storage device. Therefore a
working quantum computer may involve several kinds of qubits and
we are forced to introduce distributed quantum computing. Intercon
verting ability is also important in long distant quantum teleportation
using quantum repeaters.
(7) The ability to faithfully transmit ﬂying qubits between speciﬁed loca
tions.
Needless to say, this is an indispensable requirement for quantum com
munication such as quantum key distribution. This condition is also
important in distributed quantum computing mentioned above.
8.2. Physical realizations
There are numerous physical systems proposed as possible candidates for a
viable quantum computer to date.
44
Here is the list of the candidates;
(1) Liquidstate/Solidstate NMR and ENDOR
(2) Trapped ions
(3) Neutral atoms in optical lattice
(4) Cavity QED with atoms
(5) Linear optics
(6) Quantum dots (spinbased, chargebased)
(7) Josephson junctions (charge, ﬂux, phase qubits)
(8) Electrons on liquid helium surface
and other unique realizations. ARDA QIST roadmap
44
evaluates each of
these realizations. The roadmap is extremely valuable for the identiﬁcation
and quantiﬁcation of progress in this multidisciplinary ﬁeld.
Acknowledgements
This summer school was supported by the “Open Research Center” Project
for Private Universities: matching fund subsidy from MEXT (Ministry of
Education, Culture, Sports, Science and Technology). Special thanks are
due to the other organizers and coeditors of this lecture notes, Takashi
Aoki, Robabeh Rahimi Darabad and Akira SaiToh.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
52
References
1. M. Nakahara and T. Ohmi, Quantum Computing: From Linear Algebra to
Physical Realizations, (Taylor and Francis, 2008).
2. M. A. Neilsen and I. L. Chuang, Quantum Computation and Quantum In
formation, (Cambridge University Press, 2000).
3. E. Rieﬀel and W. Polak, ACM Computing Surveys (CSUR) 32 (2000) 300.
4. Y. Uesaka, Mathematical Principle of Quantum Computation, (Corona Pub
lishing, Tokyo, in Japanese, 2000).
5. P. A. M. Dirac, Principles of Quantum Mechanics (4th ed.), (Clarendon
Press, 1981).
6. L. I. Shiﬀ, Quantum Mechanics (3rd ed.), (McGrawHill, 1968).
7. A. Messiah, Quantum Mechanics, (Dover, 2000).
8. J. J. Sakurai, : Modern Quantum Mechanics (2nd Edition), (Addison Wesley,
Boston, 1994).
9. L. E. Ballentine, Quantum Mechanics, (World Scientiﬁc, Singapore, 1998).
10. A. Peres, Quantum Theory: Concepts and Methods, (Springer, 2006).
11. A. SaiToh, R. Rahimi and M. Nakahara, eprint quantph/0703133.
12. A. Peres, Phys. Rev. Lett. 77 (1996) 1413.
13. M. Horodecki et al., Phys. Lett. A 223 (1996) 1.
14. W. K. Wootters, and W. H. Zurek, Nature 299 (1982) 802.
15. M. A. Nielsen et al., Nature 396 (1998) 52.
16. A. Barenco et al., Phys. Rev. A 52 (1995) 3457.
17. Z. Meglicki, http://beige.ucs.indiana.edu/M743/index.html
18. D. Deutsch, Proc. Roy. Soc. Lond. A, 400 (1985) 97.
19. D. Deutsch and R. Jozsa, Proc. Roy. Soc. Lond. A, 439 (1992) 553.
20. E. Bernstein and U. Vazirani, SIAM J. Comput., 26 (1997) 1411.
21. D. R. Simon, Proc. 35th Annual Sympo. Found. Comput. Science, (IEEE
Comput. Soc. Press, Los Alamitos, 1994) 116.
22. T. Mihara and S. C. Sung, Comput. Complex. 12 (2003) 162.
23. M. A. Neilsen and I. L. Chuang, Quantum Computation and Quantum In
formation, (Cambridge University Press, 2000).
24. K. Hornberger, quantph/0612118.
25. H. Barnum, M. A. Nielsen and B. Schumacher, Phys. Rev. A 57 (1998) 4153.
26. Y. Kondo, et al., J. Phys. Soc. Jpn. 76 (2007) 074002.
27. G. Lindblad, Commun. Math. Phys. 48 (1976) 119.
28. V. Gorini, A. Kossakowski and E. C. G. Sudarshan, J. Math. Phys., 17 (1976)
821.
29. A. J. Fisher, Lecture note available at
http://www.cmmp.ucl.ac.uk/~ajf/course_notes.pdf
30. A. M. Steane, quantph/0304016.
31. P. W. Shor, Phys. Rev. A 52 (1995) 2493.
32. A. Hosoya, Lectures on Quantum Computation (Science Sha, in Japanese,
1999).
33. F. J. MacWilliams and N. J. A. Sloane, The Theory of ErrorCorrecting
Codes, (NorthHolland, Amsterdam, 1977).
34. A. R. Calderbank and P. W. Shor, Phys. Rev. A 54 (1996) 1098.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
53
35. A. M. Steane, Phys. Rev. Lett. 77 (1996) 793.
36. J Niwa, K. Matsumoto and H. Imai, quantph/0211071.
37. D. P. DiVincenzo and P. W. Shor, Phys. Rev. Lett. 77 (1996) 3260.
38. D. P. DiVincenzo, Fortschr. Phys. 48 (2000) 771.
39. M. Nakahara, S. Kanemitsu, M. M. Salomaa and S. Takagi (eds.) “Physical
Realization of Quantum Computing: Are the DiVincenzo Criteria Fulﬁlled
in 2004?” (World Scientiﬁc, Singapore, 2006).
40. J. Vartiainen et al., Phys. Rev. A 70 (2004) 012319.
41. E. Knill, R. Laﬂamme, and L. Viola, Phys. Rev. Lett. 84, 2525 (2000); P.
Zanardi, Phys. Rev. A 63, 012301 (2001); W. G. Ritter, Phys. Rev. A 72
(2005) 012305.
42. G. M. Palma, K. A. Suominen and A. K. Ekert, Proc. R. Soc. London A 452
(1996) 567; L. M. Duan and G. C. Guo, Phys. Rev. Lett. 79 (1997) 1953;
P. Zanardi and M. Rasetti, Phys. Rev. Lett. 79 (1997) 3306; D. A. Lidar, I.
L. Chuang, and K. B. Whaley, Phys. Rev. Lett. 81 (1998) 2594; P. Zanardi,
Phys. Rev. A 60 (1999) 729(R); D. Bacon, D. A. Lidar, and K. B. Whaley,
Phys. Rev. A 60 (1999) 1944.
43. D. P. DiVincenzo, Phys. Rev. A 51 (1995) 1015.
44. http://qist.lanl.gov/
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
This page intentionally left blank This page intentionally left blank
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
55
BRAID GROUP AND TOPOLOGICAL QUANTUM
COMPUTING
TAKAYOSHI OOTSUKA
Department of Physics, Kinki University, Osaka 5778502, Japan
∗
Email: ootsuka@alice.math.kindai.ac.jp
KAZUHIRO SAKUMA
Department of Mathematics, Kinki University, Osaka 5778502, Japan
∗
Email: sakuma@math.kindai.ac.jp
This is a survey article on the braid group and topological quantum computing.
The purpose of this note is to discuss relation between a unitary representation
of braid group and topological quantum computing.
Keywords: Symmetric Group, Braid Group, Presentation of a Group, Represen
tation of a Group, Alexander Theorem, Markov Theorem, Jones Polynomial,
Modular Functor, TQC.
1. Introduction
Historically in the process of development of theoretical physics, modern
geometry has constantly provided a strong method and always played an
important role at both fundamental and advanced level. For instance, it
is obviously concluded from the facts that the notion of a Riemannian
manifold, which is a direct generalization of surface theory, was employed as
a model of the space universe, and that general relativity could be discussed
as the Lorentz manifold theory. Moreover, it is impossible to discuss the
gravitation ﬁeld equation due to Einstein
R
ij
−
1
2
Rg
ij
+ Λg
ij
= T
ij
without the language of Riemannian geometry. Also, we may say that based
on the characterization and requirement from physics side, there has been
much progress in the study of geometry closely related to manifold theory.
In the Erlangen program Felix Klein declared that “for one transforma
tion group, there corresponds one geometry.” This conception by Klein has
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
56
been one of the most important teaching principles in modern geometry.
More precisely, for a given space X and a transformation group G, we can
study invariant properties of a subset A in X under a transformation in
G, which is called a geometry of X belonging to G. For example, the Eu
clidean geometry is a branch to study invariant properties of a ﬁgure A in
X = R
2
under congruent transformation or similar transformation. In its
simplest form it is the SideAngleSide theorem of middle school geometry
which characterizes triangles up to congruence or the AngleAngle theorem
up to similarity. In addition, diﬀerential topology is a branch to study in
variant properties of a manifold up to diﬀeomorphism or homeomorphism.
Thus from a modern mathematical viewpoint, geometry is a mathematics
in which one searches for invariant properties of not only ﬁgures but also
functions on the ﬁgures or maps between ﬁgures under an element of a
group G.
The ultimate purpose of group theory which is one of the branches in
algebra is to give a group structure for any set and classify all groups up
to isomorphism. Group theory was established by Galois with an excel
lent idea in order to solve an equation systematically. It gave not merely
a solution of an equation but an answer to a geometrical problem (e.g.
a regular npolygon is of possible construction if and only if the Galois
group
a
Gal(Q(e
πi
n
)/Q
∼
= (Z/2nZ)
×
has order 2
r
). Thus when we consider
a problem basically related with group theory, we see that it is tied up with
geometry. Conversely, a group naturally appears in considering a geometri
cal problem. This is because geometry is a mathematics in which one solves
equations on a geometrical object.
On one hand, a group often appears also in physics in a natural way.
As is seen in § 4, braid groups are used in describing topological quantum
computation, which is a recent work by M. H. Freedman, M. Larsen and
Z. Wang. So, in the next section we discuss braid groups together with
basic facts on group theory. For readers’ convenience, we provide an ap
pendix where we discuss basic facts on the fundamental groups including
its deﬁnition.
Acknowledgement. The authors wishes to thank Mikio Nakahara for
well organizing summer school held at August 2007 and stimulating dis
cussion. He also wishes to thank the participants at the summer school for
a
Here we do not give a deﬁnition of Galois group because we do not need it in the later
discussion
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
57
their useful comments and inspiring questions. The second author would
like to thank his student Kumi Kobata for her help in writing this paper.
2. Braid Groups
In this section we review basic concepts of braid group mainly from an
algebraic viewpoint, which will be a quick course for physicists. First we
start with recalling the deﬁnition of a group and fundamental concepts in
group theory:
Deﬁnition 2.1. For any set G, if a map (binary operation) µ : G×G → G
satisﬁes the following three conditions, then a pair (G, µ) or simply G is
called a group.
(1) For any a, b, c ∈ G we have
µ(a, µ(b, c)) = µ(µ(a, b), c).
(2) For any a ∈ G there exists an element e ∈ G such that
µ(a, e) = µ(e, a) = a,
where e is called a unit element of G.
(3) For any a ∈ G there exists an element a
∈ G such that
µ(a, a
) = µ(a
, a) = e,
where a
is called an inverse element of a and we write a
−1
∈ G.
In the above deﬁnition a binary operation µ is called a product and it is
usually written as a · b ∈ G or ab ∈ G in place of µ(a, b) for simplicity. Note
that we can make a set G to be a group with diﬀerent products in general.
Given a group (G, µ) if we have µ(a, b) = µ(b, a) for any a, b ∈ G, then
we call G an abelian group. Let H be a subset of a group (G, µ). If H is a
group under the product µ of G, H is called a subgroup of G. Note that a
subset H is a subgroup of G if and only if µ(a, b
−1
) ∈ H for any a, b ∈ H.
Let G be a group and H a subgroup of G. Then, for any g ∈ G we deﬁne
two subgroups as follows:
gH = {gh; h ∈ H}, Hg = {hg; h ∈ H}.
If gH = Hg for any g ∈ G, then H is called a normal subgroup of G. Note
that H is a normal subgroup of G if and only if g
−1
hg ∈ H for any g ∈ G
and any h ∈ H.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
58
Let H be a normal subgroup of G. Then we deﬁne g
1
∼ g
2
if and only
if g
1
g
−1
2
∈ H. This relation ∼ is an equivalence relation; gg
−1
= e ∈ H
implies g ∼ g, (g
1
g
−1
2
)
−1
= g
2
g
−1
1
∈ H implies g
2
∼ g
1
if we suppose that
g
1
∼ g
2
, and (g
1
g
−1
2
)(g
2
g
−1
3
) = g
1
g
−1
3
implies g
1
∼ g
3
if we suppose that
g
1
∼ g
2
, g
2
∼ g
3
. We denote the quotient group under this relation by
G/H. For any g ∈ G the equivalence class of g is denoted by [g] = gH =
Hg ∈ G/H, where g is called a representative. For [g
1
], [g
2
] ∈ G/H the
product is deﬁned as follows:
[g
1
][g
2
] = [g
1
g
2
] ∈ G/H.
Here note that the product is welldeﬁned, i.e., the product [g
1
g
2
] does not
depend on the choices of representatives g
1
, g
2
.
Example 2.1.
(1) Let R denote the set of real numbers. It is easy to see that R is a group
with a product deﬁned by µ(a, b) = a + b. Then we say that (R, µ)
is an additive group. However, (R, µ
) is not a group when we employ
another operation µ
(a, b) = ab. This is beacause there exists a unit
1 ∈ R but there is no inverse for 0 ∈ R. The ndimensional linear space
R
n
is naturally a group with µ(x, y) = x +y for any x, y ∈ R
n
.
(2) Let Z denote the set of integers. Then, Z is a group with µ(a, b) = a+b
and a subgroup of (R, µ). But (Z, µ
) is not a group with µ
(a, b) =
ab. The direct product Z ⊕ Z = {(x, y); x, y ∈ Z} is a group with
µ((x
1
, y
1
), (x
2
, y
2
)) = (x
1
+ x
2
, y
1
+ y
2
).
(3) Let K = Q (the set of rational numbers), R, or C (the set of com
plex numbers). Then, (K, +) is an additive group similarly as in (1).
Furthermore, it is easy to see that K − {0} is a group with a product
deﬁned by µ(a, b) = ab, which is called a multiplicative group.
(4) Let n be a natural number. It is easy to see that nZ = {nm; m ∈ Z}
is a normal subgroup of Z. Consider the quotient group Z/nZ. Then,
[a] = {a +nm; m ∈ Z} and [a] +[b] = [a +b] = {a +b +nm; m ∈ Z} is
welldeﬁned. It is often written as Z/nZ = Z
n
= {[0], [1], . . . , [n −1]}.
In order to compare the structures of given groups, we need to deﬁne a
map which preserves the products.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
59
Deﬁnition 2.2. Let (G, µ) and (G
, µ
) be groups. If a map ϕ : G → G
preserves the products, i.e.,
ϕ(µ(a, b)) = µ
(ϕ(a), ϕ(b)),
then ϕ is called a homomorphism
b
. In other words, we have the following
commutative diagram:
G×G
ϕ×ϕ
−−−−−−→ G
×G
µ
µ
G −−−−−−→
ϕ
G
.
Let e
be a unit element of G
. Given a homomorphism ϕ : G → G
, we can
deﬁne a set
Ker(ϕ) = {g ∈ G; ϕ(g) = e
},
which is called the kernel of ϕ. Moreover, we can deﬁne a set
Im(ϕ) = ϕ(G) = {ϕ(g) ∈ G
; ∀g ∈ G},
which is called the image of ϕ. Note that Ker(ϕ) is a normal subgroup of
G and Im(ϕ) a subgroup of G
.
If a homomorphism ϕ : G → G
is bijective, i.e.,
(1) if ϕ(a) = ϕ(b), then a = b, and
(2) Im(ϕ) = G
,
then ϕ is called an isomorphism. We say that G is isomorphic to G
if there
exists an isomorphism ϕ : G → G
, denoted by G
∼
= G
. When (1) holds, ϕ
is said to be injective. When (2) holds, ϕ is said to be surjective. In group
theory, two diﬀerent sets are identiﬁed even if they have diﬀerent products
as groups but there exists an isomorphism between them.
Given a homomorphism µ : G → G
, we can consider a quotient group
G/Ker(ϕ). Any element of G/Ker(ϕ) is denoted by [g] as an equivalence
class of g ∈ G. Then we deﬁne a map f : G/Ker(ϕ) → Im(ϕ) by f([g]) =
ϕ(g). It is easy to see that the map f is a welldeﬁned homomorphism and
bijective, which means that G/Ker(ϕ)
∼
= Im(ϕ).
Exercise 2.1. Let ϕ : G
1
→ G
2
be a homomorphism.
b
If we omit the notation of the products, the deﬁnition of a homomorphism is simply
given as follows: ϕ(ab) = ϕ(a)ϕ(b).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
60
(1) Show that if e ∈ G
1
is an unit element, then e uniquely exists and
ϕ(e) ∈ G
2
is also an unit element.
(2) Show that ϕ(g
−1
) = {ϕ(g)}
−1
.
(3) Prove that G
1
/Ker(ϕ)
∼
= Im(ϕ).
(4) Show that if ϕ is an isomorphism, then ϕ
−1
is also an isomorphism.
(5) Suppose that there exists a homomorphism ψ : G
2
→ G
1
such that
ψ ◦ ϕ = id
G
1
and ϕ ◦ ψ = id
G
2
. Then, show that ϕ is an isomorphism.
Deﬁnition 2.3. Let G be a group. If there are a ﬁnite (or inﬁnite) number
of elements g
1
, g
2
, . . . , g
k
, . . . ∈ G such that an arbitrary element of
g ∈ G can be expressed by a ﬁnite (or inﬁnite) number of products of those
g
1
, g
2
, . . . , g
−1
1
, g
−1
2
, . . ., then we say that g
1
, g
2
, . . . are generators of G
and that G is generated by those generators. If G is generated by a ﬁnite
number of generators, then G is said to be ﬁnitely generated.
Now we consider a sequence with a ﬁnite number of elements; a, b, . . . , n.
Such a sequence consisting of a, b, . . . , n is called a word. For example,
aa, abc, aabbcc, ababab, ccc, cab
are words deﬁned by three elements a, b, c. For two words x and y, we deﬁne
the product xy as its juxtaposition, e.g. for x = abc, y = cab we have
xy = abccab. We can consider an empty word “ ”or e, and the inverses
a
−1
, b
−1
, c
−1
for a, b, c, i.e., aa
−1
= e (empty word). We employ, namely,
such a (economical) rule that if a given word includes an empty word, then
we ignore it like the following:
abbb
−1
c = abec = ab c = abc.
Thus we should regard that abbb
−1
c and abc are same words under this
rule. We call this an irreducible word. It is easy to see that the set of all
irreducible words consisting of n elements make a group under the above
product. Note that an empty word is a unit.
Deﬁnition 2.4. The set of all irreducible words consisting of distinct 2n
elements
a
1
, a
2
, . . . , a
n
, a
−1
1
, a
−1
2
, . . . , a
−1
n
with a group structure under the above product is called a free group with
rank n, denoted by F
n
or F(a
1
, . . . , a
n
). Then, a
1
, . . . , a
n
are generators
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
61
of F
n
. For simplicity, for any x ∈ F
n
we employ the notation on k times
product of x in the following; x
k
= x· · · x
k
.
Let G and H be groups. We call a word again an element which con
sists of a ﬁnite number of arbitrarily juxtaposed elements of G and H. For
example,
g
1
h
1
g
2
g
3
h
2
h
3
h
4
(g
i
∈ G, h
i
∈ H)
is a word consisting of 7 elements. Then we introduce a rule such that if
there are adjacent elements which belong to a same group, then we replace
them by their products. For example, in the above word,
g
1
h
1
g
2
g
3
h
2
h
3
h
4
=⇒ g
1
h
1
(g
2
◦ g
3
)(h
2
◦ h
3
◦ h
4
),
where we regard the latter as the word consisting of 4 elements. Under this
rule, it is easy to see that the set of all words consisting of G and H forms
a group. We denote this group by G ∗ H or H ∗ G, which is called the free
product of G and H.
Let G be a ﬁnitely generated group with n generators g
1
, . . . , g
n
. Then
we consider a free group F
n
= F(a
1
, . . . , a
n
) with rank n and deﬁne a map
f : F
n
→ G, f(a
i
) = g
i
such that f is a homomorphism, eg. f(a
1
a
−1
2
a
3
) =
g
1
g
−1
2
g
3
. Clearly, f is a surjection, i.e., f(F
n
) = G. Then, by Exercise 2.1
(3), we have
F
n
/Ker(f)
∼
= Im(f) = G,
which means that any ﬁnitely generated group G is isomorphic to a quotient
group of a free group.
Suppose that the subgroup Ker(f) is generated by h
1
, . . . , h
r
, denoted
by Ker(f) = (h
1
, . . . , h
r
). Thus we have the isomorphic correspondence
F(a
1
, . . . , a
n
)/(h
1
, . . . , h
r
)
∼
= G, (1)
which is called a presentation of G and a
1
, . . . , a
n
are called generators, and
h
1
, . . . , h
r
relations. We usually express this like the following:
G = a
1
, . . . , a
n
 h
1
= e, . . . , h
r
= e .
Note that the way of the presentation of a given G is not unique in general.
As for a free group, we have
F
n
= a
1
, . . . , a
n

since it has no relations, but we usually write it in the following: F
n
∼
=
a
1
, . . . , a
n
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
62
Example 2.2. Let us ﬁnd the presentation of Z
3
= {[0], [1], [2]}, which
is a ﬁnite additive group of order 3. Let F(a) denote a free group with
rank 1 which is canonically isomorphic to Z. Then we deﬁne a surjective
homomorphism f : F(a) → Z
3
, f(a) = [1]. Thus we can verify that
f(a
2
) = f(a) + f(a) = [1] + [1] = [2]
f(a
3
) = f(a) + f(a) + f(a) = [1] + [1] + [1] = [0]
f(a
4
) = [4] = [1], etc.
Hence we have Ker(f) = {a
3m
; m ∈ Z}, which is generated by a
3
. Therefore
we have a presentation Z
3
∼
= F(a)/(a
3
)
∼
=
a a
3
= e
.
In general, for any natural number n we have a presentation
Z
n
∼
= a a
n
= e .
Example 2.3. F(a, b)
∼
= Z ∗ Z. In general, for G = F(a
1
, . . . , a
m
) and
H = F(b
1
, . . . , b
n
) we have
G ∗ H = F(a
1
, . . . , a
m
, b
1
, . . . , b
n
).
Exercise 2.2. Show that the following representation is true;
Z ⊕Z
∼
= a, b aba
−1
b
−1
= e.
Here let us recall the deﬁnition of a symmetric group which is closely
related to a braid group. Let X
n
= {1, 2, . . . , n} be a ﬁnite set consisting of
n elements. Then we deﬁne a set
S
n
= {σ : X
n
→ X
n
, bijection},
which is a group with a product deﬁned by composing maps. S
n
is called
an nth symmetric group, whose order is n!. Any element σ ∈ S
n
is often
written as follows:
σ =
1 2 · · · n
σ(1) σ(2) · · · σ(n)
,
and then, for τ, σ ∈ S
n
, the product is
τ ◦ σ =
1 2 · · · n
τ(σ(1)) τ(σ(2)) · · · τ(σ(n))
.
Also, e =
1 2 · · · n
1 2 · · · n
is a unit element of S
n
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
63
For example, we immediately see that S
2
= {e, σ}
∼
= a a
2
= e
∼
= Z
2
if we deﬁne f : F(a) → S
2
by f(a) = σ, where σ =
1 2
2 1
and note that
σ ◦ σ = e. Moreover, we have S
3
= {e, σ
1
, σ
2
, σ
3
, τ
1
, τ
2
}, where
σ
1
=
1 2 3
1 3 2
, σ
2
=
1 2 3
3 2 1
, σ
3
=
1 2 3
2 1 3
, τ
1
=
1 2 3
2 3 1
, τ
2
=
1 2 3
3 1 2
.
It is easy to check that
σ
1
◦ σ
2
= τ
1
, σ
2
◦ σ
1
= τ
2
, σ
1
◦ σ
2
◦ σ
1
= σ
2
◦ σ
1
◦ σ
2
= σ
3
. (2)
Note also that σ
1
◦σ
1
= σ
2
◦σ
2
= e. Thus we have seen that S
3
is generated
by σ
1
and σ
2
, and it is not apparently abelian.
Let us ﬁnd a presentation of S
3
. We deﬁne a surjective homomorphism
f : F(a, b) →S
3
as follows:
f(a) = σ
1
, f(b) = σ
2
.
Then, we see that f(a
2
) = σ
1
◦ σ
1
= e, f(b
2
) = σ
2
◦ σ
2
= e and f((ab)
3
) =
(σ
1
◦ σ
2
)
3
= (τ
1
)
3
= e. Thus f induces a surjective homomorphism
˜
f : F(a, b)/(a
2
, b
2
, (ab)
3
) →S
3
Hence in order to prove that
˜
f is an isomorphism, it suﬃces to show that
F(a, b)/(a
2
, b
2
, (ab)
3
) has order 6, which means that
˜
f is injective since S
3
also has order 6. We will leave the details to the reader as an easy exercise.
Thus, we have a presentation
S
3
∼
=
a, b  a
2
= e, b
2
= e, (ab)
3
= e
.
Here we shall put s
1
= a, s
2
= bab. Then, s
2
1
= a
2
= e, s
2
= babbab = e,
and s
1
s
2
s
1
= ababa = (ab)
3
b
−1
= b, s
2
s
1
s
2
= bababab = b(ab)
3
= b. Thus
we have another presentation
S
3
∼
=
s
1
, s
2
 s
2
1
= e, s
2
2
= e, s
1
s
2
s
1
= s
2
s
1
s
2
.
In general we have
S
n
=
s
1
, . . . , s
n−1
s
2
i
= 1 (i = 1, . . . , n − 1)
s
i
s
i+1
s
i
= s
i+1
s
i
s
i+1
(i = 1, . . . , n − 2)
s
i
s
j
= s
j
s
i
(i − j ≥ 2)
,
which is not abelian for n ≥ 3 since s
i
s
i+1
= s
i+1
s
i
.
Exercise 2.3. Consider a set G = {E, A, B, C, D, F}, where
E =
1 0 0
0 1 0
0 0 1
, A =
1 0 0
0 0 1
0 1 0
, B =
0 1 0
1 0 0
0 0 1
,
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
64
C =
0 0 1
0 1 0
1 0 0
, D =
0 0 1
1 0 0
0 1 0
, F =
0 1 0
0 0 1
1 0 0
.
(1) Show that G is a group by matrix multiplication.
(2) Show that A
2
= B
2
= C
2
= E, F
2
= D, AD = C, F
3
= E and hence
G = {A, A
2
, F, F
2
, AF, AF
2
}.
(3) Show that G is isomorphic to S
3
.
3. Knots Deﬁned by Braids
Let K be a disjoint union of circles, i.e., K = S
1
∪ · · · ∪ S
1
, where S
1
is a
circle. An embedding of K into the 3dimensional Euclidean space R
3
(or
the 3sphere S
3
), ϕ : K → R
3
, is called a link; in particular, ϕ is a knot
if K is connected, i.e., K = S
1
. Given two embeddings ϕ
1
, ϕ
2
: K → R
3
,
we say that ϕ
1
is ambiently isotopic to ϕ
2
if there exist homeomorphisms
h : K → K and H : R
3
→ R
3
such that H ◦ ϕ
1
= ϕ
2
◦ h, i.e., there is the
following commutative diagram:
K
h
−−−−−−→ K
ϕ
1
ϕ
2
R
3
−−−−−−→
H
R
3
Let L be the set of all links. Suppose that a map λ : L → Λ is deﬁned for
an appropriately calculable algebra Λ. Such a map λ that if K
1
is ambiently
isotopic to K
2
, then λ(K
1
) = λ(K
2
) for any K
1
, K
2
∈ L is called an invari
ant of links. Many invariants have been discovered and we will introduce
one of them, the Jones polynomial, later on. The ultimate purpose in knot
theory is to ﬁnd the complete invariant in which the converse implication
holds. Unfortunately, it has not been discovered yet.
Now we recall the deﬁnition of braid group which plays an important
role in knot theory. First we take 2n points in a cube C = [0, 1]×[0, 1]×[0, 1]
which is a subset of R
3
in the following:
P
1
=
1
2
,
1
n + 1
, 1
, . . . , P
n
=
1
2
,
n
n + 1
, 1
P
1
=
1
2
,
1
n + 1
, 0
, . . . , P
n
=
1
2
,
n
n + 1
, 0
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
65
Then we join the P
1
, P
2
, . . . , P
n
to P
1
, P
2
, . . . , P
n
by n arcs in C in such
a way that these n arcs do not mutually intersect each other and each arc
is not selfknotted. We say that these n curves in C are an nbraid. If we
parametrize the arc γ by (x(t), y(t), t), then γ intersects with z = t plane
at one and only one point for any t ∈ [0, 1]. Given two nbraids in C, if we
can transform one to the other by performing the elementary knot moves
on these arcs, then we say that these two nbraids are equal.
Let B
n
be the set of all nbraids. For σ, τ ∈ B
n
we deﬁne a product of
σ and τ as follows: First glue the base of the cube containing σ to the top
face of the cube containing τ. As a resultant, we have a rectangular solid in
which there are narcs composed by the vertical juxtaposition of σ and τ.
By shrinking the rectangular solid in half we can obtain an nbraid which is
the product of σ and τ, denoted by στ. Note that in general στ is not equal
to τσ, i.e., στ = τσ. We can make B
n
a group under this product and call
it the nth barid group. Let e ∈ B
n
be an nbraid obtained by joining the
P
1
, P
2
, . . . , P
n
to P
1
, P
2
, . . . , P
n
respectively by n segments (see Fig. 1),
which is called a trivial nbraid and a unit element in B
n
.
unit element of Bn
P1 P3 Pn P2 Pn1
P1 P3 Pn P2 Pn1
´ ´ ´ ´ ´
Fig. 1.
The braid group B
n
can be represented by generators and deﬁning re
lations between the generators. In order to give a presentation we shall
introduce two speciﬁc nbraids, say σ
1
, σ
−1
i
, which connect P
i
to P
i+1
and
P
i+1
to P
i
, and then connect the remaining P
j
to P
j
for j = i, i + 1 by
segments:
It is shown that any nbraid can be represented as the product of the
elements σ
±
i
for i = 1, 2, . . . , n − 1. Thus the elements σ
1
, . . . , σ
n−1
are
generators of B
n
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
66
P1 P2 Pi Pi+1 Pn
P´1 P´2 P´i P´i+1 P´n
P1 P2 Pi Pi+1 Pn
P´1 P´2 P´i P´i+1 P´n
1
Fig. 2.
For example, B
2
is generated by σ
1
or σ
−1
1
. Any element α ∈ B
2
is
uniquely determined by the number of σ
1
’s minus the number of σ
−1
1
’s in
α. Thus B
2
= σ
1
, which is isomorphic to a free group Z. Moreover, we
have
B
3
= σ
1
, σ
2
 σ
1
σ
2
σ
1
= σ
2
σ
1
σ
2
B
4
= σ
1
, σ
2
, σ
3
σ
1
σ
3
= σ
3
σ
1
, σ
1
σ
2
σ
1
= σ
2
σ
1
σ
2
, σ
3
σ
2
σ
3
= σ
2
σ
3
σ
2
.
In general, we have the following presentation:
B
n
=
σ
1
, . . . , σ
n−1
σ
i
σ
i+1
σ
i
= σ
i+1
σ
i
σ
i+1
(i = 1, . . . , n − 2)
σ
i
σ
j
= σ
j
σ
i
(i −j ≥ 2)
,
which is due to Artin (see eg. p. 51 in PrasolovSossinsky
8
). Thus we have
a (natural) surjective homomorphism Ψ : B
n
→ S
n
deﬁned by Ψ(σ
i
) = s
i
for i = 1, 2, . . . , n − 1, which just corresponds to forgetting the (under or
over) crossing in σ
i
for any i.
The braid group naturally appears in geometrical or topological context.
For example, if we set S = {(x, y, z) ∈ R
3
; x
2
+ y
2
= 1, z = 0} which is a
trivial knot in R
3
, then we have π
1
(R
3
− S)
∼
= Z (
∼
= B
2
). If K represents
a trefoil knot in R
3
(see Fig. 3), then we have π
1
(R
3
− K)
∼
= B
3
. (See
Example 1.24 in Hatcher,
5
note that a trefoil knot is a (2,3)torus knot.)
Let σ be any nbraid in B
n
. By means of a set of parallel n arcs which
lie outside the cube C, we can connect the n points P
1
, . . . , P
n
on the top
face of σ to the n points P
1
, . . . , P
n
respectively on the bottom face. Then
we have obtained a regular diagram of a knot or link from the nbraid σ.
A knot or a link constructed in this way is said to be a closed braid (or a
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
67
trefoil K
Fig. 3.
closure of σ). For example, we have a trefoil knot as a closure of 3braid σ
3
1
and a ﬁgure eight knot as that of 4braid σ
1
σ
−1
2
σ
1
σ
−1
2
.
The usefulness of this deﬁnition is based on the following fact due to
Alexander:
Theorem 3.1 (the Alexander theorem). Any knot or link can be ob
tained as a closed braid.
As for the proof, see eg. p. 55 in PrasolovSossinsky.
8
Note that we may
have such a situation that a knot or link is obtained by distinct braids. For
example, we have a trivial knot as the closure of 2braid σ
1
, the 3braid
σ
−1
1
σ
2
and the 4braid σ
1
σ
2
σ
−1
3
.
We deﬁne two operations in the union of braid groups which are called
Markov moves:
(1) (Markov move I) For b ∈ B
n
, transform b into xbx
−1
for some x ∈
B
n
.
(2) (Markov move II) For b ∈ B
n
, transform b into either of two (n+1)
braids bσ
n
or bσ
−1
in B
n+1
.
Deﬁnition 3.1. Let a, b ∈ B =
n≥1
B
n
. If we can transform a into b by a
ﬁnite number of performing the Markov moves I, II deﬁned above and their
inverses, then a is said to be Mequivalent to b, denoted by a
M
∼ b.
Theorem 3.2 (the Markov theorem). Let K
1
and K
2
be knots (or
links) obtained as the closure of a
1
and a
2
respectively in B. Then, K
1
is ambiently isotopic to K
2
if and only if a
1
M
∼ a
2
.
See the proof in Birman.
2
Let L be the set of ambient isotopy classes of
oriented links. Theorem 3.2 claims that a map Ψ : B/
M
∼ → L is a bijection,
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
68
which enables us to classify all knots or links by ﬁnding algebraic operations
which should be invariant under the Markov moves I and II.
Next we shall discuss the representation of a group. Let V be a vector
space over a ﬁeld. Let GL(V ) be the set of all linear maps of V into V
which are regular. GL(V ) is a group with a product as a composite of
linear maps, i.e., f · g = f ◦ g ∈ GL(V ) for f, g ∈ GL(V ). For a group
G a homomorphism ϕ : G → GL(V ) is called a (linear) representation of
G, V a representation space and the dimension of V the dimension of the
representation.
Example 3.1. Let us ﬁnd a one dimensional representation of B
n
. Note
that GL(C)
∼
= C−{0} and then the relation σ
i
σ
j
= σ
j
σ
i
is trivially satisﬁed
because GL(C) is an abelian group. Also, from the relation σ
i
σ
i+1
σ
i
=
σ
i+1
σ
i
σ
i+1
we can deduce that σ
1
= · · · = σ
n−1
. Hence, for any z ∈ GL(C),
the correspondence f(σ
i
) = z (i = 1, . . . , n − 1) gives a representation of
B
n
. For ∀b ∈ B
n
we deﬁne ε(b) to be the number of σ
i
’s in b minus that of
σ
−1
i
’s in b for any i, which is called a crossing number of b ∈ B
n
. Thus we
have obtained a one dimensional representation f
z
: B
n
→ GL(C) deﬁned
by f
z
(b) = z
ε(b)
. If we restrict to z ∈ U(1) = {z ∈ C; z = 1}, we have a
one dimensional unitary representation f
z
: B
n
→ U(1).
Let V = C
2
and then we consider the tensor space V ⊗ V . Take an
element
T =
1 0 0 0
0 0 −q 0
0 −q 1 − q
2
0
0 0 0 1
∈ GL(V ⊗ V ),
where q ∈ C − {0}. Moreover, we consider its n times tensor space V
⊗n
=
V ⊗ · · · ⊗ V
n
and set
T
i
= I ⊗ · · · ⊗ I
i−1
⊗ T ⊗ I ⊗ · · · ⊗ I
n−i−1
∈ GL(V
⊗n
),
where I stands for the identity map of V . Then it is easy to verify that
T
1
, T
2
, . . . , T
n−1
satisfy the relations in B
n
;
T
i
T
j
= T
j
T
i
, T
i
T
i+1
T
i
= T
i+1
T
i
T
i+1
.
So, we deﬁne ρ
n
: B
n
→ GL(V
⊗n
) by ρ
n
(σ
i
) = T
i
for each generator
σ
i
∈ B
n
and then we easily see that ρ
n
is a representation of B
n
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
69
Suppose that a representation of the nth braid group f
n
: B
n
→
GL(V
n
) is given for any n. This induces a map f :
n≥1
B
n
→
n≥1
GL(V
n
).
For any b ∈ B =
n≥1
B
n
consider its trace Tr(f(b)) of a matrix f(b). Since
f is a homomorphism, we have
Tr(f(xbx
−1
)) = Tr(f(x)f(b)f(x
−1
)) = Tr(f(x)f(b)f(x)
−1
) = Tr(f(b)),
which means that the correspondence f(b) → Tr(f(b)) is invariant under
the Markov move I. With a little more correction, we obtain a correspon
dence which is invariant both under the Markov moves I and II. For any
link K we take an nbraid b. Then we deﬁne
P
K
(q) = q
ε(b)
Tr(ρ
n
(b)h
⊗n
),
where ε(b) is the crossing number already deﬁned in Example 3.1, h =
q
−1
0
0 q
and h
⊗n
= h ⊗· · · ⊗h
n
. We can check that P
K
(q) is a Laurant
polynomial with a variable q and invariant under the Markov moves I and
II. P
K
(q) is called the Jones polynomial of K (see Jones
6
).
4. Topological Quantum Computing
Roughly speaking, quantum computing is a unitary transformation in the
process of computing (Nakahara
7
). Recalling that a representation of a
group is a homomorphism into general linear space, i.e., a linear transfor
mation will be assigned for an element of a group, we may regard that if the
representation will be realized as a unitary transformation, then knots or
links express its process of computation in quantum computing by replac
ing a unitary transformation in quantum computing with a representation
of the braid group B
n
.
Recently, FreedmanKitaevLarsenWang
4
introduced the notion of a
‘modular functor’, which is an extension of a representation of the braid
group. They have shown that a unitary modular functor is equivalent to
quantum computation. More precisely, given a problem in BQP, a linear
map which corresponds to a selfhomeomorphism of a disk with holes can
be approximated by a sequence of quantum gates for an arbitrary unitary
modular functor. Conversely, any quantum gate can be approximated by
a sequence of selfhomeomorphisms. Here, BQP is the class of decision
problems which can be solved with probability being greater than or equal
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
70
to 0.75 by an exact quantum circuit designed by a classical algorithm in
poly(L), where L is the length of the problem instance M.
We shall give a deﬁnition of a modular functor. Let X be a compact
oriented surface with boundaries. Suppose that each boundary component
of X has a label in a ﬁnite set L = {1, a, b, c, . . .} with involution ˆ , where
ˆ
1 = 1. We call ˆ a a dual of a. We denote a label set of X for all boundary
components by . Then, let C = (X, ) be a category and C
be another
category such that an object is a ﬁnite dimensional complex Hilbert space
and a morphism is a unitary transformation.
A functor F from C to C
is called a unitary topological modular functor
if F satisﬁes the following seven axioms:
(1) (Disjoint union axiom): Let X
1
and X
2
be compact oriented surfaces
with boundaries whose labels are
1
and
2
respectively.
F(X
1
X
2
,
1
2
) = F(X
1
,
1
) ⊗ F(X
2
,
2
).
(2) (Gluing axiom): Let
˜
X be a compact oriented surface obtained by
gluing together a pair of boundary circles with dual labels in X.
F(
˜
X, ) =
x∈L
F(X, ∪ {x, ˆ x}).
(3) (Duality axiom): Let X
∗
stnds for X with the reversed orientation
and labels applied ˆ .
F(X
∗
) = (F(X))
∗
, where we mean that A
∗
=
t
¯
A.
(4) (Empty surface axiom): F(∅) = C.
(5) (Disk axiom): Let D be a 2disk.
F(D, {a}) =
C if a = 1
0 if a = 1
(6) (Annulus axiom): Let A be an annulus.
F(A, {a, b}) =
C if a =
ˆ
b
0 if a =
ˆ
b
(7) (Algebraic axiom): All can be algebraically described over Q for some
bases in F(D, {a}), F(A, {a, ˆ a}) and F(P, {a, b, c}), where P is a 2
sphere with 3 holes.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
71
FreedmanLarsenWang
3
has actually given an example of a unitary
topological modular functor by using the SU(2) ChernSimonsWitten the
ory at q = e
2πi
5
. Note that we have no unitary topological modular functor
deﬁned on X with same labels all in the boundaries by the algebraic axiom
(7). We also see that a modular functor (see Chapter V in Turaev
9
) deﬁned
on X having only label 1’s all in the boundary components is nothing but
a representation of the braid group.
In physics, there are traditionally laws but no axioms. This is because
an old theory has been overtaken by a new theory through clarifying the
limit of applicability by a new experimental fact or its selfcontradiction.
However, it may be useful to clarify the axioms since we have been recently
faced with diﬃculty of experimental inspection. On one hand, in mathe
matics axiomatization is the standard method.
It will be helpful to recall an axiomatic deﬁnition of topological quan
tum ﬁeld theory (TQFT) in relation with modular functor. The TQFT in
dimension d for a ddimensional oriented compact smooth manifold Σ is
a functor which assigns to a ﬁnite dimensional complex vector space Z(Σ)
and a vector Z(Y ) ∈ Z(Σ) such that ∂Y = Σ. Atiyah
1
has given the axioms
in the following:
(1) (Duality axiom): Z(−Σ) = Z(Σ)
∗
, where Z(Σ)
∗
is the dual space of
Z(Σ).
(2) (Disjoint union axiom): Z(Σ
1
Σ
2
) = Z(Σ
1
) ⊗Z(Σ
2
).
(3) (Associativity axiom): When ∂Y
1
= −Σ
1
Σ
2
, a linear map Z(Y
1
) :
Z(Σ
1
) → Z(Σ
2
) is deﬁned and Z(Y
2
) : Z(Σ
2
) → Z(Σ
3
) is also deﬁned
when ∂Y
2
= −Σ
2
Σ
3
. Then we can construct Y = Y
1
∪
Σ
2
Y
2
by gluing
Y
1
and Y
2
along Σ
2
.
Z(Y ) = Z(Y
2
) ◦ Z(Y
1
).
(4) (Empty manifold axiom): Z(∅) = C.
(5) (Homotopy invariance axiom): Z(Σ ×[0, 1]) = id
Z(Σ)
.
Note that if Y is a closed (d +1)manifold, then a linear function Z(Y )
ia a complex number by Axiom (4), which means that the TQFT has given
a topological invariant of a closed (d + 1)manifold.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
72
Note also that a vector Z(Y ) can be regarded as a linear map. First we
recall the wellknown fact:
V
∗
⊗W
∼
= Hom(V, W) (3)
for two complex vector spaces, where V
∗
= Hom(V, C). By combining the
duality axiom, disjoint union axiom and (3), for an oriented (d +1) dimen
sional manifold Y with ∂Y = (−Σ
1
) Σ
2
we have the following:
Z(Y ) ∈ Z(∂Y ) = Z((−Σ
1
) Σ
2
) = Z(−Σ
1
) ⊗Z(Σ
2
)
= Z(Σ
1
)
∗
⊗Z(Σ
2
)
∼
= Hom(Z(Σ
1
), Z(Σ
2
)).
This isomorphism means that a vector Z(Y ) is naturally regarded as an
element of Hom(Z(Σ
1
), Z(Σ
2
)), i.e., a linear map of Z(Σ
1
) into Z(Σ
2
)).
5. Anyon Model
Anyon is a composite particle which is introduced to explain the phenom
ena of quantum hall eﬀect, a phenomenon which electron systems conﬁned
to a 2dimensional surface show at low temperature and strong magnetic
ﬁeld. The electrons conﬁned to a metal surface is in superconductive phase
under low temperature, and applying strong magnetic ﬁeld vertically, nor
mal conducting phase appears as holes on the surface. As is well known,
the magnetic ﬂux gets driven out in superconductive phase, therefore the
ﬂux penetrates into these holes of normal conducting phase, and the sys
tem exhibits a state where the ﬂux piercing these holes vertically. Around
these holes, the normal conducting electrons (holes) make cyclotron mo
tions. Since these ﬂux penetrating holes on the surface of metal and the
normal conducting electrons are correlated, one may expect to treat these
as composite particles. That means, we look upon these normal conduct
ing phase holes as a composite particle that are draped with charges and
magnetic ﬂuxes as shown in Fig. 4, given the names of Anyons.
When the charged particle rotates around the tubelike magnetic ﬁeld
generated by a solenoid, its wave function ψ deviates by e
iqΦ
ψ, which
is called the AharonovBohm eﬀect. Here, Φ is the total magnetic ﬂux
Φ =
S
BdS penetrating the tube. Similarly, since anyon is a particle wear
ing charge and magnetic ﬂuxes, when an electron rotates around the anyon,
its phase of wave function also deviates by AharonovBohm eﬀect propor
tional to the charge. Also, this deviation does not change according to the
continuous change of closed paths, i.e. it is a topological invariant.
Consider the probability amplitude ψ(x
1
, x
2
) of 2 anyons carrying
charge q and magnetic ﬂux Φ, to be observed at point x
1
and x
2
. As
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
73
B
ψ →e
iqΦ
ψ
B
Fig. 4.
sume that the 2nd anyon (at point x
2
) went around the ﬁrst anyon (at
point x
1
). Then the phase of 2nd anyon’s wave function on its ﬁrst circuit
should deviate by qΦ, but since this 2nd particle is also an anyon, it could
be interpreted that the 1st particle has rotated around the 2nd anyon. So,
the total deviation should be the twice the case an ordinary charge going
around a single anyon. Therefore the eﬀect of interchanging the 1st and
2nd anyon would be described as,
ψ(x
2
, x
1
) = e
iθ
ψ(x
1
, x
2
), θ = qΦ. (4)
The particle which its wave function obeys ψ(x
2
, x
1
) = ±ψ(x
1
, x
2
) are
known as bosons or fermions, but as its name shows, the anyon obeys the
fractional statistics, θ = 2π, π in (4). These anyons which obeys such frac
tional statistics are permitted only on a special 2 dimensional space. When
the motion of anyon particles are described on 2+1 dimensional spacetime,
one could get a spacetime diagram just like braids, but the anyon sys
tem breaks timereversal symmetry, and can memorize the entanglement
of spacetime curves. Which means, one could storage information in this
topological links of spacetime curve. The idea of topological quantum com
puting is to use the anyon system to carryout the quantum computation.
Anyon could be thought as a topological soliton solution of the ﬁeld, and
therefore is stable against perturbation. Also the information weaved into
the link is a topological value and therefore robust. Creating quantum al
gorithms could be achieved by braiding the strands. However, while the
simple abelian anyon as described above could construct braid group rep
resentation, it has only complex 1 dimensional representation by the wave
function’s phase factor deviation (4). A complete quantum calculation could
not be handled by 1 dimensional unitary representation of braid group.
Let aside the existence of nonabelian anyon, shown below is the mathe
matical model of topological quantum computing by employing nonabelian
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
74
anyon.
3,4,10–12
1) the emerging particles have types and labels
2) rules of fusion and splitting
3) braiding rules when two particles are exchanged
4) braiding rules when two particles are recoupled
(1) (Labels): Each particle is labelled by sets: L = {1, a, b, c, . . . }. 1 de
scribes no particles, vacuum, and involutionˆrepresents charge conju
gation. The antiparticle of a is labelled by ˆ a, and for the vacuum
ˆ
1 = 1
is required. The Latin letter a represents irreducible unitary represen
tation, and ˆ a is its dual representation. From the deﬁnition of modular
functor given in the previous section, the compact oriented surface X
corresponds to the metal plane, the boundaries of X to each anyons,
and its label to the types of anyon particles.
(2) (Fusion rule): The particle a and b becoming c by fusion is represented
by
a ×b =
c
N
ab
c
c.
Actually it is a composition of unitary representations, for instance,
consider 2 adjoint representation 8 of SU(3), then the rule deﬁnes that
its composition 8×8 = 1+8+8+10+
ˆ
10+27 should be taken as N
88
8
= 2.
N
ab
c
is a nonnegative integer. The space of all possible processes of
particle a and b to fusion and become a particle c is represented by V
ab
c
and called a fusion space, and could be made into a vector space by
taking an orthonormal basis,
{ab; c, µ, µ = 1, 2, . . . , N
ab
c
}.
b a
µ
c
ab; c, µ =
Fig. 5.
The direct product of these vector spaces,
c
V
ab
c
is called the full
Hilbert space for anyon pair ab. We assume orthogonality and com
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
75
pleteness between these bases,
ab; c
, µ
ab; c, µ = δ
c
c
δ
µ
µ
.
c,µ
ab; c, µab; c, µ = I
ab
.
I
ab
is an identity operator on
c
V
ab
c
. The model which satisﬁes
c
N
ab
c
≥ 2 is called the nonabelian anyon model, while those that
does not is called the abelian anyon.
(3) (Braiding rule): The swap of the positions of two particles a and b
induces an isomorphism mapping R : V
ab
c
→ V
ba
c
. R is deﬁned by the
unitary transformation
Rab; c, µ =
µ
ba; c, µ
(R
ab
c
)
µ
µ
,
and it should be the representation of braid group B
n
to satisfy the
pentagon and hexagon relation (Sec.6), together with the recoupling
map F.
b a
µ
c
R
=
µ
(R
ab
c
)
µ
µ
a b
µ
c
Fig. 6.
(4) (Recoupling rule): There are two natural decomposition for the topo
logical Hilbert space V
abc
d
, which express three anyons a,b and c fusing
to become d. That is,
V
abc
d
e
V
ab
e
⊗V
ec
d
e
V
ae
d
⊗V
bc
e
.
Above relation shows that there are two possibilities to be considered,
the ﬁrst a and b fuses and become e, and e and c to be d, and second
b and c fuses and become e
and then a and e
fuse to be d. Deﬁne the
unitary transformation F, an isomorphism of (5) as,
Fab; e, µ ⊗ec; d, ν =
e
µ
ν
ae
; d, µ
⊗bc; e
, ν
(F
abc
d
)
eµν
e
µ
ν
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
76
F =
e
µ
ν
(F
abc
d
)
eµν
e
µ
ν
c
ν
d
b a
µ
c
ν
d
b a
µ
e
e
Fig. 7.
6. Fibonacci Anyons
In the previous section, we introduced general anyon models. Here, we con
sider more concrete and simple Fibonacci anyon models.
10,12
In the Fi
bonacci model there are only two labels, L = {1, a}, where ˆ a = a. State 1
express vacuum state, and a a state where an anyon particle exist. Further,
we require fusion rules between these states,
a × a = 1 + a, a × 1 = a, 1 × 1 = 1.
Consider topological Hilbert space V
a
1
a
2
···a
n
b
. This represents a fusion space
where particles labeled a
1
, a
2
, . . . , a
n
merges and becomes particle labeled
b.
a
2
a
1
b
a
1
a
2
; b =
Fig. 8.
There is only one way of two a particles fusing to become one vacuum,
V
aa
1
=< aa; 1 >
C
. Writing the fusion space when the numbers of a parti
cles are n as V
aa···a
1
= V
(n)
1
, V
(3)
1
= V
aaa
1
=< aa; a ⊗ aa; 1 >
C
and V
(4)
1
has two possibilities such that could be shown as, therefore,
a a a a
a
a
1
a a a a
1
a
1
Fig. 9.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
77
V
(4)
1
=
aa; a ⊗ aa; a ⊗ aa; 1,
aa; 1 ⊗ 1a; a ⊗ aa; 1
C
,
similarly we obtain,
V
(5)
1
=
aa; a ⊗ aa; a ⊗ aa; a ⊗ aa; 1,
aa; a ⊗ aa; 1 ⊗ 1a; a ⊗ aa; 1,
aa; 1 ⊗ 1a; a ⊗ aa; a ⊗ aa; 1
C
.
Generally, the dimension of V
(n)
1
is given by,
dim(V
(n)
1
) = f
n−2
provided that f
n
is Fibonacci’s number, f
n+2
= f
n+1
+ f
n
, f
0
= 1, f
1
= 1,
which gives this model the name Fibonacci anyons. The recoupling opera
tor F must satisfy the following diagrams as shown in Fig. 10, named as
pentagon relations. Writing down these operations as an action to the base
x y z
w
b
c
v
d
f
e
F
F
F
F
F
x y
z w
x y z w
x
y
z w
x y
z w
b
c
f
v
v
v v
e
d
Fig. 10.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
78
vector gives,
F
xy(zw)
v
◦ F
bzw
v
((xy; b)z; c)w; v
=
de
x(y(zw; d); e); v(F
xyd
v
)
b
e
(F
bzw
v
)
c
d
= F
yzw
yzw
◦ F
x(yz)w
v
◦ F
xyz
c
((xy; b)z; c)w; v
=
def
x(y(zw; d); e); v(F
yzw
e
)
f
d
(F
xfw
v
)
c
e
(F
xyz
c
)
b
f
,
where we used the notation ((xy; b)z; c)w; v = xy; b ⊗ bz; c ⊗ cw; v
etc., so the pentagon equation is ,
(F
xyd
v
)
b
e
(F
bzw
v
)
c
d
=
f
(F
yzw
e
)
f
d
(F
xfw
v
)
c
e
(F
xyz
c
)
b
f
.
Further, the F and braiding operator R should also satisfy the following
hexagon relation as shown in Fig. 11. Similarly representing this operation
b
x y z
c
v
d
F F
F
R
R R
y z x
v
y x z
d
v
x
y
z
v
v
v
b
y x
z
y z
x
c
Fig. 11.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
79
for base vectors gives,
F
yzx
v
◦ R
x(yz)
v
◦ F
xyz
v
(xy; b)z; v
=
cd
y(zx; d); v(F
yzx
v
)
c
d
(R
xc
v
)(F
xyz
v
)
b
c
= R
xz
xz
◦ F
yxz
v
◦ R
xy
xy
(xy; b)z; v
=
d
y(zx; d); v(R
xz
d
)(F
yxz
v
)
b
d
(R
xy
b
),
therefore we derive,
c
(F
yzx
v
)
c
d
(R
xc
v
)(F
xyz
v
)
b
c
= (R
xz
d
)(F
yxz
v
)
b
d
(R
xy
b
),
the hexagon equation. Especially, for the topological Hilbert space V
(4)
1
=
V
aaaa
1
, the maps R = R
aa
a
, R
aa
1
and F = F
aaa
1
, F
1aa
1
has a solution,
R
aa
a
= −e
2πi
5
, R
aa
1
= e
4πi
5
, (F
aaa
1
)
c
d
= δ
c
a
δ
a
d
,
(F
aaa
a
)
c
d
= −τδ
c
a
δ
a
d
+
√
τδ
c
a
δ
1
d
+
√
τδ
c
1
δ
a
d
+ τδ
c
1
δ
1
d
, τ
2
+ τ = 1.
V
(4)
1
C
2
, and using the basis vectors as shown in Fig. 12, the braiding
and recoupling operator, written as a whole R, F : V
(4)
1
→ V
(4)
1
, becomes
R =
e
4πi
5
0
0 −e
2πi
5
.
, F =
τ
√
τ
√
τ −τ
.
a a a a
a
a
1
a a a a
1
a
1
= 0
= 1
Fig. 12.
The generators {σ
1
, σ
2
} of braid group B
3
on three strands can be rep
resented by
σ
1
→ R, σ
2
→ F
−1
RF.
These matrices generate a representation of the B
3
, whose image is dense
in SU(2). Similarly in V
(n)
1
, the image of the representation of the braid
group is dense in SU(f
n−1
), so we can simulate a quantum computing with
Fibonacci anyons by braiding and recoupling.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
80
Appendix A. Fundamental group
A set X (more precisely, (X, d)) is called a metric space if there is a real
valued function d : X ×X → R such that
(1) d(x, y) ≥ 0 for any x, y ∈ X; in particular d(x, y) = 0 if and only if
x = y,
(2) d(x, y) = d(y, x) for any x, y ∈ X, and
(3) d(x, z) ≤ d(x, y) + d(y, z) for any x, y, z ∈ X.
For example, (R
n
, d) is a typical metric space if we deﬁne
d(x, y) =
(x
1
−y
1
)
2
+· · · + (x
n
−y
n
)
2
for x = (x
1
, . . . , x
n
), y = (y
1
, . . . , y
n
) ∈ R
n
. The d is called an Euclidean
metric. We have another one; let C[a, b] be the set of all real valued contin
uous functions deﬁned on the closed interval [a, b]. Then we deﬁne
d(f, g) =
b
a
f(x) −g(x)dx
for any f, g ∈ C[a, b]. It is easy to check that (C[a, b], d) is a metric space.
Let (X, d) be a metric space. We deﬁne a set
B
ε
(a) = {x ∈ X; d(x, a) < ε}
for ε > 0 and a ∈ X, which is called an εneighborhood of a in X. A subset
A is called an open set (of X) if it holds that B
ε
(a) ⊂ A for any a ∈ A and
a wellchosen positive number ε. We say that B is a closed set if X − B is
an open set of X. It is clear that both the empty set ∅ and the ambient
set X are open sets and also closed sets by deﬁnition. Moreover, it is easy
to see that U = {(x, y) ∈ R
2
; 0 < y <
π
2
} is an open set of R
2
; however,
for a continuous map f : R
2
→R
2
deﬁned by f(x, y) = (xcos y, xsiny) the
image f(U) is not an open set of R
2
.
Let X be a set and suppose that there is a family of subsets of X, O,
such that
(1) ∅ ∈ O, X ∈ O,
(2) if O
1
, O
2
∈ O, then O
1
∩ O
2
∈ O, and
(3) for any family, {O
λ
}
λ∈Λ
, of subsets in O,
λ∈Λ
O
λ
∈ O.
Then, X or (X, O) is called a topological space and O ∈ O an open set of
X. Let X
1
and X
2
be topological spaces. We say that a map f : X
1
→ X
2
is continuous if f
−1
(V ) is an open of X
1
for any open set V of X
2
. Note
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
81
that the image of an open set in the source space is not necessarily open in
the target space for a given continuous map, as is already seen above.
Let X be a topological space. For any p, q ∈ X if there is a continuous
map f : [0, 1] → X such that f(0) = p, f(1) = q, then X is called arcwise
connected and such a map is called a path in X. In what follows, we always
consider an arcwise connected topological space and so we refer it simply
“a space” hereafter.
Let X be a space, and for p, q, r ∈ X let f be a path connecting p with
q and g a path connecting q with r. Then we deﬁne a product of paths,
f · g, as follows:
f · g(s) =
f(2s) (0 ≤ s ≤ 1/2)
g(2s −1) (1/2 ≤ s ≤ 1).
The f · g is a new path in X connecting p with r.
When there is a path l such that the starting point and the end point
coincide, i.e., a continuous map
l : [0, 1] →X, l(0) = p, l(1) = p
is called a loop in X with a base point p.
Deﬁnition A.1. Let X be a space. For two loops l
0
, l
1
: [0, 1] →X with a
basepoint p, there is a continuous map F : [0, 1] ×[0, 1] →X which satisﬁes
the following conditions:
F(t, 0) = l
0
(t), F(t, 1) = l
1
(t) (∀t ∈ [0, 1]),
then F is called a homotopy connecting l
0
with l
1
and we say that l
0
is
homotopic to l
1
.
Let L(X, p) be the set of all loops with a basepoint p in X. Here we
deﬁne l
0
∼ l
1
if l
0
is homotopic to l
1
. Then it is easy to see that
(1) l
0
∼ l
0
(2) l
0
∼ l
1
=⇒ l
1
∼ l
0
(3) l
0
∼ l
1
, l
1
∼ l
2
=⇒ l
0
∼ l
2
.
This means that ∼ is an equivalence relation in L(X, p). Thus we can con
sider a quotient space
π
1
(X, p) := L(X, p)/∼.
For any l ∈ L(X, p), [l] ∈ π
1
(X, p) is an equivalence class to which l belongs,
which is called a homotopy class of l.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
82
Our purpose is to introduce a group structure into this quotient set. If we
have done it and it is computable, we can study its geometrical property
of X through the group structure. For this purpose we have to deﬁne a
product in π
1
(X, p).
First for l, l
∈ L(X, p) we deﬁne a product of l and l
as follows:
l · l
=
l(2t) (0 ≤ t ≤
1
2
)
l
(2t −1) (
1
2
≤ t ≤ 1).
It is easy to see that if for [l], [l
] ∈ π
1
(X, p) we take
˜
l ∈ [l],
˜
l
∈ [l
], then
we have
l · l
∼
˜
l ·
˜
l
.
Thus we reach the following deﬁnition:
[l] ◦ [l
] = [l · l
]
for [l], [l
] ∈ π
1
(X, p), which is welldeﬁned, i.e., does not depend on the
choices of representatives.
For [α], [β], [γ] ∈ π
1
(X, p) we have
(α · β) · γ(t) =
α(4t) (0 ≤ t ≤
1
4
)
β(4t −1) (
1
4
≤ t ≤
1
2
)
γ(2t −1) (
1
2
≤ t ≤ 1)
and
α · (β · γ)(t) =
α(2t) (0 ≤ t ≤
1
2
)
β(4t −2) (
1
2
≤ t ≤
3
4
)
γ(4t −3) (
3
4
≤ t ≤ 1).
Then we can deﬁne a homotopy F : [0, 1] ×[0, 1] →X in the following:
F(t, s) =
α(
4t
s+1
) (0 ≤ t ≤
s+1
4
)
β(4t −s −1) (
s+1
4
≤ t ≤
s+2
4
)
γ(
4t−s−2
2−s
) (
s+2
4
≤ t ≤ 1)
,
which means that ∀t ∈ [0, 1]
F(t, 0) = (α · β) · γ(t), F(t, 1) = α · (β · γ)(t)
and that ∀s ∈ [0, 1]
F(0, s) = (α · β) · γ(0), F(1, s) = α · (β · γ)(1).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
83
Hence we have (α · β) · γ ∼ α · (β · γ) and the associativity
([α] ◦ [β]) ◦ [γ] = [α] ◦ ([β] ◦ [γ]).
Let e : [0, 1] → X be a constant map deﬁned by e(t) = p (∀t ∈ [0, 1]). Then
we easily see that for any [α] ∈ π
1
(X, p)
[α] ◦ [e] = [e] ◦ [α] = [α].
Thus [e] is the unit element of π
1
(X, p). Moreover, for any [α] ∈ π
1
(X, p) if
we deﬁne α
(t) = α(1 −t), then we have
[α] ◦ [α
] = [α
] ◦ [α] = [e].
Thus [α
] is the inverse element of [α], and hence we have veriﬁed that
π
1
(X, p) has a group structure. If we do not need to specify the basepoint
p ∈ X, we may omit p and write π
1
(X) for simplicity because there is an
isomorphism π
1
(X, p) → π
1
(X, q) for any p, q ∈ X, where note that this
isomorphism is not canonical
c
.
For a space X if π
1
(X, p) is trivial, then we say that X is sim
ply connected. For example, an ndimensional Euclidean space R
n
, an n
dimensional disk D
n
= {(x
1
, . . . , x
n
) ∈ R
n
; x
2
1
+ · · · + x
2
n
≤ 1}, R
n
−
{(0, . . . , 0)} with n ≥ 3, and (n −1)dimensional sphere ∂D
n
= S
n−1
with
n ≥ 3 are all simply connected. On one hand, a circle S
1
, R
2
−{(0, . . . , 0)},
and SO(n) with n = 1 are not simply connected.
Let X and Y be spaces and f : X → Y be a continuous map such that
f(x
0
) = y
0
. Take a loop l : [0, 1] → X with a basepoint x
0
and consider the
composite map
f ◦ l : [0, 1]
l
−→ X
f
−→ Y,
which is also a loop in Y with a basepoint y
0
. If a loop l
1
in X is homotopic
to l
2
, then f ◦ l
1
is also homotopic to f ◦ l
2
.
Namely, given a continuous map f : X → Y , we can obtain a map by
considering the correspondence [l] ∈ π
1
(X, x
0
) to [f ◦ l] ∈ π
1
(Y, y
0
). Then
we write this correspondence as follows:
f
∗
: π
1
(X, x
0
) −→ π
1
(Y, y
0
)
∈ ∈
[l] → [f ◦ l]
c
If we take a path connecting p with q, then this isomorphism is deﬁned. Namely, the
isomorpohism does depend on the choice of the path.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
84
We can easily check that f
∗
is a homomorphism. f
∗
is called the induced
homomorphism by f.
Furthermore, let Z be a space and g : Y → Z be a continuous map.
Then, the composite map
X
f
−→Y
g
−→Z
induces the homomorphism
(g ◦ f)
∗
: π
1
(X, x
0
) →π
1
(Z, z
0
)
and we immediately see that (g ◦ f)
∗
= g
∗
◦ f
∗
.
Let f : X →Y be a homeomorphism such that f(x
0
) = y
0
. Then, since
both f and f
−1
are continuous, we have
f
∗
◦ (f
−1
)
∗
= (f ◦ f
−1
)
∗
= (id
Y
)
∗
= id
π
1
(Y )
(f
−1
)
∗
◦ f
∗
= (f
−1
◦ f)
∗
= (id
X
)
∗
= id
π
1
(X)
.
This means that f
∗
is an isomorphism (see Exercise 2.1 (5)). Thus we have
obtained
Theorem A.1. Let X and Y be arcwise connected topological spaces. If X
is homemorphic to Y , then π
1
(X) is isomorphic to π
1
(Y ). In other words,
if π
1
(X) is not isomorphic to π
1
(Y ), then X is not homeomorphic to Y .
If X is not homeomorphic to Y , then by deﬁnition there is no homeo
morphism between them. However, it is extremely diﬃcult to show directly
that no homeomorphism exists. We emphasize here that only the diﬀer
ence of algebraic structures of fundamental groups immediately provides
homeomorphism criterion.
Moreover, the theorem has a slight generalization. Let X and Y be
spaces. If there are continuous maps f : X → Y and g : Y → X such
that g ◦ f is homotopic to id
X
and f ◦ g is homotopic to id
Y
, then we
say that X is homotopy equivalent to Y . As is easily seen, we can replace
“homeomorphic” by “homotopy equivalent” in Theorem A.1.
Further, if X is homotopy equivalent to a point, then we say that X is
contractible. Note that if X and Y are homeomorphic, then they are ho
motopy equivalent but the converse implication does not hold in general,
e.g. both R and R
2
are contractible (and hence homotopy equivalent) but
not homeomorphic. It is easy to show that R is not homeomorphic to R
2
.
Suppose that they are homeomorphic. Then, there is a homeomorphism
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
85
f : R → R
2
. Since f is an injection, there is x ∈ R such that f(x) = (0, 0).
Consider the restricted map f
R−{x}
: R − {x} → R
2
− {(0, 0)}, which is
still a homeomorphism. Then, R−{x} is not connected but R
2
−{(0, 0)} is
connected, which contradicts the fact that the map f
R−{x}
is a homeomor
phism since the connectedness is topological invariant. This can be deduced
from the calculations that π
1
(R − {x}) = {e} and π
1
(R
2
− {(0, 0)})
∼
= Z
(see Exercise 6.1 and Exercise 6.2 below). In general, the invariance of di
mension of an Euclidean space holds, i.e., R
m
is homeomorphic to R
n
if
and only if m = n, which can be proved by using the notion of “homology
group” but we omit it here because we need more pages in order to discuss
homology.
Exercise 6.1. Show that there is an isomorphism π
1
(S
1
, x
0
)
∼
= Z according
to the following orders: First we set S
1
= {(x, y) ∈ R
2
; x
2
+y
2
= 1} and take
a basepoint x
0
= (1, 0). Moreover, consider a continuous map ϕ : R → S
1
deﬁned by ϕ(x) = (cos x, sin x).
(1) Let f : [0, 1] → S
1
be a continuous map such that f(0) = x
0
. Then
there uniquely exists a map
˜
f : [0, 1] → R such that
˜
f(0) = 0, ϕ ◦
˜
f = f.
(2) Deﬁne a map ψ : π
1
(S
1
, x
0
) → Z by ψ([f]) =
˜
f(1). Show that
(a) ψ is a homomorphism.
(b) ψ is a bijection.
Exercise 6.2.
(1) We have a canonical isomorphism
π
1
(X ×Y, (x
0
, y
0
))
∼
= π
1
(X, x
0
) ×π
1
(Y, y
0
).
(2) Show that a map f : R
2
− {(0, 0)} → S
1
×R deﬁned by
f(x, y) =
x
x
2
+y
2
,
y
x
2
+y
2
, log(x
2
+y
2
)
is a diﬀeomorphism.
(3) Show that R
2
is not homotopy equivalent to R
2
−{(0, 0)} by calculating
their fundamental groups.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
86
Van Kampen’s Theorem
We would like to carry out its calculation of fundamental group for a topo
logical space. As is easily expected, a topological space with a simple geo
metrical structure has its fundamental group with a simple group structure.
However, fundamental group can be generally nonabelian by deﬁnition.
Thus we need to know better how to ﬁnd group in order to calculate fun
damental group eﬀectively for a given topological space. In this section we
further introduce Van Kampen’s theorem which is useful in calculation of
fundamental group.
Let X and Y be arcwise connected topological spaces, X ∩ Y be also.
Then, Van Kampen’s theorem tells us a relation among those fundamental
groups of
π
1
(X ∪ Y ), π
1
(X), π
1
(Y ), π
1
(X ∩ Y ).
Theorem A.2 (Van Kampen). (1) For arcwise connected topological
spaces X, Y, X∩Y , we take a base point x
0
∈ X∩Y . If π
1
(X∩Y, x
0
)
∼
=
{e}, then we have
π
1
(X ∪ Y, x
0
)
∼
= π
1
(X, x
0
) ∗ π
1
(Y, x
0
).
(2) Let X be a connected polyhedron and X ∪ D
2
be a space attaching the
2disk along the boundary circle. We take a basepoint x
0
∈ X ∩ D
2
=
∂D
2
= S
1
. Let be a homotopy class in π
1
(X, x
0
) represented by a loop
X ∩ D
2
and () denote the normal subgroup of π
1
(X, x
0
) generated by
. Then we have
π
1
(X ∪ D
2
, x
0
)
∼
= π
1
(X, x
0
)/()
1
2
X
S
D
Fig. A1.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
87
We omit the proof (see e.g. Hatcher
5
). We should emphasize here that
it is more important to master how to use Van Kampen’s theorem than to
understand its proof.
Take two circles and attach at one point x
0
(see the ﬁgure below). This
&%
'$
&%
'$
r
x
0
Fig. A2. A bouquet S
1
∨ S
1
space is called a bouquet of S
1
, denoted by S
1
∨ S
1
.
By using Theorem A.2 (1) we have π
1
(S
1
∨ S
1
, x
0
)
∼
= Z ∗ Z. Let Int D
2
be a space homeomorphic to a set {(x, y) ∈ R
2
; x
2
+ y
2
< 1}. Then, we
see that π
1
(S
1
×S
1
−Int D
2
)
∼
= Z ∗ Z since S
1
×S
1
−Int D
2
is homotopy
equivalent to S
1
∨S
1
. If we take generators α and β in π
1
(S
1
×S
1
−Int D
2
),
then we see that the boundary loop ∂D
2
represents α
−1
β
−1
αβ. Thus, by
Theorem A.2 (2) together with Exercise 2.2, we have
π
1
(S
1
×S
1
)
∼
= Z ∗ Z/(α
−1
β
−1
αβ) = α, β  α
−1
β
−1
αβ = e
∼
= Z ⊕Z,
since S
1
×S
1
= (S
1
×S
1
−Int D
2
) ∪ D
2
.
We have another typical example. Let us calculate the fundamental
group of the real projective plane RP
2
= M ∪ D
2
by using Theorem A.2,
where M is the M¨ oebius band. First note that M is homotopy equivalent
to S
1
, and hence π
1
(M)
∼
= Z. Take a generator α ∈ π
1
(M) represented by
a central circle of the band M (see the dotted line in Figure A3). Then, we
immediately see that the boundary loop ∂D
2
is α
2
. Thus, by Theorem A.2,
we have
π
1
(RP
2
)
∼
= Z/(α
2
) = α α
2
= e
∼
= Z
2
.
We will leave several handy exercises:
Exercise 6.3.
(1) Find the fundamental group of a closed orientable surface Σ
g
with genus
g ≥ 2.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
88
M bius band M
Fig. A3.
(2) Show that the fundamental group of the Klein bottle is
α, β  α
2
β
2
= e
and its abelianization, i.e.,
α, β  α
2
β
2
= e, α
−1
β
−1
αβ = e
is isomorphic to Z ⊕Z
2
.
(3) Show that π
1
(SO(3))
∼
= Z
2
and π
1
(U(2))
∼
= Z.
References
1. M. F. Atiyah, Topological Quantum Field Theories, Publ. Math. IHES 68
(1989) 175–186.
2. J. S. Birman, Braids, links and mapping class groups, Ann. of Math. Studies,
vol. 82 (Princeton Univ. Press, 1974).
3. M. H. Freedman, M. Larsen and Z. Wang, A modular functor which is uni
versal for quantum computation, Comm. Math. Phys. 227 (2002) 605–622.
4. M. H. Freedman, A. Kitaev, M. Larsen and Z. Wang, Topological Quantum
Computation, quantph/0101025.
5. A. Hatcher, Algebraic Topology (Cambridge Univ. Press, 2002).
6. V. F. Jones, A polynomial invariant for knots via von Neumann algeras, Bull.
Amer. Math. Soc. 12 (1985) 103–111.
7. M. Nakahara, Overview on Quantum Computing, Lectures in summer school
at Kinki University, August 2007.
8. V. V. Prasolov and A. B. Sossinsky, Knots, Links, Braids and 3manifolds,
Translations of Mathematical Monographs, vol. 154 (Amer. Math. Soc.,
1997).
9. V. G. Turaev, Quantum invariants of knots and 3manifolds, de Gruyter
Studies in Math., vol. 18 (BerlinNew York, 1994).
10. J. Preskill, Lecture Notes for Physics 219: Quantum Computation,
http://www.iqi.caltech.edu/preskill/ph219.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
89
11. A. Marzuoli and M. Rasetti, Computing spin neworks, Ann. Physics 318
(2005) 345–407.
12. L. H. Kauﬀman and S. J. Lomonaco Jr., qdeformed spin networks, knot
polynomials and anyonic topological quantum computation, J. Knot Theory
Ramiﬁcations 16 (2007) 267–332.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
This page intentionally left blank This page intentionally left blank
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
91
AN INTRODUCTION TO ENTANGLEMENT THEORY
DAMIAN J. H. MARKHAM
Department of Physics, Graduate School of Science, University of Tokyo, Tokyo
1130033, Japan
We introduce the theory of entanglement intended for an audience of physicists,
computer scientists and mathematicians not necessarily having a background
in quantum mechanics. We cover the main concepts of entanglement theory
such as separability, entanglement witnesses, LOCC and entanglement mea
sures. Along the way we will see many interesting questions arise spanning
mathematics, physics and information science amongst other disciplines.
Keywords: Entanglement Theory, Separability, LOCC.
1. Introduction
Entanglement arises naturally from the mathematical formalism of quan
tum mechanics. From a physical perspective this property is however pretty
surprising and has led to many longstanding debates amongst physicists as
to the validity and meaning of this formalism. More recently, entanglement
has become seen as a resource for quantum information processing, useful
for absolutely secure key distribution and quantum computation.
The theory of entanglement developed over the past two decades shows
rich mathematical structure, with many subtleties (both from physics and
a mathematics perspectives) and has drawn ideas and techniques from such
diverse ﬁelds as thermodynamics, convex analysis, information theory and
group theory. Despite the vast progress made over the years there are still
an exciting number of questions remaining in many directions, in its use in
quantum information, in the mathematical challenges it presents and in its
fundamental role as a part of physics.
The intention of this introduction is to take the student through the
main concepts of entanglement theory as clearly as possible. It will not and
cannot be a deﬁnitive guide to the subject (for a pretty good attempt at this
please see the Horodeckis’ review
1
) and of course some of the topics covered
will be inﬂuenced by my own ﬁeld of research to an extent. However I will
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
92
endeavor to cover all the basics and point out interesting open questions
when they arise, relations to other ﬁelds and further reading. I hope that
these notes will pass on some of my own enthusiasm for what I think is an
incredibly exciting ﬁeld to work in.
Mathematically entanglement is simply a consequence of composite sys
tems in quantum mechanics. This will be our starting point in these lectures.
We will begin in section 2 by introducing quantum states, statespace and
how operations on states are represented mathematically. Moving to com
posite systems we will see entanglement emerge as a property of quantum
mechanics in section 3. We will then move to the quantiﬁcation of entangle
ment in section 4. First we explain the general setting of LOCC, then move
to how this is used to deﬁne diﬀerent measures of entanglement. We will
also brieﬂy discuss methods of measuring entanglement and some subtleties
which arise in the multipartite case.
2. Quantum Mechanics and State Space
In order to facilitate a speedy review of the formalism of quantum mechanics
we will simply present the most general way to express states and evolution
of quantum systems, without too much recourse to the physical motives
behind them. We hope that this will also be more amenable to mathemati
cians and computer scientists to get the important points. From a physical
perspective this route may seem a little topsy turvy, but hopefully this is
countered by its simplicity and transparency. For a more detailed account
describing many of the physical motivations for the axioms taken, please
see Ref. 2.
2.1. State space
In quantum mechanics, to all degrees of freedom of physical objects, we
associate a Hilbert state H. For example an electron is described using L
2
for the position degree of freedom, and C
2
for the spin degree of freedom.
A particular system at any point in time is then described by a “state”.
The state describes the physical properties of the system  that is, what
would happen if we were to measure the system. Mathematically, a state is
described by a density matrix ρ living in the operator space over the Hilbert
space H of the system at hand. The set of density matrices are the set of
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
93
all operators satisfying
ρ
†
= ρ
I1 ≥ ρ ≥ 0
Tr[ρ] = 1. (1)
That is, any operator acting over Hilbert space H and satisfying the above
properties can be said to describe the state of a physical system. We refer
to the space of all density operators as state space S(H). It can easily be
checked that state space is convex  i.e. convex combinations of density
matrices are also density matrices.
All density matrices can be diagonalised to their eigenbasis
ρ =
i
p
i
e
i
e
i
, (2)
where the e
i
are the eigenstates (elements of the underlying Hilbert space
H), and the p
i
can be interpreted as probabilities since by deﬁnition we have
1 ≥ p
i
≥ 1 and
i
p
i
= 1. If the sum in (2) only consists of one element
ρ = ψψ, we say it is a pure state (in most introductory undergraduate
courses in quantum mechanics we only consider pure states). In such a case
we often simply write the unique eigenstate ψ. Thus one interpretation of
the density matrix is of a classical mixture, or ensemble, of the pure states
{p
i
, e
i
}. A state which is not pure is called mixed. A simple way to test
if a given density matrix is pure or mixed is the fact that ρ is pure if and
only if ρ
2
= ρ (as can be easily checked).
In fact any density matrix has many possible decompositions into convex
combinations of pure states. Thus, the ensemble interpretation is not unique
 physically this means that many diﬀerent physical situations may give rise
to the same physical state. For example, take two decompositions of some
state
ρ =
i
p
i
ψ
i
ψ
i

