Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Symmetry in quantum walks
(USC Thesis Other)
Symmetry in quantum walks
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
SYMMETRY IN QUANTUM WALKS
by
Hari Krovi
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ELECTRICAL ENGINEERING)
August 2007
Copyright 2007 Hari Krovi
Dedication
To my parents,
Kamala and Ganapathy Rao Krovi
ii
Acknowledgments
I am greatly indebted to Todd Brun for being an excellent advisor and mentor. Todd
gave a lot of his time especially during the initial stages of my PhD when it is so easy
to get lost. I am especially grateful for his guidance and the valuable insights he has
provided at various points which have led us to a lot of interesting results.
I would like to thank Daniel Lidar for his advice, for the useful discussions we had
on various topics and the project that I did with him. Thanks are due to Igor Devetak
for his time and help on the project that I did with him. Thanks go to Stephan Haas,
P. Vijay Kumar, Urbashi Mitra and Antonio Ortega for their help and advice in my
graduate studies.
Special thanks to Diane Demetras and Tim Boston for their help with many ad-
ministrative issues which have bothered me at times during my PhD. Thanks to Milly
Montenegro, Mayumi Thrasher and Gerrelyn Ramos for making CSI a great place to
work.
A big thank-you to Pinni, Babaiah, Ranu, Archana and Harsha for all their love and
support and for making me feel like I do the coolest thing on the planet.
iii
Table of Contents
Dedication ii
Acknowledgments iii
List Of Figures vi
Abstract viii
Chapter 1: Introduction and preliminaries 1
1.1 Quantum computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The rules of quantum mechanics . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Chapter 2: Random and quantum walks on graphs 8
2.1 Random walks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Quantum walks–discrete and continuous . . . . . . . . . . . . . . . . . . . 17
2.3 Cayley graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Chapter 3: Hitting times 25
3.1 Classical hitting time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Hitting time for quantum walks . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Other definitions of hitting time . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 Results for the Grover coin . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.5 Prior work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.6 Results for the DFT coin . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.7 Results for a quantum walk on a distorted hypercube. . . . . . . . . . . . 49
3.8 Infinite hitting times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Chapter 4: Symmetry in quantum walks 60
4.1 Automorphism groups of Cayley graphs . . . . . . . . . . . . . . . . . . . 61
4.2 Representations of finite groups . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3 Discrete-time walks on Cayley graphs . . . . . . . . . . . . . . . . . . . . 70
4.4 Continuous-time walks on Cayley graphs . . . . . . . . . . . . . . . . . . . 73
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
iv
Chapter 5: Quotient graphs 77
5.1 Action of an automorphism group . . . . . . . . . . . . . . . . . . . . . . 78
5.2 Quotient graphs and quantum walks . . . . . . . . . . . . . . . . . . . . . 81
5.3 Examples of quotient graphs. . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.4 Automorphism group of the quotient graph . . . . . . . . . . . . . . . . . 103
5.5 Hitting time on quotient graphs . . . . . . . . . . . . . . . . . . . . . . . . 105
5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Chapter 6: Quantum walks in decoherence-free subspaces 112
6.1 Decoherence-free subspaces and quantum walks . . . . . . . . . . . . . . . 113
6.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
References 121
v
List Of Figures
2.1 Examples of 2-colorable (square) and non-2-colorable (triangle) regular
graphs of degree 2, where the colors are numbered 1 and 2. . . . . . . . . 20
2.2 The hypercube in three dimensions as an example of a Cayley graph. . . . 22
3.1 Hitting times of Classical and Quantum walks (semi-log scale). . . . . . . 35
3.2 Comparison of Quantum hitting time, concurrent hitting time and the
bound on the concurrent hitting time for dimensions 10 to 20(log scale) . 36
3.3 Comparison of Quantum hitting time, concurrent hitting time and the
bound on the concurrent hitting time for dimensions 50 to 60(log scale) . 37
3.4 Hitting time (log scale) vsp for dimensionsn= 3 ton= 7 with dephasing
in the position and coin space. . . . . . . . . . . . . . . . . . . . . . . . . 40
3.5 Hitting time (log scale) vs p for dimensions n = 3 for small values of
decoherence. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.6 Hitting time (log scale) vs p for dimensions n = 3 for small values of
decoherence for an initial state which has an infinite hitting time. . . . . . 42
3.7 Hitting time (log scale) vsp for dimensionsn= 3 ton= 7 with dephasing
in the coin space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.8 Hitting time (log scale) vsp for dimensionsn= 3 ton= 7 with dephasing