=
j
q
j
φ
j
φ
j
, (3)
this could be created by randomly sampling the states {ψ
i
} according to
the probability distribution p
i
, or by randomly sampling the states {φ
j
}
according to q
j
. The freedom in decomposition occurs again and again in
entanglement theory and is the cause of some great diﬃculty when it comes
to checking the existence of entanglement, as well as its quantiﬁcation.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
94
Note that one can in fact take a more abstract mathematical approach
in terms of C* algebras, which can become important when considering
inﬁnities, but for the situation covered in these lectures (and for most of
quantum information) this reduces to the formalism described above. From
here on, unless stated otherwise we are considering ﬁnite dimensional sys
tems, where the above completely suﬃces.
In quantum information we usually do not specify the physical system
at hand. Instead we take a general two level system known as a qubit (and
leave the physics of implementation to someone else!). A general pure qubit
is written
ψ = α0 +β1, (4)
with α
2
+ β
2
= 1. Its density matrix is given by
ρ = ψψ =
α
2
αβ
∗
α
∗
β β
2
, (5)
written in the eigenbasis (i.e. ψ) and the computational basis {0, 1}.
2.2. Evolution
All evolutions of a quantum system can be mathematically expressed as
completely positive (CP) maps from one space of density matrices S
1
to
another S
2
(in fact we will usually consider S
1
= S
2
). Completely positive
maps are extremely important in the theory of entanglement as we will
see in later sections. For now we skip the deﬁnition of complete positivity
until section 2.4, and will instead give the most general way to express such
maps. These fall into two classes of operations, those that do not preserve
the trace 1 property of density matrices (or rather those which require an
additional normalisation element to the formalism), and those that do 
completely positive trace preserving (CPTP) maps.
CP maps not preserving the trace property can be thought of as result
ing from an evolution where at some point a measurement has taken place,
who’s outcome m is known
ε
m
(ρ) =
i
E
m
i
ρE
m†
i
, (6)
such that
i
E
m†
i
E
m
i
≤ 1. The E
m
i
are known as the Kraus operators. The
superscript m can be seen as indicating the classical information of some
measurement outcome in the evolution process. The trace of ε
m
(ρ) may be
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
95
less than one (if
i
E
m†
i
E
m
i
< 1), which is why we say these maps are not
trace preserving. The full normalised version (necessary for the output to
correspond to a state) is given by
ρ →
ε
m
(ρ)
Tr[ε
m
(ρ)]
. (7)
The class of CPTP maps can be understood as having resulted from
an evolution where the results m are erased or forgotten, hence we must
average over all m. Then we have
ρ →ε(ρ) =
m
ε
m
(ρ)
=
m,i
E
m
i
ρE
m†
i
, (8)
with
i,m
E
m†
i
E
m
i
= I1. Note that CPTP maps are obviously a subclass
of CP maps, so all CPTP maps can of course be written in the form of
equation (6), by simple relabelling. So in fact (6) is the most general form
of evolution possible. Hence in some sense index m acts as a dummy index
and, although the interpretation of m as a measurement outcome is very
useful (and often physically accurate), we must be a little bit careful as it
may not always be the case.
When there is only one term in the summation (6), and if the trace
remains 1, the evolution is unitary
a
. This is a clear example of a CPTP
map which does not involve a measurement. But this seems a bit cheating
as there is only one term. An important property of unitary operations is
that it is impossible to go from a pure state to a mixed state, which can
be easily checked. We can see unitary maps simply as a change of basis of
the Hilbert space. In general measurement does cause mixing (by adding
extra terms), but we must be careful to note that it is not the only way to
generate mixing. It is possible to have CPTP mixing with many terms not
via measurement as we will see in later sections  via entanglement.
2.3. POVMs, projective measurement and observables
We have seen above the most general evolution of a quantum state. Let us
now go back and extract the essential features of the measurement part.
(Usually in text books you will ﬁnd these introduced in the reverse order and
in fact the idea of CP maps can be derived from axioms about measurements
a
An operator U is unitary iﬀ U
†
U = I1.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
96
e.g. Ref. 2). Sometimes it is not necessary to know the full evolution of the
system when considering a measurement, say, for state estimation, we may
just need to know the statistics of the outcomes of the measurements (rather
than also knowing the resultant state as well). This leads to the idea of the
formalism of POVMs.
To any measurement on a state ρ we associate a set of operators {M
m
},
collectively known as the Positive Operator Valued Measure (POVM) which
obey
m
M
m
= I1
I1 ≥ M
m
≥ 0 ∀m, (9)
where m represents an outcome of the measurement, and the probability
of outcome m occurring p
m
is given by p
m
= Tr[ρM
m
].
There is no mention of how the state actually evolves, and indeed for
any given POVM, there is no unique way to implement it in general. This
can be seen if we write the POVM in terms of general evolution operators.
The measurement would evolve the system to an ensemble as
ρ → {p
m
,
i
E
m
i
ρE
m†
i
},
M
m
=
i
E
m†
i
E
m
i
. (10)
Clearly in general there are many diﬀerent Kraus operators E
m
i
which would
give the same M
m
, so the actual physical process is not unique in general.
Thus POVMs represent only the statistics of a measurement.
Relating back to the map (6) we see that the ensemble is written
{p
m
,
i
ε
m
(ρ)} and the m represents the outcomes as mentioned. Forget
ting the outcomes would then be the same as taking the ensemble to the
mixed state
m
p
m
ε
m
(ρ).
Any measurement can be written as a POVM, and in turn any set of
operators satisfying properties (9) represents the POVM of a physically
valid measurement.
If we look at the action of a unitary rotation of a state ρ → UρU
†
,
we see that the statistics of any measurement would be changed p
m
→
Tr[UρU
†
M
m
] = Tr[ρU
†
M
m
U] by the cyclic property of the trace operator.
Thus we can consider a unitary rotation of a state ρ as simply a rotation
of the basis in which we measure (this is the essence of the Heisenberg
representation of quantum evolution).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
97
An important class of POVMs that we will come across again are pro
jective measurements. These are measurements whose POVM operators P
m
satisfy
P
2
m
= P
m
. (11)
They can always be expanded as P
m
=
α
αα for some orthogonal basis
α. By comparing with (9)we see that in this case the Kraus operators are
ﬁxed, hence we have both the statistics and the outcome states of the
measurement.
To complete the review of measurements we introduce observables. An
observable is a Hermitian operator, written in the eigenbasis
A =
m
a
m
P
m
, (12)
where the eigenvalues a
m
represent the outcome of some experiment, and
the P
m
are projections onto the eigenspaces. These are the normal represen
tation of measurements in an undergraduate course on quantum mechanics.
Examples of such observables which are common in quantum information
are the Pauli spin operators. The expectation value for a given state ρ (i.e.
the average outcome) is given by
A
ρ
:=
m
p
m
a
m
= Tr[Aρ] (13)
2.4. Composite systems
To combine more than one quantum system we employ the tensor product
formalism. For systems with associated Hilbert spaces H
A
and H
B
, the
states are density matrices over H
A
⊗ H
B
.
For a composite state ρ
A,B
, if we restrict ourselves to one subsystem,
H
A
, the reduced state is given by the
ρ
A
:= Tr
B
[ρ
A,B
], (14)
where Tr
B
is the partial trace and is deﬁned as
Tr
B
[ρ
A,B
] :=
α
α
B
ρ
A,B
α
B
, (15)
where {α
B
} form an orthonormal basis over H
B
. This deﬁnition of reduced
density matrix is equivalent to deﬁning it as the operator ρ
A
which satisﬁes
Tr[ρ
A
X] = Tr[ρ
A,B
X ⊗ I1] for all X (see e.g. Ref. 3). Operationally this
basically says that it is the only way to represent a reduced state such that
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
98
measurements on one system faithfully can be seen as measurements on
half of the global system.
As a side note, with the introduction of composite systems, a simple and
elegant picture of general quantum operations emerges. It can be shown
(see e.g. Ref. 2) that all quantum evolutions, which take the form (7) can
be regarded as combinations of adding an ancillary system, performing
unitary operations, measuring of the ancilla system and tracing over the
ancilla system. We will sometimes ﬁnd this picture of evolution helpful for
descriptive reasons, but mathematically we can always use (6,8).
We are now in a position to deﬁne complete positivity. A positive map
Θ is completely positive iﬀ
Θ⊗I1 ≥ 0 (16)
for any ﬁnite dimensional identity operator I1. That is, if and only if the
positivity remains with arbitrary extension. This concept is massively im
portant in entanglement theory as we will see in later sections. An important
example of an operator that is positive, but not completely positive is the
partial transpose which we will meet in section 3.
Finally we now come to entanglement. Mathematically it arises from
the observation that not all states on a composite of N systems H = H
A
⊗
H
B
⊗... ⊗H
N
can be written in the form
ρ =
i
p
i
ρ
i
A
⊗ρ
i
B
⊗... ⊗ρ
i
N
. (17)
States that can be written in this form are called separable, and states that
cannot be written in this form are called entangled.
3. Entanglement and Separability
Now we have a deﬁnition of what entanglement is, let us look at the simple
example of a bipartite state, at ﬁrst considering pure states. This will allow
us to introduce many terms in a simple way. Following our description of
quantum states for the pure state case, the most general pure state in a
system composed of two systems A and B is written as ψ
AB
∈ H
A
⊗H
B
,
ψ
AB
=
i,j
a
i,j
i
A
⊗j
B
≡
i,j
a
i,j
ij
AB
, (18)
where {i
A
} and {j
B
} are orthonormal bases for H
A
and H
B
respectively,
the tensor product of which is often shortened to i
A
⊗j
B
≡ ij
AB
, and
a
i,j
are complex coeﬃcients satisfying
i,j
a
i,j
= 1.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
99
The deﬁnition of separability for pure states reduces to the set of states
which can be written in the form
ψ
AB
= a
A
⊗b
B
, (19)
for some states a ∈ H
A
and b ∈ H
B
. Separable pure states are often
referred to as product states.
An extreme example of an entangled pure state in C
2
⊗ C
2
is the fol
lowing
Φ
+
AB
:=
1
√
2
(00
AB
+11
AB
). (20)
It can easily be seen that there is no way of decomposing this state into a
product state as in (19) by contradiction. This state is a maximally entan
gled state of C
2
⊗C
2
b
.
If we look at the reduced state of either subsystem, we see how entan
glement implies some local mixedness. The reduced state density matrices
are
ρ
A
= ρ
B
=
1/2 0
0 1/2
. (21)
This is already in its eigenbasis, so it is clearly not a pure state. This is
the maximally mixed state of C
2
⊗ C
2
c
. In fact, the amount of mixedness
is a good measure of how entangled a pure state is (as we will see in later
sections).
In the bipartite pure state case there is a very convenient way to write
states (as far as their entanglement is concerned). For all states ψ
AB
∈
H
A
⊗ H
B
, there exist local basis i
A
∈ H
A
and i
B
∈ H
B
, such that the
state can be written
ψ
AB
=
r
i=1
√
p
i
ii
AB
. (22)
This is known as the Schmidt decomposition, the coeﬃcients p
i
are the
Schmidt coeﬃcients and the minimum r is known as the Schmidt number
(known to be a good quantiﬁer of entanglement as we will also see in later
sections). This decomposition captures all the features of the entanglement
of the system, so is very important in entanglement theory.
b
A maximally entangled state of C
d
⊗C
d
, d ≤ d
is given by Φ
+
AB
=
d−1
i=0
1
√
d
i, i
AB
.
c
A maximally mixed state of C
d
is given by ρ = I1/d.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
100
From the the symmetry of the Schmidt decomposition, for pure bipartite
states we can see that we have
ρ
A
= ρ
B
. (23)
Also a pure state is entangled iﬀ the Schmidt decomposition has only one
term (i.e. the Schmidt number is equal to one). Thus, for a global pure
bipartite state, the state is entangled iﬀ the reduced states are mixed.
Now let us turn to mixed states. The ﬁrst thing we note is that the
last observation above no longer holds  reduced states can be mixed even
without the presence of entanglement. Indeed, as we will see more in sub
sequent sections, the pure bipartite state case is rather special in the ease
with which we can analyse entanglement theory.
Our task in testing separability is to test if a state can be written in the
form
ρ
AB
=
i
p
i
ρ
i
A
⊗ρ
i
B
. (24)
For the bipartite mixed state case a useful tool can be found looking at the
action of certain maps on the state. In fact the completely positive maps
which occurred earlier. Consider a positive map acting on one part of the
system Λ
B
, which may not be completely positive. Clearly for all separable
states we have
(I1 ⊗Λ
B
(ρ
AB
)) ≥ 0, (25)
since Λ
B
(ρ
i
B
) = ρ
i
B
is a valid density matrix, thus taking any separable state
to another one just priming the right hand part of (24).
In fact it has been shown
4
that a bipartite state ρ
AB
is separable iﬀ
for all positive maps Λ : S(H
B
) → S(H
A
) the condition (I1 ⊗ Λ(ρ
AB
)) ≥ 0
holds.
Thus in the bipartite case, the question of separability becomes one of
checking this condition over all positive maps which are not completely
positive.
d
In the case of C
2
⊗C
2
and C
2
⊗C
3
, the map of partial transposition is such
a strong example of a positive positive, but not completely positive map as
d
Note that positive, but not completely positive maps are not actual physical processes.
The question then arrises: can such tests be carried out physically? Seemingly not. How
ever one approach to this has been to study approximate maps, which while CP, approx
imate well the nonCP maps above. See for example Ref. 5.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
101
to cover all possibilities. For systems residing over C
2
⊗C
2
or C
3
⊗C
2
, all
positive maps Λ : C
d
→C
d
, d = 2, d
= 2, 3, can be decomposed as
Λ = Λ
(1)
CP
+ Λ
(2)
CP
◦ T. (26)
From this it follows that taking transposition over one system is suﬃcient
to check for separability here.
Restated, for state ρ
AB
=
i,j,k,l
a
i,j,k,l
ijkl, the partial transposition
over subsystem H
B
is given by
ρ
T
B
AB
=
i,j,k,l
a
i,l,k,j
ijkl, (27)
and is dependant on the basis ij at hand. Then the above statement is
written: A state ρ
AB
∈ C
2
⊗C
2
or ρ
AB
∈ C
3
⊗C
2
is separable iﬀ its partial
transpose is positive ρ
T
B
AB
> 0.
This is known as the PPT condition for separability, initially presented
by Peres
6
as a necessary condition for separability and later used for the
above theorem. The fact that it is not in general a suﬃcient condition was
shown ﬁrst in Ref. 7 where an example of an entangled PPT state was
presented.
When it comes to higher dimensions and larger number of systems,
the problem of being able to check if a state is separable or not becomes
extremely diﬃcult, and in fact poses one of the largest open problems in
the theory of entanglement.
The task is to ﬁnd out if a given state can be written in a decomposition
in the form of (17). Since, as mentioned, there are in general inﬁnitely
many possible decompositions for a state, this is a diﬃcult task to check
and we should need some clever mathematical tricks. In the bipartite case
this can be seen as the problem of describing positive but not completely
positive maps in a concise useful way (a recognised diﬃcult problem in
mathematics). In the multipartite case it is even more diﬃcult.
Several approaches have been developed which work well for diﬀerent
sets of states. Broadly speaking there are two sets of results, those necessary
conditions (such as PPT) which are implied by separability (but do not
necessarily imply it), and those that are necessary and suﬃcient, which are
generally diﬃcult to use or of limited application. It is beyond the scope of
these lectures to go into details and we refer the reader to the Horodeckis
review
1
for a good overview.
Let us now take a small step back and review what the set of separable
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
102
states looks like. We deﬁne
SEP := {ρ
SEP
∃ decomposition ρ
SEP
=
i
p
i
ρ
i
A
⊗ ρ
i
B
⊗ ... ⊗ ρ
i
N
}. (28)
We can see that this set is also convex  a probabilistic mixture of diﬀerent
separable states is also separable. Also, it can be shown that the volume of
separable states is nonzero around the identity
8
for any ﬁnite dimensional
system.
The convexity of this set actually becomes incredibly useful and al
lows for a method of separating sets of entangled states from the separable
ones, via entanglement witnesses.
4,9
These are operators which deﬁne hy
perplanes separating some set of entangled states from the convex set of
separable states. Not only are these operators interesting from a mathe
matical point of view, they are also Hermitian and so represent physical
observables. Thus their role in entanglement theory spreads to the experi
mental validation of entanglement too.
The basis behind these witnesses is the HahnBanach theorem which
states that given a convex body and a point x
o
that does not belong to it,
then there exists a linear function f that takes positive values for all points
belonging to the convex body, while f(x
0
) < 0 (see Ref. 10 for example).
Stated in terms of density operators we deﬁne an entanglement witness
W as Hermitian operators satisfying
W ≥ 0
∀ρ
SEP
∈ SEP, Tr[Wρ
SEP
] ≥ 0
∃ρ / ∈ SEP s.t. Tr[Wρ] < 0 (29)
The Hahn Banach theorem then states that for all entangled ρ there exists
a witness which can see it.
In fact these witnesses are related to the separation of separable and
entangled states given by the positive operators mentioned above (see e.g.
Ref. 1,10). From a physics perspective, they are also related to Bell’s in
equalities  in fact Bell’s inequalities are examples of entanglement witnesses
which also witness the violation of local realism.
Entanglement witnesses are incredibly important in experimental veri
ﬁcation of entanglement. An important question that arises there is ﬁnding
the optimal decomposition into local observables. That is, in situations con
cerning entanglement we are almost by deﬁnition separate, so we would like
to measure these witnesses locally. We would then like to decompose these
into sets of local measurements in as eﬃcient a way as possible. In certain
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
103
cases optimality can be shown, and this is an important ongoing line of
investigation.
The practical usefulness of entanglement witnesses also extends to the
quantiﬁcation of entanglement. In certain cases the amount of violation of
the entanglement witness (that is, how negative is the outcome) gives a
bound to the ‘amount’ of entanglement present
11–13
(a concept which we
will expand on in the next section).
This simple and elegant approach of entanglement witnesses to the study
of entanglement is especially valuable when considering large complicated
systems such as those in condensed matter,
14,15
where often we cannot
even be sure of the state of the system exactly and typically measurements
accessible are averaged over many subsystems (in the thermodynamic limit
inﬁnitely many). A process which often irons out any quantum eﬀects.
Incredibly it is shown that even such ‘rough’ measurements as the magnetic
susceptibility can act as witnesses of entanglement.
14
4. Quantiﬁcation of Entanglement
In the last section we have seen the deﬁnition of the term entanglement,
and how, given a state we may decide whether it is entangled or not. Now
we would like to quantify how entangled a state is. For example, consider
the state
√
1 −δ00 +
√
δ11. (30)
This state is clearly entangled when δ = 1/2 and separable when δ = 0
or δ = 1. Its symmetry indicates that δ = 1/2 may be somehow the most
entangled, and so what about the in between bits?
Before we can really answer this question we must deﬁne better the
problem setting. Entanglement is a property of quantum states which is
inherently nonlocal, and has meaning when considering systems which are
separated. This idea is made precise in the paradigm of local operations
and classical communication.
4.1. Local operations and classical communication
Quantiﬁcation of entanglement is studied and deﬁned with respect to a spe
cial set of operations encapsulating the idea of what we mean by entangle
ment. These are Local Operations and Classical Communication (LOCC).
The situation is one of separate laboratories. We consider a system
spread over several subsystems, where each subsystem is acted upon sepa
rately. For a system on H
A
⊗H
B
⊗...H
N
, LOCC maps are those comprised
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
104
of combinations of
 Local CPTP maps on the individual subsystems
 Classical Communication
The role of the classical communication is to allow the generation of
classical correlations. The motivation behind this paradigm is essentially to
test which properties of the system persist beyond local quantum eﬀects,
allowing for global “classical” eﬀects  that is correlations generated by
classical communication (for a fuller discussion on the physical motivation
for these conditions, see e.g. Ref. 1,16).
It is apparent that all separable states can be made by LOCC. To see
this we only need to look at the deﬁnition of a separable state (17). Tak
ing the ensemble interpretation, this can be generated for example, by the
experimentalist at system A randomly choosing a letter i according to prob
ability distribution p
i
e
, then sending the choice to all the experimentalists
at the remaining systems, according to which i they all locally generate
the state ρ
i
x
. Then if all experimentalists erase all memory of which i they
had, the overall state is that described by (17). Similarly states which can
be generated by LOCC are separable. Thus we have an interpretation of
separable states as those which can be created by LOCC and entangled
states as those which cannot.
Let us now try to write down the Kraus operators corresponding to an
LOCC operation. We ﬁrst consider the bipartite case. We now slip into the
standard notation of calling one system H
A
, belonging to Alice, and the
other H
B
, to Bob. Without loss of generality we assume that Alice goes
ﬁrst. She may perform some measurement and send the result to Bob, so
then his action would depend on Alice’s outcome. Then, supposing he does
a measurement too, Bob could send Alice his result and she could make an
operation according to that, e.t.c. In this way the Kraus operators would
become correlated, and the map would be of the form
Λ
LOCC
(ρ) =
a
1
,b
1
,a
2
..
...(E
a
1
,b
1
,a
2
A
⊗I1)(I1 ⊗E
a
1
,b
1
B
)(E
a
1
A
⊗I1)ρ
(E
a
1
A
⊗I1)
†
(I1 ⊗E
a
1
,b
1
B
)
†
(E
a
1
,b
1
,a
2
A
⊗I1)
†
... (31)
(similar holds for the multipartite case). Given the possibility of going on
and on with the correlations, the sum can go on for arbitrarily many terms.
e
If this sounds too nonphysics we can always imagine the i as outputs of a quantum
measurement.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
105
Indeed the diﬃculty in writing LOCC operations in a concise way is one of
the most diﬃcult obstacles to overcome in entanglement theory.
For this reason we often consider simpler sets of operations, such as
separable operations (those mapping separable states to separable states),
or PPT operations (those mapping PPT states to PPT states). These op
erations have much simpler mathematical form, and so have allowed many
results to be found where restricting to the stricter set of LOCC operations
is too complicated. For now however we will stick to considering LOCC.
4.2. Entanglement measures
Now we have the situation clearly established (i.e. LOCC), we can begin
to ask what this setting would have to say about quantifying the amount
of entanglement present in some given state ρ. We can roughly split results
into two main approaches to this question; an operational approach, where
entanglement is deﬁned with respect to how useful this state would be
to some quantum information task; and an abstract approach, where we
deﬁne an entanglement measure as some functional on the set of states
which obeys certain axioms.
So far in these lectures we have taken a rather abstract approach to
entanglement theory, but this is really not how the ﬁeld began. People
became interested in entanglement in the early 90s when it was noticed
that it could be useful for certain tasks
f
. This approach led to the ﬁrst
attempts to quantify entanglement. The archetypal state used in quantum
information is the bipartite state which we already know as the maximally
entangled state
Φ
+
=
1
√
2
(00 + 11). (32)
This is the “most” useful state of two qubits in several respects: it (or local
unitary equivalent states) gives the optimum ﬁdelity in teleportation (i.e.
perfect teleportation),
17
it also allows for perfect secrecy and eﬃciency in
quantum key distribution.
18,19
On the other hand taking a state as in (30)
with δ = 1/2 these tasks cannot be done as well. Indeed the ability to do
these tasks depends on the value of δ in a very precise way. This is not a
particular feature of these tasks or this size of system, in general we will see
f
Although of course entanglement has been an issue in physics since at least 1935, en
tanglement theory proper started only then. A kind of market driven start to a research
ﬁeld.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
106
that the entanglement properties are well parameterised by the Schmidt
coeﬃcients of bipartite pure states.
With this in mind, the idea of the usefulness of a quantum state was de
veloped viewing the maximally entangled (ME) state as an ideal resource
state. The question asked was, given several copies of a quantum state
shared between Alice and Bob, how many ME pairs can they get by LOCC,
and conversely, how many ME pairs does it take to generate a given state.
Using well developed ideas of asymptotics from information theory the en
tanglement quantities of distillable entanglement E
D
and entanglement cost
E
C
were presented respectively.
20
More precisely, we deﬁne the distillable entanglement as the optimum
rate for distilling maximally entangled states in the limit of inﬁnite number
of copies, optimized over all possible LOCC protocols.
E
D
:= sup{r : lim
n→∞
( inf
Λ
LOCC
Λ
LOCC
(ρ
⊗n
) −Φ
+
2
rn
) = 0}, (33)
where Φ
+
2
rn is the ME state on C
d
⊗C
d
with dimension d = 2
rn
(which is
equivalent to rn ME pairs of qubits), . is the trace norm (see e.g. Ref. 21)
and Λ
LOCC
is an LOCC map. Similarly the entanglement cost is deﬁned as
the asymptotic rate going from ME pairs to copies of the state at hand.
E
C
:= sup{r : lim
n→∞
( inf
Λ
LOCC
ρ
⊗n
−Λ
LOCC
(Φ
+
2
rn)) = 0}. (34)
There are other operational measures of entanglement, for example the
ﬁdelity (a measure of the faithfulness of the process) of teleportation itself
can be considered as an entanglement measure. For now however, we will
leave the operational approach and go back to axiomatic thinking.
Axiomatically, we say that any entanglement measure is a map E :
S(H
A
⊗H
B
⊗... ⊗H
N
) →R
+
such that it also obeys:
1) Monotonicity under LOCC maps Λ
LOCC
E(Λ
LOCC
(ρ)) ≤ E(ρ). (35)
More precisely we ask that it decreases on average. That is for an
E(ρ) ≥
m
p
m
E (ε
m
LOCC
(ρ)) , (36)
where Λ
LOCC
(ρ) =
m
ε
m
LOCC
(ρ) takes the state to ensemble
{p
m
, ε
m
LOCC
(ρ)} where m is the (known) outcome of some measurement.
2) Vanishing on separable states, E(ρ
SEP
) = 0 for all ρ
SEP
∈ SEP.
We call any such functional obeying these properties an entanglement
monotone.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
107
This list of axioms is somehow a minimum requirement of a good en
tanglement measure and some authors also include additional axioms, e.g.
of additivity, continuity and convexity. Additivity and continuity are im
portant when it comes to considering inﬁnite limits (of dimension and of
number of copies in entanglement manipulation) and convexity sometimes
added because of its mathematical convenience.
Notice that monotonicity under LOCC implies that local unitary (LU)
invariance of entanglement monotones. We can see this since unitaries are
reversible. If an LU operation U decreases entanglement, then U
†
, which is
also LU can increase entanglement causing a contradiction. This of course
makes sense since LU operations are just local changes in the basis and
should not eﬀect global properties
g
.
An example of a measure satisfying all of the above, often used in quan
tum information is the relative entropy of entanglement
22
E
R
(ρ) = min
ω∈SEP
D(ρω), (37)
where D(ρω) = Tr[ρ(log
2
ρ−log
2
ω)] is the relative entropy. This is an ex
ample of a socalled “distancelike” measure of entanglement, where we look
for the closest separable state with respect to some distancelike function
(in this case the relative entropy). Of course we must be careful with which
function to minimise over to check that it is an entanglement monotone.
Of course there are many more entanglement measures in exsistence,
but this lecture course is too short to cover them all and rather we refer to
larger reviews.
1,16
4.3. Uniqueness of measures, order on states
A question which naturally arises is, is there a unique measure of entangle
ment within this framework. The simple answer is no (at least if we just talk
about entanglement monotones). This can be seen in the bipartite case by
the seminal result of Nielsen,
23
which states the following. Given two bipar
tite states ψ and φ, with ordered Schmidt coeﬃcients {ψ
i
} and {φ
i
}, it
is possible to convert ψ → φ by LOCC perfectly (with unit probability)
g
Note that global basis change can of course generate (or degrade) entanglement. This
may seem obvious but sometimes it can cause confusion to physicists in trying to decide
if a system can be said to be ‘entangled’ and there could be two or more basis physically
accessible, but entanglement exists only in one. We must be careful in physics to state
clearly which global basis we are talking about.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
108
iﬀ {ψ
i
} are majorized by {φ
i
}, that is iﬀ for all k > 0 we have
k
i=1
ψ
i
≤
k
i=1
φ
i
. (38)
Since majorisation is a partial order, we see that LOCC induces only a
partial order on the set of bipartite pure states. Thus there exist pairs of
states such that
ψ
LOCC
φ. (39)
It is therefore possible to deﬁne entanglement monotones such that
E
1
(ψ) > E
1
(φ)
E
2
(ψ) < E
2
(φ). (40)
However, for bipartite pure states at least, this problem can be smoothed
over when considering asymptotics. It can be shown
24
that by including
certain axioms, that for pure states all measures coincide to the entropy of
entanglement
20
E
ent
(ψ
AB
) := S(ρ
A
) = −
i
p
i
log
2
p
i
, (41)
where ρ
A
is the reduced state of subsystem A (or B equivalently), S(ρ) =
−Tr[ρ log
2
ρ] is the von Neumann entropy (which is a measure of the mixed
ness of a state) and p
i
are the Schmidt coeﬃcients of ψ
AB
.
In fact, under these additional axioms, it can be shown that for bipartite
mixed states all such measures are bounded by E
D
(Eqn. (33)) and E
C
(Eqn. (34)) (Ref. 24)
E
D
≤ E ≤ E
C
. (42)
One may ask, then, if there is a better way of deﬁning entanglement
such that there is always a unique measure (not just for the bipartite pure
state case). For example, could this be possible by adding new axioms or
changing those we have? However, such approaches have been studied a lot
in the literature and so far no good answer has been found.
So it seems for now at least that we are stuck with having to accept
that there are diﬀerent types of entanglement, corresponding to diﬀerent
measures. So we can ask, what does that mean? May it be that diﬀerent
types of entanglement are useful for diﬀerent quantum information tasks?
Or could they hold some diﬀerent signiﬁcance in diﬀerent areas of physics?
Interestingly we see then that even in our attempt to ﬁnd a clean axiomatic
mathematical approach, we are in the end drawn back to questions of the
operational value of our theory of entanglement.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
109
4.4. Measuring entanglement
Coming now back to physics for a bit, we would like to be able to know
what we can say about viewing the presence of entanglement in actual
experiments.
The most direct approach to this is total state tomography. This is
where we repeatedly prepare and measure a state to build up statistics up
to the point when we ﬁnally get all elements of the density matrix. From
this we can use the mathematical tools developed so far to answer ques
tions of separability and amount of entanglement. However this approach
is experimentally very costly and we would prefer a more eﬃcient strategy.
One approach we heard about before in section 3 concerning treat
ing nonCPTP maps experimentally. In Ref. 25 and developed in Ref. 5
(amongst others) an ingenious general approach is given to approximating
experimentally any nonlinear function on states by making global measure
ments on several copies. Such techniques can be used to check separability
and approximate the value of entanglement measures.
An even weaker (but still incredibly powerful) approach we have also
discussed before is that of entanglement witnesses. As mentioned, not only
can entanglement witnesses be used to separate entangled states from the
set of separable states, but also in certain cases their violation can be used
to bound entanglement
11,12,26
and as an entanglement measure itself.
26
The main idea is, given the expectation value of a set of witnesses {W
k
}
we want to ﬁnd the worst case corresponding value for entanglement. For
some chosen entanglement measure E, this gives the bound
(ω
1
, ..., ω
n
) := inf{E(ρ)Tr[W
k
ρ] = ω
k
∀k}. (43)
This can be phrased in terms of Legendre transformations, and has been
calculated for several witnesses and entanglement measures and further ap
plied to experiments, amazingly allowing some old results for the calculation
of a witness to give bounds on the value of entanglement.
4.5. Multipartite entanglement
We ﬁnish our review of entanglement with a discussion about the multi
partite case. It is no coincidence that most of the clean results we have
presented are for the bipartite case (if not just the bipartite pure). We have
seen the diﬃculty in deciphering the structure of statespaces induced by
entanglement in the bipartite case. It gets even worse for the multipartite
case of three or more systems.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
110
Even under a simpler set of operations of stochastic LOCC (SLOCC) 
LOCC allowing for non unit probability of success (CP LOCC maps with
known outcome)  we have inequivalent states for three or more parties.
27
For four or more there are possibly even inﬁnitely many diﬀerent classes of
states under SLOCC. Thus we have many diﬀerent ‘types’ or entanglement.
We are then forced to take steps to simplify the situation. One approach
is to consider which kinds of states are ‘normal’. That is, which kinds of
states are likely to occur in any given situation. This can lead to a much
simpler induced structure of state space (e.g. Ref. 28). However many ques
tions still remain as to how to deﬁne well the notion of ‘likely’. Typically
we take the approach of random sampling, but we would really like some
physical reason to exclude certain states (initial answers are proposed in
Ref. 29). Indeed, even if some states are ‘rare’ physically, it may be that
they are so useful to us that we would like to still use them in quantum
information.
Another approach that bares fruit is to take simple measures of en
tanglement such as distancelike measures. The simplicity of the deﬁnition
skips much of the detailed structure of the state space. Although the deﬁn
ing optimization makes calculation diﬃcult, in some cases (including many
states useful in quantum information) symmetries can be used to calculate
entanglement and entanglement properties.
Although this class of measures misses out on much of the richness of
the structure of state space under LOCC, this approach does allow some
freedom in choosing which ‘type’ of entanglement is considered. As stated in
(37), it is completely unclear where the entanglement arises in these kinds
of measures. For example a GHZ state
GHZ
ABC
=
1
√
2
(000
ABC
+ 111
ABC
), (44)
has the same relative entropy of entanglement as
Φ
ABC
=
1
√
2
(00
AB
+ 11
AB
) ⊗ φ
C
, (45)
where φ
C
is an arbitrary single qubit state, with E
R
(GHZ) = E
R
(Φ) =
1, even though the entanglement is clearly diﬀerent. By choosing E
R
with
respect to diﬀerent cuts, we can discriminate these states. Calling E
Bi(AB:C)
R
the relative entropy of entanglement with respect to the set of separable
states across the bipartite cut AB : C, we have E
Bi(AB:C)
R
GHZ = 1 and
E
Bi(AB:C)
R
Φ = 0. Hence by analysing diﬀerent cuts we can get some struc
ture from entanglement also.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
111
Of course, the simplest approach is that of entanglement witnesses men
tioned before. This is particularly successful in the multipartite case since
other approaches are so diﬃcult. It can also be adapted to consider dif
ferent types of entanglement. Again by being careful in specifying which
separable set we compare a given state to, we can witness whether a state
is entangled over all systems, or just over some subsystems.
5. Conclusions
In these notes we have attempted to present the basics of entanglement the
ory. We have taken a rather mathematical approach, and left the physical
motivations to further reading. Having said that we have seen interesting is
sues raised about the physical measurement of entanglement and mentioned
a few possible approaches.
Of course, we have only skimmed the surface of this fertile ﬁeld, and so
to ﬁnish, it is perhaps ﬁtting to list a few of the important topics we have
not covered in these notes (where direct references are not given, references
can be found in Ref. 1).
• Inﬁnite dimensional systems. In all of the results here we have only
considered ﬁnite sized systems. As hinted at, when going to inﬁnities,
there are many more subtleties that must be addressed, such as con
tinuity. Indeed entanglement may not be continuous in state space for
this case (though some continuity can be regained by certain ‘physical’
restrictions such as ﬁnite energy).
30
• Connection to quantum channels. There are a plethora of results re
lating the ability to generate/manipulate entanglement to the qual
ity of a quantum channel as an information carrier. This approach
has ﬁelded many useful results in both directions, including bounds
on how good channels are in terms of entanglement and new informa
tion based entanglement measures such as the squashed entanglement.
31
Also, through equivalence of problems it has yielded one of the largest
remaining problems of entanglement theory (next point).
• Additivity. That is given several copies of a state, how does entangle
ment scale (roughly speaking). Through connections to communication
theory mentioned above, the question of additivity of the entanglement
of formation is shown to have consequences in the ability for additivity
of communication quantities such as the Holevo bound.
• Use of entanglement in quantum information. We have only brieﬂy
touched upon the usefulness of entanglement as a resource here. There
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
112
are in fact many open questions remaining about its use in quantum
information, particularly in the multipartite case. In quantum com
munication our understanding of the role of entanglement is pretty
restricted to the bipartite case; could it be that the subtleties of multi
partite entanglement have much more to oﬀer, perhaps diﬀerent classes
of entanglement useful for diﬀerent tasks? In quantum computation we
are only making the ﬁrst steps to understand exactly the role and re
quirement of entanglement (in particular, we have recent progress in
measurement based quantum computing
32
).
• Role of entanglement in manybody physics. This question is developing
into one of the most active areas of current entanglement research. See
e.g. Ref. 33 for a recent review. Many interesting questions arise from
both physical and mathematical points of view as to the relationship
between entanglement theory and the theory of criticality in many body
physics.
• The physics of entanglement. In these notes we have taken largely an
abstract approach to the whole subject. In real physical systems there
are other considerations that must be taken into account. Firstly, as
mentioned brieﬂy, in any system, the choice of how we split our space
is a physical one, and so the deﬁnition of LOCC must depend on the
physics at hand. Secondly there are often many more constraints in
a physical system as to the allowed set of operations and states, for
example those induced by symmetries. It is another ongoing ﬁeld of
research to try to impose these extra fundamental physical laws to our
theory of entanglement.
References
1. R. Horodecki, P. Horodecki, M. Horodecki and K. Horodecki, quant
ph/0702225.
2. M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum In
formation (Cambridge University Press, 2000).
3. M. Hayashi, Quantum Information An Introduction (SpringerVerlag, 2006).
4. M. Horodecki, P. Horodecki and R. Horodecki, Phys. Lett. A 223 (1996) 1.
5. H. Carteret, Phys. Rev. Lett. 94 (2005) 040502.
6. A. Peres, Phys. Lett. A 128 (1988) 19.
7. P. Horodecki, Phys. Lett. A 232 (1997) 333.
8. K.
˙
Zyczowski, P. Horodecki, A. Sanpera and M. Lewenstein, Phys. Rev. A
58 (1998) 883.
9. B. M. Terhal, Phys. Lett. A 271 (2000) 319.
10. I. Bengtsson and K.
˙
Zyczkowski, Geometry of Quantum States (Cambridge
University Press, 2006).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
113
11. O. G¨ uhne, M. Reimpell and R. Werner, Phys. Rev. Lett. 98 (2007) 110502.
12. J. Eisert, F. B. ao and K. Audenaert, New J. Phys. 9 (2007) 46.
13. K. M. R. Audenaert and M. B. Plenio, New J. Phys. 8 (2006) 266.
14.
ˇ
C. Brukner and V. Vedral quantph/0406040.
15. G. T´oth, Phys. Rev. A 71 (2005) 010301.
16. M. Plenio and S. Virmani, Quant. Inf. Comp. 7 (2007) 1.
17. C. H. Bennett, G. Brassard, C. Crepeau, R. Jozsa, A. Peres and W. K.
Wootters, Phys. Rev. Lett 70 (1993) 1895.
18. C. H. Bennett and G. Brassard (1984), in Proc. IEEE International Confer
ence on Computers, Systems, and Signal Processing, Bangalore, India (IEEE,
New York, 1984).
19. A. Ekert, Phys. Rev. Lett. 67 (1991) 661.
20. C. H. Bennett, H. J. Bernstein, S. Popescu and B. Schumacher, Phys. Rev. A
53 (1996) 2046.
21. E. Rains, Phys. Rev. A 60 (1999) 173.
22. V. Vedral and M. B. Plenio, Phys. Rev. A 57 (1998) 1619.
23. M. A. Nielsen, Phys. Rev. Lett. 83 (1999) 436.
24. M. Donald, M. Horodecki and O. Rudolph, J. Math. Phys. 43 (2002) 4252.
25. A. Ekert, C. M. Alves, D. K. L. Oi, M. Horodecki, P. Horodecki and L. Kwek,
Phys. Rev. Lett. 88 (2002) 217901.
26. F. G. S. L. Brand˜ao, Phys. Rev. A 72 (2005) 022310.
27. W. D¨ ur, G. Vidal and J. I. Cirac, Phys. Rev. A 62 (2000) 062314.
28. P. Hayden, D. Leung and A. Winter, Comm. Math. Phys. 265 (2005) 95.
29. O. Dahlsten, R. Oliveira and M. B. Plenio, J. Phys. A: Math. Theo. 40 (2007)
8081.
30. J. Eisert, C. Simon and M. B. Plenio, J. Phys. A 35 (2002) 3911.
31. M. Christandl and A. Winter, J. Math. Phys. 45 (2004) 829.
32. M. V. den Nest, W. D¨ ur, A. Miyake and H. J. Briegel, New J. Phys. 9 (2007)
204.
33. L. Amico, R. Fazio, A. Osterloh and V. Vedral, quantph/0703044.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
This page intentionally left blank This page intentionally left blank
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
115
HOLONOMIC QUANTUM COMPUTING AND ITS
OPTIMIZATION
SHOGO TANIMURA
Department of Applied Mathematics and Physics, Graduate School of Informatics,
Kyoto University, Kyoto 6068501, Japan
Email: tanimura@i.kyotou.ac.jp
Holonomic quantum computer uses a nonAbelian adiabatic phase associated
with a controlparameter loop as a unitary gate. Holonomic computation should
be processed slowly for keeping adiabaticity. On the other hand, holonomic
computation should be executed in a short time for avoiding decoherence.
These two requests, the slow process and the quick execution, seem contradic
tory each other and lead to a kind of optimization problem, which is called the
isoholonomic problem. In this lecture we review the formulation of the holon
omy in quantum mechanics and the solution of the optimal control problem
of the holonomy. We also discuss possible implementations of the holonomic
quantum computer.
Keywords: Quantum Computation, Berry Phase, Holonomy, Optimization, Iso
holonomic Problem, Horizontal Extremal Curve.
1. Introduction
Many designs of practical quantum computers have been proposed in this
decade. Holonomic quantum computer is one of those candidates and has
been studied mostly in the theoretical aspect. A few years ago Hayashi,
Nakahara and I obtained an exact result about the optimal control prob
lem of the holonomic quantum computer.
1,2
In this lecture I aim to explain
the basic of holonomic quantum computation. I will give a quick review
on Berry phase and WilczekZee holonomy and introduce the idea of holo
nomic quantum computation. I will formulate the optimization problem
of holonomic quantum computer and show its solution. I will apply the
result to the two wellknown unitary gates; the Hadamard gate and the
discrete Fourier transformation gate. Finally, I will brieﬂy discuss the re
maining problems, which concern implementation of holonomic quantum
computation in physical systems. Most of this lecture note depends on our
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
116
papers.
1,2
2. Holonomies in Mathematics and Physics
2.1. Holonomy in Riemannian geometry
The ﬁrst example of holonomy can be found in geometry.
3,4
When a tangent
vector is parallel transported along a closed loop in a Riemannian manifold,
the vector gets a net rotation. The net rotation generated by the parallel
transportation is called the holonomy. In this case, the holonomy manifests
curvature of the manifold. The concept of parallel transportation has been
generalized in various contexts and described in terms of the connection or
the gauge ﬁeld as discussed below.
2.2. Berry phase in quantum mechanics
We shall review the adiabatic theorem
5,6
in quantum mechanics and the
formulation of Berry phase.
7
Comprehensive reviews on this subject have
been published.
8–10
Let us consider a quantum system whose Hamiltonian
H(λ) depends on a set of parameters λ = (λ
1
, . . . , λ
m
). In practice λ might
be an electric ﬁeld or a magnetic ﬁeld applied on the system. These pa
rameters can be controlled externally and are speciﬁed as functions of time
t. The whole of possible values of λ forms a parameter space M. Suppose
that the Hamiltonian has a discrete spectrum
H(λ) n; λ = ε
n
(λ) n; λ (1)
for each value of λ. The state ψ(t) evolves obeying the Schr¨ odinger equa
tion
i
d
dt
ψ(t) = H(λ(t)) ψ(t). (2)
Suppose that at t = 0 the initial state is an eigenstate of the ini
tial Hamiltonian as ψ(0) = n; λ(0). Assume that no degeneracy oc
curs along the course of the parameter λ(t). Namely, it is assumed that
ε
n
(λ(t)) = ε
n
(λ(t)) for any n
(= n) and any t. Furthermore, let us as
sume that we change the control parameter λ(t) as slowly as possible. The
slowness will be deﬁned at (7) in more rigorous words. Then the adiabatic
theorem
5,6
tells that the state ψ(t) remains an eigenstate of the instanta
neous Hamiltonian H(λ(t)) at each time t:
ψ(t) = e
iγ(t)
n; λ(t). (3)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
117
In other words, if the external parameter changes slowly enough, and if no
level crossing occurs, then transition between diﬀerent energy levels does
not occur and the quantum state remains quasistationary in some sense.
Suppose that at a time T the parameter λ(t) returns to the initial value
as λ(T) = λ(0). Then the state returns to the initial state up to a phase
factor:
ψ(T) = e
iγ(T)
ψ(0). (4)
The phase γ is given as a sum γ = γ
d
+ γ
g
with
γ
d
= −
1
T
0
ε
n
(λ(t)) dt, (5)
γ
g
= i
T
0
n; λ(t)
d
dt
n; λ(t) dt = i
C
m
µ=1
n; λ
∂
∂λ
µ
n; λ dλ
µ
. (6)
A proof of these formulae will be given around (16) in a more general
context. The phase γ
d
is called the dynamical phase. The phase γ
g
is called
the geometric phase or the Berry phase. It is independent of the time
parametrization of the curve C in the control parameter space.
Suﬃcient slowness of movement of the control parameter is character
ized by the condition
{ε
n
(λ(t)) − ε
n
(λ(t))}T 2π (7)
for n
= n during 0 ≤ t ≤ T. This inequality is called the adiabaticity
condition. When two distinct energy levels approach closely or cross, the
adiabaticity condition is violated.
2.3. WilczekZee holonomy in quantum mechanics
Immediately after Berry formulated his celebrated phase, which belongs to
the onedimensional unitary group U(1), Wilczek and Zee
11
formulated its
generalization to an arbitrary unitary group U(k). Here we shall describe
their formulation. Suppose that the Hamiltonian has a discrete spectrum
but now degenerate eigenvalues:
H(λ) n, α; λ = ε
n
(λ) n, α; λ. (8)
Here α = 1, 2, . . . , k
n
is an index to label degenerate energy eigenstates.
It is assumed that these eigenvectors are normalized as n, α; λn
, α
; λ =
δ
nn
δ
αα
. It is said that the energy eigenvalue ε
n
(λ) is k
n
fold degenerate.
Suppose that the degrees of degeneracy does not change along the course of
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
118
the controlled parameter λ(t). Then the adiabatic theorem again tells that
transition between diﬀerent energy levels occurs with negligible probability.
If the initial state is one of energy eigenstates, ψ(0) = n, α; λ(0), then it
remains a linear combination of the instantaneous energy eigenstates with
the same quantum number n:
ψ(t) =
k
n
¸
β=1
c
βα
(t) n, β; λ(t). (9)
Since the quantum number n is unchanged, we omit it in the following
equations. By substituting (9) into (2), we get
d
dt
c
βα
(t) +
k
¸
γ=1
β; λ(t)
d
dt
γ; λ(t)
c
γα
(t) = −
i
ε(γ(t))c
βα
(t), (10)
whose formal solution is given as
c
βα
(t) = exp
−
i
t
0
ε(λ(s))ds
T exp
−
t
0
A(s)ds
βα
(11)
with the matrixvalued function
A
βα
(t) =
β; λ(t)
d
dt
α; λ(t)
=
m
¸
µ=1
β; λ
∂
∂λ
µ
α; λ
dλ
µ
dt
. (12)
The symbol T indicates the timeordered product. It is easily veriﬁed that
A
∗
βα
= −A
αβ
. We introduce a u(k)valued oneform
a
A
βα
(λ) =
m
¸
µ=1
A
βα,µ
(λ) dλ
µ
=
m
¸
µ=1
β; λ
∂
∂λ
µ
α; λ
dλ
µ
(13)
in the parameter space. A is called the WilczekZee connection. Then the
unitary matrix appearing in (11) is rewritten as
Γ(t) = P exp
−
λ(t)
λ(0)
A
, (14)
where P stands for the pathordered product. It is deﬁned by the set of
diﬀerential equations
d
dt
Γ
βα
(t) = −
k
n
¸
γ=1
m
¸
µ=1
A
βγ,µ
(λ(t))Γ
γα
(t)
dλ
µ
dt
, Γ
βα
(0) = δ
βα
. (15)
a
We denote the Lie algebra of the Lie group U(k) by u(k), which is the set of k
dimensional skewHermite matrices.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
119
We assume that the control parameter λ(t) comes back to the initial point
λ(T) = λ(0). However, the state ψ(T) fails to return to the initial state
and is subject to a unitary rotation as
ψ(T) = exp
−
i
T
0
ε(λ(s))ds
k
β=1
n, β; λ(0)Γ
βα
(T). (16)
The unitary matrix
Γ[C] := Γ(T) = P exp
−
C
A
∈ U(k) (17)
is called the holonomy matrix associated with the loop λ(t). It is important
to note that the holonomy Γ[C] depends only on the shape of the loop C
in the parameter space and is independent of parametrization of the loop
by λ(t).
2.4. Examples
Here we shall demonstrate calculations of the Berry phase and the Wilczek
Zee holonomy.
2.4.1. Berry phase
This example was ﬁrst studied in Berry’s original paper.
7
Let us consider
a system of spin one half which is put in a magnetic ﬁeld
B =
B
x
B
y
B
z
= B
sin θ cos φ
sin θ sin φ
cos θ
. (18)
The Hamiltonian of the spin is
H = B
x
σ
x
+B
y
σ
y
+B
z
σ
z
=
B
z
B
x
−iB
y
B
x
+iB
y
−B
z
= B
cos θ e
−iφ
sin θ
e
iφ
sin θ −cos θ
. (19)
Its eigenvalues and eigenvectors are given as
H +; B = B+; B, +; B =
e
−iφ/2
cos(θ/2)
e
iφ/2
sin(θ/2)
, (20)
H −; B = −B−; B, −; B =
−e
−iφ/2
sin(θ/2)
e
iφ/2
cos(θ/2)
. (21)
Energy eigenvalues are nondegenerate for any nonvanishing B. Hence, by
an adiabatic change of the external parameter (B, θ, φ), we will observe
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
120
the Berry phase. The connection form A = ·d·, which is deﬁned by
(13), is calculated as follows. First, calculate the exterior derivatives of the
eigenvectors to get
d+; B = d
e
−iφ/2
cos(θ/2)
e
iφ/2
sin(θ/2)
=
i
2
−e
−iφ/2
cos(θ/2)
e
iφ/2
sin(θ/2)
dφ +
1
2
−e
−iφ/2
sin(θ/2)
e
iφ/2
cos(θ/2)
dθ, (22)
d−; B = d
−e
−iφ/2
sin(θ/2)
e
iφ/2
cos(θ/2)
=
i
2
e
−iφ/2
sin(θ/2)
e
iφ/2
cos(θ/2)
dφ −
1
2
e
−iφ/2
cos(θ/2)
e
iφ/2
sin(θ/2)
dθ. (23)
Second, calculate the inner products
A
+
(B) = +; B d+; B
= (e
iφ/2
cos(θ/2), e
−iφ/2
sin(θ/2)) d
e
−iφ/2
cos(θ/2)
e
iφ/2
sin(θ/2)
= (e
iφ/2
cos(θ/2), e
−iφ/2
sin(θ/2))
i
2
−e
−iφ/2
cos(θ/2)
e
iφ/2
sin(θ/2)
dφ +
1
2
−e
−iφ/2
sin(θ/2)
e
iφ/2
cos(θ/2)
dθ
¸
=
i
2
− cos
2
(θ/2) + sin
2
(θ/2)
dφ
+
1
2
− cos(θ/2) sin(θ/2) + sin(θ/2) cos(θ/2)
dθ
= −
i
2
cos θ dφ, (24)
A
−
(B) = −; B d−; B
= (−e
iφ/2
sin(θ/2), e
−iφ/2
cos(θ/2)) d
−e
−iφ/2
sin(θ/2)
e
iφ/2
cos(θ/2)
=
i
2
cos θ dφ. (25)
For a loop C = ∂S in the parameter space, we will observe the Berry phase
e
iγ
= exp
¸
−
C
A
+
= exp
¸
i
2
C
cos θ dφ
= exp
¸
−
i
2
S
sin θ dθ ∧ dφ
(26)
if the initial state is +; B.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
121
2.4.2. Λtype system
This system has been studied as a model of the holonomic quantum com
puter which uses atomic energy levels
21
and also as a model of the holonomic
computer which uses Josephson junctions.
22
The Hilbert space is assumed
to be C
N+1
. The Hamiltonian is
H = εee +
N
j=1
ω
j
g
j
e + ω
∗
j
eg
j

. (27)
The state e represents an excited state and g
j
(j = 1, · · · , N) repre
sent degenerate ground states before couplings between the states are in
troduced. The parameter ε ∈ R is the excitation energy and ω
j
∈ C are
coupling constants. These parameters are assumed to be controllable. The
Hamiltonian can be written in the matrix form
H =
ε ω
∗
1
ω
∗
2
· · · ω
∗
N
ω
1
0 0 · · · 0
ω
2
0 0 · · · 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
ω
N
0 0 · · · 0
. (28)
Its eigenvalues and eigenvectors are
He
+
= E
+
e
+
, e
+
= c
+
E
+
ω
1
ω
2
.
.
.
ω
N
, E
+
=
ε
2
+
ε
2
2
+ ω
2
, (29)
He
−
= E
+
e
−
, e
−
= c
−
E
−
ω
1
ω
2
.
.
.
ω
N
, E
−
=
ε
2
−
ε
2
2
+ ω
2
, (30)
Hα = 0, α =
1
ω
⊥

0
ω
⊥
, (31)
where c
±
= (E
2
±
+ ω
2
)
−
1
2
. The vector ω
⊥
∈ C
N
satisﬁes ω, ω
⊥
= 0.
Thus, the eigenvalue E = 0 is (N − 1)fold degenerate.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
122
When N = 3, we have the twofold degenerate level E = 0. The eigen
vectors are parametrized as
α
1
=
1
ω
1

2
+ ω
2

2
0
ω
∗
2
−ω
∗
1
0
, (32)
α
2
=
1
(ω
1

2
+ ω
3

2
)(ω
1

2
+ ω
2

2
+ ω
3

2
)
0
ω
1
ω
∗
3
ω
2
ω
∗
3
−(ω
1

2
+ ω
2

2
)
.(33)
The components of the connection are to be calculated according to
A =
α
1
, dα
1
α
1
, dα
2
α
2
, dα
1
α
2
, dα
2
=
6
µ=1
A
µ
dλ
µ
(34)
with ω
1
= λ
1
+iλ
2
, ω
2
= λ
3
+iλ
4
, ω
3
= λ
5
+iλ
6
. The calculation to get the
concrete forms of A
µ
is cumbersome and we do not show it here. When we
take a loop λ(t), the associated holonomy g(t), which is an (N −1)×(N−1)
matrix, is a solution of the diﬀerential equation
dg(t)
dt
= −
µ
A
µ
(λ(t)) g(t)
dλ
µ
(t)
dt
, g(0) = I. (35)
The concrete calculations of the holonomy can be found in literature.
21,22
3. Holonomic Quantum Computer
Zanardi and Rasetti,
12
and also Pachos with them
13
proposed to use the
WilczekZee holonomy (17) as a unitary gate of quantum computer. Since
the holonomy depends only on the geometry of the controlparameter loop
and it is insensitive to the detail of the dynamics of the system, the holo
nomic quantum computer is robust against noise from the environment and
tolerates incompleteness of the control. For these reasons the holonomic
quantum computer is regarded as one of promising candidates for realis
tic quantum computers. However, a useful holonomic quantum computer
should have the following properties:
(i) Controllability: a holonomic quantum computer should have a
large enough number of control parameters for generating arbitrary
holonomies.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
123
(ii) Accurate degeneracy: degeneracy of the energy levels should be kept
during the timeevolution. In an actual system degeneracy is often re
solved by a tiny asymmetric interaction or a noise. Inaccuracy of degen
eracy spoils the precision of the holonomic computation. Moreover, the
degree of degeneracy is equal to the dimension of the Hilbert space avail
able for quantum computation. Hence, to realize a large qubits system it
is necessary to ﬁnd a system which has a highly degenerate energy level.
(iii) Adiabaticity: we need to change the control parameter as slowly as pos
sible to avoid undesirable transitions between diﬀerent energy levels of
the quantum system. No level crossings should occur during the process
since a level crossing violates the adiabaticity condition.
(iv) Short execution time: an actual system cannot be completely isolated
from its environment, and the interaction between the system and the
environment causes loss of coherence of the quantum state of the sys
tem. We need to ﬁnish the whole process in a possible shortest time to
avoid the undesirable decoherence.
The requests (i) and (ii) are practically important issues. Although many
designs of working holonomic computers have been proposed, it remains
a hard task to ﬁnd a feasible design. The requests (iii) and (iv) can be
analyzed with a mathematical method. They naturally lead us to an opti
mization problem; we need to ﬁnd the shortest loop in the control parameter
space with producing the demanded holonomy. This optimization problem
is the main subject of this lecture.
4. Formulation of the Problem and its Solution
4.1. Geometrical setting
The diﬀerential geometry of connection
3,4
provides a suitable language to
describe the WilczekZee holonomy. A concise review on this subject also
given by Fujii.
14
We outline the geometrical setting of the problem here to
make this paper selfcontained.
Suppose that the state of a system is described by a ﬁnitedimensional
Hilbert space C
N
and the system has a family of Hamiltonians acting on
the Hilbert space. Assume that the ground state of each Hamiltonian is k
fold degenerate (k < N). A mathematical setting to describe this system is
the principal ﬁber bundle (S
N,k
(C), G
N,k
(C), π, U(k)), which consists of the
Stiefel manifold S
N,k
(C), the Grassmann manifold G
N,k
(C), the projection
map π : S
N,k
(C) → G
N,k
(C), and the unitary group U(k) as explained
below.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
124
The Stiefel manifold is the set of orthonormal kframes in C
N
,
S
N,k
(C) := {V ∈ M(N, k; C)  V
†
V = I
k
}, (36)
where M(N, k; C) is the set of N × k complex matrices and I
k
is the k
dimensional unit matrix. The unitary group U(k) acts on S
N,k
(C) from the
right
S
N,k
(C) ×U(k) → S
N,k
(C), (V, h) → V h (37)
by means of matrix product. It should be noted that this action is free. In
other words, h = I
k
if there exists a point V ∈ S
N,k
(C) such that V h = V .
The Grassmann manifold is deﬁned as the set of kdimensional sub
spaces in C
N
, or equivalently, as the set of projection operators of rank
k,
G
N,k
(C) := {P ∈ M(N, N; C)  P
2
= P, P
†
= P, trP = k}. (38)
The projection map π : S
N,k
(C) → G
N,k
(C) is deﬁned as
π : V → P := V V
†
. (39)
It is easily proven that the map π is surjective. Namely, for any P ∈
G
N,k
(C), there is V ∈ S
N,k
(C) such that π(V ) = P. It is easily observed
that
π(V h) = (V h)(V h)
†
= V hh
†
V
†
= V V
†
= π(V ) (40)
for any h ∈ U(k) and any V ∈ S
N,k
(C). It also can be veriﬁed that if
π(V ) = π(V
) for some V, V
∈ S
N,k
(C), there exists a unique h ∈ U(k)
such that V h = V
. Thus the Stiefel manifold S
N,k
(C) becomes a principal
bundle over G
N,k
(C) with the structure group U(k).
In particular, when k = 1, the principal bundle (S
N,k
(C), G
N,k
(C),
π, U(k)) is denoted as (S
2N−1
, CP
N−1
, π, U(1)) and is called the Hopf bun
dle. Then the Stiefel manifold becomes a (2N −1)dimensional sphere
S
N,1
(C) = S
2N−1
= {z ∈ C
N
 z
†
z = 1 }. (41)
At this time, we can introduce an equivalence relation z ∼ z
for z, z
∈
S
2N−1
as
z ∼ z
:⇐⇒ ∃θ ∈ R, e
iθ
z = z
(42)
and denote the equivalence class of z as [z] = {z
∈ S
2N−1
 z ∼ z
} =
{e
iθ
z  θ ∈ R}. Then the quotient space S
2N−1
/ ∼, which is the whole set
of the equivalence classes, is called the projective space CP
N−1
. Each point
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
125
[z] of CP
N−1
uniquely determines a onedimensional subspace in C
N
. It is
easily veriﬁed that
z ∼ z
⇐⇒ π(z) = π(z
) (43)
for the projection map π(z) = zz
†
. Thus the Grassmann manifold G
N,1
(C)
can be identiﬁed with CP
N−1
. The structure group U(1) = {e
iθ
} acts
on the ﬁber [z]. When N = 2 and k = 1, the Grassmann manifold is
isomorphic to G
2,1
(C)
∼
= CP
1
∼
= S
2
. At the same time the Stiefel manifold
is S
2,1
(C)
∼
= S
3
. Then the projection map is π : S
3
→ S
2
.
For a general values of N and k, the group U(N) acts on both S
N,k
(C)
and G
N,k
(C) as
U(N) ×S
N,k
(C) → S
N,k
(C), (g, V ) → gV, (44)
U(N) ×G
N,k
(C) → G
N,k
(C), (g, P) → gPg
†
(45)
by matrix product. It is easily veriﬁed that π(gV ) = gπ(V )g
†
. This action
is transitive, namely, there is g ∈ U(N) for any V, V
∈ S
N,k
(C) such that
V
= gV . There is also g ∈ U(N) for any P, P
∈ G
N,k
(C) such that
P
= gPg
†
. The stabilizer group of a point V ∈ S
N,k
(C) is deﬁned as
G
V
:= {g ∈ U(N)  gV = V } (46)
and it is isomorphic to U(N − k). Similarly, it can be veriﬁed that the
stabilizer group of each point in G
N,k
(C) is isomorphic to U(k) ×U(N−k).
Thus, they form the homogeneous bundle
π : S
N,k
(C)
∼
= U(N)/U(N −k) → G
N,k
(C)
∼
= U(N)/(U(k) ×U(N −k)).
(47)
The canonical connection form on S
N,k
(C) is a u(k)valued oneform
A := V
†
dV. (48)
which is a generalization of the WilczekZee connection (13). This is char
acterized as the unique connection that is invariant under the action (44).
The associated curvature twoform is then deﬁned and calculated as
F := dA + A ∧ A = dV
†
∧ dV + V
†
dV ∧ V
†
dV
= dV
†
∧ (I
N
−V V
†
)dV, (49)
where we used dV
†
V + V
†
dV = 0, which is derived from V
†
V = I
k
.
These manifolds are equipped with Riemannian metrics. We deﬁne a
metric
dV
2
:= tr (dV
†
dV ) (50)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
126
for the Stiefel manifold and
dP
2
:= tr (dPdP) (51)
for the Grassmann manifold.
4.2. The isoholonomic problem
Here we reformulate the WilczekZee holonomy in terms of the geometric
terminology introduced above. The state vector ψ(t) ∈ C
N
evolves accord
ing to the Schr¨ odinger equation
i
d
dt
ψ(t) = H(t)ψ(t). (52)
The Hamiltonian admits a spectral decomposition
H(t) =
L
n=1
ε
n
(t)P
n
(t) (53)
with projection operators P
n
(t). Therefore, the set of distinct energy eigen
values (ε
1
, . . . , ε
L
) and mutually orthogonal projectors (P
1
, . . . , P
L
) consti
tutes a complete set of control parameters of the system. Now we concen
trate on the eigenspace associated with the lowest energy, which is assumed
to be identically zero, ε
1
≡ 0. We write P
1
(t) as P(t) simply. Suppose that
the degree of degeneracy k = tr P(t) is constant. For each t, there exists
V (t) ∈ S
N,k
(C) such that P(t) = V (t)V
†
(t). By adiabatic approximation
we mean that the state vector ψ(t) belongs to the subspace spanned by the
frame V (t), which is the eigensubspace of the instantaneous Hamiltonian.
Then for the state vector ψ(t) there exists a vector φ(t) ∈ C
k
such that
ψ(t) = V (t)φ(t). (54)
This equation is equivalent to (9). Since H(t)ψ(t) = ε
1
ψ(t) = 0, the
Schr¨ odinger equation (52) becomes
dφ
dt
+ V
†
dV
dt
φ(t) = 0 (55)
and its formal solution is written as
φ(t) = P exp
−
V
†
dV
φ(0). (56)
Therefore ψ(t) is written as
ψ(t) = V (t)P exp
−
V
†
dV
V
†
(0)ψ(0). (57)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
127
In particular, when the control parameter comes back to the initial point
as P(T) = P(0), the holonomy Γ ∈ U(k) is deﬁned via
ψ(T) = V (0)Γφ(0) (58)
and is given explicitly as
Γ = V (0)
†
V (T) P exp
−
V
†
dV
. (59)
If the condition
V
†
dV
dt
= 0 (60)
is satisﬁed, the curve V (t) in S
N,k
(C) is called a horizontal lift of the curve
P(t) = π(V (t)) in G
N,k
(C). Then the holonomy (59) is reduced to
Γ = V
†
(0)V (T) ∈ U(k). (61)
Now we are ready to state the isoholonomic problem in the present
context; given a speciﬁed unitary gate U
gate
∈ U(k) and an initial point
P
0
∈ G
N,k
(C), ﬁnd the shortest loop P(t) in G
N,k
(C) with the base points
P(0) = P(T) = P
0
whose horizontal lift V (t) in S
N,k
(C) produces a holon
omy Γ that coincides with U
gate
. This problem was ﬁrst investigated sys
tematically by a mathematician, Montgomery,
15
who was motivated from
experimental studies of Berry phase.
The isoholonomic problem is formulated as a variational problem. The
length of the horizontal curve V (t) is evaluated by the functional
S[V, Ω] =
T
0
tr
dV
†
dt
dV
dt
−tr
ΩV
†
dV
dt
dt, (62)
where Ω(t) ∈ u(k) is a Lagrange multiplier to impose the horizontality
condition (60) on the curve V (t). Note that the value of the functional S is
equal to the length of the projected curve P(t) = π(V (t)),
S =
T
0
1
2
tr
dP
dt
dP
dt
dt. (63)
Thus the problem is formulated as follows; ﬁnd functions V (t) and Ω(t) that
attain an extremal value of the functional (62) and satisfy the boundary
condition (61).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
128
4.3. The solution: horizontal extremal curve
We shall derive the EulerLagrange equation associated to the functional
(62) and solve it explicitly. A variation of the curve V (t) is deﬁned by an
arbitrary smooth function η(t) ∈ u(N) such that η(0) = η(T) = 0 and an
inﬁnitesimal parameter ∈ R as
V
(t) = (1 + η(t))V (t). (64)
By substituting V
(t) into (62) and diﬀerentiating with respect to , the
extremal condition yields
0 =
dS
d
=0
=
T
0
tr
˙ η(V
˙
V
†
−
˙
V V
†
−V ΩV
†
)
dt
=
tr
η(V
˙
V
†
−
˙
V V
†
−V ΩV
†
)
t=T
t=0
−
1
0
tr
η
d
dt
(V
˙
V
†
−
˙
V V
†
− V ΩV
†
)
dt. (65)
Thus we obtain the EulerLagrange equation
d
dt
(
˙
V V
†
−V
˙
V
†
+ V ΩV
†
) = 0. (66)
The horizontality condition V
†
˙
V = 0 is rederived from the extremal con
dition with respect to Ω(t). Finally, the isoholonomic problem is reduced
to the set of equations (60) and (66), which we call a horizontal extremal
equation.
Next, we solve the equations (60) and (66). The equation (66) is inte
grable and yields
˙
V V
†
−V
˙
V
†
+ V ΩV
†
= const = X ∈ u(N). (67)
Conjugation of the horizontality condition (60) yields
˙
V
†
V = 0. Then, by
multiplying V on (67) from the right we obtain
˙
V + V Ω = XV. (68)
By multiplying V
†
on (68) from the left we obtain
Ω = V
†
XV. (69)
The equation (68) implies
˙
V = XV −V Ω, and hence the time derivative of
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
129
Ω(t) becomes
˙
Ω = V
†
X
˙
V +
˙
V
†
XV
= V
†
X(XV −V Ω) + (−V
†
X + ΩV
†
)XV
= V
†
XXV −ΩΩ −V
†
XXV + ΩΩ
= 0. (70)
We used the facts that X
†
= −X and Ω
†
= −Ω. Therefore, Ω(t) is actually
a constant. Thus the solution of (68) and (69) is
V (t) = e
tX
V
0
e
−tΩ
, Ω = V
†
0
XV
0
. (71)
We call this solution the horizontal extremal curve. Then (67) becomes
(XV −V Ω)V
†
−V (−V
†
X + ΩV
†
) +V ΩV
†
= X,
which is arranged as
X −(V V
†
X +XV V
†
−V V
†
XV V
†
) = 0, (72)
where we used (69). Since the group U(N) acts on S
N,k
(C) freely, we may
take
V
0
=
I
k
0
∈ S
N,k
(C) (73)
as the initial point without loss of generality. We can parametrize X ∈ u(N),
which satisﬁes (69), as
X =
Ω W
−W
†
Z
(74)
with W ∈ M(k, N −k; C) and Z ∈ u(N −k). Then the constraint equation
(72) forces us to choose
Z = 0. (75)
Via the above discussion we obtained a complete set of solution (71) of
the horizontal extremal equation (60) and (66). The complete solutions are
parametrized with constant matrices Ω ∈ u(k) and W ∈ M(k, N − k; C),
and are written explicitly as
V (t) = e
tX
V
0
e
−tΩ
, X =
Ω W
−W
†
0
. (76)
We call the matrix X a controller. At this time the holonomy (61) is ex
pressed as
Γ = V
†
(0)V (T) = V
†
0
e
TX
V
0
e
−TΩ
∈ U(k). (77)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
130
These results (76) and (77) coincide with Montgomery’s results
b
. We took
a diﬀerent approach from his; in the above discussion we wrote down the
EulerLagrange equation and solved it directly.
We evaluate the length of the extremal curve for later convenience by
substituting (76) into (63) as
S =
T
0
1
2
tr
dP
dt
dP
dt
dt = tr (W
†
W) T. (78)
5. The BoundaryValue Problem
Remember that we are seeking for the optimally shortest loop in the Grass
mann manifold which generates a desired holonomy. The optimality is
guaranteed by solving the horizontal extremal equation. The remaining
task is to ﬁnd a solution that satisﬁes two boundary conditions. First,
when the curve V (t) in S
N,k
(C) is projected into G
N,k
(C) via the map
π : S
N,k
(C) → G
N,k
(C), V → V V
†
, it should make a loop. Thus the curve
should satisfy the loop condition
V (T)V
†
(T) = e
TX
V
0
V
†
0
e
−TX
= V
0
V
†
0
. (79)
Second, the terminal points V (0) and V (T) of the curve V (t) deﬁne the
holonomy Γ = V
†
(0)V (T) as in (61). Thus the curve should satisfy the
holonomy condition
V
†
0
V (T) = V
†
0
e
TX
V
0
e
−TΩ
= U
gate
(80)
for a speciﬁed unitary gate U
gate
∈ U(k).
We need to ﬁnd the suitable parameters Ω and W in solutions (76) sat
isfying the above two conditions. Montgomery
15
presented this boundary
value problem as an open problem. In this section we give a scheme to solve
systematically this problem.
5.1. Equivalence class
There is a class of equivalent solutions for a given initial condition V
0
and
a given ﬁnal condition V (T) = V
0
U
gate
. Here we clarify the equivalence
relation among solutions {V (t)} that have the form (76) and satisfy (79)
and (80).
b
In his paper
15
Montgomery cited B¨ar’s theorem to complete the proof. However, B¨ar’s
paper being a diploma thesis, it is not widely available. Instead of relying B¨ar’s theorem,
here we did the explicit calculation to reach the same result.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
131
We say that two solutions V (t) and V
(t) are equivalent if there are
elements g ∈ U(N) and h ∈ U(k) such that V (t) and
V
(t) = gV (t)h
†
(81)
satisfy the same boundary conditions
gV
0
h
†
= V
0
(82)
and
hU
gate
h
†
= U
gate
. (83)
For the initial point (73), the condition (82) states that g ∈ U(N) must
have a blockdiagonal form
g =
h
1
0
0 h
2
, h = h
1
∈ U(k), h
2
∈ U(N −k). (84)
The controller X
of V
(t) are then found from
V
(t) = gV (t)h
†
= ge
tX
g
†
gV
0
h
†
he
−tΩ
h
†
= e
tgXg
†
gV
0
h
†
e
−thΩh
†
= e
tgXg
†
V
0
e
−thΩh
†
. (85)
In summary, two controllers X and X
are equivalent if and only if there
are unitary matrices h
1
∈ U(k) and h
2
∈ U(N −k) such that
X =
Ω W
−W
†
0
, X
=
h
1
Ωh
†
1
h
1
Wh
†
2
−h
2
W
†
h
†
1
0
, h
1
U
gate
h
†
1
= U
gate
.
(86)
5.2. U(1) holonomy
Here we calculate the holonomy for the case N = 2 and k = 1. In this
case the homogeneous bundle π : S
2,1
(C) → G
2,1
(C) is the Hopf bundle
π : S
3
→ S
2
with the structure group U(1) and the WilczekZee holonomy
reduces to the Berry phase. In the subsequent subsection we will use the
Berry phase as a building block to construct a nonAbelian holonomy. We
normalize the cycle time as T = 1 in the following. Using real numbers
w
1
, w
2
, w
3
∈ R we parametrize the controller as
X =
2iw
3
iw
1
+w
2
iw
1
−w
2
0
= iw
3
I +iw
1
σ
1
+iw
2
σ
2
+iw
3
σ
3
, (87)
where σ = (σ
1
, σ
2
, σ
3
) = (σ
x
, σ
y
, σ
z
) are the Pauli matrices which have
been introduced at (19). The exponentiation of X yields
e
tX
= e
itw
3
(I cos ρt +in · σ sinρt), (88)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
132
where ρ and n are deﬁned as
ρ := w =
(w
1
)
2
+ (w
2
)
2
+ (w
3
)
2
, w = wn. (89)
The associated horizontal extremal curve (71) then becomes
V (t) = e
tX
V
0
e
−tΩ
= e
−itw
3
cos ρt +in
3
sin ρt
(in
1
−n
2
) sin ρt
(90)
and the projected curve in S
2
becomes
P(t) = V (t)V
†
(t)
=
1
2
I +
1
2
σ ·
n(n · e
3
) + (e
3
−n(n · e
3
)) cos 2ρt
−(n ×e
3
) sin2ρt
, (91)
where e
3
= (0, 0, 1). We see from (91) that the point P(t) in S
2
starts at
the north pole e
3
of the sphere and moves along a small circle with the
axis n in the clockwise sense by the angle 2ρt. The point P(t) comes back
to the north pole when t satisﬁes 2ρt = 2πn with an integer n. To make a
loop, namely, to satisfy the loop condition (79) at t = T = 1, the control
parameters must satisfy
ρ = w = nπ (n = ±1, ±2, . . . ). (92)
Then, the point P(t) travels the same small circle n times during 0 ≤ t ≤ 1.
Therefore, the integer n counts the winding number of the loop. At t = 1,
cos ρ = (−1)
n
and the holonomy (80) is evaluated as
V
†
0
e
X
V
0
e
−Ω
= e
iw
3
(−1)
n
e
−2iw
3
= e
−i(w
3
−nπ)
= U
gate
= e
iγ
. (93)
Thus, to produce the holonomy U
gate
= e
iγ
, the controller parameters are
ﬁxed as
w
3
= nπ −γ, w
1
+iw
2
= e
−iφ
(nπ)
2
−(nπ −γ)
2
. (94)
This is the solution to the boundary problem deﬁned by (79) and (80). Here
the nonvanishing integer n must satisfy (nπ)
2
− (nπ − γ)
2
> 0. The real
parameter φ is not ﬁxed by the loop condition and the holonomy condition.
The phase h
2
= e
iφ
parametrizes solutions in an equivalence class as ob
served in (86). Hence each equivalence class is characterized by the integer
n.
The length of the loop, (78), is now evaluated as
S = tr (W
†
W) T = (nπ)
2
−(nπ −γ)
2
= γ(2nπ −γ). (95)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
133
For a ﬁxed γ in the range 0 ≤ γ < 2π, the simple loop with n = 1 is
the shortest one among the extremal loops. Thus, we conclude that the
controller of U
gate
= e
iγ
is
X =
2i(π −γ) ie
iφ
π
2
−(π −γ)
2
ie
−iφ
π
2
−(π −γ)
2
0
. (96)
We call this solution a small circle solution because of its geometric picture
mentioned above.
5.3. U(k) holonomy
Here we give a prescription to construct a controller matrix X that gen
erates a speciﬁc unitary gate U
gate
. It turns out that the working space
should have a dimension N ≥ 2k to apply our method. In the following we
assume that N = 2k. The time interval is normalized as T = 1 as before.
Our method consists of three steps: ﬁrst, diagonalize the unitary matrix
U
gate
to be implemented, second, construct a diagonal controller matrix
by combining small circle solutions, third, undo diagonalization of the con
troller.
In the ﬁrst step, we diagonalize a given unitary matrix U
gate
∈ U(k) as
R
†
U
gate
R = U
diag
= diag(e
iγ
1
, . . . , e
iγ
k
) (97)
with R ∈ U(k). Each eigenvalue γ
j
is taken in the range 0 ≤ γ
j
< 2π. In
the second step, we combine single loop solutions associated with the Berry
phase to construct two k ×k matrices
Ω
diag
= diag(iω
1
, . . . , iω
k
), ω
j
= 2(π −γ
j
), (98)
W
diag
= diag(iτ
1
, . . . , iτ
k
), τ
j
= e
iφ
j
π
2
−(π −γ
j
)
2
. (99)
Then we obtain a diagonal controller
X
diag
=
Ω
diag
W
diag
−W
†
diag
0
.
In the third step, we construct the controller X as
X =
R 0
0 I
k
Ω
diag
W
diag
−W
†
diag
0
R
†
0
0 I
k
=
RΩ
diag
R
†
RW
diag
−W
†
diag
R
†
0
,
(100)
which is a 2k ×2k matrix. We call the set of equations, (97), (98), (99) and
(100), constructing equations of the controller. This is our main result.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
134
The diagonal controller X
diag
is actually a direct sum of controllers (96),
which generate Berry phases {e
iγ
j
}. Hence, its holonomy is also a direct sum
of the Berry phases (93) as
V
†
0
e
X
diag
V
0
e
−Ω
diag
= U
diag
and hence we have
V
†
0
e
X
V
0
e
−Ω
= RV
†
0
e
X
diag
V
0
R
†
Re
−Ω
diag
R
†
= RU
diag
R
†
= U
gate
.
6. Examples of Unitary Gates
To illustrate our method here we construct controllers for the onequbit
Hadamard gate and the twoqubit discrete Fourier transformation (DFT)
gate. For each unitary gate U
gate
, we need to calculate the diagonalizing
matrix R. Then the constructing equations of the controller, (97)(100),
provide the desired optimal controller matrices.
6.1. Hadamard gate
The Hadamard gate is a onequbit gate deﬁned as
U
Had
=
1
√
2
1 1
1 −1
. (101)
It is diagonalized by
R =
cos
π
8
−sin
π
8
sin
π
8
cos
π
8
, cos
π
8
=
2 +
√
2
2
, sin
π
8
=
2 −
√
2
2
(102)
as
R
†
U
Had
R =
1 0
0 −1
. (103)
Therefore, we have γ
1
= 0 and γ
2
= π. We may put φ
1
= φ
2
= 0. The
ingredients of the constructing equations of the controller, (97)(100), are
Ω
diag
= diag(2iπ, 0), W
diag
= diag(0, iπ), (104)
and hence
RΩ
diag
R
†
=
iπ
√
2
√
2 + 1 1
1
√
2 − 1
, RW
diag
=
iπ
2
0 −
2 −
√
2
0
2 +
√
2
.
(105)
Substituting these into (100), we obtain the optimal controller of the
Hadamard gate.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
135
6.2. DFT2 gate
Discrete Fourier transformation (DFT) gates are important in many quan
tum algorithms including Shor’s algorithm for integer factorization. The
twoqubit DFT (DFT2) is a unitary transformation
U
DFT2
=
1
2
1 1 1 1
1 i −1 −i
1 −1 1 −1
1 −i −1 i
. (106)
It is diagonalized by
R =
1
2
1
√
2 −1 0
1 0 1 −
√
2
−1
√
2 1 0
1 0 1
√
2
(107)
as
R
†
U
DFT2
R =
1 0 0 0
0 1 0 0
0 0 −1 0
0 0 0 i
. (108)
Therefore, we have γ
1
= γ
2
= 0, γ
3
= π and γ
4
= π/2. Thus the ingredients
of the controller are
Ω
diag
= diag(2iπ, 2iπ, 0, iπ), W
diag
= diag(0, 0, iπ, iπ
√
3/2), (109)
and hence
RΩ
diag
R
†
=
iπ
2
3 1 1 1
1 2 −1 0
1 −1 3 −1
1 0 −1 2
, RW
diag
=
iπ
2
0 0 −1 0
0 0 1 −
3/2
0 0 1 0
0 0 1
3/2
.
(110)
Substituting these into (100), we ﬁnally obtain the optimal controller of the
DFT2 gate.
7. Discussions
Although we provided a concrete prescription to calculate a controller of
the holonomic quantum computer to execute an arbitrary unitary gate,
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
136
there remain several problems for building a practical holonomic quantum
computer. In the rest of this lecture we shall discuss these issues.
7.1. Restricted control parameters
To solve the isoholonomic problem we assumed that the Stiefel manifold
is a control manifold. In other words, we assumed that we can change
the projection oprator P
1
(t) appearing in the Hamiltonian (53) arbitrarily.
However, realistic systems do not allow such a large controllability. Usually
a real system admits only a smaller control manifold; the number of con
trol parameters can be less and the ranges of their allowed values can be
narrow. Finding and optimizing a loop in the control manifold to generate
a demanded holonomy become a diﬃcult task for a realistic system.
7.2. Implementation
Here we describe a (not complete) list of proposals for implementing holo
nomic quantum computers although most of them still remain theoreti
cal proposals. Mead
16
proposed a use of Kramers degeneracy in molecules
to observe the nonAbelian holonomy. Zee
17
proposed a use of nuclear
quadrupole resonance. Unanyan et al.
18
proposed a use of fourlevel atoms
manipulated with laser as a holonomic quantum computer. Jones et al.
19
implemented geometric quantum computation with NMR. Duan et al.
20
proposed a use of trapped ions. Recati et al.
21
proposed a use of neu
tral atoms and the photon ﬁeld in cavity. Faoro et al.
22
proposed a use
of Josephson junction circuits. Solinas et al.
23
studied the holonomy in a
semiconductor quantum dots.
On the other hand, a lot of theoretical studies related to holonomic quan
tum computation can be found in literature. Karle and Pachos
24
discussed
holonomy in the Grassmann manifold, which is a special case of our study.
Niskanen et al.
25,26
studied numerical optimization of holonomic quantum
computer. Zhu and Zanardi
27
studied robustness holonomic quantum com
putation against control errors. Robustness against decoherence has been
analyzed, for example, by FuentesGuridi et al.
28
and a use of decoherence
free subspace also has been studied by Wu et al.
29
However, the most important issue for making a practical holonomic
quantum computer is to build a controllable system that has a degenerate
energy level; the degeneracy must be precise and the degenerate subspace
must have large dimensions to be used as a multiqubit processor. It seems
necessary to ﬁnd a breakthrough in this direction.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
137
Acknowledgements
The exact solution of the isoholonomic problem, which was explained in
this lecture, is a result of collaborations with Daisuke Hayashi and Mikio
Nakahara. I would like to thank Mikio Nakahara for stimulating discussions
with him. I also appreciate eﬀorts of the organizers of this successful sum
mer school, Mathematical Aspects of Quantum Computing, held at Kinki
University in August, 2007. This work is ﬁnancially supported by Japan
Society for the Promotion of Science, Grant Nos. 15540277 and 17540372.
References
1. S. Tanimura, D. Hayashi, and M. Nakahara, Phys. Lett. A 325 (2004) 199.
2. S. Tanimura, M. Nakahara, and D. Hayashi, J. Math. Phys. 46 (2005) 022101.
3. S. Kobayashi and K. Nomizu, Foundations of Diﬀerential Geometry, Vol. 2
(Interscience Publishers, New York, 1969).
4. M. Nakahara, Geometry, Topology and Physics, 2nd ed. (IOP Publishing,
Bristol and New York, 2003).
5. T. Kato, J. Phys. Soc. Japan 5 (1950) 435.
6. A. Messiah, Quantum Mechanics (Dover, New York, 2000).
7. M. V. Berry, Proc. R. Soc. Lond. A 392 (1984) 45.
8. A. Shapere and F. Wilczek (editors), Geometric Phases in Physics (World
Scientiﬁc, Singapore, 1989).
9. S. Tanimura, Tom and Berry (A thesis paper on holonomies in dynamical
systems written in Japanese), Soryushiron Kenkyu 85 (1992) 1.
10. A. Bohm, A. Mostafazadeh, H. Koizumi, Q. Niu, and J. Zwanziger, The
Geometric Phase in Quantum Systems (Springer, 2003).
11. F. Wilczek and A. Zee, Phys. Rev. Lett. 52 (1984) 2111.
12. P. Zanardi and M. Rasetti, Phys. Lett. A 264 (1999) 94.
13. J. Pachos, P. Zanardi, and M. Rasetti, Phys. Rev. A 61 (1999) 010305(R).
14. K. Fujii, J. Math. Phys. 41 (2000) 4406.
15. R. Montgomery, Commun. Math. Phys. 128 (1991) 565.
16. C. A. Mead, Phys. Rev. Lett. 59 (1987) 161.
17. A. Zee, Phys. Rev. A 38 (1988) 1.
18. R. G. Unanyan, B. W. Shore, and K. Bergmann, Phys. Rev. A 59 (1999)
2910.
19. J. Jones, V. Vedral, A. K. Ekert, and C. Castagnoli, Nature 403 (2000) 869.
20. L.M. Duan, J. I. Cirac, and P. Soller, Science 292 (2001) 1695.
21. A. Recati, T. Calarco, P. Zanardi, J. I. Cirac, and P. Zoller, Phys. Rev. A 66
(2002) 032309.
22. L. Faoro, J. Siewert, and R. Fazio, Phys. Rev. Lett. 90 (2003) 28301.
23. P. Solinas, P. Zanardi, N. Zanghi, and F. Rossi, Phys. Rev. B 67 (2003)
121307.
24. R. Karle and J. Pachos, J. Math. Phys. 44 (2003) 2463.
25. A. O. Niskanen, M. Nakahara, and M. M. Salomaa, Quantum Inf. Comput.
2 (2002) 560.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
138
26. A. O. Niskanen, M. Nakahara, and M. M. Salomaa, Phys. Rev. A 67 (2003)
012319.
27. S.L. Zhu and P. Zanardi, Phys. Rev. A 72 (2005) 020301.
28. I. FuentesGuridi, F. Girelli, and E. Livine, Phys. Rev. Lett. 94 (2005) 020503.
29. L.A. Wu, P. Zanardi, and D. A. Lidar, Phys. Rev. Lett. 95 (2005) 130501.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
139
PLAYING GAMES IN QUANTUM MECHANICAL
SETTINGS: FEATURES OF QUANTUM GAMES
S¸AHIN KAYA
¨
OZDEMIR, JUNICHI SHIMAMURA, NOBUYUKI IMOTO
Division of Materials Physics, Department of Materials Engineering Science, Graduate
School of Engineering Science, Osaka University, Toyonaka, Osaka 5608531, Japan
SORST Research Team for Interacting Carrier Electronics, 418 Honmachi,
Kawaguchi, Saitama 3310012, Japan
CREST Photonic Quantum Information Project, 418 Honmachi, Kawaguchi,
Saitama 3310012, Japan
Email: ozdemir@qi.mp.es.osakau.ac.jp
In this lecture note, we present the implications of playing classical games in
quantum mechanical settings where the quantum mechanical toolbox consisting
of entanglement, quantum operations and measurement is used. After a brief
introduction to the concepts of classical game theory and quantum mechanics,
we study quantum games and their corresponding classical analogues to deter
mine the novelties. In addition, we introduce a benchmark which attempts to
make a fair comparison of classical games and their quantum extensions. This
benchmark exploits the fact that in special settings a classical game should be
reproduced as a subgame of its quantum extension. We obtained a rather sur
prising result that this requirement prevents the use of a large set of entangled
states in quantum extension of classical games.
Keywords: Game Theory, Quantum Mechanics, Entanglement, Quantum
Games.
1. Introduction
Game theory is a mathematical modelling tool to formally describe con
ﬂict, cooperation or competition situations among rational and intelligent
decision makers who have diﬀerent preferences. Each of the decision makers
is motivated to increase his/her utility which is not solely determined by
his/her action but by the interaction of the strategies of all the decision
makers.
1
The start of modern game theory dates back to 1928 when J. von Neu
mann deﬁned and completely solved twoperson zerosum games.
2
Later
on, von Neumann introduced the concepts of linear programming and the
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
140
fundamentals of game theory which later became the building blocks for his
book with Morgenstern.
3,4
Although von Neumann made important contri
butions to quantum mechanics and game theory around the same years, the
two were developed as separate ﬁelds with no or very little interaction, and
with no common ground for applications. His attempts to elucidate the eco
nomical problems using physical examples faced strong criticism based on
the argument that social sciences cannot be modelled after physics because
the social sciences should take psychology into account. J. von Neumann’s
response to these criticism was ”such statements are at least premature.”.
3
Critiques, which say physics and social sciences are distinct ﬁelds and should
not mix, are still alive, and so is the response of von Neumann.
Game theory and its theoretical toolbox have been traditionally limited
to economical and evolutionary biology problems. However, recently math
ematical models and techniques of game theory have increasingly been used
by computer and information scientists,
5–11
Game theoretical models of eco
nomics, evolutionary biology, population dynamics, large scale distribution
systems, learning models and resource allocation problems have been stud
ied. Proposals have been done for modelling complex physical phenomena
such as decoherence and irreversibility using game theoretical models.
12,13
If we bear in mind that a game is about the transfer of information, then it
is not a surprise that all those distinct areas meet in game theory: Informa
tion constitutes a common ground and a strong connection among them.
14
Then one naturally wonders what happens if the information carriers are
taken to be quantum systems. This leads to the introduction of the quan
tum mechanical toolbox into game theory to combine two distinct interests
of von Neumann in an environment to mutually beneﬁt from both, and to
see what new features will arise from this combination.
Quantum mechanical toolbox (unitary operators, superposition and en
tangled states, measurement) was introduced in game theory, for the ﬁrst
time, by Meyer
15
who showed that in a coin tossing game a player utilizing
quantum superposition could win with certainty against a classical player.
Following this ﬁrst paper, Eisert et al. studied the wellknown Prisoners’
dilemma (PD) using a general protocol for two playertwo strategy games
where the players share entanglement.
16
These two works raised an excite
ment in the quantum mechanics community mostly due to emergence of a
new ﬁeld where quantum mechanics shows its advantage and also due to the
hope of using game theoretical toolbox to study quantum information tasks
and quantum algorithms. Since then most of the works has been devoted to
the analysis of dilemma containing classical game dynamics using Eisert’s
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
141
scheme and its multiparty extension. Although some quantum communi
cation protocols and algorithms
17,18
were reformulated in the language of
game theory, this area still begs for more exploration. In the mean time,
quantum versions of some important theories, e.g., minmax theorem, have
been developed.
19
In this review, we will brieﬂy introduce the terminology and basic con
cepts of classical game theory. Then we will discuss Eisert’s scheme and
focus on game dynamics with respect to the changes in the quantum me
chanical resources used. We will derive a condition that should be satisﬁed
for an entangled state to be useful in quantum version of classical games.
Finally, we will outline the results.
2. A Brief Review of Classical Game Theory
The ingredients of a game theory problem include:
1
(i) Players: Two or
more goal oriented, rational and intelligent decision makers. (ii) Strategies:
Feasible action space of players. Players can choose to play either in pure
or mixed strategies. In pure strategy, each player chooses an action with
unit probability, whereas in mixed strategy the players randomize among
their actions. (iii) Payoﬀs: Beneﬁts, prizes, or awards distributed to players
depending on the choices of all players. Payoﬀ to a player is determined by
the joint action of all players.
Rationality of players and common knowledge are two important as
sumptions in game theory. Rationality implies that each player aims at
maximizing his own payoﬀ, and that all rational players search for an equi
librium on which all can settle down. The common knowledge assumption
requires that all the players know the rules of the game and that all the
players know that all players know the rules of the game.
In game theory, games are generally modelled either in strategic form or
in extensive form. If the players take their actions simultaneously or without
observing the actions taken by other players, the strategic form, which is
best represented by a payoﬀ matrix, is preferred. On the other hand, if the
players take their actions sequentially after observing the actions taken by
the players before them, it is best to model using extensive form, which is
represented by a game tree. The extensive form requires more speciﬁcations
such as which player moves when doing what and with what information.
In this work, we consider only the games with strategic form.
Games are usually grouped as cooperative or noncooperative games. In
the ﬁrst one, players are allowed to form coalitions and combine their de
cision making problems with binding commitments, whereas in the latter
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
142
Table 1. Twoplayer twostrategy game
in strategic form.
Bob:B
1
Bob: B
2
Alice: A
1
(a, w) (b, x)
Alice: A
2
(c, y) (d, z)
each player pursues his/her own interests with no intention to negotiate.
Games can also be classiﬁed according to the structure of their payoﬀ
matrices as symmetric or asymmetric, zero sum or nonzero sum, and co
ordination (at least one equilibrium in pure strategies) or discoordination
games. In Table 1, payoﬀ matrix of a generic twoplayer twostrategy game
is given. The players, Alice (row player) and Bob (column player), have the
action sets {A
1
, A
2
} and {B
1
, B
2
}. The payoﬀs are represented as (.., ..)
with the ﬁrst element being Alice’s payoﬀ and the second Bob’s. This pay
oﬀ matrix represents a symmetric game if a = w, b = y, c = x, d = z,, a
zerosum game if a + w = b + x = c + y = d + z = 0, a constantsum game
if a + w = b + x = c + y = d + z = K = 0 and a discoordination game if
c > a, b > d, w > x, z > y or a > c, d > b, x > w, y > z.
A systematic description of the outcomes in a gametheoretic problem
is called a solution. Optimality and equilibria are the important solution
concepts. In some games, there is an optimal way of taking an action inde
pendent of the actions of other players. A strategy is Pareto optimal (PO) if
the only way each player has to increase his/her payoﬀ is decreasing some
one else’s payoﬀ. Nash Equilibrium (NE) is the most widely used solution
concept in game theory. Nash Equilibrium is a position where all the strate
gies are mutually optimal responses. This means that no unilateral change
of strategy will give a higher payoﬀ to the corresponding player. Therefore,
once it is reached, no individual has an incentive to deviate from it unilat
erally. However NE concept, too, has diﬃculties: (i) Not every game has an
NE in pure strategies, (ii) An NE need not be the best solution, or even a
reasonable solution for a game. It is merely a stable solution against unilat
eral moves by a single player, and (iii) NE need not be unique. This prevents
a sharp prediction with regard to the actual play of the game. In Fig.1, we
give examples of classical games with diﬀerent types of dilemmas.
1,20
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
143
(3,3)
(5,0)
(0,5)
(1,1)
Prisoners’ Dilemma
(1,2)
(0,0)
(2,1)
(0,0)
Battle of Sexes
(3,2)
(1,1)
(1,3)
(0,0)
Samaritan s Dilemma
* *
*
♦
(3,3)
(5,0)
(0,5)
(1,1) (1,2)
(0,0)
(2,1)
(0,0)
(3,2)
(1,1)
(1,3)
(0,0)
(3,2)
(1,1)
(1,3)
(0,0)
* *
*
♦
Fig. 1. Payoﬀ matrices of games with diﬀerent types of dilemmas: Samaritan’s dilemma
(SD) with no Nash equilibrium (NE), Prisoners’ dilemma (PD) where NE and Pareto
optimality (PO) do not coincide, and the Battle of Sexes (BoS) with two NE’s. Horizontal
(vertical) arrows correspond to the best responses of the column (row) player for a given
strategy of the row (column) player. The entries pointed by two arrows and labelled with
∗ are the Nash equilibria. The Pareto optimal entry in the PD game is labelled with .
3. Basic Concepts of Quantum Mechanics for a Study of
Games
In this section, we will introduce some basic concepts of quantum mechan
ical toolbox needed for the study of quantum versions of classical games.
Reference
21
is an excellent resource for a detailed study of quantum me
chanics and quantum information science.
3.1. Quantum bits
The state of a quantum system is described by a vector in Hilbert space. The
simplest quantum mechanical system has a two dimensional state space, and
can be represented as  ψ = α 0 + β 1 where 0 and 1 are basis states
(vectors), and α and β are complex numbers satisfying α
2
+β
2
= 1. This
simplest system is called as quantum bit or qubit. While the classical bits can
be in one of two states 0 or 1 which can in principle be distinguished, a qubit
is in a continuum of states described by α and β and it is not possible, in
general, to reliably distinguish nonorthogonal states of a quantum system.
If a measurement is performed on the qubit, the outcome will be  0 with
probability α
2
, and it will be  1 with probability β
2
. A discussion of
this measurement is given in subsection 3.4. Note that the normalization
ensures that some result is obtained with probability 1.
3.2. Density matrices
This is a frequently used formalism in quantum mechanics. Suppose we have
an ensemble of possible states { ψ
1
,  ψ
2
, . . . ,  ψ
n
} with probabilities p
i
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
144
This ensemble is a mixture of all  ψ
i
’s and its density matrix is deﬁned as
ˆ ρ =
n
i=1
p
i
 ψ
i
ψ
i
. (1)
It is important to notice that this is a mixture but not a superposition of
 ψ
i
’s, but each of  ψ
i
’s may be a superposition state. A density matrix
may be formed from many diﬀerent ensembles, e.g., the ensemble of  0 and
 1 with equal probabilities and the ensemble of the superposition states
( 0 +  1)/
√
2 and ( 0 −  1)/
√
2 with equal probabilities have the same
density matrix. No measurement can distinguish these two ensembles if no
additional information is given about the preparation.
3.3. Unitary transformation
The evolution of an isolated quantum system is governed by unitary oper
ator
ˆ
U which takes the state  ψ at time t = t
0
to the state  ψ
at time
t = t
1
. This evolution can be represented as  ψ
=
ˆ
U ψ. The evolution of
a density matrix ˆ ρ is described by ˆ ρ
=
ˆ
Uˆ ρ
ˆ
U
†
.
A unitary operator on a qubit can be written as
ˆ
U = e
iφ
e
iγ
cos(
θ
2
) e
iδ
sin(
θ
2
)
−e
−iδ
sin(
θ
2
) e
−iγ
cos(
θ
2
)
. (2)
Unitary operators without the global phase e
iφ
belong to SU(2) operators
set which will be mainly used in this study. The operators known as Pauli
operators play an important role in quantum information, computation and
games. These operators are
ˆ σ
0
=
1 0
0 1
, ˆ σ
x
=
0 1
1 0
, ˆ σ
y
=
0 − i
i 0
, ˆ σ
z
=
1 0
0 − 1
. (3)
The identity operator
ˆ
I = ˆ σ
0
leaves the qubits invariant. The bit ﬂip oper
ator ˆ σ
x
transforms  0 to  1, and  1 to  0. ˆ σ
z
is the phase ﬂip operator
which takes  1 to − 1 and does nothing on  0. The ˆ σ
y
Pauli operator
performs both a phase ﬂip and a bit ﬂip which takes  0 to i 1 and  1
to −i 0. Another important operator that we will frequently use is the
Hadamard operator
ˆ
H =
1
√
2
1 1
1 −1
, (4)
which takes  0 to ( 0 +  1)/
√
2 and  1 to ( 0 −  1)/
√
2.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
145
3.4. Measurement
In order to gather information from a quantum mechanical system we have
to perform measurement. Quantum measurements are described by a set
of measurement operators {M
i
} acting on the state space to be measured.
The index i is associated with the measurement outcome . If the state is
 ψ before measurement, then the post measurement state is
 ψ
i
=
M
i
 ψ
ψ M
†
i
M
i
 ψ
(5)
with the measurement result i occuring with probability p
i
= ψ M
†
i
M
i
 ψ.
The measurement operators satisfy the completeness equation
i
M
†
i
M
i
=
ˆ
I. As an example let us measure the qubit  ψ = α 0 +β 1 in the compu
tational basis which is described by two measurement operators M
0
= 00
and M
1
= 11. First, observe that the measurement operators are Hermi
tian, M
0
= M
†
0
and M
1
= M
†
1
, and they satisfy the completeness equation
M
†
0
M
0
+M
†
1
M
1
= M
2
0
+M
2
1
= M
0
+M
1
=
ˆ
I. Then the probability of obtain
ing zero as the measurement outcome is p
0
= ψ M
†
0
M
0
 ψ = ψ M
0
 ψ =
α
2
. In the same way, the probability of obtaining the measurement out
come is p
1
= ψ M
†
1
M
1
 ψ = ψ M
1
 ψ = β
2
, just as given in subsection
3.1. The post measurement states then are the  0 for M
0
and  1 for M
1
.
A special class of the general measurement described above is the pro
jective measurements. A projective measurement is deﬁned by an observ
able which is a hermitian operator M acting on the state space to be
measured. Upon measuring the system  ψ with M, we obtain an out
come m
i
which is an eigenvalue of M. If the projector onto the eigenspace
of M with eigenvalue m
i
is P
i
, then the observable has a spectral de
composition M =
i
m
i
P
i
. The probability of getting the result m
i
is given by p
i
= ψ P
i
 ψ. Then the post measurement state becomes
 ψ
i
= P
i
 ψ/
√
p
i
.
For some applications, we are not very much interested in the post
measurement state but rather on the probabilities of the respective measure
ment outcomes. For such situations, a mathematical tool called as POVM
(positive operator valued measure) is used. POVM is a collection of positive
operators {E
i
} with the completeness condition
i
E
i
=
ˆ
I. The probability
of the measurement outcome is then given by p
i
= Tr[ ψψ E
i
].
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
146
3.5. Correlations in quantum mechanical systems
Analysis of quantum games presumes that the players initially share some
kind of correlation but they are not allowed to have any type of communica
tion. Thus, it is needed to deﬁne the correlations in a quantum mechanical
system. We start by giving the mathematical representation of an uncorre
lated state deﬁned as ˆ ρ
u
= ˆ ρ
1
⊗ ˆ ρ
2
⊗· · · ⊗ ˆ ρ
n−1
⊗ ˆ ρ
n
where measurement of
observables on respective subsystems always factorize. On the other hand,
a classically correlated system is described as ˆ ρ
c
=
i
p
i
ˆ ρ
(i)
1
⊗· · · ⊗ˆ ρ
(i)
n
with
i
p
i
= 1. Unlike the uncorrelated system, joint measurement of observ
ables for such a system do not factorize.
In quantum mechanical system, there is another type of correlation
which cannot be seen in classical systems. This is the quantum correla
tion or socalled entanglement. A state is entangled if it cannot be writ
ten as a convex combination of product states. An entangled state cannot
be prepared by local interactions even if the preparers can communicate
classically. In entangled systems, one cannot acquire all information on
the system by accessing only some of the subsytems. For a bipartite sys
tem, there are four maximally entangled states (MES) which are deﬁned
as Φ
±
= (00 ± 11)/
√
2 and Ψ
±
= (01 ± 10)/
√
2. Starting from
one of these bipartite MES, any bipartite state can be prepared using local
operations and classical communication (LOCC).
On the other hand, in the multipartite case (N ≥ 3), it is known that
there are inequivalent classes of states such as those represented by the N
particle Wstates  W
N
=  N −1, 1/
√
N and the Nparticle Greenberger
HorneZeilinger (GHZ) states  GHZ
N
= ( N, 0 +  0, N)/
√
2. Here
 N − k, k corresponds to a symmetric state with N − k zeros and k
ones, e.g.,  W
3
=  2, 1/
√
3 = ( 001 + 010 + 100)/
√
3 and  GHZ
3
=
( 3, 0 +  0, 3)/
√
2 = ( 000 +  111)/
√
2. In addition to these classes,
there is the class of cluster states  χ
N
which includes  GHZ
3
but forms
another inequivalent class for N ≥ 4. Those inequivalent classes cannot
be transformed into each other by LOCC. Although the total amount of
entanglement contained in  W
N
is smaller than the other two classes, it
has higher persistency of entanglement which is deﬁned as the minimum
number of local measurements to completely destroy entanglement. It has
been shown that  W
N
is optimal in the amount of pairwise entanglement
when N − 2 particles are discarded,
22,23
and has a persistency of N − 1,
whereas GHZ and cluster states have one and N/2, respectively.
24
The states  Ψ
+
and  W
N
are the members of a more general class of
symmetric states called as Dicke states,  N − m, m/
√
N
C
m
where
N
C
m
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
147
denotes the binomial coeﬃcient.
25,26
One can easily see that GHZ states
can be written as a superposition of two Dicke states  N, 0 and  0, N .
4. A Brief Review of Quantum Game Theory: Models and
Present Status
In this section, we will present a short historical background of quantum
game theory, and introduce, in detail, two quantization schemes which have
been the milestones in this area of research. These schemes are Meyer’s
scheme for coin tossing game
15
and the Eisert’s scheme for Prisoners’
dilemma (PD).
16
The transfer of the classical games into the quantum
domain has been done using the resources allowed by the quantum me
chanical toolbox which includes unitary operators, superposition states,
entanglement and quantum measurement.
Q wins (+1)
Q loses (1)
Flip /
No Flip
Q
P Q
H,
T,
Coin: T
Flip /
No Flip
Flip /
No Flip
{
Q always wins
(+1)
Q
P Q
Coin:
T
Fig. 2. Classical (upper) and quantum (lower) versions of the coin ﬂip game. Coin is
initially prepared in tails state (T). After the turns of P and Q, coin is analyzed. If it is
tails (T), Q wins; otherwise Q loses. While in classical version strategy set players have
the actions of Flip and No Flip, in the quantum version Q has the quantum strategy (op
erator)
ˆ
H while P has the classical strategies with the corresponding quantum operators
{
ˆ
X,
ˆ
I}.
4.1. Coin ﬂip game
The ﬁrst strong argument for a quantum game theory was presented by
Meyer who introduced the coin ﬂip game between two Star Trek characters,
Q and P.
15
In this simple game (see Fig. 2) P prepares a coin in the ”tails”
state and puts it into a box. Then Q can choose either to ﬂip the coin or
leave it unchanged without knowing the initial state of the coin. In the same
way, P, without knowing the action of Q, can choose either of the actions.
After the second turn of Q, the coin is examined. If the coin is tails, Q
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
148
wins and if it is heads Q loses. A classical coin clearly gives both players an
equal probability of success. As can be seen from the game payoﬀ matrix in
Table 2, there is no Nash equilibrium in the pure strategies, and the best
the players can do is a mixed strategy.
In Meyers description of coin ﬂip game, a quantum coin with states
deﬁned as T =  0 and H =  1 is used, and P is restricted to classical
strategies {Flip =
ˆ
X, No Flip =
ˆ
I} while Q has the quantum strategies
{
ˆ
X,
ˆ
H}. Then the game sequences as follows
 0
ˆ
H
−→
Q
2
−1/2
( 0 +  1)
ˆ
X or
ˆ
I
−→
P
2
−1/2
( 0 +  1)
ˆ
H
−→
Q
 0 (6)
from which it is seen that just applying a Hadamard gate before and after
P’s action, Q brings the coin into the tail and hence wins the game regard
less of the action of P. This work clearly shows that players with quantum
strategies perform better than the players restricted to classical strategies.
Table 2. Payoﬀ matrix for the classical coin ﬂip game where
‡ depicts the best strategy combination of Q for a given action
of P.
Q:NN Q:NF Q:FN Q:FF
P: N (1, −1) (−1, 1)‡ (−1, 1)‡ (1, −1)
P: F (−1, 1)‡ (1, −1) (1, −1) (−1, 1)‡
4.2. Eisert’s model  Prisoners’ Dilemma
In the same year as Meyer’s work,
15
Eisert et al. studied a quantum version
of PD where two players sharing an entangled state use quantum strate
gies.
16
Before going into the quantization of this game, let us ﬁrst review
the classical one: The socalled Prisoners’ Dilemma game is an important
and wellstudied two player two strategy symmetric game where the play
ers Alice and Bob independently without knowing the action of the other
player decide whether to ”cooperate” (C) or ”defect” (D). According to
their combined strategy choices, they obtain a certain payoﬀ determined
by the game payoﬀ matrix (see Fig. 1 for the payoﬀ matrix). They then
face a dilemma since as rational players they should look for an equilibrium
which dictates that they should both defect. On the other hand, they would
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
149
Referee
Entangled state preparation
Referee
Disentangling and measurement
Bob's action
Alice's action
Fig. 3. Schematic conﬁguration of Eisert’s scheme to play twoplayer strategic games.
The strategy set of the players are unitary operators
ˆ
U
A
and
ˆ
U
B
, respectively for the
players Alice and Bob. An entangled state is prepared by the referee using the entangling
operator
ˆ
J, and is distributed to players. After the action of the players referee performs
measurement to distribute payoﬀs.
both beneﬁt from cooperation which is the Pareto optimal. Their dilemma
originates from the fact that NE and PO do not coincide.
In Eisert’s model (see Fig.3), a twoplayer twostrategy game is played
in the quantum mechanical settings as follows: (i) A referee prepares the
maximally entangled state  Ψ
i
by applying the entangling operator
ˆ
J on
a product state  fg where
ˆ
Jfg =
1
√
2
[ fg + i(−1)
(f+g)
(1 − f)(1 − g) ] (7)
with f, g = 0, 1. Then the referee sends this state to two players, Alice and
Bob, one qubit for each player. (ii) As their strategies, Alice and Bob carry
out unitary operations U
A
, U
B
∈ SU(2) described by
ˆ
U(θ, φ, ψ) =
e
iφ
cos(
θ
2
) e
iψ
sin(
θ
2
)
−e
−iψ
sin(
θ
2
) e
−iφ
cos(
θ
2
)
(8)
with 0 ≤ θ ≤ π, 0 ≤ φ ≤ π/2, and 0 ≤ ψ ≤ π/2. They then send their
qubits to the referees. (iii) After the referee receives the state, he oper
ates
ˆ
J
†
obtaining  Ψ
f
=
ˆ
J
†
(
ˆ
U
A
⊗
ˆ
U
B
)
ˆ
J fg. (iv) The referee makes a
projective measurement Π
n
= jj
{j,=0,1}
with n = 2j + correspond
ing to the projection onto the orthonormal basis {00, 01, 10, 11}. In
PD game, the measurement outcome { 00,  01,  10,  11} corresponds to
{(C, C), (C, D), (D, C), (D, D)} where left and right entry represent Alice’s
and Bob’s decision, respectively. (v) According to the measurement outcome
n, the referee assigns to each player the payoﬀ a
n
and b
n
chosen from the
payoﬀ matrix of the original classical game, e.g., a
{n=0,1,2,3}
= {3, 0, 5, 1}
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
150
and b
{n=0,1,2,3}
= {3, 5, 0, 1} in the PD game for Alice and Bob, respec
tively. Then the average payoﬀ of the players can be written as
$
A
=
n
a
n
Tr(Π
n
ˆ
J
†
ˆ ρ
f
ˆ
J)
P
j
,
$
B
=
n
b
n
Tr(Π
n
ˆ
J
†
ˆ ρ
f
ˆ
J)
P
j
(9)
with ˆ ρ
f
=  Ψ
f
Ψ
f
 and P
j
representing the probability of obtaining
the measurement outcome n. When the referee starts with the input state
 fg =  00, the probabilities P
j
is found as
P
00
= x
cos(φ
A
+ φ
B
) − y
sin(ϕ
A
+ ϕ
B
)
2
P
01
= xsin(φ
B
− ϕ
A
) − y cos(φ
A
− ϕ
B
)
2
P
10
= xcos(φ
B
− ϕ
A
) − y sin(φ
A
− ϕ
B
)
2
P
11
= x
sin(φ
A
+ φ
B
) + y
cos(ϕ
A
+ ϕ
B
)
2
(10)
where x = sin
θ
A
2
cos
θ
B
2
, y = sin
θ
B
2
cos
θ
A
2
, x
= cos
θ
A
2
cos
θ
B
2
, and y
=
sin
θ
A
2
sin
θ
B
2
.
Note that original classical games should be simulated in a model of
its quantum version, and in Eisert’s model this is satisﬁed by imposing the
following conditions:
[
ˆ
J, ˆ σ
y
⊗ iˆ σ
y
] = 0, [
ˆ
J, ˆ σ
0
⊗ iˆ σ
y
] = 0, [
ˆ
J, iˆ σ
y
⊗ ˆ σ
0
] = 0 (11)
These conditions imply that the unitary operator
ˆ
U(θ, 0, 0) = cos(θ/2)ˆ σ
0
+
i sin(θ/2)ˆ σ
y
and
ˆ
J commute. Hence when players’ strategies are restricted
to
ˆ
U(θ), the order of operations
ˆ
J and
ˆ
U
A
⊗
ˆ
U
B
can be changed and then
ˆ
J and
ˆ
J
†
are cancelled, resulting in the state  Ψ
f
= (
ˆ
U
A
⊗
ˆ
U
B
) 00. In
this case, the joint probability P
j
factorize P
j
= p
A
(j)p
B
(). For PD
game p(C) = cos
2
(θ/2) and p(D) = 1 − p(C), and the payoﬀ function
takes the form of the classical payoﬀ function. This implies that ˆ σ
0
and iˆ σ
y
correspond to the pure strategies C and D, and the setting can simulate
the classical game faithfully. Therefore, when the players take
ˆ
U(θ, 0, 0) as
their strategies, the original Prisoners’ dilemma is reproduced.
In this model, when players’ strategy space is restricted to only one
parameter unitary operator
ˆ
U(θ, 0, 0), the payoﬀ functions become identical
to those of the classical Prisoners Dilemma with mixed strategies. This
equilibrium point gives 1 to each player which is less than the optimal
one with 3 to each. On the other hand, when the players are allowed to use
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
151
twoparameter unitary operators
ˆ
U(θ, φ, 0), a unique Nash equilibrium with
payoﬀ (3,3), which is also the Pareto optimal, emerges for their strategies
ˆ
U(θ = 0, φ = π/2, 0). Thus, it is claimed that by introducing quantum
toolbox, the dilemma in this game disappears and players can achieve the
Pareto optimal solution.
Benjamin and Hayden, however, criticized this restricted operation set
to be unnatural and pointed out that if two players are allowed the full
range of quantum strategies, there exists no Nash equilibrium in the game
because the eﬀect of one player’s unitary transformation can be cancelled by
the other locally using a counterstrategy due to the shared bipartite maxi
mally entangled state.
27
This leads to a point that there is no compromised
strategy. However, their argument does not hold when a nonmaximally en
tangled state is shared between the players, and when the game is extended
to multiplayer form. Benjamin and Hayden have also examined three and
four player quantum games which are strategically richer than the two
player ones.
28
They constructed a threeplayer PD game and showed that
the game has an NE even when the players are allowed to use full range
of unitary operators. We should note that despite the critiques, the new
dynamics brought into the games by quantum operators and entanglement
in Eisert’s model is clear though sometimes results are rather surprising.
4.3. Present status in quantum game theory
It is important to realize that quantization of a classical game is not unique;
one can introduce quantumness into a game in diﬀerent ways. The model
proposed by Eisert et al. has been the widely accepted and studied quanti
zation scheme despite some negative critiques as we have discussed above.
Another way of achieving results similar to those of Eisert’s scheme was
introduced by Marinatto and Weber
29
who removed the disentangling op
erator
ˆ
J
†
and simply hypothesized various initial states. In this model,
strategies are chosen as the probabilistic mixtures of identity and ﬂip op
erators. Although the scheme is much simpler it does not allow probing
the whole set of quantum strategies. Later, Nawaz and Toor
30
proposed a
more general model which includes both the Eisert and Marinatto models.
In this model, they introduced two entangling parameters (one for
ˆ
J and
the other for
ˆ
J
†
) which can be tuned. If the two parameters are equal we
obtain the Eisert scheme, and when the parameter for
ˆ
J
†
is zero, we end
up with the Marinatto schemes. Cheon and Tsutsui introduced a simple
and general formulation of quantum game theory by accommodating all
possible strategies in the Hilbert space, and solved two strategy quantum
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
152
games.
31
Recently, Ichikawa et al. proposed a formulation of quantum game
theory based on the Schmidt decomposition where they could quantify the
entanglement of quantum strategies.
32
The model proposed by Eisert et al. is a general one that can be applied
to any twoplayer twostrategy game, with the generalization to Nplayer
twostrategy games as well as twoplayer nstrategy games. While in the
former, the players share multipartite entangled state and use SU(2) op
erators, in the latter they share bipartite entanglement but armed with
SU(n) operators to represent the players’ actions. Multiplayer games ex
hibit a richer dynamics than the two player games. While such games can
be realized by preparing initially a multipartite entangled state, Chen et
al. has proposed to play using only bipartite entangled states shared by
neighboring players.
33
In order to investigate the property of quantum version of dynamic
games, Iqbal and Toor analyzed quantum version of repeated Prisoners’
dilemma game where the players decide to cooperate in the ﬁrst stage while
knowing that both will defect in the second stage.
34
Kay et al. studied
evolutionary quantum games and demonstrated that length of memory is
crucial to obtain and maintain the quantum advantage.
35
Iqbal has also
considered evolutionarily stable strategies (ESS) in quantum versions of
both the prisoners dilemma and the battle of the sexes and concluded that
entanglement can be made to produce or eliminate ESS’s while retaining
the same set of NE.
36
He also showed that it is only when entanglement
is employed, a mutant strategy with quantum tools can easily invade a
classical ESS. In these models the replicator dynamic takes a quantum
form.
37,38
Various quantum advantages over the classical have been found in the
series of studies on games in quantum settings. Comparative studies by
Shimamura et al.
20
and Ozdemir et al.
39
on the eﬀects of classical/quantum
correlations and operators not only conﬁrmed the known eﬀects of using the
quantum mechanical toolbox in game theory but also revealed that in some
circumstances one need not stick to quantum operations or correlations
but rather use simple classical correlations and operations. Based on the
requirement that classical game should be reproduced in its quantum model,
they derived a necessary and suﬃcient condition for the entangled states
and quantum operations which are allowed in playing games in quantum
mechanical settings.
40,41
Piotrowski and Sladkowski have proposed a quantum game theoretic
approach to economics.
42–44
In their model of market, transactions are de
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
153
scribed as projective operations on Hilbert spaces of strategies of traders,
and a quantum strategy is a superposition of trading actions and hence can
achieve outcomes not realizable by classical means. They also speculated
that markets will be more eﬃcient with quantum algorithms and dramatic
market reversals can be avoided.
There has been a growing interest in the quantum version of the well
known Parrondo’s games,
13,45,46
where two games that are losing when
played separately become a winning game when played in combination
(their convex combination). It is expected that Parrondo games have the po
tential use in increasing the eﬃciencies of quantum algorithms. Flitney and
Abbott have developed models and studied the history dependent quantum
versions of these games.
47,48
Physical realizations of Eisert’s scheme of PD have been achieved ﬁrst in
a two qubit NMR computer
49
and recently in an alloptical oneway quan
tum computer.
50
The NMR experiment followed the original proposal by
Eisert, and showed good agreement with theory. The amount of entangle
ment between the players was varied and the transition between classical
and quantum regions were clearly observed. The alloptical realization was
performed following the scheme of Paternostro et al.
51
This scheme is a
hybrid one combining the quantum circuit approach and the cluster state
model. The quantum circuit was realized by a 4qubit boxcluster conﬁgu
ration, and the local strategies of the players by measurements performed
on the physical qubits of the cluster.
Here, we gave only a short summary of what has been done in the ﬁeld.
Detailed and excellent reviews of quantum game theory can be found in the
references.
40,52–54
5. Study of a Discoordination Game  Samaritan’s
Dilemma  Using Eisert’s Model: Eﬀects of Shared
Correlation on the Game Dynamics
Samaritan’s Dilemma (SD), which is a discoordination game with no Nash
equilibrium in classical pure strategies, is an interesting game where one
can systematically analyze the changes in the game dynamics for various
quantum mechanical settings. Games with no NE are interesting because
they represent situations in which individual players might never settle
down to a stable solution. Therefore, the important step in such games
is to ﬁnd an NE on which the players can settle down. In the following,
we will investigate this game under diﬀerent settings of Eisert’s model. In
particular we will see how the game payoﬀ matrix and the strategies of the
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
154
players evolve for diﬀerent strategy sets and shared correlation between the
players. A more detailed analysis can be found in References 39 and 40.
In the game of SD, the “Samaritan” (Alice) voluntarily wants to help
the person ”in need” (Bob) who cannot help himself. However, she suﬀers a
welfare loss due to the selﬁsh behavior of Bob who may inﬂuence or create
situations which will evoke Alice’s help. Then, a dilemma arises because Al
ice wants to help, however, Bob acts so strategically that the amount of help
increases which is not desirable for Alice. Moreover, Alice cannot retaliate
to minimize or stop this exploitation because doing so is a punishment for
Bob and this will harm Alice’s own interests in the short run.
55
The game
and the payoﬀ matrix (see Fig. 1) studied in this paper are taken from Ref.1
where the speciﬁc game is named as The Welfare Game. In this game, Alice
(row player in the game matrix) wishes to aid Bob (column player in the
matrix) if he searches for work but not otherwise. On the other hand, Bob
searches for work if he cannot get aid from Alice. The action set of Alice is
{Aid (A), No Aid (N)} whereas Bob’s is {Work (W), Loaf (L)}. Thus their
strategy combinations can be listed as (A, W), (A, L), (N, W), and (N, L).
When this game is played classically, there is no dominant strategy for
neither of the players and there is no NE for pure strategies: (A,W) is not
an NE because if Alice chooses A, Bob can respond with strategy L where
he gets a better payoﬀ, 3, as shown with arrow in Fig. 1. (A,L) is not an
NE because, in this case, Alice will switch to N. The strategies (N,L) and
(N,W) are not NE either, because for the former one Bob will switch to
W to get payoﬀ one, whereas for the latter case Alice will switch to A to
increase her payoﬀ from −1 to 3. Therefore, this game has no NE when
played with pure classical strategies.
In mixed classical strategies, Alice randomize between her strategies A
and N with the probabilities p and (1 − p), respectively. In the same way,
Bob’s chooses between W and L with probabilities q and (1 −q). Then the
payoﬀs for Alice and Bob is written as
$
A
= 3pq − p(1 − q) − q(1 − p),
$
B
= 2pq + 3p(1 − q) + q(1 − p). (12)
The strategy combination with p = 0.5 and q = 0.2 corresponds to the NE
for the game with average payoﬀs given as $
A
= −0.2 and $
B
= 1.5. In this
case, the payoﬀ of Alice is negative which is not a desirable result for her.
(A,L) and (N,L) emerge as the most probable strategies with probabilities
0.4.
In the study of this game, we ﬁrst look at the problem whether a unique
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
155
NE exists for strategies utilizing classical or quantum mechanical toolboxes.
If there exists at least one NE then we look at how the payoﬀ of the players
compare and to what extent this NE strategy can resolve the dilemma. For
SD, there are three cases: CASE I, insuﬃcient solution with $
A
< 0, CASE
II, weak solution with 0 ≤ $
A
≤ $
B
, and CASE III, the strong solution
solution with 0 ≤ $
B
< $
A
. If at the NE, both players receive positive
payoﬀs then this implies both are satisﬁed with the outcome and there is
no loss of resources for Alice (CASE II and III). The most desirable solution
of the dilemma for Alice is represented in CASE III.
5.1. Quantum operations and quantum correlations
When the SD game is played with quantum operations but with no shared
correlated state between the players, the situation produces the classical
game with mixed strategies. Moreover, we know from Eisert’s study that
operators chosen from three parameter SU(2) set do not resolve the dilem
mas in the games. Therefore, in the following we will not study these two
cases. We will consider two types of MES: ψ
1
= [00 + i11]/
√
2 and
ψ
2
= [01 − i10]/
√
2 obtained when the referee applies the entangling
operator
ˆ
J on the initial states 00 and 11, respectively.
Oneparameter SU(2) operators  Operators for Alice and Bob are ob
tained by setting φ
A
= ϕ
A
= 0 and φ
B
= ϕ
B
= 0 in Eq.(8). With the
shared entangled state  ψ
1
and the oneparameter SU(2) operators, the
payoﬀs for Alice and Bob become
$
A
=
1
4
[1 + 3(cos θ
A
+ cos θ
B
) + 5 cos θ
A
cos θ
B
],
$
B
=
1
2
[3 + 2 cos θ
A
−cos θ
A
cos θ
B
]. (13)
From Eq.(13), we see that while $
B
is always positive, $
A
may be positive
or negative depending on Bob’s action. This implies that the samaritan,
Alice, cannot make her payoﬀ positive by acting unilaterally, thus Bob’s
choice of action becomes important to resolve Alice’s dilemma. After some
straightforward calculations, we ﬁnd that an NE emerges for the operators
ˆ
U
A
= (ˆ σ
0
+ iˆ σ
y
)/
√
2 and
ˆ
U
B
= (ˆ σ
0
+ i2ˆ σ
y
)/
√
5. At this unique NE, the
payoﬀs of the players become $
A
= −0.2 and $
B
= 1.5. On the other
hand, when they share ψ
2
, an NE is found for
ˆ
U
A
= (ˆ σ
0
+ iˆ σ
y
)/
√
2 and
ˆ
U
B
= (2ˆ σ
0
+ iˆ σ
y
)/
√
5 with the payoﬀs $
A
= −0.2 and $
B
= 1.5. It is seen
that this strategy set reproduced the results of the classical mixed strategy,
and does not bring any quantum advantage. This strategy set resolves the
dilemma, but with an insuﬃcient solution for Alice (CASE I).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
156
Twoparameter SU(2) operators  This set of quantum operations is ob
tained from Eq. (8) by setting ϕ
A
= 0 and ϕ
B
= 0. For ψ
1
, the expressions
for the payoﬀs are found as $
A
= 3P
00
−P
01
−P
10
and $
B
= 2P
00
+3P
01
+P
10
with
P
00
= cos
2
(θ
A
/2) cos
2
(θ
B
/2) cos
2
(φ
A
+φ
B
),
P
01
= xsin φ
B
−y cos φ
A

2
,
P
10
= xcos φ
B
−y sin φ
A

2
(14)
where P
j
is calculated from Eq. (9). We ﬁnd that there is a unique NE
which appears at
ˆ
U
A
=
ˆ
U
B
= iˆ σ
z
with the payoﬀ ($
A
, $
B
) = (3, 2). Note
that this NE is a new one which cannot be seen in the classical version of the
game. Since, this new NE emerges in pure quantum strategies we conclude
that original discoordination game transforms into a coordination game.
This is the NE where Alice receives the highest payoﬀ available for her in
this game, and where both players beneﬁt. Therefore, Alice’s dilemma is
resolved in the stronger sense (CASE III).
In the case of ψ
2
, four NE’s with equal payoﬀs ($
A
, $
B
) = (3, 2)
emerge for the quantum operators (
ˆ
U
A
,
ˆ
U
B
) equal to (N, P) = (iˆ σ
y
, iˆ σ
z
),
(T, Q) =
(ˆ σ
0
+iˆ σ
y
)
√
2
,
i(ˆ σ
z
+ˆ σ
y
)
√
2
), (Y, R) =
ˆ σ
0
+i
√
3ˆ σ
y
2
,
i(
√
3ˆ σ
z
+ˆ σ
y
)
2
), and (Z, S) =
(γ
0
(ˆ σ
0
+ iγ
1
ˆ σ
y
), iδ
0
(ˆ σ
z
+ δ
1
ˆ σ
y
)) where γ
0
= cos(3π/8), γ
1
= tan(3π/8),
δ
0
= cos(π/8), and δ
1
= tan(π/8). At these NEs players receive payoﬀs
higher than those when a classical mixed strategy is used. However, Alice’s
dilemma survives because players cannot coordinate their moves to decide
on which NE point to choose. For example, if Alice thinks Bob will play
iˆ σ
z
then she will play iˆ σ
y
to reach at the ﬁrst NE. However, since this is a
simultaneous move game and there is no classical communication between
the players, Bob may play i(ˆ σ
z
+ˆ σ
y
)/
√
2 (because this action will take him
to the second NE point) while Alice plays iˆ σ
y
. Such a case will result in the
case $
A
< $
B
and will lower the payoﬀs of both players. Therefore, still a
dilemma exists in the game; however the nature of dilemma has changed.
In the classical version, the dilemma is due to welfare losses of Alice and
lack of NE. However, now the dilemma is the existence of multiple NEs.
Implications  Now let us review what we have learned from the above
analysis:
(1) Twoparameter SU(2) operators and shared entanglement are necessary
to see the novel features introduced into game theory by the quantum
mechanical toolbox. Note that oneparameter SU(2) operator set with
shared entanglement corresponds to the classical mixed strategies.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
157
(2) The number of solutions that is the number of NE’s depends on the
shared entangled state as well as on the payoﬀ matrix of the original
classical game. In SD game, switching from the entangled state ψ
1
to
ψ
2
increases the number of NE from one to four. On the other hand,
for the PD game, the number of NE’s is one for both cases.
(3) The quantum mechanical toolbox does not necessarily resolve dilemma
in classical games. Note that for the PD and SD played with ψ
1
the
dilemmas in the original games are resolved, however in SD played with
ψ
2
, a new type of dilemma appears. While in the original game the
dilemma was the absence of an NE where Alice is not exploited, in
the quantum version played with ψ
2
the dilemma is the existence of
multiple NE’s with the same payoﬀs. This new dilemma is similar to
that of the Battle of Sexes (BoS) game, that is Alice and Bob cannot
decide on which of the multiple NE’s to choose.
(4) Quantum version of a classical game can be represented by an en
larged classical payoﬀ matrix by including the unitary operators lead
ing to equilibrium solutions into the payoﬀ matrix of the initial classical
game.
56
The originally 2 × 2 classical games of PD and SD are trans
formed into new games which can be described with the new 3×3 payoﬀ
matrices. In case of SD, another representation is the 5 × 6 payoﬀ ma
trix obtained for ψ
2
. Evolution of the payoﬀ matrix of the SD game
under diﬀerent correlated states is depicted in Fig. 4. Note that the
new payoﬀ matrix includes the payoﬀ matrix of the original classical
game as its submatrix.
(5) An important quantum advantage arises when we consider the commu
nication cost of playing games. For a game represented by the 2 × 2
payoﬀ matrix, each player needs 1 classical bit (cbits) to classically
communicate their preferred strategy to the referee. Thus, the total
classical communication cost is 2 cbits. When we play the quantum
version of the game, the task is completed using four qubits (two pre
pared by the referee and distributed, and two qubits sent by the players
back to the referee after being operated on) when shared entanglement
is used. This corresponds to a communication cost of 2 ebits (1 ebit
for the referee and 1 ebit for the players). Now, if write down the
payoﬀ matrix for this quantum version, we end of with 3 × 3 payoﬀ
matrices for the PD and SD with ψ
1
, and a 5 ×6 matrix for SD with
ψ
2
. The cost of communicating players’ choices in these new payoﬀ
matrices classically (playing the game represented with those matrices)
amounts to respectively to 4cbits and 6cbits. However, in quantum
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
158
versions what is needed is still 2 ebits.
(6) In classical mixed strategies, the payoﬀs become continuous in the mix
ing probability of the players’ actions, and a compromise becomes pos
sible between the players which assures the existence of an NE. In
games played by quantum operations, the inclusion of new strategies
and moves into the game is an essential feature. Once these new moves
are incorporated, both the strategy space and the payoﬀs become con
tinuous. Then the players can ﬁnely adjust their strategies and come
closer to ﬁnding the best responses to each other’s actions which leads
to one or more NE’s.
5.2. Quantum operations and classical correlations
The shared entanglement introduced in the quantum version of the SD
game transforms the original 2 × 2 noncooperative game into a coopera
tive one. This new game thus has its own rules which are diﬀerent than
the original game: While in the original classical game, no communication
of any form is allowed between the players, in the quantum version the
players share entanglement which could be considered as a kind of “spooky
communication”.
57
In order to clarify further the role of the operators and correlations in
the obtained results above, we now consider that the players share classical
correlations. This will make it easier to evaluate the power of quantum oper
ators. Here, we consider that classical correlations are due to phase damping
acting on the MES during its distribution (see Fig. 5). Phase damping trans
forms the initially prepared MES ψ
1
and ψ
2
into the classically corre
lated states ˆ ρ
in,1
= [0000 +1111]/2 and ˆ ρ
in,2
= [0101 +1010]/2,
respectively. In the following, we will see how these correlations aﬀect the
games.
Oneparameter SU(2) operators  It is straightforward to show that
when the players are restricted to oneparameter SU(2) operators, they get
the payoﬀ (0.25, 1.5) at the NE with
ˆ
U
A
=
ˆ
U
B
= (ˆ σ
0
+iˆ σ
y
)/
√
2 independent
of the classical correlation they share. Alice’s payoﬀ in this case is much
better than the one she receives when the players share MES and use one
parameter SU(2) operators.
Twoparameter SU(2) operators  In this new setting of the problem,
M
a
r
c
h
1
1
,
2
0
0
8
8
:
5
9
W
S
P
C

P
r
o
c
e
e
d
i
n
g
s
T
r
i
m
S
i
z
e
:
9
i
n
x
6
i
n
M
a
t
h
e
m
a
t
i
c
a
l
A
s
p
e
c
t
s
1
5
9
(a)
(b)
(c)
(d)
Fig. 4. Evolution of the payoﬀ matrix of Samaritan’s dilemma (SD) game for shared quantum and classical correlation between players
using twoparameter SU(2) operators: (a) and (b) for quantum correlations  ψ
2
and  ψ
1
, respectively, and (c) and (d) for classical
correlations ˆ ρ
in,1
and ˆ ρ
in,2
, respectively. These payoﬀ matrices are obtained by adding to the original game matrix (2 × 2 payoﬀ matrix
in the middle) the strategies leading to Nash equilibria (NE). Refer to the text for the operators corresponding to the strategies shown
by letters in these payoﬀ matrices.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
160
Referee
channel
noise
Referee Referee
Fig. 5. Quantum version of a classical game played with classical correlations ˆ ρ
in
ob
tained from an initial maximally entangled state (MES) after it is subjected to noise.
for ˆ ρ
in,1
the payoﬀs become
$
A
=
1
4
[1 + 5 cos θ
A
cos θ
B
− 3 sinθ
A
sinθ
B
sin(φ
A
+ φ
B
)],
$
B
=
1
2
[3 − cos θ
A
cos θ
B
− 2 sinθ
A
sin θ
B
cos φ
A
sin φ
B
] (15)
from which we see that Alice cannot obtain a positive payoﬀ by acting
unilaterally so she should look for settlement at an equilibrium. It is easy
to show that an NE with a payoﬀ (0.25, 1.5) appears for
ˆ
U
B
= (ˆ σ
0
+iˆ σ
y
)/
√
2
and (φ
A
= 0 and {∀θ
A
: 0 ≤ θ
A
≤ π}). This is a selfenforcing NE solution
for the rational players Alice and Bob because neither of the players can
make her/his payoﬀ arbitrarily high and positive independent of the action
of the other player. This NE provides a weak solution (CASE II).
For the shared classical correlation of ˆ ρ
in,2
, the the payoﬀ functions are
obtained from Eq. (15) by replacing all the − by + and vice verse except
the + in the sin(..) term. Then the operators
ˆ
U
A
= (ˆ σ
0
+ iˆ σ
y
)/
√
2 and
ˆ
U
B
= i(ˆ σ
y
+ ˆ σ
z
)/
√
2 leads to an NE with the payoﬀ ($
A
, $
B
) = (1, 2.5).
This is a selfforcing NE, too, because Alice assures positive payoﬀs for
both Bob and herself regardless of Bob’s action. For this action of Alice,
the rational Bob will go for the above operator which maximizes his payoﬀ.
Then, the dilemma is again resolved with a weak solution (CASE II).
Another possible correlation that may be shared by the players is the full
rank classical correlation ˆ ρ
in,3
= [0000 +1010 +0101 +1111]/4.
In this case, regardless of their quantum operators players arrive at a weak
solution (CASE II) with constant payoﬀ (0.25, 1.5).
Threeparameter SU(2) operators  It is only when the players share a
MES that the threeparameter operator set does not give quantum advan
tage. It is then interesting to discuss whether this set of SU(2) operators
aﬀect the game dynamics in the case of shared classical correlation. For
ˆ ρ
in,1
, the payoﬀ functions can be obtained from Eq. (15) by replacing φ
A
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
161
and φ
B
by φ
A
+ϕ
A
and φ
B
+ϕ
B
, respectively. Then, an NE with payoﬀs
($
A
= 1, $
B
= 2.5) emerges for the operators (θ
A
= π/2, φ
A
+ ϕ
A
= π)
and (θ
B
= π/2, φ
B
+ ϕ
B
= π/2). The extra parameter ϕ allows the play
ers to increase their payoﬀs beyond those obtained with twoparameter
SU(2) operators. On the other hand, for ˆ ρ
in,2
, we ﬁnd that the strate
gies of the players to achieve an NE becomes (θ
A
= π/2, φ
A
+ ϕ
A
= 0)
and (θ
B
= π/2, φ
B
+ ϕ
B
= π/2). These operators give them the payoﬀs
($
A
= 1, $
B
= 2.5) which are the same as they obtained with twoparameter
operators. Note that even though the payoﬀs may increase, the solution is
still a weak solution (CASE II).
Implications  Our ﬁndings in this section on the eﬀects of classical
correlations on the outcome can be summarized as follows:
(1) Entanglement is not necessary to ﬁnd an NE that may solve the
dilemma of the original game. Shared classical correlations between
players with quantum operations leads to NE solutions which provide
better payoﬀs for Alice compared to those she obtains for classical
mixed strategies.
(2) Although a solution can be provided with the classical correlations
considered in this study (ˆ ρ
in,1
, ˆ ρ
in,2
and ˆ ρ
in,3
), the dilemma in SD is
not resolved in the stronger sense (CASE III). The solution with the
considered classical correlations is weak (CASE II). Note that with
twoparameter SU(2) operators and shared MES ψ
2
, players obtain a
strong solution.
(3) Three parameter SU(2) operators which cannot be used with shared
MES can be utilized to ﬁnd an NE with better payoﬀs. Note that the
most cited criticism against quantum versions of classical games is that
restricting the players to twoparameter SU(2) operators is not reason
able. Now it is possible to get rid of this. By decreasing the amount
of shared entanglement, the players can exploit whole SU(2) without
any restriction. It is important to keep in mind that this conclusion
strongly depends on the payoﬀ matrix and the dilemma type of the
original classical game.
5.3. Classical operations and classical correlations
Classical operations are a subset of quantum operations, and in Eisert model
they should commute with the entangling operator
ˆ
J. For the SD game,
quantum operators corresponding to classical strategies are the identity
operator ˆ σ
0
for A and W, and the ﬂip operator iˆ σ
y
for N and L . Here
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
162
Table 3. The new payoﬀ matrix for the SD
game when players share the classical corre
lation ˆ ρ
in
= [0000 + 1111]/2, and use
classical operations.
Bob: ˆ σ
0
Bob: iˆ σ
y
Alice: ˆ σ
0
(1.5, 1) → (−1, 2)
↑ ↓
Alice: iˆ σ
y
(−1, 2) ← (1.5, 1)
we consider the classical correlations as in the above subsection. When the
classical correlation is ˆ ρ
in,1
, the diagonal and oﬀdiagonal elements of the
payoﬀ matrix of the original game (see Fig. 1) are averaged out separately,
resulting in a new game with the payoﬀ matrix given in Table 3 without
any correlation. From Table 3, it is seen that the game is still a discoor
dination game with no NE and the dilemma survives. When the classical
correlation is ˆ ρ
in,2
, the payoﬀ matrix has the same properties as in Table
3 with diagonal and oﬀdiagonal elements interchanged. Thus, we conclude
that classical correlation does not change the structure of the SD game
except the scaling of the payoﬀs when players choose to play in classical
pure strategies. However, when they switch to mixed strategies, that is they
randomize between their pure strategies, we ﬁnd that there is a unique NE
when p = 0.5 and q = 0.2 with payoﬀ (0.25,1.5). The values of p and q
are the same as those in the classical mixed strategy of the original game
without any shared correlation. But since the payoﬀs were scaled with the
introduction of the classical correlation, now we see that there is an in
crease in Alice’s payoﬀ from −0.2 to 0.25. This is interesting because an
NE appears where both Alice and Bob have positive payoﬀ, implying that
the dilemma is resolved in the weaker sense (CASE II). Note that this was
not possible for the classical game played without correlations.
5.4. Classical operations and quantum correlations
As pointed out above, in Eisert’s scheme, the classical operators of the
players should commute with the entangling operators. This requirement
reduces the game using classical operators to the case where players apply
their operators to the initial product state from which the referee prepares
the maximally entangled state, and then makes a projective measurement
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
163
onto the four basis {00, 01, 10, 11} to calculate their payoﬀ. This is
the same as the original classical game played in pure and mixed strategies.
5.5. Classical Bob versus quantum Alice
The dilemma in SD game is a onesided dilemma of only Alice, thus one
may wonder whether Alice can do better and resolve her dilemma if Bob
is restricted to classical strategies, and Alice is allowed to use quantum
operations from SU(2) set when they share the MES, ψ
1
. In this case,
Bob’s operators will be
ˆ
U
0
B
= ˆ σ
0
with probability p and
ˆ
U
1
B
= iˆ σ
y
with
probability 1 − p. The state after the actions of Alice and Bob becomes
ˆ ρ
out
=
1
k=0
p
k
(
ˆ
U
A
⊗
ˆ
U
k
B
)ˆ ρ
in
(
ˆ
U
†
A
⊗
ˆ
U
k†
B
). (16)
Then average payoﬀs are calculated as
$
A
=
n
a
n
(pP
j
+ (1 − p)P
j
),
$
B
=
n
b
n
(pP
j
+ (1 − p)P
j
). (17)
where
P
00
= p cos
2
(
θ
A
2
) cos
2
φ
A
+ (1 − p) sin
2
(
θ
A
2
) sin
2
ϕ
A
,
P
01
= (1 − p) cos
2
(
θ
A
2
) cos
2
φ
A
+ p sin
2
(
θ
A
2
) sin
2
ϕ
A
,
P
10
= (1 − p) cos
2
(
θ
A
2
) sin
2
φ
A
+ p sin
2
(
θ
A
2
) cos
2
ϕ
A
. (18)
Oneparameter SU(2) operators  Setting φ
A
= ϕ
A
= 0, an NE with
payoﬀs ($
A
, $
B
) = (−0.2, 1.5) emerges when Alice chooses θ
A
= π/2 corre
sponding to
ˆ
U
A
= (ˆ σ
0
+iˆ σ
y
)/
√
2 and Bob chooses p = 0.2. This is the same
result as the case for quantum Alice and Bob with oneparameter SU(2)
operators. Therefore, results of classical mixed strategies are reproduced.
This is not very desirable for Alice due to her negative payoﬀ while Bob’s
is positive (CASE I).
Twoparameter SU(2) operators  Setting ϕ
A
= 0, we see that Alice
can never achieve a positive payoﬀ independent of Bob’s strategy deﬁned
by p. For any strategy of Alice, Bob can ﬁnd a p value that will minimize
Alice’s payoﬀ and vice verse. This leads to the fact that there is no NE
where both players can settle.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
164
Threeparameter SU(2) operators  The situation changes dramat
ically in this case because Alice now has an extra parameter to con
trol the outcome. When she chooses
ˆ
U
A
= [ˆ σ
0
+ i(ˆ σ
x
+ ˆ σ
y
+ ˆ σ
z
)]/2 or
ˆ
U
A
= (ˆ σ
0
+ iˆ σ
x
)/
√
2, the payoﬀs become independent of Bob’s action.
Thus Alice can control the game but this does not let her to resolve the
dilemma with a strong solution. The best she can do is a weaker solution
(case II) with the corresponding payoﬀs (0.25, 1.5) and (1, 2.5).
It is thus understood that by acting unilaterally Alice can resolve her
dilemma only when she is allowed to use operators chosen from three
parameter SU(2) set and Bob is restricted to classical operations. Even
in this case, only a solution satisfying CASE II is achieved.
6. Decoherence and Quantum Version of Classical Games
It is wellknown that quantum systems are easily aﬀected by their environ
ment, and physical schemes are usually far from ideal in practical situations.
Therefore, it is important to study whether the advantage of players and
dynamics of games with quantum operations and entanglement survive in
the presence of noise and nonideal components. In the previous sections,
we assumed that the sources used in the realization of the games were ideal
and free of noise but the distribution channel induced phase damping.
Since the initial state from which the referee prepares the entangled
state is a crucial parameter in Eisert’s scheme, any deviation from the ide
ality of the source will change the dynamics and outcomes of the game.
Consequently, the analysis of situations where the source is corrupt is nec
essary to shed a light in understanding the game dynamics in the presence
of imperfections. In this section, we consider a corrupt source and analyze
its eﬀects:
59
(i) Is there a critical corruption rate above which the players
cannot maintain their quantum advantage if they are unaware of the action
of the noise on the source, and (ii) How can the players adopt their actions
if they have information on the corruption rate of the source. The ﬁrst of
such studies was carried out by Johnson
58
who has considered a threeplayer
PD game and shown that a crossover from quantum to classical advantage
occurs as the corruption rate increases.
Here, we consider a more general corrupt source model (see Fig. 6),
where the initial state input to the entangler in Eisert’s model is ﬂipped
with some probability, and study its eﬀect on PD and SD games.
59
In the
comparison of the classical and quantum versions with and without cor
ruption in the sources, we assume that in the quantum version the players
choose their optimal operators from the twoparameter SU(2) operators set
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
165
Fig. 6. Model of the corrupt source. Two identical sources S
1
and S
2
prepare the initial
product state fg. In ideal case, they deterministically prepare the spindown state 0
resulting in fg = 00. The corrupt sources, however, prepare the spinup and down
states 1 and 0 with probabilities r and 1 − r, respectively. Therefore, the input state
of the entangler, denoted by
ˆ
J, becomes a mixture of spinup and spindown states.
while in the classical version they apply either ˆ σ
0
or iˆ σ
y
.
The state prepared by the corrupt source model shown in Fig. 6 can
be written as ˆ ρ
1
= ˆ ρ
2
= (1 − r)00 + r11. Then the combined state
generated and sent to the entangler becomes ˆ ρ
1
⊗ ˆ ρ
2
= (1 − r)
2
0000 +
r
2
1111+r(1−r)(0101+1010). This results in a mixture of the four
possible maximally entangled states (1−r)
2
ψ
+
ψ
+
 +r
2
ψ
−
ψ
−
 +r(1−
r)(φ
+
φ
+
 +φ
−
φ
−
), where ψ
∓
= 00 ∓i11 and φ
∓
= 01 ∓i10.
This is the state on which the players will perform their unitary operators.
The players assume ideal source while it is not  In this case, Alice
and Bob are not aware of the corruption in the source, and they apply the
optimal strategies to resolve their dilemma based on the assumption that
the source is ideal and the referee distributes the MES ψ
+
.
For PD with ideal sources, we know that the best strategy of the players
is to apply
ˆ
U
A
=
ˆ
U
B
= iˆ σ
y
in the classical version, and
ˆ
U
A
=
ˆ
U
B
= iˆ σ
z
in
the quantum version. Then a quantum advantage is observed in the payoﬀs
and dynamics of the game: an NE coinciding with the Pareto optimal is
obtained. In quantum version, players receive $
A
= $
B
= 3 which is better
than the payoﬀ $
A
= $
B
= 1 in classical version. In the corrupt source
model, we see that with increasing corruption rate r the payoﬀs in the
quantum version decrease while those in the classical version increase. Then
the payoﬀs with classical and quantum strategies become equal to 2.25 when
r = r
= 1/2. If r satisﬁes 0 ≤ r < 1/2, the quantum version of the game
always does better than the classical one. Otherwise, the classical game is
better. Thus, if r > r
, then the players would rather apply their classical
strategies than the quantum ones. This can be explained as follows: Since
classical operators commute with the entangling operator, in the classical
version the game is played with no entanglement. Then the players apply
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
166
iˆ σ
y
directly on the state prepared by the source. If r = 0, then they operate
on  00 and receive $
A
= $
B
= 1. On the other hand, if r = 1, they operate
on  11 and receive $
A
= $
B
= 3. Thus, when the players apply the classical
operator iˆ σ
y
, their payoﬀs continuously increase from one to three with the
increasing corruption rate from r = 0 to r = 1. In the quantum version,
on the other hand, when r = 0, the optimal strategy is
ˆ
U
A
=
ˆ
U
B
= iˆ σ
z
;
however it will not be optimal when r = 1. The optimal strategy in this case
is when θ
A
= θ
B
= 0 and φ
A
= φ
B
= π/4. If the players insist applying iˆ σ
z
when r = 1 then they will end up with the payoﬀs of the classical version.
The results of the analysis are depicted in Fig.7.
In SD, we know that the best strategy for the players in the classical
version is to use a mixed strategy which delivers ($
A
, $
B
) = (−0.2, 1.5)
at the NE. In this strategy, while Alice randomizes between her operators
with equal probabilities, Bob uses a biased randomization and applies ˆ σ
0
with probability 0.2. In the quantum version with ideal sources, we know
that the most desired solution ($
A
, $
B
) = (3, 2) is obtained at the NE when
both players apply iˆ σ
z
to ψ
+
. The dynamics of the payoﬀs in this game
with the corrupt source when the players stick to their operators iˆ σ
z
and its
comparison with their classical mixed strategy are depicted in Fig. 7. In this
asymmetric game the payoﬀs of the players become equal for the corrupt
sources when r = 1/7 and r = 1. For these corrupt sources, they receive
the payoﬀs 96/49 and 0, respectively. It is observed that the transition
from the quantum advantage to classical advantage occurs at r
= 1/2
for both players. While for increasing r, $
B
monotonously decreases from
two to zero, $
A
reaches its minimum of −0.2 at r = 0.8, where it starts
increasing to the value of zero at r = 1. It is worth noting that when
the players apply their classical mixed strategies in this physical scheme,
$
B
is always constant and independent of the corruption rate, whereas $
A
increases linearly for 0 ≤ r ≤ 1. We identify three regions of r corresponding
to the three classes of solution to SD game: CASE I when 0.6 ≤ r ≤ 1,
CASE II when 1/7 ≤ r < 0.6, and ﬁnally CASE III for r < 1/7.
Players know the corruption in the source  We assume that the referee
knows the characteristics of the corruption in the source, and informs the
players on r. The question now is whether given r the players can ﬁnd a
unique NE; and if they can, does this NE resolve their dilemma in the game
or not. One can easily see that for r = 1/2, the state shared between the
players become ˆ ρ =
ˆ
I/4. Then independent of what action they choose, the
players receive constant payoﬀs determined by averaging the payoﬀ entries
of the game matrix. For the symmetric game PD, both players receive the
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
167
0 0.2 0.4 0.6 0.8 1
 0.5
0
0.5
1
1.5
2
2.5
3
Corruption Rate (r)
P
a
y
o
f
f
a
b
c
Quantum Alice
Quantum Bob
Classical Alice
Classical Bob
0 0.2 0.4 0.6 0.8 1
0.5
1
1.5
2
2.5
3
a
Corruption Rate (r)
P
a
y
o
f
f
Quantum
Alice & Bob
Classical
Alice & Bob
Prisoners' Dilemma Samaritan's Dilemma
Fig. 7. The payoﬀs received by the players in Prisoners’ dilemma (PD) and Samaritan’s
dilemma (SD) as a function of corruption rate, r in the source, for their quantum, and
classical strategies. In the PD, the point labelled as a at r = 1/2 is the transition from
quantum advantage to classical advantage. The horizontal dashed line denotes the payoﬀs
of both players for the classical strategy without corruption. In SD, the labelled points
are as follows: a, transition point from $
A
> $
B
to $
B
> $
A
(Case III ↔ Case II), b and
c at r = 1/2, transitions from quantum advantage to classical advantage for Bob and
Alice, respectively. The horizontal solid line denotes the payoﬀ Alice receives in classical
strategies when the source is ideal. For Bob, classical strategies with and without noise
coincide and are depicted with the horizontal dotted line.
payoﬀ 9/4; for the asymmetric game SD the payoﬀs become 1/4 and 3/2
respectively for Alice and Bob. We see from Table 4 that if the players know
r then they can solve the dilemma for both r = 0 and r = 1 obtaining the
payoﬀs 3. For the completely corrupt source r = 1, they should choose
ˆ
U
A
=
ˆ
U
B
= (ˆ σ
0
+ iˆ σ
z
)/
√
2. This is however not true for the SD game: for
r = 0 the dilemma is resolved with the best possible solution (CASE III)
contrary to the situation for r = 1 where there is no unique NE (see Table
5). There are a number of NE’s with equal payoﬀs which prevents players
making sharp decisions.
Implications  The study on the corrupt source model reveals that:
(1) In a game with corrupt source with no side information to the players,
quantum advantage no longer survives if the corruption rate is above a
critical value.
(2) The strength of corruption may not only decrease the payoﬀs but also
may lead to multiple NE’s depending on the game even if the players
know the corruption in the source.
(3) If the players are given the corruption rate in the source, they can either
adapt to new optimal operators or choose a riskfree action, i.e. For PD
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
168
Table 4. Strategies leading to NE and the corresponding payoﬀs
in the PD game for the source corruption rate r.
r = 0 r = 1/4 r = 3/4 r = 1
ˆ
U
A
(θ
A
, φ
A
) (0,
π
2
) (0,
π
2
) (0,
π
4
) (0,
π
4
)
ˆ
U
B
(θ
B
, φ
B
) (0,
π
2
) (0,
π
2
) (0,
π
4
) (0,
π
4
)
($
A
, $
B
) (3, 3) (
43
16
,
43
16
) (
43
16
,
43
16
) (3, 3)
Table 5. Strategies leading to NE and the corresponding payoﬀs in
the SD game for the source corruption rate r.
1
∀φ
A
, ∀φ
B
∈ [0, π/2]
and
2
φ ∈ [0, π/4]
r = 0 r = 1/4 r = 3/4 r = 1
ˆ
U
A
(θ
A
, φ
A
) (0,
π
2
) (0,
π
2
) (0, φ)
1
(0, φ)
2
ˆ
U
B
(θ
B
, φ
B
) (0,
π
2
) (0,
π
2
) (0,
π
2
−φ)
1
(0,
π
2
−φ)
2
($
A
, $
B
) (3, 2) (
21
16
,
15
8
) (
21
16
,
15
8
) (3, 2)
and SD games, they may choose the operators
ˆ
U
A
=
ˆ
U
B
= (ˆ σ
0
+
iˆ σ
y
)/
√
2 where the payoﬀs of the players become constant independent
of corruption rate.
(4) Although the quantum players have high potential gains, there is a
large potential loss if the source is deviated from the ideal one. Classical
players are more robust to corruption in the source.
It is worth to note here that with a decoherence model in the distribution
channels in various two player quantum games in Eisert’s model, Flitney et
al. showed that the quantum player maintains an advantage over a player
with classical strategies provided that some level of coherence remains.
60
In a recent study, Flitney and Hollenberg studied the Minority game in
the presence of decoherence, and showed that the NE payoﬀ is reduced as
the decoherence is increased while the NE remains the same provided the
decoherence probability is less than 1/2.
61
The decrease rate in payoﬀs is
much higher as the number of players in the game increases. They also
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
169
showed that all players are equally disadvantaged by decoherence in one
of the qubits. Hence no player, or group of players, can gain an advantage
over the remainder by utilizing quantum error correction to reduce the error
probability of their qubit.
7. Entanglement and Reproducibility of Multiparty
Classical Games in Quantum Mechanical Settings
The expectation that quantum game theory may beneﬁt from the rich dy
namics of physical systems with many interacting particles ignited the re
search of multiplayer games. This was also encouraged by the observation
that entanglement leads to equilibrium solutions which are not observed in
the classical versions of the games.
The ﬁrst study of multiplayer quantum games was carried out by Ben
jamin and Hayden who showed by studying the PD and the Minority game
that multiplayer games can exhibit forms of pure quantum equilibrium that
can be observed neither in the classical versions nor in their twoplayer
forms.
28
They also concluded that shared entanglement leads to diﬀerent
cooperative behavior and acts a kind of contract preventing players betray
ing each other. In their model, they considered GHZlike states. Following
this work, YJ Han et al. performed a comparative study of GHZ and W
states in a three player PD game showing that (i) players prefer to form
threeperson coalition when the game is played with W state, and as a re
sult they receive the highest payoﬀ, and (ii) when the game is played with
GHZ state, players prefer twoperson coalitions and the players in the coali
tion receive payoﬀ better than the third player who is not in the coalition.
62
They further compared their results with the classical counterparts to show
the advantage of quantum games over the classical ones.
With the above two studies, it became generally accepted that quantum
versions can be easily extended to Nplayer situations by simply allowing
Npartite entangled states. In this section, however, we show that this is
not generally true because the reproducibility of classical tasks in quantum
domain imposes limitations on the type of entanglement and quantum op
erators. While in the case of GHZ state, the original classical game payoﬀ
matrix in pure strategies becomes a subset of the quantum version, this is
not true for Wstates. When the operators corresponding to the classical
strategies are applied, the payoﬀs delivered to the players are unique entries
from the classical payoﬀ matrix for the GHZ state. However, for Wstate
the payoﬀs to the players become a probability distribution over the entries
of the payoﬀ matrix. Therefore, comparing the games played with these
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
170
two states with each other and with the classical games in terms of the
payoﬀs is not fair. For twostrategy games, we derived the necessary and
suﬃcient conditions for a physical realization.
41,63,64
We also gave examples
of entangled states that can and cannot be used, and the characteristics of
quantum operators used as strategies.
Fig. 8. The model for the multiplayer quantum games.
7.1. Reproducibility criterion to play games in quantum
mechanical settings
In order to make a fair comparison between the classical and quantum
versions of a game as well as its dynamics with diﬀerent shared entangled
states, the physical scheme in the quantum mechanical setting should be
able to reproduce the results of the classical version, too. That is with the
proper choices of operators the classical game should pop up as a subset of
its quantum version. In Eisert’s model of quantum games, reproducibility
criterion is satisﬁed by choosing the quantum operators corresponding to
classical strategies in such a way that they commute with the entangling
operator
ˆ
J. Therefore, we require that in any model of multiplayer games
this criterion should be satisﬁed.
An Nplayer twostrategy classical game should be reproduced in the
quantum version when each player’s strategy set is restricted to two unitary
operators, {ˆ u
1
i
, ˆ u
2
i
}, corresponding to the two pure strategies in the classi
cal game. Then the combined pure strategy of the players is represented
by ˆ x
k
= ˆ u
l
1
1
⊗ ˆ u
l
2
2
⊗ · · · ⊗ ˆ u
l
N
N
with l
i
= {1, 2} and k =
N
i=1
(l
i
− 1)2
i−1
.
Thus the output state becomes  Φ
k
= ˆ x
k
 Ψ with k = {1, 2, . . . , 2
N
}. For
the strategy combination ˆ x
k
, expected payoﬀ for the ith player becomes
$
i
(ˆ u
1
, · · · , ˆ u
N
) = Tr
Σ
j
a
i
j
ˆ
P
j
ˆ x
k
 ΨΨˆ x
†
k
. In this section, we will look
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
171
for an answer to the question: Can we ﬁnd {ˆ u
1
i
, ˆ u
2
i
} for any Npartite en
tangled state and for any game?
Reproducibility can be discussed in two cases: (i) The referee should
be able to identify the strategy played by each player deterministically re
gardless of the structure of the payoﬀ matrix, (ii) The referee should be
able to reproduce the expected payoﬀ of the classical game in the quantum
version.
63
In (i), the referee should identify all possible outcomes  Φ
k
de
terministically which requires that all possible states formed by the joint
applications of the players’ operators should be mutually orthogonal to each
other,
Φ
α
 Φ
β
= δ
αβ
∀α, β. (19)
This strong criterion should be satisﬁed independent of the structure of
the game payoﬀ matrix. In (ii), however, the referee needs to distinguish
between the subsets of outcomes with the same payoﬀ. If two output
states  Φ
j
and  Φ
k
have the same payoﬀs then they should be grouped
into the same set. If all possible output states are grouped into sets
S
j
= {Φ
1j
, Φ
2j
, . . . Φ
nj
} and S
k
= {Φ
1k
, Φ
2k
, . . . Φ
n
k
} then the
referee should deterministically discriminate between these sets which is
possible iﬀ the state space spanned by the elements of each set are orthogo
nal. Hence for every element of S
j
and S
k
, we have Φ
nj
Φ
n
k
= 0, ∀j = k,
that is all the elements of S
j
and S
k
must be orthogonal to each other, too.
This weak criterion is very much dependent on the structure of the game
payoﬀ matrix.
When a twoplayer twostrategy game is extended to Nplayer game
(N > 2), the new payoﬀ matrix is formed by summing the payoﬀs that
each player would have received in simultaneously playing the twoplayer
game with N − 1 players. Hence, in their Nplayer extensions, Prisoners’
dilemma (PD), Samaritan’s dilemma (SD), Boxed Pigs (BP), Modeller’s
dilemma (MD), DeadLock (DL), and Ranked Coordination (RC) have pay
oﬀ matrices where all the outcomes have diﬀerent payoﬀ vectors.
1,41
Thus,
these games should be evaluated according to the strong criterion. On the
other hand StagHunt (SH), Chicken Game (CG), Battle of Sexes (BoS),
Battle of Bismarck (BB), Matching Pennies (MP), and Alphonse & Gaston
Coordination Game (AG), Minority, majority and coordination games have
payoﬀ matrices where some of the outcomes have the same payoﬀ vectors,
thus these games can be evaluated by the weak criterion.
1,41
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
172
7.2. Entangled states and strong reproducibility criterion
In this subsection, ﬁrst as an example we will show that Wstates cannot
be used in quantum games because they do not satisfy the strong repro
ducibility criterion. Then we will derive a general form of entangled states
that satisfy the criterion and hence can be used in quantum versions of
classical games.
W states do not satisfy the strong reproducibility criterion  Let us
assume that the Npartite W state  W
N
has been prepared and distributed
to the players, and the ith player has the strategy set {ˆ u
1
i
, ˆ u
2
i
}. Acting on
the shared state  W
N
with all possible strategy combinations of Nplayers
will generate the output state set {Φ
1
, Φ
2
, .., Φ
2
N} such that
Φ
1
=
ˆ u
1
1
⊗ ˆ u
1
2
⊗ ˆ u
1
3
⊗... ⊗ ˆ u
1
N
W
N
Φ
2
=
ˆ u
2
1
⊗ ˆ u
1
2
⊗ ˆ u
1
3
⊗... ⊗ ˆ u
1
N
W
N
Φ
3
=
ˆ u
2
1
⊗ ˆ u
2
2
⊗ ˆ u
1
3
⊗... ⊗ ˆ u
1
N
W
N
.. ....
.. ....
Φ
2
N =
ˆ u
2
1
⊗ ˆ u
2
2
⊗ ˆ u
2
3
⊗... ⊗ ˆ u
2
N
W
n
. (20)
Since the strong criterion of reproducibility requires that Φ
α
Φ
β
= δ
αβ
for ∀α, β ∈ {1, .., 2
n
}, we check the mutually orthogonality of all the 2
N
output states listed above. We ﬁrst take any two output states for which the
strategies of only one player diﬀer, e.g.,Φ
1
and Φ
2
. The inner products
of these states becomes
Φ
1
 Φ
2
=
1
N
[(N − 1)0 ˆ u
1†
1
ˆ u
2
1
 0 + 1 ˆ u
1†
1
ˆ u
2
1
 1]. (21)
Since ˆ u
1
1
and ˆ u
2
1
are SU(2) operators, so is the ˆ u
1†
1
ˆ u
2
1
which can be expressed
as
x y
y
∗
−x
∗
(22)
with x
2
+ y
2
= 1. Substituting this in Eq. (21), we ﬁnd that ˆ u
1†
1
ˆ u
2
1
=
ˆ σ
x
ˆ
R
z
(2φ
1
) where the rotation operator
ˆ
R
z
(γ) is given as
ˆ
R
z
(γ) = e
−iγˆ σ
z
/2
.
Note also that ˆ u
1†
1
ˆ u
2
1
is a normal operator. Similar expression can be easily
obtained for the ith player just considering two output states diﬀering in
the operators of the ith player. Next, we take two output states diﬀering
in the operators of two players, e.g.,Φ
1
and Φ
3
. Inner product of these
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
173
states is
Φ
1
 Φ
3
=
1
N
(01  + 10 ) (ˆ u
1†
1
ˆ u
2
1
) ⊗ (ˆ u
1†
2
ˆ u
2
2
) ( 01 +  10)
=
1
N
e
i(φ
1
−φ
2
)
+ e
−i(φ
1
−φ
2
)
=
2
N
cos(φ
1
− φ
2
). (23)
The condition of orthogonality requires cos(φ
1
− φ
2
) = 0, thus φ
1
− φ
2
=
nπ+π/2 where n is an integer. Thus, we have φ
j
−φ
k
= nπ+π/2 for any two
output states diﬀerent only in the strategies of j and kth players. Then for
any three player, we have the set of equations χ
jkm
= {φ
j
−φ
k
= nπ+π/2,
φ
m
− φ
j
= n
π + π/2, φ
k
− φ
m
= n
π + π/2} with n, n
and n
being
integers. One can easily notice that the sum of the three equations in this
set leads to a contradiction: The sum of the three equations in χ
jkm
results
in 3π/2+m
π = 0 which is impossible for integer m
= n+n
+n
. Thus, we
conclude that there exists no two operators corresponding to the classical
strategies of the players when  W
N
is shared in a twostrategy Nplayer
game. Therefore, the classical game cannot be reproduced and hence the
W state cannot be used.
General form of states satisfying the strong reproducibility criterion 
Now suppose that for an Nqubit entangled state  Φ we have found two
operators {ˆ u
1
i
, ˆ u
2
i
} implying that we have a situation where the strong repro
ducibility criterion is satisﬁed. Let us again take the output states diﬀering
in the operators of only one player say Φ
1
=
ˆ u
1
1
⊗ ˆ u
1
2
⊗ ... ⊗ ˆ u
1
N
Φ and
Φ
2
=
ˆ u
2
1
⊗ ˆ u
1
2
⊗ ... ⊗ ˆ u
1
N
Φ. Taking the inner product and imposing
the reproducibility criterion, we ﬁnd Φˆ u
1†
1
ˆ u
2
1
⊗
ˆ
I ⊗ ... ⊗
ˆ
IΦ = 0. Since
ˆ u
1†
1
ˆ u
2
1
is a normal operator, it can be diagonalized by a unitary operator,
say ˆ z
1
. Thus, we can write
ˆ
D
1
= ˆ z
1
ˆ u
1†
1
ˆ u
2
1
ˆ z
†
1
=
ˆ
R
z
(−2φ
1
) where e
∓iφ
1
are
the eigenvalues of ˆ u
1†
1
ˆ u
2
1
, and consequently we have
Φˆ z
†
1
ˆ z
1
ˆ u
1†
1
ˆ u
2
1
ˆ z
†
1
ˆ z
1
⊗
ˆ
I ⊗ · · · ⊗
ˆ
I Φ = Φ

ˆ
D
1
⊗
ˆ
I ⊗ · · · ⊗
ˆ
I Φ
= Φ

ˆ
R
z
(−2φ
1
) ⊗
ˆ
I ⊗ · · · ⊗
ˆ
I Φ
= 0, (24)
where  Φ
= ˆ z
1
⊗
ˆ
I⊗· · ·⊗
ˆ
I Φ. Now writing the state  Φ
in computational
basis as  Φ
= Σ
i
j
∈{0,1}
c
i
1
i
2
...i
N
 i
1
 i
2
· · ·  i
N
, and substituting into Eq.
(24), we ﬁnd
cos φ
1
+ i
2Σ
i
j
∈{0,1}
c
0 i
2
...i
N

2
− 1
sin φ
1
= 0 (25)
which is satisﬁed for cos φ
1
= 0 and Σ
i
j
∈{0,1}
c
0 i
2
...i
N

2
= 1/2. From
cos φ
1
= 0, we see that φ
1
= nπ + pi/2 implying that
ˆ
D
1
= iσ
z
. This
result can be extended for the other players, therefore we have
ˆ
D
k
= iσ
z
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
174
The above equations imply that if the Nqubit state  Φ and the op
erators {ˆ u
1
k
, ˆ u
2
k
} satisfy reproducibility criterion, then the state  Φ
=
ˆ z
1
⊗ ˆ z
2
⊗· · · ⊗ ˆ z
N
 Φ and the unitary operators {
ˆ
D,
ˆ
I} should satisfy, too.
Neglecting the irrelevant global phase of i of
ˆ
D, and substituting it into
all possible output states written as in Eq. (24), we end up with 2
N
− 1
equalities to be satisﬁed:
Φ
ˆ σ
z
⊗
ˆ
I ⊗
ˆ
I ⊗ · · · ⊗
ˆ
I Φ
= 0,
Φ

ˆ
I ⊗ ˆ σ
z
⊗
ˆ
I ⊗ · · · ⊗
ˆ
I Φ
= 0,
.
.
.
Φ
ˆ σ
z
⊗ ˆ σ
z
⊗ · · · ⊗ ˆ σ
z
 Φ
= 0. (26)
Rewriting these equations in the matrix form yields
1 . . . 1 −1 . . . −1
. . . . . . . . .
.
.
.
.
.
.
1 . . . 1 1 . . . 1
c
00...0

2
c
00...1

2
.
.
.
.
.
.
c
11...1

2
=
0
0
.
.
.
.
.
.
1
, (27)
where the last row is the normalization condition. The row vector corre
sponds to the diagonal elements of ˆ σ
{0,1}
z
⊗· · ·⊗ˆ σ
{0,1}
z
where ˆ σ
0
z
is deﬁned as
ˆ
I. Consider the operators ˆ x, ˆ y ∈ (ˆ σ
{0,1}
z
)
⊗N
where ˆ xˆ y ∈ (ˆ σ
{0,1}
z
)
⊗N
. Since
Tr[ˆ σ
z
] = 0, for ˆ x = ˆ y, we have Tr[ˆ x ˆ y] = Tr[ˆ x] Tr[ˆ y] = 0. Thus any two row
vectors are orthogonal to each other, thus the matrix in Eq. (27) has an
inverse, and c
i
1
i
2
...i
N