in the position space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.9 Slope of the hitting time vs p for the dimension n = 3 with dephasing in
the position and coin space. . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.10 Comparison of hitting times on the regular and distorted hypercubes . . . 50
4.1 The direction-preserving automorphism group of the n=2 hypercube. . . . 63
vi
4.2 Automorphisms which interchange directions for the n=2 hypercube. . . . 64
5.1 The graph Γ(S
3
,{(1,2),(2,3)}) and its quotient graph. . . . . . . . . . . . 88
5.2 The graph Γ(S
3
,{(1,2),(1,3),(2,3)}) and its quotient graphs. . . . . . . . 92
5.3 ThegraphΓ(S
4
,{(1,2),(1,3),(1,4)})anditquotientgraphunderthesub-
group H. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.4 The n =3 hypercube and its quotient graphs. . . . . . . . . . . . . . . . . 97
5.5 The quotient graph of a general hypercube under the group S
n−1
. . . . . 99
5.6 The glued trees graph and its quotient graph. . . . . . . . . . . . . . . . . 102
vii
Abstract
A discrete-time quantum walk on a graph is the repeated application of a unitary evolu-
tion operator to a Hilbert space corresponding to the graph. Hitting times for discrete
quantum walks on graphs give an average time before the walk reaches an ending con-
dition. We derive an expression for hitting time using superoperators, and numerically
evaluate it for the walk on the hypercube for various coins and decoherence models. The
hitting time for a classical random walk on a connected graph will always be finite. We
show that, by contrast, quantum walks can have infinite hitting times for some initial
states. We seek criteria to determine if a given walk on a graph will have infinite hit-
ting times, and find a sufficient condition, which for discrete time quantum walks is that
the degeneracy of the evolution operator be greater than the degree of the graph. The
phenomenon of infinite hitting times is in general a consequence of the symmetry of the
graph and its automorphism group.
Symmetries of a graph, given by its automorphism group, can be inherited by the
evolution operator. Using the irreducible representations of the automorphism group, we
derive conditions such that quantum walks defined on this graph must have infinite hit-
ting times for some initial states. Symmetry can cause the walk to also be confined to a
subspace of the original Hilbert space for cartain initial states. We show that a quantum
viii
walk confined to the subspace corresponding to this symmetry group can be seen as a
different quantum walk on a smaller quotient graph. We give an explicit construction of
the quotient graph for any subgroupH of the automorphism group. The automorphisms
of the quotient graph which are inherited from the original graph are the original auto-
morphism group modulo the subgroup H used to construct it. We conjecture that the
existence of a small quotient graph with finite hitting times is necessary for a walk to ex-
hibit a quantum speed-up. Finally, we use symmetry and the theory of decoherence-free
subspaces to determine when the subspace of the quotient graph is a decoherence-free
subspace of the dynamics.
ix
Chapter 1
Introduction and preliminaries
1.1 Quantum computing
In the course of the 20th century, quantum mechanics was established as a fundamental
theory of physics. Every physical theory is built (or needs to be built) within the frame-
work of quantum mechanics. But quantum mechanics can seem very counter-intuitive
even to an expert. Superposition, interference and entanglement are some of the main
and bizarre aspects of quantum mechanics. These aspects can lead to many spectacular
and often puzzling effects in physical systems. One of the main goals of the field of quan-
tum computing and quantum information is to use these non-classical aspects to build
more powerful computers, communication devices and cryptographic systems.
The invention of the transistor by Bardeen, Brattain and Shockley has revolutionized
computer hardware. Computers have since been becoming more powerful as more and
more transistors are made to fit on an integrated circuit (IC). In 1965, Gordon Moore
characterized this growth in his famous Moore’s law which states that the number of
transistors on an IC doubles in number roughly every two years. This means that the
1
power of the computer doubles for the same cost roughly every two years. This law has
held sway for many years and seems to be valid even today. However, experts predict
that its rule will have run its time sometime in the next twenty years. As electronic
devices become smaller they will eventually run into fundamental difficulties of size due
to quantum effects. Conventional fabrication techniques are going to have a problem
when these quantum effects come into play. New technologies will have to replace them
to make progress. Quantum computing is a paradigm which can provide a solution to
this problem. This is another reason why research in quantum computing and quantum
information is critical.