2
are uniquely determined as 1/N. This implies that
if a state satisﬁes strong reproducibility criterion, then it should be trans
formed by local unitary operators into the state which contains all possible
terms with the same magnitude but diﬀerent relative phases:
 Φ
= Σ
i
j
∈{0,1}
N
−1/2
e
iφ
i
1
i
2
..i
N
 i
1
 i
2
· · ·  i
N
. (28)
One can easily show that the product state can be transformed into the
form of Eq. (28) when all the players apply the Hadamard operator,
ˆ
H =
( ˆ σ
x
+ ˆ σ
z
)/
√
2. The GHZ state, which also satisﬁes the strong reproducibility
criterion, can be transformed into the same form by (e
i
π
4
ˆ
I+e
−i
π
4
ˆ σ
z
+ˆ σ
y
)/
√
2
for one player and
ˆ
H for the others. Thus, Bell states, product states and
 GHZ
N
states can be used in quantum versions of classical twostrategy
multiplayer games. In the same way, one can easily show that  W
N
and
Dicke states  N − m, m except  2, 2 and  1, 1 cannot be written in the
form Eq. (28) and hence cannot be used. The proof for Dicke states can
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
175
also be obtained in the way we have done in the previous subsection for
Wstates. We ﬁnd that the Dicke state  2, 2 can be used in a quantum
game by assigning the classical strategies to the operators ˆ u
1
k=1,2,3,4
=
ˆ
I,
ˆ u
2
k=1,2,3
= i(
√
2ˆ σ
z
+ ˆ σ
x
)/
√
3, and ˆ u
2
4
= iˆ σ
y
. The eigenvalues of ˆ u
k
1†
ˆ u
2
k
are i
and −i and they are already in the diagonalized form. For GHZ state, the
operators are ˆ u
1
k
=
ˆ
I and ˆ u
2
k
= iˆ σ
y
which can also be written in the form
ˆ
D.
Weak reproducibility criterion and W states  In this section we check
whether the above results for strong reproducibility criterion are valid or
not for the games where payoﬀs are the same for some of the possible output
states after the players perform their operators. We will consider only the
Wstates here but a detailed analysis for Dicke states can be found in.
41
The results of such a study requires a classiﬁcation of multiplayer games
because the results would be very much dependent on the structure of
payoﬀ matrices. The following observations for the Wstates from the above
analysis make our task easier:
(i) If the mutual orthogonality condition of the subsets (each subset in
cludes the all output states with the same payoﬀs) leads to the operator
form ˆ u
1†
k
ˆ u
2
k
= ˆ σ
x
ˆ
R
z
(2φ
k
) for any two output states diﬀering only in the
strategy of kth player, then there will be contradiction if we obtain
the set χ
jkm
for any threeplayercombination (j, k, m). Presence of at
least one such set is enough to conclude that there is contradiction.
(ii) If the mutual orthogonality condition of the subsets leads to the oper
ator form as in ˆ u
1†
k
ˆ u
2
k
= ˆ σ
x
ˆ
R
z
(2φ
k
) for one and only one player, then
there will be no contradiction because there will be at least one missing
equation in χ
jkm
for all possible threeplayercombinations (j, k, m).
Note that such a situation occurs iﬀ 2
N
possible outputs are divided
into two subsets with equal number of elements. Then the only equa
tions we will obtain are φ
1
−φ
j
= nπ +π/2 for all j = 2, · · · , N.
(iii) If the number of elements in any of the subsets in a game is an odd
number, then there will always be contradiction
(iv) If there is a set with only two elements which are the outcomes when
all the players choose the same strategy, there will be contradiction.
If we encounter a contradiction in the analysis of a multiplayer game
played with Wstates, we conclude that Wstates cannot be used in the
quantum version of that game. Checking these observation on various
games, we found out that Wstates cannot be used in (1) multiparty exten
sions of originally twoplayer twostrategy games such as SH, CG, BoS, BB,
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
176
MP, and AG due to contradiction explained in (iii), (2) minority and major
ity games due to contradiction explained in (i), (3) coordination games due
to (iiv), (4) zero sum gates due to (iv), and (5) symmetric games among
which we can list PD, MD, RC, CG, SH, and AG up to six players due to
(iii).
As an example, let us analyze the minority game with fourplayers. In
this game, each player chooses between two strategies. The choices are then
compared and the player(s) who have made the minority decision are re
warded by one point. If there is an even split, or if all players have made
the same choice, then there is no reward. The players with the majority
decision loses one point. The structure of this game reﬂects many com
mon social dilemmas, for example choosing a route in rush hour, choosing
which evening to visit an overcrowded bar, or trading in a ﬁnancial market.
Since we are dealing with a fourparty case, there are 16 possible outcomes,
{ Φ
L=1,..,16
} where the subscript is L = 1 +
4
i=1
i
2
4−i
with
i
= 0 (or
i
= 1) when the kth player chooses the ﬁrst operator ˆ u
1
k
(or the second
operator ˆ u
2
k
). For example, for the joint action ˆ u
1
1
ˆ u
2
2
ˆ u
2
3
ˆ u
1
4
where the ﬁrst
and fourth players choose the ﬁrst, and the second and third players choose
the second operators, we ﬁnd
1,4
= 0 and
2,3
= 1 and consequently L = 7
implying that  Φ
7
is the output state received by the referee. In the same
way, if the output state is  Φ
2
, then we know that L = 2 which is calcu
lated using
1,2,3
= 0 (ﬁrst, second and third players have chosen the ﬁrst
operator), and
4
= 1 (the fourth player has chosen the second operator).
For eight output states out of sixteen possible outputs for the four
player minority game, there is an evensplit (half of the players prefers
ﬁrst operator and the other half prefers the second operator) and hence
all the players are rewarded with payoﬀ zero. We group these eight states
into the set S
1
= { Φ
1,4,6,7,10,11,13,16
}. Each player is minority in two
cases, therefore the rest of the eight output states can be grouped into
S
2
= { Φ
2,15
}, S
3
= { Φ
3,14
}, S
4
= { Φ
5,12
}, and S
5
= { Φ
8,9
} where
fourth, third, second and ﬁrst players are minority players, respectively.
Note also that in each of the sets we have output states diﬀering in the
operators of all the players. From the setpair (S
1
, S
3
) using Φ
3
Φ
1
, from
the setpair (S
1
, S
4
) using Φ
5
Φ
1
, from the setpair (S
1
, S
5
) using Φ
9
Φ
1
,
and from the setpair (S
1
, S
2
) using Φ
2
Φ
1
, we ﬁnd ˆ u
1†
k
ˆ u
2
k
= ˆ σ
x
ˆ
R
z
(2φ
k
) for
∀k. Substituting ˆ u
1†
k
ˆ u
2
k
into the expressions obtained from the orthogonality
relations of the sets (S
2
, S
5
), (S
3
, S
5
) and (S
4
, S
5
), we see that the set
of expressions represented by χ
234
in (i) is obtained. Therefore, similar
contradiction is observed.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
177
For a zerosum game with N players it is easy to see that a contradiction
similar to the one in (iv) occurs. In a zerosum game with a competitive
advantage ξ, there is no winner or loser so all the players receive zero
payoﬀs. Otherwise, t players choosing the ﬁrst strategy gets ξ/t, and the
rest N−t players receive ξ/(N−t). Then the output states can be grouped
into 2
N
− 1 diﬀerent sets. One of these sets will contain the two output
states when all players choose the same strategy, and each of the rest will
contain only one state. Such an output state partition is exactly what leads
to the contradiction in (iv). Therefore, Wstates cannot be used in zerosum
games.
Our studies have revealed that the results obtained under the strong
reproducibility condition for the W, GHZ, Dicke and Bell states are valid
for a large class of classical games whose classical outcomes correspond to
the same payoﬀ. We restricted our results to the speciﬁc entangled states
which have already been created in experiments. The reproducibility crite
rion provides a fair basis to compare quantum versions of games with their
classical counterparts because only the states and games which satisfy the
criterion in a physical scheme can be compared and played. The analysis
shows that a large class of multipartite entangled states cannot be used in
the quantum version of classical games; and that the operators that might
be used should have a special diagonalized form.
8. Conclusion
In this lecture note, we have attempted to clarify some of the dynamics in
the games when the classical games are transferred to quantum domain.
The widely used Eisert’s scheme is discussed and applied to speciﬁc games
to show the eﬀects of shared correlations and the set of operators available
to the players on the game dynamics. One interesting ﬁnding is that for
some games entanglement is not needed to arrive at an NE and resolve
the dilemma in the game; however it is necessary to ﬁnd the solution with
the highest payoﬀ distribution to the players. Classical correlations in a
game may lead to NE and in some cases resolve the dilemma however they
never lead to payoﬀs as high as those in the case of entanglement. The
comparison of various games also revealed that quantum advantage does
not necessarily take place in all the games; it is strongly dependent on the
game payoﬀ matrix. In an attempt to form a fair basis to evaluate the
entanglement in games and to compare the game dynamics, we suggested
to impose a reproducibility condition which states that when a quantum
mechanical setting is used to play the game the classical game should be
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
178
reproduced in that speciﬁc setting. This condition lead us to give general
forms of the entangled states and quantum operators which can be used in
quantum versions of classical games.
In the rapidly growing quantum information science, it is inescapable
to deal with conﬂict and competitive situations. The expectation is that
quantum game theory may help when dealing with such situations. It is also
expected that development and understanding of quantum game theory will
pave the way to develop eﬃcient quantum algorithms and protocols. It is
still not clear how we can utilize the present knowledge learned from the
study of game theory in the past years to meet the above expectations.
Therefore, there is still much work to be done to fully understand the
mechanisms behind the interesting observations in quantum games, and to
develop a more fundamental game theory which is associated with quantum
mechanics.
Acknowledgements
The authors thank Prof. M. Koashi, Dr. T. Yamamoto and Dr. F. Morikoshi
for fruitful discussions, advices and support. S¸. K.
¨
Ozdemir thanks Prof. M.
Nakahara, Dr. R. Rahimi and Dr. T. Aoki for their invitation and kind hos
pitality during the Summer School on Mathematical Aspects of Quantum
Computing 2007 held at Kinki University, HigashiOsaka, Japan.
References
1. E. Rasmusen, Games and Information: An Introduction to Game Theory
(Blackwell Pub, Oxford, 2001); R. B. Myerson, Game Theory: Analysis of
Conﬂict (Harvard Univ. Press, Cambridge MA, 1997).
2. Neumann John von, Mathematische Annalen 100 (1928) 295.
3. Neumann John von and Oscar Morgenstern, The Theory of Games and Eco
nomic Behavior, (New York: Wiley, 1944).
4. J. Orlin Grabbe, quantph/0506219
5. P Moulin, and A. Ivanovic, Proc. of Int. Conf. on Image Process. 3 (2001)
975.
6. A. S. Cohen, A. Lapidoth, IEEE Trans. on Inf. Theory 48 (2002) 1639.
7. J. Conway, N. Sloane, IEEE Trans. on Inf. Theory 32 (1986) 337.
8. X. M. Shen, L. Deng, IEEE Trans. on Signal Process. 45 (1997) 1092.
9. J. M. Ettinger, in 2nd Inf. Hiding Workshop, Portland, USA, Apr. 1517
(1998).
10. S. Pateux, G. Le Guelvouit, Signal Process.: Image Comm. 18 (2003) 283.
11. L. A. DaSilva, and V. Srivastava, in The First Workshop on Games and
Emergent Behaviors in Distributed Computing Environments, Birmingham,
UK, September (2004)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
179
12. N. F. Johnson et. al. Fluct. and Noise Lett. 2 (2002) L305.
13. J.M.R. Parrondo et. al., Phy. Rev. Lett. 85 (2000) 5226.
14. C. F. Lee, N. F. Johnson, Phys. World 15(10) (2002) 259.
15. D. A. Meyer, Phys. Rev. Lett. 82 (1999) 1052.
16. J. Eisert, M. Wilkens and M. Lewenstein, Phys. Rev. Lett. 83 (1999) 3077;
ibid 87 (2001) 069802.
17. C. F. Lee, and N. F. Johnson, Phys. Lett. A 319 (2003) 429.
18. C. F. Lee, and N. F. Johnson, quantph/0203049.
19. C. F. Lee, and N. F. Johnson, Phys. Rev. A 67 (2003) 022311.
20. J. Shimamura S. K. Ozdemir, F. Morikoshi and N. Imoto, Int. J. of Quant.
Inf. 2/1 (2004) 79.
21. M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum In
formation, (Cambridge university Press, 2000).
22. W. D¨ ur, G. Vidal, and J. I. Cirac, Phys. Rev. A 62 (2000) 062314.
23. M. Koashi, V. Buˇ zek, and N. Imoto, Phys. Rev. A 62 (2000) 050302(R).
24. H. J. Briegel and R. Raussendorf, Phys. Rev. Lett. 86 (2001) 910.
25. R. H. Dicke, Phys. Rev. 93 (1954) 99.
26. J. K. Stockton et. al., Phys. Rev. A 67 (2003) 022112.
27. S. C. Benjamin and P. M. Hayden, Phys. Rev. Lett. 87 (2001) 069801.
28. S. C. Benjamin, P. M. Hayden, Phys. Rev. A 64(3) (2001) 030301(R).
29. L. Marinatto and T. Weber, Phys. Lett. A 272 (2000) 291.
30. A. Nawaaz, and A. H. Toor, quantph/0409046.
31. T. Cheon, and I. Tsutsui, Phys. Lett. A 348 (2006) 147.
32. T. Ichikawa, I. Tsutsui, and T. Cheon, quantph/0702167.
33. K. Y. Chen, T. Hogg, and R. Beausoleil, Quant. Inf. Proc 1 (2002) 449.
34. A. Iqbal and A. H. Toor, Phys. Lett. A 300 (2002) 537.
35. R. Kay, N. F. Johnson, and S. C. Benjamin, quantph/0102008.
36. A. Iqbal and A. H. Toor, Phys. Lett. A 280 (2001) 249.
37. A. Iqbal and A. H. Toor, Phys. Lett. A 294 (2002) 261.
38. A. Iqbal and A. H. Toor, quantph/0503176
39. S. K. Ozdemir, J. Shimamura, F. Morikoshi, and N. Imoto, Phys. Lett. A
313 (2004) 218.
40. J. Shimamura, Playing games in quantum realm, PhD Dissertation, Osaka
University, Osaka, Japan, March (2005).
41. S. K. Ozdemir, J. Shimamura, and N. Imoto, New J. Phys. 9 (2007) 43.
42. E. W. Piotrowski, and J. Sladkowski, Physica A 312 (2002) 208.
43. E. W. Piotrowski, and J. Sladkowski, Physica A bf 318 (2003) 505.
44. E. W. Piotrowski, and J. Sladkowski, Quant. Finance 4 (2004) 1.
45. G. P. Harmer and D. Abbott, Nature 402 (1999) 864.
46. A. Allison and D. Abbott, Fluct. Noise. Lett. 2 (2002) L327.
47. A. P. Flitney, J. Ng, and D. Abbott, Physica A 314 (2002) 35.
48. A. P. Flitney, and D. Abbott, Physica A 324 (2003) 152.
49. Du J, Li H, Xu X, Shi M, Wu J and Han R, Phys. Rev. Lett. 88 (2002)
137902.
50. R. Prevedel, A. Stefanov, P. Walther, and A. Zeilinger, New J. Phys. 9 (2007)
205.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
180
51. M. Paternostro, M. S. Tame, M. S. Kim, New J. Phys. 7 (2005) 226.
52. A. P. Flitney, Aspects of quantum game theory, PhD Disserta
tion, University of Adelaide, Australia, January (2005). Available at
http://ariic.library.unsw.edu.au/adelaide/adtSUA20051027072026/.
53. A. Iqbal, Studies in the theory of quantum games, PhD Dissertation, Quaidi
azam University, Islamabad, Pakistan, September (2004). Available at arXiv:
quantph/0503176.
54. A. F. HuertasRosero, Classiﬁcation of quantum symmetric nonzero sum 2×2
games in the Eisert scheme, MSc Thesis, Universidad de Los Andes, Bogota,
Colombia, February (2004). Available at arXiv: quantph/0402117.
55. J. M. Buchanan, “The Samaritan’s Dilemma”, in Altruism, Morality, and
Economic Theory, edited by Edmund S. Phelps, (Russell Sage, New York,
1975).
56. S. J. van Enk and R. Pike, Phys. Rev. A 66 (2002) 024306.
57. G. Brassard, quantph/0101005.
58. N. F. Johnson, Phys. Rev. A 63 (2001) 020302.
59. S. K. Ozdemir, J. Shimamura, F. Morikoshi, and N. Imoto, Phys. Lett. A
325 (2004) 104.
60. A. P. Flitney and D. Abbott, J. Phys. A 38 (2005) 449.
61. A. P. Flitney and L. C. L. Hollenberg, quantph/0510108.
62. YJ. Han, YS. Zhang, and G. C. Guo, Phys. Lett. A 295 (2002) 61.
63. J. Shimamura, S. K. Ozdemir, F. Morikoshi and N. Imoto, Phys. Lett. A 328
(2004) 20.
64. J. Shimamura, S. K. Ozdemir, F. Morikoshi and N. Imoto, quantph/0508105.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
181
QUANTUM ERRORCORRECTING CODES
MANABU HAGIWARA
National Institute of Advanced Industrial Science and Technology,
AkihabaraDaibiru 11F, 11813 Sotokanda,
Chiyodaku, Tokyo, 1010021 Japan
∗
Email: hagiwara.hagiwara@aist.go.jp
In this review, we discuss quantum errorcorrecting codes by emphasizing the
generalization of classical errorcorrecting codes. Starting with the explanations
of general notions of classical errorcorrecting codes, we ﬁnish the discussion
with the construction of quantum codes which are constructed by a pair of
practical and trendy classical codes.
Keywords: Quantum ErrorCorrecting Codes, CSS Codes, Stabilizer Codes,
LDPC Codes, QuasiCyclic LDPC Codes, Group Theory, Algebraic Combina
torics.
1. Introduction
Quantum errorcorrecting codes provide detection or correction of the errors
which occur in a communication through a noisy quantum channel. Hence
the codes protect integrity of a message sent from a sender to a receiver.
In this review, we discuss quantum errorcorrecting codes in view of gen
eralization of classical errorcorrecting codes. We give examples of basic the
ory of classical errorcorrecting codes and then introduce related quantum
counterparts. One of the research trends is to generalize algebraic geomet
rical codes, e.g. BCH codes, ReedSolomon codes, and others,
1
to quantum
codes in view of mathematical aspects of quantum errorcorrection codes.
In this paper, we focus not on algebraic geometrical codes but on LDPC
codes.
2
In particular, the main target of a class of LDPC codes which are
generalized to quantum codes is called a class of CSS codes.
3
For construc
tion, the nearly random paritycheck matrix is required to classical LDPC
codes. On the other hand, a constructive but interesting condition, called
twisted condition, is required to CSS codes, which is one of the most popular
classes of quantum codes. Thus LDPC codes does not seem to be suitable
for generalizations to quantum errorcorrecting codes. Recently, the notion
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
182
of quasicyclic LDPC codes has been proposed, which changed the reseach
trend of quantum errorcorrecting theory. One of the advantages of quasi
cyclic LDPC code is theoretical one that there are various mathematical
tools, e.g. discrete mathematics, ﬁnite rings, and others, to describe the
properties of LDPC codes. Another advantage of a quasicyclic LDPC code
is practical one that it has high errorcorrecting performance with short
length. In this review we aim to study the construction of quantum error
correcting code in the framework of generalization of classical quasicyclic
LDPC codes.
Throughout this paper, we assume that a classical errorcorrecting code
is a binary code and a quantum errorcorrecting code consists of sequences
of two level states.
2. Classical ErrorCorrecting Codes
2.1. Introduction to classical errorcorrecting codes
Let us consider the following of examples of classical errorcorrecting codes:
A sender would like to transmit a single bit x ∈ {0, 1} to a receiver. It is
natural to assume that the communication channel is noisy. Let us assume
that the bitﬂip error occurs with small but nonnegligible probability over
the communication channel. The sender transmits x itself over the noisy
channel and the receiver obtains a single bit y. It is impossible to distinguish
for the receiver that the received message y is the original message x or the
bitﬂipped message x + 1 because of noise. Here + is a bitwise addition
modulo 2.
Let us introduce an errorcorrecting code to this scenario. Imagine that
the sender transmits a word xxx instead of the single bit x to the receiver
over the noisy channel. This implies that the original word is an element
of C = {000, 111}. If one of the three bits is ﬂipped, then the received
word is not an element of the set C. Thus the receiver detects the bitﬂip
error. Furthermore, the receiver corrects the error by a majority decision.
The 3repetition code is useful for classical bitﬂip errorcorrection if the
probability of double bitﬂip errors in a word over the channel is negligible.
The probability that double bitﬂip errors occurring in the word depends
on a communication channel. If it is not negligible, we should choose a
suitable encoding for communication. For example, the 5repetition code is
one of the candidates. The point of encoding is to introduce some form of
redundancy to the original information. Study of good encoding is a fun
damental and an interesting but an important problem for errorcorrecting
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
183
codes. Not only to ﬁnd a good encoding but also to ﬁnd a good decoding
algorithm is an important problem.
Shannon shows that high errorcorrecting performance is achieved by
a random code, where a random code is a code whose codewords are cho
sen randomly.
4,5
However, there remain some problems. The random code
requires huge length and it is impossible to correct errors over a random
code in realtime. Shannon’s result is quite interesting and meaningful from
theoretical point of view. However we would like to develop a good error cor
recting code in realistic meaning, easy encoding, easy decoding, and good
errorcorrecting performance with shorter length. The purpose of error
correcting code theory is to develop a code with “easy encoding”, “easy
decoding method” and “shorter length”.
2.2. Linear codes and paritycheck matrices
Let F
2
be a prime ﬁeld with characteristic 2, i.e. F
2
= {0, 1}. A subset of
F
2
∪F
2
2
∪F
3
2
∪. . . is called a code. An element of F
2
∪F
2
2
∪F
3
2
∪. . . is called
a bit sequence. Hence a code (or a code space) means a set which consists
of bit sequences and no mathematical structure is required. If a code C is
given, an element of a code is called a codeword of C. If a codeword (resp.
a bit sequence) belongs to F
n
2
for some positive integer n, then the length
of the codeword (resp. the bit sequence) is said to be n. For example, the
length of (0, 0, 0, 0, 0) is 5.
A code is called a block code if the length of any element of the code
is independent of a codeword. Thus a block code C is a subset of F
n
2
for a
positive integer n. Note that a block code is a ﬁnite set. The positive integer
n is called the length of the code C. Note that we can consider F
n
2
a vector
space over F
n
. We call a block code C(= ∅) linear if C is a linear subspace
of F
n
2
, i.e. c + c
∈ C for any c, c
∈ C. Then the code C is a vector space
over F
2
. We should note that it is not required that a code be a block code
to deﬁne a linear code. In fact, we can deﬁne an inﬁnite vector space over
F
2
, and an inﬁnite linear errorcorrecting code. In this section, we focus on
ﬁnite linear codes among classical linear codes.
The linearity provides us with beneﬁts in studying errorcorrecting
codes. For example, we might apply theory of linear algebra to error
correcting codes. Now let us deﬁne the rank, or dimension, of an error
correcting code. If the dimension of a linear code is k as a vector space over
F
2
, then the dimension of a linear code is said to be k. We call a linear
code C an [n, k] code if C ⊂ F
n
2
and the rank of C is k. Another example
of beneﬁts of linearity is to develop lowcost errordetection. Let us intro
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
184
duce a paritycheck matrix of a linear code to explain the realization of the
lowcost errordetection. Since C is a vector subspace of F
n
2
, there exists a
linear mapping H
C
from F
n
2
to F
n−k
2
such that
C = ker H
C
:= {x ∈ F
n
2
H
C
x
T
= 0}
as a binary matrix of size n × (n − k) and rankH
C
= n − k, where x
T
is
the transposed vector of x. We can represent H
C
as a binary matrix of size
(n−k)×n, which is called a paritycheck matrix of C. Conversely, a given
binary matrix H
C
deﬁnes a linear code C by putting C := {x ∈ F
n
2
H
C
x
T
=
0}. The paritycheck matrix H
C
is not uniquely determined by a given code
C. On the other hand, a linear code C is characterized by the paritycheck
matrix H
C
. In other words, an nbit sequence x is a codeword if and only if
H
C
x
T
= 0. Thus we have a characterization for x ∈ F
n
2
to be a codeword.
Let us go back to errordetection theory. For a received word y, we can
detect error by calculating H
C
y
T
. The vector H
C
y
T
∈ F
n−k
2
is called the
syndrome of y associated with H
C
. If the syndrome is not 0, then we ﬁnd
that there is at least one error in the received message. An algorithm to
calculate the syndrome H
C
y
T
and an algorithm to compare the syndrome
H
C
y
T
with a zerovector 0 work in practicaltime if the length is not so
long.
2.3. Minimum distance and [n, k, d] codes
The weight wt(x) of x = (x
1
, x
2
, . . . , x
n
) ∈ F
n
2
is deﬁned by
wt(x) = #{ix
i
= 1}.
The weight wt(x) is called the Hamming weight of x. For example,
wt(1, 0, 0, 0, 1, 1, 0) = 3. Next we introduce a distance in F
n
2
. For x, y ∈ F
n
2
,
we deﬁne the distance d(x, y) between x and y by
d(x, y) := wt(x −y).
The distance is called the Hamming distance. For example, the Hamming
distance between (1, 1, 0, 0, 0) and (1, 0, 0, 0, 1) is 2, since (1, 1, 0, 0, 0) −
(1, 0, 0, 0, 1) = (0, 1, 0, 0, 1). Note that the triangle inequality d(x, z) ≤
d(x, y) + d(y, z) holds for any x, y, z ∈ F
n
2
.
Let S be a subset of F
n
2
. For S, the minimum distance d(S) of S is
deﬁned by
d(S) := min{d(c, c
)c, c
∈ S, c = c
}.
For example, if we put S = {(1, 1, 0, 0, 0), (0, 1, 1, 0, 1), (0, 0, 1, 1, 0)}, the
minimum distance d(S) is 3.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
185
Proposition 2.1. Let C be a linear code. Then
d(C) = min{wt(c)c ∈ C, c = 0}.
Proof. Let c, c
be codewords of C such that d(C) = d(c, c
). By the linear
ity of C, c −c
is a codeword of C. Therefore d(C) = d(c, c
) = wt(c −c
) ≥
min{wt(c)c ∈ C, c = 0}.
Let x be a codeword of C such that wt(x) = min{wt(c)c ∈ C, c = 0}.
Then d(C) = min{d(c, c
)c, c
∈ C, c = c
} ≤ d(x, 0) = wt(x) =
min{wt(c)c ∈ C, c = 0}.
Consequently, d(C) = min{wt(c)c ∈ C, c = 0} holds.
A linear code C is called an [n, k, d] code if C is a block linear code
with the length n, dimension k, and minimum distance d. For example, the
3repetition code (appearing in Sec. 2.1) is a [3, 1, 3] code.
Example 2.1. Let us deﬁne a paritycheck matrix H
C
by
H
C
=
0 0 0 1 1 1 1
0 1 1 0 0 1 1
1 0 1 0 1 0 1
.
Note that the i th column of H
C
is equal to the binary expression of i. The
code deﬁned by H
C
is called a Hamming code.
We show that the Hamming code is a [7, 4, 3] code. Clearly, the length
is 7. Because the rows of H
C
are linearly independent, the dimension of
the code is 4. Thus the Hamming code is a [7, 4] code. We show that the
minimum distance of the Hamming code is 3. In other words, there is no
codeword c of the Hamming code such that 1 ≤ wt(c) ≤ 2 (Proposition 2.1).
Assume that there is a codeword c such that wt(c) = 1. Then c = e
i
, where
e
i
denotes a vector such that the i th entry is 1 and all other entries are
zero. The syndrome H
C
e
T
i
is the binary expression of i, in particular, the
syndrome is a nonzero vector, which is a contradiction. Assume that there
is a codeword c such that wt(c) = 2. Then c is a sum of two elementary
vectors e
i
and e
j
for some i, j. H
C
e
T
i
and H
C
e
T
j
are linearly independent
for any 1 ≤ i < j ≤ 7. Thus H
C
(e
i
+e
j
)
T
is a nonzero vector. On the other
hand, there is a codeword c such that wt(c) = 3. In fact, H
C
(e
1
+e
2
+e
3
)
T
=
0. Therefore, the minimum distance of the Hamming code is 3.
Remark 2.1. In general, it is diﬃcult to determine the minimum distance
of a given code. In fact, determination of the minimum distance for a given
linear code is known as a NPhard problem.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
186
2.4. Encoding
Proposition 2.2. Let C be a linear code. Then there is a binary matrix G
such that C = {xGx ∈ F
k
2
}. The matrix G is called a generator matrix.
Proof. Let k be the dimension of C. Then there are k linearly independent
vectors g
1
, g
2
, . . . , g
k
. Deﬁne a matrix G by
G =
g
1
g
2
.
.
.
g
k
.
It is easy to verify that g
1
, g
2
, . . . , g
k
form a basis of C. In other words,
g
1
, g
2
, . . . , g
k
span the code C. Therefore G is a generator matrix.
By using a generator matrix G of a linear code C, we have a simple encoding
which is multiplication of a message x and the generator matrix G.
An [n, k] code is called systematic if the ﬁrst k bits of a codeword
coinside with the original message. For a systematic code, it is easy to
recover the original message from a codeword after errorcorrection. In fact,
we have only to delete the last n − k bits to recover the original message.
The last n − k bits of a codeword of an [n, k] systematic code are called
check symbols.
If the rightside block of a paritycheck matrix is a lower triangular
matrix, then the linear code C of the paritycheck matrix is systematic. We
can encode C as a systematic code by using the paritycheck matrix. Now,
we show an example of encoding by using the paritycheck matrix. Let us
permute the columns of the paritycheck matrix in Example 2.1 as follows:
H
=
0 1 1 1 1 0 0
1 0 1 1 0 1 0
0 1 0 1 0 1 1
.
Then the linear code C
associated with H
is a [7, 4] code. Thus, the in
formation bits consist of four bits. Let x = (x
1
, x
2
, x
3
, x
4
) be a fourbit
sequence. We deﬁne a paritybit x
5
by putting x
5
= x
2
+ x
3
+ x
4
. This
deﬁnition of x
5
comes from the ﬁrst row of H
. By a similar way, we deﬁne
recursively x
6
:= x
1
+x
3
+x
4
and x
7
:= x
2
+x
4
+x
6
. By the construction
of x
5
, x
6
, x
7
, (x
1
, x
2
, . . . , x
7
) is a codeword of C
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
187
2.5. Decoding
In this review, we use the term “decoding method” to denote error
correction algorithm in this article. We should remember that the decoding
method in a restricted sense is to recover the information bits from the code
word. It is known that to decode a given linear code is a NPhard problem.
Thus the decoding algorithm does not work in real time if the length of a
given code is very long. However, if we assume some structures for a code,
then we might ﬁnd the decoding algorithm which works practically. First
we introduce the decoding algorithms for the following two linear codes:
the 3repetition code and the Hamming code (Example 2.1). Next, we in
troduce the two general decoding algorithms, called the bounded distance
decoding algorithm and the maximum likelihood decoding algorithm. Both
decoding algorithms are applicable to any linear code. We should remember
that it might not work practically for a general linear code.
Let us explain a decoding algorithm for the 3repetition code which
appeared in Sec. 2.1.
• Decoding algorithm for the 3repetition code.
Input Received word y, which consists of three bits.
Output 000 or 111.
Step 1 Count the number of 1 in the input y.
Step 2 If 1 appears two times or three times in y, then the output is 111.
Step 3 If otherwise, then the output is 000.
As was mentioned in Sec. 2.1, the decoding algorithm works successfully
against a single bitﬂip error.
The standard decoding algorithm of the Hamming code is to determine
the errorposition from a syndrome.
• Decoding algorithm for the Hamming code.
Input Received word y which consists of seven bits.
Output Estimatederror position: null or positive integer i, where 1 ≤ i ≤ 7.
Step 1 Calculate the syndrome of y associated with H (See Example 2.1).
Step 2 If the syndrome y is a zerovector, then output null.
Step 3 If the syndrome is a binary expression of some integer 1 ≤ i ≤ 7, then
output i.
If the output is null, we regard that the received word y does not contain
errors. If the output is a positive integer i, then we regard that the received
word y contains a single bit error e
i
. We shall explain the relation between
the output of the algorithm and the estimated errorvector. Assume that
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
188
the single bitﬂip error e
i
occurs on a codeword c. Then the syndrome is:
H(c + e
i
) = Hc + He
i
= He
i
= the binary expression of i.
Thus the above decoding algorithm works as a single bitﬂip error correction
algorithm. We note that we can correct a single bitﬂip error without the
received word if the syndrome is given.
Next we introduce the bounded distance decoding. Let C be a linear
code with its minimum distance d(C) = d.
• Bounded Distance Decoding
Input Received word y.
Output Estimated codeword or null.
Step 1 Choose a codeword c ∈ C.
Step 2 Calculate the Hamming distance d(c, y).
Step 3 If the Hamming distance d(c, y) is less than or equal to (d−1)/2, then
output c and stop this algorithm. If not, go to Step 4.
Step 4 If all of the codewords in C are tested by Steps 2 and 3 and failed to
produce any output in Step 3, then output null and stop this algorithm.
Step 5 Go back to Step 1.
For a linear code C with its minimum distance d, the bounded distance
decoding works well as an errorcorrection algorithm. We explain the reason
why the bounded distance decoding works (See Figure 1). For any c ∈ C, we
put a set P
c
:= {x ∈ F
n
2
wt(c −x) ≤ (d −1)/2}. We claim that P
c
∩P
c
= ∅
if c = c
. If there is a vector x ∈ P
c
∩ P
c
for some c, c
, then wt(c − c
) =
d(c, c
) ≤ d(c, x)+d(x, c
) ≤ (d−1)/2+(d−1)/2 < d. This is a contradiction.
Therefore, for a codeword c ∈ C and an errorvector e, if the received word
is y = c + e and wt(e) ≤ (d − 1)/2, then the algorithm outputs c.
Next we introduce the maximum likelihood decoding. In general, a
communication channel is deﬁned by using the term of probability distri
bution of noise. Denote the probability that a variable x is equal to d by
Pr[x = d]. The posteriori probability Pr[y = d
x = d] is deﬁned by
Pr[y = d
x = d] =
Pr[x = d, y = d
]
Pr[x = d]
.
For example, a binary symmetric channel (BSC) is a channel with a
probability distribution of bitﬂip error such that the probability distribu
tion satisﬁes
Pr[y = 0x = 0] = Pr[y = 1x = 1],
Pr[y = 1x = 0] = Pr[y = 0x = 1],
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
189
Fig. 1. Schematic idea of bounded distance decoding: Each circle is P
c
for some code
word c.
Pr[y = 0x = 0] + Pr[y = 1x = 0] = 1,
Pr[y = 0x = 1] + Pr[y = 1x = 1] = 1,
and each bitﬂip error occurs independently (See Figure 2), where x is the
input bit and y is the received bit. The maximum likelihood decoding is to
Fig. 2. Image for Binary Symmetric Channel
ﬁnd the code word c for the received word y such that
Pr [received word = yx = c] = max
c
∈C
Pr [received word = yx = c
] .
The maximum likelihood decoding seems to be the best decoding method
for errorcorrection. However, for a given linear code, to calculate the pos
teriori probability for all of codeword of a code is impossible in real time if
the code has numbers of codewords.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
190
2.6. Bit errorrate and block errorrate
The way of evaluation for errorcorrecting codes depends on the applica
tion. Let us consider a mobile phone, for example. Here we do not require
the ﬁdelity of the encoded voice. What is essential for the receiver is to un
derstand the content of the speaker’s message. Even if the users are under
noisy circumstances, the receiver may understand what the speaker says.
On the other hand, the accuracy of each bit is required for a magnetic disk
of a personal computer If only one bit in a magnetic disk is broken, then
we might never read important data in the disk again.
Two of typical (or widely used) criteria of evaluations for error
correcting codes are values of bit errorrate and block errorrate. The bit
errorrate (BER) is the expectation value of the bitﬂip error probability
for a bit in the estimated word after a decoding algorithm. In other words,
BER = Ex
c=(c
1
,c
2
,...,c
n
)∈C
n
i=1
Pr[c
i
= c
i
]
n
,
where n is the length of the code and c
i
is the i th bit of the estimated word
with the received word c. The block errorrate (BLER) is the expectation
value of the error probability for a decoding algorithm. In other words,
BLER = Ex
c∈C
[Pr[c = c
]] ,
where c is the original encoded message, and c
is the estimated word.
Remark 2.2. It is known that BLER of a classical code is closely related
to the ﬁdelity of quantum communication appearing in a quantum code
called a CSS code. A CSS code is deﬁned and is investigated later in this
review (Sec. 4).
Example 2.2. Let us calculate BLER of the 3repetition code (see Sec. 2.1)
over BSC channel with the crossover probability p. If no bitﬂip error occurs
in a transmission, then a block error does not occur. If a single bitﬂip error
occurs in a transmission, then a block error does not occur either. On the
other hand, if two or more error occur in a transmission, then a block error
occurs. Thus BLER is:
1 −(1 −p)
3
−3p(1 −p)
2
= p
3
+ 3p
2
(1 −p).
We leave a calculation of BER of the 3repetition code to for readers.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
191
3. How to Generalize Classical ErrorCorrecting Codes to
Quantum Codes
In this section, we introduce widely used terminologies to study quantum
errorcorrecting codes. Important classical objects such as paritycheck ma
trix, code space, syndrome, and others, which appeared in the previous sec
tion, are generalized to quantum errorcorrecting codes in a natural way.
Throughout this paper, matrices I, X, Z, and Y denote the following:
I =
1 0
0 1
, X =
0 1
1 0
,
Z =
1 0
0 −1
, and Y = iXZ.
3.1. Paritycheck measurements and stabilizer codes
Let H
1
, H
2
, . . . , H
n−k
be Hermitean operators such that
H
i
= H
i,1
⊗H
i,2
⊗· · · ⊗H
i,n
and
H
i
H
i
= H
i
H
i
for any 1 ≤ i, i
≤ n−k, where H
i,j
∈ {I, X, Z, Y } and ⊗ is the tensor prod
uct operation. The set of the Hermitean operators H
1
, H
2
, . . . , H
n−k
are
called a paritycheck measurement. By the condition of H
i
, its eigenval
ues are 1 and −1. A stabilizer code, which is a quantum errorcorrecting
code, is an eigenspace such that the measurement outcomes of all H
i
are 1.
Thus a stabilizer code is a subspace of C
2⊗n
. For a quantum state in C
2⊗n
,
the outcomes of a paritycheck measurement are called the syndrome.
As the names suggest, “paritycheck measurement” (resp. “syndrome”) is
closely related to “paritycheck matrix” (resp. “syndrome”) in the previous
section. We explain relations between these terms later (Proposition 3.2).
Next proposition shows the dimension of a stabilizer code and its mother
set. In classical errorcorrecting codes, the dimension of an [n, k] code is k
and the dimension of its mother set is n. For quantum errorcorrecting
codes, the dimension of the mother set C
2⊗n
is 2
n
and the dimension of a
stabilizer code is shown in the following:
Proposition 3.1. If each of the Hermitean operators H
i
is not included in
a group generated by H
j
(j = i), then the dimension of an eigenspace with
respect to a ﬁxed syndrome is 2
k
. In particular, the dimension of the related
stabilizer code is 2
k
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
192
Proof. Denote the dimension of eigenspace V
s
associated with a syndrome
s := (s
1
, s
2
, . . . , s
n−k
) by dim(V
s
), where s
i
is the measurement outcome of
the operator H
i
. For I ⊂ {1, 2, . . . , n −k}, we deﬁne a Hermitian operator
H
I
by putting H
I
:=
i∈I
H
i
. Similarly, we put s
I
:=
i∈I
s
i
.
There are two eigenspaces of H
I
and their eigenvalues are 1 and −1.
The eigenspaces of H
I
associated with an eigenvalue e is the linear subspace
V
I,e
spanned by {V
s
s
I
= e}. Thus
dimV
I,e
=
s,s
I
=e
dim(V
s
).
The dimension of V
I,1
is equal to the dimension of V
I,−1
for I = ∅, since
the trace of H
I
= dimV
I,1
−dimV
I,−1
= 0. It implies:
dimV
I,1
= dimV
I,−1
= 2
n−1
.
Next, let us calculate
I⊂{1,...,n−k},I=∅
dimV
I,s
I
. Since dimV
I,1
=
dimV
I,−1
= 2
n−1
, we have
I⊂{1,...,n−k},I=∅
dimV
I,s
I
= 2
n−1
(2
n−k
−1).
On the other hand,
I⊂{1,...,n−k},I=∅
dimV
I,s
I
=
I⊂{1,...,n−k},I=∅
(
r∈{1,−1}
n−k
,r
I
=s
I
dim(V
r
))
=
I⊂{1,...,n−k},I=∅
(dim(V
s
) +
r∈{1,−1}
n−k
,r
I
=s
I
,r=s
dim(V
r
))
= (2
n−k
−1) dim(V
s
) +
r=s
(2
n−k−1
−1) dim(V
r
)
= 2
n−k−1
dim(V
s
) + (2
n−k−1
−1)2
n
.
Thus dim(V
s
) = 2
k
.
By the following proposition, the readers may agree that a paritycheck
measurement is regarded as a generalization of a paritycheck matrix of a
classical code.
Proposition 3.2. Let H
1
= H
1,1
⊗H
1,2
⊗· · · ⊗H
1,n
, H
2
= H
2,1
⊗H
2,2
⊗
· · · ⊗ H
2,n
, . . . , H
n−k
= H
n−k,1
⊗ H
n−k,2
⊗ · · · ⊗ H
n−k,n
be a paritycheck
measurement such that H
i,j
∈ {I, Z}. Take a (classical) paritycheck matrix
H = (H
i,j
) by H
i,j
= 1 if H
i,j
= Z and H
i,j
= 0 if H
i,j
= I. Then the
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
193
stabilizer code associated with the paritycheck measurement consists of a
linear combination of quantum states c, where c is a codeword with a
paritycheck matrix H.
Proof. It is obvious that a quantum state c is an element of the stabilizer
code. Thus a linear subspace spanned by {cc ∈ C} is contained in the
stabilizer code. These spaces are identiﬁed by comparing the dimension of
them.
3.2. Encoding
Encoding should be a unitary operation. Thus the dimension of information
qubits must be equal to the dimension of encoded qubits. On the other hand,
encoding is a map from a 2
k
dimensional space to C
2⊗n
. Therefore we need
to add n − k initial states 0 to the information (i.e. preencoded) qubits.
We would like to leave the following simple question as an exercise for
readers: Generalize a classical encoding by the lower triangle paritycheck
matrix to an encoding for a quantum errorcorrection code by a paritycheck
measurement.
3.3. Decoding
By the law of quantum mechanics, it is impossible to measure the received
state itself which is transmitted over a noisy channel. Instead of the received
state which is a quantum codeword of a stabilizer code, we can measure a
syndrome associated with the paritycheck measurement. We have already
seen an errorcorrection method by using syndrome in Example 2.1. It ex
pects that we might correct quantum errors by using the syndrome, if the
stabilizer code is well designed.
4. CSS Codes
A stabilizer code is called a CSS code if there exists a paritycheck mea
surement such that each observable H
i
consists of tensor product of either
I, X or one of I, Z.
4.1. Twisted condition
By the deﬁnition of a CSS code, we can separate the observables to two sets
{H
i
}
i=1...t
and {H
i
}
i=t+1...t+t
for some t, t
such that H
i
= Z
hc
i,1
⊗Z
hc
i,2
⊗
· · · ⊗ Z
hc
i,n
for 1 ≤ i ≤ t and H
t+i
= X
hd
i
,1
⊗ X
hd
i
,2
⊗ · · · ⊗ X
hd
i
,n
for
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
194
1 ≤ i
≤ t
, where hc
i,j
and hd
i
,j
∈ F
2
. Then we can construct two classical
paritycheck matrices H
C
= {hc
i,j
}
i=1...t
, H
D
= {hd
i,j
}
i=1...t
.
Let C and D be classical linear codes with the paritycheck matrices
H
C
and H
D
respectively. By the commutativity of the observables, the
classical linear codes C and D satisfy the following condition: each row
of H
C
and each row of H
D
are orthogonal with the inner product , ,
where (a
1
, a
2
, . . . , a
n
), (b
1
, b
2
, . . . , b
n
) =
a
i
b
i
. In fact, observables H
i
and H
t+j
are commutative if and only if Z of H
i
and X of H
t+j
meet even
times. It is equivalent to claim that there are even positions p such that
hc
i,p
= hd
j,p
= 1. In other words,
H
C
×H
T
D
= 0,
where A
T
is the transposed matrix of a matrix A.
Let us recall that the (classical) dual code of a linear code. For a linear
code C, the dual code C
⊥
is deﬁned by
C
⊥
= {x ∈ F
n
2
x, c = 0, ∀c ∈ C}.
In other words, the dual code is a linear code generated by rows of the
paritycheck matrix. Finally, we conclude that the linear codes C and D
obtained from paritycheck measurements of CSS codes above satisfy the
following condition:
D
⊥
⊂ C(⇔ C
⊥
⊂ D).
We call this condition the twisted condition.
Proposition 4.1. Let C (resp. D) be linear codes with a paritycheck ma
trix H
C
(resp. H
D
). C and D satisfy the twisted condition if and only if
H
C
and H
D
are orthogonal i.e. H
C
×H
T
D
= 0.
Proof.
D
⊥
⊂ C
⇐⇒ ∀x ∈ D
⊥
, x ∈ C
⇐⇒ ∀x ∈ D
⊥
, H
C
x
T
= 0
⇐⇒ ∀ row x of H
D
, H
C
x
T
= 0
⇐⇒ H
C
×H
T
D
= 0
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
195
4.2. Characterization of code space
Lemma 4.1. Let D be a classical linear code and D
⊥
the dual code of D.
For x ∈ F
n
2
,
d
∈D
⊥
(−1)
x,d
= #D
⊥
(resp. = 0), if x ∈ D (resp. if
otherwise), where #X is the cadinarity of a set X.
Proof. If x ∈ D, then x, d
= 0 for all d
∈ D
⊥
. Thus
d
∈D
⊥
(−1)
x,d
= #D
⊥
.
If otherwise, there is d
0
∈ D
⊥
such that x, d
0
= 1. Since {d
0
+ dd ∈
D
⊥
} = D
⊥
,
d
∈D
⊥
(−1)
x,d
=
d
∈D
⊥
(−1)
x,d
+d
0
= (−1)
d
∈D
⊥
(−1)
x,d
.
Therefore
d
∈D
⊥
(−1)
x,d
= 0.
Theorem 4.1. Let C and D be classical linear codes which satisfy twisted
condition. Then the code space of the CSS code, deﬁned by C and D, is
spanned by the following quantum states:
{
d
∈D
⊥
c +d
}
c∈C
.
Proof. Denote the Hadamard operator by H. (Do not confuse the
Hadamard operator H and a paritycheck matrix H
C
.)
Let Q be a spanned by {
d
∈D
⊥
c + d
c ∈ C}. By the deﬁnition,
the dimension of Q is equal to dimC/D
⊥
= dimC − dimD
⊥
. Remember
that the number of observables of the paritycheck measurement is (n −
dimC) + dimD
⊥
. By Theorem 3.1, the dimension of the stabilizer code
is equal to that of Q. Thus it is suﬃcient to show that a quantum state
φ :=
d
∈D
⊥
c +d
is a state of the CSS code.
Since D
⊥
⊂ C and c +d ∈ D
⊥
, it is obvious that H
i
c +d = c +d for
the observable H
i
associated with H
C
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
196
Now, we act a Hadamard operator to each qubit of φ as H
⊗n
φ.
H
⊗n
φ =
d
∈D
⊥
H
⊗n
c +d
= 2
−n/2
d
∈D
⊥
(0 + (−1)
c
i
1)
= 2
−n/2
x∈F
n
2
d
∈D
⊥
(−1)
x,c+d
x
= 2
−n/2
x∈F
n
2
(−1)
x,c
d
∈D
⊥
(−1)
x,d
x
= 2
−n/2+dimD
⊥
d∈D
(−1)
c,d
d.
In particular, the last term above is a linear combination of d with d ∈ D.
Thus for the observable H
i
associated with H
D
, we have
H
i
φ = H
⊗n
(H
⊗n
H
i
H
⊗n
)(H
⊗n
φ) = H
⊗n
(H
⊗n
φ) = φ.
Hence, Q is included in the CSS code.
4.3. Correctable error
Let E and R be quantum operators. We call E a set of correctable errors
associated with R if there exists α ∈ C such that (R ◦ E)(ρ) = αρ for
any codeword state ρ of quantum errorcorrecting code. The operator R
is called the recovery operator. By a recovery operator, it measures a
syndrome of ρ or its partial information. And then, it performs quantum
operations which depend on the syndrome or the partial information.
Proposition 4.2. Let E
X
be a subset of F
n
2
and E
X
a subset of F
n
2
deﬁned
by E
X
= {X
⊗e
= X
e
1
⊗X
e
2
⊗. . . X
e
n
e = (e
1
, . . . , e
n
) ∈ E
X
}. It is a subset
of bitﬂip errors associated with E
X
.
Then E
X
is a set of correctable bitﬂip errors associated with some re
covery operator R if and only if e −e
is not in C \ D
⊥
for any e, e
∈ E
X
.
Proof. If e −e
is not an element of C for any e
∈ E
X
, then the syndrome
H
C
e
T
and H
C
(e
)
T
are diﬀerent from each other. Thus we can deﬁne a
map r from the syndrome space to E
X
such that r(H
C
e
T
) = e. Therefore
we can construct a recovery operator R such that R ◦ X
e
(ρ) = ρ for any
X
e
∈ E
X
.
Conversely, assume that there exists e
∈ E
X
such that e −e
∈ C. Then
the syndrome of e − e
associated with H
C
is 0 because e − e
∈ C. This
implies H
C
e = H
C
e
. Since the behavior of a recovery operator depends on
the syndrome, (R ◦ X
e
)(ρ) = (R ◦ X
e
)(ρ) holds for any encoded state ρ.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
197
If e − e
∈ D
⊥
⊂ C, then X
e
ρ = X
e
ρ holds by the structure of a CSS
code. Thus (R ◦ X
e
)(ρ) = (R ◦ X
e
)(ρ) holds.
Assume e−e
is not in D
⊥
but e−e
∈ C and assume there is a recovery
operator R such that (R◦X
e
)(ρ) = ρ and (R◦X
e
)(ρ) = ρ for any encoded
state ρ. Put φ =
d
∈D
⊥
d
and ψ =
d
∈D
⊥
d
+e −e
. Since e −e
is not in D
⊥
, φ = ψ hold. On the other hand, φ = (R ◦ X
e
)(φ) =
(R ◦ X
e
)(ψ) = ψ because X
e
φ = X
e
ψ holds. It is a contradiction.
By a similar argument, we have the following:
Proposition 4.3. Let E
Z
be a subset of F
n
2
and E
Z
a subset of F
n
2
deﬁned
by E
Z
= {Z
⊗e
= Z
e
1
⊗ Z
e
2
⊗ . . . Z
e
n
e = (e
1
, . . . , e
n
) ∈ E
Z
} is a subset of
bitﬂip errors associated with E
Z
.
Then E
Z
is a set of correctable phaseﬂip errors if and only if e − e
is
not in D \ C
⊥
for any e, e
∈ E
Z
.
Because of the independence of syndromes associated with H
C
and H
D
,
we have the following:
Proposition 4.4. Let Q be a CSS code. Let E
X
and E
Z
be subsets of F
n
2
.
Deﬁne E
X
:= {X
⊗e
e ∈ E
X
} and E
Z
:= {Z
⊗e
e ∈ E
Z
}. If E
X
and E
Z
are
sets of correctable errors for Q, then E := {X
⊗e
Z
⊗e
e ∈ E
X
, e
∈ E
Z
} is
also a set of correctable errors for Q.
By the linearity of a recovery operator, we have the following:
Theorem 4.2. Let E = {E
i
} be a set of correctable errors and R a set of
recovery operators for a quantum code Q. Deﬁne a set F :=
i
m
i
E
i
for
m
i
∈ C. Then F is a correctable errors by the same recovery operator R.
Remember that any unitary operator over C
2
is a linear combination of
I, X, Z and Y . By Theorem 4.2, if we prove that the errors I, X, Z and Y
are correctable for a ﬁxed position, then any unitary error for the position is
correctable. Furthermore, if any tbit error is correctable on classical error
correcting codes C and D, then any tquantum error is correctable on the
CSS code associated with C and D.
4.4. 7Qubit code (quantum Hamming code)
In this subsection, we would like to introduce an example of a CSS code.
The example is a generalization of the classical Hamming code.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
198
Proposition 4.5. Let C and D be Hamming codes (See Example 2.1).
Then C and D satisfy the twisted condition.
Proof. Let H be a paritycheck matrix of the Hamming code in Example
2.1. Then the matrix H is a paritycheck matrix H
C
(resp. H
D
) of C (resp.
D). It is easy to verify H
C
×H
T
D
= 0.
The Hamming code is a [7, 4, 3] code. In particular, it can correct any
single bitﬂip error. Thus it should be possible to correct any quantum
single position error with a CSS code Q derived from a pair of Hamming
codes. We call the CSS code a 7qubit code (or a quantum Hamming
code).
In other words, X
i
is a tensor product of seven unitary matrices with
X in the i th position and the others are identity matrices. By a similar
way, we deﬁne Z
i
and (XZ)
i
. Let E be a single quantum error in the 7
qubit code on the i th position. Then E = aI
⊗7
+ bX
i
+ cZ
i
+ dXZ
i
, for
some a, b, c, d ∈ C. For a quantum message φ, the received message with
a quantum error E is
Eφ = aφ + bX
i
φ + cZ
i
φ + dXZ
i
φ.
The syndrome s
X
of a paritycheck measurement associated with H
C
is
(1, 1, 1) or ((−1)
i1
, (−1)
i2
, (−1)
i3
), where (i
1
, i
2
, i
3
) is the binary expression
of i. Similarly, the syndrome s
Z
of a paritycheck measurement associated
with H
D
is also (1, 1, 1) or ((−1)
i1
, (−1)
i2
, (−1)
i3
). Thus there are four
possibility of the syndrome. For each possibility, we can recover the original
state φ by the following:
• If s
X
= s
Z
= (1, 1, 1), then the quantum state after the measurement
is φ. Thus we recover the original state.
• If s
X
= ((−1)
i1
, (−1)
i2
, (−1)
i3
) and s
Z
= (1, 1, 1), then the quantum
state after the measurement is X
i
φ. Thus we can recover the original
state by multiplying X
i
to the state.
• If s
X
= (1, 1, 1) and s
Z
= ((−1)
i1
, (−1)
i2
, (−1)
i3
), then the quantum
state after the measurement is Z
i
φ. Thus we can recoever the original
state by multiplying Z
i
to the state.
• If s
X
= ((−1)
i1
, (−1)
i2
, (−1)
i3
) and s
Z
= ((−1)
i1
, (−1)
i2
, (−1)
i3
),
then the quantum state after the measurement is XZ
i
φ. Thus we
can recover the original state by multiplying XZ
i
to the state.
We would like to leave the following excise for readers: Let Q be a 7qubit
code. Put Q
i,0
:= X
i
Q, Q
0,j
:= Z
j
Q and Q
i,j
:= X
i
Z
j
Q for 1 ≤ i, j ≤ 7.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
199
Then 64 quantum subspaces Q, Q
i,0
, Q
0,j
, Q
i,j
(1 ≤ i, j ≤ 7) are orthogonal
to each other.
5. Quantum LDPC Codes
5.1. LDPC codes and sumproduct decoding
LDPC codes, which are introduced by Gallager, are classical linear codes
and their broadersense deﬁnition is a code whose paritycheck matrix has
fewer 1’s than 0’s.
2
LDPC stands for “LowDensity ParityCheck”. LDPC
codes achieve outstanding performance. Its performance approaches the
maximum likelihood decoding over some channels with practical assump
tions.
The standard decoding algorithm for LDPC codes is called sum
product decoding. We introduce the sumproduct decoding and explain
that it is applicable to a CSS code with a sparse paritycheck measurement.
We do not mention why the combination of LDPC codes and sumproduct
decoding works well. Interested readers may verify the highperformance of
this combination by computer simulations.
We denote the (m, n)entry of a binary matrix H by H
m,n
, and we put
A(m) := {nH
m,n
= 1}, B(n) := {m : H
m,n
= 1}. We assume that we
know the error probability distribution of a communication channel, the
noise of the channel does not depend on the original message and the noise
occurs independently in each bit. By the syndrome of a received message,
sumproduct (syndrome) decoding outputs an error vector. The algorithm
of sumproduct (syndrome) decoding is the following:
• Input: Paritycheck matrix H, and syndrome s = (s
1
, s
2
, . . . , s
M
)
T
of
the received word y i.e. s = Hy
T
.
• Output: Estimated error vector (e
1
, e
2
, . . . , e
N
).
• Step 1: For a pair (m, n) with H
m,n
= 1, put q
m,n
(0) = 1/2, q
m,n
(1) =
1/2, set the iteration counter to l = 1 and introduce the maximum
number of iteration l
max
.
• Step 2: For m = 1, 2, . . . , M and all of the pairs (m, n) with H
m,n
= 1,
update r
m,n
(0), r
m,n
(1) by the following formulae:
r
m,n
(0) = K
c
i
∈{0,1},
i∈A
m,n
\n,
c
i
=s
m
n
∈A(m)\n
q
m,n
(c
n
)Pr(0x
n
= c
n
)
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
200
r
m,n
(1) = K
c
i
∈{0,1},
i∈A
m,n
\n,
c
i
=1−s
m
n
∈A(m)\n
q
m,n
(c
n
)Pr(0x
n
= c
n
),
where K is a constant such that r
m,n
(0) +r
m,n
(1) = 1.
• Step 3: For n = 1, 2, . . . , N and all of pairs (m, n) with H
m,n
= 1,
update q
m,n
(0) and q
m,n
(1) by the following formula: q
m,n
(i) =
K
m
∈B(n)\m
r
m
,n
(i), where i = 0, 1 and K
is a constant such that
q
m,n
(0) +q
m,n
(1) = 1.
• Step 4: For n = 1, 2, . . . , N, calculate the following:
Q
n
(i) = K
Pr(y
n
x
n
= i)
m
∈B(n)
r
m
,n
(i)
ˆ e
n
:= 0, if Q
n
(0) ≥ Q
n
(1)
ˆ e
n
:= 1, otherwise,
where i = 0, 1 and K
is a constant such that Q
n
(0) +Q
n
(1) = 1. We
call (ˆ e
1
, ˆ e
2
, . . . , ˆ e
N
) a temporary estimated error vector.
• Step 5: Check if the temporary estimated error vector satisﬁes syn
drome check. In other words, if
(ˆ e
1
, ˆ e
2
, . . . , ˆ e
N
)H
T
= (s
1
, s
2
, . . . , s
M
)
holds then the algorithm outputs (ˆ c
1
, ˆ c
2
, . . . , ˆ c
N
) as an estimated error
vector and the algorithm stops.
• Step 6: If l < l
max
, then l := l +1 and return to step 2. If l = l
max
, then
the algorithm outputs (ˆ e
1
, ˆ e
2
, . . . , ˆ e
N
) as an estimated error vector and
the algorithm stops.
By the deﬁnition of sumproduct algorithm, the output depends on a parity
check matrix. The errorcorrecting performance changes according to the
paritycheck matrix, even when the code space is the same.
5.2. Application of the sumproduct algorithm for the
errorcorrection of CSS codes
Let C and D be LDPC codes which satisfy the twisted condition and let
H
C
and H
D
be their lowdensity paritycheck matrices, respectively. As we
have remarked, the sumproduct algorithm outputs an errorvector from a
syndrome. By the deﬁnition of a CSS code, we can decompose the parity
check measurement into two parts, which are the Zpart associated with
H
C
, and the Xpart associated with H
D
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
201
After receiving a quantum message, we measure it by the paritycheck
measurement and obtain the syndrome. We decompose the syndrome into
two parts s
Z
and s
X
according to H
C
and H
D
. We input the paritycheck
matrix H
C
and one of the syndrome s
Z
to the sumproduct algorithm. Then
the algorithm outputs the estimated errorvector eC = (eC
1
, eC
2
, . . . , eC
n
).
We act X
eC
= X
eC
1
⊗ X
eC
2
⊗ X
eC
n
to the received quantum state. This
implies we expect to correct the bitﬂip error. Next we input H
D
and s
Z
to the sumproduct algorithm and obtain the estimated errorvector eD.
Then we act Z
eD
to the received quantum state. Consequently we expect
that the received state is corrected to the original quantum message.
5.3. Classical and quantum quasicyclic LDPC codes
Let I(∞) be a zero matrix and I(1) a circulant permutation matrix, i.e.
I(1) =
¸
¸
¸
¸
¸
¸
1
1
.
.
.
1
1
¸
.
Put I(b) = I(1)
b
for an integer b.
If a paritycheck matrix H
C
of C has the following form:
H
C
=
I(c
0,0
) I(c
0,1
) . . . I(c
0,L−1
)
I(c
1,0
) I(c
1,1
) . . . I(c
1,L−1
)
.
.
.
.
.
.
.
.
.
.
.
.
I(c
J−1,0
) I(c
J−1,1
) . . . I(c
J−1,L−1
)
¸
¸
¸
¸
¸
,
then C is called a quasicyclic (QC) LDPC code, where c
i,j
∈ [P
∞
] :=
{0, 1, . . . , P −1} ∪ {∞} and P is the size of a circulant matrix I(1).
Let us denote a matrix which consists of the indices of H
C
by H
C
, in
other words,
H
C
=
c
0,0
c
0,1
. . . c
0,L−1
c
1,0
c
1,1
. . . c
1,L−1
.
.
.
.
.
.
.
.
.
.
.
.
c
J−1,0
c
J−1,1
. . . c
J−1,L−1
¸
¸
¸
¸
¸
.
We call H
C
the model matrix of H
C
.
Imagine that we have a random lowdensity paritycheck matrix. Be
cause of the randomness, we need large memory to store the paritycheck
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
202
matrix. On the other hand, we do not need large memory to keep quasi
cyclic paritycheck matrix because the paritycheck matrix is reconstructed
by its model matrix. Quasicyclic LDPC codes are good not only from the
viewpoint of memory, but also from the viewpoint of errorcorrecting per
formance. In particular, with sumproduct decoding, it is expected that
quasicyclic LDPC codes have good performance for short length codes.
6
5.4. Twisted condition for quasicyclic LDPC codes
We give some necessary and suﬃcient conditions for QCLDPC codes C
and D to satisfy the twisted condition. Let us denote rows of index ma
trices H
C
and H
D
of quasicyclic LDPC matrices H
C
and H
D
by c
j
and d
k
, respectively, in other words, c
j
:= (c
j,0
, c
j,1
, . . . , c
j,L−1
), d
k
:=
(d
k,0
, d
k,1
, . . . , d
k,L−1
). For Example 5.1, we obtain c
0
= (1, 2, 4, 2, 4, 1) and
d
1
= (6, 5, 3, 5, 3, 6).
For an integer sequence x = (x
0
, x
1
, . . . , x
L−1
), we call x multiplic
ity even if each entry appears even times in x
0
, x
1
, . . . , x
L−1
except for
the symbol ∞. For example, (0, 1, 1, 0, 3, 3, 3, 3, ∞) is multiplicity even, but
(0, 2, 2, 4, 4, 4, 0) is not.
Let us deﬁne the diﬀerence operation “−” over [P
∞
] by as follows: for
x, y ∈ {0, 1, . . . , P −1}, put
x −y := x −y (mod P),
and
x −∞= ∞−x = ∞−∞:= ∞.
Theorem 5.1. Let C and D be QCLDPC codes with model matrices H
C
and H
D
respectively such that their size of the circulant matrices, which are
components of the paritycheck matrices, are the same. The codes C and D
satisfy the twisted condition if and only if c
j
− d
k
is multiplicity even for
any row c
j
of H
C
and any row d
k
of H
D
.
Proof. We divide H
C
into J rowblocks H
C
1
, H
C
2
, . . . , H
C
(J−1)
:
H
C
j
:= (I(c
j,0
), I(c
j,1
), . . . , I(c
j,L−1
)), 0 ≤ j < J.
Similarly, we divide H
D
into K rowblocks. These codes C and D satisfy
the twisted condition if and only if H
C
× H
T
D
= 0 (Proposition 4.1). It is
easy to verify that any row of H
C
and any row of H
D
are orthogonal if
and only if any row of H
C
j
and any row of H
D
k
are orthogonal for any
0 ≤ j < J and 0 ≤ k < K.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
203
Denote ath row of H
C
j
by C
a
and bth row of H
D
k
by D
b
. Thus C
a
and
D
b
are binary vectors. Deﬁne a binary matrix X := (x
a,b
)
0≤a,b<P
, where
x
a,b
:= C
a
, D
b
= C
a
× D
T
b
. Then for 0 ≤ j < J and 0 ≤ k < K, X = 0
if and only if any row of H
C
j
and any row of H
D
k
are orthogonal to each
other.
By the deﬁnition of X, X = H
Cj
× H
T
Dk
. And by I(i)
T
= I(−i), we
have H
Cj
× H
T
Dk
=
0≤l<L
I(c
j,l
− d
k,l
). Since I(x) is a binary circulant
permutation matrix for any integer x, X = 0 if and only if c
j
− d
k
is
multiplicity even.
5.5. CSS QCLDPC codes from rightshifted matrices
We call a matrix M a rightshifted matrix if M has the following form:
M =
m
0
m
1
. . . m
L−1
m
L−1
m
0
. . . m
L−2
.
.
.
.
.
.
m
1
m
2
. . . m
0
,
where m
i
∈ [P
∞
].
Let A and B be rightshifted matrices over [P
∞
]. Let us deﬁne model
matrices H
C
and H
D
of linear codes C and D respectively by putting:
H
C
:= [A, B] and H
D
:= [−B
T
, −A
T
].
Then we have the following propositions:
Proposition 5.1. The QCLDPC codes C and D above satisfy the twisted
relation.
Proof. By theorem 5.1, a pair of LDPC codes C and D satisﬁes the twisted
relation if and only if c
j
−d
k
is a multiplicity even vector for any pair of a
row c
j
of H
C
and a row d
k
of H
D
.
Denote the ﬁrst rows of A and B by (a
0
, a
1
, . . . , a
L−1
) and
(b
0
, b
1
, . . . , b
L−1
), respectively. Then we can rewrite c
j
− d
k
= (a
1−j
+
b
k−1
, a
2−j
+b
k−2
, . . . , a
L−j
+b
k−L
, b
1−j
+a
k−1
, b
2−j
+a
k−2
, . . . , b
L−j
+a
k−L
).
Then the entries of the lefthalf of c
j
−d
k
are
{a
0
+b
k−j
, a
1
+b
k−j−1
, . . . , a
L−1
+b
k−j−L+1
},
and the entries of the righthalf of c
k
−d
j
are
{a
0
+b
k−j
, a
1
+b
k−j−1
, . . . , a
L−1
+b
k−j−L+1
}.
Thus c
j
−d
k
is a multiplicity even vector.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
204
Two linear codes C and D are called equivalent if one can be obtained
from the other by a permutation of the coordinates.
Proposition 5.2. The linear codes C and D in Proposition 5.1 are equiv
alent and C/D
⊥
and D/C
⊥
are equivalent also.
Proof. For the paritycheck matrices H
C
= (hc
i,j
) and H
D
= (hd
i,j
),
where hc
i,j
, hd
i,j
∈ {0, 1} and hc
i,j
= hd
L−i,2L−j
holds. Thus C and D are
equivalent.
Let us denote the permutation by σ such that C = σ(D). We note that
σ(D
⊥
) = σ(D)
⊥
.
5.6. MacKay’s code
The study of LDPC quantum codes was developed by MacKay, who pro
posed the following construction of a LDPC CSS code:
7
H
C
= H
D
=
A A
T
,
where A is a binary rightshifted matrix.
MacKay’s code is a special case of the codes derived from Proposition
5.1. In fact, we obtain MacKay’s code by putting the size of the circulant
matrix P = 1 and B = A
T
.
5.7. Regular LDPC codes
An LDPC code is called (ρ, λ)regular if the numbers of 1’s in any rows and
any columns of a paritycheck matrix are constants λ and ρ, respectively.
The parameter λ is called the column weight and ρ is called the row
weight. At the beginning of the study of LDPC codes, the construction
research has focused on regular LDPC codes. In the design of regular LDPC
codes, the following three parameters are regarded important: the column
weight λ, the row weight ρ, and the girth.
The Tanner graph of a paritycheck matrix {h
i,j
}
i,j
is a bipartite graph
with the vertices of rows and columns of the paritycheck matrix and the
edges between the ith row and the jth column with h
i,j
= 1. The girth
is the length of the shortest cycle of the Tanner graph, which is deﬁned
below, of a lowdensity paritycheck matrix. It is thought generally that a
small girth has bad eﬀects on the performance of sumproduct decoding.
If the Tanner graph has a short cycle, then the high performance of sum
product decoding is not expected. If the Tanner graph has no cycle, i.e. the
graph is a tree, then it is expected that the performance of the sumproduct
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
205
algorithm is the same as that of a maximally likelihood decoding. By the
regularity of a paritycheck matrix associated with a regular LDPC code, it
is unavoidable to contain cycles in the Tanner graph. Thus it is important
to attain a big girth in the Tanner graph as one of the research directions
(See Refs. 8 and 9 for details).
We would like to leave the following exercise to readers: compare error
correcting performance of girth4 LDPC codes with that of girth6 LDPC
codes by computer simulation.
If C = D and C
⊥
⊂ C, the CSS code obtained by C is called self
dual containing.
7,10
For example, the Mackay’s code, appearing in the
previous subsection, is a selfdual containing code. With quasicyclic LDPC
conditions, we have the following negative result for selfdual containing.
Proposition 5.3. Let C be a (λ, ρ) LDPC code with λ ≥ 2. If C is selfdual
containing, then the girth of the LDPC code is 4.
Proof. C
⊥
⊂ C is equivalent to H
C
× H
T
C
, where H
C
is a paritycheck
matrix of C. If λ ≥ 2, then there exist rows h
i
= (h
i,1
, h
i,2
. . . , ) and
h
j
= (h
j,1
, h
j,2
, . . . ) of H
C
such that h
i,k
= h
j,k
= 1 for some k. By
assumption C
⊥
⊂ C, we have h
i
× h
T
j
= 0. On the other hand, h
i
× h
T
j
=
p
h
i,p
h
j,p
= 1 +
p=k
h
i,p
h
j,p
. Thus there must be an index k
such that
k = k
and h
i,k
= h
j,k
= 1.
Remark 5.1. Let C (resp. D) be a classical code associated with MacKay’s
code and H
C
(resp. H
D
) a paritycheck matrix of the MacKay’s code. Then
C
⊥
⊂ C holds. By Proposition 5.3, the girth of the Tanner graph associated
with H
C
is 4.
On the other hand, the girth of the Tanner graph derived from the
construction in Proposition 5.1 is not limited to 4. In fact, we realize the
girth6 codes which satisfy the twisted condition (Theorem 5.3).
In order to construct a code with good errorcorrecting performance,
we should avoid having the size4 cycle, i.e. the girth to be 4. If this is the
case, we denote the following condition by (G):
(G) The girth of the Tanner graph of a paritycheck matrix is greater than
or equal to 6.
Proposition 5.4 (Ref. 11). A necessary and suﬃcient condition to have
girth ≥ 6 in the Tanner graph representation of the paritycheck matrix of
a QCLDPC code with the index matrix {c
i,j
}
i,j
is c
j
1
,l
1
− c
j
1
,l
2
+ c
k
1
,l
2
−
c
k
1
,l
1
= 0 for any 0 ≤ j
1
< j
2
< J, 0 ≤ l
1
< l
2
< L.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
206
Proof. If there is a sequence c
j
1
,l
1
, c
j
1
,l
2
, c
j
2
,l
2
, and c
j
2
,l
1
such that c
j
1
,l
1
−
c
j
1
,l
2
+ c
k
1
,l
2
−c
k
1
,l
1
= 0 , then there is a cycle of length 4. In fact, entries
H
1+Pj
1
,c
j
1
,l
1
+Pl
1
, H
1+Pj
1
,c
j
1
,l
2
+Pl
2
, H
1+Pj
2
,c
j
2
,l
2
+Pl
2
, H
1+Pj
2
,c
j
2
,l
1
+Pl
1
are all 1’s in the paritycheck matrix and form a cycle of length 4.
Conversely, if there is a cycle of length 4, which consists of
H
a
1
+Pj
1
,b
1
+Pl
1
, H
a
2
+Pj
1
,b
2
+Pl
2
, H
a
3
+Pj
2
,b
3
+Pl
2
, and H
a
4
+Pj
2
,b
4
+Pl
1
,
then c
j
1
,l
1
−c
j
1
,l
2
+ c
k
1
,l
2
−c
k
1
,l
1
= 0 holds.
It is obvious that Proposition 5.4 is equivalent to Proposition 5.5 be
low. A term “multiplicity free” means all the entries of a given vector
are diﬀerent. For example, (0, 1, 2, 3, 4, 5, 6, 7) is multiplicity free, however,
(0, 0, 1, 2, 3, 4, 5) is not.
Proposition 5.5. A necessary and suﬃcient condition for a QCLDPC
code C to have girth ≥ 6 is c
j
1
− c
j
2
be multiplicity free for any 0 ≤ j
1
<
j
2
< J.
Theorem 5.2. There is no selfdual containing QCLDPC code which sat
isﬁes simultaneously (G) and the twisted condition without ∞ in the index
matrix.
Proof. This statement is already proven in Proposition 5.3. We give an
other proof in terms of QuasiCyclic LDPC code theory.
Assume that C and D are a pair of linear codes which are the ingredients
of a selfdual containing CSS code. It implies D = C. By Proposition 5.1,
c
j
− c
j
must be multiplicity even for any 0 ≤ j, j
< J. It follows that
the diﬀerence cannot by multiplicity free. On the other hand, c
j
−c
j
must
be multiplicity free if (G) holds by Proposition 5.5. Therefore (G) and the
twisted condition cannot hold simultaneously.
5.8. Construction of CSS LDPC codes
The integers module P, denoted Z
P
, is the set of (equivalence classes of)
integers {0, 1, . . . , P − 1}. Addition, subtraction, and multiplication in Z
P
are performed module P.
Proposition 5.6 (Ref. 12). For any positive integer P ≥ 2, Z
∗
P
:= {z ∈
Z
P
z
−1
, which is the inverse element of z i.e. z(z
−1
) = (z
−1
)z = 1, exists}
is an abelian (i.e. commutative) group with multiple operation of Z
P
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
207
Proof. The commutativity holds obviously. The identity is 1 ∈ Z
∗
P
. For
z ∈ Z
∗
P
, the inverse z
−1
is in Z
∗
P
because of z = (z
−1
)
−1
. Finally, for
x, y ∈ Z
∗
P
, there exists (xy)
−1
= x
−1
y
−1
.
Proposition 5.7 (Ref. 12). Let G be a group and H a subgroup of G.
Put [g] := {hgh ∈ H} for g ∈ G. Then we have
[g] ∩ [g
] = ∅ ⇐⇒ [g] = [g
] ⇐⇒ gg
−1
∈ H ⇐⇒ g
g
−1
∈ H.
Proof. Let g, g
∈ G. Assume [g] ∩ [g
] = ∅. Then there exist h, h
∈ H
such that hg = h
g
. This implies g = h
−1
h
g
. For any g
0
∈ G, there exists
h
0
∈ H such that g
0
= h
0
g by the deﬁnition of [g]. By g ∈ [g
], we have
g
0
= h
0
g = h
0
h
−1
h
g
∈ [g
]. Thus [g] ⊂ [g
] holds. By a similar argument,
[g] ⊃ [g
] holds. Therefore [g] = [g
].
If [g] = [g
], then there is h
∈ H such that g = h
g
. This implies
g/g
= h
∈ H.
If g/g
∈ H, then
(g/g
)
−1
= g
/g ∈ H.
If g
/g ∈ H, then there is h ∈ H such that
g
/g = h.
Because g
∈ [g
] and hg ∈ [g], g
= hg ∈ [g
] ∩ [g].
Theorem 5.3. Let P be an integer > 2 and σ an element of Z
∗
P
with the
even order L := ord(σ) = #Z
∗
P
, where ord(σ) := min{m > 0σ
m
= 1}.
Pick any τ ∈ Z
∗
P
\ {1, σ, σ
2
, . . . }. Put
c
j,l
:=
σ
−j+l
0 ≤ l < L/2
τσ
−j−1+l
L/2 ≤ l < L,
d
k,l
:=
−τσ
−k−1+l
0 ≤ l < L/2
−σ
−k+l
L/2 ≤ l < L
and put H
C
= [c
j,l
]
0≤j<J,0≤l<L
, H
D
= [d
k,l
]
0≤k<K,0≤l<L
, where 1 ≤ J, K ≤
L/2. Then C and D satisfy (G) and the twisted condition.
Example 5.1. Let P = 7, (Z
∗
P
= {1, 2, 3, 4, 5, 6}) and choose σ = 2 and
τ = 3. Then L/2 = ord(2) = 3 and we put J = K = L/2 (= 3). Then the
paritycheck matrices are obtained from Theorem 5.3 as:
H
C
=
1 2 4 3 6 5
4 1 2 5 3 6
2 4 1 6 5 3
, H
D
=
4 1 2 6 5 3
2 4 1 3 6 5
1 2 4 5 3 6
.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
208
Proof. By Proposition 5.1, the twisted condition holds.
Next, we prove that (G) holds for C by Proposition 5.5. Fix any 1 ≤
a < b ≤ J. Denote the ﬁrst L/2 entries of c
a
−c
b
by X, and the remaining
L/2 entries by Y . Then we have X = (σ
−a
− σ
−b
) × (σ, σ
2
, σ
3
, . . . , σ
L/2
)
and Y = (σ
−a
−σ
−b
)τ ×(1, σ, σ
2
, . . . , σ
L/2−1
).
Thus the condition that c
a
− c
b
be multiplicity free is equivalent to
[σ
−a
−σ
−b
] = [τ(σ
−a
−σ
−b
)] with notation in Proposition 5.7, It is obvious
that [σ
−a
−σ
−b
] = [τ(σ
−a
−σ
−b
)] if and only if [1] = [τ]. By the choice of
τ, [1] = [τ] holds.
We quote the following fundamental facts from group theory and num
ber theory:
Proposition 5.8 (Ref. 12). If P is a prime, then Z
∗
P
is a cyclic group.
In other words, there exists z ∈ Z
∗
P
such that Z
∗
P
= {1, z, z
2
, . . . z
P−2
}.
Theorem 5.4 (Dirichlet’s Theorem
13
). Let A
n
:= b + an. Then there
are inﬁnitely many primes in the series A
1
, A
2
, . . . provided gcd(a, b) = 1,
where gcd(a, b) is the greatest common divisor for a and b.
Theorem 5.5 (Ref. 14). Let C and D be QCLDPC codes without ∞ in
index matrices.
(i) For any odd L > 0, there is no pair C and D such that both matrices
satisfy (G) and the twisted condition simultaneously.
(ii) For any even L > 0 and any 1 ≤ J, K ≤ L/2, there exists P such
that C is a regular (J, L) QCLDPC code, D is a regular (K, L) QCLDPC
code with the circulant matrix size P, and C and D satisfy both (G) and
the twisted condition.
Proof. (i) Let L be an odd integer and C and D QCLDPC codes. To
satisfy the twisted condition for C and D, it is necessary for c
j
− d
k
be
multiplicity even by Proposition 5.1. However, it is impossible since L is
odd.
(ii) Let L be an even integer and J and K integers with 1 ≤ J, K < L/2.
In case L = 2, put P = 3, σ = 1, and τ = 2. Then Theorem 5.3 provides
us with a pair of linear codes which satisfy (G) and the twisted condition.
From here on, we assume L ≥ 4. Theorem 5.5 holds if there exists a
prime P = 1 + (L/2)n, x > 2, Since Z
∗
P
is a cyclic group of order (L/2)n
(Proposition 5.8), there is a generator z. Put σ := z
n
, then we have ord(σ) =
L/2 and a set Z
∗
P
\ {1, σ, σ
2
, . . . } is not empty. Thus we can pick τ up from
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
209
Z
∗
P
\ {1, σ, σ
2
, . . . }. Theorem 5.5 is a consequence of Theorem 5.3. On the
other hand, such a prime P exists by Dirichlet’s theorem (Theorem 5.4).
5.9. Performance of errorcorrection and minimum
distance
We performed simulations for evaluating errorcorrecting performance of
our codes with sumproduct decoding, which is generally used as a decoding
method for LDPC codes, and with some parameters (J, L, P, σ, τ), where
the symbols J, L, P, σ, τ are deﬁned in Theorem 5.3. We assumed that the
channel is binary symmetric and chosen the maximal iteration l
max
= 128.
Fig. 3. Simulation Results: Block error rate (BLER) and Crossover probability of BSC
Figure 3 shows simulation results for our codes C. The list of the pa
rameters (J, L, P, σ, τ) are written bottom right part of the ﬁgure. The code
length is deﬁned by L × P. The vertical axis shows the block errorrate of
the errorcorrecting performance. Recall that the block errorrate is deﬁned
by:
the num. of “errorcorrection fails”
the total num. of simulations
.
The reasons why our simulation results shows the blockerror rate of
C/D
⊥
over the bitﬂip binary symmetric channel is the followings:
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
210
• The code D/C
⊥
is equivalent to C/D
⊥
up to the bitpositions. Thus
the simulation results of the blockerror rate of D/C
⊥
over the phase
error symmetric channel is the same as the blockerror rate of C/D
⊥
.
• If we assume that the bitﬂip error and the phaseﬂip error occur inde
pendently, the ﬁdelity of the quantum communication over the noisy
channel by the quantum CSS code is equal to
B
+
P
−
B
P
, where
B
and
P
are the blockerror rates of C/D
⊥
and D/C
⊥
over given
crossover probabilities respectively.
If J < L/2 or K < L/2, the minimum distances of classical codes
C/D
⊥
and D/C
⊥
derived from Theorem 5.3 and the related quantum codes
cannot exceed L. We know a codeword with the weight L in the paritycheck
matrix H
C
or H
D
. Imagine the block errorrate of the related quantum
code over the bitﬂip crossover rate 0.03 and the phaseﬂip crossover rate
0.03. If we assume the both errors occur independently, it is expected that
approximately 6% of the qubits are ﬂipped by both or one of the errors.
It is known that only half of the minimum distance of the code is cor
rectable by the minimum distance decoding. Remember that the length
of our code with ID (4, 10, 61, 9, 49) is 610, the expected number of error
is 0.06 × 610 36, and the minimum distance is at most 10. The mini
mum distance should be larger than 72. However, we do not have to worry
about it since our decoding method is not the minimum distance decod
ing but the sumproduct decoding. See the block errorrate of the code
ID (4, 10, 61, 9, 49) in Figure 3. The block error rate of C/D
⊥
is less than
0.0048 below the crossover probability 0.03 of the channel. Hence the block
errorrate of the quantum CSS code is less than 0.01. Thus it is expected
that the block error of the related quantum code occurs only one time
per 100 times communications over the bitﬂip crossover rate 0.03 and the
phaseﬂip crossover rate 0.03.
6. Postscript
In this review, we discuss quantum errorcorrecting codes from the view
point of a generalization of classical errorcorrecting codes. In particular,
we have focused on the CSS LDPC codes in the second half of this article.
Quasicyclic LDPC codes are one of the most important classes of classical
LDPC codes. In the framework of Quasicyclic LDPC codes, it is expected
to develop quantum errorcorrecting coding theory by ring theory, discrete
mathematics, combinatorics and others. It is quite a new and interesting
subject.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
211
There are several topics on QECC left out in this article. For example,
the bound of codes is a meaningful topic for coding theory. Quantum
singleton bound is the one of famous bounds for quantum errorcorrecting
codes.
3
How many qubits do we need to encode a single qubit such that the
quantum errorcorrecting code overcomes any single quantum error? The
answer is n = 5.
3
As we have shown in Section 2.1, the answer for classical
case is three. Thus the quantum bound is diﬀerent from the classical bound.
The LDPC codes in this paper are constructed by two classical LDPC
codes. There is another generalization of classical LDPC codes to quantum
errorcorrecting codes. That is the class of stabilizer LDPC codes. This
is a natural generalization of LDPC codes in the view of paritycheck mea
surement of stabilizer code. The interesting problem of stabilizer LDPC
codes is a decoding algorithm. At a glance, we ﬁnd many of cycles of length
4 in the paritycheck measurement because of computatibity of observables.
As was mentioned in Section 5.7, many cycles of length 4 cause the decod
ing algorithm failure. Good references of stabilizer codes are Refs. 15,16
and 17.
If an LDPC code is not regular, then the code is called an irregu
lar LDPC code. It is known that welldesigned irregular classical LDPC
codes have better errorcorrecting performance than regular ones. In fact,
it achieves almost the theoretical bound, called the Shannon limit. A simi
lar result is expected for quantum irregular LDPC codes. Indeed, quantum
irregular LDPC codes proposed in Refs. 18 show better errorcorrecting
performance than the codes in Section 5.7.
Acknowledgments
The author would like to thank to the staﬀ of Summer School on Mathe
matical Aspects of Quantum Computing, 27(Mon)29(Wed) August 2007,
at Kinki University, to Dr. Chinen, and to the chief organizer Prof. Naka
hara.
This research was partially supported by GrantsinAid for Young Sci
entists (B), 18700017, 2006.
References
1. F. J. MacWilliams and N. J. A. Sloane, The Theory of ErrorCorrecting
Codes, NorthHolland Mathematical Library (NorthHolland, 1977), Chap
ters 9 and 10.
2. R. G. Gallager, LowDensity ParityCheck Codes (Cambridge, MA: MIT,
1963).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
212
3. M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Infor
mation, Cambridge Series on Information and the Natural Sciences (Cam
bridge University Press, 2001).
4. C. E. Shannon, A Mathematical Theory Of Communication (Parts 1 and 2),
Bell System Technical Journal 27 (1948) 379–423.
5. J. H. van Lint, Introduction to Coding Theory, Graduate Texts in Mathe
matics vol. 86 (2nd Edition, Springer Verlag, 1992).
6. LANMAN Standards Commitee, http://ieee802.org/
7. D. MacKay, G. Mitchison, and P. McFadden, Sparse Graph Codes for Quan
tum ErrorCorrection, quantph/0304161; IEEE Trans. Inform. Theory 50
(10) (2004) 2315–2330.
8. X.Y. Hu, E. Eleftheriou, and D. M. Arnold, Progressive edgegrowth Tan
ner graphs, in Proceedings of Global Telecommunications Conference 2001
(GLOBECOM ’01), vol. 2 (IEEE, 2001) 995–1001.
9. R. M. Tanner, D. Sridhara, and T. Fuja, A Class of GroupStructured LDPC
Codes, in Proceedings of International Symposium on Communication The
ory and Applications (ISCTA, Ambleside, England), (2001).
10. M. Hamada, Information Rates Achievable with Algebraic Codes on Quantum
Discrete Memoryless Channels, IEEE Trans. Inform. Theory 51 (12) (2005)
4263–4277.
11. M. Fossorier, QuasiCyclic LowDensity ParityCheck Codes From Circulant
Permutation Matrices, IEEE Trans. Inform. Theory 50 (8) (2004) 1788–1793.
12. W. K. Nicholson, Introduction to Abstract Algebra (3rd Edition, Wiley,
2006).
13. M. R. Schroeder, Number Theory in Science and Communication (4th Edi
tion, Springer, 2005).
14. M. Hagiwara and H. Imai, Quantum QuasiCyclic LDPC Codes, in Proceed
ings of 2007 IEEE International Symposium on Information Theory (ISIT)
(2007).
15. T. Camara, H. Ollivier, and J.P. Tillich, Constructions and performance of
classes of quantum LDPC codes, quantph/0502086.
16. J.P. Tillich, T. Camara, and H. Ollivier, A class of quantum LDPC code:
construction and performances under iterative decoding, in Proceedings of
2007 IEEE International Symposium on Information Theory (ISIT) (2007).
17. P. Tan and T. J. Li, On Construction of Two Classes of Eﬃcient Quantum
ErrorCorrection Codes, in Proceedings of 2007 IEEE International Sympo
sium on Information Theory (ISIT) (2007).
18. M. Hagiwara and H. Imai, A Simple Construction of Quantum QuasiCyclic
LDPC Codes, in Proceedings of 2007 Hawaii and SITA Joint Conference on
Information Theory (2007).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
213
CONTROLED TELEPORTATION OF AN ARBITRARY
UNKNOWN TWOQUBIT ENTANGLED STATE
VAHIDEH EBRAHIMI
, ROBABEH RAHIMI
†
, MIKIO NAKAHARA
‡
1
Interdiciplinary Graduate School of Science and Engineering, Kinki University,
HigashiOsaka, Osaka 5778502, Japan
1
Department of Physics, Kinki Univeristy, HigashiOsaka 5778502, Japan
Email: ebrahimi@alice.math.kindai.ac.jp
†
Email: rahimi@alice.math.kindai.ac.jp
‡
Email: nakahara@math.kindai.ac.jp
We investigate controlled quantum teleportation of an arbitrary entangled
state. It is shown [Man et al., J. Phys. B 40 (2007) 1767] that by using EPR
and GHZ states, an unknown Nqubit entangled state can be teleported from
a sender to a receiver, under control of a third party. This protocol is studied
in more details in our work. The protocol is simulated for a small number of
qubits. It is found that a W state, instead of the GHZ state, should be used
for a successful teleportation of an entangled state. The diﬀerence between
the initial state to be teleported and the ﬁnal state reproduced is studied for
diﬀerent amount of entanglement of the Wlike state.
Keywords: Teleportation, Entanglement.
1. Introduction
Quantum information processing is performed by using classical and quan
tum resources. A classical communication channel is relatively cheap com
pared to quantum resources, such as entanglement. Therefore, we have to
reduce the required quantum resources as much as possible. The quantum
resources required are protocol dependent. The majority of the protocols
are hard to implement in practice since production of multiqubit entangled
states is not easy at present and the diﬃculty increases considerably with
the number of qubits taking part in entanglement.
1
Recently, Man et al.
2
have shown that an unknown Nqubit entangled
state can be teleported from a sender to a receiver, under the control of a
third party. By simulating this protocol for a small number of qubits, we
have found that a W state is required so that the protocol of the controlled
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
214
teleportation works ﬁne. For further investigation, we have used Wlike
entangled states for this protocol. The diﬀerence between the initial state
for controlled teleportation and the ﬁnal state reproduced by the receiver
is studied for a range of a maximally entangled state to a maximally mixed
state. Reasonably, we have found the best agreement between the initial
state and the ﬁnal state is attained if a maximallyentangled W state is
employed.
2. Controlled Teleportation of Bipartite Entanglement
Suppose Alice teleports her Nqubit state to Charlie, the receiver, under
the M controllers Bob
1
, Bob
2
, . . ., Bob
M
. This protocol works for M < N
and consumes M GHZ states plus (N − M) EPR pairs.
2
In the controlled
teleportation of a twoqubit entangled state according to the original pro
posal,
2
Alice, Bob, and Charlie share a GHZ state of which the ﬁrst (the
second and the third) qubit is held by Alice (Bob and Charlie) while an
EPR pair is shared only between Alice and Charlie.
Analysis of the designed quantum circuit based on the original proposal
2
revealed that Charlie cannot receive entangled two qubits from Alice under
the control of Bob, as far as the initial tripartite shared state is a GHZ
state. It is due to this fact that tracing out Bob’s qubit in the GHZ state
leaves the other twoqubit system in a mixed classical state. This result
shows that the controlled teleportation of bipartite entanglement cannot
be done by using the proposal provided in Ref. 2, while it works correctly
by using the other maximally entanglement tripartite state (the W state):
Tracing out Bob’s qubit in the W state leaves an EPR state shared by Alice
and Charlie.
We have found the best agreement between the initial state to be tele
ported and the ﬁnal state reproduced by the receiver when the original
shared multipartite state is a W state. As the amount of the entanglement
of a Wlike state decreases, the ﬁnal state deviates from the expected initial
EPR state to be teleported.
References
1. F. G. Deng, C. Y. Li, Y. S. Li, H. Y. Zhou, Y. Wang, Phys. Rev. A 72
(2005) 022338.
2. Z. X. Man, Y. J. Xia, N. B. An, J. Phys. B: At. Mol. Opt. Phys. 40 (2007)
1767.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
215
NOTES ON THE D
¨
UR–CIRAC CLASSIFICATION
YUKIHIRO OTA
Department of Physics, Kinki university,
Higashi–Osaka, 577–8502, Japan
Email: yota@alice.math.kindai.ac.jp
MOTOYUKI YOSHIDA
∗,§
, ICHIRO OHBA
∗,†,‡,¶
∗
Department of Physics, Waseda University,
Tokyo, 169–8555, Japan
†
Research Institute for Materials Science and Technology, Waseda University,
Tokyo, 169–0051, Japan
‡
Research Institute for Science and Engineering, Waseda University,
Email:
§
motoyuki@hep.phys.waseda.ac.jp
¶
ohba@waseda.jp
One cannot always obtain information about entanglement by the D¨ ur–Cirac
(DC) method. The impracticality is attributed to the decrease of entanglement
by local operations in the DC method. Even in 2–qubit systems, there exist
states whose entangled property the DC method never evaluates. This result
is closely related to the discussion of puriﬁcation protocols. In addition, we
discuss its eﬀectiveness in multiqubit systems.
Keywords: Multiparticle Entanglement, Local Operations, Fully Entangled
Fraction, Logarithmic Negativity.
1. Introduction
A systematic and eﬀective way to classify multiparticle entanglement in
N–qubit systems was proposed by D¨ ur and Cirac .
1
We call it the D¨ ur–
Cirac (DC) method. The main idea is explained below. Using a sequence
of local operations (LO), one can transform an arbitrary density matrix of
an N–qubit system into a state whose entangled property is easily exam
ined. Note that entanglement cannot increase through LO. If the density
matrix transformed by LO is entangled, the original one represents an en
tangled state. However, one cannot always obtain an entangled property by
it. We suggested there exists an impracticality in the DC method through
an example .
2
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
216
2. Results
First, we explain there exists such impracticality even in 2–qubit systems .
3
One can never determine whether a quantum state is entangled or not by
the DC method if the fully entangled fraction
4
is less than or equal to 1/2.
This result is closely related to puriﬁcation protocols. We concentrate on
the LO in the recurrence method .
5
The arbitrary density matrix of a 2–
qubit system can be transformed into the one which is almost the same as
in the DC method. It can be distilled by this method if and only if its fully
entangled fraction is greater than 1/2. Next, let us examine its eﬀectiveness
in an N–qubit system (N ≥ 3). We focus on the following type of density
matrices: ρ
+
BD
=
2
N−1
−1
j=0
µ
+
j
ψ
+
j
ψ
+
j
, where µ
+
j
≥ 0. The corresponding
eigenvectors are given by ψ
+
j
= (0j + 1¯ )/
√
2 (
¯
j = 2
N−1
− 1 − j).
Let us consider the density matrix transformed from ρ
BD
by the LO in
the DC method. With respect to an arbitrary bipartition of the system,
we can ﬁnd its logarithmic negativity, which is known as an upper bound
of the distillable entanglement ,
6
is not vanishing, as far as the maximal
eigenvalues of ρ
+
BD
are not degenerate. Nevertheless, the strict posotivity of
the distillable entanglement for ρ
+
BD
has not yet shown.
3. Summary
In a 2–qubit system, one can’t always obtain the entangled property of a
quantum state by the DC method, if the fully entangled fraction is less
than or equal to 1/2. This result is consistent with the consideration on
workability of a puriﬁcation protocol for a 2–qubit. The situation is more
complicated in a multiqubit system. We have partially understood its eﬃ
ciency; one may obtain the entangled property of ρ
+
BD
.
References
1. W. D¨ ur and J. I. Cirac, Phys. Rev. A 61 (2000) 042314.
2. Y. Ota, S. Mikami, M. Yoshida and I. Ohba, Accepted for publication in J.
Phys. A; quantph/0612158.
3. Y. Ota, M. Yoshida and I. Ohba, arXiv:0704.1375 [quantph].
4. G. Alber, T. Beth, M. Horodecki, P. Horodecki, R. Horodecki, M. R¨otteler,
H. Weinfurter, R. Werner and A. Zeilinger, Quantum Information: An In
troduction to Basic Theoretical Concepts and Experiments (Springer, Berlin,
2001).
5. C. H. Bennett, D. P. DiVincenzo, J. A. Smolin and W. K. Wootters, Phys.
Rev. A 54 (1996) 3824.
6. G. Vidal and R. F. Werner, Phys. Rev. A 65 (2002) 032314.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
217
BANGBANG CONTROL OF ENTANGLEMENT IN
SPINBUSBOSON MODEL
∗
ROBABEH RAHIMI
1
, AKIRA SAITOH
2†
, MIKIO NAKAHARA
1‡
1
Interdiciplinary Graduate School of Science and Engineering, Kinki University,
HigashiOsaka, Osaka 5778502, Japan
Email: rahimi@alice.math.kindai.ac.jp
‡
Email: nakahara@math.kindai.ac.jp
2
Graduate School of Engineering Science, Osaka University,
Toyonaka, Osaka 5608531, Japan
†
Email: saitoh@qc.ee.es.osakau.ac.jp
Electron spin bus molecules have been used for implementations of quantum
computing, including realizing entanglement. However, systems consisting elec
tron spins are usually unstable. We have studied electron spin bus molecular
systems in order to introduce methods with which the produced states i.e. en
tangled states, stay stable. According to the operational evidences, the problem
is numerically studied by introducing a model of decoherence and applying fast
multipulses for suppressing decoherence. We found that bangbang pulses with
realistic duty ratios can somehow suppress decoherence.
Keywords: Entanglement, Decoherence, Electron Spin Bus Quantum Comput
ing.
1. Introduction
Molecular systems involving electron spins have been used for implemen
tations of quantum computing, specially for realizing quantum entangle
ment.
1–3
One of the advantages of this system for realizing a quantum
computer is that entanglement between an electron spin and a nuclear spin
can be achieved under milder experimental conditions compared to the case
of an eg. NMR system where entanglement is being realized between two
nuclear spins. However, the physical system involving an electron spin is
usually very fragile since an electron spin possesses very short decoherence
time. In an attempt to remove this drawback , we have studied the electron
spin bus systems.
∗
A precise discussion can be found in J. Phys. Soc. Japan 76 (2007) 114007.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
218
2. The Model of Decoherence
In a conventional model boson modes are distributed uniformly over all
spin systems in a solid material. In fact, this model ignores the natural
dissipation of bosons through a cavity and/or a sample holder. Then under
the inﬂuence of noise, by using a conventional noise model, oscillations of
entanglement rather than a decay of entanglement is found since the total
system involving the spins and the bosonic bath is a closed system, which
is integrable. We consider another model in which bosons dissipate into an
extra bath of bosons of a cavity and/or a sample holder. According to the
operational evidences, the dynamic of entanglement for an electron spin
bus system is numerically solved by employing the introduced model of
decoherence and applying fast multipulses for suppressing decoherence.
3. Results
It is mentioned that the system we deal with is particularly useful for mak
ing an entangled state. This is due to the fact that high spin polarization
can be transferred from a polarized electron spin to a less polarized nuclear
spin. The total state after transmission, would be highly polarized, thus
appropriate for generalizing entanglement. Here, we study two cases of one
with polarization transfer and else without any polarization transfer pulse.
In a numerical study, we take a set of realistic parameters representing
a commercial Qband machine (magnetic ﬁeld of 35 GHz for an electron
nuclear double resonance, ENDOR, machine), that is currently available for
our experiment. In case no polarization transfer is used, a high negativity
is achieved when the temperature is less than 10 mK; if, in contrast, the
ideal polarization transfer is used, a high negativity is achieved when the
temperature is even at around 10
3
mK.
Totally we conclude that decoherence matter is rather more challenging,
while entanglement is comparatively easy to be established for an electron
spin bus molecule system for quantum computing. When number of qubits
is small, we still ﬁnd some regions of parameters where the quantum state
can be stable, if careful tuning of bangbang pulses are made.
References
1. M. Mehring, J. Mende, W. Scherer, Phys. Rev. Lett. 90 (2003) 153001.
2. R. Rahimi: PhD Thesis, Osaka Univ., 2006, quantph/0609063.
3. W. G. Morley, J. van Tol, A. Ardavan, K. Porfyrakis, J. Zhang,
G. A. D. Briggs, quantph/0611276.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
219
NUMERICAL COMPUTATION OF TIMEDEPENDENT
MULTIPARTITE NONCLASSICAL CORRELATION
AKIRA SAITOH, ROBABEH RAHIMI
†
,
MIKIO NAKAHARA
†
, MASAHIRO KITAGAWA
Department of Systems Innovation, Graduate School of Engineering Science, Osaka
University, 13 Machikaneyama, Toyonaka, Osaka 5608531, Japan
†
Interdisciplinary Graduate School of Science and Engineering, Kinki University,
341 Kowakae, Higashi Osaka, Osaka 5778502, Japan
Nonclassical correlation of a multipartite quantum system is an indis
pensable constituent of quantum information processing (QIP).
1
There is a
commonly recognized paradigm
2
in which a multipartite quantum system
described by a density matrix having no product eigenbasis is considered to
possess nonclassical correlation. The set of such density matrices includes
the set of entangled states as shown in Fig. 1(a). Eﬀective QIP requires
quantumness to be stable against noise.
There is a longterm study of suppressing decoherence in quantumness.
Here, we focus on the bangbang control scheme
3
together with a sort of
spinboson linear coupling models. Consider a spin system S (consisting of
Separable
Density matrices
Having a product
eigenbasis
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0 50 100 150 200
t [µs]
N without BB
D without BB
N x = 10
2
D x = 10
2
N x = 10
4
D x = 10
4
(a) (b)
Fig. 1. (a) Set of density matrices having no product eigenbasis (shaded region).
Its complementary set is nonconvex. (b) Time evolutions of D and negativity (N)
with/without bangbang control (x is the duty ratio tried in the numerical simulation).
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
220
n spins 1/2) coupled with a onresonance bosonic system B replaced with
a thermal one with dissipation probability p for a time interval τ. The sys
tem Hamiltonian is set to H = H
S
+ H
B
+ H
c
with H
S
=
n
j=1
f
j
S
j
+
j,k,j<k
A
jk
S
j
S
k
(spin system Hamiltonian), H
B
=
n
j=1
f
j
a
†
j
a
j
(Hamil
tonian of bosonic modes), and H
c
=
n
j=1
c
j
S
j
(a
†
j
+ a
j
). Here, f
j
is the
precession frequency of the jth spin (f
j
= f
k
for j = k), A
jk
is a spinspin
coupling constant, c
j
is a spinboson onresonance coupling constant. We
consider a bangbang pulse irradiation with pulse width L and duty ratio
x [i.e., each X
π
pulse is irradiated with the duration xL followed by the
absence time interval (1 − x)L]. The bangbang pulses for individual spins
are synchronized.
We numerically evaluate the eﬀect of bangbang pulses on the time
evolution of a nonclassicalcorrelation measure deﬁned as follows. Consider
a density matrix ρ
[1,...,m]
of an mpartite system. Consider local complete
orthonormal bases {e
[1]
j
}
j
, . . . , {e
[m]
z
}
z
. Then, a measure of nonclassical
correlation of our interest is deﬁned by
D(ρ
[1,...,m]
) = min
local bases
(−
j,...,z
p
j,...,z
log
2
p
j,...,z
) − S
vN
(ρ
[1,...,m]
)
with p
j,...,z
= e
[1]
j
e
[2]
k
 · · · e
[m]
z
ρ
[1,...,m]
e
[1]
j
e
[2]
k
· · · e
[m]
z
. The value of
D(ρ
[1,...,m]
) is zero if ρ
[1,...,m]
has a (fully) product eigenbasis. In addition,
D(ρ
[1,...,m]
) is invariant under local unitary operations.
One of simulation results is illustrated in Fig. 1(b). For this simulation,
we set the following parameter values: n = 2, f
1
= 3.40 × 10
10
[Hz], f
2
=
4.87 × 10
7
[Hz], A
12
= 1.00 × 10
7
[Hz], temperature 1mK, τ = 1 × 10
−6
[s],
L = 1.0 × 10
−7
[s], and p = 0.1. We apply an entangling operation, an
Hadamard gate followed by a CNOT gate, to the spins originally in the
thermal state to make the initial state for t = 0. Time evolutions of D
and negativity (one of entanglement measures) with/without bangbang
control are shown in the ﬁgure. Suppression of decoherence is signiﬁcant
for the duty ratio x = 1.0 × 10
−4
.
References
1. J. Gruska, Quantum Computing (McGrawHill, London, 1999); M. A. Nielsen
and I. L. Chuang, Quantum Computation and Quantum Information (Cam
bridge University Press, Cambridge, 2000).
2. J. Oppenheim, M. Horodecki, P. Horodecki, and R. Horodecki, Phys. Rev.
Lett. 89 (2002) 180402.
3. M. Ban, J. Mod. Opt. 45 (1998) 2315; L. Viola and S. Lloyd, Phys. Rev. A
58 (1998) 2733.
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
221
ON CLASSICAL NOCLONING THEOREM UNDER
LIOUVILLE DYNAMICS AND DISTANCES
TAKUYA YAMANO
∗
, OSAMU IGUCHI
Department of Physics, Ochanomizu University,
211 Otsuka, Bunkyoku, Tokyo 1128610, Japan
∗
Email: yamano@skycosmos.phys.ocha.ac.jp
We have reported that the classical counterpart of the quantum nocloning,
which has been shown for the KullbackLeibler distance under Liouville dy
namics, can be proven also by non Csisz´ar fdivergence type distance.
Keywords: NoCloning, Liouville Dynamics, Distance Measure.
Nocloning theorem
1
tells that no operation O which allows devices to
copy an arbitrary unknown states  Ψ
s
to a target states  0
t
exists,
 m⊗  Ψ
s
⊗  0
t
O
−→ m
Ψ
⊗  Ψ
s
⊗  Ψ
s
(1)
where the symbols s,m, and t denote source, machine and target, respec
tively. It has been proven that this theorem has a classical analogue,
2,3
where the crucial logic of the proof is to show a contradiction. That is, the
KullbackLeibler (KL) distance (a special case of the Csisz´ar fdivergence
4
)
is conserved under the Liouvile dynamics (LD) ∂
t
P = −∇· (vP), where v is
the velocity associated with generalized coordinates in phase space, on the
contrary, the KL distances before and after the process were shown to be not
equivalent. To what extent is this analogue robust? Is this peculiar to the
KL distance? The situation is not so obvious as the quantum case, where the
von Neumann relative entropy S(σ  ρ) = Tr(σ log σ −σ log ρ) between two
states σ and ρ is conserved under unitary evolution ρ(t) = U(t)ρ(0)U
†
(t)
etc. To answer this question, we have investigated whether or not the con
tradiction exists for the diﬀerent class (non Csisz´ar fdivergence type) of
distance.
The distance is deﬁned as C
α
= −log[
dxP
α
1
P
1−α
2
], where 0 ≤ α ≤ 1
and P
1,2
are two diﬀerent statistical ensemble distributions of the sys
tem with phase space coordinate x = (x
m
, x
s
, x
t
). C
α
(P
1
, P
2
) ≥ 0 holds
March 11, 2008 8:59 WSPC  Proceedings Trim Size: 9in x 6in MathematicalAspects
222
with equality if and only if P
1
= P
2
when α = 0 and α = 1. We can
obtain that the derivative of C
α
vanishes when P
1,2
obey the LD, i.e.,
dC
α
(P
1
, P
2
)/dt = 0.
5
This means that we can expect the conservation
of the distance as C
α
(P
1
, P
2
) = C
α
(Q
1
, Q
2
) for two corresponding dif
ferent ﬁnal (after copying) distributions Q
1,2
. This expectation, however,
is found to be unachievable. The factorization of the initial distributions
P
1,2
= P
m
(x
m
)P
1,2
s
(x
s
)P
t
(x
t
) are assumed in the copying process. The
marginalization for the ﬁnal distribution with respect to the coordinate m
is a premise i.e.,
dx
m
Q
1,2
= P
1,2
s
P
1,2
t
. The successful cloning produces
Q
1,2
= Q
m
P
1,2
s
P
1,2
s
as the ﬁnal distribution, which corresponds to Eq.(1).
Noting that for the two factorized distributions the total distance is a sum
of the distances of component systems, we obtain C
α
(P
1
, P
2
) = C
α
(P
1
s
, P
2
s
)
by C
α
(P
m
, P
m
) = 0 and C
α
(P
t
, P
t
) = 0. This means that the distance be
tween P
1
and P
2
is attributed to the distance between the probabilities of
the sources. With the help of the H¨older inequality, the distance after the
copying is then evaluated as
C
α
(Q
1
, Q
2
) ≥ −log
dx
s,t
dx
m
Q
1
α
dx
m
Q
1
1−α
= −2 log
dx
s
(P
1
s
)
α
(P
2
s
)
1−α
= 2C
α
(P
1
, P
2
). (2)
This clearly shows the discrepancy. In summary, the assertion of the clas
sical counterpart of the nocloning theorem is achieved based on a non
Csiszar’s fdivergence distance under the Liouville dynamics. The copying
machine cannot produce a statistical ensemble clone under this dynamics.
The more details of the present consideration will be provided elsewhere.
5
Acknowledgments
The authors thank Dr. T. Ootsuka for inviting to participate in the Summer
school at Kinki University, Japan.
References
1. W. K. Wootters, W. H. Zurek, Nature 299 (1982) 802.
2. A. R. Plastino, A. Daﬀertshofer, Phys. Rev. Lett. 93 (2004) 138701.
3. A. Daﬀertshofer, A. R. Plastino, A. Plastino, Phys. Rev. Lett. 88 (2002)
210601.
4. I. Csisz´ar, Studia Math. Hungarica 2 (1967) 299.
5. T. Yamano, O. Iguchi, Classical nocloning under Liouville dynamics by non
Csisz´ ar fdivergence, in preparation.
Mathematical Aspects of Quantum Computing 2007
Kinki University Series on Quantum Computing
EditorinChief: Mikio Nakahara (Kinki University, Japan) ISSN: 17937299
Published Vol. 1 Mathematical Aspects of Quantum Computing 2007 edited by Mikio Nakahara, Robabeh Rahimi (Kinki Univ., Japan) & Akira SaiToh (Osaka Univ., Japan)
Kinki University Series on Quantum Computing – Vol. 1
editors
Mikio Nakahara Robabeh Rahimi
Kinki University, Japan
Akira SaiToh
Osaka University, Japan
Mathematical Aspects of Quantum Computing 2007
World Scientific
NEW JERSEY
•
LONDON
•
SINGAPORE
•
BEIJING
•
SHANGHAI
•
HONG KONG
•
TA I P E I
•
CHENNAI
NJ 07601 UK office: 57 Shelton Street. MA 01923. . Singapore 596224 USA office: 27 Warren Street. ISBN13 9789812814470 ISBN10 9812814477 Printed in Singapore. recording or any information storage and retrieval system now known or to be invented. Hackensack. All rights reserved. 5 Toh Tuck Link. Suite 401402. For photocopying of material in this volume. This book. 1 Copyright © 2008 by World Scientific Publishing Co. Pte. or parts thereof. 222 Rosewood Drive. please pay a copying fee through the Copyright Clearance Center. MATHEMATICAL ASPECTS OF QUANTUM COMPUTING 2007 Kinki University Series on Quantum Computing — Vol. Ltd..Published by World Scientific Publishing Co. without written permission from the Publisher. Pte. including photocopying. electronic or mechanical. may not be reproduced in any form or by any means. Inc. In this case permission to photocopy is not required from the publisher. Covent Garden. USA. London WC2H 9HE British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library. Ltd. Danvers.
Invited speakers to the summer school were asked to prepare their lecture notes in a selfcontained way and therefore we expect that each contribution will be useful for students and researchers even with less background to the topic. we would like to thank Ms Zhang Ji of World Scientiﬁc for her excellent editorial work. 2007 at Kinki University in Osaka. Finally. Osaka. Sports. Japan. Culture. We yearn for continued outstanding success in this ﬁeld and wish to expand summer schools and conferences further in the ﬁeld of quantum computing based at Kinki University. and to motivate students to tackle the advanced subjects of quantum computing. This summer school was supported by “Open Research Center” Project for Private Universities: matching fund subsidy from MEXT (Ministry of Education. The aims of this summer school were to exchange and share ideas among researchers working in various ﬁelds of quantum computing particularly from a mathematical point of view.v PREFACE This volume contains lecture notes and poster contributions presented at the summer school “Mathematical Aspects of Quantum Computing”. January 2008 Mikio Nakahara Robabeh Rahimi Akira SaiToh . Science and Technology). held from 27 to 29 August.
This page intentionally left blank .
Japan) An Introduction to Entanglement Theory Shogo Tanimura (Kyoto University. Japan) Quantum Error Correction Codes Poster Presentations Vahideh Ebrahimi (Kinki University. Japan 27 – 29 August 2007 27 August Mikio Nakahara (Kinki Univerisity. Japan) Holonomic Quantum Computing and Its Optimization 29 August Sahin Kaya Ozdemir (Osaka University. Japan) Controlled Teleportation of an Arbitrary Unknown TwoQubit Entangled State .vii Summer School on Mathematical Aspects of Quantum Computing Kinki Univerisity. Japan) Playing Games in Quantum Mechanical Settings: Features of Quantum Games Manabu Hagiwara (National Institute of Advanced Industrial Science and Technology. Japan) Braid Group and Topological Quantum Computing 28 August Damian J. Markham (Tokyo University. Osaka. Japan) Quantum Computing: An Overview Kazuhiro Sakuma (Kinki University. H.
Japan) Notes on D¨ rCirac Classiﬁcation u Robabeh Rahimi (Kinki University.viii Yukihiro Ota (Kinki University. Japan) BangBang Control of Entanglement in SpinBusBoson Model Akira SaiToh (Osaka University. Japan) On Classical NoCloning Theorem under Liouville Dynamics and Distances . Japan) Numerical Computation of TimeDependent Multipartite Nonclassical Correlation Takuya Yamano (Ochanomizu University.
ix .
x .
xi .
This page intentionally left blank .
Kaiki Taro Kikuta. Japan Kinki University. Toshiyuki Kobata. Yasushi Markham. Japan Kyoto University. Kaori Nagaoka. Japan National Institute of Advanced Industrial Science and Technology. Japan Waseda University. Akiko Yoshida. Yukihiro Ozdemir. Japan Kinki University. Japan Kinki University. Makoto Yamano. Japan Osaka University. Japan Yokohama National University. Japan Kinki University. Hiroyuki Unoki. Japan Tokyo University. Japan . Kumi Kondo. Japan Osaka University. Shogo Tomita. Etsuo Sugita. Koji Ebrahimi Bakhtavar. Shoyu Nakagawa. Japan Kinki University. Nobuo Nakahara. Robabeh SaiToh. Vahideh Hagiwara. Japan Kinki University. Japan Osaka City University. Ayumu Tanimura. Damian James Harold Minami. Sahin Kaya Rahimi Darabad. Japan Osaka City University. Takuya Yamanouchi. Manabu Ikeda. Japan Waseda University. Japan Kinki University. Japan Kinki University. Japan Kinki University. Takashi Chinen.xiii LIST OF PARTICIPANTS Aoki. Japan ITOCHU TechnoSolutions Corporations. Motoyuki Kinki University. Takayoshi Ota. Yasushi Inoue. Japan Kinki University. Japan Kinki University. Japan Kyoto University. Hiroshi Segawa. Japan Kinki University. Akira Sakuma. Japan Kinki University. Kazuhiro Sasano. Japan Ochanomizu University. Mikio Ootsuka. Japan Kinki University. Japan Kinki University.
This page intentionally left blank .
Shimamura. R. Ohba 213 215 . Hagiwara v 1 55 91 115 139 181 —Poster Summaries— Controled Teleportation of an Arbitrary Unknown TwoQubit Entangled State V. J. Nakahara Notes on the D¨ r–Cirac Classiﬁcation u Y. Ota. Ozdemir. M. Imoto Quantum ErrorCorrecting Codes M. Ebrahimi. Nakahara Braid Group and Topological Quantum Computing T. Yoshida. Rahimi. M. K. N.xv CONTENTS Preface Quantum Computing: An Overview M. J. Sakuma An Introduction to Entanglement Theory D. K. Ootsuka. Markham Holonomic Quantum Computing and Its Optimization S. H. I. Tanimura Playing Games in Quantum Mechanical Settings: Features of Quantum Games ¨ ¸ S.
Kitagawa On Classical NoCloning Theorem Under Liouville Dynamics and Distances T. Rahimi. M. Nakahara.xvi BangBang Control of Entanglement in SpinBusBoson Model R. O. M. M. R. Iguchi 217 219 221 . A. Rahimi. Yamano. SaiToh. SaiToh. Nakahara Numerical Computation of TimeDependent Multipartite Nonclassical Correlation A.
most quantum states cannot be described by . they can be described by specifying each object separately. mp3 players. personality and so on of each constituent person. Keywords: Quantum Physics. This means that an AND circuit. quantum error correcting codes and physical realizations.1 QUANTUM COMPUTING: AN OVERVIEW MIKIO NAKAHARA Department of Physics. In a quantum world. qubits. mobile phones. Quantum Algorithms. One of the most remarkable aspects of the principles of quantum physics is the superposition principle by which a quantum system can take several diﬀerent states simultaneously. Introduction Quantum computing and quantum information processing are emerging disciplines in which the principles of quantum physics are employed to store and process information. Subjects inclulde quantum physics. which is far beyond the classical description. Kinki University. The input for a quantum computing device may be a superposition of many possible inputs. this group can be described by specifying the height. quantum algorithms. is entanglement. Qubits. Even though quantum mechanics is used in the design of devices such as LSI. color of eyes. and accordingly the output is also a superposition of the corresponding output states.kindai. 1. for example. the logic is purely classical. for example. HigashiOsaka 5778502. Given several objects in a classical world. decoherece. In other words. Presentations of these subjects are as pedagogical as possible. however. Given a group of ﬁve people.jp Elements of quantum computing and quantum infromation processing are introduced for nonspecialists. Quantum Gates. Some sections are meant to be brief introductions to contributions by other lecturers.ac. We use the classical digital technology at almost every moment in our lives: computers. Another aspect of quantum physics. only a very small subset of all possible states can be described by such individual speciﬁcations. quantum gates. Japan Email: nakahara@math. produces deﬁnitely 1 when the inputs are 1 and 1. just to name a few.
The inner product of x and α is n αx = i=1 αi xi . σy = 0 −i i 0 . xn )t ↔ x = (x∗ . . thereby being called “entangled”. 1 n by which an inner product of two vectors are deﬁned as xy = i=1 x∗ yi . .2 such individual speciﬁcations. xi ∈ C .1. A vector in Cn is called a ket vector or a ket and is denoted as x1 . σz = 1 0 0 −1 in the basis in which σz is diagonalized. xn while a vector in the dual space Cn∗ is called a bra vector or a bra and denoted α = (α1 . . . . . . x = . . Symbols X = σx . A part of this lecture note is based on our forthcoming book. . . 2. Why and how these two features give rise to the enormous computational power in quantum computing and quantum information processing will be explained in this contribution. Notation and conventions We will exclusively work with a ﬁnitedimensional complex vector space Cn with an inner product . This inner product naturally introduces a correspondence x = (x1 . Y = −iσy and Z = σz are also employed. x∗ ). Quantum Physics 2. Pauli matrices are generators of su(2) and denoted n σx = 01 10 . αn ) αi ∈ C. . (Hilbert spaces). i The inner product naturally deﬁnes the norm of a vector x as x = xx . . . Index i sometimes runs from 0 to n − 1.1 General references are [2–4].
. It should be noted that a measurement produces one outcome λi and the probability of obtaining it is experimentally evaluated only after repeating measurements with many copies of the same state. (1) ∂t where is a physical constant known as the Planck constant and H is a Hermitian operator (matrix) corresponding to the energy of the system and is called the Hamiltonian. the outcome is one of the eigenvalues λj of A.3 Let A be an m × n matrix and B be a p × q matrix. . . 2). the interpretation of the wave function remains an open question. a1n B a B. Suppose we prepare many copies of the state c1 λ1 + c2 λ2 . a12 B. Here we adopt the most popular one. These statements are easily generalized to superposition states of more than two states. with i=1 ci 2 = 1. . Let λ1 and λ2 be two eigenvalues of A: Aλi = λi λi . the state undergoes an abrupt change (wave function collapse) to one of the eigenstates λi corresponding to the observed eigenvalue λi . . a22 B. 2. a2n B A ⊗ B = 21 . When a measurement of a is made.. The probability of collapsing to the state λi is given by ci 2 (i = 1. . If we measure a in this state.. A 1 A pure state in quantum mechanics is represented by a normalized vector ψ in a Hilbert space H associated with the system. . .2. Consider a superposition state c1 λ1 + c2 λ2 . . amn B is an (mp) × (nq) matrix called the tensor product of A and B. Then a11 B. am1 B. The complex coeﬃcient ci is called the probability amplitude in this sense. A 2 For any physical quantity (observable) a. . is also a possible state of the same system (superposition principle). there exists a corresponding Hermitian operator A acting on H. . If two states ψ1 and ψ2 are physical states of the system. i . their linear superpo2 sition c1 ψ1 + c2 ψ2 (ck ∈ C). am2 B. A 3 The time dependence of a state is governed by the Schr¨dinger equation o ∂ψ = Hψ . Axioms of quantum mechanics Quantum mechanics was discovered roughly a century ago. called the Copenhagen interpretation. .5–10 In spite of its long history.
• Axiom A 2 may be formulated in a diﬀerent but equivalent way as follows. where T is the timeordering operator. conversely. Let the spectral decomposition of the corresponding operator A be A= i λi λi λi . According to A 2. The operator U (t) : ψ(0) → ψ(t) .j c∗ ci λj Aλi = j i. Unitarity of U (t) guarantees that the norm of ψ(t) is conserved: ψ(0)U † (t)U (t)ψ(0) = ψ(0)ψ(0) = 1 (∀t > 0).4 Several comments are in order. we will obtain the same result since ψAψ = i. called the timeevolution operator.j λi c∗ ci δij = j i λi ci 2 . the phase of the vector may be chosen arbitrarily. Eq. the probability of observing λi upon measurement of a is ci 2 and therefore the expectation value after many measurements is i λi ci 2 . is unitary. The state at t > 0 is ψ(t) = U (t)ψ(0) . This measurement is called the projective measurement. ψ in fact represents the “ray” {eiα ψ α ∈ R}. The overall phase is not observable and has no physical meaning. If. (3) where Pi = λi λi  is the projection operator and the state immediately after the measurement is λi or equivalently Pi ψ / ψPi ψ . where Aλi = λi λi . (2) is employed. This is called the ray representation. while ψ(t) = T exp − i 0 t (5) H(t)dt ψ(0) (6) if H depends on t. Then the expectation value A of a after measurements with respect to many copies of ψ is A = ψAψ . (2) Let us expand ψ in terms of λi as ψ = i ci λi . • In Axiom A 1. Any particular outcome λi will be found with the probability ci 2 = ψPi ψ . . (4) • The Schr¨dinger equation (1) in Axiom A 3 is formally solved to yield o ψ(t) = e−iHt/ ψ(0) . Suppose we would like to measure an observable a. if the Hamiltonian H is timeindependent.
they do not commute. 2 ˆ where n is a unit vector in R3 . B]ψ . Let A = ψAψ and B = ψBψ be their respective expectation (A − A )2 and ∆B = (B − B )2 be respective values and ∆A = standard deviations. If. Let us consider a timeindependent Hamiltonian H = − ωσx (8) 2 acting on the spin Hilbert space C2 . ψ(0) =  ↑ . They form a basis of a vector space C2 . (12) . It is common to assign components  ↑ = (1. in contrast.5 Two mutually commuting operators A and B have simultaneous eigenstates. which we call spinup state  ↑ and spindown state  ↓ .3. Consider a more general Hamiltonian ˆ (10) H = − ω n · σ. A spin1/2 particle has two states. 2 (7) ω ω t ↑ + i sin t ↓ . Suppose the system is in the eigenstate of σz with the eigenvalue +1 at time t = 0. the measurement outcomes of these operators on any state ψ satisfy the following uncertainty relations. The wave function ψ(t) (t > 0) is then found from Eq. 1)t . The timeevolution operator is readily obtained. Simple example Examples to clarify the axioms introduced in the previous subsection are given. 0)t and  ↓ = (0. by making use of a well known formula ˆ ˆ (11) eiα(n·σ) = cos αI + i(n · σ) sin α as ˆ U (t) = exp(−iHt/ ) = cos ωt/2 I + i(n · σ) sin ωt/2. (5) as ω ψ(t) = exp i σx t ψ(0) = 2 = cos ωt/2 i sin ωt/2 = cos cos ωt/2 i sin ωt/2 i sin ωt/2 cos ωt/2 1 0 (9) 1  ψ[A. They are used to controll quantum states in physical realizations of a quantum computer. The spin is found spinup with probability P↑ (t) = cos2 (ωt/2) and spindown with probability P↓ (t) = sin2 (ωt/2). 2 2 Suppose we measure σz in ψ(t) . Then they satisfy ∆A∆B ≥ 2.
0)t for example. c1 d1 = c2 d2 = 1/ 2 simultaneously and it is clear that the above equations have no common solution. What are the missing states then? Let us consider a spin state 1 ψ = √ ( ↑ ⊗  ↑ +  ↓ ⊗  ↓ ) 2 of two electrons.j cij e1. However this decomposition is not possible since we must have c1 d2 = √ c2 d1 = 0. 2) are large. This number is considerably larger than the dimension of the sparable states when dim Ha (a = 1. Suppose ψ may be decomposed as ψ = (c1  ↑ + c2  ↓ ) ⊗ (d1  ↑ + d2  ↓ ) = c1 d1  ↑ ⊗  ↑ + c1 d2  ↑ ⊗  ↓ + c2 d1  ↓ ⊗  ↑ + c2 d2  ↓ ⊗  ↓ . Note. (15) . (14) where {ea. however. tensor product and entangled state So far. that the total space H has diﬀerent dimension than this: dim H = dim H1 dim H2 . The system as a whole lives in a Hilbert space H = H1 ⊗ H2 . A state ψ ∈ H written as a tensor product of two vectors as ψ = ψ1 ⊗ψ2 .i } (a = 1. Such nonseparable states are called entangled. whose general vector is written as ψ = i.j cij 2 = 1. at a later time t > 0. ψ(t) = U (t)ψ(0) = cos(ωt/2) + inz sin(ωt/2) i(nx + iny ) sin(ωt/2) . A system composed of two separate components is called bipartite. A separable state admits a classical interpretation “The ﬁrst system is in the state ψ1 while the second system is in ψ2 ”.i ⊗ e2. one lives in a Hilbert space H1 and the other in H2 . Then we ﬁnd. Entangled states refuse classical descriptions. 2) is an orthonormal basis in Ha and i.j .4. It is clear that the set of separable state has dimension dim H1 + dim H2 . Multipartite system. (13) 2. (ψa ∈ Ha ) is called a separable state or a tensor product state. showing ψ is not separable.6 Suppose the initial state is ψ(0) = (1. Suppose a system is made of two components. we have implictly assumed that the system is made of a single component. Entanglement is used extensively as a powerful computational resource in the following.
(16) where si > 0 are called the Schmidt coeﬃcients satisfying i si = 1 and {fa.5. Let H = H1 ⊗ H2 be the Hilbert space of a bipartite system. Mixed states and density matrices It might happen in some cases that a quantum system under considertation is in the state ψi with a probability pi . (18) Then Eq. The criterion is given by the Schmidt decomposition of ψ .1. A particular state ψi ∈ H appears with probability pi in an ensemble of a mixed state. See1 for the proof. It is easy to show all the eigenvalues of . Let us introduce the density matrix by N ρ= i=1 pi ψi ψi . (17) is rewritten in a compact form as A = Tr(ρA). A pure state is a special case of a mixed state in which pi = 1 for some i and pj = 0 (j = i). Let A be a Hermitian matrix. Theorem 2. Such a system is said to be in a mixed state while a system whose vector is uniquely speciﬁed is in a pure state. 2.i .i ⊗ f2. We are interested in when the state is separable and when entangled. we cannot say deﬁnitely which state the system is in. Therefore some random nature comes into the description of the system. It follows from the above theorem that a bipartite state ψ is separable if and only if its Schmidt number r is 1.7 Suppose a bipartite state (14) is given.i } is an orthonormal set of Ha . (17) where N is the number of available states. The number r ∈ N is called the Schmidt number of ψ . In other words. Then a vector ψ ∈ H admits the Schmidt decomposition r ψ = i=1 √ si f1. The mean value of a averaged over the ensemble is then given by N A = i=1 pi ψi Aψi . in which case the expectation value of the observable a is ψi Aψi . A is called positivesemideﬁnite if ψAψ ≥ 0 for any ψ ∈ H.
Let us consider a beam of photons. • The density matrix (18) is Hermitian since pi ∈ R. A density matrix is a positivesemideﬁnite Hermitian operator with tr ρ = 1. i Several remarks are in order. see remarks below. Its Hermitian conjugate is −i dt ψi  = ψi H. the density matrix is given by ρ= 1 1 1 e1 e1  + e2 e2  = 2 2 2 10 01 = 1 I. ρ]. Observe that A = tr ρA = i ei ψ ψAei = ψA i ei ei ψ = ψAψ . a Hermitian matrix A whose every eigenvalue is nonnegative is positivesemideﬁnite. Properties which a density matrix ρ satisﬁes are very much like axioms for pure states. (19) A 3’ The temporal evolution of the density matrix follows the Liouvillevon Neumann equation d ρ = [H. ρ] dt where H is the system Hamiltonian. A 1’ A physical state of a system.1. A pure state ψ is a special case in which the corresponding density matrix is ρ = ψ ψ. whose Hilbert space is H. We denote the set of all possible density matrices as S(H). It is positivesemideﬁnite since ψρψ = i pi  ψi ψ 2 ≥ 0. We prove the Liouvillevon Neumann equation from these equations as i d d ρ=i dt dt pi ψi ψi  = i i (20) pi Hψi ψi − i pi ψi ψi H = [H. Therefore ρ is nothing but the projection operator onto the state. If the photons are a totally uniform mixture of two polarized states. Example 2.8 a positivesemideﬁnite Hermitian matrix are nonnegative. We take a horizontally polarized state e1 =  ↔ and a vertically polarized state e2 =  as orthonormal basis vectors. see remarks below. A 2’ The mean value of an observable a is given by A = tr (ρA). d o • Each ψi follows the Schr¨dinger equation i dt ψi = Hψi in a closed d quantum system. is completely speciﬁed by its associated density matrix ρ : H → H. 2 . Conversely. where {ei } is an orthonormal set.
all its eigenvalues λi (1 ≤ i ≤ dim H) are real and the corresponding eigenvectors {λi } are made orthonormal. A state ρ is called separable if it is written in the form ρ= i pi ρ1.2. It is important to realize that only inseparable states have quantum correlations analogous to entangled pure states.11 In the next subsection. We use a bipartite system in the deﬁnition but generalization to multipartitle systems should be obvious. with {ei } as basis.9 This state is called a maximally mixed state. We are interested in when ρ represents a pure state or a mixed state. (21) where 0 ≤ pi ≤ 1 and i pi = 1. A state ρ is pure if and only if tr ρ2 = 1.1.kj . Proof: Since ρ is Hermitian.i . (22) . Then ρ2 = i. The converse is trivial. similarly to the classiﬁcation of pure states into separable states and entangled states. where λmax is the largest eigenvalue of ρ.j λi λj λi λi λj λj  = i λ2 λi λi . if ρ does not admit the decomposition (21). 2. Negativity Let ρ be a bipartite state and deﬁne the partial transpose ρpt of ρ with respect to the second Hilbert space as ρij.6.i ⊗ ρ2. √ If photons are in a pure state ψ = (e1 +e2 )/ 2. Deﬁnition 2. is ρ = ψ ψ = 1 2 11 11 . Therefore tr ρ2 = i λ2 ≤ i i λmax i λi = λmax ≤ 1. We classify mixed states into three classes. we discuss how to ﬁnd whether a given bipartite density matrix is separable or inseparable. Therefore tr ρ2 = 1 implies λmax = 1 and all the other eigenvalues are zero. Theorem 2. It is called inseparable. It does not necessarily imply that a separable state has no nonclassical correlation.kl → ρil. the density matrix. It is pointed out that useful nonclassical correlation exists in the subset of separable states.
The characteristic equation of ρpt is D(λ) = det(ρpt − λI) = λ− p+1 4 3 λ− 1 − 3p 4 = 0. Conversely.i (23) Note here that ρt for any density matrix ρ is again a density matrix since it is still positive semideﬁnite Hermitian with unit trace.k } is the orthonormal basis of the ﬁrst system while {e2. 2. Then the partial transpose yields ρpt = i pi ρ1. e1. This shows that ρpt is an unphysical state for 1/3 < p ≤ 1.2 e2. if the partial transpose of ρ of these systems is not a density matrix. e1. e1. It was conjectured by Peres12 and subsequently proven by the Hordecki family13 that positivity of the partially transposed density matrix is necessary and suﬃcient condition for ρ to be separable in the cases of C2 ⊗C2 systems and C2 ⊗ C3 systems.2 .1 .10 where ρij.l ). Instead of giving the proof.1 e2. Example 2. Therefore the partial transposed density matrix (23) is another density matrix. Here {e1. Let us consider the Werner state 1−p 0 0 0 4 0 1+p − p 0 4 2 ρ= 0 − p 1+p 0 . Suppose ρ takes a separable form (21). then ρ is inseparable.i ⊗ ρt .j ) ρ (e1. ρ is inseparable.2 e2.k } of the second system. If this is the case.kl = ( e1. 0 1−p 4 0 ρpt must have nonnegative eigenvalues to be a physically acceptable state.i ⊗ e2. Partial transpose of ρ yields 1−p 0 ρpt = 0 −p 2 4 (24) 0 1+p 4 0 0 1+p 4 0 0 −p 2 0 . Here the basis vectors are arranged in the order e1. we look at the following example.2. There are threefold degenerate eigenvalue λ = (1 + p)/4 and nondegenerate eigenvalue λ = (1 − 3p)/4. .1 . 2 4 0 0 0 1−p 4 where 0 ≤ p ≤ 1.1 e2.k ⊗ e2.2 .
ρ= 2 0 0 0 0 1001 where the basis vectors are ordered as {e1 e1 . The partial trace of ρ is ρ1 = tr 2 ρ = (I ⊗ ei )ρ(I ⊗ ei ) = i=1. Then the partial trace allows us to “forget” about the second system. . The corresponding density matrix is 1001 1 0 0 0 0 . However there is a class of inseparable states which are not characterized by negativity. Let ρ = ψ ψ ∈ S(H) be a density matrix of a pure state ψ . consider a pure state 1 ψ = √ (e1 e1 + e2 e2 ). (27) Note that a pure state ψ is mapped to a maximally mixed state ρ1 . 2 where {ei } is an orthonormal basis of C2 . e2 e2 }. entangled states are characterized by nonvanishing negativity deﬁned as N (ρ) ≡ 1 ( 2 λi  − 1). e1 e2 . e2 e1 . i (25) Note that negativity vanishes if and only if all the eigenvalues of ρpt are nonnegative.11 From the above observation. Partial trace and puriﬁcation Let H = H1 ⊗ H2 be a Hilbert space of a bipartite system made of components 1 and 2 and let A be an arbitrary operator acting on H. In other words.2 1 2 10 01 . 2. the partial trace quantiﬁes our ignarance on the second sytem.7. To be concrete. (26) We will be concerned with the partial trace of a density matrix in practical applications. The partial trace of A over H2 generates an operator acting on H1 deﬁned as A1 = tr 2 A ≡ k (I ⊗ k)A(I ⊗ k ). Suppose we are interested only in the ﬁrst system and have no access to the second system.
k = k pk ψk ψk  = ρ1 . 0 =  ↑ and 1 =  ↓ . 3.j.1. 1)t . How about the converse? Given a mixed state density matrix. and it constitutes the building block of the classical information theory. on the other hand. We ﬁnd √ (I ⊗ φi ) pj pk ψj φj ψk  φk  (I ⊗ φi ) tr 2 Ψ Ψ = i. 3. 0)t and 1 = (0. (28) Ψ = k where {φk } is an orthonormal basis of H2 . Now let us introduce the second Hilbert space H2 whose dimension is the same as that of H1 . Puriﬁcation is far from unique. is it always possible to ﬁnd a pure state density matrix whose partial trace over the extra Hilbert space yields the given density matrix? The answer is yes and the process to ﬁnd the pure state is called the puriﬁcation. is based on qubits. (30) What these vectors physically mean depends on the physical realization employed for quantum information processing. . Qubits A (Boolean) bit assumes two distinct values. (29) It is always possible to purify a mixed state by tensoring an extra Hilbert space of the same dimension as that of the original Hilbert space. Then formally introduce a normalized vector √ pk ψk ⊗ φk . Quantum information theory. Electrons are replaced by nuclei with spin 1/2 in NMR (Nuclear Magnetic Resonance).12 We have seen above that the partial trace of a purestate density matrix of a bipartite system over one of the constituent Hilbert spaces yields a mixed state. One qubit A qubit is a (unit) vector in the vector space C2 . Let ρ1 = k pk ψk ψk  be a general density matrix of a system 1 with the Hilbert space H1 . whose basis vectors are denoted as 0 = (1. 0 and 1. They might represent spin states of an electron.
sin θ sin φ. Bloch sphere It is useful. We may assign 0 to the ground state and 1 to the ﬁrst excited state of an atom or an ion. cos θ)t . (31) If we measure ψ to see whether it is in 0 or 1 . the outcome will be 0 (1) with the probability a2 (b2 ) and the state immediately after the measurement is 0 (1 ). Alternatively they might correspond to photons polarized in diﬀerent directions. with a2 + b2 = 1. it should be kept in mind that we can extract from it as the same amount of information as that of a classical bit. 0 1 = √2 ( + ↔ ) while 1 represents may represent a polarization state  1 = √2 ( −  ↔ ). Although a qubit may take inﬁnitely many diﬀerent states. which we call abstractly 0 and 1. to express a state of a single qubit graphically. the basis is written in an abstract form as {0 . where σ = (σx . the system immediately after the measurement is in 0 . In the following. φ) · σ)ψ(θ. Moreover a qubit may be in a superposition state: ↔ ↔ ψ = a0 + b1 . 1 }.13 In some cases. 3. a state  Truncated two states from many levels may be employed as a qubit. In any case. the state vector ‘collapses’ to the eigenvector that corresponds to the eigenvalue observed. Let us parameterize a onequbit pure state ψ with θ and φ as θ θ ψ(θ. This happens with probability ψ0 0ψ = a2 . φ) is a real unit vector called the Bloch vector with components ˆ n(θ. σy . φ) . for many purposes. It is convenient to assume the vector 0 corresponds to the classical bit 0. we have to ﬁx a set of basis vectors when we carry out quantum information processing. When we measure a qubit.2. Suppose a spin is in the state a0 + b1 . If we observe that the zcomponent of the spin is +1/2. φ) = (sin θ cos φ. φ) is expressed as a unit . For example. The measurement outcome of a qubit is always one of the eigenvalues. 0 stands for a vertically polarized photon  while 1 represents a horizontally polarized photon  ↔ . σz ) ˆ and n(θ. φ) so that ψ(θ. Information can be extracted only through measurements. It is therefore natural to assign ˆ n(θ. unless otherwise stated. 2 2 (32) The phase of ψ is ﬁxed in such a way that the coeﬃcient of 0 is real. It is ˆ easy to verify that (n(θ. while 1 to 1. φ) = ψ(θ. φ) = cos 0 + eiφ sin 1 . φ) to a state vector ψ(θ.
. 3.1 (35) ai1 i2 . As an example. n . Each of the qubit is then described by a 2d complex vector of the form ai 0 +bi 1 and we need 2n complex numbers {ai . An nqubit system is often called a (quantum) register in the context of quantum computing. The ball is called the Bloch ball and the vector u is also called the Bloch vector. its most general form is 1 ui σi . (34) ρ = I + 2 i=x. (33) A density matrix ρ of a qubit can be represented as a point on a unit ball. (34) satisﬁes σ = tr (ρσ) = u. The eigenvalue λ− vanishes in 2 1± case u = 1. The converse is also shown easily. . The reality follows from the Hermiticity requirement and tr ρ = 1 is obvious. for which rank ρ = 1. ⊗ (an 0 + bn 1 ) ∈ C2n . In contrast. all the points u inside a unit ball correspond to mixed states. Such a system behaves quite diﬀerently from a classical one and this diﬀerence gives a distinguishing aspect to quantum information theory. Multiqubit systems and entangled states Let us consider a group of many (n) qubits next.3. .. Therefore the surface of the unit sphere corresponds to pure states. It is easily veriﬁed that ρ given by Eq. . This correspondence is onetoone if the ranges of θ and φ are restricted to 0 ≤ θ ≤ π and 0 ≤ φ < 2π. Suppose we specify the state of each qubit separately like a classical case. φ) = n(θ.. bi }1≤i≤n to specify the state. φ)σψ(θ.14 ˆ vector n(θ.y. called the Bloch sphere. If the system is treated in a fully quantummechanical way.in i1 ⊗ i2 ⊗ . Since ρ is a positive semideﬁnite Hermitian matrix with unit trace. φ).z where u ∈ R3 satisﬁes u ≤ 1. This corresponds the a tensor product state (a1 0 + b1 1 ) ⊗ . ⊗ in ∈ C2 . however. The eigenvalues of ρ are λ± = 1 u /2 and therefore nonnegative. φ) on the surface of the unit sphere. let us consider an nqubit register. It is veriﬁed that state (32) satisﬁes ˆ ψ(θ. a general state vector of the register is represented as ψ = ik =0.
. Φ− = √ (00 − 11 ). a quantum algorithm. . Ψ− = √ (01 − 10 )} 2 2 is an orthonormal basis of a twoqubit system and is called the Bell basis. 4. U. Each vector is called the Bell state or the Bell vector. Deﬁnition 4.1. More complicated quantum circuits are composed of these simple gates. 3 } with this decimal notation. . 1}. The system has a binary basis {00 . + b0 . in analogy with those in classical logic circuits.15 Note that 2n 2n for a large number n. The ratio 2n /2n is ∼ 10298 for n = 1000. . . . bn−2 . . More generally. these operations are called gates. A collection of quantum circuits for executing a complicated algorithm. b0 . The basis for a 2qubit system may be written also as {0 .1. Most quantum states in a Hilbert space with large n are entangled having no classical analogues. Entanglement is an extremely powerful resource for quantum computation and quantum communication. which is taylored to execute a given quantum algorithm and (3) Measurements to extract information we need. where n H = C2 is the Hilbert space of an nqubit register. instead of bn−1 bn−2 . We write x . b0 ∈ {0. Quantum Circuit and Quantum Computation 4. b0 }. Quantum Gates. (2) A unitary matrix u. 2 . 11 }. a basis for a system of n qubits may be {bn−1 bn−2 . If they are simple. . Note that all the Bell states are entangled. {Mm}}. where bn−1 . Let us consider a 2qubit system for deﬁniteness. is a part of a quantum computation. . 2 2 (36) 1 1 Ψ+ = √ (01 + 10 ). More formally. it is time to consider operations acting on them. Introduction Now that we have introduced qubits to store information. . It is also possible to express the basis in terms of the decimal system. 10 . or quantum gates. 01 . The set 1 1 {Φ+ = √ (00 + 11 ). 1 . where x = bn−1 2n−1 + bn−2 2n−2 + . a quantum computation is the set {H. U ∈ U (2n ) represents . (Quantum Computation) A quantum computation is a collection of the following three elements: (1) A register or a set of registers.
(37) Similarly we introduce X : 0 → 1 . Linearity guarantees that the action of a gate is completely speciﬁed if its action on the basis {0 . 1 → −1 by X = 1 0 + 0 1 = Y = 0 1 − 1 0 = Z = 0 0 − 1 1 = 01 10 0 −1 1 0 1 0 0 −1 = σx . 1 → 0 . Simple quantum gates Examples of quantum gates which transform a onequbit state are given below. = σz . When measurements are not made. We call them onequbit gates in the following. Quantum gates We have so far studied the change of a state upon measurements. = −iσy . The matrix expression of this gate is I = 0 0 + 1 1 = 10 01 . 1 } is given. the time evolution of a state is described by the Schr¨dinger equation. Suppose the register is set to a ﬁducial initial state. while X is the negation (NOT). We will be free from the Schr¨dinger equation in the following and assume there always exist unitary matrices which we need. Information is extracted from ψout by appropriate measurements. 4. One of the important conclusions derived from the unitarity of gates is that the computational process is reversible. . 1 → 0 and Z : 0 → 0 . Operation of Ualg on ψin yields the output state ψout = Ualg ψin . The hardware (1) is called a quantum computer. 0 for example.16 a quantum algorithm and {Mm } is the set of measurement operators. . (38) (39) (40) The transformation I is the identity transformation. The time evolution operator U is unitary: o o U U † = U † U = I. Y : 0 → −1 . Z the phase shift and Y = XZ the combination thereof.2.2.1. . 1 → 1 . A unitary matrix Ualg is generated by an algorithm which we want to execute. ψin = 00 . Consider the gate I whose action on the basis vectors is I : 0 → 0 . 4.
0. 0) . There may be many control bits (see CCNOT gate below). Let {00 . 0. 0) . in which the target bit is acted on by a unitary transformation U only when the control bit is 1 . 11 = (0. 1. More generally. 11 → 10 . 0. is UCNOT : 00 → 00 . while denotes the conditional negation. where i ⊕ j is an addition mod 2. while leaving the second bit unchanged when the ﬁrst bit is 0 . j is written as i. 10 → 11 . 10 = (0. 0) . 11 } be a basis for the 2qubit system. 0. A 1qubit gate whose unitary matrix is U is graphically depicted as The left horizontal line is the input qubit while the right horizontal line is the output qubit: time ﬂows from the left to the right. 10 . 01 → 01 . The action of CNOT gate. 0. which plays an important role.17 CNOT (controlledNOT) gate is a 2qubit gate. 1. 0 0 1 010 (41) t t t t UCNOT (42) Let {i } be the basis vectors. 0. This gate is denoted graphically as . whose matrix expression will be written as UCNOT . i ⊕ j . The gate ﬂips the second qubit (the target qubit) when the ﬁrst qubit (the control qubit) is 1 . It has two equivalent expressions UCNOT = 00 00 + 01 01 + 11 10 + 10 11 = 0 0 ⊗ I + 1 1 ⊗ X. where i ∈ {0. The action of CNOT on the input state i. V = 0 0 ⊗ I + 1 1 ⊗ U . 1) . 1}. A CNOT gate is expressed as where • denotes the control bit. We use the standard basis vectors with components 00 = (1. we consider a controlledU gate. 01 = (0. having a matrix form 1 0 = 0 0 000 1 0 0 . 01 .
(45) (44) There are numerous important applications of the Hadamard transformation.18 CCNOT (ControlledControlledNOT) gate has three inputs and the third qubit ﬂips only when the ﬁrst two qubits are both in the state 1 . WalshHadamard transformation The Hadamard gate or the Hadamard transformation H is an important unitary transformation deﬁned by 1 UH : 0 → √ (0 + 1 ) 2 1 √ (0 − 1 ). : 1 → 2 The matrix representation of H is 1 1 1 UH = √ (0 + 1 ) 0 + √ (0 − 1 ) 1 = √ 2 2 2 A Hadamard gate is depicted as 1 1 1 −1 . All possible 2n states are generated when UH is applied on each . The explicit form of the CCNOT gate is UCCNOT = (00 00 + 01 01 + 10 10) ⊗ I + 11 11 ⊗ X.2. This gate is graphically expressed as (43) 4.2.
The explict form of USWAP is given by 10 0 0 = 00 00 + 01 10 + 10 01 + 11 11 = 0 1 00 00 1 0 . (48) 0 0 01 (47) USWAP The SWAP gate is expressed as Note that the SWAP gate is a special gate which maps an arbitrary tensor product state to a tensor product state. . The quantum circuit is used to generate Bell states from inputs 00 . The transformation UH is called the Walsh transformation. SWAP gate and Fredkin gate The SWAP gate acts on a tensor product state as USWAPψ1 .3.19 qubit of the state 00 . or WalshHadamard transformation and denoted as Wn . x=0 (46) Therefore. The controlledSWAP gate . . ψ2 = ψ2 . . . ⊗ UH )00 . . 4. we produce a superposition of all the states x with 0 ≤ x ≤ ⊗n 2n − 1 simultaneously. In contrast. . most 2qubit gates map a tensor product state to an entangled state. 0 1 1 1 = √ (0 + 1 ) ⊗ √ (0 + 1 ) ⊗ . 0 : (UH ⊗ UH ⊗ . √ (0 + 1 ) 2 2 2 1 = √ 2n 2n −1 x . . . 10 and 11 . ψ1 .2. 01 .
we must also have 1 U ψ0 = ψψ = (ϕϕ + ϕφ + φϕ + φφ ). for any state ϕ . Proof: Suppose there would exist a unitary transformation U that makes a clone of a quantum system.20 is also called the Fredkin gate. Its explicit form is UFredkin = 0 0 ⊗ I4 + 1 1 ⊗ USWAP . For these cases. . there does not exist a unitary cloning transformation. 2 which contradicts the previous result. Let ϕ and φ be two states that are linearly independent. An example of such U is U = (00 00 + 11 10) + (01 01 + 10 11). It ﬂips the second (middle) and the third (bottom) qubits only when the ﬁrst (top) qubit is in the state 1 . What we have to keep in mind is only that U be unitary.1. Then we should have U ϕ0 = ϕϕ and U φ0 = φφ by deﬁnition. as U : ϕ0 → ϕϕ . (Wootters and Zurek 14 ) An unknown quantum system cannot be cloned by unitary transformations. Namely. We immediately notice that U is nothing but the CNOT gate. (49) where the ﬁrst set of operators renders U the cloning operator and the second set is added just to make U unitary. Therefore. Then 1 the action of U on ψ = √ (ϕ + φ ) yields 2 1 1 U ψ0 = √ (U ϕ0 + U φ0 ) = √ (ϕϕ + φφ ). the copying operator U should work as U : 00 → 00 . Nocloning theorem Theorem 4. suppose U acts. We can assign arbitrary action of U on a state whose second input is 1 since this case will never happen. 4. 2 2 If U were a cloning transformation.3. Note however that the theorem does not apply if the states to be cloned are limited to 0 and 1 . : 10 → 11 .
4. She wishes to send Bob the quantum state of this qubit through a classical communication channel. Quantum teleportation The purpose of quantum teleportation is to transmit an unknown quantum state of a qubit using two classical bits in such a way that the recipient reproduces the same state as the original qubit state.4. 1. if the data under consideration is limited within 0 and 1 . Alice sends Bob two classical bits so that Bob reproduces a qubit state Alice initially had. whose state she does not know. In quantum teleportation. which results in (UH ⊗ I ⊗ I)(UCNOT ⊗ I)(φ ⊗ Φ+ ) 1 = [00 (a0 + b1 ) + 01 (a1 + b0 ) 2 +10 (a0 − b1 ) + 11 (a1 − b0 )].21 Therefore. 2 where Alice possesses the ﬁrst two qubits while Bob has the third. we can copy the qubit states even in a quantum computer. Fig. (51) . The original state is destroyed such that quantum teleportation is not in contradiction with the nocloning theorem. Both of them have been given one of the qubits of the entangled pair 1 Φ+ = √ (00 + 11 ) 2 in advance. They start with the state 1 (50) φ ⊗ Φ+ = √ (a000 + a011 + b100 + b111 ) . This fact is used to construct quantum error correcting codes. Alice: Alice has a qubit. Let φ = a0 + b1 be the state of the qubit. Alice applies UCNOT ⊗ I followed by UH ⊗ I ⊗ I to this state. Note that the qubit itself is not transported but the information required to reproduce the quantum state is transmitted.
much larger than . Such a set of gates is called the universal set of gates. Bob knows the state of the qubit in his hand. however. AND. Suppose Alice sends Bob the classical bits 10. for example. It can be shown that the CCNOT gate simulates these classical gates. Figure 2 shows the actual quantum circuit for quantum teleportation. 4. received bits Bob’s state decoding 00 a0 + b1 I 01 a1 + b0 X 10 a0 − b1 Z 11 a1 − b0 Y (52) Bob reconstructs the intial state φ by applying the decoding process shown above. 10 or 11 with equal probability 1/4.5. a1 + b0 . depending on the result of Alice’s measurement. 2. respectively. 01 . Then Bob applies Z on his qubit to reconstruct φ as Z : (a0 −b1 ) → (a0 +b1 ) = φ . The set of quantum gates is. and hence quantum circuits simulate any classical circuits. Alice then sends Bob her result of the measurement using two classical bits.22 If Alice measures the 2 qubits in her hand. Quantum circuit implementation of quantum teleportation. she will obtain one of the states 00 . NOT and XOR for example. Bob’s qubit (one of the EPR pair previously) collapses to a0 + b1 . a0 − b1 or a1 − b0 . Universal quantum gates It can be shown that any classical logic gate can be constructed by using a small set of gates. Bob: After receiving two classical bits. Fig.
An example of a twolevel unitary matrix is ∗ α 0 0 β∗ 0 10 0 2 2 V = 0 0 1 0 . Vab . (54) .1. It can be shown that (1) the set of single qubit gates and (2) CNOT gate form a universal set of quantum circuits (universality theorem). Vab . Now we need to prove the universality theorem for twolevel unitary matrices. It can be shown that any U ∈ U(n) is written as a product of N twolevel unitary matrices. We need to implement the controlledV gate whose target qubit is the one on which V acts. which is certainly simpler than the general proof. The proof is highly technical and is not given here.2. it acts on all the constituent vectors of the superposition simultaneously. Implementation of the controlledV gate requires generalized CNOT gates and several U(2) gates. Vba and Vbb so that they acts on a single qubit in the register.6. Suppose Uf acts on an input which is a superposition of many x . instead. where N ≤ n(n − 1)/2 and a twolevel unitary matrix is a unit matrix In in which only four components Vaa . Quantum parallelism and entanglement Given an input x. Thus we want to ﬁnd a universal set of quantum gates from which any quantum circuits can be constructed. Uf : x x 0 → x x f (x) .2. a typical quantum computer “computes” f (x) as Uf : x 0 → x f (x) . Since Uf is a linear operator. Moreover V = (Vij ) is an element of U(2). By employing CNOT gates and their generalizations. (α + β = 1) −β 0 0 α where a = 1 and b = 4.16 We.1.23 those classical gates. sketch the proof in several lines. The output is also a superposition of all the results. Vba and Vbb are diﬀerent from In . it is possible to move the elements Vaa .16 4. (53) where Uf is a unitary matrix which implements the function f .
This state is a superposition of vectors encoding all the integers between 0 and 2n − 1. This superposition is prepared by the action of the WalshHadamard transformation on an nqubit register in the initial √ 2n −1 state 00 . 2 By operating CCNOT on this state. 2 (57) This output may be thought of as the truth table of AND: x. Suppose the input state of the ﬁrst and the second qubits is a superposition of all possible states while the third qubit is ﬁxed to 0 . A quantum computer is advantageous over a classical counterpart in that it makes use of this quantum parallelism and also entanglement. . gives quantum computer an enormous power.e. The third output qubit state is x ∧ y . . 0 = 0 ⊗ 0 ⊗ . This is because only one result may be obtained by a single set of . Let us ﬁx the third input qubit to 0 . Then the linearlity of Uf leads to Uf 1 √ 2n 2n −1 x 0 x=0 1 =√ 2n 2n −1 x=0 1 Uf x 0 = √ 2n 2n −1 x f (x) . . where x and y are the ﬁrst and the second input qubits respectively. . This gate ﬂips the third qubit if and only if the ﬁrst and the second qubits are both in the state 1 while it leaves the third qubit unchanged otherwise. What about the limitation of a quantum computer3 ? Let us consider the CCNOT gate for example. x ∧ y . It is extremely important to note that the output is an entangled state and the measurement projects the state to one line of the truth table. a single term in the RHS of Eq. which makes quantum computation exponentially faster than classical counterpart in a certain kind of computation.. (55) x=0 Note that the superposition is made of 2n = en ln 2 states. (57). y. called the quantum parallelism. i. This can be achieved by the WalshHadamard transformation as 1 1 UH 0 ⊗ UH 0 ⊗ 0 = √ (0 + 1 ) ⊗ √ (0 + 1 ) ⊗ 0 2 2 1 (56) = (000 + 010 + 100 + 110 ).24 This feature. A unitary transformation acts on a superposition of all possible states in most quantum algorithms. There is no advantage of quantum computation over classical one at this stage. ⊗ 0 resulting in x=0 x / 2n . we obtain UCCNOT (UH 0 ⊗ UH 0 ⊗ 0 ) = 1 (000 + 010 + 100 + 111 ).
This idea was employed in the quantum Fourier transform to ﬁnd the ordera of f in the Shor’s factoring algorithm. N ∈ N (m < N ) be numbers coprime to each other. Since each qubit has two basis states {0 . Now we consider the power of entanglement. i. ∀x ∈ N.25 measurements. instead.1. consider how many diﬀerent basis vectors are required for this system. . there are 2n basis states. of the vector that we want to observe. a Let m. Imagine that we have a single quantum system. we cannot choose a speciﬁc vector x. n 0 ’s and n 1 ’s. This single system must have an enormous number. and hence the probability. x∧y at our will! Thus any quantum algorithm should be programmed so that the particular vector we want to observe should have larger probability to be measured compared to other vectors. 5. involved to span this Hilbert space. Deutsch algorithm The Deutsch algorithm is one of the ﬁrst quantum algorithms which showed quantum algorithms may be more eﬃcient than their classical counterparts. 5. The smallest such number P is called the period or the order..e. (2) to ﬁnd a common property of all the f (x). This strategy is employed in the Grover’s database search algorithm. which has the same Hilbert space. It is easily seen that mx+P ≡ mx (mod N ). One might think that the system may do the same quantum computation as the nqubit register does. The programming strategies to deal with this feature are (1) to amplify the amplitude. Even worse. What is worse. whose Hilbert space is 2n dimensional. 1 }. One possible problem is that one cannot “measure the kth digit” leaving other digits unaﬀected. 2n . Then there exists P ∈ N such that mP ≡ 1 (mod N ). Suppose we have an nqubit register. of basis vectors! Multipartite implementation of a quantum algorithm requires exponentially smaller number of basis vectors than monopartite implementation since the former makes use of entanglement as a computational resource. Simple Quantum Algorithms Let us introduce a few simple quantum algorithms which will be of help to understand how quantum algorithms are diﬀerent from and superior to classical algorithms. y.
f2 : 0 → 1. respectively. f3 and f4 . Note that there are only four possible f . the controlledNOT gate with the control bit f (x). as was shown by Deutsch. There is a quantum algorithm. Let f : {0. Let us consider a quantum circuit which implements the Deutsch algorithm. where ⊕ is an addition mod 2. we obtain 1 (0.. Subsequently we apply the Hadamard gate on the ﬁrst qubit to obtain ψ1 = Uf ψ0 = ψ2 = UH ψ1 1 = √ [(0 + 1 )(f (0) − ¬f (0) ) + (0 − 1 )(f (1) − ¬f (1) )] 2 2 The wave function reduces to 1 ψ2 = √ 0 (f (0) − ¬f (0) ) 2 in case f is constant. 2 where ¬ stands for negation. 1 → 0. are called constant. y ⊕ f (x) .18 Let 0 and 1 correspond to classical bits 0 and 1. i.while the rest. Therefore the measurement of the ﬁrst qubit tells us whether f is constant or balanced. 1 → 1. To be explicit.If we only have classical resources. for which ¬f (0) = f (1) . and 1 ψ2 = √ 1 (f (0) − f (1) ) 2 (59) (58) if f is balanced. f1 and f2 . namely f1 : 0 → 0. We apply f on this 2 state in terms of the unitary operator Uf : x.e. we need to evaluate f twice to tell if f is constant or balanced. 1 → 0.26 In spite of its simplicty. f (1) − 1. We ﬁrst apply the WalshHadamard transformation W2 = UH ⊗ UH on 01 to obtain ψ0 . full usage of superposition principle and entanglement has been made here. for which f (0) = f (1) . whose action is . 1 → 1. with which it is possible to tell if f is constant or balanced with a single evaluation of f . and consider the state ψ0 = 1 (00 − 01 + 10 − 11 ). f4 : 0 → 1. y → x. ¬f (1) ). First two cases. ¬f (0) + 1. Therefore this operation is nothing but the CNOT gate with the control bit f (x). 1} → {0. f (0) − 0. 1} be a binary function. f3 : 0 → 0. the target bit y is ﬂipped if and only if f (x) = 1 and left unchanged otherwise. are balanced. We need to introduce a conditional gate Uf . in contrast.
. When it is balanaced the value f (x) for a half of x ∈ Sn is 0 while it is 1 for the rest of x. 1. 3 with the given matrix Uf and apply the circuit on the input state 01 .2. Such a black box is often called an oracle. It is clear that we need at least 2n−1 + 1 steps. we will not consider such cases here. (60) We require f be either constantor balancedas before. It will be shown below that the number of steps reduces to a single step if we are allowed to use a quantum algorithm. Its implementation is given only after f is speciﬁed. In the quauntum circuit. Fig. We might think it is a kind of subroutine.27 Uf : x. y ⊕ f (x) . Figure 3 depicts this implementation. it takes a constant value 0 or 1 irrespetive of the input value x. . When f is constant. y → x. Implementation of the Deutsch algorithm. to make sure if f (x) is constant or balanced with 100 % conﬁdence. 1}. we assume the gate Uf is a black box for which we do not ask the explicit implementation.19 Let us ﬁrst deﬁne the DeutschJozsa problem. Then the Hadamard gate is applied on the ﬁrst qubit before it is measured. you may contruct the circuit in Fig. Then what is the merit of the Deutsch algorithm? Suppose your friend gives you a unitary matrix Uf and asks you to tell if f is constant or balanced. 2n − 1} → {0. . 5. Although there are functions which are neither constant nor balanced. The gate Uf is called the Deutsch oracle. Our task is to ﬁnd an algorithm which tells if f is constant or balanced with the least possible number of evaluations of f . 3. Suppose there is a binary function f : Sn ≡ {0. in the worst case with classical manipulations. . Instead of applying 0 and 1 separately. DeutschJozsa algorithm The Deutsch algorithm introduced in the previous section may be generalized to the DeutschJozsa algorithm. Then you can tell your friend whether f is constant or balanced with a single use of Uf . .
(2) Apply the WalshHadamard transforamtion to the register. it is applied to all the nqubit states x simultaneously. Therefore we need a Uf gate which evaluates f (x) and acts on the register as Uf x c = x c ⊕ f (x) . 1}. Then we have the state 1 1 ⊗n+1 ψ0 = √ (0 + 1 )⊗n ⊗ √ (0 − 1 ) ψ1 = UH n 2 2 1 = √ 2n 2n −1 x=0 1 x ⊗ √ (0 − 1 ). (63) 2 x=0 It is instructive to write the action of the onequbit Hadamard gate as 1 1 UH x = √ (0 + (−1)x 1 ) = √ (−1)xy y . The action of the Walsh .1} where x ∈ {0. First n qubits work as input qubits while the (n + 1)st qubit serves as a “scratch pad”. We obtain ψ3 = (Wn ⊗ I)ψ2 1 =√ 2n 2n −1 1 ⊗n (−1)f (x) UH x √ (0 − 1 ). 2 (61) (3) Apply the f (x)controlledNOT gate on the register. Observe that c is ﬂipped if and only if f (x) = 1 and left unchanged otherwise. to ﬁnd the resulting state. which are neither input qubits nor output qubits. which ﬂips the (n+1)st qubit if and only if f (x) = 1 for the input x. 2 Although the gate Uf is applied once for all. We then obtain a state ψ2 = Uf ψ1 1 = √ 2n 1 = √ 2n (−1) x 2n −1 x=0 1 x √ (f (x) − ¬f (x) ) 2 (62) f (x) 1 x √ (0 − 1 ). but work as a scratch pad to store temporary information are called ancillas or ancillary qubits.28 The algorithm is divided into the following steps: (1) Prepare an (n + 1)qubit register in the state ψ0 = 0 ⊗n ⊗ 1 . Such qubits. where c is the onequbit state of the (n + 1)st qubit. (4) The WalshHadamard transformation (46) is applied on the ﬁrst n qubits next. 2 2 y∈{0.
Now the state reduces to ψ3 = 0 ⊗n √2 (0 − 1 ) and the measurement outcome of the ﬁrst n qubits is always 00 . Therefore the probability of obtaining measurement outcome 00 . . we obtain ψ3 = 1 2n 2n −1 (−1)f (x) (−1)x·y y x. . The system under consideration is not a closed system when interaction with outside world is in action.. Let us consider the summation 21 n x=0 (−1) for a ﬁxed y ∈ Sn .y . . The probability amplitude of y = 0 in ψ3 is pro2n −1 2n −1 portional to x=0 (−1)f (x) (−1)x·y = x=0 (−1)f (x) = 0. . (UH x0 ) 1 = √ (−1)xn−1 yn−1 +xn−2 yn−2 +. 2 (65) (5) The ﬁrst n qubits are measured.+x0 y0 2n y . the function f is constant if we obtain 00 . which causes loss of information encoded in this system..1} n−1 n−2 0 ×yn−1 yn−2 . 6. (63). . . Decoherence A quantum system is always in interaction with its environment.29 Hadamard transformation on x = xn−1 . .. .⊕ x0 y0 . . . Suppose f (x) is constant. . . Suppose f (x) is balanced next.y ∈{0. We formulate the theory of open quantum system in this chapter by regarding the combined system of the quantum system and its environment as a closed system and subsequently trace out the environment degrees of . 0 upon the meaurement of the ﬁrst n qubits in the state ψ3 and it is balanced otherwise.. .. y=0 (64) where x·y = xn−1 yn−1 ⊕ xn−2 yn−2 ⊕ . Therefore the summation yields 1 δy0 .. Clearly it vanishes since x · y is 0 for half of x and 1 for the other half of x unless y = 0. Then ψ3 is put in the form ψ3 = 1 2n 1 (−1)x·y y √ (0 − 1 ) 2 x.y 2n −1 x·y up to an overall phase. y0 1 =√ 2n 2n −1 (−1)x·y y . x1 x0 yields Wn x = (UH xn−1 )(UH xn−2 ) . 0. In conclusion. 0 for the ﬁrst n qubits vanishes. . Substituting this result into Eq. This interaction inevitably alter the state of the quantum system.y=0 1 √ (0 − 1 ).
Decoherence is a process in which environment causes various changes in the quantum system. Then the corresponding quantum map E is deﬁned as E(ρS ) = U (t)ρS U (t)† . the environment Hamiltonian and their interaction Hamiltonian.2. Quantum operations and Kraus operators Suppose a system of interest is coupled with its environment. HE and HSE be the system Hamiltonian. We assume the systemenvironment interaction is weak enough so that this separation into the system and its environment makes sense. Even when the initial state is an uncorrelated state ρS ⊗ ρE . We must specify the details of the environment and the coupling between the system and the environment to study the eﬀect of the environment on the behavior of the system. respectively. the systemenvironment interaction entangles the total system so that the total state develops to an inseparable entangled state in general. One of the simplest quantum operations is a unitary time evolution of a closed system.1. To avoid confusion. Let ρS and ρE be the initial density matrices of the system and the environment.1. for example. The condition of weak systemenvironment interaction may be lifted in some cases. Let us consider a qubit propagating through a noisy quantum channel. which is described by a powerful tool called a quantum operation. respectively. which manifests itself as undesirable noise. respectively. we denote the system Hilbert space and the environment Hilbert space as HS and HE . (67) Correspondingly. We are interested in a general evolution of a quantum system. Open quantum system Let us start our exposition with some mathematical background materials. (66) One of our primary aims in this section is to generalize this map to cases of open quantum systems.1. Let HS . The total Hamiltonian HT is then HT = HS + HE + HSE .24 We deal with general quantum states described by density matrices. and the total Hilbert space as HT = HS ⊗ HE . “Propagating” does not necessarily . 6.1. we often call the system of interest the principal system.30 freedom. 6. Let ρS be a density matrix of a closed system at t = 0 and let U (t) be the time evolution operator.
ρ(0) = ρS ⊗ ρE . however. the input and the output qubit states belong to a well deﬁned Hilbert space HS and the above separation of the Hamiltonian is perfectly acceptable even for strongly interacting cases. in the following. We have already noted that the unitary time evolution is an example of a quantum operation. (70) We may forget about the environment by taking a trace over HE . We may take the basis of HT to be {ej ⊗ εa }. We are interested in extracting information on the state of the principal system at some later time t > 0. We consider. When the noise is localized in space and time. (69) Note that the resulting state is not a tensor product state in general. ρE = a ra εa εa . εb . Action of the time evolution operator on a basis vector of HT is explicitly written as U (t)ej . The initial density matrices may be written as ρS = j pj ej ej .ja ek . (68) where ρS (ρE ) is the initial density matrix of the principal system (environment). The state of the total system is described by a density matrix ρ. E(ρS ) = ρS (t). Suppose ρ is uncorrelated initally at time t = 0. how the principal system state ρS at t = 0 evolves in time in the presence of its environment. A map which describes a general change of the state from ρS to E(ρS ) is called a quantum operation. Let {ej } be a basis of the system Hilbert space while {εa } be that of the environment Hilbert space. (71) . This is an example of a quantum operation. The qubit may be spatially ﬁxed and subject to timedependent noise. The total system is assumed to be closed and to evolve with a unitary matrix U (t) as ρ(t) = U (t)(ρS ⊗ ρE )U (t)† . The latter quantum map is our primary interest in this chapter. Other quantum operations include state change associated with measurement and state change due to noise. εa = k.b Ukb. we may still deﬁne the system density matrix ρS (t) by taking partial trace of ρ(t) over the environment Hilbert space as ρS (t) = trE [U (t)(ρS ⊗ ρE )U (t)† ]. Even under these circumstances.31 mean propagating in space.
a.32 where ej .k.7. which we take. This is equivalent with the trace† preserving property of E as 1 = trS ρS (t) = trS (E(ρS )) = trS a Ea Ea ρS for any ρS ∈ S(HS ). εa ej .b pj k √ ra Ukb. εa U (t)† ∗ pj ra Ukb. ρS (t) is written as ρS (t) = trE [U (t)(ρS ⊗ ε0 ε0 )U (t)† ] = a (I ⊗ εa )U (t)(ρS ⊗ ε0 ε0 )U (t)† (I ⊗ εa ) (I ⊗ εa )U (t)(I ⊗ ε0 )ρS (I ⊗ ε0 )U (t)† (I ⊗ εa ). εc Ulc. j.b. a = We will drop I⊗ from I ⊗ εa  hereafter. ρE = ε0 ε0 . Then we may write E(ρS ) = ρS (t) = a (74) Ea (t)ρS Ea (t)† . the density matrix ρ(t) is written as U (t)(ρS ⊗ ρE )U (t)† = j. εa = ej ⊗ εa for example.l ∗ pj ra Ukb. (73) To write down the quantum operation in a closed form.a pj ra U (t)ej . Completeness relation and tracepreserving property .c = (72) The partial trace over HE is carried out to yield ρS (t) = trE [U (t)(ρS ⊗ ρE )U (t)† ] = j. (76) where I is the unit matrix in HS . (75) This is called the operatorsum representation (OSR) of a quantum operation E.ja ek l √ ∗ ra el Ulb. Let us deﬁne the Kraus operator Ea (t) : HS → HS by Ea (t) = εa U (t)ε0 . we assume the initial environment state is a pure state.ja ek el Ulb. Using this expression.b.k. whenever it does not cause confusion. see § 2. εb el .ja . With this assumption. without loss of generality.a.ja = j.ja . Even when ρE is a mixed state.a.ja ek . we may always complement HE with a ﬁctitious Hilbert space to “purify” ρE . Note that {Ea } satisﬁes the completeness relation Ea (t)† Ea (t) a kl = a ε0 U (t)† εa εa U (t)ε0 kl = δkl .l.
2. it is called a monoid. for example. Here we give an interpretation of OSR as a noisy quantum channel. Unitarity condition may be relaxed when meaρ10 ρ11 surements are included as quantum operations. Let us consider a twoqubit system on which the CNOT gate acts. Then we ﬁnd E0 = (I ⊗ 0)UCNOT (I ⊗ 0 ) = P0 . Operatorsum representation and noisy quantum channel Operatorsum representation (OSR) introduced in the previous subsection seems to be rather abstract. we deﬁne the expectation value of the resulting density matrix as M(ρS ) = a † p a U a ρS U a . where Pi = i i. for example. By choosing Ua randomly with probability pa and applying it to ρS . . and hence no partial trace is involved. such that ea = ae = a. it turns out to be useful to relax the condition that U (t) be a time evolution operator. ∀a ∈ S. Given an initial principal system state ρS . we assume U be any operator including an arbitrary unitary gate. Note that no enviroment has been introduced in the above deﬁnition. If S has a unit element e. Instead.33 are satisﬁed since our total system is a closed system. b A set S is called a semigroup if S is closed under a product satisfying associativity (ab)c = a(bc). A general quantum map does not necessarily satisfy these properties. E1 = (I ⊗ 1)UCNOT (I ⊗ 0 ) = P1 . Therefore even though it is possible to compose two quantum operations. b 6.25 At this stage. (77) ρ00 ρ01 . and consequently E(ρS ) = P0 ρS P0 + P1 ρS P1 = ρ00 P0 + ρ11 P1 = where ρS = ρ00 0 0 ρ11 . Suppose we have a set of unitary matrices {Ua } and a set of nonnegative real numbers {pa } such that a pa = 1. there are inﬁnitely many U that yield the same E(ρS ).26 This occurs when a ﬂying qubit is sent through a noisy quantum channel which transforms the density matrix by Ua with probability pa . the set of quantum operations is not a group but merely a semigroup. Suppose the principal system is the control qubit while the environment is the target qubit.1. Tracing out the extra degrees of freedom makes it impossible to invert a quantum operation. (78) which we call a mixing process.
Operators Ea are identiﬁed with the Kraus operators. the space of density matrices is not a linear vector space. It is easily veriﬁed from the orthonormality of {εa } that U is indeed a unitary matrix.3. also acts on the space of density matrices and it maps a density matrix to another density matrix. . Then Eq. Let Λ be a superoperator acting on the system density matrices. Let HE = Span(εa ) be a Hilbert space with the dimension equal to the number of the unitary matrices {Ua }. 6. A quantum operation maps a density matrix to another density matrix linearly. Completely positive maps All linear operators we have encountered so far map vectors to vectors. Let √ us deﬁne Ea ≡ pa Ua . acting on the vector space of Hermitian matrices. under the action of Ua . Deﬁne formally the environment density matrix ρE = a pa εa εa  and U≡ a Ua ⊗ εa εa  (81) which acts on HS ⊗ HE .c A linear operator of this kind is called a superoperator. pa } we may introduce an “environment” with the Hilbert space HE as follows. What is meant hear is a linear operator. (80) Conversely.34 Now the correspondence between E(ρS ) and M(ρS ) should be clear. as † † ρS → Ea ρS Ea /tr Ea ρS Ea . where {εa } is an orthonormal basis. Partial trace over HE then yields E(ρS ) = trE [U (ρS ⊗ ρE )U † ] = a (I ⊗ εa ) b Ub ⊗ εb εb  ρS ⊗ c pc εc εc  × d Ud ⊗ εd εd  (I ⊗ εa ) † pa Ua ρS Ua = M(ρS ) = a (82) showing that the mixing process is also decribed by a quantum operation with a ﬁctitious environment.1. Λ : S(HS ) → c Of course. The system transforms. (78) is rewritten as M(ρS ) = a † Ea ρS Ea (79) and the equivalence has been shown. given a noisy quantum channel {Ua .
that ΛT is not necessarily a map S(HT ) → S(HT ). (83) We require not only that Λ be CP but also Λ(ρ) be a density matrix: tr Λ(ρS ) = tr a † Ea ρEa = tr a † Ea Ea ρ = 1. Moreover. Measurements as quantum operations We have already seen that a unitary evoluation ρS → U ρS U † and a mixing process ρS → i pi Ui ρS Ui† are quantum operations. any quantum operation obtained by tracing out the environment degrees of freedom is CP and preserves trace. .35 S(HS ). It may happen that ΛT (ρ) is not a density matrix any more. we have seen that Λt ⊗ I : ρ12 → ρpt deﬁned by Eq. We have already encountered this situation when we have introduced partial transpose operation in § 2. (84) This condition is satisﬁed for any ρ if and only if † Ea Ea = IS . It is clear that the transpose operation Λt : ρ1 → ρt on a singlequbit 1 state ρ1 preserves the density matrix properties. This section deals with measurements as quantum operations. Let HT = H1 ⊗ H2 be a twoqubit Hilbert space. this is not always the case. a (85) Therefore. A map Λ which maps a positive operator acting on HS to another positive operator on HS is said to be positive. In fact. Note. however. For a twoqubit density matrix ρ12 . 6.1. if its extension ΛT = Λ ⊗ In remains a positive operator for an arbitrary n ∈ N. however. The operator Λ is easily extended to an operator acting on HT by ΛT = Λ ⊗ IE . which acts on S(HS ⊗ HE ). A linear map Λ is CP if and only if there exists a set of operators {Ea } such that Λ(ρS ) can be written as Λ(ρS ) = a † Ea ρS Ea . We will see further examples of quantum operations in this section and the next. it is called a completely positive map (CP map). Theorem 6.2. where Hk is the kth qubit Hilbert space.7. (22) maps a density matrix to a matrix 12 which is not a density matrix when ρ12 is inseparable.
ψ e0 U † U ψ e0 = † † i. i (87) where the set {Pi } satisifes the completeness relation i Pi Pi† = I. 6. Thus we may regard the measurement process as a quantum operation ρS → i p(i) Pi ρS Pi = p(i) Pi ρS Pi . Operators {Mi† Mi } are said to form a POVM (positive operatorvalued measure). We have seen in Chapter 2 that the probability of observing λi upon a measurement of A in a state ρ is p(i) = λi ρλi = tr (Pi ρ) (86) and the state changes as ρ → Pi ρPi /p(i).1. Suppose a system and an environment. it should be noted that they are not unique type of measurements. Here we will deal with the most general framework of measurement and show that it is a quantum operation.j ψ ei Mi Mj ⊗ Iψ ej = ψ i Mi Mi ψ = 1.36 6. POVM We have been concerned with projective measurements so far. are acted by a unitary operator U . where Pi = λi λi  is the projection operator corresponding to the eigenvector λi . at the same time. However. which applies an operator Mi on the system and. prepared initially in a product state ψ e0 . Since ψ is arbitrary. This process happens with a probability p(i). Projective measurements Suppose we measure an observable A = i λi Pi . .2.2. (88) The system and its environment are correlated in this way.2. It is written explicitly as Ψ = U ψ e0 = i Mi ψ ei . The projective measurement is a special case of a quantum operation in which the Kraus operators are Ei = Pi . This state must satisfy the normalization condition since U is unitary. we must have Mi† Mi = IS . put the environment to ei for various i. i (89) where IS is the unit matrix acting on the system Hilbert space HS .
The combined system immediately after the measurement is 1 1 Mi ψ ei (IS ⊗ ek ek )U ψ e0 = (IS ⊗ ek ek ) p(k) p(k) i 1 = (91) Mk ψ ek . 6. Bitﬂip channel Consider a closed twoqubit system with a Hilbert space C2 ⊗ C2 . We obtain a measurement outcome λk with a probability p(k) = Ψ(IS ⊗ ek ek )Ψ = i.3. (90) where Ψ = U ψ e0 .1. It follows from the above observation for a pure state ψ ψ that the reduced density matrix immediately after the measurement is p(k) k † M k ρS M k = p(k) † M k ρS M k . in which {Mi } are the projective operators. (93) . p(k) Let ρS = i pi ψi ψi  be an arbitrary density matrix of the principal system. Decoherence appears as an error in quantum information processing.j † ψ ei Mi† (IS ⊗ ek ek )Mj ψ ej = ψMk Mk ψ . We call the ﬁrst qubit the “(principal) system” while the second qubit the “environment”.3. The next chapter is devoted to strategies to ﬁght against errors introduced in this section. k (92) This shows that POVM measurement is a quantum operation in which the Kraus operators are given by the generalized measurement operators {Mi }. Examples Now we examine several important examples which have relevance in quantum information theory. 6.37 Suppose we measure the environment with a measurement operator O = IS ⊗ i λi ei ei  = i λi (IS ⊗ ei ei ) . A bitﬂip channel is deﬁned by a quantum operation E(ρS ) = (1 − p)ρS + pσx ρS σx . The projective measurement is a special class of POVM. 0 ≤ p ≤ 1.
z . The Kraus operators are read oﬀ as E0 = 1 − pI.y. (ck ∈ R) (98) ρS = I+ 2 k=x. The circuit depicted in Fig. 4 is E(ρS ) = trE [V ρS ⊗ ψE ψE V † ] = (1 − p)ρS + pσx ρS σx . for example. The circuit is nothing but the inverted CNOT gate V = I ⊗ 0 0 + σx ⊗ 1 1. from which we obtain E(ρS ) = (1 − p)ρS + pσx ρS σx (96) (95) after tracing over the environment Hilbert space. Quantum circuit modelling a bitﬂip channel. 4 models the bitﬂip channel provided that the second qubit is in a mixed state (1 − p)0 0 + p1 1. (97) producing the same result as before. The choice of the second qubit input state is far from unique and so is the choice of the circuit. Suppose the initial state of the environment is a √ √ pure state ψE = 1 − p0 + p1 . 4. The gate is the inverted CNOT gate I ⊗ 0 0 + σx ⊗ 1 1. Then the output of the circuit in Fig. (94) Fig.38 The input ρS is bitﬂipped with a probability p while it remains in its input state with a probability 1 − p. Let us see what transformation this quantum operation brings about in ρS . E1 = √ pσx . The output of this circuit is V (ρS ⊗ [(1 − p)0 0 + p1 1]) V † = (1 − p)ρS ⊗ 0 0 + pσx ρS σx 1 1. We parametrize ρS using the Bloch vector as 1 ck σk .
Bloch sphere of the input state ρS (a) and output states of (b) bitﬂip channel and (c) phaseﬂip channel. .5 0 0. −cz ) with weights 1 − p and p respectively.5 0 0.2 is common to both channels. cy .5 0 0. Equation (99) shows that the quantum operation has produced a mixture of the Bloch vector states (cx .5 1 1 0. The Bloch sphere shrinks along the y. cz ) and (cx . (100) The input ρS is phaseﬂipped (0 → 0 and 1 → −1 ) with a probability p while it remains in its input state with a probability 1 − p.5 1 1 0. 5.5 1 Fig. Phaseﬂip channel Consider again a closed twoqubit system with the “(principal) system” and its “environment”.3.and zaxes. E1 = pσz . which results in the ellipsoid shown in Fig. (99) = 2 cx + i(1 − 2p)cy 1 − (1 − 2p)cz Observe that the radius of the Bloch sphere is reduced along the y. 1 1 0. 5 (b). Figure 5 (a) shows the Bloch sphere which represents the input qubit states.39 where 2 k ck ≤ 1.5 1 1 0.5 0. The probability p = 0.5 1 0 0. −cy . 0 ≤ p ≤ 1.5 0 0.5 0 0. 6. The corresponding Kraus operators are √ (101) E0 = 1 − pI. We obtain E(ρS ) = (1 − p)ρS + pσx ρS σx 1−p p (I + cx σx + cy σy + cz σz ) + (I + cx σx − cy σy − cz σz ) = 2 2 1 1 + (1 − 2p)cz cx − i(1 − 2p)cy .and the zaxes so that the radius in these directions is 1 − 2p. The phaseﬂip channel is deﬁned by a quantum operation E(ρS ) = (1 − p)ρS + pσz ρS σz .5 1 0.2.
40 Fig. The second qubit input state may be a pure state √ ψE = 1 − p0 + p1 . 6. 6. (107) = 1 − cz 2 (1 − 2p)(cx + icy ) Observe that the oﬀdiagonal components decay while the diagonal components remain the same. Equation (107) shows that the quantum operation has produced a mixture of the Bloch vector states (cx . (106) Let us work out the transformation this quantum operation brings about to ρS . cz ) and . Quantum circuit modelling a phaseﬂip channel. (102) (103) (104) (105) where the Kraus operators are E0 = 0V ψE = 1 − pI. E1 = 1V ψE = √ pσz . We obtain E(ρS ) = (1 − p)ρS + pσz ρS σz 1−p p (I + cx σx + cy σy + cz σz ) + (I − cx σx − cy σy + cz σz ) = 2 2 1 1 + cz (1 − 2p)(−cx − icy ) . The output of this circuit is V (ρS ⊗ [(1 − p)0 0 + p1 1]) V † = (1 − p)ρS ⊗ 0 0 + pσz ρS σz ⊗ 1 1. Let ρS be the ﬁrst qubit input state while (1 − p)0 0 + p1 1 be the second qubit input state. from which we obtain E(ρS ) = (1 − p)ρS + pσz ρS σz . for example. Then we ﬁnd † † E(ρS ) = trE [V ρS ⊗ ψE ψE V † ] = E0 ρS E0 + E1 ρS E1 . We parametrize ρS using the Bloch vector as before. The gate is the inverted controlledσz gate. A quantum circuit which models the phaseﬂip channel is shown in Fig. The circuit is the inverted controlledσz gate V = I ⊗ 0 0 + σz ⊗ 1 1. cy .
The phase after the quantum operaition is applied is a mixture of states with φ and φ + π. We have to somehow think out the way to overcome this theorem. or alternatively. When 000 is sent through this channel. This method cannot be applicable to qubits. To reduce channel errors.2]. due to nocloning theorem. Before we introduce quantum error correcting codes.41 (−cx . Namely. 101 or 110 with probability 3p2 (1 − p) and ﬁnally as 111 with probability p3 . .and the yaxes as 1 → 1 − 2p. for example. as 100. By taking the majority vote. We closely follow Steane30 here. 7. To reduce such errors. Introduction It has been shown in the previous chapter that interactions between a quantum system with environment cause undesirable changes in the state of the quantum system.2. cz ) with weights 1 − p and p respectively. we ﬁnd p0 = 0. Other examples will be found in [1. The initial state has a deﬁnite phase φ = tan−1 (cy /cx ) in the oﬀdiagonal components. as 011.2. In fact. it will be received as 000 with probability (1−p)3 . Each bit is assumed to ﬂip independently with a probability p. they appear as bitﬂip and phaseﬂip errors. however. we encode logical 0 by 000 and 1 by 111. 010 or 001 with probability 3p(1−p)2 . Threequbit bitﬂip code: the simplest example It is instructive to introduce a simple example of quantum error correcting codes (QECC).1.1. Suppose we transmit a serise of 0’s and 1’s through a noisy classical channel. 7. if we use more bits to encode 0 or 1. Thus a bit 0 sent through the channel will be received as 0 with probability 1 − p and as 1 with probability p. for example.972 and p1 = 0. The success probability p0 increases as p approaches to 0. We obtain p0 p ≥ 0. we must implement some sort of error correcting mechanism in the algorithm. The radius of the Bloch sphere is reduced along the x. Quantum Error Correcting Codes 7. In the case of qubits. −cy . This process is called the phase relaxation process. Figure 5 (c) shows the eﬀect of the phaseﬂip channel on the Bloch sphere for p = 0. we correctly reproduce the desired result 0 with probability p0 = (1 − p)3 + 3p(1 − p)2 = (1 − p)2 (1 + 2p) while fails with probability p1 for suﬃciently small p1 = 3p2 (1 − p) + p3 = (3 − 2p)p2 . we have a brief look at the simplest version of error correcting code in classical bits. or the T2 process in the context of NMR. we may invoke to majority vote.028 for p = 0.
j ∈ {0. 7. It is immediately found that the output of this circuit is ψ E = a000 + b111 as promised. Let ψ = a0 +b1 be the state she wants to send. A quantum circuit which implements the encoding (109) is easily found from our experience in CNOT gate. Bitﬂip QECC Suppose Alice wants to send a qubit or a series of qubits to Bob through a noisy quantum channel. the operator X is applied to the qubit with probability p and consequently the state is mapped to ψ → ψ = Xψ = a1 + b0 . a2 + b2 = 1} (110) the code and each member of C a codeword. We use this fact to triplicate the basis vectors as ψ 00 = (a0 + b1 )00 → ψ E = a000 + b111 . In other words. We borrow terminologies from classical error correcting code (ECC) and call the set C = {a000 + b111 a.3. 1} when the target bit is initially set to 0 . Let us recall that the action of a CNOT gate is CNOT : j0 → jj . If she is to transmit a serise of qubits. b ∈ C. 7 (a) whose input state is ψ 00 . This redundancy makes it possible to detect errors in ψ E and correct them as we see below.42 7.2. The state ψ E is called the logical qubit while each constituent qubit is called the physical qubit.2. 1} and therefore it duplicates the control bit j ∈ {0.1. Let us consider the circuit shown in Fig. 7. It is important to note that the state ψ is not triplicated but only the basis vectors are triplicated.2. we want to mimic somehow the classical counterpart without using a clone machine. (108) We have already seen in the previous section that this channel is described by a quantum operation (93). (109) where ψ E denotes the encoded state. she sends them one by one and the following argument applies to each of the qubits. Let p be the probability with which a qubit is ﬂipped and we assume there are no other types of errors in the channel. Transmission Now the state ψ E is sent through a quantum channel which introduces bitﬂip error with a rate p for each qubit independently. Encoding To reduce the error probability.2. We assume p is .
State Bob receives a000 + b111 a100 + b011 a010 + b101 a001 + b110 a110 + b001 a101 + b010 a011 + b100 a111 + b000 Probability (1 − p)3 p(1 − p)2 p(1 − p)2 p(1 − p)2 p2 (1 − p) p2 (1 − p) p2 (1 − p) p3 7. Quantum circuits to (a) encode. Table 1 lists possible received states and the probabilities with which these states are received.2. (c) make correction to a relevant qubit and (d) decode. 7. For this purpose. Error syndrome dectection and correction Now Bob has to extract from the received state which error occurred during qubits transmission. Bob prepares two ancillary qubits in the state 00 as depicted in Fig. State Bob receives and the probability which this may happen. (b) detect bitﬂip error syndrome. The gate NX stands for the bitﬂip noise.43 Fig.4. suﬃciently small so that not many errors occur during qubit transmission. The received state depends on in which physical qubit(s) the bitﬂip error occurred. Table 1. 7 (b) and apply four CNOT operations whose control bits are the encoded qubits while the target qubits are Bob’s .
It is instructive to visualize what errors do to the encoded basis vectors. An action of Xi . j. there are exactly two states with the same ancilla state.44 two ancillary qubits. When the ancillary state is 10 . Therefore the latter probability is negligible compared to the former for suﬃciently small p. j. under which the vectors 000 and 111 correspond to diagonally separated vertices. 7 (b) that A = x1 ⊕x2 and B = x1 ⊕x3 . The vertices of the cube have coordinates (i. It is important to realize that (i) the syndrome is independent of a and b and (ii) the received state a100 + b011 remains intact. Both of the ancillary qubits are ﬂipped since x1 ⊕ x2 = x1 ⊕ x3 = 1 for both 100 and 011 . j. Does it mean this error extraction scheme does not work? Now let us compare the probabilities associated with the same ancillary state. It is seen from Fig. Note that the former is received with probability p(1 − p)2 while that latter with p2 (1 − p). Let x1 x2 x3 be a basis vectors Bob has received and let A (B) be the output state of the ﬁrst (second) ancilla qubit. State after error syndrome extraction (a000 + b111 )00 (a100 + b011 )11 (a010 + b101 )10 (a001 + b110 )01 (a110 + b001 )01 (a101 + b010 )10 (a011 + b100 )11 (a111 + b000 )00 Probability (1 − p)3 p(1 − p)2 p(1 − p)2 p(1 − p)2 p2 (1 − p) p2 (1 − p) p2 (1 − p) p3 possible states. k ∈ {0. sends these basis vectors to the nearest neighbor vetices. we have detected an error without measuring the received state! These features are common to all QECC. Note that among eight Table 2. the operator X = σx acting on the ith qubit. These diﬀerence are detected by the pairs of CNOT gates in Fig. We list the result of other cases in Table 2. k). The error extracting sequence transforms the anicillary qubits as (a100 + b011 )00 → a10011 + b01111 = (a100 + b011 )11 . Note that the ﬁrst qubit state in both of the basis vectors is diﬀerent from the second and the third qubit states. We assign a vector ijk to the vertex (i. which . States after error extraction is made and the probabilities with which these states are produced. 1}. k) where i. Consider a cube with the unit length. Let a100 +b011 be the received logical qubit for example. for example. 7 (b). there are two possible reseived states a010 + b101 and a101 + b010 .
This is nothing but the inverse transformation of the . In contrast. we immediately ﬁnd that the following action must be taken: error syndrome correction to be made 00 identity operation (nothing is required) 01 apply σx to the third qubit 10 apply σx to the second qubit 11 apply σx to the ﬁrst qubit Suppose the syndrome is 01.972. while the success probability is enhanced from 0.1. The state Bob received is likely to be a001 + b110 . If Bob receives the state a110 + b001 . Decoding Now that Bob has corrected an error.9 to 0. he will obtain (I ⊗ I ⊗ σx )(a110 + b001 ) = a111 + b000 . for any error sydrome. for example.45 diﬀer from the correct basis vectors in the ith position. for example. Then the error rate is lowered to P (error) = 0. Bob recovers the initial state Alice has sent by applying I ⊗ I ⊗ σx on the received state: (I ⊗ I ⊗ σx )(a001 + b110 ) = a000 + b111 . The set of two bits is called the (error) syndrome and it tells Bob in which physical qubit the error occurred during transmission. The intersection of the sets of vectors obtained by a single action of Xi on 000 and 111 is an empty set. The probability with which multiple error occurs is found from Table 1 as P (error) = 3p2 (1 − p) + p3 = 3p2 − 2p3 . Let p = 0.5. Bob applies correcting procedure to the received state according to the error sydrome he has obtained. Ignoring extra error states with small probabilities. 7. In fact. and hence it is less likely to happen for suﬃciently small error rate p. The latter case occurs if and only if more than one qubit are ﬂipped. Therefore an action of a single error operator X can be corrected with no ambiguity. Now Bob measures his ancillary qubits and obtains two bits of classical information. unfortunately. what is left for him is to decode the encoded state. Bob obtaines either a000 +b111 or a111 + b000 . (111) This error rate is less than p if p < 1/2.028. success probability has been enhanced from 1 − p to 1 − P (error) = 1 − 3p2 + 2p3 .2.
7. Then the encoded qubits are sent through a noisy channel. Note that everything is done without knowing what the origial state is. (112) 7.6. which maps a state ψ to Uα ψ = cos αψ + i sin αXψ . which may be subject to an error. syndrome measurement singles out a particular error state which produces the observed syndrome. albeit simple. is then entangled with ancillary qubits. Miracle of entanglement This example. for example.2.2. (115) (114) .46 encoding (109). 7 (d) that CNOT12 CNOT13 (a000 + b111 ) = a000 + b100 = (a0 + b1 )00 . Once syndrome is found. The received state. whose state reﬂects the error which occurred during the state transmission. In a sense. We prepare some redundant qubits which somehow “triplicate” the original qubit state to be sent without violating nocloning theorem. which causes a bitﬂip in at most one of the qubits. it is an easy task to transform the received state back to the original state. upon the measurement of the ancillary qubits. contains almost all fundamental ingredients of QECC. Bob then receives (Uα ⊗ I ⊗ I)(a000 + b111 ) = cos α(a000 + b111 ) + i sin α(a100 + b011 ). Suppose noise in the channel is characterized by a contiuous paramter α as Uα = eiαX = cos αI + iX sin α. Continuous rotations We have considered noise X so far. This results in an entangled state A bitﬂip error in the kth qubit ⊗ corresponding error syndrome . collapses to a state with a bitﬂip error corresponding to the observed error syndrome.7. It can be seen from Fig. k (113) The wave function. Suppose Uα acts on the ﬁrst qubit.