Theoriginalideaofaquantumcomputercanbetracedbacktosimulationofquantum
systems. Any physical system adheres to the laws of quantum mechanics and hence it
is inherently a quantum system. Thus, understanding quantum systems is essential to
physics. Butanyreasonableandlarge-scalemodelofsuchasystemishardtosimulateon
a classical computer since it requires exponential resources. In the 1980s, it was realized
that an appropriate system which might efficiently simulate any quantum system is one
which is based on quantum mechanics. The essential idea of a such a quantum computer
was suggested by Manin [56], Benioff [11] and Feynman [31]. This idea was taken to the
next level by Deutsch, who introduced the notion of a quantum computer as a universal
computing machine [23, 24]. An early application of these concepts was in quantum
cryptography [12], where a cryptographic protocol was proposed which has since been
shown to be unconditionally secure [62]. The first instance of a computational problem
that could be solved faster by a quantum computer than by a classical computer was
givenbyDeutschandJosza[25]. Thenextproblemwhichhasafasterquantumalgorithm
2
than a classical one, was given by Simon [75]. Though these algorithms demonstrated
the power of quantum computing, the problems they solve are rather artificial in nature.
The discovery of quantum algorithms for prime factoring and discrete-logarithm by Shor
[74] was the first example where quantum algorithms work faster than the best known
classical algorithms for problems of great practical importance. Since then there has
been a surge of interest in quantum computing and quantum information which led to
the development of more areas, such as quantum Shannon theory, which was built from
theoriginalideasofquantumteleportationandsuperdensecoding[13]andquantumerror
correction.
Oneofthemostimportantgoalsofquantumcomputingisthedesignoffastalgorithms
for computational problems. A quantum algorithm for database search which works
faster than any classical algorithm was given by Grover [35]. The algorithms of Grover
[35] and Shor [74] are among the famous examples of quantum algorithms which have a
speed-up over classical algorithms. These two algorithms are very different in structure:
Grover’s algorithm exploits an invariant two dimensional subspace within the search
space, while Shor’s algorithm exploits the properties of the quantum Fourier transform
(QFT). Efficient algorithms for a class of problems called the hidden subgroup problem
(HSP)(factoringanddiscrete-logarithmbelongtothisclass)usetheQFT[53]. TheQFT
isusefulfortheAbelianversionoftheHSPandforsomenon-Abeliangroups[68]. Butfor
other non-Abelian groups, most notably, the symmetric group and the associated HSP–
the graph isomorhism problem, the power of the QFT seems to be limited. Grover’s
algorithm, although very useful in many search problems, gives only a quadratic speed
up, and hence a straightforward application of this algorithm is not very efficient for the
3
HSP. This is because it ignores structure in the problem which can be used to obtain a
speed up. New concepts and tools might be needed to design algorithms to solve these
problems. In this context, quantum walks might provide such tools. A goal of this thesis
is to explore the properties of quantum walks which might be useful in designing fast
algorithms.
We begin by giving a brief introduction to the basic concepts and definitions of quan-
tum computing which are used in this thesis. To get a more detailed explanation refer to
the book by Nielsen and Chuang [61].
1.2 The rules of quantum mechanics
Statespace. Anisolatedquantum system can be described by a quantum state whichis
a vector in a complex vector space endowed with an inner product (i.e., a Hilbert space).
Given an orthonormal basis {|ii} i ∈ S, for this Hilbert space, the state of the system
|αi can be written as |αi =
P
i
a
i
|ii, such that
P
i
|a
i
|
2
= 1. hα| denotes the complex
conjugate of the state and hα|βi denotes the inner product between two states |αi and
|βi.
Unitary evolution. The evolution of a closed system can be described by a unitary
transformation. In other words, the state of the system at two different times can be
related by a unitary operator. If|αi and|βi are the states at timest
1
andt
2
respectively,
then
|βi=
ˆ
U|αi, (1.1)
4
where
ˆ
U is unitary i.e.,
ˆ
U
ˆ
U
†
=
ˆ
I,
ˆ
U
†
is the transpose complex conjugate of
ˆ
U. Note that
the norm of the state does not change after the evolution, i.e., |hβ|βi| =|hα|
ˆ
U
†
ˆ
U|αi| =
|hα|αi|.
Measurement. Generalized quantum measurements are described by a set of mea-
surementoperators{M
i
}whichcanactontheHilbertspaceofthesystemandsatisfythe
completeness relation
P
i
M
†
i
M
i
=I. The indexi denotes the measurement outcome and
when a measurement is performed any one of these outcomes can occur with a certain
probability. Given a state of the system, say|αi before the measurement, an outcome k
occurs with a probability hα|M
†
k
M
k
|αi and if this outcome k occurs, then the resulting
state of the system is,
M
k
|αi
q
hα|M
†
k
M
k
|αi
. (1.2)
Projective measurements are a special case of generalized measurements in which the
measurement operators are projectors. An operator P is a projector if P
2
= P and
P = P
†
. Thus, in such measurements, the completeness relation becomes
P
i
P
i
= I,
where P
i
are the measurement operators. The operators P
i
also satisfy orthogonality
relations P
i
P
j
=δ
ij
.
Density operators. Suppose the exact quantum state is not known, but rather it
is known that the state belongs to a set |ψ
i
i where i indexes the set. If each |ψ
i
i could
be the state with a probability p
i
, then the density operator (also known as the density
matrix) for the system is defined as
ρ=
X
i
p
i
|ψ
i
ihψ
i
|. (1.3)
5
If the state of a quantum system is known exactly (|ψi, say), it is called a pure state. The
density operator corresponding to a pure state|ψi isρ=|ψihψ|. If the state is not pure,
it is called mixed. The unitary evolution and measurement postulates can be written in
the density operator picture in the following way. Given that the density matrices of a
closed quantum system at times t
1
and t
2
are ρ
1
and ρ
2
respectively, this evolution can
be related by a unitary operator as
ρ
2
=
ˆ
Uρ
1
ˆ
U
†
. (1.4)
Now,givenameasurementdescribedbymeasurementoperators{M
i
}(suchthat
P
i
M
†
i
M
i
=
I),theoutcomekoccurswithaprobabilityp(k) =Tr(M
†
k
M
k
ρ)andthestateofthesystem
is thenM
k
ρM
†
k
/Tr(M
†
k
M
k
ρ). For a projective measurement with operators{P
i
}, an out-
comek occurs with probabilityp(k) =Tr(P
k
ρ) and the resulting state isP
k
ρP
k
/Tr(P
k
ρ).
Density operators are characterized by the following two properties.
1. The trace of a density matrix is unity.
2. A density matrix is a positive matrix.
1.3 Outline
The rest of this thesis is organized as follows. In Chapter 2, we define classical random
walks and then give the definitions of quantum walks. Then we define Cayley graphs and
give a detailed description of the hypercube as an example of a Cayley graph.
6
In Chapter 3, we define a notion of hitting time and provide a formula for the hitting
time of a quantum walk on a graph. We then give the results of simulations for various
cases with and without decoherence for the hitting time on the hypercube. We also
provide evidence of infinite hitting times and then derive rigorously the conditions under
which a quantum walk has infinite hitting times.
In Chapter 4, we relate the notion of infinite hitting times to symmetry of the graph
and give examples of graphs with sufficient symmetry to have infinite hitting times. We
also provide the necessary definitions and results in the representation theory of finite
groups used in the analysis.
In Chapter 5, we use symmetry again to explain fast hitting times. We develop
the notion of a quotient graph and use it to show that quantum walks on a graph with
symmetryareactuallywalksonasmallerquotientgraph. Quantumwalkswithsymmetry
have invariant subspaces which can be used for algorithm design.
In Chapter 6, we make use of symmetry to find conditions on the decoherence such
that the subspace of a quotient graph will lie in a decoherence-free subspaces of the
dynamics.
7
Chapter 2
Random and quantum walks on graphs
Quantum walks were formulated in studies involving the dynamics of quantum diffusion
[30], but the analysis of quantum walks for use in quantum algorithms was first done by
Farhi and Gutmann [27]. The motivation for the development of quantum walks is three-
fold. First, algorithmdesigninclassicalcomputersciencehasbenefittedenormouslywith
the advent of randomized algorithms. The best known algorithms for some important
problems like 3-SAT (3-satisfiability) are based on random walks [59]. A natural question
iswhetherquantumwalkswouldperformbetterfortheseproblems. Second,asmentioned
above existing quantum algorithms for the so called hidden subgroup problem, based on
the quantum Fourier transform, like Shor’s algorithm and related algorithms [74], do not
seemtobeeffectiveforcertainnon-Abelianproblemslikethegraphisomorphismproblem.