8. linearity of a quantum circuit guarantees that any QECC. Continuous rotation Uα for any α may be corrected in this way. This new discipline called quantum information processing (QIP) is expected to solve a certain class of problems that current digital computers cannot solve in a practical time scale. DiVincenzo Criteria We have learned so far that information may be encoded and processed in a quantummechanical way. (1) A scalable physical system with well characterized qubits. is an entangled state cos α(a000 + b111 )00 + i sin α(a100 + b111 )11 . DiVincenzo criteria In his inﬂuential article. Recall that a classical computer also requires memory to . To begin with. Measurement of the error syndrome yields either 00 or 11.38 DiVincenzo proposed ﬁve criteria that any physical system must satisfy to be a viable quantum computer.1. the received state collapses to Xψ = a100 + b011 and this happens with a probability sin2 α. DiVincenzo proposed necessary conditions. Although a small scale quantum information processor is already available commercially. corrects continuous error Uα . 8. which corrects the bitﬂip error X. In general. In the latter case. we outline these conditions as well as two additional criteria for networkability. A quantum computer should have at least 102 ∼ 103 qubits to be able to execute algorithms that are more eﬃcient than their classical counterparts. Bob applies I (X) to the ﬁrst qubit to correct the error when the syndrome readout is 00 (11). socalled the DiVincenzo criteria that any physical system has to fulﬁll to be a candidate for a viable quantum computer.38 In the next section. We summarize the relevant parts of these criteria in this section. It is clear that error Uα may act on the second or the third qubit. we need a quantum register made of many qubits to store information. on the other hand. before the syndrome measurement is made. In the former case the state collapses to ψ = a000 + b111 and this happens with a probability cos2 α.47 The output of the error syndrome detection circuit. physical realization of large scale quantum information processors is still beyond the scope of our currently available technology. (116) see Table 2.
Suppose you are not able to reset your (classical) computer. A similar scenario is also proposed for Josephson junction qubits. Then you will never trust the output of some computation even though processing is done correctly. we observe the system to be in an undesired state upon such measurement. Moreover it should be scalable up to a large number of qubits. A system may be made of several diﬀerent kinds of qubits. For example. The system is in the ground state with a good precision at low temperatures satis∆E. Then we may transform the system to the desired ﬁducial state by applying appropriate gates. Simultaneous usage of several types of qubits may be the most promising way to achieving a viable quantum computer.48 store information. a spin 1/2 nucleus or two mutually orthogonal polarization states (horizontal and vertical. 0 and 1 . (2) The ability to initialize the state of the qubits to a simple ﬁducial state. we may use projective measurement fying kB T to project the system onto a desired state. in which two ﬂux qubits are coupled through a quantized LC circuit. . Alternatively. of a multidimensional Hilbert space. initialization may be done simply by cooling to bring the system into its ground state. may be deﬁned as: (1) hyperﬁne/Zeeman sublevels in the electronic ground state of ions (2) a ground state and an excited state of a weakly allowed optical transition and (3) normal mode of ion oscillation. an electron. The twodimensional vector space of a qubit may be extended to threedimensional (qutrit) or. it is im . We may also employ a twodimensional subspace. In some cases. for instance. Let ∆E be the diﬀerence between energies of the ﬁrst excited state and the ground state. For some realizations. however. Each qubit must be separately addressable. such as liquid state NMR. In many realizations. The simplest way to realize a qubit physically is to use a twolevel quantum system. for example) of a single photon can be a qubit. α2 + β2 = 1. ddimensional (qudit). Therefore initialization is an important part of both quantum and classical information processors. such as atomic energy levels. . A multiqubit state is expanded in terms of the tensor products of these basis vectors. the two states are identiﬁed as the basis vectors. more generally. of the Hilbert space so that a general single qubit state takes the form ψ = α0 + β1 . In any case. such as 00 . such as the ground state and the ﬁrst excited state. Qubits in an ion trap quantum computer. 0 .
49
possible to cool the system down to extremely low temperatures. In those cases, we are forced to use a thermally populated state as an initial state. This seemingly diﬃcult problem may be amended by several methods if some computational resources are sacriﬁced. We then obtain an “eﬀective” pure state, socalled the pseudopure state, which works as an initial state for most purposes. Continuous fresh supply of qubits in a speciﬁed state, such as 0 , is also an important requirement for successful quantum error correction. as we have seen in Section 7. (3) Long decoherence times, much longer than the gate operation time. Hardware of a classical computer lasts long, for on the order of 10 years. Things are totally diﬀerent for a quantum computer, which is fragile against external disturbance called decoherence, see Section 6. Decoherence is probably the hardest obstacle to building a viable quantum computer. Decoherence means many aspects of quantum state degradation due to interactions of the system with the environment and sets the maximum time available for quantum computation. Decoherence time itself is not very important. What matters is the ratio “decoherence time/gate operation time”. For some realizations, decoherence time may be as short as ∼ µs. This is not necessarily a big problem provided that the gate operation time, determined by the Rabi oscillation period and the qubitcoupling strength, for example, is much shorter than the decoherence time. If the typical gate operation time is ∼ ps, say, the system may execute 1012−6 = 106 gate operations before the quantum state decays. We quote the number ∼ 105 of gates required to factor 21 into 3 and 7 by using Shor’s algorithm.40 There are several ways to eﬀectively prolong decoherence time. A closedloop control method incorporates QECC, while an openloop control method incorporates noiseless subsystem41 and decoherence free subspace (DFS).42 (4) A “universal” set of quantum gates. Suppose you have a classical computer with a big memory. Now you have to manipulate the data encoded in the memory by applying various logic gates. You must be able to apply arbitrary logic operations on the memory bits to carry out useful information processing. It is known that the NAND gate is universal, i.e., any logic gates may be implemented with NAND gates. Let H(γ(t)) be the Hamiltonian of an nqubit system under consideration, where γ(t) collectively denotes the control parameters in
50
the Hamiltonian. The timedevelopment operator of the system is T U [γ(t)] = T exp − i H(γ(t))dt ∈ U(2n ), where T is the timeordering operator. Our task is to ﬁnd the set of control parameters γ(t), which implements the desired gate Ugate as U [γ(t)] = Ugate . Although this “inverse problem” seems to be diﬃcult to solve, a theorem by Barenco et al. guarantees that any U(2n ) gate may be decomposed into singlequbit gates ∈ U(2) and CNOT gates.16 Therefore it suﬃces to ﬁnd the control sequences to implement U(2) gates and a CNOT gate to construct an arbitrary gate. Naturally, implementation of a CNOT gate in any realization is considered to be a milestone in this respect. Note, however, that any twoqubit gates, which are neither a tensor product of two onequbit gates nor a SWAP gate, work as a component of a universal set of gates.43 (5) A qubitspeciﬁc measurement capability. The result of classical computation must be displayed on a screen or printed on a sheet of paper to readout the result. Although the readout process in a classical computer is regarded as too trivial a part of computation, it is a vital part in quantum computing. The state at the end of an execution of quantum algorithm must be measured to extract the result of the computation. The measurement process depends heavily on the physical system under consideration. For most realizations, projective measurements are the primary method to extract the outcome of a computation. In liquid state NMR, in contrast, a projective measurement is impossible, and we have to resort to ensemble averaged measurements. Measurement in general has no 100% eﬃciency due to decoherence, gate operation error and many more reasons. If this is the case, we have to repeat the same computation many times to achieve reasonably high reliability. Moreover, we should be able to send and store quantum information to construct a quantum data processing network. This “networkability” requires following two additional criteria to be satisﬁed. (6) The ability to interconvert stationary and ﬂying qubits. Some realizations are excellent in storing quantum information while long distant transmission of quantum information might require diﬀerent physical resources. It may happen that some system has a Hamiltonian which is easily controllable and is advantageous in executing
51
quantum algorithms. Compare this with a current digital computer, in which the CPU and the system memory are made of semiconductors while a hard disk drive is used as a mass storage device. Therefore a working quantum computer may involve several kinds of qubits and we are forced to introduce distributed quantum computing. Interconverting ability is also important in long distant quantum teleportation using quantum repeaters. (7) The ability to faithfully transmit ﬂying qubits between speciﬁed locations. Needless to say, this is an indispensable requirement for quantum communication such as quantum key distribution. This condition is also important in distributed quantum computing mentioned above.
8.2. Physical realizations There are numerous physical systems proposed as possible candidates for a viable quantum computer to date.44 Here is the list of the candidates; (1) (2) (3) (4) (5) (6) (7) (8) Liquidstate/Solidstate NMR and ENDOR Trapped ions Neutral atoms in optical lattice Cavity QED with atoms Linear optics Quantum dots (spinbased, chargebased) Josephson junctions (charge, ﬂux, phase qubits) Electrons on liquid helium surface
and other unique realizations. ARDA QIST roadmap44 evaluates each of these realizations. The roadmap is extremely valuable for the identiﬁcation and quantiﬁcation of progress in this multidisciplinary ﬁeld.
Acknowledgements This summer school was supported by the “Open Research Center” Project for Private Universities: matching fund subsidy from MEXT (Ministry of Education, Culture, Sports, Science and Technology). Special thanks are due to the other organizers and coeditors of this lecture notes, Takashi Aoki, Robabeh Rahimi Darabad and Akira SaiToh.
52
References
1. M. Nakahara and T. Ohmi, Quantum Computing: From Linear Algebra to Physical Realizations, (Taylor and Francis, 2008). 2. M. A. Neilsen and I. L. Chuang, Quantum Computation and Quantum Information, (Cambridge University Press, 2000). 3. E. Rieﬀel and W. Polak, ACM Computing Surveys (CSUR) 32 (2000) 300. 4. Y. Uesaka, Mathematical Principle of Quantum Computation, (Corona Publishing, Tokyo, in Japanese, 2000). 5. P. A. M. Dirac, Principles of Quantum Mechanics (4th ed.), (Clarendon Press, 1981). 6. L. I. Shiﬀ, Quantum Mechanics (3rd ed.), (McGrawHill, 1968). 7. A. Messiah, Quantum Mechanics, (Dover, 2000). 8. J. J. Sakurai, : Modern Quantum Mechanics (2nd Edition), (Addison Wesley, Boston, 1994). 9. L. E. Ballentine, Quantum Mechanics, (World Scientiﬁc, Singapore, 1998). 10. A. Peres, Quantum Theory: Concepts and Methods, (Springer, 2006). 11. A. SaiToh, R. Rahimi and M. Nakahara, eprint quantph/0703133. 12. A. Peres, Phys. Rev. Lett. 77 (1996) 1413. 13. M. Horodecki et al., Phys. Lett. A 223 (1996) 1. 14. W. K. Wootters, and W. H. Zurek, Nature 299 (1982) 802. 15. M. A. Nielsen et al., Nature 396 (1998) 52. 16. A. Barenco et al., Phys. Rev. A 52 (1995) 3457. 17. Z. Meglicki, http://beige.ucs.indiana.edu/M743/index.html 18. D. Deutsch, Proc. Roy. Soc. Lond. A, 400 (1985) 97. 19. D. Deutsch and R. Jozsa, Proc. Roy. Soc. Lond. A, 439 (1992) 553. 20. E. Bernstein and U. Vazirani, SIAM J. Comput., 26 (1997) 1411. 21. D. R. Simon, Proc. 35th Annual Sympo. Found. Comput. Science, (IEEE Comput. Soc. Press, Los Alamitos, 1994) 116. 22. T. Mihara and S. C. Sung, Comput. Complex. 12 (2003) 162. 23. M. A. Neilsen and I. L. Chuang, Quantum Computation and Quantum Information, (Cambridge University Press, 2000). 24. K. Hornberger, quantph/0612118. 25. H. Barnum, M. A. Nielsen and B. Schumacher, Phys. Rev. A 57 (1998) 4153. 26. Y. Kondo, et al., J. Phys. Soc. Jpn. 76 (2007) 074002. 27. G. Lindblad, Commun. Math. Phys. 48 (1976) 119. 28. V. Gorini, A. Kossakowski and E. C. G. Sudarshan, J. Math. Phys., 17 (1976) 821. 29. A. J. Fisher, Lecture note available at http://www.cmmp.ucl.ac.uk/~ajf/course_notes.pdf 30. A. M. Steane, quantph/0304016. 31. P. W. Shor, Phys. Rev. A 52 (1995) 2493. 32. A. Hosoya, Lectures on Quantum Computation (Science Sha, in Japanese, 1999). 33. F. J. MacWilliams and N. J. A. Sloane, The Theory of ErrorCorrecting Codes, (NorthHolland, Amsterdam, 1977). 34. A. R. Calderbank and P. W. Shor, Phys. Rev. A 54 (1996) 1098.
53 35.. Salomaa and S. 42. A 70 (2004) 012319. quantph/0211071. P. Lett. Rev. B. P. K. Vartiainen et al. Lett. 79 (1997) 3306. Imai. Whaley. J. Zanardi and M. DiVincenzo. A. 43. Rev. Rev. Rev. Phys. Nakahara. M. M. Zanardi. 36. A 60 (1999) 729(R). Soc. Proc. R. P. Rev. Rev. Lett. Knill. London A 452 (1996) 567. Phys. Whaley. Rev. Takagi (eds.) “Physical Realization of Quantum Computing: Are the DiVincenzo Criteria Fulﬁlled in 2004?” (World Scientiﬁc. 77 (1996) 793. Phys. M. Guo. Singapore. Shor. Steane. Matsumoto and H. B. 41. Phys. P. D. I. Rev. and K. 2525 (2000). J Niwa.gov/ . 012301 (2001). 2006). G. Zanardi. Phys. Phys. Rev. Ekert. P. Lidar. E. 39. P. Duan and G. Lett. L. A 51 (1995) 1015. Phys. K. W. Suominen and A. Ritter. M. Rev. 77 (1996) 3260. A 60 (1999) 1944. Phys. C. M. Rev. Palma. W. Phys. A 72 (2005) 012305. D. and L. A.lanl. DiVincenzo and P. Phys. A. Phys. S. Laﬂamme. D. Lidar. A 63. Rev. Fortschr. http://qist. Kanemitsu. D. 81 (1998) 2594. L. Chuang. 37. 84. Phys. 38. M. A. G. 40. Lett. Lett. 79 (1997) 1953. 44. Viola. R. DiVincenzo. Bacon. K. D. Rasetti. D. and K. 48 (2000) 771. Phys.
This page intentionally left blank .
ac. Introduction Historically in the process of development of theoretical physics.math. Braid Group.55 BRAID GROUP AND TOPOLOGICAL QUANTUM COMPUTING TAKAYOSHI OOTSUKA Department of Physics. Also. it is impossible to discuss the gravitation ﬁeld equation due to Einstein 1 Rij − Rgij + Λgij = Tij 2 without the language of Riemannian geometry. Kinki University. TQC. Moreover. Presentation of a Group. For instance. there corresponds one geometry. Markov Theorem. Osaka 5778502. and that general relativity could be discussed as the Lorentz manifold theory.jp KAZUHIRO SAKUMA Department of Mathematics. was employed as a model of the space universe. The purpose of this note is to discuss relation between a unitary representation of braid group and topological quantum computing. Jones Polynomial. modern geometry has constantly provided a strong method and always played an important role at both fundamental and advanced level. In the Erlangen program Felix Klein declared that “for one transformation group. Japan ∗ Email: sakuma@math.ac. we may say that based on the characterization and requirement from physics side. Keywords: Symmetric Group.” This conception by Klein has . there has been much progress in the study of geometry closely related to manifold theory. which is a direct generalization of surface theory.jp This is a survey article on the braid group and topological quantum computing. Kinki University. Modular Functor. Osaka 5778502.kindai. Japan ∗ Email: ootsuka@alice.kindai. it is obviously concluded from the facts that the notion of a Riemannian manifold. Representation of a Group. 1. Alexander Theorem.
a group often appears also in physics in a natural way. For example. For readers’ convenience. In its simplest form it is the SideAngleSide theorem of middle school geometry which characterizes triangles up to congruence or the AngleAngle theorem up to similarity. In addition. It gave not merely a solution of an equation but an answer to a geometrical problem (e. which is called a geometry of X belonging to G. braid groups are used in describing topological quantum computation. a regular npolygon is of possible construction if and only if the Galois πi group a Gal(Q(e n )/Q ∼ (Z/2nZ)× has order 2r ). Larsen and Z. On one hand. for a given space X and a transformation group G. a group naturally appears in considering a geometrical problem. More precisely. Wang. Conversely. Acknowledgement. geometry is a mathematics in which one searches for invariant properties of not only ﬁgures but also functions on the ﬁgures or maps between ﬁgures under an element of a group G. Thus when we consider = a problem basically related with group theory. So. He also wishes to thank the participants at the summer school for a Here we do not give a deﬁnition of Galois group because we do not need it in the later discussion .g. the Euclidean geometry is a branch to study invariant properties of a ﬁgure A in X = R2 under congruent transformation or similar transformation. H. M.56 been one of the most important teaching principles in modern geometry. As is seen in § 4. we provide an appendix where we discuss basic facts on the fundamental groups including its deﬁnition. Group theory was established by Galois with an excellent idea in order to solve an equation systematically. The authors wishes to thank Mikio Nakahara for well organizing summer school held at August 2007 and stimulating discussion. This is because geometry is a mathematics in which one solves equations on a geometrical object. in the next section we discuss braid groups together with basic facts on group theory. we see that it is tied up with geometry. we can study invariant properties of a subset A in X under a transformation in G. Thus from a modern mathematical viewpoint. The ultimate purpose of group theory which is one of the branches in algebra is to give a group structure for any set and classify all groups up to isomorphism. diﬀerential topology is a branch to study invariant properties of a manifold up to diﬀeomorphism or homeomorphism. Freedman. which is a recent work by M.
c)) = µ(µ(a. µ(b. Hg = {hg. First we start with recalling the deﬁnition of a group and fundamental concepts in group theory: Deﬁnition 2. b. If H is a group under the product µ of G. Braid Groups In this section we review basic concepts of braid group mainly from an algebraic viewpoint. if a map (binary operation) µ : G × G → G satisﬁes the following three conditions. Note that we can make a set G to be a group with diﬀerent products in general. b ∈ G. Let H be a subset of a group (G. a) = e. In the above deﬁnition a binary operation µ is called a product and it is usually written as a · b ∈ G or ab ∈ G in place of µ(a. a ) = µ(a . (2) For any a ∈ G there exists an element e ∈ G such that µ(a. If gH = Hg for any g ∈ G. µ). h ∈ H}. . which will be a quick course for physicists. where a is called an inverse element of a and we write a−1 ∈ G. (3) For any a ∈ G there exists an element a ∈ G such that µ(a. µ) or simply G is called a group. Let G be a group and H a subgroup of G. The second author would like to thank his student Kumi Kobata for her help in writing this paper. e) = µ(e.57 their useful comments and inspiring questions. then H is called a normal subgroup of G. Note that H is a normal subgroup of G if and only if g −1 hg ∈ H for any g ∈ G and any h ∈ H. where e is called a unit element of G. c ∈ G we have µ(a. then a pair (G. b). b−1 ) ∈ H for any a. b) for simplicity. c). a) for any a. a) = a. µ) if we have µ(a. 2. Then. (1) For any a. for any g ∈ G we deﬁne two subgroups as follows: gH = {gh. Given a group (G. H is called a subgroup of G.1. For any set G. b ∈ H. Note that a subset H is a subgroup of G if and only if µ(a. then we call G an abelian group. b) = µ(b. h ∈ H}.
. b) = ab. Furthermore. m ∈ Z} is a normal subgroup of Z. y) = x + y for any x. b) = a + b and a subgroup of (R. Then we say that (R. . y2 )) = (x1 + x2 .. +) is an additive group similarly as in (1). The ndimensional linear space Rn is naturally a group with µ(x. [1]. m ∈ Z} and [a] + [b] = [a + b] = {a + b + nm. (1) Let R denote the set of real numbers. b) = ab. For [g1 ]. (K. x. i. y ∈ Rn . Then. y ∈ Z} is a group with µ((x1 . we need to deﬁne a map which preserves the products. where g is called a representative. In order to compare the structures of given groups. It is often written as Z/nZ = Zn = {[0]. For any g ∈ G the equivalence class of g is denoted by [g] = gH = Hg ∈ G/H. the product [g1 g2 ] does not depend on the choices of representatives g1 . y1 ).1. (x2 . Then we deﬁne g1 ∼ g2 if and only −1 if g1 g2 ∈ H. The direct product Z ⊕ Z = {(x. µ ) is not a group when we employ another operation µ (a. (g1 g2 )−1 = g2 g1 ∈ H implies g2 ∼ g1 if we suppose that −1 −1 −1 g1 ∼ g2 . Example 2. It is easy to see that R is a group with a product deﬁned by µ(a. µ ) is not a group with µ (a. (R. y). (2) Let Z denote the set of integers. This relation ∼ is an equivalence relation. and (g1 g2 )(g2 g3 ) = g1 g3 implies g1 ∼ g3 if we suppose that g1 ∼ g2 .e. But (Z. Here note that the product is welldeﬁned. µ). Consider the quotient group Z/nZ. b) = a + b. However. or C (the set of complex numbers). µ) is an additive group. . m ∈ Z} is welldeﬁned. [n − 1]}. Then. [a] = {a + nm. gg −1 = e ∈ H −1 −1 implies g ∼ g. Then.58 Let H be a normal subgroup of G. R. g2 ∼ g3 . (3) Let K = Q (the set of rational numbers). This is beacause there exists a unit 1 ∈ R but there is no inverse for 0 ∈ R. [g2 ] ∈ G/H the product is deﬁned as follows: [g1 ][g2 ] = [g1 g2 ] ∈ G/H. g2 . b) = ab. It is easy to see that nZ = {nm. . (4) Let n be a natural number. We denote the quotient group under this relation by G/H. which is called a multiplicative group. y1 + y2 ). . Z is a group with µ(a. it is easy to see that K − {0} is a group with a product deﬁned by µ(a.
2. ϕ = is said to be injective. Let ϕ : G1 → G2 be a homomorphism. In other words. If a homomorphism ϕ : G → G is bijective. Given a homomorphism µ : G → G . It is easy to see that the map f is a welldeﬁned homomorphism and bijective. i. then ϕ is called a homomorphismb . . ϕ(g) = e }.e. ∀g ∈ G}. (1) if ϕ(a) = ϕ(b). Then we deﬁne a map f : G/Ker(ϕ) → Im(ϕ) by f ([g]) = ϕ(g). then a = b. = Exercise 2. b If we omit the notation of the products. we can consider a quotient group G/Ker(ϕ). then ϕ is called an isomorphism. ϕ is said to be surjective. two diﬀerent sets are identiﬁed even if they have diﬀerent products as groups but there exists an isomorphism between them. we can deﬁne a set Im(ϕ) = ϕ(G) = {ϕ(g) ∈ G . which is called the image of ϕ. we can deﬁne a set Ker(ϕ) = {g ∈ G.1. ϕ(µ(a. i. µ ) be groups.e. we have the following commutative diagram: −−− G×G − − − → G µ G −−−→ −−− ϕ ϕ×ϕ ×G µ G . which is called the kernel of ϕ.. When (2) holds. Given a homomorphism ϕ : G → G . and (2) Im(ϕ) = G .. We say that G is isomorphic to G if there exists an isomorphism ϕ : G → G . Moreover. Note that Ker(ϕ) is a normal subgroup of G and Im(ϕ) a subgroup of G . the deﬁnition of a homomorphism is simply given as follows: ϕ(ab) = ϕ(a)ϕ(b). b)) = µ (ϕ(a). µ) and (G . denoted by G ∼ G . In group theory. If a map ϕ : G → G preserves the products. Let (G. When (1) holds. ϕ(b)). Any element of G/Ker(ϕ) is denoted by [g] as an equivalence class of g ∈ G. which means that G/Ker(ϕ) ∼ Im(ϕ).59 Deﬁnition 2. Let e be a unit element of G .
g1 . If there are a ﬁnite (or inﬁnite) number of elements g1 . . an ). are generators of G and that G is generated by those generators. . for x = abc. . . . . . . Then.e. . e. . . aa−1 = e (empty word). = (4) Show that if ϕ is an isomorphism. . If G is generated by a ﬁnite number of generators.. . We employ. g2 . (3) Prove that G1 /Ker(ϕ) ∼ Im(ϕ). a2 . b. show that ϕ is an isomorphism. . .60 (1) Show that if e ∈ G1 is an unit element. Deﬁnition 2. . aa. abc. g2 . . denoted by Fn or F (a1 . then ϕ−1 is also an isomorphism. a−1 n 1 2 with a group structure under the above product is called a free group with rank n. . c−1 for a. . (2) Show that ϕ(g −1 ) = {ϕ(g)}−1 . .4. then we say that g1 . . We call this an irreducible word. . For example. For two words x and y. a−1 . Let G be a group. . It is easy to see that the set of all irreducible words consisting of n elements make a group under the above product. then e uniquely exists and ϕ(e) ∈ G2 is also an unit element. . The set of all irreducible words consisting of distinct 2n elements a1 . and the inverses a−1 . namely.3. We can consider an empty word “ ”or e. ccc. . aabbcc. . . b. we deﬁne the product xy as its juxtaposition. ababab. an . g2 . a. . ∈ G such that an arbitrary element of g ∈ G can be expressed by a ﬁnite (or inﬁnite) number of products of those −1 −1 g1 . . c. g2 . b. a1 . . an are generators . Now we consider a sequence with a ﬁnite number of elements. i. . Such a sequence consisting of a. a−1 . b. n. (5) Suppose that there exists a homomorphism ψ : G2 → G1 such that ψ ◦ ϕ = idG1 and ϕ ◦ ψ = idG2 . then we ignore it like the following: abbb−1 c = abec = ab c = abc. cab are words deﬁned by three elements a. . . . n is called a word. Note that an empty word is a unit. Thus we should regard that abbb−1 c and abc are same words under this rule.. . . such a (economical) rule that if a given word includes an empty word. gk .g. y = cab we have xy = abccab. . Deﬁnition 2. . then G is said to be ﬁnitely generated. . b−1 . Then. . . c.
. . . . = (1) which is called a presentation of G and a1 . For example. . . . an are called generators. . . . for any x ∈ Fn we employ the notation on k times product of x in the following. . gn . Then we introduce a rule such that if there are adjacent elements which belong to a same group.1 (3). g 1 h1 g 2 g 3 h2 h3 h4 =⇒ g1 h1 (g2 ◦ g3 )(h2 ◦ h3 ◦ h4 ). We call a word again an element which consists of a ﬁnite number of arbitrarily juxtaposed elements of G and H.. g 1 h1 g 2 g 3 h2 h3 h4 (gi ∈ G. . . hr . i. Then we consider a free group Fn = F (a1 . . an ) with rank n and deﬁne a map f : Fn → G. . Clearly. it is easy to see that the set of all words consisting of G and H forms a group. We denote this group by G ∗ H or H ∗ G. an  h1 = e. . f (Fn ) = G. an  since it has no relations. . . we have Fn = a1 . . .61 of Fn . . . xk = x · · · x. where we regard the latter as the word consisting of 4 elements. . . . eg. . . Under this rule. hr ) ∼ G. . k Let G and H be groups. denoted by Ker(f ) = (h1 . . hr = e . . . . and h1 . Suppose that the subgroup Ker(f ) is generated by h1 . f (a1 a−1 a3 ) = 2 −1 g1 g2 g3 . We usually express this like the following: G = a1 . For example. . Thus we have the isomorphic correspondence F (a1 . f (ai ) = gi such that f is a homomorphism. hr relations. but we usually write it in the following: Fn ∼ = a1 . . f is a surjection. . hr ). . which is called the free product of G and H. = which means that any ﬁnitely generated group G is isomorphic to a quotient group of a free group. by Exercise 2. As for a free group. . then we replace them by their products. . . . Note that the way of the presentation of a given G is not unique in general. an )/(h1 . . For simplicity. . we have Fn /Ker(f ) ∼ Im(f ) = G. . hi ∈ H) is a word consisting of 7 elements.e. . Then. . . . in the above word. an . . . Let G be a ﬁnitely generated group with n generators g1 .
. bijection}. . . n} be a ﬁnite set consisting of n elements. Z ⊕ Z ∼ a. . [1]. . . . σ ∈ Sn . bn ). Show that the following representation is true. . bn ) we have G ∗ H = F (a1 . which is a ﬁnite additive group of order 3. which is generated by a3 . . for any natural number n we have a presentation Zn ∼ a an = e . . m ∈ Z}.2. Any element σ ∈ Sn is often written as follows: σ= 1 2 ··· n . which is a group with a product deﬁned by composing maps. In general. b aba−1 b−1 = e . the product is τ ◦σ = Also. σ(1) σ(2) · · · σ(n) and then. for G = F (a1 . . . [2]}. Sn is called an nth symmetric group. am ) and = H = F (b1 . e = 1 2 ··· n 1 2 ··· n 1 2 ··· n . for τ. . = = In general. f (a) = [1]. Exercise 2. Let Xn = {1. b1 . etc. Let F (a) denote a free group with rank 1 which is canonically isomorphic to Z.62 Example 2. Hence we have Ker(f ) = {a3m . F (a.2. . . Thus we can verify that f (a2 ) = f (a) + f (a) = [1] + [1] = [2] f (a3 ) = f (a) + f (a) + f (a) = [1] + [1] + [1] = [0] f (a4 ) = [4] = [1]. Then we deﬁne a surjective homomorphism f : F (a) → Z3 . . . Therefore we have a presentation Z3 ∼ F (a)/(a3 ) ∼ a a3 = e . = Example 2. b) ∼ Z ∗ Z. . Then we deﬁne a set Sn = {σ : Xn → Xn . . am . . whose order is n!. . Let us ﬁnd the presentation of Z3 = {[0]. = Here let us recall the deﬁnition of a symmetric group which is closely related to a braid group. τ (σ(1)) τ (σ(2)) · · · τ (σ(n)) is a unit element of Sn . 2.3.
. .3. . σ3 = 321 123 . and it is not apparently abelian. C. . Note also that σ1 ◦σ1 = σ2 ◦σ2 = e. which is not abelian for n ≥ 3 since si si+1 = si+1 si . τ2 }. σ1 . . . it suﬃces to show that 2 2 3 ˜ F (a. . sn−1 s2 = 1 (i = 1. Thus. B. A = 0 0 1 . τ1 = 213 123 . where 100 100 010 E = 0 1 0 . s1 s2 s1 = s2 s1 s2 . . 001 010 001 . Then. Let us ﬁnd a presentation of S3 . b2 . σ} ∼ a a2 = e ∼ Z2 = = 12 if we deﬁne f : F (a) → S2 by f (a) = σ.63 For example. b  a2 = e. Thus we have another presentation S3 ∼ s1 . σ1 ◦ σ2 ◦ σ1 = σ2 ◦ σ1 ◦ σ2 = σ3 . where σ = and note that 21 σ ◦ σ = e. . n − 2) (i − j ≥ 2) s i sj = sj si . (ab)3 = e . σ2 = 132 123 . we have a presentation S3 ∼ a. τ2 = 231 123 . f (b2 ) = σ2 ◦ σ2 = e and f ((ab)3 ) = (σ1 ◦ σ2 )3 = (τ1 )3 = e. b . . s2  s2 = e. s2 = bab. F }. (ab) ) has order 6. where σ1 = 123 . which means that f is injective since S3 also has order 6. We deﬁne a surjective homomorphism f : F (a. 2 f (b) = σ2 . s2 = babbab = e. s2 = a2 = e. σ2 . b)/(a . b)/(a2 . σ2 ◦ σ1 = τ2 . n − 1) i si si+1 si = si+1 si si+1 (i = 1. 1 and s1 s2 s1 = ababa = (ab)3 b−1 = b. Moreover. . = Here we shall put s1 = a. Consider a set G = {E. Exercise 2. D. B = 1 0 0 . We will leave the details to the reader as an easy exercise. b2 = e. 312 (2) It is easy to check that σ1 ◦ σ2 = τ1 . . (ab)3 ) → S3 ˜ Hence in order to prove that f is an isomorphism. we immediately see that S2 = {e. σ3 . τ1 . Thus f induces a surjective homomorphism ˜ f : F (a. A. b) → S3 as follows: f (a) = σ1 . we have S3 = {e. we see that f (a ) = σ1 ◦ σ1 = e. Then. = 1 2 In general we have Sn = s1 . s2 s1 s2 = bababab = b(ab)3 = b. s2 = e. Thus we have seen that S3 is generated by σ1 and σ2 .
is called a link. (3) Show that G is isomorphic to S3 . AF 2 }. A2 . K = S 1 ∪ · · · ∪ S 1 . Now we recall the deﬁnition of braid group which plays an important role in knot theory. Pn = 2 n+1 n 1 . F. . later on.64 001 001 010 C = 0 1 0 . ϕ : K → R3 . . 0 . it has not been discovered yet. we say that ϕ1 is ambiently isotopic to ϕ2 if there exist homeomorphisms h : K → K and H : R3 → R3 such that H ◦ ϕ1 = ϕ2 ◦ h. F 3 = E and hence G = {A. . . An embedding of K into the 3dimensional Euclidean space R3 (or the 3sphere S 3 )..e. 1] which is a subset of R3 in the following: P1 = 1 1 . in particular. The ultimate purpose in knot theory is to ﬁnd the complete invariant in which the converse implication holds. the Jones polynomial. 1]×[0. i. (2) Show that A2 = B 2 = C 2 = E.e. K = S 1 . ϕ is a knot if K is connected. 1 . i. Given two embeddings ϕ1 . AD = C. 1]×[0. then λ(K1 ) = λ(K2 ) for any K1 . First we take 2n points in a cube C = [0. F = 0 0 1 . Suppose that a map λ : L → Λ is deﬁned for an appropriately calculable algebra Λ. ϕ2 : K → R3 . .. 100 010 100 (1) Show that G is a group by matrix multiplication. i.0 .e. . AF. F 2 = D. . Unfortunately. Pn = 2 n+1 1 1 . . . 3.. F 2 . Such a map λ that if K1 is ambiently isotopic to K2 . there is the following commutative diagram: K −−−→ −−− ϕ1 H h K ϕ2 R3 − − − → R3 −−− Let L be the set of all links. where S 1 is a circle. D = 1 0 0 . . Many invariants have been discovered and we will introduce one of them.1 2 n+1 n 1 . . Knots Deﬁned by Braids Let K be a disjoint union of circles. 2 n+1 P1 = . . K2 ∈ L is called an invariant of links.
Note that in general στ is not equal to τ σ. . 2. τ ∈ Bn we deﬁne a product of σ and τ as follows: First glue the base of the cube containing σ to the top face of the cube containing τ . if we can transform one to the other by performing the elementary knot moves on these arcs. i. . σi . If we parametrize the arc γ by (x(t). i + 1 by segments: It is shown that any nbraid can be represented as the product of the ± elements σi for i = 1. 1. . .. . . say σ1 . σn−1 are generators of Bn . P2 . Let e ∈ Bn be an nbraid obtained by joining the P1 . then γ intersects with z = t plane at one and only one point for any t ∈ [0. . Pn respectively by n segments (see Fig. n − 1. 1). We say that these n curves in C are an nbraid. As a resultant. . Pn by n arcs in C in such a way that these n arcs do not mutually intersect each other and each arc is not selfknotted. Given two nbraids in C. The braid group Bn can be represented by generators and deﬁning relations between the generators. Pn to P1 . P2 .e.65 Then we join the P1 . . which is called a trivial nbraid and a unit element in Bn . . . which connect Pi to Pi+1 and Pi+1 to Pi . . . . . στ = τ σ. 1]. . . P1 P2 P3 Pn1 Pn P´ P´ P´ 1 2 3 P´ P´ n1 n unit element of Bn Fig. and then connect the remaining Pj to Pj for j = i. t). By shrinking the rectangular solid in half we can obtain an nbraid which is the product of σ and τ . . We can make Bn a group under this product and call it the nth barid group. Let Bn be the set of all nbraids. . . then we say that these two nbraids are equal. . y(t). . . . P2 . In order to give a presentation we shall −1 introduce two speciﬁc nbraids. . P2 . Thus the elements σ1 . For σ. denoted by στ . we have a rectangular solid in which there are narcs composed by the vertical juxtaposition of σ and τ . Pn to P1 .
which just corresponds to forgetting the (under. . If K represents = = a trefoil knot in R3 (see Fig. B2 is generated by σ1 or σ1 . Thus we have a (natural) surjective homomorphism Ψ : Bn → Sn deﬁned by Ψ(σi ) = si for i = 1. . y. σn−1 σi σi+1 σi = σi+1 σi σi+1 (i = 1. . . . . . . σ2  σ1 σ2 σ1 = σ2 σ1 σ2 B4 = σ1 .66 P1 P2 Pi Pi+1 Pn P 1 P2 Pi Pi+1 Pn P´1 P´2 P´i P´i+1 P´n P´1 P´2 P´i P´i+1 1 P´n Fig. . A knot or a link constructed in this way is said to be a closed braid (or a . Pn on the top face of σ to the n points P1 . which is due to Artin (see eg. 2. . . . x2 + y 2 = 1. 3).or over) crossing in σi for any i. . Then we have obtained a regular diagram of a knot or link from the nbraid σ. Moreover. . 51 in PrasolovSossinsky8). . σ3 σ1 σ3 = σ3 σ1 . For example. . p. By means of a set of parallel n arcs which lie outside the cube C. we have B3 = σ1 . In general. . 2. .24 in Hatcher.3)torus knot.5 note that a trefoil knot is a (2. n − 1. then we have π1 (R3 − S) ∼ Z (∼ B2 ). σ2 . (See = Example 1. −1 For example. . we have the following presentation: Bn = σ1 . Thus B2 = σ1 . then we have π1 (R3 − K) ∼ B3 . σ1 σ2 σ1 = σ2 σ1 σ2 . n − 2) (i − j ≥ 2) σi σj = σj σi . . if we set S = {(x.) Let σ be any nbraid in Bn . which is isomorphic to a free group Z. Pn respectively on the bottom face. we can connect the n points P1 . Any element α ∈ B2 is −1 uniquely determined by the number of σ1 ’s minus the number of σ1 ’s in α. z) ∈ R3 . z = 0} which is a trivial knot in R3 . The braid group naturally appears in geometrical or topological context. σ3 σ2 σ3 = σ2 σ3 σ2 .
see eg.2 Let L be the set of ambient isotopy classes of M oriented links. We deﬁne two operations in the union of braid groups which are called Markov moves: (1) (Markov move I) For b ∈ Bn . we have a trivial knot as the closure of 2braid σ1 . For example. For example. (2) (Markov move II) For b ∈ Bn . we have a trefoil knot as a closure of 3braid σ1 −1 −1 and a ﬁgure eight knot as that of 4braid σ1 σ2 σ1 σ2 . 3. denoted by a ∼ b.2 claims that a map Ψ : B/∼ → L is a bijection. 3 closure of σ). Let a.1 (the Alexander theorem). Any knot or link can be obtained as a closed braid.67 trefoil K Fig. then a is said to be Mequivalent to b. 55 in PrasolovSossinsky. b ∈ B = n≥1 Bn . the 3braid −1 −1 σ1 σ2 and the 4braid σ1 σ2 σ3 . Let K1 and K2 be knots (or links) obtained as the closure of a1 and a2 respectively in B. The usefulness of this deﬁnition is based on the following fact due to Alexander: Theorem 3. Theorem 3.1. K1 is ambiently isotopic to K2 if and only if a1 ∼ a2 . If we can transform a into b by a ﬁnite number of performing the Markov moves I. II deﬁned above and their M inverses. Then. transform b into either of two (n + 1)braids bσn or bσ −1 in Bn+1 . Theorem 3. p. See the proof in Birman.8 Note that we may have such a situation that a knot or link is obtained by distinct braids. As for the proof.2 (the Markov theorem). M . Deﬁnition 3. transform b into xbx−1 for some x ∈ Bn .
For ∀b ∈ Bn we deﬁne ε(b) to be the number of σi ’s in b minus that of −1 σi ’s in b for any i. . from the relation σi σi+1 σi = σi+1 σi σi+1 we can deduce that σ1 = · · · = σn−1 . Moreover. we consider its n times tensor space V ⊗n = V ⊗ · · · ⊗ V and set n Ti = I ⊗ · · · ⊗ I ⊗ T ⊗ I ⊗ · · · ⊗ I ∈ GL(V ⊗n ). For a group G a homomorphism ϕ : G → GL(V ) is called a (linear) representation of G.68 which enables us to classify all knots or links by ﬁnding algebraic operations which should be invariant under the Markov moves I and II. Note ∼ that GL(C) = C−{0} and then the relation σi σj = σj σi is trivially satisﬁed because GL(C) is an abelian group. Ti Ti+1 Ti = Ti+1 Ti Ti+1 . . Let V = C2 and then element 1 0 T = 0 0 we consider the tensor space V ⊗ V . So. Thus we have obtained a one dimensional representation fz : Bn → GL(C) deﬁned by fz (b) = z ε(b) . we have a one dimensional unitary representation fz : Bn → U (1). GL(V ) is a group with a product as a composite of linear maps. i−1 n−i−1 where I stands for the identity map of V . T2 . Ti Tj = Tj Ti . i. the correspondence f (σi ) = z (i = 1. . . Example 3. Also. . Let V be a vector space over a ﬁeld. Let us ﬁnd a one dimensional representation of Bn . Hence. If we restrict to z ∈ U (1) = {z ∈ C. Then it is easy to verify that T1 . n − 1) gives a representation of Bn . we deﬁne ρn : Bn → GL(V ⊗n ) by ρn (σi ) = Ti for each generator σi ∈ Bn and then we easily see that ρn is a representation of Bn . Take an 0 0 0 −q −q 1 − q 2 0 0 0 0 ∈ GL(V ⊗ V ). . .. Let GL(V ) be the set of all linear maps of V into V which are regular. which is called a crossing number of b ∈ Bn . 0 1 where q ∈ C − {0}. for any z ∈ GL(C). .1. Next we shall discuss the representation of a group. V a representation space and the dimension of V the dimension of the representation. z = 1}. g ∈ GL(V ). Tn−1 satisfy the relations in Bn .e. . f · g = f ◦ g ∈ GL(V ) for f.
We can check that PK (q) is a Laurant 0 q n polynomial with a variable q and invariant under the Markov moves I and II. we have Tr(f (xbx−1 )) = Tr(f (x)f (b)f (x−1 )) = Tr(f (x)f (b)f (x)−1 ) = Tr(f (b)). Conversely. then knots or links express its process of computation in quantum computing by replacing a unitary transformation in quantum computing with a representation of the braid group Bn . GL(Vn ) is given for any n. we may regard that if the representation will be realized as a unitary transformation. Since f is a homomorphism. given a problem in BQP. any quantum gate can be approximated by a sequence of selfhomeomorphisms. Recently.e. a linear transformation will be assigned for an element of a group. i. which is an extension of a representation of the braid group. BQP is the class of decision problems which can be solved with probability being greater than or equal .. Topological Quantum Computing Roughly speaking.69 Suppose that a representation of the nth braid group fn : Bn → Bn → GL(Vn ). With a little more correction. PK (q) is called the Jones polynomial of K (see Jones6 ). a linear map which corresponds to a selfhomeomorphism of a disk with holes can be approximated by a sequence of quantum gates for an arbitrary unitary modular functor. More precisely. quantum computing is a unitary transformation in the process of computing (Nakahara7). For any link K we take an nbraid b. Recalling that a representation of a group is a homomorphism into general linear space.1. This induces a map f : n≥1 n≥1 For any b ∈ B = n≥1 Bn consider its trace Tr(f (b)) of a matrix f (b). 4. They have shown that a unitary modular functor is equivalent to quantum computation. Here. FreedmanKitaevLarsenWang4 introduced the notion of a ‘modular functor’. which means that the correspondence f (b) → Tr(f (b)) is invariant under the Markov move I. h = q −1 0 and h⊗n = h ⊗ · · · ⊗ h. we obtain a correspondence which is invariant both under the Markov moves I and II. Then we deﬁne PK (q) = q ε(b) Tr(ρn (b)h⊗n ). where ε(b) is the crossing number already deﬁned in Example 3.
A functor F from C to C is called a unitary topological modular functor if F satisﬁes the following seven axioms: (1) (Disjoint union axiom): Let X1 and X2 be compact oriented surfaces with boundaries whose labels are 1 and 2 respectively. let C = (X. ˜ F (X. F (A. where L is the length of the problem instance M . ∪ {x. . {a}) = C 0 if a = 1 if a = 1 (6) (Annulus axiom): Let A be an annulus. ) be a category and C be another category such that an object is a ﬁnite dimensional complex Hilbert space and a morphism is a unitary transformation. . ) = x∈L F (X. {a. . where P is a 2ˆ sphere with 3 holes. ˜ (2) (Gluing axiom): Let X be a compact oriented surface obtained by gluing together a pair of boundary circles with dual labels in X. 1) ⊗ F (X2 . (5) (Disk axiom): Let D be a 2disk. 1 2) = F (X1 . 2 ). a. a}) and F (P. ˆ (3) (Duality axiom): Let X ∗ stnds for X with the reversed orientation and labels applied ˆ .} with involution ˆ . We shall give a deﬁnition of a modular functor. F (D. x}). Then. {a}). where we mean that A∗ = t A. We call a a dual of a. F (X1 X2 . c. {a. where ˆ = 1. b.75 by an exact quantum circuit designed by a classical algorithm in poly(L). Let X be a compact oriented surface with boundaries. b}) = C 0 if a = ˆ b if a = ˆ b (7) (Algebraic axiom): All can be algebraically described over Q for some bases in F (D.70 to 0. Suppose that each boundary component of X has a label in a ﬁnite set L = {1. ¯ F (X ∗ ) = (F (X))∗ . b. {a. . (4) (Empty surface axiom): F (∅) = C. c}). We denote a label set of X for all boundary 1 ˆ components by . F (A.
then a linear function Z(Y ) ia a complex number by Axiom (4). This is because an old theory has been overtaken by a new theory through clarifying the limit of applicability by a new experimental fact or its selfcontradiction. Then we can construct Y = Y1 ∪Σ2 Y2 by gluing Y1 and Y2 along Σ2 . (2) (Disjoint union axiom): Z(Σ1 Σ2 ) = Z(Σ1 ) ⊗ Z(Σ2 ). a linear map Z(Y1 ) : Z(Σ1 ) → Z(Σ2 ) is deﬁned and Z(Y2 ) : Z(Σ2 ) → Z(Σ3 ) is also deﬁned when ∂Y2 = −Σ2 Σ3 . . Note that we have no unitary topological modular functor deﬁned on X with same labels all in the boundaries by the algebraic axiom (7). where Z(Σ)∗ is the dual space of Z(Σ). We also see that a modular functor (see Chapter V in Turaev9 ) deﬁned on X having only label 1’s all in the boundary components is nothing but a representation of the braid group. However. there are traditionally laws but no axioms.71 FreedmanLarsenWang3 has actually given an example of a unitary topological modular functor by using the SU (2) ChernSimonsWitten the2πi ory at q = e 5 . (5) (Homotopy invariance axiom): Z(Σ × [0. On one hand. The TQFT in dimension d for a ddimensional oriented compact smooth manifold Σ is a functor which assigns to a ﬁnite dimensional complex vector space Z(Σ) and a vector Z(Y ) ∈ Z(Σ) such that ∂Y = Σ. In physics. Z(Y ) = Z(Y2 ) ◦ Z(Y1 ). Note that if Y is a closed (d + 1)manifold. Atiyah1 has given the axioms in the following: (1) (Duality axiom): Z(−Σ) = Z(Σ)∗ . in mathematics axiomatization is the standard method. It will be helpful to recall an axiomatic deﬁnition of topological quantum ﬁeld theory (TQFT) in relation with modular functor. (4) (Empty manifold axiom): Z(∅) = C. which means that the TQFT has given a topological invariant of a closed (d + 1)manifold. (3) (Associativity axiom): When ∂Y1 = −Σ1 Σ2 . it may be useful to clarify the axioms since we have been recently faced with diﬃculty of experimental inspection. 1]) = idZ(Σ) .
Consider the probability amplitude ψ(x1 . That means. a phenomenon which electron systems conﬁned to a 2dimensional surface show at low temperature and strong magnetic ﬁeld. this deviation does not change according to the continuous change of closed paths. we look upon these normal conducting phase holes as a composite particle that are draped with charges and magnetic ﬂuxes as shown in Fig. By combining the duality axiom. Since these ﬂux penetrating holes on the surface of metal and the normal conducting electrons are correlated. Similarly. therefore the ﬂux penetrates into these holes of normal conducting phase. C). to be observed at point x1 and x2 . where V = Hom(V. it is a topological invariant. Also. W ) = ∗ (3) for two complex vector spaces. Z(Σ2 )). the magnetic ﬂux gets driven out in superconductive phase. The electrons conﬁned to a metal surface is in superconductive phase under low temperature. given the names of Anyons.72 Note also that a vector Z(Y ) can be regarded as a linear map. i. when an electron rotates around the anyon. which is called the AharonovBohm eﬀect. x2 ) of 2 anyons carrying charge q and magnetic ﬂux Φ. normal conducting phase appears as holes on the surface. 5. its phase of wave function also deviates by AharonovBohm eﬀect proportional to the charge. Around these holes. disjoint union axiom and (3). Here. First we recall the wellknown fact: V ∗ ⊗ W ∼ Hom(V. When the charged particle rotates around the tubelike magnetic ﬁeld generated by a solenoid. since anyon is a particle wearing charge and magnetic ﬂuxes. As is well known. for an oriented (d + 1) dimensional manifold Y with ∂Y = (−Σ1 ) Σ2 we have the following: Z(Y ) ∈ Z(∂Y ) = Z((−Σ1 ) ∗ Σ2 ) = Z(−Σ1 ) ⊗ Z(Σ2 ) ∼ = Z(Σ1 ) ⊗ Z(Σ2 ) = Hom(Z(Σ1 ). and applying strong magnetic ﬁeld vertically. its wave function ψ deviates by eiqΦ ψ.e. and the system exhibits a state where the ﬂux piercing these holes vertically. As .e. one may expect to treat these as composite particles. This isomorphism means that a vector Z(Y ) is naturally regarded as an element of Hom(Z(Σ1 ). a linear map of Z(Σ1 ) into Z(Σ2 )).. Z(Σ2 )). 4. Φ is the total magnetic ﬂux Φ = S BdS penetrating the tube. Anyon Model Anyon is a composite particle which is introduced to explain the phenomena of quantum hall eﬀect. the normal conducting electrons (holes) make cyclotron motions. i.
and therefore is stable against perturbation. Also the information weaved into the link is a topological value and therefore robust. it has only complex 1 dimensional representation by the wave function’s phase factor deviation (4). one could get a spacetime diagram just like braids. Therefore the eﬀect of interchanging the 1st and 2nd anyon would be described as. x2 ) are known as bosons or fermions. π in (4). So. sume that the 2nd anyon (at point x2 ) went around the ﬁrst anyon (at point x1 ).73 B ψ → eiqΦ ψ B Fig. Anyon could be thought as a topological soliton solution of the ﬁeld. while the simple abelian anyon as described above could construct braid group representation. θ = 2π. but the anyon system breaks timereversal symmetry. Let aside the existence of nonabelian anyon. it could be interpreted that the 1st particle has rotated around the 2nd anyon. 4. but since this 2nd particle is also an anyon. ψ(x2 . and can memorize the entanglement of spacetime curves. (4) The particle which its wave function obeys ψ(x2 . shown below is the mathematical model of topological quantum computing by employing nonabelian . However. These anyons which obeys such fractional statistics are permitted only on a special 2 dimensional space. The idea of topological quantum computing is to use the anyon system to carryout the quantum computation. x1 ) = ±ψ(x1 . the anyon obeys the fractional statistics. θ = qΦ. x2 ). A complete quantum calculation could not be handled by 1 dimensional unitary representation of braid group. When the motion of anyon particles are described on 2 + 1 dimensional spacetime. Which means. x1 ) = eiθ ψ(x1 . one could storage information in this topological links of spacetime curve. the total deviation should be the twice the case an ordinary charge going around a single anyon. Then the phase of 2nd anyon’s wave function on its ﬁrst circuit should deviate by qΦ. but as its name shows. Creating quantum algorithms could be achieved by braiding the strands.
consider 2 adjoint representation 8 of SU (3). and involution ˆ represents charge conjugation. c. We assume orthogonality and com . vacuum. The Latin letter a represents irreducible unitary representation. {ab. and its label to the types of anyon particles. the boundaries of X to each anyons. 5. ab Nc is a nonnegative integer. µ = Fig. . (2) (Fusion rule): The particle a and b becoming c by fusion is represented by a×b= c ab Nc c. a b µ c ab. b. . . 1 describes no particles. c. }. c. a. . . The antiparticle of a is labelled by a. .10–12 1) the emerging particles have types and labels 2) rules of fusion and splitting 3) braiding rules when two particles are exchanged 4) braiding rules when two particles are recoupled (1) (Labels): Each particle is labelled by sets: L = {1.74 anyon. and for the vacuum ˆ = 1 ˆ 1 is required. for instance. is called the full c Vc Hilbert space for anyon pair ab. µ .4. Actually it is a composition of unitary representations.3. and a is its dual representation. and could be made into a vector space by taking an orthonormal basis. 2. the compact oriented surface X corresponds to the metal plane. ab µ = 1. Nc }. The space of all possible processes of particle a and b to fusion and become a particle c is represented by Vcab and called a fusion space. ab The direct product of these vector spaces. then the rule deﬁnes that 88 ˆ its composition 8×8 = 1+8+8+10+10+27 should be taken as N8 = 2. . From the deﬁnition of modular ˆ functor given in the previous section.
The model which satisﬁes ab ≥ 2 is called the nonabelian anyon model. µ and it should be the representation of braid group Bn to satisfy the pentagon and hexagon relation (Sec. d.b and c fusing to become d. (3) (Braiding rule): The swap of the positions of two particles a and b induces an isomorphism mapping R : Vcab → Vcba . e . Rab.6). R is deﬁned by the unitary transformation ab. while those that c Nc does not is called the abelian anyon. c. c. 6. µ = Iab . an isomorphism of (5) as. Deﬁne the unitary transformation F . µ = µ ab ba.75 pleteness between these bases. d. c. µ ab. That is. and e and c to be d. Vdabc e Veab ⊗ Vdec e Vdae ⊗ Vebc . c. µ ⊗ bc. ν = eµν abc ae . c . c. ν (Fd )eµν ν e µ .µ ab Iab is an identity operator on c Vc . together with the recoupling map F . a b µ c b a µ R = µ Fig. µ (Rc )µ . F ab. and second b and c fuses and become e and then a and e fuse to be d. ab (Rc )µ µ c (4) (Recoupling rule): There are two natural decomposition for the topological Hilbert space Vdabc . µ = δc c δµ µ . µ ⊗ ec. Above relation shows that there are two possibilities to be considered. c. e. µ ab. ab. which express three anyons a. the ﬁrst a and b fuses and become e.
10. L = {1. b = b Fig. V1aa a a a a a a 1 Fig. Fibonacci Anyons In the previous section. . a 1 a a a a 1 . 7. . Consider topological Hilbert space Vba1 a2 ···an . . a × 1 = a. a1 a2 a1 a2 . This represents a fusion space where particles labeled a1 . we require fusion rules between these states. we introduced general anyon models. =< aa.76 a µ b c a b µ c F e ν = eµν abc (Fd )eµν ν eµ e ν d d Fig. a}. Writing the fusion space when the numbers of a parti(n) (3) (4) cles are n as V1aa···a = V1 . 9. 6. a × a = 1 + a.12 In the Fibonacci model there are only two labels. an merges and becomes particle labeled b. 1 × 1 = 1. 8. V1 = V1aaa =< aa. where a = a. Further. State 1 ˆ express vacuum state. therefore. and a a state where an anyon particle exist. Here. 1 >C . a ⊗ aa. . There is only one way of two a particles fusing to become one vacuum. a2 . 1 >C and V1 has two possibilities such that could be shown as. we consider more concrete and simple Fibonacci anyon models.
the dimension of V1 is given by. a ⊗ aa. C similarly we obtain. a ⊗ aa. a ⊗ aa. a ⊗ aa. aa. 1 . a ⊗ aa. a ⊗ aa. dim(V1 ) = fn−2 provided that fn is Fibonacci’s number. 1 ⊗ 1a. 1 ⊗ 1a. which gives this model the name Fibonacci anyons. a ⊗ aa. f1 = 1. C Generally. 1 . a ⊗ aa.77 V1 (4) = aa. aa. a ⊗ aa. 1 ⊗ 1a. a ⊗ aa. f0 = 1. named as pentagon relations. The recoupling operator F must satisfy the following diagrams as shown in Fig. fn+2 = fn+1 + fn . 1 . 10. Writing down these operations as an action to the base x x b c x y z w b y z d w x y z d w (n) F v y f z w v x y f e F z w v e F c F F v Fig. v . 1 (n) (5) V1 = . 10. aa. 1 . aa.
11.. c)w. e f d Further. e). b)z. e f d where we used the notation ((xy. e). v = de xyd bzw x(y(zw. v = xy. R v . Similarly representing this operation y c y z d z v x z x y c z x x b y z F y v b v x z R v y x F R d F v Fig. c)w. c)w. so the pentagon equation is . the F and braiding operator R should also satisfy the following hexagon relation as shown in Fig. v = def yzw xf xyz x(y(zw. b)z. 11. d). xyd bzw (Fv )b (Fv )c = e d f yzw xf xyz (Fe )f (Fv w )c (Fc )b . v (Fe )f (Fv w )c (Fc )b . xy(zw) bzw Fv ◦ Fv ((xy. d). c ⊗ cw.78 vector gives. b)z. b ⊗ bz. v etc. v (Fv )b (Fv )c e d yzw x(yz)w xyz = Fyzw ◦ Fv ◦ Fc ((xy.
(F1 )c = δa δd . so we can simulate a quantum computing with Fibonacci anyons by braiding and recoupling. v = = d xy xz yxz y(zx. the image of the representation of the braid group is dense in SU (fn−1 ). xy yzx xc xyz xz yxz (Fv )c (Rv )(Fv )b = (Rd )(Fv )b (Rb ). Similarly in V1 . 12. whose image is dense (n) in SU (2). R= . d aa Ra = −e 2πi 5 = . b)z.79 for base vectors gives. aa aaa c a R1 = e 5 . 12. F1 has a solution. F : V1 → V1 . for the topological Hilbert space V1 aa aa aaa 1aa V1aaaa . v (Rd )(Fv )b (Rb ). τ −τ 0 −e (4) a a a a a a 1 Fig. σ2 → F −1 RF. d √ c 1 √ c a aaa c a c 1 (Fa )c = −τ δa δd + τ δa δd + τ δ1 δd + τ δ1 δd . a a a a 1 a 1 = 0 = 1 The generators {σ1 . d therefore we derive. d). 4πi C2 . d). the braiding V1 (4) (4) and recoupling operator. becomes √ 4πi e 5 0 τ τ . the maps R = Ra . These matrices generate a representation of the B3 . F = √ 2πi 5 . written as a whole R. σ2 } of braid group B3 on three strands can be represented by σ1 → R. yzx x(yz) xyz Fv ◦ Rv ◦ Fv (xy. b)z. τ 2 + τ = 1. Especially. v (Fv )c (Rv )(Fv )b d c cd xz yxz xy = Rxz ◦ Fv ◦ Rxy (xy. v yzx xc xyz y(zx. . R1 and F = F1 . and using the basis vectors as shown in Fig. d c d c (4) the hexagon equation.
0 < y < π } is an open set of R2 . . let C[a. For example. Then we deﬁne b d(f. (2) if O1 . d) is a metric space. (2) d(x. y) = (x cos y.80 Appendix A. (Rn . . O) is called a topological space and O ∈ O an open set of X. X ∈ O. . of subsets in O. however. Then. {Oλ }λ∈Λ . O. a) < ε} for ε > 0 and a ∈ X. y = (y1 . O2 ∈ O. It is clear that both the empty set ∅ and the ambient set X are open sets and also closed sets by deﬁnition. z) for any x. y) = 0 if and only if x = y. then O1 ∩ O2 ∈ O. b]. in particular d(x. It is easy to check that (C[a. d) is a typical metric space if we deﬁne d(x. . We say that a map f : X1 → X2 is continuous if f −1 (V ) is an open of X1 for any open set V of X2 . such that (1) ∅ ∈ O. xn ). (X. We have another one. g ∈ C[a. . b] be the set of all real valued continuous functions deﬁned on the closed interval [a. it is easy to see that U = {(x. . x sin y) the image f (U ) is not an open set of R2 . y. x) for any x. X or (X. λ∈Λ Oλ ∈ O. We deﬁne a set Bε (a) = {x ∈ X. Fundamental group A set X (more precisely. Moreover. y) + d(y. b]. d) be a metric space. b]. The d is called an Euclidean metric. which is called an εneighborhood of a in X. y) ∈ R2 . y) ≥ 0 for any x. z) ≤ d(x. y ∈ X. y ∈ X. . and (3) for any family. . d)) is called a metric space if there is a real valued function d : X × X → R such that (1) d(x. g) = a f (x) − g(x)dx for any f. We say that B is a closed set if X − B is an open set of X. y) = (x1 − y1 )2 + · · · + (xn − yn )2 for x = (x1 . Let (X. and (3) d(x. y) = d(y. Let X1 and X2 be topological spaces. z ∈ X. 2 for a continuous map f : R2 → R2 deﬁned by f (x. A subset A is called an open set (of X) if it holds that Bε (a) ⊂ A for any a ∈ A and a wellchosen positive number ε. Let X be a set and suppose that there is a family of subsets of X. yn ) ∈ Rn . d(x. Note .
1] → X. 0) = l0 (t). Deﬁnition A. f · g. Let X be a space. p)/∼. p) is an equivalence class to which l belongs. For any l ∈ L(X. Let L(X. l1 : [0. l(0) = p. In what follows.1. l1 ∼ l2 =⇒ l0 ∼ l2 . For two loops l0 .. 1] → X such that f (0) = p. For any p. and for p. Let X be a topological space. q ∈ X if there is a continuous map f : [0. F (t.81 that the image of an open set in the source space is not necessarily open in the target space for a given continuous map. p). Then it is easy to see that (1) l0 ∼ l0 (2) l0 ∼ l1 =⇒ l1 ∼ l0 (3) l0 ∼ l1 . Thus we can consider a quotient space π1 (X. p) be the set of all loops with a basepoint p in X. then X is called arcwise connected and such a map is called a path in X. there is a continuous map F : [0. 1]). 1) = l1 (t) (∀t ∈ [0. i. p) := L(X. Let X be a space. then F is called a homotopy connecting l0 with l1 and we say that l0 is homotopic to l1 . which is called a homotopy class of l. 1] → X with a basepoint p. [l] ∈ π1 (X. Here we deﬁne l0 ∼ l1 if l0 is homotopic to l1 . f (1) = q. l(1) = p is called a loop in X with a base point p. a continuous map l : [0. This means that ∼ is an equivalence relation in L(X. 1] × [0. as follows: f · g(s) = f (2s) g(2s − 1) (0 ≤ s ≤ 1/2) (1/2 ≤ s ≤ 1).e. When there is a path l such that the starting point and the end point coincide. we always consider an arcwise connected topological space and so we refer it simply “a space” hereafter. p). . Then we deﬁne a product of paths. as is already seen above. The f · g is a new path in X connecting p with r. q. r ∈ X let f be a path connecting p with q and g a path connecting q with r. 1] → X which satisﬁes the following conditions: F (t.
. then l we have l ˜ l · l ∼ ˜· l . [β]. p). Thus we reach the following deﬁnition: [l] ◦ [l ] = [l · l ] for [l].e. [γ] ∈ π1 (X. choices of representatives. 2 ˜ It is easy to see that if for [l]. s) = (α · β) · γ(0). l ∈ [l ]. p). which is welldeﬁned. F (t. s) = α · (β · γ)(1). p) we deﬁne a product of l and l as follows: l·l = l(2t) l (2t − 1) (0 ≤ t ≤ 1 ) 2 ( 1 ≤ t ≤ 1).82 Our purpose is to introduce a group structure into this quotient set. First for l. For this purpose we have to deﬁne a product in π1 (X. 1] × [0. 1] F (t. [l ] ∈ π1 (X. l ∈ L(X. does not depend on the (0 ≤ t ≤ 1 ) 4 (1 ≤ t ≤ 1) 4 2 ( 1 ≤ t ≤ 1) 2 (0 ≤ t ≤ 1 ) 2 (1 ≤ t ≤ 3) 2 4 ( 3 ≤ t ≤ 1). p) we have α(4t) (α · β) · γ(t) = β(4t − 1) γ(2t − 1) and α(2t) α · (β · γ)(t) = β(4t − 2) γ(4t − 3) i. 0) = (α · β) · γ(t). [l ] ∈ π1 (X. 1) = α · (β · γ)(t) . p) we take ˜ ∈ [l]. 4 Then we can deﬁne a homotopy F : [0. 1] → X in the following: 4t (0 ≤ t ≤ s+1 ) α( s+1 ) 4 F (t. If we have done it and it is computable. s) = β(4t − s − 1) ( s+1 ≤ t ≤ s+2 ) . we can study its geometrical property of X through the group structure. 1] F (0. F (1. For [α]. and that ∀s ∈ [0. 4 4 4t−s−2 s+2 ( 4 ≤ t ≤ 1) γ( 2−s ) which means that ∀t ∈ [0.
.83 Hence we have (α · β) · γ ∼ α · (β · γ) and the associativity ([α] ◦ [β]) ◦ [γ] = [α] ◦ ([β] ◦ [γ]). q ∈ X. R2 − {(0. . . 0)} with n ≥ 3. Take a loop l : [0. for any [α] ∈ π1 (X. an ndimensional Euclidean space Rn . Then we easily see that for any [α] ∈ π1 (X. then f ◦ l1 is also homotopic to f ◦ l2 . p) is trivial. y0 ) ∈ [l] c If l f → [f ◦ l] we take a path connecting p with q. given a continuous map f : X → Y . y0 ). . For example. . Let e : [0. . and SO(n) with n = 1 are not simply connected. 0)}. q) for any p. a circle S 1 . where note that this isomorphism is not canonicalc . and hence we have veriﬁed that π1 (X. Thus [e] is the unit element of π1 (X. . then this isomorphism is deﬁned. then we have [α] ◦ [α ] = [α ] ◦ [α] = [e]. the isomorpohism does depend on the choice of the path. . x2 + · · · + x2 ≤ 1}. Then we write this correspondence as follows: f∗ : π1 (X. Let X and Y be spaces and f : X → Y be a continuous map such that f (x0 ) = y0 . 1]). p) has a group structure. Thus [α ] is the inverse element of [α]. x0 ) −→ π1 (Y. we can obtain a map by considering the correspondence [l] ∈ π1 (X. . . ∈ . p) [α] ◦ [e] = [e] ◦ [α] = [α]. Namely. Namely. . If a loop l1 in X is homotopic to l2 . p). then we say that X is simply connected. we may omit p and write π1 (X) for simplicity because there is an isomorphism π1 (X. 1] → X with a basepoint x0 and consider the composite map f ◦ l : [0. If we do not need to specify the basepoint p ∈ X. . p) if we deﬁne α (t) = α(1 − t). On one hand. p) → π1 (X. an ndimensional disk Dn = {(x1 . For a space X if π1 (X. 1] → X be a constant map deﬁned by e(t) = p (∀t ∈ [0. Rn − 1 n {(0. Moreover. xn ) ∈ Rn . and (n − 1)dimensional sphere ∂Dn = S n−1 with n ≥ 3 are all simply connected. 1] −→ X −→ Y. x0 ) to [f ◦ l] ∈ π1 (Y. which is also a loop in Y with a basepoint y0 .
84 We can easily check that f∗ is a homomorphism. This means that f∗ is an isomorphism (see Exercise 2. Further. since both f and f −1 are continuous. Furthermore. However. Let X and Y be spaces. it is extremely diﬃcult to show directly that no homeomorphism exists. let Z be a space and g : Y → Z be a continuous map. then we say that X is contractible. if X is homotopy equivalent to a point. we can replace “homeomorphic” by “homotopy equivalent” in Theorem A. Suppose that they are homeomorphic. the composite map X −→ Y −→ Z induces the homomorphism (g ◦ f )∗ : π1 (X.1 (5)). Let f : X → Y be a homeomorphism such that f (x0 ) = y0 . then π1 (X) is isomorphic to π1 (Y ). the theorem has a slight generalization. then X is not homeomorphic to Y . we have f∗ ◦ (f −1 )∗ = (f ◦ f −1 )∗ = (idY )∗ = idπ1 (Y ) (f −1 )∗ ◦ f∗ = (f −1 ◦ f )∗ = (idX )∗ = idπ1 (X) . We emphasize here that only the diﬀerence of algebraic structures of fundamental groups immediately provides homeomorphism criterion. If there are continuous maps f : X → Y and g : Y → X such that g ◦ f is homotopic to idX and f ◦ g is homotopic to idY . Then. then we say that X is homotopy equivalent to Y . Thus we have obtained Theorem A. Then. then by deﬁnition there is no homeomorphism between them. As is easily seen. z0 ) and we immediately see that (g ◦ f )∗ = g∗ ◦ f∗ .g. then they are homotopy equivalent but the converse implication does not hold in general. If X is homemorphic to Y .1. It is easy to show that R is not homeomorphic to R2 . if π1 (X) is not isomorphic to π1 (Y ). e. there is a homeomorphism f g . both R and R2 are contractible (and hence homotopy equivalent) but not homeomorphic. f∗ is called the induced homomorphism by f . Then. If X is not homeomorphic to Y . Let X and Y be arcwise connected topological spaces.1. Note that if X and Y are homeomorphic. x0 ) → π1 (Z. In other words. Moreover.
(b) ψ is a bijection. This can be deduced from the calculations that π1 (R − {x}) = {e} and π1 (R2 − {(0. 0)}.85 f : R → R2 . 0). (3) Show that R2 is not homotopy equivalent to R2 −{(0. which contradicts the fact that the map f R−{x} is a homeomorphism since the connectedness is topological invariant.e. Rm is homeomorphic to Rn if and only if m = n. (x0 . 0). there is x ∈ R such that f (x) = (0. (1) Let f : [0. 0)}) ∼ Z = (see Exercise 6. which can be proved by using the notion of “homology group” but we omit it here because we need more pages in order to discuss homology. Moreover. y0 ). x2 +y 2 = 1} and take a basepoint x0 = (1. Then ˜ there uniquely exists a map f : [0. y) ∈ R2 . y x2 + y2 .2. y) = x x2 + y2 . ˜ (2) Deﬁne a map ψ : π1 (S 1 .1 and Exercise 6. which is still a homeomorphism.. Then. 1] → S 1 be a continuous map such that f (0) = x0 . Exercise 6. Consider the restricted map f R−{x} : R − {x} → R2 − {(0. x0 ) × π1 (Y. 0)} by calculating their fundamental groups. log(x2 + y 2 ) is a diﬀeomorphism.2 below). = (2) Show that a map f : R2 − {(0. R − {x} is not connected but R2 − {(0. . In general. 1] → R such that ˜ f (0) = 0. x0 ) → Z by ψ([f ]) = f (1). Show that (a) ψ is a homomorphism. 0)} is connected. x0 ) ∼ Z according = to the following orders: First we set S 1 = {(x. ˜ ϕ ◦ f = f. Show that there is an isomorphism π1 (S 1 . y0 )) ∼ π1 (X. Since f is an injection. 0)} → S 1 × R deﬁned by f (x. i.1. consider a continuous map ϕ : R → S 1 deﬁned by ϕ(x) = (cos x. sin x). (1) We have a canonical isomorphism π1 (X × Y. Exercise 6. the invariance of dimension of an Euclidean space holds.
x0 ) ∼ = {e}. A1. π1 (X ∩ Y ). x0 ) ∗ π1 (Y. x0 )/( ) = D2 S1 X Fig. However. then we have π1 (X ∪ Y. x0 ). Thus we need to know better how to ﬁnd group in order to calculate fundamental group eﬀectively for a given topological space. Let be a homotopy class in π1 (X. . Van Kampen’s theorem tells us a relation among those fundamental groups of π1 (X ∪ Y ). In this section we further introduce Van Kampen’s theorem which is useful in calculation of fundamental group. fundamental group can be generally nonabelian by deﬁnition. x0 ) represented by a loop X ∩ D2 and ( ) denote the normal subgroup of π1 (X. we take a base point x0 ∈ X ∩Y .86 Van Kampen’s Theorem We would like to carry out its calculation of fundamental group for a topological space. Then we have π1 (X ∪ D2 . If π1 (X ∩Y. Let X and Y be arcwise connected topological spaces. Then. Y. Theorem A. x0 ) generated by . a topological space with a simple geometrical structure has its fundamental group with a simple group structure. x0 ) ∼ π1 (X. π1 (Y ). We take a basepoint x0 ∈ X ∩ D2 = ∂D2 = S 1 . X ∩Y . (1) For arcwise connected topological spaces X. π1 (X). As is easily expected. = (2) Let X be a connected polyhedron and X ∪ D2 be a space attaching the 2disk along the boundary circle.2 (Van Kampen). X ∩ Y be also. x0 ) ∼ π1 (X.
y) ∈ R2 . x0 ) ∼ Z ∗ Z. Take a generator α ∈ π1 (M ) represented by = a central circle of the band M (see the dotted line in Figure A3). Thus. Then. Hatcher5 ). Thus. Let us calculate the fundamental group of the real projective plane RP 2 = M ∪ D2 by using Theorem A. where M is the M¨ebius band. we see that π1 (S 1 × S 1 − Int D2 ) ∼ Z ∗ Z since S 1 × S 1 − Int D2 is homotopy = equivalent to S 1 ∨S 1 . x2 + y 2 < 1}. then we see that the boundary loop ∂D2 represents α−1 β −1 αβ. This '$ '$ x0 r &% &% Fig.87 We omit the proof (see e.2. β  α−1 β −1 αβ = e ∼ Z ⊕ Z. we have π1 (RP 2 ) ∼ Z/(α2 ) = α  α2 = e ∼ Z2 . Then. Let Int D2 = be a space homeomorphic to a set {(x.2 (1) we have π1 (S 1 ∨ S 1 . By using Theorem A. Take two circles and attach at one point x0 (see the ﬁgure below). We have another typical example.2. (1) Find the fundamental group of a closed orientable surface Σg with genus g ≥ 2. A bouquet S 1 ∨ S 1 space is called a bouquet of S 1 . we have π1 (S 1 × S 1 ) ∼ Z ∗ Z/(α−1 β −1 αβ) = α. If we take generators α and β in π1 (S 1 × S 1 − Int D2 ).2 (2) together with Exercise 2. . denoted by S 1 ∨ S 1 .3. = = since S 1 × S 1 = (S 1 × S 1 − Int D2 ) ∪ D2 . and hence π1 (M ) ∼ Z.2. by Theorem A. we immediately see that the boundary loop ∂D2 is α2 . First note that M is homotopy equivalent o to S 1 . We should emphasize here that it is more important to master how to use Van Kampen’s theorem than to understand its proof.g. A2. = = We will leave several handy exercises: Exercise 6. by Theorem A.
A polynomial invariant for knots via von Neumann algeras. of Math. 1994). Larsen and Z. (2) Show that the fundamental group of the Klein bottle is α. Jones. 227 (2002) 605–622. Math. links and mapping class groups. (3) Show that π1 (SO(3)) ∼ Z2 and π1 (U (2)) ∼ Z. Publ. August 2007. Knots. M. V. M. J. Atiyah. 2. M. = = References 1.iqi. B. Birman. G. Wang.caltech. Topological Quantum Field Theories. Overview on Quantum Computing. 12 (1985) 103–111. A. Freedman. Larsen and Z. Preskill. Press.e. M. http://www. α. Hatcher. H. A modular functor which is universal for quantum computation. 82 (Princeton Univ. β  α2 β 2 = e. Math. 7. M. Soc. 1974). de Gruyter Studies in Math. . Studies. β  α2 β 2 = e and its abelianization. Soc. Sossinsky. vol. vol. F. IHES 68 (1989) 175–186. V. Kitaev. quantph/0101025. Links. 3. F. Freedman. Turaev. Nakahara. Comm. A3. 9. Translations of Mathematical Monographs. Lectures in summer school at Kinki University. A. Braids and 3manifolds. Math. Lecture Notes for Physics 219: Quantum Computation. Press.edu/preskill/ph219. α−1 β −1 αβ = e is isomorphic to Z ⊕ Z2 .88 M bius band M Fig. Braids. 4.. Prasolov and A. Amer. 5.. Quantum invariants of knots and 3manifolds.. 8. Topological Quantum Computation. M. V. 2002). Ann. Wang. i. 154 (Amer. H. Algebraic Topology (Cambridge Univ. Phys. J. 1997). 6. V. Bull. 10. vol. Math. 18 (BerlinNew York. S.
Lomonaco Jr. Kauﬀman and S. . knot polynomials and anyonic topological quantum computation. Physics 318 (2005) 345–407. J. Ann. H. Rasetti. L.. Computing spin neworks. Knot Theory Ramiﬁcations 16 (2007) 267–332. qdeformed spin networks. Marzuoli and M.89 11. J. 12. A.
This page intentionally left blank .
information theory and group theory. Introduction Entanglement arises naturally from the mathematical formalism of quantum mechanics. The intention of this introduction is to take the student through the main concepts of entanglement theory as clearly as possible. entanglement witnesses. Separability. University of Tokyo. computer scientists and mathematicians not necessarily having a background in quantum mechanics. physics and information science amongst other disciplines. 1. The theory of entanglement developed over the past two decades shows rich mathematical structure. MARKHAM Department of Physics. H. We cover the main concepts of entanglement theory such as separability. entanglement has become seen as a resource for quantum information processing. with many subtleties (both from physics and a mathematics perspectives) and has drawn ideas and techniques from such diverse ﬁelds as thermodynamics. useful for absolutely secure key distribution and quantum computation. LOCC and entanglement measures. LOCC. Keywords: Entanglement Theory. Graduate School of Science. However I will . Tokyo 1130033. Japan We introduce the theory of entanglement intended for an audience of physicists. From a physical perspective this property is however pretty surprising and has led to many longstanding debates amongst physicists as to the validity and meaning of this formalism. It will not and cannot be a deﬁnitive guide to the subject (for a pretty good attempt at this please see the Horodeckis’ review1 ) and of course some of the topics covered will be inﬂuenced by my own ﬁeld of research to an extent.91 AN INTRODUCTION TO ENTANGLEMENT THEORY DAMIAN J. in its use in quantum information. Despite the vast progress made over the years there are still an exciting number of questions remaining in many directions. More recently. convex analysis. in the mathematical challenges it presents and in its fundamental role as a part of physics. Along the way we will see many interesting questions arise spanning mathematics.
we associate a Hilbert state H. Mathematically entanglement is simply a consequence of composite systems in quantum mechanics. to all degrees of freedom of physical objects. We will then move to the quantiﬁcation of entanglement in section 4. First we explain the general setting of LOCC. Quantum Mechanics and State Space In order to facilitate a speedy review of the formalism of quantum mechanics we will simply present the most general way to express states and evolution of quantum systems. We hope that this will also be more amenable to mathematicians and computer scientists to get the important points. a state is described by a density matrix ρ living in the operator space over the Hilbert space H of the system at hand. A particular system at any point in time is then described by a “state”. I hope that these notes will pass on some of my own enthusiasm for what I think is an incredibly exciting ﬁeld to work in. Mathematically. Moving to composite systems we will see entanglement emerge as a property of quantum mechanics in section 3. The set of density matrices are the set of . We will also brieﬂy discuss methods of measuring entanglement and some subtleties which arise in the multipartite case. statespace and how operations on states are represented mathematically. The state describes the physical properties of the system . without too much recourse to the physical motives behind them. what would happen if we were to measure the system. For example an electron is described using L2 for the position degree of freedom. 2. please see Ref. This will be our starting point in these lectures. From a physical perspective this route may seem a little topsy turvy.1. We will begin in section 2 by introducing quantum states. relations to other ﬁelds and further reading.92 endeavor to cover all the basics and point out interesting open questions when they arise. but hopefully this is countered by its simplicity and transparency. State space In quantum mechanics. and C2 for the spin degree of freedom. 2.that is. then move to how this is used to deﬁne diﬀerent measures of entanglement. 2. For a more detailed account describing many of the physical motivations for the axioms taken.
or ensemble. and the pi can be interpreted as probabilities since by deﬁnition we have 1 ≥ pi ≥ 1 and i pi = 1. or by randomly sampling the states {φj } according to qj . (1) That is. ei }. any operator acting over Hilbert space H and satisfying the above properties can be said to describe the state of a physical system. Thus.93 all operators satisfying ρ† = ρ I ≥ ρ ≥0 1 Tr[ρ] = 1. We refer to the space of all density operators as state space S(H). If the sum in (2) only consists of one element ρ = ψ ψ. of the pure states {pi .e. For example. we say it is a pure state (in most introductory undergraduate courses in quantum mechanics we only consider pure states). Thus one interpretation of the density matrix is of a classical mixture. A simple way to test if a given density matrix is pure or mixed is the fact that ρ is pure if and only if ρ2 = ρ (as can be easily checked). A state which is not pure is called mixed. All density matrices can be diagonalised to their eigenbasis ρ= i pi ei ei . (2) where the ei are the eigenstates (elements of the underlying Hilbert space H). take two decompositions of some state ρ= i pi ψi ψi  qj φj φj . convex combinations of density matrices are also density matrices. the ensemble interpretation is not unique . In such a case we often simply write the unique eigenstate ψ . The freedom in decomposition occurs again and again in entanglement theory and is the cause of some great diﬃculty when it comes to checking the existence of entanglement.physically this means that many diﬀerent physical situations may give rise to the same physical state. In fact any density matrix has many possible decompositions into convex combinations of pure states.i. as well as its quantiﬁcation. It can easily be checked that state space is convex . . j = (3) this could be created by randomly sampling the states {ψi } according to the probability distribution pi .
94 Note that one can in fact take a more abstract mathematical approach in terms of C* algebras. with α2 + β2 = 1. For now we skip the deﬁnition of complete positivity until section 2. and will instead give the most general way to express such maps. The Ei are known as the Kraus operators. Instead we take a general two level system known as a qubit (and leave the physics of implementation to someone else!). These fall into two classes of operations. (6) . From here on. Completely positive maps are extremely important in the theory of entanglement as we will see in later sections. Evolution All evolutions of a quantum system can be mathematically expressed as completely positive (CP) maps from one space of density matrices S1 to another S2 (in fact we will usually consider S1 = S2 ). unless stated otherwise we are considering ﬁnite dimensional systems. 1 }. ψ ) and the computational basis {0 . (5) (4) written in the eigenbasis (i. In quantum information we usually do not specify the physical system at hand. but for the situation covered in these lectures (and for most of quantum information) this reduces to the formalism described above. those that do not preserve the trace 1 property of density matrices (or rather those which require an additional normalisation element to the formalism).4. The trace of εm (ρ) may be m† m Ei ρEi .2. A general pure qubit is written ψ = α0 + β1 .e. and those that do completely positive trace preserving (CPTP) maps. where the above completely suﬃces. 2. who’s outcome m is known εm (ρ) = i m† m m such that i Ei Ei ≤ 1. The superscript m can be seen as indicating the classical information of some measurement outcome in the evolution process. CP maps not preserving the trace property can be thought of as resulting from an evolution where at some point a measurement has taken place. which can become important when considering inﬁnities. Its density matrix is given by ρ = ψ ψ = α2 αβ ∗ α∗ β β2 .
we must be a little bit careful as it may not always be the case. The full normalised version (necessary for the output to correspond to a state) is given by ρ→ εm (ρ) . (Usually in text books you will ﬁnd these introduced in the reverse order and in fact the idea of CP maps can be derived from axioms about measurements a An operator U is unitary iﬀ U † U = I . We can see unitary maps simply as a change of basis of the Hilbert space. It is possible to have CPTP mixing with many terms not via measurement as we will see in later sections . m. POVMs. projective measurement and observables We have seen above the most general evolution of a quantum state. the evolution is unitarya . But this seems a bit cheating as there is only one term. 2. When there is only one term in the summation (6). although the interpretation of m as a measurement outcome is very useful (and often physically accurate). which is why we say these maps are not trace preserving. Tr[εm (ρ)] (7) The class of CPTP maps can be understood as having resulted from an evolution where the results m are erased or forgotten. In general measurement does cause mixing (by adding extra terms).i = (8) m† m 1 with i. so all CPTP maps can of course be written in the form of equation (6). This is a clear example of a CPTP map which does not involve a measurement.via entanglement. but we must be careful to note that it is not the only way to generate mixing. Note that CPTP maps are obviously a subclass of CP maps. hence we must average over all m. by simple relabelling.m Ei Ei = I . An important property of unitary operations is that it is impossible to go from a pure state to a mixed state. 1 . Hence in some sense index m acts as a dummy index and. So in fact (6) is the most general form of evolution possible. which can be easily checked. and if the trace remains 1. Then we have ρ → ε(ρ) = m εm (ρ) m† m Ei ρEi . Let us now go back and extract the essential features of the measurement part.95 m† m less than one (if i Ei Ei < 1).3.
and indeed for any given POVM. Mm = (10) . If we look at the action of a unitary rotation of a state ρ → U ρU † . collectively known as the Positive Operator Valued Measure (POVM) which obey Mm = I 1 m I ≥ Mm ≥ 0 ∀m. so the actual physical process is not unique in general. 1 (9) where m represents an outcome of the measurement. Sometimes it is not necessary to know the full evolution of the system when considering a measurement. we see that the statistics of any measurement would be changed pm → Tr[U ρU † Mm ] = Tr[ρU † Mm U ] by the cyclic property of the trace operator. The measurement would evolve the system to an ensemble as ρ → {pm . This can be seen if we write the POVM in terms of general evolution operators. i εm (ρ)} and the m represents the outcomes as mentioned.96 e. i m† m Ei ρEi }. 2). for state estimation. and in turn any set of operators satisfying properties (9) represents the POVM of a physically valid measurement. and the probability of outcome m occurring pm is given by pm = Tr[ρMm ]. Ref. To any measurement on a state ρ we associate a set of operators {Mm }. Any measurement can be written as a POVM. there is no unique way to implement it in general. Relating back to the map (6) we see that the ensemble is written {pm . Thus we can consider a unitary rotation of a state ρ as simply a rotation of the basis in which we measure (this is the essence of the Heisenberg representation of quantum evolution). There is no mention of how the state actually evolves. This leads to the idea of the formalism of POVMs. Forgetting the outcomes would then be the same as taking the ensemble to the mixed state m pm εm (ρ). i m Clearly in general there are many diﬀerent Kraus operators Ei which would give the same Mm . say.g. m† m Ei Ei . Thus POVMs represent only the statistics of a measurement. we may just need to know the statistics of the outcomes of the measurements (rather than also knowing the resultant state as well).
Operationally this basically says that it is the only way to represent a reduced state such that .B X ⊗ I ] for all X (see e. the average outcome) is given by A ρ := m pm am = Tr[Aρ] (13) 2.4. Composite systems To combine more than one quantum system we employ the tensor product formalism. Ref. These are the normal representation of measurements in an undergraduate course on quantum mechanics.B ] := α (14) αB ρA. For a composite state ρA. Examples of such observables which are common in quantum information are the Pauli spin operators. These are measurements whose POVM operators Pm satisfy 2 Pm = Pm . The expectation value for a given state ρ (i. if we restrict ourselves to one subsystem. To complete the review of measurements we introduce observables. By comparing with (9)we see that in this case the Kraus operators are ﬁxed.g. and the Pm are projections onto the eigenspaces. This deﬁnition of reduced density matrix is equivalent to deﬁning it as the operator ρA which satisﬁes 1 Tr[ρA X] = Tr[ρA. the states are density matrices over HA ⊗ HB . (12) where the eigenvalues am represent the outcome of some experiment. For systems with associated Hilbert spaces HA and HB .B . 3). hence we have both the statistics and the outcome states of the measurement. HA . written in the eigenbasis A= m am Pm .B α B . (11) They can always be expanded as Pm = α α α for some orthogonal basis α . An observable is a Hermitian operator.B ]. where TrB is the partial trace and is deﬁned as TrB [ρA. (15) where {α B } form an orthonormal basis over HB . the reduced state is given by the ρA := TrB [ρA.97 An important class of POVMs that we will come across again are projective measurements.e.
98 measurements on one system faithfully can be seen as measurements on half of the global system. A B N (17) States that can be written in this form are called separable. Finally we now come to entanglement. a simple and elegant picture of general quantum operations emerges. That is. measuring of the ancilla system and tracing over the ancilla system.j are complex coeﬃcients satisfying i. We are now in a position to deﬁne complete positivity. (18) where {i A } and {j B } are orthonormal bases for HA and HB respectively. and states that cannot be written in this form are called entangled.j ai.8). the tensor product of which is often shortened to i A ⊗ j B ≡ ij AB . Mathematically it arises from the observation that not all states on a composite of N systems H = HA ⊗ HB ⊗ . performing unitary operations. but mathematically we can always use (6. let us look at the simple example of a bipartite state. This will allow us to introduce many terms in a simple way. the most general pure state in a system composed of two systems A and B is written as ψ AB ∈ HA ⊗ HB . It can be shown (see e..j ai..g. This concept is massively important in entanglement theory as we will see in later sections. Entanglement and Separability Now we have a deﬁnition of what entanglement is. . 2) that all quantum evolutions.j ai. ⊗ HN can be written in the form ρ= i pi ρi ⊗ ρi ⊗ . An important example of an operator that is positive. but not completely positive is the partial transpose which we will meet in section 3.. We will sometimes ﬁnd this picture of evolution helpful for descriptive reasons. A positive map Θ is completely positive iﬀ Θ⊗I ≥0 1 (16) for any ﬁnite dimensional identity operator I . which take the form (7) can be regarded as combinations of adding an ancillary system. 3. at ﬁrst considering pure states. ψ AB = i. Following our description of quantum states for the pure state case. As a side note.j i A ⊗ j B ≡ i. and ai. Ref.j ij AB . if and only if the 1 positivity remains with arbitrary extension.. with the introduction of composite systems.j = 1. ⊗ ρi .
(20) It can easily be seen that there is no way of decomposing this state into a product state as in (19) by contradiction.99 The deﬁnition of separability for pure states reduces to the set of states which can be written in the form ψ AB = a A ⊗ b B . The reduced state density matrices are ρ A = ρB = 1/2 0 0 1/2 . i AB . such that the state can be written r ψ AB = i=1 √ pi ii AB . (21) This is already in its eigenbasis. An extreme example of an entangled pure state in C2 ⊗ C2 is the following Φ+ AB 1 := √ (00 2 AB + 11 AB ). In fact. (22) This is known as the Schmidt decomposition. This state is a maximally entangled state of C2 ⊗ C2 b . we see how entanglement implies some local mixedness. 1 AB = d−1 √ 1 i. i=0 d . This decomposition captures all the features of the entanglement of the system. so is very important in entanglement theory. the coeﬃcients pi are the Schmidt coeﬃcients and the minimum r is known as the Schmidt number (known to be a good quantiﬁer of entanglement as we will also see in later sections). (19) for some states a ∈ HA and b ∈ HB . there exist local basis i A ∈ HA and i B ∈ HB . so it is clearly not a pure state. In the bipartite pure state case there is a very convenient way to write states (as far as their entanglement is concerned). Separable pure states are often referred to as product states. d ≤ d is given by Φ+ maximally mixed state of Cd is given by ρ = I /d. This is the maximally mixed state of C2 ⊗ C2 c . If we look at the reduced state of either subsystem. bA cA maximally entangled state of Cd ⊗Cd . For all states ψ AB ∈ HA ⊗ HB . the amount of mixedness is a good measure of how entangled a pure state is (as we will see in later sections).
1 (25) since ΛB (ρi ) = ρi is a valid density matrix. thus taking any separable state B B to another one just priming the right hand part of (24). but not completely positive map as d Note that positive. Now let us turn to mixed states. Our task in testing separability is to test if a state can be written in the form ρAB = i p i ρ i ⊗ ρi . 5. . Thus in the bipartite case. the state is entangled iﬀ the reduced states are mixed.d In the case of C2 ⊗C2 and C2 ⊗C3 . Consider a positive map acting on one part of the system ΛB . Clearly for all separable states we have (I ⊗ ΛB (ρAB )) ≥ 0. the question of separability becomes one of checking this condition over all positive maps which are not completely positive. In fact it has been shown4 that a bipartite state ρAB is separable iﬀ 1 for all positive maps Λ : S(HB ) → S(HA ) the condition (I ⊗ Λ(ρAB )) ≥ 0 holds. A B (24) For the bipartite mixed state case a useful tool can be found looking at the action of certain maps on the state.reduced states can be mixed even without the presence of entanglement. for pure bipartite states we can see that we have ρ A = ρB .e. the map of partial transposition is such a strong example of a positive positive. the Schmidt number is equal to one). In fact the completely positive maps which occurred earlier. See for example Ref. for a global pure bipartite state. However one approach to this has been to study approximate maps. The question then arrises: can such tests be carried out physically? Seemingly not. approximate well the nonCP maps above. which may not be completely positive. The ﬁrst thing we note is that the last observation above no longer holds . Thus.100 From the the symmetry of the Schmidt decomposition. which while CP. as we will see more in subsequent sections. (23) Also a pure state is entangled iﬀ the Schmidt decomposition has only one term (i. Indeed. but not completely positive maps are not actual physical processes. the pure bipartite state case is rather special in the ease with which we can analyse entanglement theory.
k.j. can be decomposed as Λ = ΛCP + ΛCP ◦ T. Broadly speaking there are two sets of results.k.k. AB This is known as the PPT condition for separability.j. For systems residing over C2 ⊗ C2 or C3 ⊗ C2 . those necessary conditions (such as PPT) which are implied by separability (but do not necessarily imply it). It is beyond the scope of these lectures to go into details and we refer the reader to the Horodeckis review1 for a good overview. for state ρAB = i. Let us now take a small step back and review what the set of separable . 3. all positive maps Λ : Cd → Cd .101 to cover all possibilities. When it comes to higher dimensions and larger number of systems.l ai. (27) and is dependant on the basis ij at hand. and in fact poses one of the largest open problems in the theory of entanglement. as mentioned. the problem of being able to check if a state is separable or not becomes extremely diﬃcult.j ij kl. (1) (2) (26) From this it follows that taking transposition over one system is suﬃcient to check for separability here. this is a diﬃcult task to check and we should need some clever mathematical tricks. In the bipartite case this can be seen as the problem of describing positive but not completely positive maps in a concise useful way (a recognised diﬃcult problem in mathematics). there are in general inﬁnitely many possible decompositions for a state. The task is to ﬁnd out if a given state can be written in a decomposition in the form of (17).l ij kl. The fact that it is not in general a suﬃcient condition was shown ﬁrst in Ref. Several approaches have been developed which work well for diﬀerent sets of states. which are generally diﬃcult to use or of limited application. initially presented by Peres6 as a necessary condition for separability and later used for the above theorem. 7 where an example of an entangled PPT state was presented. d = 2.l ai. d = 2. and those that are necessary and suﬃcient. the partial transposition over subsystem HB is given by ρTB = AB i. Then the above statement is written: A state ρAB ∈ C2 ⊗ C2 or ρAB ∈ C3 ⊗ C2 is separable iﬀ its partial transpose is positive ρTB > 0.k. Restated. In the multipartite case it is even more diﬃcult.j.l. Since.
Not only are these operators interesting from a mathematical point of view. Also..4. ⊗ ρi }. (28) A B N We can see that this set is also convex . The basis behind these witnesses is the HahnBanach theorem which states that given a convex body and a point xo that does not belong to it. Tr[W ρ] < 0 / (29) The Hahn Banach theorem then states that for all entangled ρ there exists a witness which can see it.10). via entanglement witnesses. We would then like to decompose these into sets of local measurements in as eﬃcient a way as possible. In fact these witnesses are related to the separation of separable and entangled states given by the positive operators mentioned above (see e. 1. while f (x0 ) < 0 (see Ref. they are also related to Bell’s inequalities . 10 for example). Entanglement witnesses are incredibly important in experimental veriﬁcation of entanglement.in fact Bell’s inequalities are examples of entanglement witnesses which also witness the violation of local realism. in situations concerning entanglement we are almost by deﬁnition separate. Stated in terms of density operators we deﬁne an entanglement witness W as Hermitian operators satisfying W ≥0 ∀ρSEP ∈ SEP.102 states looks like.9 These are operators which deﬁne hyperplanes separating some set of entangled states from the convex set of separable states. Thus their role in entanglement theory spreads to the experimental validation of entanglement too.. We deﬁne SEP := {ρSEP ∃ decomposition ρSEP = i pi ρi ⊗ ρi ⊗ . That is. The convexity of this set actually becomes incredibly useful and allows for a method of separating sets of entangled states from the separable ones.g. Ref. Tr[W ρSEP ] ≥ 0 ∃ρ ∈ SEP s. then there exists a linear function f that takes positive values for all points belonging to the convex body. An important question that arises there is ﬁnding the optimal decomposition into local observables.t. From a physics perspective. In certain . so we would like to measure these witnesses locally. they are also Hermitian and so represent physical observables. it can be shown that the volume of separable states is nonzero around the identity8 for any ﬁnite dimensional system.a probabilistic mixture of diﬀerent separable states is also separable.
. (30) This state is clearly entangled when δ = 1/2 and separable when δ = 0 or δ = 1. consider the state √ √ 1 − δ00 + δ11 . For example. 4. For a system on HA ⊗ HB ⊗ . how negative is the outcome) gives a bound to the ‘amount’ of entanglement present11–13 (a concept which we will expand on in the next section). A process which often irons out any quantum eﬀects.HN . This idea is made precise in the paradigm of local operations and classical communication. and how. This simple and elegant approach of entanglement witnesses to the study of entanglement is especially valuable when considering large complicated systems such as those in condensed matter. Quantiﬁcation of Entanglement In the last section we have seen the deﬁnition of the term entanglement. The situation is one of separate laboratories.14. Its symmetry indicates that δ = 1/2 may be somehow the most entangled. The practical usefulness of entanglement witnesses also extends to the quantiﬁcation of entanglement.103 cases optimality can be shown.. where each subsystem is acted upon separately. LOCC maps are those comprised . Entanglement is a property of quantum states which is inherently nonlocal. Local operations and classical communication Quantiﬁcation of entanglement is studied and deﬁned with respect to a special set of operations encapsulating the idea of what we mean by entanglement. and so what about the in between bits? Before we can really answer this question we must deﬁne better the problem setting. These are Local Operations and Classical Communication (LOCC). and has meaning when considering systems which are separated.15 where often we cannot even be sure of the state of the system exactly and typically measurements accessible are averaged over many subsystems (in the thermodynamic limit inﬁnitely many).14 4.1. Now we would like to quantify how entangled a state is. In certain cases the amount of violation of the entanglement witness (that is. Incredibly it is shown that even such ‘rough’ measurements as the magnetic susceptibility can act as witnesses of entanglement. We consider a system spread over several subsystems. and this is an important ongoing line of investigation. given a state we may decide whether it is entangled or not.
a a a . Thus we have an interpretation of separable states as those which can be created by LOCC and entangled states as those which cannot.16). Then. Bob could send Alice his result and she could make an operation according to that.that is correlations generated by classical communication (for a fuller discussion on the physical motivation for these conditions. We ﬁrst consider the bipartite case. belonging to Alice. the sum can go on for arbitrarily many terms.g. Taking the ensemble interpretation.. allowing for global “classical” eﬀects . and the other HB . Ref. see e. .Local CPTP maps on the individual subsystems .Classical Communication The role of the classical communication is to allow the generation of classical correlations.a2 ⊗ I )† . To see this we only need to look at the deﬁnition of a separable state (17). Given the possibility of going on and on with the correlations. Let us now try to write down the Kraus operators corresponding to an LOCC operation. by the experimentalist at system A randomly choosing a letter i according to probability distribution pi e . a a a 1 1 1 (EA1 ⊗ I )† (I ⊗ EB1 . The motivation behind this paradigm is essentially to test which properties of the system persist beyond local quantum eﬀects.104 of combinations of . Similarly states which can be generated by LOCC are separable. 1. We now slip into the standard notation of calling one system HA . and the map would be of the form ΛLOCC (ρ) = a1 . according to which i they all locally generate the state ρi . supposing he does a measurement too. then sending the choice to all the experimentalists at the remaining systems. She may perform some measurement and send the result to Bob.b1 )(EA1 ⊗ I )ρ 1 1 1 (31) (similar holds for the multipartite case). this can be generated for example.a2 ⊗ I )(I ⊗ EB1 .a2 . so then his action would depend on Alice’s outcome.b1 .b1 ..b1 )† (EA1 . e If this sounds too nonphysics we can always imagine the i as outputs of a quantum measurement.b1 . to Bob. Then if all experimentalists erase all memory of which i they x had.. the overall state is that described by (17). Without loss of generality we assume that Alice goes ﬁrst. It is apparent that all separable states can be made by LOCC. e... In this way the Kraus operators would become correlated.(EA1 .t.c.
entanglement theory proper started only then.19 On the other hand taking a state as in (30) with δ = 1/2 these tasks cannot be done as well. where we deﬁne an entanglement measure as some functional on the set of states which obeys certain axioms. So far in these lectures we have taken a rather abstract approach to entanglement theory. and an abstract approach.e. or PPT operations (those mapping PPT states to PPT states). in general we will see f Although of course entanglement has been an issue in physics since at least 1935. LOCC). For now however we will stick to considering LOCC.17 it also allows for perfect secrecy and eﬃciency in quantum key distribution. an operational approach. People became interested in entanglement in the early 90s when it was noticed that it could be useful for certain tasks f . perfect teleportation). where entanglement is deﬁned with respect to how useful this state would be to some quantum information task. Entanglement measures Now we have the situation clearly established (i. A kind of market driven start to a research ﬁeld.2. we can begin to ask what this setting would have to say about quantifying the amount of entanglement present in some given state ρ. such as separable operations (those mapping separable states to separable states). We can roughly split results into two main approaches to this question. This approach led to the ﬁrst attempts to quantify entanglement. This is not a particular feature of these tasks or this size of system. . 2 (32) This is the “most” useful state of two qubits in several respects: it (or local unitary equivalent states) gives the optimum ﬁdelity in teleportation (i. For this reason we often consider simpler sets of operations.105 Indeed the diﬃculty in writing LOCC operations in a concise way is one of the most diﬃcult obstacles to overcome in entanglement theory. but this is really not how the ﬁeld began. These operations have much simpler mathematical form. and so have allowed many results to be found where restricting to the stricter set of LOCC operations is too complicated.e. The archetypal state used in quantum information is the bipartite state which we already know as the maximally entangled state 1 Φ+ = √ (00 + 11 ).18. 4. Indeed the ability to do these tasks depends on the value of δ in a very precise way.
Using well developed ideas of asymptotics from information theory the entanglement quantities of distillable entanglement ED and entanglement cost EC were presented respectively. how many ME pairs does it take to generate a given state. With this in mind. ⊗ HN ) → R+ such that it also obeys: 1) Monotonicity under LOCC maps ΛLOCC E(ΛLOCC (ρ)) ≤ E(ρ). εLOCC (ρ)} where m is the (known) outcome of some measurement. is the trace norm (see e. E(ρSEP ) = 0 for all ρSEP ∈ SEP. for example the ﬁdelity (a measure of the faithfulness of the process) of teleportation itself can be considered as an entanglement measure. More precisely we ask that it decreases on average. (35) pm E (εm LOCC (ρ)) . For now however..106 that the entanglement properties are well parameterised by the Schmidt coeﬃcients of bipartite pure states. the idea of the usefulness of a quantum state was developed viewing the maximally entangled (ME) state as an ideal resource state. Ref. we deﬁne the distillable entanglement as the optimum rate for distilling maximally entangled states in the limit of inﬁnite number of copies. ED := sup{r : lim ( inf ΛLOCC (ρ⊗n ) − Φ+ ) = 0}. . and conversely. The question asked was. how many ME pairs can they get by LOCC.g. EC := sup{r : lim ( inf ρ⊗n − ΛLOCC (Φ+ )) = 0}. (36) . optimized over all possible LOCC protocols. That is for an E(ρ) ≥ m m where ΛLOCC (ρ) = m εLOCC (ρ) takes the state to ensemble m {pm . 2rn n→∞ ΛLOCC (33) where Φ+ is the ME state on Cd ⊗ Cd with dimension d = 2rn (which is 2rn equivalent to rn ME pairs of qubits). Axiomatically. 21) and ΛLOCC is an LOCC map.. We call any such functional obeying these properties an entanglement monotone. given several copies of a quantum state shared between Alice and Bob. Similarly the entanglement cost is deﬁned as the asymptotic rate going from ME pairs to copies of the state at hand. we say that any entanglement measure is a map E : S(HA ⊗ HB ⊗ . we will leave the operational approach and go back to axiomatic thinking. 2) Vanishing on separable states.20 More precisely. 2rn n→∞ ΛLOCC (34) There are other operational measures of entanglement.
it is possible to convert ψ → φ by LOCC perfectly (with unit probability) g Note that global basis change can of course generate (or degrade) entanglement.23 which states the following. Additivity and continuity are important when it comes to considering inﬁnite limits (of dimension and of number of copies in entanglement manipulation) and convexity sometimes added because of its mathematical convenience. This may seem obvious but sometimes it can cause confusion to physicists in trying to decide if a system can be said to be ‘entangled’ and there could be two or more basis physically accessible. but entanglement exists only in one. but this lecture course is too short to cover them all and rather we refer to larger reviews. If an LU operation U decreases entanglement. Of course we must be careful with which function to minimise over to check that it is an entanglement monotone. ω∈SEP (37) where D(ρω) = Tr[ρ(log2 ρ − log2 ω)] is the relative entropy. This of course makes sense since LU operations are just local changes in the basis and should not eﬀect global propertiesg .16 4. An example of a measure satisfying all of the above. e. often used in quantum information is the relative entropy of entanglement22 ER (ρ) = min D(ρω).3. Notice that monotonicity under LOCC implies that local unitary (LU) invariance of entanglement monotones.g. The simple answer is no (at least if we just talk about entanglement monotones). of additivity. order on states A question which naturally arises is. with ordered Schmidt coeﬃcients {ψi } and {φi }. Given two bipartite states ψ and φ . then U † . continuity and convexity.1. We must be careful in physics to state clearly which global basis we are talking about. We can see this since unitaries are reversible. which is also LU can increase entanglement causing a contradiction. This is an example of a socalled “distancelike” measure of entanglement. Uniqueness of measures. Of course there are many more entanglement measures in exsistence. This can be seen in the bipartite case by the seminal result of Nielsen. where we look for the closest separable state with respect to some distancelike function (in this case the relative entropy). .107 This list of axioms is somehow a minimum requirement of a good entanglement measure and some authors also include additional axioms. is there a unique measure of entanglement within this framework.
In fact. It can be shown24 that by including certain axioms. that is iﬀ for all k > 0 we have k k ψi ≤ i=1 i=1 φi . (40) However. it can be shown that for bipartite mixed states all such measures are bounded by ED (Eqn. (34)) (Ref. under these additional axioms. (33)) and EC (Eqn. . 24) ED ≤ E ≤ EC . for bipartite pure states at least. For example. we are in the end drawn back to questions of the operational value of our theory of entanglement. this problem can be smoothed over when considering asymptotics. corresponding to diﬀerent measures. we see that LOCC induces only a partial order on the set of bipartite pure states.108 iﬀ {ψi } are majorized by {φi }. (41) where ρA is the reduced state of subsystem A (or B equivalently). S(ρ) = −Tr[ρ log2 ρ] is the von Neumann entropy (which is a measure of the mixedness of a state) and pi are the Schmidt coeﬃcients of ψ AB . if there is a better way of deﬁning entanglement such that there is always a unique measure (not just for the bipartite pure state case). then. could this be possible by adding new axioms or changing those we have? However. what does that mean? May it be that diﬀerent types of entanglement are useful for diﬀerent quantum information tasks? Or could they hold some diﬀerent signiﬁcance in diﬀerent areas of physics? Interestingly we see then that even in our attempt to ﬁnd a clean axiomatic mathematical approach. Thus there exist pairs of states such that ψ LOCC φ . that for pure states all measures coincide to the entropy of entanglement20 Eent (ψ AB ) := S(ρA ) = − i pi log2 pi . So it seems for now at least that we are stuck with having to accept that there are diﬀerent types of entanglement. (38) Since majorisation is a partial order. (39) It is therefore possible to deﬁne entanglement monotones such that E1 (ψ ) > E1 (φ ) E2 (ψ ) < E2 (φ ). So we can ask. such approaches have been studied a lot in the literature and so far no good answer has been found. (42) One may ask.
Measuring entanglement Coming now back to physics for a bit. this gives the bound (ω1 . As mentioned.4. 25 and developed in Ref.. and has been calculated for several witnesses and entanglement measures and further applied to experiments. ωn ) := inf{E(ρ)Tr[Wk ρ] = ωk ∀k}. From this we can use the mathematical tools developed so far to answer questions of separability and amount of entanglement. we would like to be able to know what we can say about viewing the presence of entanglement in actual experiments. 5 (amongst others) an ingenious general approach is given to approximating experimentally any nonlinear function on states by making global measurements on several copies. It is no coincidence that most of the clean results we have presented are for the bipartite case (if not just the bipartite pure). It gets even worse for the multipartite case of three or more systems. However this approach is experimentally very costly and we would prefer a more eﬃcient strategy.26 The main idea is. The most direct approach to this is total state tomography. not only can entanglement witnesses be used to separate entangled states from the set of separable states.5. 4. but also in certain cases their violation can be used to bound entanglement11. Multipartite entanglement We ﬁnish our review of entanglement with a discussion about the multipartite case. Such techniques can be used to check separability and approximate the value of entanglement measures. An even weaker (but still incredibly powerful) approach we have also discussed before is that of entanglement witnesses..26 and as an entanglement measure itself. (43) This can be phrased in terms of Legendre transformations. This is where we repeatedly prepare and measure a state to build up statistics up to the point when we ﬁnally get all elements of the density matrix. One approach we heard about before in section 3 concerning treating nonCPTP maps experimentally. In Ref.109 4. For some chosen entanglement measure E. given the expectation value of a set of witnesses {Wk } we want to ﬁnd the worst case corresponding value for entanglement. .12.. We have seen the diﬃculty in deciphering the structure of statespaces induced by entanglement in the bipartite case. . amazingly allowing some old results for the calculation of a witness to give bounds on the value of entanglement.
However many questions still remain as to how to deﬁne well the notion of ‘likely’.27 For four or more there are possibly even inﬁnitely many diﬀerent classes of states under SLOCC. but we would really like some physical reason to exclude certain states (initial answers are proposed in Ref. it is completely unclear where the entanglement arises in these kinds of measures. even though the entanglement is clearly diﬀerent. (44) has the same relative entropy of entanglement as Φ ABC 1 = √ (00 2 AB + 11 AB ) ⊗ φ C. this approach does allow some freedom in choosing which ‘type’ of entanglement is considered. One approach is to consider which kinds of states are ‘normal’. The simplicity of the deﬁnition skips much of the detailed structure of the state space.g. . That is. Although this class of measures misses out on much of the richness of the structure of state space under LOCC. even if some states are ‘rare’ physically. Ref.we have inequivalent states for three or more parties. we have ER Bi(AB:C) ER Φ = 0. As stated in (37). Hence by analysing diﬀerent cuts we can get some structure from entanglement also. 28). For example a GHZ state GHZ ABC 1 = √ (000 2 ABC + 111 ABC ). This can lead to a much simpler induced structure of state space (e. we can discriminate these states. which kinds of states are likely to occur in any given situation.110 Even under a simpler set of operations of stochastic LOCC (SLOCC) LOCC allowing for non unit probability of success (CP LOCC maps with known outcome) . We are then forced to take steps to simplify the situation. Indeed. Another approach that bares fruit is to take simple measures of entanglement such as distancelike measures. Typically we take the approach of random sampling. it may be that they are so useful to us that we would like to still use them in quantum information. Thus we have many diﬀerent ‘types’ or entanglement. By choosing ER with Bi(AB:C) respect to diﬀerent cuts. Calling ER the relative entropy of entanglement with respect to the set of separable Bi(AB:C) GHZ = 1 and states across the bipartite cut AB : C. with ER (GHZ ) = ER (Φ ) = 1. 29). in some cases (including many states useful in quantum information) symmetries can be used to calculate entanglement and entanglement properties. (45) where φ C is an arbitrary single qubit state. Although the deﬁning optimization makes calculation diﬃcult.
such as continuity. There are a plethora of results relating the ability to generate/manipulate entanglement to the quality of a quantum channel as an information carrier. Through connections to communication theory mentioned above. when going to inﬁnities. 5. Having said that we have seen interesting issues raised about the physical measurement of entanglement and mentioned a few possible approaches. we have only skimmed the surface of this fertile ﬁeld. • Additivity.30 • Connection to quantum channels. In all of the results here we have only considered ﬁnite sized systems. and so to ﬁnish. This is particularly successful in the multipartite case since other approaches are so diﬃcult. or just over some subsystems. • Inﬁnite dimensional systems. it is perhaps ﬁtting to list a few of the important topics we have not covered in these notes (where direct references are not given. Indeed entanglement may not be continuous in state space for this case (though some continuity can be regained by certain ‘physical’ restrictions such as ﬁnite energy). and left the physical motivations to further reading.111 Of course. This approach has ﬁelded many useful results in both directions. 1). There . It can also be adapted to consider different types of entanglement. there are many more subtleties that must be addressed. Again by being careful in specifying which separable set we compare a given state to. We have only brieﬂy touched upon the usefulness of entanglement as a resource here. Of course. That is given several copies of a state. • Use of entanglement in quantum information. including bounds on how good channels are in terms of entanglement and new information based entanglement measures such as the squashed entanglement. the question of additivity of the entanglement of formation is shown to have consequences in the ability for additivity of communication quantities such as the Holevo bound. Conclusions In these notes we have attempted to present the basics of entanglement theory. through equivalence of problems it has yielded one of the largest remaining problems of entanglement theory (next point). We have taken a rather mathematical approach. how does entanglement scale (roughly speaking). we can witness whether a state is entangled over all systems.31 Also. references can be found in Ref. As hinted at. the simplest approach is that of entanglement witnesses mentioned before.
Lett. In quantum communication our understanding of the role of entanglement is pretty restricted to the bipartite case.112 are in fact many open questions remaining about its use in quantum information. A 58 (1998) 883. L. . Horodecki. Carteret. Lewenstein. Horodecki and R. P. Phys. Horodecki and K. Chuang. Secondly there are often many more constraints in a physical system as to the allowed set of operations and states. R. 4. This question is developing into one of the most active areas of current entanglement research. A 271 (2000) 319. P. Nielsen and I. could it be that the subtleties of multipartite entanglement have much more to oﬀer. Quantum Information An Introduction (SpringerVerlag. Horodecki. Lett. M. ˙ 10. A. the choice of how we split our space is a physical one. as mentioned brieﬂy. P. Many interesting questions arise from both physical and mathematical points of view as to the relationship between entanglement theory and the theory of criticality in many body physics. Horodecki. ˙ 8. Phys. 5. H. Zyczkowski. I. M. 6. Lett. we have recent progress in measurement based quantum computing32 ). Geometry of Quantum States (Cambridge University Press. Sanpera and M. 3. Bengtsson and K. Zyczowski. A. Hayashi. A. References 1. A 223 (1996) 1. 2. Horodecki. Rev. 33 for a recent review. perhaps diﬀerent classes of entanglement useful for diﬀerent tasks? In quantum computation we are only making the ﬁrst steps to understand exactly the role and requirement of entanglement (in particular. In these notes we have taken largely an abstract approach to the whole subject. A 128 (1988) 19. Lett. and so the deﬁnition of LOCC must depend on the physics at hand. for example those induced by symmetries. In real physical systems there are other considerations that must be taken into account. M. Phys. Terhal. Horodecki. Ref. Horodecki. 2006). • The physics of entanglement. 7. quantph/0702225. K. Horodecki. Rev. Phys. • Role of entanglement in manybody physics. particularly in the multipartite case. Quantum Computation and Quantum Information (Cambridge University Press. M. 2006). Firstly.g. B. P. A 232 (1997) 333. 9. Lett. See e. M. in any system. Peres. 94 (2005) 040502. 2000). Phys. It is another ongoing ﬁeld of research to try to impose these extra fundamental physical laws to our theory of entanglement. Phys.
Fazio. M. ˇ C. 19. 22. Brand˜o. C. H. Rev. Comm. New J. in Proc. Horodecki and L. Systems. Brassard. Phys. Vedral quantph/0406040. Rev. Phys. Phys. New J. G. 15. A 71 (2005) 010301. M. Eisert. J. Audenaert and M. 32. Winter. Vedral. Werner. Ekert. 17. A. 28. Popescu and B. Phys. Dahlsten. 8 (2006) 266. a W. R. E. Math. 83 (1999) 436. 9 (2007) u 204. 16. 13. 1984). G. u P. Hayden. Alves. H. K. J. 45 (2004) 829. M. I. J. F. B. M. Winter. 40 (2007) 8081. Crepeau. A. C. Plenio and S. Miyake and H. Oi. Plenio. Rudolph. B. Phys. Christandl and A. A 57 (1998) 1619. O. 25. J. Lett. P. Virmani. Comp. D¨r. 43 (2002) 4252. Phys. quantph/0703044. Cirac. K. o M. C. M. New York. M. Rev. Phys. A 72 (2005) 022310. Inf. Rev. Donald. 88 (2002) 217901. 30. J. Plenio. Bennett and G. Phys. . Plenio. Lett 70 (1993) 1895. A. 26. Amico. 27. Brukner and V. G. Bangalore. K. Phys. A 53 (1996) 2046. F. Plenio. M. Brassard (1984). Rev. Lett. H. Math. M. Bernstein. Kwek. Vedral and M. Jozsa. and Signal Processing. A. R. Phys. Wootters. C. Phys. A. Phys. Rev. Peres and W. Phys. L. 21. D¨r. Oliveira and M. Simon and M. G. Phys. D. V. W. M. B. Rev. Quant. u J. 14. Phys. Bennett. 98 (2007) 110502. Reimpell and R. C. New J. J. L. Phys. 24. Phys. Math. C. 33. Rains.113 11. Theo. 20. S. J. Lett. Rev. H. R. Osterloh and V. 9 (2007) 46. Schumacher. 67 (1991) 661. Rev. India (IEEE. L. 12. Briegel. Eisert. T´th. Rev. Audenaert. 265 (2005) 95. Phys. A 60 (1999) 173. Horodecki. Vidal and J. Lett. Rev. 18. 29. 31. D. B. 7 (2007) 1. Bennett. Horodecki and O. den Nest. V. Leung and A. Nielsen. A 35 (2002) 3911. IEEE International Conference on Computers. 23. Ekert. S. A 62 (2000) 062314. B. R. A. A: Math. G¨hne. O. Phys. ao and K.
This page intentionally left blank .
Keywords: Quantum Computation. I will brieﬂy discuss the remaining problems. Introduction Many designs of practical quantum computers have been proposed in this decade.ac. We also discuss possible implementations of the holonomic quantum computer. 1. I will formulate the optimization problem of holonomic quantum computer and show its solution. A few years ago Hayashi. Finally. Holonomic quantum computer is one of those candidates and has been studied mostly in the theoretical aspect. holonomic computation should be executed in a short time for avoiding decoherence. Graduate School of Informatics.1. I will apply the result to the two wellknown unitary gates. Holonomy. Kyoto University.115 HOLONOMIC QUANTUM COMPUTING AND ITS OPTIMIZATION SHOGO TANIMURA Department of Applied Mathematics and Physics. Most of this lecture note depends on our .2 In this lecture I aim to explain the basic of holonomic quantum computation. which is called the isoholonomic problem.kyotou. the slow process and the quick execution. Horizontal Extremal Curve. Berry Phase. Holonomic computation should be processed slowly for keeping adiabaticity.jp Holonomic quantum computer uses a nonAbelian adiabatic phase associated with a controlparameter loop as a unitary gate. which concern implementation of holonomic quantum computation in physical systems. On the other hand. Optimization. Isoholonomic Problem. the Hadamard gate and the discrete Fourier transformation gate. Japan Email: tanimura@i. Nakahara and I obtained an exact result about the optimal control problem of the holonomic quantum computer. Kyoto 6068501. I will give a quick review on Berry phase and WilczekZee holonomy and introduce the idea of holonomic quantum computation. seem contradictory each other and lead to a kind of optimization problem. These two requests. In this lecture we review the formulation of the holonomy in quantum mechanics and the solution of the optimal control problem of the holonomy.
1. The whole of possible values of λ forms a parameter space M . 2. λ(t) . The state ψ(t) evolves obeying the Schr¨dinger equao tion i d ψ(t) = H(λ(t)) ψ(t) . . Holonomy in Riemannian geometry The ﬁrst example of holonomy can be found in geometry. (3) .1. . Then the adiabatic theorem5. The concept of parallel transportation has been generalized in various contexts and described in terms of the connection or the gauge ﬁeld as discussed below. These parameters can be controlled externally and are speciﬁed as functions of time t. In this case.4 When a tangent vector is parallel transported along a closed loop in a Riemannian manifold. λ = εn (λ) n. The slowness will be deﬁned at (7) in more rigorous words. let us assume that we change the control parameter λ(t) as slowly as possible. it is assumed that εn (λ(t)) = εn (λ(t)) for any n (= n) and any t.7 Comprehensive reviews on this subject have been published. Assume that no degeneracy occurs along the course of the parameter λ(t).6 tells that the state ψ(t) remains an eigenstate of the instantaneous Hamiltonian H(λ(t)) at each time t: ψ(t) = eiγ(t) n.6 in quantum mechanics and the formulation of Berry phase. The net rotation generated by the parallel transportation is called the holonomy. dt (2) Suppose that at t = 0 the initial state is an eigenstate of the initial Hamiltonian as ψ(0) = n. λ(0) . . λ (1) for each value of λ. the vector gets a net rotation. Furthermore.3. . Suppose that the Hamiltonian has a discrete spectrum H(λ) n.8–10 Let us consider a quantum system whose Hamiltonian H(λ) depends on a set of parameters λ = (λ1 . Berry phase in quantum mechanics We shall review the adiabatic theorem5. Holonomies in Mathematics and Physics 2.116 papers.2. In practice λ might be an electric ﬁeld or a magnetic ﬁeld applied on the system. λm ). the holonomy manifests curvature of the manifold. Namely.2 2.
The phase γ is given as a sum γ = γd + γg with γd = − γg = i 0 (4) 1 0 T T εn (λ(t)) dt. λ C µ=1 ∂ n. λ(t) d n.117 In other words. 2. WilczekZee holonomy in quantum mechanics Immediately after Berry formulated his celebrated phase. α . if the external parameter changes slowly enough. When two distinct energy levels approach closely or cross. Wilczek and Zee11 formulated its generalization to an arbitrary unitary group U (k). α. . This inequality is called the adiabaticity condition. then transition between diﬀerent energy levels does not occur and the quantum state remains quasistationary in some sense. Then the state returns to the initial state up to a phase factor: ψ(T ) = eiγ(T ) ψ(0) . Suppose that the degrees of degeneracy does not change along the course of . Suﬃcient slowness of movement of the control parameter is characterized by the condition {εn (λ(t)) − εn (λ(t))}T 2π (7) for n = n during 0 ≤ t ≤ T . . which belongs to the onedimensional unitary group U (1). the adiabaticity condition is violated. and if no level crossing occurs. λ = δnn δαα . The phase γd is called the dynamical phase. Suppose that the Hamiltonian has a discrete spectrum but now degenerate eigenvalues: H(λ) n. . It is assumed that these eigenvectors are normalized as n. λ dλµ . Here we shall describe their formulation. α. It is said that the energy eigenvalue εn (λ) is kn fold degenerate. Suppose that at a time T the parameter λ(t) returns to the initial value as λ(T ) = λ(0). n. (8) Here α = 1. 2. kn is an index to label degenerate energy eigenstates. α.3. λ . The phase γg is called the geometric phase or the Berry phase. ∂λµ (6) A proof of these formulae will be given around (16) in a more general context. λ(t) dt = i dt m (5) n. λn . λ = εn (λ) n. It is independent of the timeparametrization of the curve C in the control parameter space. .
λ ∂ α. λ ∂ α. Aβγ.µ (λ(t))Γγα (t) dt dt γ=1 µ=1 k m Γβα (0) = δβα . λ(t) d α. β. A is called the WilczekZee connection. By substituting (9) into (2). (9) Since the quantum number n is unchanged. dt (12) The symbol T indicates the timeordered product. we omit it in the following equations.µ (λ) dλµ = µ=1 β. λ(0) (14) where P stands for the pathordered product. λ(t) dt dt γ=1 whose formal solution is given as cβα (t) = exp − i 0 t t 0 k (10) ε(λ(s))ds T exp − A(s)ds βα (11) with the matrixvalued function Aβα (t) = β. . We introduce a u(k)valued oneforma βα m m Aβα (λ) = µ=1 Aβα. It is easily veriﬁed that A∗ = −Aαβ . we get d d i cβα (t) + γ.118 the controlled parameter λ(t). It is deﬁned by the set of diﬀerential equations n dλµ d Γβα (t) = − . which is the set of kdimensional skewHermite matrices. λ ∂λµ dλµ . β. Then the unitary matrix appearing in (11) is rewritten as Γ(t) = P exp − λ(t) A . ψ(0) = n. λ(t) dt m = µ=1 β. λ dλµ ∂λµ (13) in the parameter space. (15) a We denote the Lie algebra of the Lie group U (k) by u(k). λ(0) . λ(t) . λ(t) cγα (t) = − ε(γ(t))cβα (t). If the initial state is one of energy eigenstates. then it remains a linear combination of the instantaneous energy eigenstates with the same quantum number n: kn ψ(t) = β=1 cβα (t) n. Then the adiabatic theorem again tells that transition between diﬀerent energy levels occurs with negligible probability. α.
119 We assume that the control parameter λ(t) comes back to the initial point λ(T ) = λ(0). +. (18) Bz cos θ The Hamiltonian of the spin is H = Bx σx + By σy + Bz σz = Bx − iBy Bz Bx + iBy −Bz =B cos θ e−iφ sin θ . H −.7 Let us consider a system of spin one half which is put in a magnetic ﬁeld Bx sin θ cos φ B = By = B sin θ sin φ .4.4.1. However. θ. Examples Here we shall demonstrate calculations of the Berry phase and the WilczekZee holonomy. B = e−iφ/2 cos(θ/2) . Hence. B . (16) A ∈ U (k) (17) is called the holonomy matrix associated with the loop λ(t). φ). B = −. 2. B = B +. B . we will observe . λ(0) Γβα (T ). the state ψ(T ) fails to return to the initial state and is subject to a unitary rotation as ψ(T ) = exp − The unitary matrix Γ[C] := Γ(T ) = P exp − C i 0 T k ε(λ(s))ds β=1 n. Berry phase This example was ﬁrst studied in Berry’s original paper. eiφ/2 cos(θ/2) (20) (21) Energy eigenvalues are nondegenerate for any nonvanishing B. eiφ/2 sin(θ/2) −e−iφ/2 sin(θ/2) . e sin θ − cos θ iφ (19) Its eigenvalues and eigenvectors are given as H +. 2. β. It is important to note that the holonomy Γ[C] depends only on the shape of the loop C in the parameter space and is independent of parametrization of the loop by λ(t). by an adiabatic change of the external parameter (B. B = −B −.
2 −e−iφ/2 sin(θ/2) eiφ/2 cos(θ/2) (24) (25) For a loop C = ∂S in the parameter space. B = (eiφ/2 cos(θ/2). B = d = i 2 e−iφ/2 cos(θ/2) eiφ/2 sin(θ/2) 1 −e−iφ/2 cos(θ/2) dφ + eiφ/2 sin(θ/2) 2 −e−iφ/2 sin(θ/2) eiφ/2 cos(θ/2) 1 e−iφ/2 sin(θ/2) dφ − eiφ/2 cos(θ/2) 2 e−iφ/2 cos(θ/2) dθ. (22) eiφ/2 cos(θ/2) d−. which is deﬁned by (13). The connection form A = ·d· . . calculate the inner products A+ (B) = +. eiφ/2 sin(θ/2) (23) −e−iφ/2 sin(θ/2) dθ. B d+. First. B = (−eiφ/2 sin(θ/2). B = d = i 2 Second.120 the Berry phase. B d−. e−iφ/2 sin(θ/2)) i 2 = 1 −e−iφ/2 cos(θ/2) dφ + eiφ/2 sin(θ/2) 2 −e−iφ/2 sin(θ/2) dθ eiφ/2 cos(θ/2) e−iφ/2 cos(θ/2) eiφ/2 sin(θ/2) i − cos2 (θ/2) + sin2 (θ/2) dφ 2 1 − cos(θ/2) sin(θ/2) + sin(θ/2) cos(θ/2) dθ + 2 i = − cos θ dφ. B . is calculated as follows. e−iφ/2 sin(θ/2)) d = (eiφ/2 cos(θ/2). 2 A− (B) = −. we will observe the Berry phase eiγ = exp − C A+ = exp i 2 cos θ dφ = exp − C i 2 sin θ dθ ∧ dφ S (26) if the initial state is +. e−iφ/2 cos(θ/2)) d = i cos θ dφ. calculate the exterior derivatives of the eigenvectors to get d+.
121 2.22 The Hilbert space is assumed to be CN +1 . . . ω⊥ (31) 2 where c± = (E± + ω2 )− 2 . . (30) 0 . . . The Hamiltonian is N H = εe e + j=1 ∗ ωj gj e + ωj e gj  . ω⊥ = 0. (27) The state e represents an excited state and gj (j = 1.2. ωN Hα = 0. . . (29) He− = E+ e− . E− = ε − 2 ε 2 2 + ω2 .. . . α= 1 ω⊥  1 He+ = E+ e+ . These parameters are assumed to be controllable.4. ωN 0 0 · · · 0 Its eigenvalues and eigenvectors are E+ ω1 e+ = c+ ω 2 . the eigenvalue E = 0 is (N − 1)fold degenerate. . . . . . The parameter ε ∈ R is the excitation energy and ωj ∈ C are coupling constants. Thus. E+ = ε + 2 ε 2 2 + ω2 . . N ) represent degenerate ground states before couplings between the states are introduced. · · · . . ωN E− ω1 e− = c− ω2 . . . The vector ω ⊥ ∈ CN satisﬁes ω. . . Λtype system This system has been studied as a model of the holonomic quantum computer which uses atomic energy levels21 and also as a model of the holonomic computer which uses Josephson junctions. The Hamiltonian can be written in the matrix form ∗ ∗ ∗ ε ω1 ω2 · · · ωN ω1 0 0 · · · 0 (28) H = ω2 0 0 · · · 0 . .
dα1 α1 . dα2 6 = µ=1 Aµ dλµ (34) with ω1 = λ1 + iλ2 . For these reasons the holonomic quantum computer is regarded as one of promising candidates for realistic quantum computers.22 3. we have the twofold degenerate level E = 0.12 and also Pachos with them13 proposed to use the WilczekZee holonomy (17) as a unitary gate of quantum computer. (35) The concrete calculations of the holonomy can be found in literature.122 When N = 3. . The calculation to get the concrete forms of Aµ is cumbersome and we do not show it here. ω2 = λ3 + iλ4 . dα2 α2 . which is an (N −1)×(N −1) matrix. Holonomic Quantum Computer Zanardi and Rasetti.21. a useful holonomic quantum computer should have the following properties: (i) Controllability: a holonomic quantum computer should have a large enough number of control parameters for generating arbitrary holonomies. the holonomic quantum computer is robust against noise from the environment and tolerates incompleteness of the control. dα1 α2 . Since the holonomy depends only on the geometry of the controlparameter loop and it is insensitive to the detail of the dynamics of the system. ω3 = λ5 + iλ6 . the associated holonomy g(t).(33) α2 = ∗ ω2 ω3 (ω1 2 + ω3 2 )(ω1 2 + ω2 2 + ω3 2 ) −(ω1 2 + ω2 2 ) The components of the connection are to be calculated according to A= α1 . is a solution of the diﬀerential equation dg(t) =− dt Aµ (λ(t)) g(t) µ dλµ (t) . However. (32) α1 = ∗ ω1 2 + ω2 2 −ω1 0 0 ∗ 1 ω1 ω3 . The eigenvectors are parametrized as 0 ∗ ω2 1 . When we take a loop λ(t). dt g(0) = I.
1. The requests (iii) and (iv) can be analyzed with a mathematical method. .14 We outline the geometrical setting of the problem here to make this paper selfcontained.k (C). and the interaction between the system and the environment causes loss of coherence of the quantum state of the system. We need to ﬁnish the whole process in a possible shortest time to avoid the undesirable decoherence. it remains a hard task to ﬁnd a feasible design. the Grassmann manifold GN. Moreover. and the unitary group U (k) as explained below. U (k)). The requests (i) and (ii) are practically important issues. This optimization problem is the main subject of this lecture.k (C) → GN. which consists of the Stiefel manifold SN. the projection map π : SN. A mathematical setting to describe this system is the principal ﬁber bundle (SN. Hence. (iv) Short execution time: an actual system cannot be completely isolated from its environment.k (C). A concise review on this subject also given by Fujii. They naturally lead us to an optimization problem. to realize a large qubits system it is necessary to ﬁnd a system which has a highly degenerate energy level. Formulation of the Problem and its Solution 4. 4. Inaccuracy of degeneracy spoils the precision of the holonomic computation. π. Although many designs of working holonomic computers have been proposed. No level crossings should occur during the process since a level crossing violates the adiabaticity condition. the degree of degeneracy is equal to the dimension of the Hilbert space available for quantum computation. Assume that the ground state of each Hamiltonian is kfold degenerate (k < N ). (iii) Adiabaticity: we need to change the control parameter as slowly as possible to avoid undesirable transitions between diﬀerent energy levels of the quantum system.k (C).123 (ii) Accurate degeneracy: degeneracy of the energy levels should be kept during the timeevolution. we need to ﬁnd the shortest loop in the control parameter space with producing the demanded holonomy.k (C). Suppose that the state of a system is described by a ﬁnitedimensional Hilbert space CN and the system has a family of Hamiltonians acting on the Hilbert space.k (C). Geometrical setting The diﬀerential geometry of connection3. GN.4 provides a suitable language to describe the WilczekZee holonomy. In an actual system degeneracy is often resolved by a tiny asymmetric interaction or a noise.
124
The Stiefel manifold is the set of orthonormal kframes in CN , SN,k (C) := {V ∈ M (N, k; C)  V † V = Ik }, (36)
where M (N, k; C) is the set of N × k complex matrices and Ik is the kdimensional unit matrix. The unitary group U (k) acts on SN,k (C) from the right SN,k (C) × U (k) → SN,k (C), (V, h) → V h (37)
by means of matrix product. It should be noted that this action is free. In other words, h = Ik if there exists a point V ∈ SN,k (C) such that V h = V . The Grassmann manifold is deﬁned as the set of kdimensional subspaces in CN , or equivalently, as the set of projection operators of rank k, GN,k (C) := {P ∈ M (N, N ; C)  P 2 = P, P † = P, trP = k}. The projection map π : SN,k (C) → GN,k (C) is deﬁned as π : V → P := V V † . (39) (38)
It is easily proven that the map π is surjective. Namely, for any P ∈ GN,k (C), there is V ∈ SN,k (C) such that π(V ) = P . It is easily observed that π(V h) = (V h)(V h)† = V hh† V † = V V † = π(V ) (40)
for any h ∈ U (k) and any V ∈ SN,k (C). It also can be veriﬁed that if π(V ) = π(V ) for some V, V ∈ SN,k (C), there exists a unique h ∈ U (k) such that V h = V . Thus the Stiefel manifold SN,k (C) becomes a principal bundle over GN,k (C) with the structure group U (k). In particular, when k = 1, the principal bundle (SN,k (C), GN,k (C), π, U (k)) is denoted as (S 2N −1 , CP N −1 , π, U (1)) and is called the Hopf bundle. Then the Stiefel manifold becomes a (2N − 1)dimensional sphere SN,1 (C) = S 2N −1 = {z ∈ CN  z † z = 1 }. (41)
At this time, we can introduce an equivalence relation z ∼ z for z, z ∈ S 2N −1 as z ∼ z :⇐⇒ ∃θ ∈ R, eiθ z = z (42)
and denote the equivalence class of z as [z] = {z ∈ S 2N −1  z ∼ z } = {eiθ z  θ ∈ R}. Then the quotient space S 2N −1 / ∼, which is the whole set of the equivalence classes, is called the projective space CP N −1 . Each point
125
[z] of CP N −1 uniquely determines a onedimensional subspace in CN . It is easily veriﬁed that z ∼ z ⇐⇒ π(z) = π(z ) (43)
for the projection map π(z) = zz † . Thus the Grassmann manifold GN,1 (C) can be identiﬁed with CP N −1 . The structure group U (1) = {eiθ } acts on the ﬁber [z]. When N = 2 and k = 1, the Grassmann manifold is isomorphic to G2,1 (C) ∼ CP 1 ∼ S 2 . At the same time the Stiefel manifold = = is S2,1 (C) ∼ S 3 . Then the projection map is π : S 3 → S 2 . = For a general values of N and k, the group U (N ) acts on both SN,k (C) and GN,k (C) as U (N ) × SN,k (C) → SN,k (C), U (N ) × GN,k (C) → GN,k (C), (g, V ) → gV, (g, P ) → gP g
†
(44) (45)
by matrix product. It is easily veriﬁed that π(gV ) = gπ(V )g † . This action is transitive, namely, there is g ∈ U (N ) for any V, V ∈ SN,k (C) such that V = gV . There is also g ∈ U (N ) for any P, P ∈ GN,k (C) such that P = gP g † . The stabilizer group of a point V ∈ SN,k (C) is deﬁned as GV := {g ∈ U (N )  gV = V } (46)
and it is isomorphic to U (N − k). Similarly, it can be veriﬁed that the stabilizer group of each point in GN,k (C) is isomorphic to U (k) × U (N − k). Thus, they form the homogeneous bundle π : SN,k (C) ∼ U (N )/U (N − k) → GN,k (C) ∼ U (N )/(U (k) × U (N − k)). = = (47) The canonical connection form on SN,k (C) is a u(k)valued oneform A := V † dV. (48)
which is a generalization of the WilczekZee connection (13). This is characterized as the unique connection that is invariant under the action (44). The associated curvature twoform is then deﬁned and calculated as F := dA + A ∧ A = dV † ∧ dV + V † dV ∧ V † dV = dV † ∧ (IN − V V † )dV, (49) where we used dV † V + V † dV = 0, which is derived from V † V = Ik . These manifolds are equipped with Riemannian metrics. We deﬁne a metric dV
2
:= tr (dV † dV )
(50)
126
for the Stiefel manifold and dP for the Grassmann manifold. 4.2. The isoholonomic problem Here we reformulate the WilczekZee holonomy in terms of the geometric terminology introduced above. The state vector ψ(t) ∈ CN evolves according to the Schr¨dinger equation o d ψ(t) = H(t)ψ(t). dt The Hamiltonian admits a spectral decomposition i
L 2
:= tr (dP dP )
(51)
(52)
H(t) =
n=1
εn (t)Pn (t)
(53)
with projection operators Pn (t). Therefore, the set of distinct energy eigenvalues (ε1 , . . . , εL ) and mutually orthogonal projectors (P1 , . . . , PL ) constitutes a complete set of control parameters of the system. Now we concentrate on the eigenspace associated with the lowest energy, which is assumed to be identically zero, ε1 ≡ 0. We write P1 (t) as P (t) simply. Suppose that the degree of degeneracy k = tr P (t) is constant. For each t, there exists V (t) ∈ SN,k (C) such that P (t) = V (t)V † (t). By adiabatic approximation we mean that the state vector ψ(t) belongs to the subspace spanned by the frame V (t), which is the eigensubspace of the instantaneous Hamiltonian. Then for the state vector ψ(t) there exists a vector φ(t) ∈ Ck such that ψ(t) = V (t)φ(t). (54)
This equation is equivalent to (9). Since H(t)ψ(t) = ε1 ψ(t) = 0, the Schr¨dinger equation (52) becomes o dφ dV +V† φ(t) = 0 dt dt and its formal solution is written as φ(t) = P exp − Therefore ψ(t) is written as ψ(t) = V (t)P exp − V † dV V † (0)ψ(0). (57) V † dV φ(0). (55)
(56)
127
In particular, when the control parameter comes back to the initial point as P (T ) = P (0), the holonomy Γ ∈ U (k) is deﬁned via ψ(T ) = V (0)Γ φ(0) and is given explicitly as Γ = V (0)† V (T ) P exp − If the condition V† dV =0 dt (60) V † dV . (59) (58)
is satisﬁed, the curve V (t) in SN,k (C) is called a horizontal lift of the curve P (t) = π(V (t)) in GN,k (C). Then the holonomy (59) is reduced to Γ = V † (0)V (T ) ∈ U (k). (61)
Now we are ready to state the isoholonomic problem in the present context; given a speciﬁed unitary gate Ugate ∈ U (k) and an initial point P0 ∈ GN,k (C), ﬁnd the shortest loop P (t) in GN,k (C) with the base points P (0) = P (T ) = P0 whose horizontal lift V (t) in SN,k (C) produces a holonomy Γ that coincides with Ugate . This problem was ﬁrst investigated systematically by a mathematician, Montgomery,15 who was motivated from experimental studies of Berry phase. The isoholonomic problem is formulated as a variational problem. The length of the horizontal curve V (t) is evaluated by the functional
T
S[V, Ω] =
0
tr
dV † dV dt dt
− tr Ω V †
dV dt
dt,
(62)
where Ω(t) ∈ u(k) is a Lagrange multiplier to impose the horizontality condition (60) on the curve V (t). Note that the value of the functional S is equal to the length of the projected curve P (t) = π(V (t)),
T
S=
0
1 dP dP tr 2 dt dt
dt.
(63)
Thus the problem is formulated as follows; ﬁnd functions V (t) and Ω(t) that attain an extremal value of the functional (62) and satisfy the boundary condition (61).
128
4.3. The solution: horizontal extremal curve We shall derive the EulerLagrange equation associated to the functional (62) and solve it explicitly. A variation of the curve V (t) is deﬁned by an arbitrary smooth function η(t) ∈ u(N ) such that η(0) = η(T ) = 0 and an inﬁnitesimal parameter ∈ R as V (t) = (1 + η(t))V (t). (64)
By substituting V (t) into (62) and diﬀerentiating with respect to , the extremal condition yields 0= dS d
T
=
=0 0
˙ ˙ tr η(V V † − V V † − V ΩV † ) dt ˙
t=T t=0
˙ ˙ = tr η(V V † − V V † − V ΩV † ) −
0 1
d ˙ ˙ tr η (V V † − V V † − V ΩV † ) dt. dt
(65)
Thus we obtain the EulerLagrange equation d ˙ † ˙ (V V − V V † + V ΩV † ) = 0. dt (66)
˙ The horizontality condition V † V = 0 is rederived from the extremal condition with respect to Ω(t). Finally, the isoholonomic problem is reduced to the set of equations (60) and (66), which we call a horizontal extremal equation. Next, we solve the equations (60) and (66). The equation (66) is integrable and yields ˙ ˙ V V † − V V † + V ΩV † = const = X ∈ u(N ). (67)
˙ Conjugation of the horizontality condition (60) yields V † V = 0. Then, by multiplying V on (67) from the right we obtain ˙ V + V Ω = XV. By multiplying V † on (68) from the left we obtain Ω = V † XV. (69) (68)
˙ The equation (68) implies V = XV − V Ω, and hence the time derivative of
129
Ω(t) becomes ˙ ˙ ˙ Ω = V † X V + V † XV = V † X(XV − V Ω) + (−V † X + ΩV † )XV = V † XXV − ΩΩ − V † XXV + ΩΩ = 0. (70) We used the facts that X † = −X and Ω† = −Ω. Therefore, Ω(t) is actually a constant. Thus the solution of (68) and (69) is V (t) = etX V0 e−tΩ , Ω = V0† XV0 . (71)
We call this solution the horizontal extremal curve. Then (67) becomes (XV − V Ω)V † − V (−V † X + ΩV † ) + V ΩV † = X, which is arranged as X − (V V † X + XV V † − V V † XV V † ) = 0, (72)
where we used (69). Since the group U (N ) acts on SN,k (C) freely, we may take V0 = Ik 0 ∈ SN,k (C) (73)
as the initial point without loss of generality. We can parametrize X ∈ u(N ), which satisﬁes (69), as X= Ω W −W † Z (74)
with W ∈ M (k, N − k; C) and Z ∈ u(N − k). Then the constraint equation (72) forces us to choose Z = 0. (75)
Via the above discussion we obtained a complete set of solution (71) of the horizontal extremal equation (60) and (66). The complete solutions are parametrized with constant matrices Ω ∈ u(k) and W ∈ M (k, N − k; C), and are written explicitly as V (t) = etX V0 e−tΩ , X= Ω W −W † 0 . (76)
We call the matrix X a controller. At this time the holonomy (61) is expressed as Γ = V † (0)V (T ) = V0† eT X V0 e−T Ω ∈ U (k). (77)
130
These results (76) and (77) coincide with Montgomery’s resultsb . We took a diﬀerent approach from his; in the above discussion we wrote down the EulerLagrange equation and solved it directly. We evaluate the length of the extremal curve for later convenience by substituting (76) into (63) as
T
S=
0
1 dP dP tr 2 dt dt
dt = tr (W † W ) T.
(78)
5. The BoundaryValue Problem Remember that we are seeking for the optimally shortest loop in the Grassmann manifold which generates a desired holonomy. The optimality is guaranteed by solving the horizontal extremal equation. The remaining task is to ﬁnd a solution that satisﬁes two boundary conditions. First, when the curve V (t) in SN,k (C) is projected into GN,k (C) via the map π : SN,k (C) → GN,k (C), V → V V † , it should make a loop. Thus the curve should satisfy the loop condition V (T )V † (T ) = eT X V0 V0† e−T X = V0 V0† . (79)
Second, the terminal points V (0) and V (T ) of the curve V (t) deﬁne the holonomy Γ = V † (0)V (T ) as in (61). Thus the curve should satisfy the holonomy condition V0† V (T ) = V0† eT X V0 e−T Ω = Ugate (80)
for a speciﬁed unitary gate Ugate ∈ U (k). We need to ﬁnd the suitable parameters Ω and W in solutions (76) satisfying the above two conditions. Montgomery15 presented this boundaryvalue problem as an open problem. In this section we give a scheme to solve systematically this problem. 5.1. Equivalence class There is a class of equivalent solutions for a given initial condition V0 and a given ﬁnal condition V (T ) = V0 Ugate . Here we clarify the equivalence relation among solutions {V (t)} that have the form (76) and satisfy (79) and (80).
b In his paper15 Montgomery cited B¨r’s theorem to complete the proof. However, B¨r’s a a paper being a diploma thesis, it is not widely available. Instead of relying B¨r’s theorem, a here we did the explicit calculation to reach the same result.
131
We say that two solutions V (t) and V (t) are equivalent if there are elements g ∈ U (N ) and h ∈ U (k) such that V (t) and V (t) = gV (t)h† satisfy the same boundary conditions gV0 h† = V0 and hUgate h† = Ugate . (83) For the initial point (73), the condition (82) states that g ∈ U (N ) must have a blockdiagonal form g= h1 0 0 h2 , h = h1 ∈ U (k), h2 ∈ U (N − k). (84) (82) (81)
The controller X of V (t) are then found from V (t) = gV (t)h† = getX g † gV0 h† he−tΩ h† = etgXg gV0 h† e−thΩh = etgXg V0 e−thΩh .
† † † †
(85)
In summary, two controllers X and X are equivalent if and only if there are unitary matrices h1 ∈ U (k) and h2 ∈ U (N − k) such that X= Ω W −W † 0 , X = h1 Ωh† h1 W h† 1 2 −h2 W † h† 0 1 , h1 Ugate h† = Ugate . 1 (86) 5.2. U (1) holonomy Here we calculate the holonomy for the case N = 2 and k = 1. In this case the homogeneous bundle π : S2,1 (C) → G2,1 (C) is the Hopf bundle π : S 3 → S 2 with the structure group U (1) and the WilczekZee holonomy reduces to the Berry phase. In the subsequent subsection we will use the Berry phase as a building block to construct a nonAbelian holonomy. We normalize the cycle time as T = 1 in the following. Using real numbers w1 , w2 , w3 ∈ R we parametrize the controller as X= 2iw3 iw1 + w2 iw1 − w2 0 = iw3 I + iw1 σ1 + iw2 σ2 + iw3 σ3 , (87)
where σ = (σ1 , σ2 , σ3 ) = (σx , σy , σz ) are the Pauli matrices which have been introduced at (19). The exponentiation of X yields etX = eitw3 (I cos ρt + in · σ sin ρt), (88)
the controller parameters are ﬁxed as w3 = nπ − γ. The real parameter φ is not ﬁxed by the loop condition and the holonomy condition. (93) Thus. is now evaluated as S = tr (W † W ) T = (nπ)2 − (nπ − γ)2 = γ(2nπ − γ). To make a loop. 0. . The phase h2 = eiφ parametrizes solutions in an equivalence class as observed in (86). . ). . w = w n. the control parameters must satisfy ρ = w = nπ (n = ±1. the integer n counts the winding number of the loop. The length of the loop. The point P (t) comes back to the north pole when t satisﬁes 2ρt = 2πn with an integer n. 1). cos ρ = (−1)n and the holonomy (80) is evaluated as V0† eX V0 e−Ω = eiw3 (−1)n e−2iw3 = e−i(w3 −nπ) = Ugate = eiγ . (91) cos ρt + in3 sin ρt (in1 − n2 ) sin ρt (90) where e3 = (0. ±2. to produce the holonomy Ugate = eiγ . (78). w1 + iw2 = e−iφ (nπ)2 − (nπ − γ)2 .132 where ρ and n are deﬁned as ρ := w = (w1 )2 + (w2 )2 + (w3 )2 . (92) Then. We see from (91) that the point P (t) in S 2 starts at the north pole e3 of the sphere and moves along a small circle with the axis n in the clockwise sense by the angle 2ρt. Hence each equivalence class is characterized by the integer n. namely. (95) . to satisfy the loop condition (79) at t = T = 1. (94) This is the solution to the boundary problem deﬁned by (79) and (80). the point P (t) travels the same small circle n times during 0 ≤ t ≤ 1. Here the nonvanishing integer n must satisfy (nπ)2 − (nπ − γ)2 > 0. Therefore. (89) The associated horizontal extremal curve (71) then becomes V (t) = etX V0 e−tΩ = e−itw3 and the projected curve in S 2 becomes P (t) = V (t)V † (t) 1 1 = I + σ · n(n · e3 ) + (e3 − n(n · e3 )) cos 2ρt 2 2 −(n × e3 ) sin 2ρt . At t = 1.
In the ﬁrst step. . . third. It turns out that the working space should have a dimension N ≥ 2k to apply our method. . constructing equations of the controller. (99) and (100). we conclude that the controller of Ugate = eiγ is X= ie−iφ 2i(π − γ) ieiφ π 2 − (π − γ)2 π 2 − (π − γ)2 0 . . 5. We call the set of equations. In the following we assume that N = 2k. (98). ωj = 2(π − γj ). Then we obtain a diagonal controller Xdiag = Ωdiag Wdiag † −Wdiag 0 . (96) We call this solution a small circle solution because of its geometric picture mentioned above. . . we diagonalize a given unitary matrix Ugate ∈ U (k) as R† Ugate R = Udiag = diag(eiγ1 . . (100) which is a 2k × 2k matrix. we combine single loop solutions associated with the Berry phase to construct two k × k matrices Ωdiag = diag(iω1 . Each eigenvalue γj is taken in the range 0 ≤ γj < 2π. diagonalize the unitary matrix Ugate to be implemented. . Our method consists of three steps: ﬁrst. . τj = eiφj π 2 − (π − γj )2 . This is our main result. (97). iωk ). second. Wdiag = diag(iτ1 . . iτk ). In the second step. (98) (99) In the third step. The time interval is normalized as T = 1 as before. . eiγk ) (97) with R ∈ U (k). the simple loop with n = 1 is the shortest one among the extremal loops. we construct the controller X as X= R 0 0 Ik Ωdiag Wdiag † −Wdiag 0 R† 0 0 Ik = RΩdiag R† RWdiag † −Wdiag R† 0 . Thus.133 For a ﬁxed γ in the range 0 ≤ γ < 2π. construct a diagonal controller matrix by combining small circle solutions. . . U (k) holonomy Here we give a prescription to construct a controller matrix X that generates a speciﬁc unitary gate Ugate .3. undo diagonalization of the controller.
134 The diagonal controller Xdiag is actually a direct sum of controllers (96). Hadamard gate The Hadamard gate is a onequbit gate deﬁned as 1 UHad = √ 2 It is diagonalized by R= as R† UHad R = 1 0 0 −1 . we need to calculate the diagonalizing matrix R. we have γ1 = 0 and γ2 = π. 6. we obtain the optimal controller of the Hadamard gate. Examples of Unitary Gates To illustrate our method here we construct controllers for the onequbit Hadamard gate and the twoqubit discrete Fourier transformation (DFT) gate. (97)(100). √ 2 (101) . √ 2+1 √ 1 2−1 1 . its holonomy is also a direct sum of the Berry phases (93) as V0† eXdiag V0 e−Ωdiag = Udiag and hence we have V0† eX V0 e−Ω = RV0† eXdiag V0 R† R e−Ωdiag R† = RUdiag R† = Ugate . We may put φ1 = φ2 = 0. provide the desired optimal controller matrices. are Ωdiag = diag(2iπ. RWdiag iπ = 2 2 . 0 − 2− 0 2+ √ √ (104) .1. Then the constructing equations of the controller. (103) cos π − sin π 8 8 sin π 8 cos π 8 . π cos = 8 2+ 2 1 1 1 −1 √ 2 . For each unitary gate Ugate . Hence. π sin = 8 2− 2 (102) Therefore. 6. iπ). (97)(100). 2 (105) Substituting these into (100). and hence iπ RΩdiag R = √ 2 † Wdiag = diag(0. The ingredients of the constructing equations of the controller. 0). which generate Berry phases {eiγj }.
3 iπ 1 RΩdiag R† = 2 1 1 1 1 1 2 −1 0 . 0. 2iπ. 0 −1 0 0 0 i (107) 1 0 R† UDFT2 R = 0 0 (108) Therefore. The twoqubit DFT (DFT2) is a unitary transformation 1 1 1 1 1 1 i −1 −i . Discussions Although we provided a concrete prescription to calculate a controller of the holonomic quantum computer to execute an arbitrary unitary gate. we have γ1 = γ2 = 0. DFT2 gate Discrete Fourier transformation (DFT) gates are important in many quantum algorithms including Shor’s algorithm for integer factorization. . iπ). Thus the ingredients of the controller are √ Ωdiag = diag(2iπ.135 6. (109) and hence −1 0 1 − 3/2 . RWdiag 1 0 3/2 1 (110) Substituting these into (100). −1 3 −1 0 −1 2 00 iπ 0 0 = 2 0 0 00 7. we ﬁnally obtain the optimal controller of the DFT2 gate. γ3 = π and γ4 = π/2.2. UDFT2 = (106) 2 1 −1 1 −1 1 −i −1 i It is diagonalized by √ 1 2 1 1 0 √ R= 2 −1 2 1 0 as −1 √ 0 1 − 2 1 √ 0 2 1 0 0 0 1 0 0 . Wdiag = diag(0. iπ. iπ 3/2). 0.
21 proposed a use of neutral atoms and the photon ﬁeld in cavity.20 proposed a use of trapped ions. 7. we assumed that we can change the projection oprator P1 (t) appearing in the Hamiltonian (53) arbitrarily. Usually a real system admits only a smaller control manifold. Recati et al. Unanyan et al.2. Zhu and Zanardi27 studied robustness holonomic quantum computation against control errors.28 and a use of decoherencefree subspace also has been studied by Wu et al. which is a special case of our study. the degeneracy must be precise and the degenerate subspace must have large dimensions to be used as a multiqubit processor. by FuentesGuridi et al. Faoro et al. Mead16 proposed a use of Kramers degeneracy in molecules to observe the nonAbelian holonomy. . Restricted control parameters To solve the isoholonomic problem we assumed that the Stiefel manifold is a control manifold.23 studied the holonomy in a semiconductor quantum dots. In the rest of this lecture we shall discuss these issues. Finding and optimizing a loop in the control manifold to generate a demanded holonomy become a diﬃcult task for a realistic system. a lot of theoretical studies related to holonomic quantum computation can be found in literature. Duan et al. On the other hand. In other words.136 there remain several problems for building a practical holonomic quantum computer.26 studied numerical optimization of holonomic quantum computer.25. Niskanen et al. However. the most important issue for making a practical holonomic quantum computer is to build a controllable system that has a degenerate energy level. Zee17 proposed a use of nuclear quadrupole resonance.22 proposed a use of Josephson junction circuits. Implementation Here we describe a (not complete) list of proposals for implementing holonomic quantum computers although most of them still remain theoretical proposals. for example. Solinas et al. 7. the number of control parameters can be less and the ranges of their allowed values can be narrow.18 proposed a use of fourlevel atoms manipulated with laser as a holonomic quantum computer.29 However. Karle and Pachos24 discussed holonomy in the Grassmann manifold. Robustness against decoherence has been analyzed. realistic systems do not allow such a large controllability. Jones et al. It seems necessary to ﬁnd a breakthrough in this direction.19 implemented geometric quantum computation with NMR.1.
Math. 11. Phys. Phys. W. 2003). 2. Math. S. Zee. Phys. Duan. Tanimura. A 264 (1999) 94. Geometry. 8. J. V. C. Phys. Math. Rasetti. 59 (1987) 161. Cirac. 16. 15. Recati. I. Q. 2000). Grant Nos. P. Rev. P. P. R. 9. Lond. 2 (Interscience Publishers. Zee. B. 21. Ekert. Phys. Nakahara. References 1. and K. G. Shapere and F. T. M. Calarco. K. Castagnoli. 1969). Faoro. Cirac. A. 2 (2002) 560. Geometric Phases in Physics (World Scientiﬁc. J. Fujii. A 38 (1988) 1. A 59 (1999) 2910. Phys. 14. Niskanen. Fazio. 1989). This work is ﬁnancially supported by Japan Society for the Promotion of Science. Phys. Rev. Phys. and P. Rossi. Nakahara. Nakahara. Soryushiron Kenkyu 85 (1992) 1. 2007. Nakahara. 25. 128 (1991) 565. 41 (2000) 4406. Topology and Physics.M. Siewert. and J. Wilczek (editors). Soller. Zanardi. Mathematical Aspects of Quantum Computing. A 61 (1999) 010305(R). 90 (2003) 28301. Kato. A. O. Zoller. Quantum Inf. M. Koizumi. held at Kinki University in August. I also appreciate eﬀorts of the organizers of this successful summer school. A. 18. I would like to thank Mikio Nakahara for stimulating discussions with him. Proc. B 67 (2003) 121307. Salomaa. Pachos. Lett. Mostafazadeh. 7. Foundations of Diﬀerential Geometry. S. Phys. 5. Kobayashi and K. Nomizu. H. Soc. Rev. Quantum Mechanics (Dover. Zwanziger. M. Montgomery. Zanghi. Singapore. and C. A. S. R. 13. Lett. Rev. and M. A. 24. Commun. Rev. Bohm. Phys. Phys. which was explained in this lecture. and F. F. 3. Mead. Berry. 15540277 and 17540372. Jones. M. 22. A 325 (2004) 199. Zanardi and M. A. A 66 (2002) 032309. 23. V. The Geometric Phase in Quantum Systems (Springer. Soc. Solinas. N. 20. 17. and R. Japan 5 (1950) 435. Rasetti. Shore. A. L. 10. Phys. 2nd ed. Vol. . T. Rev. and M. I. is a result of collaborations with Daisuke Hayashi and Mikio Nakahara. 19. J. L. Karle and J. and M. Lett. Hayashi. 6. J. Tanimura. 52 (1984) 2111. Rev. J. Bristol and New York. 46 (2005) 022101. R. Messiah. K. Unanyan. M. Comput. Vedral. A 392 (1984) 45. and P. 12. Pachos. Lett. Niu. P. Phys. 2003). A. New York. D. A. J. 4. Wilczek and A. Math. R. J. Tanimura. Phys. Phys. J. Nature 403 (2000) 869. P. Bergmann. and D. Rev. (IOP Publishing. Science 292 (2001) 1695. Tom and Berry (A thesis paper on holonomies in dynamical systems written in Japanese). Lett. S. Zanardi. New York. 44 (2003) 2463. Hayashi.137 Acknowledgements The exact solution of the isoholonomic problem. J. Zanardi.
F. S. Rev. A. and D. Rev. Rev. Livine. P. Phys. Wu. Phys. 29. M. O.138 26. Lidar. Phys. A 67 (2003) 012319. Zanardi.A. A. Lett. Niskanen. and E. 28. Zhu and P. Rev. I. Girelli. A 72 (2005) 020301. Phys. L. Lett. FuentesGuridi. Nakahara. 27. 94 (2005) 020503. and M.L. . M. 95 (2005) 130501. Zanardi. Salomaa.
1.1 The start of modern game theory dates back to 1928 when J. This benchmark exploits the fact that in special settings a classical game should be reproduced as a subgame of its quantum extension. NOBUYUKI IMOTO ¸ Division of Materials Physics. In addition. Quantum Mechanics.mp. Keywords: Game Theory. Each of the decision makers is motivated to increase his/her utility which is not solely determined by his/her action but by the interaction of the strategies of all the decision makers. Osaka University. Entanglement. Graduate School of Engineering Science.2 Later on. Japan CREST Photonic Quantum Information Project.jp In this lecture note. quantum operations and measurement is used. Saitama 3310012. we present the implications of playing classical games in quantum mechanical settings where the quantum mechanical toolbox consisting of entanglement. Introduction Game theory is a mathematical modelling tool to formally describe conﬂict. 418 Honmachi. Saitama 3310012. von Neumann introduced the concepts of linear programming and the . Kawaguchi. Toyonaka. Japan Email: ozdemir@qi. we study quantum games and their corresponding classical analogues to determine the novelties.139 PLAYING GAMES IN QUANTUM MECHANICAL SETTINGS: FEATURES OF QUANTUM GAMES ¨ SAHIN KAYA OZDEMIR. we introduce a benchmark which attempts to make a fair comparison of classical games and their quantum extensions. We obtained a rather surprising result that this requirement prevents the use of a large set of entangled states in quantum extension of classical games. von Neumann deﬁned and completely solved twoperson zerosum games. After a brief introduction to the concepts of classical game theory and quantum mechanics.ac. Quantum Games.es. JUNICHI SHIMAMURA. Osaka 5608531. Kawaguchi. 418 Honmachi. Japan SORST Research Team for Interacting Carrier Electronics. Department of Materials Engineering Science.osakau. cooperation or competition situations among rational and intelligent decision makers who have diﬀerent preferences.
recently mathematical models and techniques of game theory have increasingly been used by computer and information scientists. then it is not a surprise that all those distinct areas meet in game theory: Information constitutes a common ground and a strong connection among them.12. Proposals have been done for modelling complex physical phenomena such as decoherence and irreversibility using game theoretical models. which say physics and social sciences are distinct ﬁelds and should not mix. the two were developed as separate ﬁelds with no or very little interaction. measurement) was introduced in game theory. Following this ﬁrst paper.16 These two works raised an excitement in the quantum mechanics community mostly due to emergence of a new ﬁeld where quantum mechanics shows its advantage and also due to the hope of using game theoretical toolbox to study quantum information tasks and quantum algorithms. large scale distribution systems.3 Critiques. von Neumann’s response to these criticism was ”such statements are at least premature. Quantum mechanical toolbox (unitary operators. studied the wellknown Prisoners’ dilemma (PD) using a general protocol for two playertwo strategy games where the players share entanglement. for the ﬁrst time. Eisert et al. and with no common ground for applications. population dynamics. However. and so is the response of von Neumann.3. His attempts to elucidate the economical problems using physical examples faced strong criticism based on the argument that social sciences cannot be modelled after physics because the social sciences should take psychology into account.4 Although von Neumann made important contributions to quantum mechanics and game theory around the same years. Since then most of the works has been devoted to the analysis of dilemma containing classical game dynamics using Eisert’s .140 fundamentals of game theory which later became the building blocks for his book with Morgenstern. J. superposition and entangled states.5–11 Game theoretical models of economics.”. This leads to the introduction of the quantum mechanical toolbox into game theory to combine two distinct interests of von Neumann in an environment to mutually beneﬁt from both. and to see what new features will arise from this combination. learning models and resource allocation problems have been studied. are still alive. by Meyer15 who showed that in a coin tossing game a player utilizing quantum superposition could win with certainty against a classical player. evolutionary biology. Game theory and its theoretical toolbox have been traditionally limited to economical and evolutionary biology problems.14 Then one naturally wonders what happens if the information carriers are taken to be quantum systems.13 If we bear in mind that a game is about the transfer of information.
In game theory. or awards distributed to players depending on the choices of all players. this area still begs for more exploration. prizes. rational and intelligent decision makers.g. (iii) Payoﬀs: Beneﬁts. Rationality of players and common knowledge are two important assumptions in game theory. In this work. players are allowed to form coalitions and combine their decision making problems with binding commitments. e. Finally. games are generally modelled either in strategic form or in extensive form. Players can choose to play either in pure or mixed strategies. In pure strategy. each player chooses an action with unit probability. which is represented by a game tree. We will derive a condition that should be satisﬁed for an entangled state to be useful in quantum version of classical games.. we will brieﬂy introduce the terminology and basic concepts of classical game theory. which is best represented by a payoﬀ matrix. If the players take their actions simultaneously or without observing the actions taken by other players.19 In this review. Games are usually grouped as cooperative or noncooperative games.18 were reformulated in the language of game theory. Rationality implies that each player aims at maximizing his own payoﬀ. whereas in the latter . is preferred. (ii) Strategies: Feasible action space of players.141 scheme and its multiparty extension. In the mean time. we consider only the games with strategic form. whereas in mixed strategy the players randomize among their actions. The common knowledge assumption requires that all the players know the rules of the game and that all the players know that all players know the rules of the game. the strategic form. it is best to model using extensive form. A Brief Review of Classical Game Theory The ingredients of a game theory problem include:1 (i) Players: Two or more goal oriented. Although some quantum communication protocols and algorithms17. The extensive form requires more speciﬁcations such as which player moves when doing what and with what information. 2. and that all rational players search for an equilibrium on which all can settle down. In the ﬁrst one. Payoﬀ to a player is determined by the joint action of all players. quantum versions of some important theories. Then we will discuss Eisert’s scheme and focus on game dynamics with respect to the changes in the quantum mechanical resources used. if the players take their actions sequentially after observing the actions taken by the players before them. we will outline the results. On the other hand. have been developed. minmax theorem.
x) (d.142 Table 1. B2 }. b > d. there is an optimal way of taking an action independent of the actions of other players. Nash Equilibrium is a position where all the strategies are mutually optimal responses. In Fig. However NE concept. once it is reached.. y) (b. no individual has an incentive to deviate from it unilaterally. x > w. Alice (row player) and Bob (column player).. y > z. w > x. Twoplayer twostrategy game in strategic form. or even a reasonable solution for a game. This means that no unilateral change of strategy will give a higher payoﬀ to the corresponding player. The players. Games can also be classiﬁed according to the structure of their payoﬀ matrices as symmetric or asymmetric. z) each player pursues his/her own interests with no intention to negotiate. w) (c. Therefore. a constantsum game if a + w = b + x = c + y = d + z = K = 0 and a discoordination game if c > a. The payoﬀs are represented as (.. a zerosum game if a + w = b + x = c + y = d + z = 0. This payoﬀ matrix represents a symmetric game if a = w. . d = z. we give examples of classical games with diﬀerent types of dilemmas. Optimality and equilibria are the important solution concepts. Nash Equilibrium (NE) is the most widely used solution concept in game theory. c = x.) with the ﬁrst element being Alice’s payoﬀ and the second Bob’s.1. have the action sets {A1 . payoﬀ matrix of a generic twoplayer twostrategy game is given. z > y or a > c.1. zero sum or nonzero sum. In Table 1. d > b. and coordination (at least one equilibrium in pure strategies) or discoordination games. Bob:B1 Bob: B2 Alice: A1 Alice: A2 (a.20 . and (iii) NE need not be unique. This prevents a sharp prediction with regard to the actual play of the game. A systematic description of the outcomes in a gametheoretic problem is called a solution.. too. b = y. A2 } and {B1 . has diﬃculties: (i) Not every game has an NE in pure strategies. (ii) An NE need not be the best solution. A strategy is Pareto optimal (PO) if the only way each player has to increase his/her payoﬀ is decreasing someone else’s payoﬀ. It is merely a stable solution against unilateral moves by a single player. In some games.
and α and β are complex numbers satisfying α2 + β2 = 1. 3. the outcome will be  0 with probability α2 . and it will be  1 with probability β2 . A discussion of this measurement is given in subsection 3. . Quantum bits The state of a quantum system is described by a vector in Hilbert space. While the classical bits can be in one of two states 0 or 1 which can in principle be distinguished.143 (3.1) (0. Note that the normalization ensures that some result is obtained with probability 1. The entries pointed by two arrows and labelled with ∗ are the Nash equilibria.3) (0.1. .0) (1.1) (0. If a measurement is performed on the qubit. Prisoners’ dilemma (PD) where NE and Pareto optimality (PO) do not coincide. . Basic Concepts of Quantum Mechanics for a Study of Games In this section. Density matrices This is a frequently used formalism in quantum mechanics. .5) *(2.2. 3.1) * (1.0) (5. we will introduce some basic concepts of quantum mechanical toolbox needed for the study of quantum versions of classical games. and can be represented as  ψ = α 0 + β 1 where 0 and 1 are basis states (vectors). 1.0) (0.2) Samaritan s Dilemma Prisoners’ Dilemma Battle of Sexes Fig.  ψn } with probabilities pi .4.2) (1. Suppose we have an ensemble of possible states { ψ1 . a qubit is in a continuum of states described by α and β and it is not possible. Reference21 is an excellent resource for a detailed study of quantum mechanics and quantum information science. Horizontal (vertical) arrows correspond to the best responses of the column (row) player for a given strategy of the row (column) player. in general. Payoﬀ matrices of games with diﬀerent types of dilemmas: Samaritan’s dilemma (SD) with no Nash equilibrium (NE).0) * (1. The simplest quantum mechanical system has a two dimensional state space.  ψ2 . The Pareto optimal entry in the PD game is labelled with . to reliably distinguish nonorthogonal states of a quantum system.3) ♦ (3. 3. . and the Battle of Sexes (BoS) with two NE’s. This simplest system is called as quantum bit or qubit.
Unitary transformation The evolution of an isolated quantum system is governed by unitary operˆ ator U which takes the state  ψ at time t = t0 to the state  ψ at time ˆ t = t1 . Another important operator that we will frequently use is the Hadamard operator 1 1 1 ˆ . σz is the phase ﬂip operator ˆ which takes  1 to − 1 and does nothing on  0 . A density matrix may be formed from many diﬀerent ensembles. (1) It is important to notice that this is a mixture but not a superposition of  ψi ’s. The bit ﬂip operˆ ator σx transforms  0 to  1 .3..144 This ensemble is a mixture of all  ψi ’s and its density matrix is deﬁned as n ρ= ˆ i=1 pi  ψi ψi . The evolution of † ˆ ˆˆ a density matrix ρ is described by ρ = U ρU . σz = ˆ 1 0 0 −1 . The operators known as Pauli operators play an important role in quantum information. computation and games. 3.g. The σy Pauli operator ˆ performs both a phase ﬂip and a bit ﬂip which takes  0 to i 1 and  1 to −i 0 . No measurement can distinguish these two ensembles if no additional information is given about the preparation. ˆ ˆ A unitary operator on a qubit can be written as ˆ U = eiφ eiγ cos( θ ) eiδ sin( θ ) 2 2 −e−iδ sin( θ ) e−iγ cos( θ ) 2 2 . σy = ˆ 0 −i i 0 . the ensemble of  0 and  1 with equal probabilities and the ensemble of the superposition states √ √ ( 0 +  1 )/ 2 and ( 0 −  1 )/ 2 with equal probabilities have the same density matrix. These operators are σ0 = ˆ 1 0 0 1 . (4) . and  1 to  0 . e. but each of  ψi ’s may be a superposition state. (2) Unitary operators without the global phase eiφ belong to SU(2) operators set which will be mainly used in this study. H= √ 2 1 −1 √ √ which takes  0 to ( 0 +  1 )/ 2 and  1 to ( 0 −  1 )/ 2. This evolution can be represented as  ψ = U  ψ . (3) ˆ ˆ The identity operator I = σ0 leaves the qubits invariant. σx = ˆ 0 1 1 0 .
the probability of obtaining the measurement out† come is p1 = ψ M1 M1  ψ = ψ M1  ψ = β2 . In the same way. observe that the measurement operators are Hermi† † tian. then the post measurement state is  ψi = Mi  ψ ψ Mi† Mi  ψ (5) with the measurement result i occuring with probability pi = ψ Mi† Mi  ψ . Upon measuring the system  ψ with M . For some applications. For such situations. . Then the post measurement state becomes √  ψi = Pi  ψ / pi . The index i is associated with the measurement outcome . then the observable has a spectral decomposition M = i mi Pi . The probability of the measurement outcome is then given by pi = T r[ ψ ψ Ei ]. M0 = M0 and M1 = M1 . If the state is  ψ before measurement. we are not very much interested in the postmeasurement state but rather on the probabilities of the respective measurement outcomes. First. As an example let us measure the qubit  ψ = α 0 + β 1 in the computational basis which is described by two measurement operators M0 = 0 0 and M1 = 1 1. a mathematical tool called as POVM (positive operator valued measure) is used.145 3. The measurement operators satisfy the completeness equation i Mi† Mi = ˆ I. just as given in subsection 3. POVM is a collection of positive ˆ operators {Ei } with the completeness condition i Ei = I. If the projector onto the eigenspace of M with eigenvalue mi is Pi .4. The probability of getting the result mi is given by pi = ψ Pi  ψ . A special class of the general measurement described above is the projective measurements. A projective measurement is deﬁned by an observable which is a hermitian operator M acting on the state space to be measured. and they satisfy the completeness equation † † 2 2 ˆ M0 M0 +M1 M1 = M0 +M1 = M0 +M1 = I.1. The post measurement states then are the  0 for M0 and  1 for M1 . Measurement In order to gather information from a quantum mechanical system we have to perform measurement. Then the probability of obtain† ing zero as the measurement outcome is p0 = ψ M0 M0  ψ = ψ M0  ψ = α2 . we obtain an outcome mi which is an eigenvalue of M . Quantum measurements are described by a set of measurement operators {Mi } acting on the state space to be measured.
 N − m. respectively. Here  N − k. In addition to these classes. whereas GHZ and cluster states have one and N/2.g. there is the class of cluster states  χN which includes  GHZ3 but forms another inequivalent class for N ≥ 4. This is the quantum correlation or socalled entanglement. Starting from one of these bipartite MES. For a bipartite system. any bipartite state can be prepared using local operations and classical communication (LOCC). It has been shown that  WN is optimal in the amount of pairwise entanglement when N − 2 particles are discarded. 3 )/ 2 = ( 000 +  111 )/ 2. m / N Cm where N Cm . 1 / 3 = ( 001 + √ √ ( 3.24 The states  Ψ+ and  WN are the members of a more general class of √ symmetric states called as Dicke states.23 and has a persistency of N − 1. Thus. In entangled systems. it is needed to deﬁne the correlations in a quantum mechanical system. 0 +  0. there is another type of correlation which cannot be seen in classical systems. Although the total amount of entanglement contained in  WN is smaller than the other two classes. one cannot acquire all information on the system by accessing only some of the subsytems. 1 / N and the N particle Greenberger√ HorneZeilinger (GHZ) states  GHZN = ( N. An entangled state cannot be prepared by local interactions even if the preparers can communicate classically.5. A state is entangled if it cannot be written as a convex combination of product states. in the multipartite case (N ≥ 3).. there are four maximally entangled states (MES) √ which are deﬁned √ as Φ± = (00 ± 11 )/ 2 and Ψ± = (01 ± 10 )/ 2. 0 +  0. On the other hand.22. Unlike the uncorrelated system. Those inequivalent classes cannot be transformed into each other by LOCC. In quantum mechanical system. Correlations in quantum mechanical systems Analysis of quantum games presumes that the players initially share some kind of correlation but they are not allowed to have any type of communication. e. k corresponds to a symmetric state with N√ k zeros and k − √ 010 +  100 )/ 3 and  GHZ3 = ones.146 3. it has higher persistency of entanglement which is deﬁned as the minimum number of local measurements to completely destroy entanglement. N )/ 2. We start by giving the mathematical representation of an uncorreˆ ˆ ˆ ˆ lated state deﬁned as ρu = ρ1 ⊗ ρ2 ⊗ · · · ⊗ ρn−1 ⊗ ρn where measurement of ˆ observables on respective subsystems always factorize. it is known that there are inequivalent classes of states√ such as those represented by the N particle Wstates  WN =  N − 1. (i) (i) ˆ ˆ a classically correlated system is described as ρc = i pi ρ1 ⊗ · · ·⊗ ρn with ˆ i pi = 1. On the other hand.  W3 =  2. joint measurement of observables for such a system do not factorize.
superposition states. entanglement and quantum measurement. Coin ﬂip game The ﬁrst strong argument for a quantum game theory was presented by Meyer who introduced the coin ﬂip game between two Star Trek characters. If it is tails (T). we will present a short historical background of quantum game theory. Coin is initially prepared in tails state (T). P. Q wins. otherwise Q loses. Q P Flip / No Flip Q Flip / No Flip Coin: T Flip / No Flip { T.1. in the quantum version Q has the quantum strategy (opˆ erator) H while P has the classical strategies with the corresponding quantum operators ˆ ˆ {X. While in classical version strategy set players have the actions of Flip and No Flip. These schemes are Meyer’s scheme for coin tossing game15 and the Eisert’s scheme for Prisoners’ dilemma (PD). without knowing the action of Q. two quantization schemes which have been the milestones in this area of research. In the same way.147 denotes the binomial coeﬃcient. 4. coin is analyzed. Q loses (1) Coin: Q P Q T Q always wins (+1) Fig. 2. the coin is examined. Q and P. Q wins (+1) H.16 The transfer of the classical games into the quantum domain has been done using the resources allowed by the quantum mechanical toolbox which includes unitary operators. If the coin is tails. Classical (upper) and quantum (lower) versions of the coin ﬂip game. After the turns of P and Q. Then Q can choose either to ﬂip the coin or leave it unchanged without knowing the initial state of the coin. can choose either of the actions.25. Q . 0 and  0.15 In this simple game (see Fig.26 One can easily see that GHZ states can be written as a superposition of two Dicke states  N. and introduce. in detail. After the second turn of Q. 4. N . 2) P prepares a coin in the ”tails” state and puts it into a box. A Brief Review of Quantum Game Theory: Models and Present Status In this section. I}.
H}.2. they obtain a certain payoﬀ determined by the game payoﬀ matrix (see Fig. 1)‡ 4. 1)‡ (1.16 Before going into the quantization of this game. As can be seen from the game payoﬀ matrix in Table 2. No Flip = I} while Q has the quantum strategies ˆ ˆ {X. Table 2. In Meyers description of coin ﬂip game. −1) (−1.15 Eisert et al. a quantum coin with states deﬁned as T =  0 and H =  1 is used. 1)‡ (1. and P is restricted to classical ˆ ˆ strategies {Flip = X. 1 for the payoﬀ matrix). Then the game sequences as follows  0 −→ 2−1/2 ( 0 +  1 ) −→ 2−1/2 ( 0 +  1 ) −→  0 Q P Q ˆ H ˆ ˆ X or I ˆ H (6) from which it is seen that just applying a Hadamard gate before and after P’s action. Q:N N Q:N F Q:F N Q:F F P: N P: F (1.Prisoners’ Dilemma In the same year as Meyer’s work. On the other hand. Payoﬀ matrix for the classical coin ﬂip game where ‡ depicts the best strategy combination of Q for a given action of P. Eisert’s model . They then face a dilemma since as rational players they should look for an equilibrium which dictates that they should both defect. and the best the players can do is a mixed strategy. let us ﬁrst review the classical one: The socalled Prisoners’ Dilemma game is an important and wellstudied two player two strategy symmetric game where the players Alice and Bob independently without knowing the action of the other player decide whether to ”cooperate” (C) or ”defect” (D). 1)‡ (−1. −1) (−1. This work clearly shows that players with quantum strategies perform better than the players restricted to classical strategies. studied a quantum version of PD where two players sharing an entangled state use quantum strategies. there is no Nash equilibrium in the pure strategies. A classical coin clearly gives both players an equal probability of success. −1) (1. According to their combined strategy choices. Q brings the coin into the tail and hence wins the game regardless of the action of P. they would . −1) (−1.148 wins and if it is heads Q loses.
1} with n = 2j + corresponding to the projection onto the orthonormal basis {00 . (v) According to the measurement outcome n. 10 . ψ) = eiφ cos( θ ) eiψ sin( θ ) 2 2 −e−iψ sin( θ ) e−iφ cos( θ ) 2 2 (8) with 0 ≤ θ ≤ π. 1} .  10 . Schematic conﬁguration of Eisert’s scheme to play twoplayer strategic games. An entangled state is prepared by the referee using the entangling ˆ operator J. respectively for the players Alice and Bob. Alice and Bob carry out unitary operations UA . Their dilemma originates from the fact that NE and PO do not coincide. 0..  01 . (iii) After the referee receives the state. respectively. D)} where left and right entry represent Alice’s and Bob’s decision. the referee assigns to each player the payoﬀ an and bn chosen from the payoﬀ matrix of the original classical game. (D. (iv) The referee makes a projective measurement Πn = j j {j. UB ∈ SU (2) described by ˆ U (θ. 1. and is distributed to players. (ii) As their strategies. (C. In PD game. 3. 11 }. 01 . (D. and 0 ≤ ψ ≤ π/2.3} = {3. C). e. 0 ≤ φ ≤ π/2. both beneﬁt from cooperation which is the Pareto optimal.  11 } corresponds to {(C. =0. In Eisert’s model (see Fig.2. 5. the measurement outcome { 00 . he operˆ ˆ ˆ ˆ ˆ ates J † obtaining  Ψf = J † (UA ⊗ UB )J f g . After the action of the players referee performs measurement to distribute payoﬀs. Then the referee sends this state to two players. Alice and Bob.149 Entangled state preparation Disentangling and measurement Alice's action Referee Bob's action Referee Fig. g = 0. ˆ ˆ The strategy set of the players are unitary operators UA and UB .1.g. They then send their qubits to the referees.3). one qubit for each player. φ. a{n=0. a twoplayer twostrategy game is played in the quantum mechanical settings as follows: (i) A referee prepares the ˆ maximally entangled state  Ψi by applying the entangling operator J on a product state  f g where 1 ˆ Jf g = √ [ f g + i(−1)(f +g) (1 − f )(1 − g) ] 2 (7) with f. C). D).
iˆy ⊗ σ0 ] = 0 (11) ˆ These conditions imply that the unitary operator U (θ. [J. [J. and the setting can simulate ˆ the classical game faithfully. the probabilities Pj is found as P00 = x cos(φA + φB ) − y sin(ϕA + ϕB )2 P01 = x sin(φB − ϕA ) − y cos(φA − ϕB )2 P10 = x cos(φB − ϕA ) − y sin(φA − ϕB )2 P11 = x sin(φA + φB ) + y cos(ϕA + ϕB )2 (10) A B B A A B where x = sin θ2 cos θ2 . 0.3} = {3. the original Prisoners’ dilemma is reproduced. 0. In this model. when players’ strategy space is restricted to only oneˆ parameter unitary operator U (θ.150 and b{n=0. This equilibrium point gives 1 to each player which is less than the optimal one with 3 to each. x = cos θ2 cos θ2 . 0).1. Note that original classical games should be simulated in a model of its quantum version. 0. For PD game p(C) = cos2 (θ/2) and p(D) = 1 − p(C). resulting in the state  Ψf = (UA ⊗ UB ) 00 . σy ⊗ iˆy ] = 0. σ0 ⊗ iˆy ] = 0. when the players are allowed to use . On the other hand. y = sin θ2 cos θ2 . 5. Pj $B = n ˆ ˆ J) bn Tr(Πn J† ρf ˆ Pj (9) with ρf =  Ψf Ψf  and Pj representing the probability of obtaining ˆ the measurement outcome n. 0) as their strategies. and in Eisert’s model this is satisﬁed by imposing the following conditions: ˆ ˆ ˆ σ ˆ ˆ σ σ ˆ [J. In ˆ J this case. Hence when players’ strategies are restricted σ ˆ ˆ ˆ ˆ to U (θ). and y = θA θB sin 2 sin 2 . and the payoﬀ function σ takes the form of the classical payoﬀ function. 0) = cos(θ/2)ˆ0 + σ ˆ i sin(θ/2)ˆy and J commute. Then the average payoﬀ of the players can be written as $A = n ˆ ˆ J) an Tr(Πn J† ρf ˆ . the payoﬀ functions become identical to those of the classical Prisoners Dilemma with mixed strategies. when the players take U (θ.2. Therefore. the joint probability Pj factorize Pj = pA (j)pB ( ). the order of operations J and UA ⊗ UB can be changed and then ˆ ˆ ˆ and J † are cancelled. This implies that σ0 and iˆy ˆ correspond to the pure strategies C and D. respectively. 1} in the PD game for Alice and Bob. When the referee starts with the input state  f g =  00 . 0.
it is claimed that by introducing quantum toolbox. φ. the dilemma in this game disappears and players can achieve the Pareto optimal solution. 4. has been the widely accepted and studied quantization scheme despite some negative critiques as we have discussed above. their argument does not hold when a nonmaximally entangled state is shared between the players. criticized this restricted operation set to be unnatural and pointed out that if two players are allowed the full range of quantum strategies. Present status in quantum game theory It is important to realize that quantization of a classical game is not unique. 0). We should note that despite the critiques. Although the scheme is much simpler it does not allow probing the whole set of quantum strategies. they introduced two entangling parameters (one for J and ˆ† ) which can be tuned. If the two parameters are equal we the other for J ˆ obtain the Eisert scheme. 0).3. Cheon and Tsutsui introduced a simple and general formulation of quantum game theory by accommodating all possible strategies in the Hilbert space. there exists no Nash equilibrium in the game because the eﬀect of one player’s unitary transformation can be cancelled by the other locally using a counterstrategy due to the shared bipartite maximally entangled state.28 They constructed a threeplayer PD game and showed that the game has an NE even when the players are allowed to use full range of unitary operators. and when the parameter for J † is zero.27 This leads to a point that there is no compromised strategy. however. we end up with the Marinatto schemes. Another way of achieving results similar to those of Eisert’s scheme was introduced by Marinatto and Weber29 who removed the disentangling opˆ erator J † and simply hypothesized various initial states. Benjamin and Hayden have also examined three and four player quantum games which are strategically richer than the two player ones. Thus.151 ˆ twoparameter unitary operators U(θ. The model proposed by Eisert et al. and when the game is extended to multiplayer form. In this model. Nawaz and Toor30 proposed a more general model which includes both the Eisert and Marinatto models. and solved two strategy quantum . a unique Nash equilibrium with payoﬀ (3. However. ˆ In this model. Benjamin and Hayden. strategies are chosen as the probabilistic mixtures of identity and ﬂip operators.3). Later. emerges for their strategies ˆ U (θ = 0. one can introduce quantumness into a game in diﬀerent ways. which is also the Pareto optimal. the new dynamics brought into the games by quantum operators and entanglement in Eisert’s model is clear though sometimes results are rather surprising. φ = π/2.