Therefore, there is a need for a new class of algorithms to tackle these problems, and
quantum walk based algorithms may provide a new approach. Finally, a large number of
problems in computer science can be reformulated as graph related problems and having
a quantum formalism which is specific for graphs may be useful in designing algorithms
for these problems.
8
Quantum walks have been already been used in quantum algorithms which have a
speed-up over the corresponding classical algorithms. Childs et al. showed in [19] that
a continuous quantum walk would find a certain final node exponentially faster than a
classical random walk on the so-called “glued trees” graph. The classical walk tends to
take exponentially longer due to the large number of vertices in the middle of the graph,
where as a quantum walk would traverse it in a superposition of paths and move toward
the final vertex much faster. Shenvi et al. [72] showed that a quantum walk on an
unsorted database (represented as a hypercube) has a quadratic speed up over a classical
algorithm. The algorithm begins in a superposition of all the vertices and coin states and
proceeds by applying the “Grover” coin at every vertex except the final vertex where it
applies another coin (−I, which they point out in the paper, is an arbitrary choice since
other coins worked just as well in numerical simulations.) This makes the walk converge
to the final vertex in O(
√
N) steps, where N is the number of vertices. Quantum search
for a marked item in N items arranged on a d-dimensional N
1/d
×N
1/d
×...N
1/d
grid
was analyzed using a continuous-time walk by Childs and Goldstone in [20] and using
a discrete-time walk by Ambainis et al. in [7]. This is one of the few places where
discrete walks are better. An algorithm based on a continuous-time walk worked in
O(
√
N) time steps for d≥ 5, in O(
√
NlogN) for d = 4 and had no speed up in d = 2,3
whereas a discrete walk took O(
√
N) for d ≥ 3 and O(
√
NlogN) for d = 2. Ambainis
[4] has applied quantum walks to the element distinctness problem i.e., the problem of
determining whether N elements in a given set are all distinct or not. This algorithm is
based on a discrete-time quantum walk on the Johnson graph. Other algorithms based
9
on quantum walks are an algorithm for matrix product verification [17], triangle finding
[54] and group commutativity testing [55].
Several quantities of interest have been defined for quantum walks analogous to clas-
sical walks in [2], such as mixing time, sampling time, filling time and dispersion time. In
[2], a lower bound on the mixing time of a discrete quantum walk on theN-cycle is found
and shown to be at most polynomially faster than that of the classical walk. Hitting time
is another important quantity for classical walks on graphs. Two definitions of hitting
time are given in [38] and an upper bound for one of them was found for the walk on
a hypercube. A different definition of hitting time is given in [47], where the unitary
evolution of the discrete walk is replaced by a measured walk. In such a walk, after the
application of the unitary evolution operator, a measurement is performed to see if the
particle is in the final vertex or not. In [48] the phenomenon of infinite hitting times is
analyzed and it was shown that graphs with sufficient symmetry can have infinite hitting
times for certain initial states. This is a purely quantum phenomenon and does not have
a classical analogue. Cayley graphs on the symmetric group are shown to be examples
of graphs which have this symmetry and hence have an infinite hitting time for certain
starting states. The theory of irreducible representations is used to estimate the amount
of degeneracy that a given group of symmetries produces. The use of representation
theory to explain aspects of quantum walks on certain classes of graphs was also done
in [33], where the behavior of mixing times of Cayley graphs on the symmetric group is
explained based on its irreducible representations.
Quantum walks have applications other than in the design of new algorithms. In
[21,22],quantumwalksonweightedgraphshavebeenusedtoefficientlytransferquantum
10
states with perfect fidelity. Here and in [28], symmetry has been used to demonstrate a
class of graphs on which a continuous time quantum walk reduces to the walk on a line
(quantum wire) like in the case of the glued-trees graph.