is a general one that can be applied to any twoplayer twostrategy game.152 games. Multiplayer games exhibit a richer dynamics than the two player games. proposed a formulation of quantum game theory based on the Schmidt decomposition where they could quantify the entanglement of quantum strategies.37.34 Kay et al. transactions are de . While in the former. a mutant strategy with quantum tools can easily invade a classical ESS.33 In order to investigate the property of quantum version of dynamic games. Iqbal and Toor analyzed quantum version of repeated Prisoners’ dilemma game where the players decide to cooperate in the ﬁrst stage while knowing that both will defect in the second stage.32 The model proposed by Eisert et al. they derived a necessary and suﬃcient condition for the entangled states and quantum operations which are allowed in playing games in quantum mechanical settings. Chen et al. studied evolutionary quantum games and demonstrated that length of memory is crucial to obtain and maintain the quantum advantage.42–44 In their model of market.39 on the eﬀects of classical/quantum correlations and operators not only conﬁrmed the known eﬀects of using the quantum mechanical toolbox in game theory but also revealed that in some circumstances one need not stick to quantum operations or correlations but rather use simple classical correlations and operations.41 Piotrowski and Sladkowski have proposed a quantum game theoretic approach to economics. has proposed to play using only bipartite entangled states shared by neighboring players. Based on the requirement that classical game should be reproduced in its quantum model.38 Various quantum advantages over the classical have been found in the series of studies on games in quantum settings.36 He also showed that it is only when entanglement is employed. in the latter they share bipartite entanglement but armed with SU(n) operators to represent the players’ actions.31 Recently. Comparative studies by Shimamura et al.40. Ichikawa et al.20 and Ozdemir et al. While such games can be realized by preparing initially a multipartite entangled state. In these models the replicator dynamic takes a quantum form.35 Iqbal has also considered evolutionarily stable strategies (ESS) in quantum versions of both the prisoners dilemma and the battle of the sexes and concluded that entanglement can be made to produce or eliminate ESS’s while retaining the same set of NE. with the generalization to N player twostrategy games as well as twoplayer nstrategy games. the players share multipartite entangled state and use SU(2) operators.
Samaritan’s Dilemma . There has been a growing interest in the quantum version of the wellknown Parrondo’s games. Here. The amount of entanglement between the players was varied and the transition between classical and quantum regions were clearly observed. Flitney and Abbott have developed models and studied the history dependent quantum versions of these games. Detailed and excellent reviews of quantum game theory can be found in the references.48 Physical realizations of Eisert’s scheme of PD have been achieved ﬁrst in a two qubit NMR computer49 and recently in an alloptical oneway quantum computer. The quantum circuit was realized by a 4qubit boxcluster conﬁguration. Study of a Discoordination Game . Therefore.51 This scheme is a hybrid one combining the quantum circuit approach and the cluster state model. we gave only a short summary of what has been done in the ﬁeld. Games with no NE are interesting because they represent situations in which individual players might never settle down to a stable solution. which is a discoordination game with no Nash equilibrium in classical pure strategies. They also speculated that markets will be more eﬃcient with quantum algorithms and dramatic market reversals can be avoided.40.50 The NMR experiment followed the original proposal by Eisert.52–54 5. and the local strategies of the players by measurements performed on the physical qubits of the cluster. the important step in such games is to ﬁnd an NE on which the players can settle down.Using Eisert’s Model: Eﬀects of Shared Correlation on the Game Dynamics Samaritan’s Dilemma (SD). The alloptical realization was performed following the scheme of Paternostro et al.153 scribed as projective operations on Hilbert spaces of strategies of traders. In particular we will see how the game payoﬀ matrix and the strategies of the . and a quantum strategy is a superposition of trading actions and hence can achieve outcomes not realizable by classical means.46 where two games that are losing when played separately become a winning game when played in combination (their convex combination). In the following.13. It is expected that Parrondo games have the potential use in increasing the eﬃciencies of quantum algorithms. we will investigate this game under diﬀerent settings of Eisert’s model.45.47. and showed good agreement with theory. is an interesting game where one can systematically analyze the changes in the game dynamics for various quantum mechanical settings.
W ). 3. 1) studied in this paper are taken from Ref.4.2 corresponds to the NE for the game with average payoﬀs given as $A = −0. in this case. the “Samaritan” (Alice) voluntarily wants to help the person ”in need” (Bob) who cannot help himself. In the game of SD. No Aid (N)} whereas Bob’s is {Work (W). as shown with arrow in Fig.L) and (N. whereas for the latter case Alice will switch to A to increase her payoﬀ from −1 to 3. she suﬀers a welfare loss due to the selﬁsh behavior of Bob who may inﬂuence or create situations which will evoke Alice’s help. (A. In mixed classical strategies.2 and $B = 1. (A. The action set of Alice is {Aid (A). Moreover. L).L) emerge as the most probable strategies with probabilities 0.W) are not NE either. (N. In this game. When this game is played classically. there is no dominant strategy for neither of the players and there is no NE for pure strategies: (A. Alice (row player in the game matrix) wishes to aid Bob (column player in the matrix) if he searches for work but not otherwise.5. Alice randomize between her strategies A and N with the probabilities p and (1 − p). However.5 and q = 0. Bob’s chooses between W and L with probabilities q and (1 − q). the payoﬀ of Alice is negative which is not a desirable result for her. In the study of this game. Bob searches for work if he cannot get aid from Alice. Alice will switch to N.55 The game and the payoﬀ matrix (see Fig. L). this game has no NE when played with pure classical strategies. In this case.154 players evolve for diﬀerent strategy sets and shared correlation between the players.1 where the speciﬁc game is named as The Welfare Game. Then the payoﬀs for Alice and Bob is written as $A = 3pq − p(1 − q) − q(1 − p).L) and (N. On the other hand. Bob acts so strategically that the amount of help increases which is not desirable for Alice.W) is not an NE because if Alice chooses A. Therefore. Bob can respond with strategy L where he gets a better payoﬀ. Then. Loaf (L)}. (12) The strategy combination with p = 0.L) is not an NE because. 1. The strategies (N. however. a dilemma arises because Alice wants to help. Alice cannot retaliate to minimize or stop this exploitation because doing so is a punishment for Bob and this will harm Alice’s own interests in the short run. we ﬁrst look at the problem whether a unique . In the same way. because for the former one Bob will switch to W to get payoﬀ one. respectively. A more detailed analysis can be found in References 39 and 40. and (N. W ). (A. Thus their strategy combinations can be listed as (A. $B = 2pq + 3p(1 − q) + q(1 − p).
Moreover. there are three cases: CASE I. If there exists at least one NE then we look at how the payoﬀ of the players compare and to what extent this NE strategy can resolve the dilemma. 5. insuﬃcient solution with $A < 0. we see that while $B is always positive.1. CASE II.2 and $B = 1. It is seen that this strategy set reproduced the results of the classical mixed strategy.Operators for Alice and Bob are obtained by setting φA = ϕA = 0 and φB = ϕB = 0 in Eq. This strategy set resolves the dilemma. the strong solution solution with 0 ≤ $B < $A . both players receive positive payoﬀs then this implies both are satisﬁed with the outcome and there is no loss of resources for Alice (CASE II and III). we know from Eisert’s study that operators chosen from three parameter SU(2) set do not resolve the dilemmas in the games. but with an insuﬃcient solution for Alice (CASE I). the payoﬀs for Alice and Bob become 1 $A = [1 + 3(cos θA + cos θB ) + 5 cos θA cos θB ].5. thus Bob’s choice of action becomes important to resolve Alice’s dilemma. The most desirable solution of the dilemma for Alice is represented in CASE III. Alice. an NE is found for UA = (ˆ0 + iˆy )/ 2 and √ ˆ σ σ UB = (2ˆ0 + iˆy )/ 5 with the payoﬀs $A = −0. For SD. in the following we will not study these two √ cases. the payoﬀs of the players become $A = −0.155 NE exists for strategies utilizing classical or quantum mechanical toolboxes. the situation produces the classical game with mixed strategies. At this unique NE. . After some straightforward calculations.(13).5.(8). $A may be positive or negative depending on Bob’s action. cannot make her payoﬀ positive by acting unilaterally. Oneparameter SU(2) operators . Quantum operations and quantum correlations When the SD game is played with quantum operations but with no shared correlated state between the players. weak solution with 0 ≤ $A ≤ $B .2 and $B = 1. (13) 2 From Eq. This implies that the samaritan. Therefore. respectively. With the shared entangled state  ψ1 and the oneparameter SU(2) operators. and does not bring any quantum advantage. 4 1 $B = [3 + 2 cos θA − cos θA cos θB ]. and CASE III. when they share ψ2 . We will consider two types of MES: ψ1 = [00 + i11 ]/ 2 and √ ψ2 = [01 − i10 ]/ 2 obtained when the referee applies the entangling ˆ operator J on the initial states 00 and 11 . On the other √ ˆ σ σ hand. we ﬁnd that an NE emerges for the operators √ √ ˆ ˆ σ σ σ σ UA = (ˆ0 + iˆy )/ 2 and UB = (ˆ0 + i2ˆy )/ 5. If at the NE.
156
Twoparameter SU(2) operators  This set of quantum operations is obtained from Eq. (8) by setting ϕA = 0 and ϕB = 0. For ψ1 , the expressions for the payoﬀs are found as $A = 3P00 −P01 −P10 and $B = 2P00 +3P01 +P10 with P00 = cos2 (θA /2) cos2 (θB /2) cos2 (φA + φB ), P01 = x sin φB − y cos φA 2 , P10 = x cos φB − y sin φA 2 (14)
where Pj is calculated from Eq. (9). We ﬁnd that there is a unique NE ˆ ˆ σ which appears at UA = UB = iˆz with the payoﬀ ($A , $B ) = (3, 2). Note that this NE is a new one which cannot be seen in the classical version of the game. Since, this new NE emerges in pure quantum strategies we conclude that original discoordination game transforms into a coordination game. This is the NE where Alice receives the highest payoﬀ available for her in this game, and where both players beneﬁt. Therefore, Alice’s dilemma is resolved in the stronger sense (CASE III). In the case of ψ2 , four NE’s with equal payoﬀs ($A , $B ) = (3, 2) ˆ ˆ σ σ emerge for the quantum operators (UA , UB ) equal to (N, P) = (iˆy , iˆz ), √ √
(ˆ +iˆ ) i(ˆ +ˆ ) σ σ σ σ σ +i 3ˆ ˆ σ i( 3ˆ +ˆ ) σ σ
z y √ ), and (Z, S) = (T, Q) = 0√2 y , z 2 y ), (Y, R) = 0 2 y , 2 σ ˆ σ ˆ (γ0 (ˆ0 + iγ1 σy ), iδ0 (ˆz + δ1 σy )) where γ0 = cos(3π/8), γ1 = tan(3π/8), δ0 = cos(π/8), and δ1 = tan(π/8). At these NEs players receive payoﬀs higher than those when a classical mixed strategy is used. However, Alice’s dilemma survives because players cannot coordinate their moves to decide on which NE point to choose. For example, if Alice thinks Bob will play σ iˆz then she will play iˆy to reach at the ﬁrst NE. However, since this is a σ simultaneous move game and there is√ classical communication between no ˆ the players, Bob may play i(ˆz + σy )/ 2 (because this action will take him σ to the second NE point) while Alice plays iˆy . Such a case will result in the σ case $A < $B and will lower the payoﬀs of both players. Therefore, still a dilemma exists in the game; however the nature of dilemma has changed. In the classical version, the dilemma is due to welfare losses of Alice and lack of NE. However, now the dilemma is the existence of multiple NEs. Implications  Now let us review what we have learned from the above analysis:
(1) Twoparameter SU(2) operators and shared entanglement are necessary to see the novel features introduced into game theory by the quantum mechanical toolbox. Note that oneparameter SU(2) operator set with shared entanglement corresponds to the classical mixed strategies.
157
(2) The number of solutions that is the number of NE’s depends on the shared entangled state as well as on the payoﬀ matrix of the original classical game. In SD game, switching from the entangled state ψ1 to ψ2 increases the number of NE from one to four. On the other hand, for the PD game, the number of NE’s is one for both cases. (3) The quantum mechanical toolbox does not necessarily resolve dilemma in classical games. Note that for the PD and SD played with ψ1 the dilemmas in the original games are resolved, however in SD played with ψ2 , a new type of dilemma appears. While in the original game the dilemma was the absence of an NE where Alice is not exploited, in the quantum version played with ψ2 the dilemma is the existence of multiple NE’s with the same payoﬀs. This new dilemma is similar to that of the Battle of Sexes (BoS) game, that is Alice and Bob cannot decide on which of the multiple NE’s to choose. (4) Quantum version of a classical game can be represented by an enlarged classical payoﬀ matrix by including the unitary operators leading to equilibrium solutions into the payoﬀ matrix of the initial classical game.56 The originally 2 × 2 classical games of PD and SD are transformed into new games which can be described with the new 3×3 payoﬀ matrices. In case of SD, another representation is the 5 × 6 payoﬀ matrix obtained for ψ2 . Evolution of the payoﬀ matrix of the SD game under diﬀerent correlated states is depicted in Fig. 4. Note that the new payoﬀ matrix includes the payoﬀ matrix of the original classical game as its submatrix. (5) An important quantum advantage arises when we consider the communication cost of playing games. For a game represented by the 2 × 2 payoﬀ matrix, each player needs 1 classical bit (cbits) to classically communicate their preferred strategy to the referee. Thus, the total classical communication cost is 2 cbits. When we play the quantum version of the game, the task is completed using four qubits (two prepared by the referee and distributed, and two qubits sent by the players back to the referee after being operated on) when shared entanglement is used. This corresponds to a communication cost of 2 ebits (1 ebit for the referee and 1 ebit for the players). Now, if write down the payoﬀ matrix for this quantum version, we end of with 3 × 3 payoﬀ matrices for the PD and SD with ψ1 , and a 5 × 6 matrix for SD with ψ2 . The cost of communicating players’ choices in these new payoﬀ matrices classically (playing the game represented with those matrices) amounts to respectively to 4cbits and 6cbits. However, in quantum
158
versions what is needed is still 2 ebits. (6) In classical mixed strategies, the payoﬀs become continuous in the mixing probability of the players’ actions, and a compromise becomes possible between the players which assures the existence of an NE. In games played by quantum operations, the inclusion of new strategies and moves into the game is an essential feature. Once these new moves are incorporated, both the strategy space and the payoﬀs become continuous. Then the players can ﬁnely adjust their strategies and come closer to ﬁnding the best responses to each other’s actions which leads to one or more NE’s.
5.2. Quantum operations and classical correlations The shared entanglement introduced in the quantum version of the SD game transforms the original 2 × 2 noncooperative game into a cooperative one. This new game thus has its own rules which are diﬀerent than the original game: While in the original classical game, no communication of any form is allowed between the players, in the quantum version the players share entanglement which could be considered as a kind of “spooky communication”.57 In order to clarify further the role of the operators and correlations in the obtained results above, we now consider that the players share classical correlations. This will make it easier to evaluate the power of quantum operators. Here, we consider that classical correlations are due to phase damping acting on the MES during its distribution (see Fig. 5). Phase damping transforms the initially prepared MES ψ1 and ψ2 into the classically correˆ lated states ρin,1 = [00 00 + 11 11]/2 and ρin,2 = [01 01 + 10 10]/2, ˆ respectively. In the following, we will see how these correlations aﬀect the games. Oneparameter SU(2) operators  It is straightforward to show that when the players are restricted to oneparameter SU(2) operators, they get √ ˆ ˆ σ σ the payoﬀ (0.25, 1.5) at the NE with UA = UB = (ˆ0 +iˆy )/ 2 independent of the classical correlation they share. Alice’s payoﬀ in this case is much better than the one she receives when the players share MES and use oneparameter SU(2) operators. Twoparameter SU(2) operators  In this new setting of the problem,
(a)
(b) (c)
(d)
Fig. 4. Evolution of the payoﬀ matrix of Samaritan’s dilemma (SD) game for shared quantum and classical correlation between players using twoparameter SU(2) operators: (a) and (b) for quantum correlations  ψ2 and  ψ1 , respectively, and (c) and (d) for classical ˆ correlations ρin,1 and ρin,2 , respectively. These payoﬀ matrices are obtained by adding to the original game matrix (2 × 2 payoﬀ matrix ˆ in the middle) the strategies leading to Nash equilibria (NE). Refer to the text for the operators corresponding to the strategies shown by letters in these payoﬀ matrices.
159
It is easy √ ˆ to show that an NE with a payoﬀ (0.5) appears for UB = (ˆ0 +iˆy )/ 2 σ σ and (φA = 0 and {∀θA : 0 ≤ θA ≤ π}).25.3 = [00 00 + 10 10 + 01 01 + 11 11]/4.1 .5). Then the operators UA = (ˆ0 + iˆy )/ 2 and √ ˆ σ ˆ UB = i(ˆy + σz )/ 2 leads to an NE with the payoﬀ ($A . regardless of their quantum operators players arrive at a weak solution (CASE II) with constant payoﬀ (0.) term. This is a selfforcing NE.160 channel noise Referee Referee Referee Fig. Threeparameter SU(2) operators . too. For this action of Alice. 1. the payoﬀ functions can be obtained from Eq. for ρin. This NE provides a weak solution (CASE II).It is only when the players share a MES that the threeparameter operator set does not give quantum advantage. 2.25..5). Another possible correlation that may be shared by the players is the full rank classical correlation ρin. Quantum version of a classical game played with classical correlations ρin obˆ tained from an initial maximally entangled state (MES) after it is subjected to noise. For the shared classical correlation of ρin. It is then interesting to discuss whether this set of SU(2) operators aﬀect the game dynamics in the case of shared classical correlation. (15) by replacing all the − by + and vice verse except √ ˆ σ σ the + in the sin(. the dilemma is again resolved with a weak solution (CASE II).1 the payoﬀs become ˆ 1 [1 + 5 cos θA cos θB − 3 sin θA sin θB sin(φA + φB )]. This is a selfenforcing NE solution for the rational players Alice and Bob because neither of the players can make her/his payoﬀ arbitrarily high and positive independent of the action of the other player. the rational Bob will go for the above operator which maximizes his payoﬀ. $B ) = (1. 4 1 $B = [3 − cos θA cos θB − 2 sin θA sin θB cos φA sin φB ] (15) 2 from which we see that Alice cannot obtain a positive payoﬀ by acting unilaterally so she should look for settlement at an equilibrium. 1.2 . ˆ In this case. the the payoﬀ functions are ˆ obtained from Eq. (15) by replacing φA ˆ $A = . Then. because Alice assures positive payoﬀs for both Bob and herself regardless of Bob’s action. For ρin. 5.
5. $B = 2. Note that even though the payoﬀs may increase. The solution with the considered classical correlations is weak (CASE II). the dilemma in SD is ρ not resolved in the stronger sense (CASE III). Classical operations and classical correlations Classical operations are a subset of quantum operations.161 and φB by φA + ϕA and φB + ϕB . φA + ϕA = 0) and (θB = π/2.3 ). On the other hand. Note that the most cited criticism against quantum versions of classical games is that restricting the players to twoparameter SU(2) operators is not reasonable. φA + ϕA = π) and (θB = π/2.5) emerges for the operators (θA = π/2. φB + ϕB = π/2). Here ˆ . an NE with payoﬀs ($A = 1. Now it is possible to get rid of this. φB + ϕB = π/2). By decreasing the amount of shared entanglement. Then. The extra parameter ϕ allows the players to increase their payoﬀs beyond those obtained with twoparameter SU(2) operators. Note that with twoparameter SU(2) operators and shared MES ψ2 .1 . the solution is still a weak solution (CASE II). $B = 2. Shared classical correlations between players with quantum operations leads to NE solutions which provide better payoﬀs for Alice compared to those she obtains for classical mixed strategies.2 . (2) Although a solution can be provided with the classical correlations ˆ ˆ considered in this study (ˆin. we ﬁnd that the strateˆ gies of the players to achieve an NE becomes (θA = π/2. (3) Three parameter SU(2) operators which cannot be used with shared MES can be utilized to ﬁnd an NE with better payoﬀs. quantum operators corresponding to classical strategies are the identity σ operator σ0 for A and W . players obtain a strong solution. ρin.Our ﬁndings in this section on the eﬀects of classical correlations on the outcome can be summarized as follows: (1) Entanglement is not necessary to ﬁnd an NE that may solve the dilemma of the original game. For the SD game. Implications . These operators give them the payoﬀs ($A = 1.2 and ρin. the players can exploit whole SU(2) without any restriction. for ρin. and in Eisert model ˆ they should commute with the entangling operator J. It is important to keep in mind that this conclusion strongly depends on the payoﬀ matrix and the dilemma type of the original classical game. respectively. and the ﬂip operator iˆy for N and L .5) which are the same as they obtained with twoparameter operators.3.
This is interesting because an NE appears where both Alice and Bob have positive payoﬀ. 2) (−1. in Eisert’s scheme.25. When the classical correlation is ρin. the payoﬀ matrix has the same properties as in Table ˆ 3 with diagonal and oﬀdiagonal elements interchanged. implying that the dilemma is resolved in the weaker sense (CASE II).162 Table 3. now we see that there is an increase in Alice’s payoﬀ from −0. 2) ↓ (1. and then makes a projective measurement . Note that this was not possible for the classical game played without correlations.5).2 with payoﬀ (0.5 and q = 0. 1) ↑ (−1. and use ˆ classical operations. Bob: σ0 ˆ → ← Bob: iˆy σ Alice: σ0 ˆ (1. we conclude that classical correlation does not change the structure of the SD game except the scaling of the payoﬀs when players choose to play in classical pure strategies.5. the diagonal and oﬀdiagonal elements of the ˆ payoﬀ matrix of the original game (see Fig. 1) are averaged out separately. that is they randomize between their pure strategies.2 to 0. The values of p and q are the same as those in the classical mixed strategy of the original game without any shared correlation.2 .1.25. From Table 3.1 . The new payoﬀ matrix for the SD game when players share the classical correlation ρin = [00 00 + 11 11]/2. But since the payoﬀs were scaled with the introduction of the classical correlation. the classical operators of the players should commute with the entangling operators. when they switch to mixed strategies. resulting in a new game with the payoﬀ matrix given in Table 3 without any correlation. it is seen that the game is still a discoordination game with no NE and the dilemma survives.5. 1) Alice: iˆy σ we consider the classical correlations as in the above subsection. This requirement reduces the game using classical operators to the case where players apply their operators to the initial product state from which the referee prepares the maximally entangled state. Thus. However.4. Classical operations and quantum correlations As pointed out above. 5. we ﬁnd that there is a unique NE when p = 0. When the classical correlation is ρin.
5)√ ˆ σ σ sponding to UA = (ˆ0 + iˆy )/ 2 and Bob chooses p = 0. thus one may wonder whether Alice can do better and resolve her dilemma if Bob is restricted to classical strategies. Classical Bob versus quantum Alice The dilemma in SD game is a onesided dilemma of only Alice. 1. . $B ) = (−0.2. bn (pPj + (1 − p)Pj ). 10 .5. In this case. (16) Then average payoﬀs are calculated as $A = n an (pPj + (1 − p)Pj ). Therefore. and Alice is allowed to use quantum operations from SU(2) set when they share the MES. 2 2 (18) Oneparameter SU(2) operators . 2 2 2 θA 2 θA 2 = (1 − p) cos ( ) cos φA + p sin ( ) sin2 ϕA .163 onto the four basis {00 . ˆ1 ˆ0 ˆ σ Bob’s operators will be UB = σ0 with probability p and UB = iˆy with probability 1 − p. This leads to the fact that there is no NE where both players can settle. This is not very desirable for Alice due to her negative payoﬀ while Bob’s is positive (CASE I). 01 . Bob can ﬁnd a p value that will minimize Alice’s payoﬀ and vice verse. 2 2 2 2 θA 2 θA = (1 − p) cos ( ) sin φA + p sin ( ) cos2 ϕA . 11 } to calculate their payoﬀ.Setting ϕA = 0. This is the same result as the case for quantum Alice and Bob with oneparameter SU(2) operators. The state after the actions of Alice and Bob becomes 1 ρout = ˆ k=0 ˆ ˆk ρ ˆ† ˆ k† pk (UA ⊗ UB )ˆin (UA ⊗ UB ). ψ1 . results of classical mixed strategies are reproduced.2. For any strategy of Alice. an NE with emerges when Alice chooses θA = π/2 correpayoﬀs ($A . we see that Alice can never achieve a positive payoﬀ independent of Bob’s strategy deﬁned by p. 5. n $B = where P00 = p cos2 ( P01 P10 (17) θA θA ) cos2 φA + (1 − p) sin2 ( ) sin2 ϕA .Setting φA = ϕA = 0. Twoparameter SU(2) operators . This is the same as the original classical game played in pure and mixed strategies.
The situation changes dramatically in this case because Alice now has an extra parameter to conˆ σ σ ˆ ˆ trol the outcome. In this section. Thus Alice can control the game but this does not let her to resolve the dilemma with a strong solution. we consider a corrupt source and analyze its eﬀects:59 (i) Is there a critical corruption rate above which the players cannot maintain their quantum advantage if they are unaware of the action of the noise on the source. and (ii) How can the players adopt their actions if they have information on the corruption rate of the source. When she chooses UA = [ˆ0 + i(ˆx + σy + σz )]/2 or √ ˆ σ σ UA = (ˆ0 + iˆx )/ 2. 2. Since the initial state from which the referee prepares the entangled state is a crucial parameter in Eisert’s scheme.59 In the comparison of the classical and quantum versions with and without corruption in the sources. we assume that in the quantum version the players choose their optimal operators from the twoparameter SU(2) operators set . Here. we assumed that the sources used in the realization of the games were ideal and free of noise but the distribution channel induced phase damping.5). Therefore. Even in this case. the analysis of situations where the source is corrupt is necessary to shed a light in understanding the game dynamics in the presence of imperfections. In the previous sections. the payoﬀs become independent of Bob’s action. and study its eﬀect on PD and SD games. and physical schemes are usually far from ideal in practical situations. any deviation from the ideality of the source will change the dynamics and outcomes of the game. Decoherence and Quantum Version of Classical Games It is wellknown that quantum systems are easily aﬀected by their environment.25.164 Threeparameter SU(2) operators . The ﬁrst of such studies was carried out by Johnson58 who has considered a threeplayer PD game and shown that a crossover from quantum to classical advantage occurs as the corruption rate increases. 6). we consider a more general corrupt source model (see Fig. It is thus understood that by acting unilaterally Alice can resolve her dilemma only when she is allowed to use operators chosen from threeparameter SU(2) set and Bob is restricted to classical operations. only a solution satisfying CASE II is achieved. 1. where the initial state input to the entangler in Eisert’s model is ﬂipped with some probability. Consequently. The best she can do is a weaker solution (case II) with the corresponding payoﬀs (0. it is important to study whether the advantage of players and dynamics of games with quantum operations and entanglement survive in the presence of noise and nonideal components. 6.5) and (1.
the classical game is better. while in the classical version they apply either σ0 or iˆy . Then the payoﬀs with classical and quantum strategies become equal to 2. Then the combined state ˆ ˆ generated and sent to the entangler becomes ρ1 ⊗ ρ2 = (1 − r)2 00 00 + ˆ r2 11 11+r(1−r)(01 01+10 10). however. they deterministically prepare the spindown state 0 resulting in f g = 0 0 . Otherwise. In ideal case. The corrupt sources. This results in a mixture of the four possible maximally entangled states (1 − r)2 ψ + ψ +  + r2 ψ − ψ −  + r(1 − r)(φ+ φ+  + φ− φ− ). then the players would rather apply their classical strategies than the quantum ones. becomes a mixture of spinup and spindown states. in the classical version the game is played with no entanglement.25 when r = r = 1/2. Then a quantum advantage is observed in the payoﬀs and dynamics of the game: an NE coinciding with the Pareto optimal is obtained. where ψ ∓ = 00 ∓ i11 and φ∓ = 01 ∓ i10 . we see that with increasing corruption rate r the payoﬀs in the quantum version decrease while those in the classical version increase. Two identical sources S1 and S2 prepare the initial product state f g . Then the players apply . ˆ σ The state prepared by the corrupt source model shown in Fig. Thus.165 Fig. we know that the best strategy of the players ˆ ˆ ˆ ˆ σ σ is to apply UA = UB = iˆy in the classical version. and they apply the optimal strategies to resolve their dilemma based on the assumption that the source is ideal and the referee distributes the MES ψ + . respectively. Alice and Bob are not aware of the corruption in the source. if r > r . In quantum version. the input state ˆ of the entangler. For PD with ideal sources. players receive $A = $B = 3 which is better than the payoﬀ $A = $B = 1 in classical version. 6. 6 can ˆ be written as ρ1 = ρ2 = (1 − r)0 0 + r1 1. In the corrupt source model. the quantum version of the game always does better than the classical one. The players assume ideal source while it is not . This is the state on which the players will perform their unitary operators. This can be explained as follows: Since classical operators commute with the entangling operator. Model of the corrupt source. Therefore.In this case. denoted by J. and UA = UB = iˆz in the quantum version. prepare the spinup and down states 1 and 0 with probabilities r and 1 − r. If r satisﬁes 0 ≤ r < 1/2.
In the quantum version with ideal sources. does this NE resolve their dilemma in the game or not. the state shared between the players become ρ = I/4.7.2 at r = 0. Bob uses a biased randomization and applies σ0 ˆ with probability 0. then they operate σ on  00 and receive $A = $B = 1. the optimal strategy is UA = UB = iˆz . we know that the most desired solution ($A . and informs the players on r. $A reaches its minimum of −0. For these corrupt sources. they receive the payoﬀs 96/49 and 0. $B ) = (3. In the quantum version. they operate on  11 and receive $A = $B = 3. It is worth noting that when the players apply their classical mixed strategies in this physical scheme. respectively. CASE II when 1/7 ≤ r < 0. however it will not be optimal when r = 1. For the symmetric game PD.166 iˆy directly on the state prepared by the source. while Alice randomizes between her operators with equal probabilities. The question now is whether given r the players can ﬁnd a unique NE. The results of the analysis are depicted in Fig. $B ) = (−0. The dynamics of the payoﬀs in this game σ with the corrupt source when the players stick to their operators iˆz and its σ comparison with their classical mixed strategy are depicted in Fig. both players receive the . when the players apply the classical operator iˆy . whereas $A increases linearly for 0 ≤ r ≤ 1. On the other hand. It is observed that the transition from the quantum advantage to classical advantage occurs at r = 1/2 for both players. when r = 0. In this strategy. we know that the best strategy for the players in the classical version is to use a mixed strategy which delivers ($A . $B is always constant and independent of the corruption rate. Then independent of what action they choose.2. 1. and if they can. While for increasing r. In this asymmetric game the payoﬀs of the players become equal for the corrupt sources when r = 1/7 and r = 1. where it starts increasing to the value of zero at r = 1. if r = 1. ˆ ˆ σ on the other hand. 2) is obtained at the NE when both players apply iˆz to ψ + . If r = 0.8. their payoﬀs continuously increase from one to three with the σ increasing corruption rate from r = 0 to r = 1. If the players insist applying iˆz when r = 1 then they will end up with the payoﬀs of the classical version. In SD.We assume that the referee knows the characteristics of the corruption in the source. 7. The optimal strategy in this case σ is when θA = θB = 0 and φA = φB = π/4.5) at the NE. Thus.2. and ﬁnally CASE III for r < 1/7. $B monotonously decreases from two to zero. the ˆ ˆ players receive constant payoﬀs determined by averaging the payoﬀ entries of the game matrix.6. Players know the corruption in the source . We identify three regions of r corresponding to the three classes of solution to SD game: CASE I when 0.6 ≤ r ≤ 1. One can easily see that for r = 1/2.
r in the source. transition point from $A > $B to $B > $A (Case III ↔ Case II).2 0. for their quantum. In the PD. quantum advantage no longer survives if the corruption rate is above a critical value.The study on the corrupt source model reveals that: (1) In a game with corrupt source with no side information to the players.2 0. For PD . (2) The strength of corruption may not only decrease the payoﬀs but also may lead to multiple NE’s depending on the game even if the players know the corruption in the source. For the completely corrupt source r = 1. For Bob. The payoﬀs received by the players in Prisoners’ dilemma (PD) and Samaritan’s dilemma (SD) as a function of corruption rate.8 1 Corruption Rate (r) Corruption Rate (r) Fig. There are a number of NE’s with equal payoﬀs which prevents players making sharp decisions.6 0. The horizontal dashed line denotes the payoﬀs of both players for the classical strategy without corruption. the labelled points are as follows: a.8 1 0. payoﬀ 9/4. the point labelled as a at r = 1/2 is the transition from quantum advantage to classical advantage. i. they can either adapt to new optimal operators or choose a riskfree action. they should choose √ ˆ ˆ σ σ UA = UB = (ˆ0 + iˆz )/ 2.4 0. transitions from quantum advantage to classical advantage for Bob and Alice. b and c at r = 1/2. respectively. and classical strategies.5 1 0. 7.5 a 2 a Quantum Bob Classical Bob Payoff 2 1. for the asymmetric game SD the payoﬀs become 1/4 and 3/2 respectively for Alice and Bob.0. Implications .5 Classical Alice b c 0 . In SD. (3) If the players are given the corruption rate in the source.167 Prisoners' Dilemma 3 Quantum 2.e. classical strategies with and without noise coincide and are depicted with the horizontal dotted line. We see from Table 4 that if the players know r then they can solve the dilemma for both r = 0 and r = 1 obtaining the payoﬀs 3.5 0 0.5 Alice & Bob Samaritan's Dilemma 3 Quantum Alice 2.4 0. This is however not true for the SD game: for r = 0 the dilemma is resolved with the best possible solution (CASE III) contrary to the situation for r = 1 where there is no unique NE (see Table 5).6 0. The horizontal solid line denotes the payoﬀ Alice receives in classical strategies when the source is ideal.5 1 0.5 0 Classical Alice & Bob Payoff 1.
π ) 2 π ) 2 (0. φB ) ($A . Strategies leading to NE and the corresponding payoﬀs in the SD game for the source corruption rate r. (0. ( 21 . Flitney et al. ( 43 . 2) ˆ ˆ and SD games.60 In a recent study. r=0 r = 1/4 r = 3/4 r=1 ˆ UA (θA . and showed that the NE payoﬀ is reduced as the decoherence is increased while the NE remains the same provided the decoherence probability is less than 1/2. φA ) ˆ UB (θB . $B ) (0. (0. π ) 4 π ) 4 (3. Strategies leading to NE and the corresponding payoﬀs in the PD game for the source corruption rate r. φ)1 (0. 2) ( 21 . 1 ∀φA . π ) 2 π ) 2 (0. (0.61 The decrease rate in payoﬀs is much higher as the number of players in the game increases. 16 (3. 16 π ) 2 π ) 2 15 ) 8 (0. π 2 (0. ∀φB ∈ [0. (0. π 2 − φ)1 15 ) 8 − φ)2 (3. ( 43 . φB ) ($A . 3) (3. $B ) (0. Flitney and Hollenberg studied the Minority game in the presence of decoherence. (0. 16 π ) 4 π ) 4 43 ) 16 (0.168 Table 4. they may choose the operators UA = UB = (ˆ0 + σ √ iˆy )/ 2 where the payoﬀs of the players become constant independent σ of corruption rate. 16 π ) 2 π ) 2 43 ) 16 (0. 3) Table 5. there is a large potential loss if the source is deviated from the ideal one. (0. showed that the quantum player maintains an advantage over a player with classical strategies provided that some level of coherence remains. φ)2 (0. φA ) ˆ UB (θB . π/4] r=0 r = 1/4 r = 3/4 r=1 ˆ UA (θA . They also . It is worth to note here that with a decoherence model in the distribution channels in various two player quantum games in Eisert’s model. (4) Although the quantum players have high potential gains. π/2] and 2 φ ∈ [0. Classical players are more robust to corruption in the source.
for Wstate the payoﬀs to the players become a probability distribution over the entries of the payoﬀ matrix. 7. and (ii) when the game is played with GHZ state. Hence no player. comparing the games played with these . and as a result they receive the highest payoﬀ. the original classical game payoﬀ matrix in pure strategies becomes a subset of the quantum version. Entanglement and Reproducibility of Multiparty Classical Games in Quantum Mechanical Settings The expectation that quantum game theory may beneﬁt from the rich dynamics of physical systems with many interacting particles ignited the research of multiplayer games. the payoﬀs delivered to the players are unique entries from the classical payoﬀ matrix for the GHZ state. players prefer twoperson coalitions and the players in the coalition receive payoﬀ better than the third player who is not in the coalition. In this section. however. can gain an advantage over the remainder by utilizing quantum error correction to reduce the error probability of their qubit. The ﬁrst study of multiplayer quantum games was carried out by Benjamin and Hayden who showed by studying the PD and the Minority game that multiplayer games can exhibit forms of pure quantum equilibrium that can be observed neither in the classical versions nor in their twoplayer forms. it became generally accepted that quantum versions can be easily extended to N player situations by simply allowing N partite entangled states. This was also encouraged by the observation that entanglement leads to equilibrium solutions which are not observed in the classical versions of the games. However.28 They also concluded that shared entanglement leads to diﬀerent cooperative behavior and acts a kind of contract preventing players betraying each other. performed a comparative study of GHZ and Wstates in a three player PD game showing that (i) players prefer to form threeperson coalition when the game is played with W state. or group of players. With the above two studies. YJ Han et al. In their model. Therefore. we show that this is not generally true because the reproducibility of classical tasks in quantum domain imposes limitations on the type of entanglement and quantum operators. they considered GHZlike states. When the operators corresponding to the classical strategies are applied.62 They further compared their results with the classical counterparts to show the advantage of quantum games over the classical ones. While in the case of GHZ state. Following this work.169 showed that all players are equally disadvantaged by decoherence in one of the qubits. this is not true for Wstates.
Reproducibility criterion to play games in quantum mechanical settings In order to make a fair comparison between the classical and quantum versions of a game as well as its dynamics with diﬀerent shared entangled states. Fig. Therefore.63. 8. For twostrategy games. For the strategy combination xk . we will look u ˆ xk j . corresponding to the two pure strategies in the classiui ˆ i cal game. 7. In this section. · · · .41. and the characteristics of quantum operators used as strategies. 2N }. In Eisert’s model of quantum games. we require that in any model of multiplayer games this criterion should be satisﬁed. . Then the combined pure strategy of the players is represented N ˆ1 ˆ2 ˆN by xk = ul1 ⊗ ul2 ⊗ · · · ⊗ ulN with li = {1. An N player twostrategy classical game should be reproduced in the quantum version when each player’s strategy set is restricted to two unitary operators. u2 }. ˆ ˆ Thus the output state becomes  Φk = xk  Ψ with k = {1.170 two states with each other and with the classical games in terms of the payoﬀs is not fair.1. . too. expected payoﬀ for the ith player becomes ˆ ˆ ˆ $i (ˆ1 . 2. reproducibility criterion is satisﬁed by choosing the quantum operators corresponding to classical strategies in such a way that they commute with the entangling ˆ operator J. 2} and k = i=1 (li − 1)2i−1 .64 We also gave examples of entangled states that can and cannot be used. The model for the multiplayer quantum games. . . That is with the proper choices of operators the classical game should pop up as a subset of its quantum version. we derived the necessary and suﬃcient conditions for a physical realization. the physical scheme in the quantum mechanical setting should be able to reproduce the results of the classical version. uN ) = Tr Σj ai Pj xk  Ψ Ψ ˆ† . {ˆ1 .
we have Φnj Φn k = 0. Matching Pennies (MP). in their N player extensions. and Alphonse & Gaston Coordination Game (AG). (19) This strong criterion should be satisﬁed independent of the structure of the game payoﬀ matrix. that is all the elements of Sj and Sk must be orthogonal to each other. Hence. In (ii). On the other hand StagHunt (SH). Battle of Sexes (BoS). ∀j = k. Chicken Game (CG). Φn k } then the referee should deterministically discriminate between these sets which is possible iﬀ the state space spanned by the elements of each set are orthogonal. DeadLock (DL). This weak criterion is very much dependent on the structure of the game payoﬀ matrix. the new payoﬀ matrix is formed by summing the payoﬀs that each player would have received in simultaneously playing the twoplayer game with N − 1 players. . .1.63 In (i). When a twoplayer twostrategy game is extended to N player game (N > 2). majority and coordination games have payoﬀ matrices where some of the outcomes have the same payoﬀ vectors. thus these games can be evaluated by the weak criterion. If two output states  Φj and  Φk have the same payoﬀs then they should be grouped into the same set. Boxed Pigs (BP). Hence for every element of Sj and Sk . . Φ2j . the referee should identify all possible outcomes  Φk deterministically which requires that all possible states formed by the joint applications of the players’ operators should be mutually orthogonal to each other. however. Minority. and Ranked Coordination (RC) have payoﬀ matrices where all the outcomes have diﬀerent payoﬀ vectors. Φnj } and Sk = {Φ1k . the referee needs to distinguish between the subsets of outcomes with the same payoﬀ. (ii) The referee should be able to reproduce the expected payoﬀ of the classical game in the quantum version. u2 } for any Npartite enui ˆ i tangled state and for any game? Reproducibility can be discussed in two cases: (i) The referee should be able to identify the strategy played by each player deterministically regardless of the structure of the payoﬀ matrix. Samaritan’s dilemma (SD). Φ2k . Battle of Bismarck (BB).41 .1. . Modeller’s dilemma (MD). . Prisoners’ dilemma (PD).41 Thus. β. . If all possible output states are grouped into sets Sj = {Φ1j . Φα  Φβ = δαβ ∀α. these games should be evaluated according to the strong criterion.171 for an answer to the question: Can we ﬁnd {ˆ1 . too.
.172 7. ⊗ u2 Wn .. ⊗ u1 WN . ﬁrst as an example we will show that Wstates cannot be used in quantum games because they do not satisfy the strong reproducibility criterion. We ﬁrst take any two output states for which the strategies of only one player diﬀer. and the ith player has the strategy set {ˆ1 .. ⊗ u1 WN ˆ1 ˆ2 ˆ 3 ˆN Φ2 = u2 ⊗ u1 ⊗ u1 ⊗ ... (20) ˆ1 ˆ2 ˆ3 ˆN Φ2N = u2 ⊗ u2 ⊗ u2 ⊗ . . . u2 }. Φ2N } such that ˆ1 ˆ2 ˆ 3 ˆN Φ1 = u1 ⊗ u1 ⊗ u1 ⊗ . Since the strong criterion of reproducibility requires that Φα Φβ = δαβ for ∀α. e.Φ1 and Φ2 . we take two output states diﬀering in the operators of two players... The inner products of these states becomes Φ 1  Φ2 = 1 [(N − 1) 0 ˆ1† u2  0 + 1 ˆ1† u2  1 ]... . so is the u1† u2 which can be expressed ˆ1 ˆ1 ˆ1 ˆ 1 as x y y ∗ −x∗ (22) ˆ1 ˆ 1 with x2 + y2 = 1.... . Acting on ui ˆ i the shared state  WN with all possible strategy combinations of Nplayers will generate the output state set {Φ1 . u1 ˆ 1 u1 ˆ 1 N (21) Since u1 and u2 are SU(2) operators. Then we will derive a general form of entangled states that satisfy the criterion and hence can be used in quantum versions of classical games..g.. ⊗ u1 WN ˆ1 ˆ2 ˆ 3 ˆN Φ3 = u2 ⊗ u2 ⊗ u1 ⊗ .. Inner product of these . (21)... Substituting this in Eq. σx R ˆ 1† 2 Note also that u1 u1 is a normal operator. Entangled states and strong reproducibility criterion In this subsection.Let us assume that the Npartite W state  WN has been prepared and distributed to the players. 2n}. e.2. Φ2 .. β ∈ {1.Φ1 and Φ3 . we check the mutually orthogonality of all the 2N output states listed above. .... Similar expression can be easily ˆ ˆ obtained for the ith player just considering two output states diﬀering in the operators of the ith player.g. W states do not satisfy the strong reproducibility criterion . we ﬁnd that u1† u2 = ˆ ˆ ˆ ˆ z (2φ1 ) where the rotation operator Rz (γ) is given as Rz (γ) = e−iγ σz /2 .. Next.
. the classical game cannot be reproduced and hence the W state cannot be used. (24).. we see that φ1 = nπ + pi/2 implying that D1 = iσz . Taking the inner product and imposing ˆ the reproducibility criterion. General form of states satisfying the strong reproducibility criterion Now suppose that for an Nqubit entangled state  Φ we have found two operators {ˆ1 . . therefore we have Dk = iσz .. Since u1 ˆ1 ˆ 1† 2 u1 u1 is a normal operator. we ﬁnd Φˆ1† u2 ⊗ I ⊗ ..173 states is Φ 1  Φ3 = 1 ( 01  + 10 ) (ˆ1† u2 ) ⊗ (ˆ1† u2 ) ( 01 +  10 ) u1 ˆ 1 u2 ˆ 2 N 2 1 ei(φ1 −φ2 ) + e−i(φ1 −φ2 ) = cos(φ1 − φ2 ). (24) ˆ where  Φ = z1 ⊗I ⊗· · ·⊗I Φ .. Thus..iN  i1  i2 · · ·  iN . From ˆ cos φ1 = 0.. Then for any three player. we ﬁnd cos φ1 + i 2Σij ∈{0. and consequently we have ˆ ˆ ˆ ˆ ˆ ˆ ˆ Φ ˆ1 z1 u1† u2 z1 z1 ⊗ I ⊗ · · · ⊗ I Φ = Φ D1 ⊗ I ⊗ · · · ⊗ I Φ z † ˆ ˆ1 ˆ1 ˆ† ˆ ˆ ˆ ˆ = Φ Rz (−2φ1 ) ⊗ I ⊗ · · · ⊗ I Φ = 0. One can easily notice that the sum of the three equations in this set leads to a contradiction: The sum of the three equations in χjkm results in 3π/2+m π = 0 which is impossible for integer m = n+n +n .. ˆ ˆ ˆ ˆ ˆ ˆ1 ˆ1 ˆ† say z1 . thus φ1 − φ2 = nπ+π/2 where n is an integer. Therefore. ⊗ IΦ = 0. n and n being integers. = N N (23) The condition of orthogonality requires cos(φ1 − φ2 ) = 0.. we can write D1 = z1 u1† u2 z1 = Rz (−2φ1 ) where e∓iφ1 are ˆ 1† 2 the eigenvalues of u1 u1 . u2 } implying that we have a situation where the strong reproui ˆi ducibility criterion is satisﬁed. it can be diagonalized by a unitary operator.1} c0 i2 . Thus. Now writing the state  Φ in computational ˆ ˆ basis as  Φ = Σij ∈{0.iN 2 − 1 sin φ1 = 0 (25) which is satisﬁed for cos φ1 = 0 and Σij ∈{0.. ⊗ u1 Φ .1} ci1 i2 . φm − φj = n π + π/2.. ⊗ u1 Φ and ˆ1 ˆ2 ˆN Φ2 = u2 ⊗ u1 ⊗ . Let us again take the output states diﬀering ˆ1 ˆ2 ˆN in the operators of only one player say Φ1 = u1 ⊗ u1 ⊗ . we have φj −φk = nπ+π/2 for any two output states diﬀerent only in the strategies of j. we have the set of equations χjkm = {φj − φk = nπ + π/2. and substituting into Eq. φk − φm = n π + π/2} with n.iN 2 = 1/2. we conclude that there exists no two operators corresponding to the classical strategies of the players when  WN is shared in a twostrategy Nplayer game. Thus. This ˆ result can be extended for the other players.1} c0 i2 ..and kth players.
and substituting it into all possible output states written as in Eq..1} {0. ˆ ˆ Neglecting the irrelevant global phase of i of D.1  (26) (27) where the last row is the normalization condition.1} sponds to the diagonal elements of σz ˆ ⊗· · ·⊗ σz ˆ where σz is deﬁned as ˆ0 {0. Thus. σ ˆ ˆ Φ ˆz ⊗ σz ⊗ · · · ⊗ σz  Φ = 0... ˆ ˆ ˆ ˆ Φ I ⊗ σz ⊗ I ⊗ · · · ⊗ I Φ = 0. . H = √ (σx + σz )/ 2. too. .1} ⊗N ˆ Consider the operators x. then it should be transformed by local unitary operators into the state which contains all possible terms with the same magnitude but diﬀerent relative phases:  Φ = Σij ∈{0. one can easily show that  WN and Dicke states  N − m.. c00. . y ∈ (ˆz ) where xy ∈ (ˆz ˆˆ σ ) . The GHZ state. .. which also satisﬁes the strong reproducibility ˆ ˆ √ π π ˆ ˆ σ criterion. . for x = y . 1 c11.1  0 . we end up with 2N − 1 equalities to be satisﬁed: ˆ ˆ ˆ σ Φ ˆz ⊗ I ⊗ I ⊗ · · · ⊗ I Φ = 0. . 1 1 .. (28) when all the players apply the Hadamard operator. m except  2. . . . This implies that if a state satisﬁes strong reproducibility criterion. (27) has an inverse. . .. .1} N −1/2 eiφi1 i2 . . we have Tr[ˆ y ] = Tr[ˆ] Tr[ˆ] = 0. ...iN 2 are uniquely determined as 1/N . Bell states.iN  i1  i2 · · ·  iN . . . = . (28) One can easily show that the product state can be transformed into the ˆ form of Eq. and ci1 i2 . The row vector corre{0. The proof for Dicke states can . can be transformed into the same form by (ei 4 I+e−i 4 σz +ˆy )/ 2 ˆ for one player and H for the others. . . In the same way.1} ⊗N {0. . . u2 } satisfy reproducibility criterion... . then the state  Φ = uk ˆ k ˆ ˆ ˆ ˆ z1 ⊗ z2 ⊗ · · · ⊗ zN  Φ and the unitary operators {D. 1 cannot be written in the form Eq. ˆ ˆ σ ˆ ˆ xˆ x y Tr[ˆz ] = 0. product states and  GHZN states can be used in quantum versions of classical twostrategy multiplayer games.. 2 and  1. .174 The above equations imply that if the Nqubit state  Φ and the operators {ˆ1 . 2 1 1 . . thus the matrix in Eq..0 2 2 . ... −1 c00. . I} should satisfy. 1 −1 .. Thus any two row σ vectors are orthogonal to each other. . Rewriting these equations in the matrix form yields 0 1 . (24). (28) and hence cannot be used. . Since I.. .
CG. then there will always be contradiction (iv) If there is a set with only two elements which are the outcomes when all the players choose the same strategy.3. m).2. Checking these observation on various games. √ √ 1† 2 2 2 uk=1.4 = I. then there will be contradiction if we obtain the set χjkm for any threeplayercombination (j. Weak reproducibility criterion and W states . We will consider only the Wstates here but a detailed analysis for Dicke states can be found in. BoS. we found out that Wstates cannot be used in (1) multiparty extensions of originally twoplayer twostrategy games such as SH. The eigenvalues of uk uk are i ˆ σ ˆ ˆ σ ˆ ˆ and −i and they are already in the diagonalized form. the ˆ ˆk σ operators are u1 = I and u2 = iˆy which can also be written in the form ˆk ˆ D. N . 2 can be used in a quantum ˆ game by assigning the classical strategies to the operators u1 ˆk=1.3 = i( 2ˆz + σx )/ 3. m). there will be contradiction. Then the only equations we will obtain are φ1 − φj = nπ + π/2 for all j = 2. Note that such a situation occurs iﬀ 2N possible outputs are divided into two subsets with equal number of elements. We ﬁnd that the Dicke state  2. The following observations for the Wstates from the above analysis make our task easier: (i) If the mutual orthogonality condition of the subsets (each subset includes the all output states with the same payoﬀs) leads to the operator ˆ ˆ form u1† u2 = σx Rz (2φk ) for any two output states diﬀering only in the ˆk ˆ k strategy of kth player. k.41 The results of such a study requires a classiﬁcation of multiplayer games because the results would be very much dependent on the structure of payoﬀ matrices. (iii) If the number of elements in any of the subsets in a game is an odd number. (ii) If the mutual orthogonality condition of the subsets leads to the operˆ ˆ ator form as in u1† u2 = σx Rz (2φk ) for one and only one player. then ˆk ˆk there will be no contradiction because there will be at least one missing equation in χjkm for all possible threeplayercombinations (j. · · · . Presence of at least one such set is enough to conclude that there is contradiction. BB. k. and u4 = iˆy .175 also be obtained in the way we have done in the previous subsection for Wstates. If we encounter a contradiction in the analysis of a multiplayer game played with Wstates.2. we conclude that Wstates cannot be used in the quantum version of that game. .In this section we check whether the above results for strong reproducibility criterion are valid or not for the games where payoﬀs are the same for some of the possible output states after the players perform their operators. For GHZ state.
From the setpair (S1 . or if all players have made the same choice. For example. ˆk ˆk ˆ ˆ and from the setpair (S1 . Since we are dealing with a fourparty case. and 4 = 1 (the fourth player has chosen the second operator). from the setpair (S1 . S5 ) and (S4 . If there is an even split. and (5) symmetric games among which we can list PD. (3) coordination games due to (iiv). we ﬁnd 1. The players with the majority decision loses one point. third. (2) minority and majority games due to contradiction explained in (i). similar contradiction is observed.3 = 1 and consequently L = 7 implying that  Φ7 is the output state received by the referee..16 } where the subscript is L = 1 + 4 i=1 i 2 = 1) when the kth player chooses the ﬁrst operator u1 (or the second ˆk i ˆ 1 ˆ 2 ˆ 3 ˆ4 operator u2 ).10. S2 ) using Φ2 Φ1 . Substituting uk uk into the expressions obtained from the orthogonality ˆ ˆ relations of the sets (S2 . and S5 = { Φ8.16 }. 4−i with i = 0 (or { ΦL=1. or trading in a ﬁnancial market. then there is no reward. we see that the set of expressions represented by χ234 in (i) is obtained. For eight output states out of sixteen possible outputs for the fourplayer minority game. The structure of this game reﬂects many common social dilemmas.4. The choices are then compared and the player(s) who have made the minority decision are rewarded by one point.12 }. S4 ) using Φ5 Φ1 .13. and the second and third players choose the second operators. let us analyze the minority game with fourplayers. S5 ) using Φ9 Φ1 . and AG up to six players due to (iii). (4) zero sum gates due to (iv). from the setpair (S1 . S5 ). Each player is minority in two cases. and AG due to contradiction explained in (iii). We group these eight states into the set S1 = { Φ1.4 = 0 and 2. there is an evensplit (half of the players prefers ﬁrst operator and the other half prefers the second operator) and hence all the players are rewarded with payoﬀ zero. RC.2.11. SH. (S3 . for example choosing a route in rush hour. S4 = { Φ5. In this game.. As an example. Note also that in each of the sets we have output states diﬀering in the operators of all the players. S3 = { Φ3. S3 ) using Φ3 Φ1 .9 } where fourth.6. if the output state is  Φ2 . each player chooses between two strategies.7. there are 16 possible outcomes. for the joint action u1 u2 u2 u1 where the ﬁrst ˆk and fourth players choose the ﬁrst. second and ﬁrst players are minority players. S5 ). In the same way. then we know that L = 2 which is calculated using 1.14 }. we ﬁnd u1† u2 = σx Rz (2φk ) for 1† 2 ∀k. second and third players have chosen the ﬁrst operator)..3 = 0 (ﬁrst.176 MP. therefore the rest of the eight output states can be grouped into S2 = { Φ2. MD. Therefore. choosing which evening to visit an overcrowded bar. respectively. .15 }. CG.
177 For a zerosum game with N players it is easy to see that a contradiction similar to the one in (iv) occurs. The reproducibility criterion provides a fair basis to compare quantum versions of games with their classical counterparts because only the states and games which satisfy the criterion in a physical scheme can be compared and played. Dicke and Bell states are valid for a large class of classical games whose classical outcomes correspond to the same payoﬀ. t players choosing the ﬁrst strategy gets ξ/t. and each of the rest will contain only one state. and that the operators that might be used should have a special diagonalized form. One interesting ﬁnding is that for some games entanglement is not needed to arrive at an NE and resolve the dilemma in the game. 8. Our studies have revealed that the results obtained under the strong reproducibility condition for the W. One of these sets will contain the two output states when all players choose the same strategy. In a zerosum game with a competitive advantage ξ. and the rest N − t players receive ξ/(N − t). The analysis shows that a large class of multipartite entangled states cannot be used in the quantum version of classical games. Otherwise. The widely used Eisert’s scheme is discussed and applied to speciﬁc games to show the eﬀects of shared correlations and the set of operators available to the players on the game dynamics. we have attempted to clarify some of the dynamics in the games when the classical games are transferred to quantum domain. there is no winner or loser so all the players receive zero payoﬀs. GHZ. Such an output state partition is exactly what leads to the contradiction in (iv). however it is necessary to ﬁnd the solution with the highest payoﬀ distribution to the players. In an attempt to form a fair basis to evaluate the entanglement in games and to compare the game dynamics. Wstates cannot be used in zerosum games. We restricted our results to the speciﬁc entangled states which have already been created in experiments. Then the output states can be grouped into 2N − 1 diﬀerent sets. we suggested to impose a reproducibility condition which states that when a quantum mechanical setting is used to play the game the classical game should be . Conclusion In this lecture note. The comparison of various games also revealed that quantum advantage does not necessarily take place in all the games. it is strongly dependent on the game payoﬀ matrix. Therefore. Classical correlations in a game may lead to NE and in some cases resolve the dilemma however they never lead to payoﬀs as high as those in the case of entanglement.
P Moulin. UK. and to develop a more fundamental game theory which is associated with quantum mechanics. 11. 45 (1997) 1092. M. Conway. Press. A. Ettinger. L. J. IEEE Trans. Japan. Proc. Conf. on Inf. on Inf. 2. 4. M. J. Srivastava. F. on Image Process. Rahimi and Dr. there is still much work to be done to fully understand the mechanisms behind the interesting observations in quantum games. Pateux. of Int. A. J. 7. Shen. in 2nd Inf. Aoki for their invitation and kind hospitality during the Summer School on Mathematical Aspects of Quantum Computing 2007 held at Kinki University. Koashi. 18 (2003) 283. Signal Process. 3. This condition lead us to give general forms of the entangled states and quantum operators which can be used in quantum versions of classical games. In the rapidly growing quantum information science. Sloane. M. Mathematische Annalen 100 (1928) 295. K. G. Theory 48 (2002) 1639. Cambridge MA. (New York: Wiley. IEEE Trans. Yamamoto and Dr. Ivanovic. on Signal Process. 1997). 1517 (1998). 8. HigashiOsaka. Oxford. Orlin Grabbe. Rasmusen. advices and support. The expectation is that quantum game theory may help when dealing with such situations. in The First Workshop on Games and Emergent Behaviors in Distributed Computing Environments. Lapidoth. Morikoshi ¨ for fruitful discussions. quantph/0506219 5. S. Games and Information: An Introduction to Game Theory (Blackwell Pub. Portland. A. Therefore. T. R. 9. X. 2001). and V. Hiding Workshop. Dr. Myerson. it is inescapable to deal with conﬂict and competitive situations. L. 6. B. Acknowledgements The authors thank Prof. USA. and A. Apr. M. 1944). Neumann John von. T. Neumann John von and Oscar Morgenstern.178 reproduced in that speciﬁc setting. 10. S. September (2004) . Cohen. The Theory of Games and Economic Behavior. ¸ Nakahara. It is also expected that development and understanding of quantum game theory will pave the way to develop eﬃcient quantum algorithms and protocols. Birmingham. S. It is still not clear how we can utilize the present knowledge learned from the study of game theory in the past years to meet the above expectations. E. Theory 32 (1986) 337. N. Le Guelvouit. Game Theory: Analysis of Conﬂict (Harvard Univ. References 1. Deng.: Image Comm. R. IEEE Trans. DaSilva. 3 (2001) 975. Ozdemir thanks Prof. Dr.
and N. Y.179 12. Abbott. Parrondo et. Physica A bf 318 (2003) 505. W. Tsutsui. and J. Phys. Benjamin. Wu J and Han R. Phys. 19. Tsutsui. S. Phys. A 348 (2006) 147. Chen. and N. A. Sladkowski. March (2005). Ozdemir. K. Piotrowski. T. Phys. 43. Rev. 49. D¨r. Stefanov. E. Shi M. and N. 2 (2002) L327. A. 2000). Lett. 88 (2002) 137902. J. Physica A 312 (2002) 208. and I. J. quantph/0702167. J. 26. 42. Toor. 2/1 (2004) 79. Lett. M. C. Rev. I. u M. 47. F. K. Dicke. and N. Rev. Hayden. J. Kay. 24. 50.. W. Lee. C. A 300 (2002) 537. Flitney. 14. Piotrowski. 30. Rev. G. 18. Johnson et. Iqbal and A. 41.. A. Marinatto and T. Shimamura. Benjamin and P. World 15(10) (2002) 259. 45. and D. Imoto. Imoto. Ichikawa. quantph/0503176 S. 16. Xu X. A. M. 23. 27. P. Sladkowski. Lett. K. Lett. Johnson. 21. Walther.R. Koashi. Phy. Inf. quantph/0409046. 15. H. K. and D. 82 (1999) 1052. of Quant. 25. J. 28. A. Abbott. K. R. A 62 (2000) 050302(R). F. C. A 313 (2004) 218. 86 (2001) 910.M. I. Toor. al. F. Noise. Eisert. F. A. Hayden. Phys. Cirac. and S. ibid 87 (2001) 069802. H. Lett. Physica A 314 (2002) 35. Chuang. Rev. Fluct. S. Beausoleil. quantph/0203049. C. A. z H. Flitney. Quant. 34. 83 (1999) 3077. Wilkens and M. Iqbal and A. Benjamin. Phys. T. Fluct. Lee. 22. Lett. al. N. Du J. P. A. Raussendorf. Li H. Weber. Rev. A 272 (2000) 291. Briegel and R. C. Johnson. Lett. R. Cheon. Ozdemir. 9 (2007) 43. F. F. H. A 67 (2003) 022112. Phys. Toor. 87 (2001) 069801. Int. Nawaaz. H. 13. A. Finance 4 (2004) 1. 17. E. Shimamura. Quantum Computation and Quantum Information. N. F. and A. Morikoshi and N. Rev. Imoto. W. and R. New J. al. P. 85 (2000) 5226. Shimamura. Nielsen and I. Lett. Hogg. Lett. J. 2 (2002) L305. P. Lett. F. Phys. 33. F. H. Lett. 39. A. 29. Playing games in quantum realm. and N. Physica A 324 (2003) 152. Phys. quantph/0102008. 32. Ng. 37. PhD Dissertation. Johnson. 9 (2007) 205. Lewenstein. A 294 (2002) 261. Cheon. and N. Rev. (Cambridge university Press. 93 (1954) 99. R. Buˇek. Rev. 44. E. Abbott. N. Phys. Rev. S. Rev. 40. Inf. M. Phys. . C. Phys. W. A 62 (2000) 062314. 38. T. Johnson. C. G. 35. J. Prevedel. D. Phys. Osaka. 46. Quant. Iqbal and A. Vidal. F. Lett. Phys. A 280 (2001) 249. Stockton et. and A. Lee. Shimamura S. Nature 402 (1999) 864. Phys. and J. Phys. V. Ozdemir. F. Zeilinger. P. 20. and T. Phys. F. Lett. J. 36. Imoto. Morikoshi. Phys. H. Abbott. Phys. Johnson. Sladkowski. Toor. and Noise Lett. Piotrowski. Rev. J. L. Lee. A 64(3) (2001) 030301(R). Phys. L. A 319 (2003) 429. Osaka University. and J. and J. Iqbal and A. Allison and D. A 67 (2003) 022311. Harmer and D. Phys. Toor. J. A. Lett. Meyer. Japan. M. 31. New J. 48. Proc 1 (2002) 449.
Johnson. and Economic Theory. Iqbal. September (2004). C. University of Adelaide. Lett. YJ. 54. Morality. N. Shimamura. A 66 (2002) 024306. J. A 295 (2002) 61. Ozdemir. Zhang. 62. Morikoshi and N. quantph/0508105. Australia. Pakistan. New J. A. Ozdemir. Studies in the theory of quantum games. Imoto. J. Hollenberg. F. S. Lett. Abbott. edited by Edmund S. 1975). Paternostro. K. Imoto. Universidad de Los Andes. MSc Thesis.au/adelaide/adtSUA20051027072026/. Morikoshi and N. January (2005). Phys. Phys. S. Available at arXiv: quantph/0402117. 57. 60. quantph/0101005. Phys. Aspects of quantum game theory. A. 63. Phys. 52. PhD Dissertation. Lett. A. Shimamura. A 38 (2005) 449. A 328 (2004) 20. Rev. Available at http://ariic. A.180 51. quantph/0510108. Morikoshi. New York. Bogota. Classiﬁcation of quantum symmetric nonzero sum 2×2 games in the Eisert scheme. F. and N. P. K. F. Phys. Guo. M. Tame. Flitney and D.library. Kim. Rev. 64. P. Phelps. February (2004).unsw. F. A.edu. Ozdemir. Phys. M. S. F. “The Samaritan’s Dilemma”. Islamabad. . Colombia. C. Flitney. Quaidiazam University. Pike. Han. PhD Dissertation. in Altruism. Buchanan. M. J. van Enk and R. Phys. J. Brassard. 58. S. M. Available at arXiv: quantph/0503176. L. and G. YS. Flitney and L. 59. K. Imoto. G. (Russell Sage. J. P. 55. A 325 (2004) 104. 56. Shimamura. S. 61. HuertasRosero. S. 7 (2005) 226. 53. J. A 63 (2001) 020302.
Chiyodaku. In this paper. On the other hand. QuasiCyclic LDPC Codes. Tokyo. ReedSolomon codes.181 QUANTUM ERRORCORRECTING CODES MANABU HAGIWARA National Institute of Advanced Industrial Science and Technology. Introduction Quantum errorcorrecting codes provide detection or correction of the errors which occur in a communication through a noisy quantum channel. Hence the codes protect integrity of a message sent from a sender to a receiver. CSS Codes. e. and others. One of the research trends is to generalize algebraic geometrical codes. the main target of a class of LDPC codes which are generalized to quantum codes is called a class of CSS codes. Thus LDPC codes does not seem to be suitable for generalizations to quantum errorcorrecting codes. we discuss quantum errorcorrecting codes by emphasizing the generalization of classical errorcorrecting codes.g. is required to CSS codes. we focus not on algebraic geometrical codes but on LDPC codes.1 to quantum codes in view of mathematical aspects of quantum errorcorrection codes. Starting with the explanations of general notions of classical errorcorrecting codes. we discuss quantum errorcorrecting codes in view of generalization of classical errorcorrecting codes. called twisted condition.go. we ﬁnish the discussion with the construction of quantum codes which are constructed by a pair of practical and trendy classical codes. which is one of the most popular classes of quantum codes. 11813 Sotokanda. the nearly random paritycheck matrix is required to classical LDPC codes. Keywords: Quantum ErrorCorrecting Codes. 1010021 Japan ∗ Email: hagiwara.2 In particular. BCH codes. Algebraic Combinatorics. Group Theory. the notion . LDPC Codes.3 For construction. In this review. Recently. 1. a constructive but interesting condition. We give examples of basic theory of classical errorcorrecting codes and then introduce related quantum counterparts. Stabilizer Codes. AkihabaraDaibiru 11F.hagiwara@aist.jp In this review.
1. The 3repetition code is useful for classical bitﬂip errorcorrection if the probability of double bitﬂip errors in a word over the channel is negligible. ﬁnite rings. The sender transmits x itself over the noisy channel and the receiver obtains a single bit y. we should choose a suitable encoding for communication. 1} to a receiver. Another advantage of a quasicyclic LDPC code is practical one that it has high errorcorrecting performance with short length. The point of encoding is to introduce some form of redundancy to the original information. The probability that double bitﬂip errors occurring in the word depends on a communication channel. Throughout this paper. Study of good encoding is a fundamental and an interesting but an important problem for errorcorrecting . Classical ErrorCorrecting Codes 2. discrete mathematics. and others. to describe the properties of LDPC codes.g. Here + is a bitwise addition modulo 2. If one of the three bits is ﬂipped. the receiver corrects the error by a majority decision. If it is not negligible. Introduction to classical errorcorrecting codes Let us consider the following of examples of classical errorcorrecting codes: A sender would like to transmit a single bit x ∈ {0. Thus the receiver detects the bitﬂip error. the 5repetition code is one of the candidates. 111}. Let us assume that the bitﬂip error occurs with small but nonnegligible probability over the communication channel. which changed the reseach trend of quantum errorcorrecting theory. then the received word is not an element of the set C. In this review we aim to study the construction of quantum errorcorrecting code in the framework of generalization of classical quasicyclic LDPC codes.182 of quasicyclic LDPC codes has been proposed. 2. It is natural to assume that the communication channel is noisy. Imagine that the sender transmits a word xxx instead of the single bit x to the receiver over the noisy channel. e. we assume that a classical errorcorrecting code is a binary code and a quantum errorcorrecting code consists of sequences of two level states. This implies that the original word is an element of C = {000. Let us introduce an errorcorrecting code to this scenario. Furthermore. It is impossible to distinguish for the receiver that the received message y is the original message x or the bitﬂipped message x + 1 because of noise. One of the advantages of quasicyclic LDPC code is theoretical one that there are various mathematical tools. For example.
of an errorcorrecting code. .e. The linearity provides us with beneﬁts in studying errorcorrecting codes. Shannon’s result is quite interesting and meaningful from theoretical point of view. We should note that it is not required that a code be a block code to deﬁne a linear code. . i. the length of (0. 0) is 5. easy encoding. the bit sequence) is said to be n. In this section. where a random code is a code whose codewords are chosen randomly. We call a linear code C an [n. F2 = {0. and good errorcorrecting performance with shorter length. . The positive integer n is called the length of the code C. Hence a code (or a code space) means a set which consists of bit sequences and no mathematical structure is required. 0. then the length 2 of the codeword (resp. Linear codes and paritycheck matrices Let F2 be a prime ﬁeld with characteristic 2. In fact. Thus a block code C is a subset of Fn for a 2 positive integer n. 0. If a code C is given. .183 codes. For example. However we would like to develop a good error correcting code in realistic meaning. If the dimension of a linear code is k as a vector space over F2 . A code is called a block code if the length of any element of the code is independent of a codeword. A subset of F2 ∪ F2 ∪ F3 ∪ . Not only to ﬁnd a good encoding but also to ﬁnd a good decoding algorithm is an important problem. 0.5 However. then the dimension of a linear code is said to be k. Now let us deﬁne the rank. a bit sequence) belongs to Fn for some positive integer n. we focus on ﬁnite linear codes among classical linear codes. we might apply theory of linear algebra to errorcorrecting codes. Then the code C is a vector space 2 over F2 . there remain some problems. or dimension. 1}. If a codeword (resp. We call a block code C(= ∅) linear if C is a linear subspace of Fn . we can deﬁne an inﬁnite vector space over F2 . An element of F2 ∪ F2 ∪ F3 ∪ . easy decoding. Let us intro . c ∈ C. k] code if C ⊂ Fn and the rank of C is k. The random code requires huge length and it is impossible to correct errors over a random code in realtime. Another example 2 of beneﬁts of linearity is to develop lowcost errordetection. Note that we can consider Fn a vector 2 space over Fn . Shannon shows that high errorcorrecting performance is achieved by a random code. For example.e. is called 2 2 2 2 a bit sequence.4. i. 2. The purpose of errorcorrecting code theory is to develop a code with “easy encoding”. Note that a block code is a ﬁnite set. is called a code. an element of a code is called a codeword of C. c + c ∈ C for any c. “easy decoding method” and “shorter length”. and an inﬁnite linear errorcorrecting code.2.
2 Let us go back to errordetection theory. (0. 1. 0. 0. 1). z) ≤ d(x. a given binary matrix HC deﬁnes a linear code C by putting C := {x ∈ Fn HC xT = 2 0}. 1. In other words. 2 2 we deﬁne the distance d(x. 0. Since C is a vector subspace of Fn . if we put S = {(1. We can represent HC as a binary matrix of size (n− k)× n. . 1. there exists a 2 linear mapping HC from Fn to Fn−k such that 2 2 C = ker HC := {x ∈ Fn HC xT = 0} 2 as a binary matrix of size n × (n − k) and rankHC = n − k. 0). The paritycheck matrix HC is not uniquely determined by a given code C. an nbit sequence x is a codeword if and only if HC xT = 0. 1. Thus we have a characterization for x ∈ Fn to be a codeword. 0. 0) and (1. 0. which is called a paritycheck matrix of C. 0. y) + d(y. The weight wt(x) is called the Hamming weight of x. c = c }. 0)}. Conversely. 1. a linear code C is characterized by the paritycheck matrix HC . For x. 1. On the other hand. the minimum distance d(S) is 3. y ∈ Fn . d] codes The weight wt(x) of x = (x1 . c ∈ S. y. 1). 2 Let S be a subset of Fn . wt(1. 0. 0. the minimum distance d(S) of S is 2 deﬁned by d(S) := min{d(c. 0. we can detect error by calculating HC y T . (0. 0.3. 1. 1. z) holds for any x. z ∈ Fn . 0. k. For example. 0. x2 . 0. For example. then we ﬁnd that there is at least one error in the received message. 0) − (1. y) between x and y by d(x. 0) = 3. 1. 0. Next we introduce a distance in Fn . . 0. 1) is 2. The distance is called the Hamming distance. An algorithm to calculate the syndrome HC y T and an algorithm to compare the syndrome HC y T with a zerovector 0 work in practicaltime if the length is not so long. y) := wt(x − y). 2. 0. For S.184 duce a paritycheck matrix of a linear code to explain the realization of the lowcost errordetection. 0. For example. Note that the triangle inequality d(x. 0. 1) = (0. If the syndrome is not 0. where xT is the transposed vector of x. The vector HC y T ∈ Fn−k is called the 2 syndrome of y associated with HC . For a received word y. 0. the Hamming distance between (1. Minimum distance and [n. c )c. . xn ) ∈ Fn is deﬁned by 2 wt(x) = #{ixi = 1}. . . since (1. 1.
We show that the minimum distance of the Hamming code is 3. and minimum distance d. In other words. 1.1. Remark 2. the length is 7. 3] code. j. Thus HC (ei + ej )T is a nonzero vector. k. Example 2. 4.1). the minimum distance of the Hamming code is 3. In general. On the other hand. c = c } ≤ d(x. HC eT and HC eT are linearly independent i j for any 1 ≤ i < j ≤ 7. Therefore. d(C) = min{wt(c)c ∈ C. Let c. c = 0}. c ). c )c. Let x be a codeword of C such that wt(x) = min{wt(c)c ∈ C. determination of the minimum distance for a given linear code is known as a NPhard problem. c = 0} holds. Consequently. Proof. Assume that there is a codeword c such that wt(c) = 2. Therefore d(C) = d(c.185 Proposition 2. which is a contradiction. A linear code C is called an [n.1. c ∈ C. 4] code. Assume that there is a codeword c such that wt(c) = 1. in particular. For example. By the linearity of C. the i syndrome is a nonzero vector. Then d(C) = min{wt(c)c ∈ C. We show that the Hamming code is a [7. The code deﬁned by HC is called a Hamming code. c − c is a codeword of C. Thus the Hamming code is a [7. the 3repetition code (appearing in Sec. c = 0}. dimension k. it is diﬃcult to determine the minimum distance of a given code. 2. there is a codeword c such that wt(c) = 3. Clearly. HC (e1 +e2 +e3 )T = 0. d] code if C is a block linear code with the length n. In fact. where ei denotes a vector such that the i th entry is 1 and all other entries are zero. c be codewords of C such that d(C) = d(c. Then c is a sum of two elementary vectors ei and ej for some i. c = 0}. there is no codeword c of the Hamming code such that 1 ≤ wt(c) ≤ 2 (Proposition 2. In fact. Then c = ei . Let us deﬁne a paritycheck matrix HC by 0001111 HC = 0 1 1 0 0 1 1 . the dimension of the code is 4.1) is a [3. 1010101 Note that the i th column of HC is equal to the binary expression of i. Then d(C) = min{d(c. Because the rows of HC are linearly independent. 0) = wt(x) = min{wt(c)c ∈ C. . c = 0}. 3] code. Let C be a linear code. c ) = wt(c − c ) ≥ min{wt(c)c ∈ C.1. The syndrome HC eT is the binary expression of i.
gk form a basis of C. Therefore G is a generator matrix. . the information bits consist of four bits.186 2. . gk span the code C. By a similar way. x7 ) is a codeword of C . In other words. x2 . . Then there are k linearly independent vectors g1 . . Deﬁne a matrix G by g1 g2 G = . We deﬁne a paritybit x5 by putting x5 = x2 + x3 + x4 . The last n − k bits of a codeword of an [n. . . . Now. We can encode C as a systematic code by using the paritycheck matrix. 0101011 Then the linear code C associated with H is a [7. By using a generator matrix G of a linear code C. . Let us permute the columns of the paritycheck matrix in Example 2. Thus. . x6 . g1 . 4] code. Let C be a linear code. Let x = (x1 .1 as follows: 0111100 H = 1 0 1 1 0 1 0. (x1 . . If the rightside block of a paritycheck matrix is a lower triangular matrix. g2 . In fact. . x4 ) be a fourbit sequence. g2 . . .4. k] code is called systematic if the ﬁrst k bits of a codeword coinside with the original message. An [n. . . 2 Proof. we have a simple encoding which is multiplication of a message x and the generator matrix G. Let k be the dimension of C. . x3 . we show an example of encoding by using the paritycheck matrix. gk . we deﬁne recursively x6 := x1 + x3 + x4 and x7 := x2 + x4 + x6 . This deﬁnition of x5 comes from the ﬁrst row of H . x7 . k] systematic code are called check symbols. gk It is easy to verify that g1 . Encoding Proposition 2. By the construction of x5 . then the linear code C of the paritycheck matrix is systematic. we have only to delete the last n − k bits to recover the original message. . Then there is a binary matrix G such that C = {xGx ∈ Fk }. . x2 . For a systematic code. The matrix G is called a generator matrix. it is easy to recover the original message from a codeword after errorcorrection. g2 .2. . .
Decoding In this review.1. Next. which consists of three bits. Let us explain a decoding algorithm for the 3repetition code which appeared in Sec. we regard that the received word y does not contain errors. Count the number of 1 in the input y. then output i. If the syndrome y is a zerovector. If otherwise. However. As was mentioned in Sec. we introduce the two general decoding algorithms.1). we use the term “decoding method” to denote errorcorrection algorithm in this article. First we introduce the decoding algorithms for the following two linear codes: the 3repetition code and the Hamming code (Example 2. The standard decoding algorithm of the Hamming code is to determine the errorposition from a syndrome. Received word y. If 1 appears two times or three times in y. 000 or 111. We should remember that it might not work practically for a general linear code.1). Both decoding algorithms are applicable to any linear code. 2. Estimatederror position: null or positive integer i. • Input Output Step 1 Step 2 Step 3 Decoding algorithm for the Hamming code. Calculate the syndrome of y associated with H (See Example 2.187 2. 2.1. It is known that to decode a given linear code is a NPhard problem.5. Assume that . if we assume some structures for a code. called the bounded distance decoding algorithm and the maximum likelihood decoding algorithm. Received word y which consists of seven bits. the decoding algorithm works successfully against a single bitﬂip error. then we might ﬁnd the decoding algorithm which works practically. then we regard that the received word y contains a single bit error ei . then output null. If the output is a positive integer i. We should remember that the decoding method in a restricted sense is to recover the information bits from the codeword. then the output is 111. • Input Output Step 1 Step 2 Step 3 Decoding algorithm for the 3repetition code. If the output is null. Thus the decoding algorithm does not work in real time if the length of a given code is very long. then the output is 000. If the syndrome is a binary expression of some integer 1 ≤ i ≤ 7. We shall explain the relation between the output of the algorithm and the estimated errorvector. where 1 ≤ i ≤ 7.
Therefore. We claim that Pc ∩ Pc = ∅ 2 if c = c . c ) ≤ d(c. Denote the probability that a variable x is equal to d by Pr[x = d]. Pr[y = 1x = 0] = Pr[y = 0x = 1].188 the single bitﬂip error ei occurs on a codeword c. Pr[x = d] For example. Estimated codeword or null. In general. Step 5 Go back to Step 1. go to Step 4. a communication channel is deﬁned by using the term of probability distribution of noise. then output c and stop this algorithm. then the algorithm outputs c. x)+d(x. y = d ] . For a linear code C with its minimum distance d. Let C be a linear code with its minimum distance d(C) = d. If the Hamming distance d(c. The posteriori probability Pr[y = d x = d] is deﬁned by Pr[y = d x = d] = Pr[x = d. c . y). y) is less than or equal to (d − 1)/2. If there is a vector x ∈ Pc ∩ Pc for some c. then wt(c − c ) = d(c. a binary symmetric channel (BSC) is a channel with a probability distribution of bitﬂip error such that the probability distribution satisﬁes Pr[y = 0x = 0] = Pr[y = 1x = 1]. Step 4 If all of the codewords in C are tested by Steps 2 and 3 and failed to produce any output in Step 3. for a codeword c ∈ C and an errorvector e. Next we introduce the maximum likelihood decoding. . This is a contradiction. the bounded distance decoding works well as an errorcorrection algorithm. then output null and stop this algorithm. We explain the reason why the bounded distance decoding works (See Figure 1). For any c ∈ C. we put a set Pc := {x ∈ Fn wt(c − x) ≤ (d − 1)/2}. c ) ≤ (d−1)/2+(d−1)/2 < d. Thus the above decoding algorithm works as a single bitﬂip error correction algorithm. We note that we can correct a single bitﬂip error without the received word if the syndrome is given. if the received word is y = c + e and wt(e) ≤ (d − 1)/2. If not. • Input Output Step 1 Step 2 Step 3 Bounded Distance Decoding Received word y. Choose a codeword c ∈ C. Calculate the Hamming distance d(c. Next we introduce the bounded distance decoding. Then the syndrome is: H(c + ei ) = Hc + Hei = Hei = the binary expression of i.
1. The maximum likelihood decoding is to Fig. for a given linear code. to calculate the posteriori probability for all of codeword of a code is impossible in real time if the code has numbers of codewords. where x is the input bit and y is the received bit. Pr[y = 0x = 1] + Pr[y = 1x = 1] = 1. Pr[y = 0x = 0] + Pr[y = 1x = 0] = 1. Image for Binary Symmetric Channel ﬁnd the code word c for the received word y such that Pr [received word = yx = c] = max Pr [received word = yx = c ] .189 Fig. Schematic idea of bounded distance decoding: Each circle is Pc for some codeword c. However. 2. and each bitﬂip error occurs independently (See Figure 2). c ∈C The maximum likelihood decoding seems to be the best decoding method for errorcorrection. .
Even if the users are under noisy circumstances.190 2. then we might never read important data in the disk again. Let us calculate BLER of the 3repetition code (see Sec.c2 . if two or more error occur in a transmission. A CSS code is deﬁned and is investigated later in this review (Sec.6.cn )∈C n i=1 Pr[ci = ci ] . 4). If no bitﬂip error occurs in a transmission. 2. On the other hand. We leave a calculation of BER of the 3repetition code to for readers. The block errorrate (BLER) is the expectation value of the error probability for a decoding algorithm. the receiver may understand what the speaker says. then a block error does not occur. Here we do not require the ﬁdelity of the encoded voice...2. BER = Exc=(c1 . then a block error does not occur either. On the other hand. Remark 2. Let us consider a mobile phone.1) over BSC channel with the crossover probability p. for example. In other words.. It is known that BLER of a classical code is closely related to the ﬁdelity of quantum communication appearing in a quantum code called a CSS code. Bit errorrate and block errorrate The way of evaluation for errorcorrecting codes depends on the application. If a single bitﬂip error occurs in a transmission. the accuracy of each bit is required for a magnetic disk of a personal computer If only one bit in a magnetic disk is broken. Example 2. The bit errorrate (BER) is the expectation value of the bitﬂip error probability for a bit in the estimated word after a decoding algorithm.. and c is the estimated word. BLER = Exc∈C [Pr[c = c ]] . Two of typical (or widely used) criteria of evaluations for errorcorrecting codes are values of bit errorrate and block errorrate. In other words. where c is the original encoded message. then a block error occurs. What is essential for the receiver is to understand the content of the speaker’s message.2. n where n is the length of the code and ci is the i th bit of the estimated word with the received word c. Thus BLER is: 1 − (1 − p)3 − 3p(1 − p)2 = p3 + 3p2 (1 − p). .
Thus a stabilizer code is a subspace of C2⊗n . “syndrome”) is closely related to “paritycheck matrix” (resp. In classical errorcorrecting codes. . We explain relations between these terms later (Proposition 3. Important classical objects such as paritycheck matrix. Z= . X. Throughout this paper. . H2 . which appeared in the previous section. Y } and ⊗ is the tensor product operation. i ≤ n−k. and Y denote the following: I= 10 01 1 0 0 −1 . X.1. In particular. which is a quantum errorcorrecting code. As the names suggest. syndrome. are generalized to quantum errorcorrecting codes in a natural way. How to Generalize Classical ErrorCorrecting Codes to Quantum Codes In this section. then the dimension of an eigenspace with respect to a ﬁxed syndrome is 2k . the outcomes of a paritycheck measurement are called the syndrome. . A stabilizer code. the dimension of an [n. its eigenvalues are 1 and −1. H2 . Z. . For quantum errorcorrecting codes.n and Hi Hi = Hi Hi for any 1 ≤ i.2). and others.191 3. By the condition of Hi . the dimension of the mother set C2⊗n is 2n and the dimension of a stabilizer code is shown in the following: Proposition 3. we introduce widely used terminologies to study quantum errorcorrecting codes. . 3. Paritycheck measurements and stabilizer codes Let H1 . the dimension of the related stabilizer code is 2k .1. Hn−k are called a paritycheck measurement. For a quantum state in C2⊗n . . “syndrome”) in the previous section.j ∈ {I. .2 ⊗ · · · ⊗ Hi. where Hi. The set of the Hermitean operators H1 . and Y = iXZ. is an eigenspace such that the measurement outcomes of all Hi are 1. k] code is k and the dimension of its mother set is n. If each of the Hermitean operators Hi is not included in a group generated by Hj (j = i). Next proposition shows the dimension of a stabilizer code and its mother set. “paritycheck measurement” (resp.X = 01 10 .1 ⊗ Hi. . matrices I. . Hn−k be Hermitean operators such that Hi = Hi. Z. code space.
.n be a paritycheck measurement such that Hi..−1 = 0.n−k}. It implies: dim VI.n−k}.I=∅ On the other hand.−1}n−k .j ∈ {I. Proposition 3.j ) by Hi.n−k}.I=∅ r∈{1. ..sI =e dim(Vs ).−1}n−k . where si is the measurement outcome of the operator Hi . ..1 = dim VI.sI = 2n−1 (2n−k − 1).I=∅ dim VI. Hn−k = Hn−k. I⊂{1.n−k}.192 Proof. For I ⊂ {1. s2 .. .... . dim VI.2. . Thus dim(Vs ) = 2k . ..rI =sI dim(Vr )) = (dim(Vs ) + I⊂{1. 2.I=∅ r∈{1.−1 = 2n−1 .−1 for I = ∅. Then the . Thus dim VI... . since the trace of HI = dim VI.r=s dim(Vr )) = (2n−k − 1) dim(Vs ) + r=s (2n−k−1 − 1) dim(Vr ) = 2n−k−1 dim(Vs ) + (2n−k−1 − 1)2n . we deﬁne a Hermitian operator HI by putting HI := i∈I Hi .1 ⊗ Hn−k.1 is equal to the dimension of VI.. .n−k}. The dimension of VI. .I=∅ ( I⊂{1. Since dim VI.n . let us calculate dim VI.. Next. Let H1 = H1..sI .sI = I⊂{1.2 ⊗ · · · ⊗ H2.−1 = 2n−1 . .. By the following proposition.. the readers may agree that a paritycheck measurement is regarded as a generalization of a paritycheck matrix of a classical code.j = 1 if Hi. Take a (classical) paritycheck matrix H = (Hi.e spanned by {Vs sI = e}. Denote the dimension of eigenspace Vs associated with a syndrome s := (s1 ...j = 0 if Hi.1 ⊗ H1. . n − k}.j = I.rI =sI . we have I⊂{1. Z}. The eigenspaces of HI associated with an eigenvalue e is the linear subspace VI.e = s.n .1 ⊗ H2.1 = dim VI. Similarly.2 ⊗ · · · ⊗ H1.2 ⊗ · · · ⊗ Hn−k. we put sI := i∈I si .j = Z and Hi.. sn−k ) by dim(Vs ). H2 = H2..1 − dim VI.. There are two eigenspaces of HI and their eigenvalues are 1 and −1.
We would like to leave the following simple question as an exercise for readers: Generalize a classical encoding by the lower triangle paritycheck matrix to an encoding for a quantum errorcorrection code by a paritycheck measurement. It is obvious that a quantum state c is an element of the stabilizer code. 4.1 ⊗Z hci. CSS Codes A stabilizer code is called a CSS code if there exists a paritycheck measurement such that each observable Hi consists of tensor product of either I.2. 3. encoding is a map from a 2k dimensional space to C2⊗n . X or one of I... We have already seen an errorcorrection method by using syndrome in Example 2. Instead of the received state which is a quantum codeword of a stabilizer code. Encoding Encoding should be a unitary operation. where c is a codeword with a paritycheck matrix H. Twisted condition By the deﬁnition of a CSS code.193 stabilizer code associated with the paritycheck measurement consists of a linear combination of quantum states c . Thus the dimension of information qubits must be equal to the dimension of encoded qubits. we can measure a syndrome associated with the paritycheck measurement. t such that Hi = Z hci.e. Thus a linear subspace spanned by {c c ∈ C} is contained in the stabilizer code. Therefore we need to add n − k initial states 0 to the information (i.3.1. On the other hand... These spaces are identiﬁed by comparing the dimension of them.2 ⊗ · · · ⊗ Z hci. we can separate the observables to two sets {Hi }i=1. it is impossible to measure the received state itself which is transmitted over a noisy channel.t and {Hi }i=t+1.t+t for some t. Z. if the stabilizer code is well designed. 3. It expects that we might correct quantum errors by using the syndrome. Proof.n for 1 ≤ i ≤ t and Ht+i = X hdi . preencoded) qubits.1. 4.2 ⊗ · · · ⊗ X hdi .1 ⊗ X hdi .n for . Decoding By the law of quantum mechanics.
Let C and D be classical linear codes with the paritycheck matrices HC and HD respectively. C and D satisfy the twisted condition if and only if T HC and HD are orthogonal i.1. Proof. In other words. the dual code C ⊥ is deﬁned by C ⊥ = {x ∈ Fn  x. . where AT is the transposed matrix of a matrix A. (b1 . . By the commutativity of the observables. T HC × HD = 0. we conclude that the linear codes C and D obtained from paritycheck measurements of CSS codes above satisfy the following condition: D⊥ ⊂ C(⇔ C ⊥ ⊂ D). In fact. D) be linear codes with a paritycheck matrix HC (resp. . HC × HD = 0. 2 In other words. ∀c ∈ C}. Proposition 4. the dual code is a linear code generated by rows of the paritycheck matrix. HC xT = 0 ⇐⇒ ∀ row x of HD .j }i=1..p = 1. .. We call this condition the twisted condition. x ∈ C ⇐⇒ ∀x ∈ D⊥ . b2 . . observables Hi where (a1 . a2 . where hci. Let us recall that the (classical) dual code of a linear code. bn ) = and Ht+j are commutative if and only if Z of Hi and X of Ht+j meet even times.. c = 0. D⊥ ⊂ C ⇐⇒ ∀x ∈ D⊥ .j }i=1. Finally. ai bi .194 1 ≤ i ≤ t . .t . an ).t .e. For a linear code C. Let C (resp.j and hdi . It is equivalent to claim that there are even positions p such that hci. .. . . HD = {hdi. HC xT = 0 T ⇐⇒ HC × HD = 0 . Then we can construct two classical paritycheck matrices HC = {hci.j ∈ F2 .p = hdj. the classical linear codes C and D satisfy the following condition: each row of HC and each row of HD are orthogonal with the inner product . HD ).
Therefore d ∈D⊥ (−1) x. Theorem 4.d = d ∈D⊥ (−1) x. Proof. where #X is the cadinarity of a set X. the dimension of the stabilizer code is equal to that of Q.1. Since D⊥ ⊂ C and c + d ∈ D⊥ . there is d0 ∈ D⊥ such that x.d = #D⊥ . 2 d ∈D⊥ (−1) otherwise). Proof.195 4. then x. Thus (−1) d ∈D⊥ x. if For x ∈ Fn . By Theorem 3. Characterization of code space Lemma 4. (Do not confuse the Hadamard operator H and a paritycheck matrix HC . d = 0 for all d ∈ D⊥ . Since {d0 + dd ∈ D⊥ } = D⊥ . x. deﬁned by C and D. the dimension of Q is equal to dim C/D⊥ = dim C − dim D⊥ . Let C and D be classical linear codes which satisfy twisted condition. Remember that the number of observables of the paritycheck measurement is (n − dim C) + dim D⊥ . is spanned by the following quantum states: { d ∈D⊥ c + d }c∈C . it is obvious that Hi c + d = c + d for the observable Hi associated with HC . If oth