Quantum walks come in two distinct flavors: discrete-time and continuous-time. The
main difference between them is that discrete time walks require a “coin”—which is just
any unitary matrix—plus an extra Hilbert space on which the coin acts, while continuous
time walks do not need this extra Hilbert space. Aside from this, these two versions
are similar to their classical counterparts. Discrete-time quantum walks evolve by the
application of a unitary evolution operator at discrete time intervals, and continuous-
time walks evolve under a (usually time-independent) Hamiltonian. Unlike the classical
case, the extra Hilbert space for discrete-time quantum walks means that one cannot
obtain the continuous quantum walk from the discrete walk by taking a limit as the time
stepgoestozero. Althoughthereisnonaturallimittogofromthediscretetocontinuous
walks for general graphs, for the quantum walk on the line [77] offers a treatment of this
limit, where it is possible to meaningfully extract the continuous-time walk as a limit of
thediscrete-timewalk. Thedynamicsofquantumwalksofbothtypeshasbeenstudiedin
detailforwalksonaninfiniteline—forthecontinuous-timecaseinRefs.[19,27,18,45,46]
and for the discrete-time case in [60, 10, 14, 15, 16]. There has also been considerable
work on other regular graphs. The N-cycle is treated in [2, 79], and the hypercube in
[72, 57, 38, 47, 48, 49, 50]. Quantum walks on general undirected graphs are defined in
[42,5], andondirectedgraphsin[58]. Reviewsofquantumwalksincludeanintroductory
review by Kempe in [37], and a review from the perspective of algorithms by Ambainis
in [5]. The role of symmetry in quantum walks has been analyzed in [47, 48, 49]. It has
11
been shown in [48] that the evolution operator of a quantum walk inherits symmetries
from the automorphisms of the graph and this leads to degeneracy in the operator. This
degeneracy can determine subspaces to which the walk remains confined. Thus, the walk
never explores some parts of the Hilbert space, leading to the phenomenon of infinite
hitting times. In [49], it was shown that due to the symmetry of the graph, certain
subspaces which confine the walk have fast hitting times, and moreover these subspaces
are the corresponding Hilbert spaces of a different graph - a quotient graph. Thus, the
walk is confined to smaller graph due to symmetry. The exponential speed-up observed
for quantum walks in [19, 72] on the “glued-trees” graph and the hypercube respectively
is due to the fact that the quantum walk is on a smaller quotient graph.
Decoherence in quantum systems can be broadly defined as any process that destroys
quantum coherence. Aharanov [1] first considered quantum walks with measurements.
Decoherence was later considered in quantum walks mainly to demonstrate the classical
behavior of such decohering quantum walks. Quantum to classical transition due to
decoherence has been considered for the walk on the line in Refs. [14, 15] and for the
hypercube in [3, 41]. Kendon et al [43] provide a treatment of this transition for any
quantum walk. More recently, decoherence has been shown to be beneficial for fast
mixing behavior of quantum walks. For the continuous-time walk on the hypercube,
decohering quantum walks have been shown to mix faster [63, 64]. Kendon [44] gives a
review of the work done in this field so far, focusing mainly on decoherence. Studies of
decoherence in quantum walks is useful not only to understand the quantum behavior of
the walk but also for possible implementations. Any implementation of a quantum walk
will involve having to deal with decoherence. Such studies may be useful in determining
12
the kind of decoherence that has little or no effect on the useful properties of the walk.
Some implementation schemes proposed so far are in Refs. [26, 29, 43, 70, 69].
2.1 Random walks
A simple random walk on an undirected graph is defined as the repeated application of a
stochastic matrixP, whereP(i,j) =1/d
i
ifi andj are connected and 0 otherwise, where
P(i,j) is the probability to go from vertex i to vertex j and d
i
and d
j
are the degrees
of these vertices. Such a process is called a Markov chain. If the graph is connected
and non-bipartite then a fundamental property of Markov chains is that the distribution
tendstoastationarydistribution(π)whichisindependentoftheinitialdistribution(p
0
).
If the graph is regular (if every vertex is connected to the same number of other vertices),
then this final distribution is uniform over all the vertices. By contrast, quantum walks
do not converge to a final distribution since norms of states do not change under a
unitary operation and hence the distance between the states describing the system does
not converge to zero ([2], [37]). The induced probability distribution does not converge
either, but it turns out that the time-averaged probability distribution converges. The
time averaged distribution can be defined as
¯
P
T
(v|α
0
) =
1
T
T−1
X
t=0
P
t
(v|α
0
), (2.1)
where|α
0
i is the initial state, P
t
is the instantaneous distribution and v is any vertex of
the graph.
13
The rate of convergence to this final distribution can expressed in terms of many
quantities, but the one used commonly is mixing time. Mixing time is defined as
M
= min{T|∀t≥T,p
0
:||p
t
−π||≤}, (2.2)
where||p−q||=
P
i
|p
i
−q
i
|isthetotalvariationdistancebetweenthedistributionspand
q. Mixing time for a classical walk is related to the spectral gap (the difference between
the largest and the second largest eigenvalues of P) in the following way:
λ
2
(λ
1
−λ
2
)log2
≤M
≤
1
(λ
1
−λ
2
)
(max
i
logπ
−1
i
+log
−1
), (2.3)
where λ
1
is the largest eigenvalue and λ
2
is the second largest eigenvalue. If the graph
is regular, it turns out that the largest eigenvalue of P (i.e., λ
1
) is 1. The above relation
connects mixing time and the second largest eigenvalue of the stochastic matrix. The
classical mixing time has been found for a number of graphs. For theN-cycle the mixing
time of a simple random walk can be calculated to be O(N
2
log(1/)). The mixing time
for the hypercube turns out to be O(dlogdlog(1/)), where d is the dimension of the
hypercube. The hypercube is an example of a class of graphs called expander graphs for
which it has been shown that a classical random walk is rapidly mixing. But quantum
walks on the other hand do not have good mixing properties on the hypercube. The
discrete-time quantum walk has a mixing time of at least O(d
3/2
/) [57], where d again
is the dimension of the hypercube. But a quantity that looks promising for quantum
walks on the hypercube is hitting time. Hitting time measures the average time it takes
14
for a walk to reach a certain vertex from a given starting vertex. The analysis of hitting
times under different situations is one of the aspects of this thesis. In order to motivate
the analysis of hitting times of quantum walks on the hypercube, we briefly present the
following classical random walk based algorithm.
Classical random walks have been used in many randomized algorithms in computer
science. Oneoftheapplicationsofrandomwalkstoalgorithmsisforthe3-satisfiability(3-
SAT) problem where the best known classical algorithm is based on the hitting time of a
randomwalkonthehypercube. The3-SATproblemcanbedefinedasfollows. Considera
setofnliterals{x
1
,x
2
,...,x
n
},eachtakingavalueof0or1,andasetofclausessuchthat
each clause contains a logical OR of only three literals (or their negations). The problem
of 3-satisfiability consists of finding an assignment for the literals such that each of the
clausesissatisfied(i.e.,eachhasavalue1). Forexample, (x
1
∨¬x
2
∨¬x
3
)∧(¬x
1
∨x
2
∨x
4
)
is an expression with two clauses, where ∧ and ∨ denote the binary operations of AND
and OR and ¬x denotes the negation of x. (x
1
,x
2
,x
3
,x
4
) = (1,0,0,1) is an assignment
of the literals that satisfies both the clauses and thus represents a solution. The classical
random-walk based algorithm takes O(1.329...
n
poly(n)) steps [65, 73]. It consists of the
following steps.
1. Choose a random initial assignment for the literals.
2. Repeat 3n times
(a) If all the clauses are satisfied, then stop.
(b) If not, pick one unsatisfied clause, choose one literal uniformly at random and
flip it.
15
This algorithm is a random walk on the hypercube since an initial assignment of the
literals represents a vertex on the hypercube and flipping a bit of this vertex represents
moving along an edge connected to it. Thus analyzing the behavior of hitting times
of quantum walks on the hypercube may give us important clues to design quantum
algorithms for the 3-SAT. However, in the above algorithm, the walk is on a directed
hypercube since after moving along an edge to a new vertex, the literal that was flipped
need not occur in an unsatisfied clause. This means that it cannot be flipped at the new
vertexandtherefore,thereisnopathfromthisvertextotheoriginalvertex. Thisdirected
nature of the graph causes problems when we try to make this a quantum algorithm. A
quantum walk as with any quantum computation, needs to be unitary. It has been
shown in [58] that on a directed graph, it is possible to define a unitary walk if and only
if the graph is reversible. A directed edge of a graph going from vertex v
i
to v
j
is called
reversible if there is a path fromv
j
tov
i
. A graph is called reversible if every edge in it is
reversible. Thus, givenanexpressionfor3-SAT,onecoulddefineaunitaryquantumwalk
if the corresponding directed graph of the problem is reversible. This may be overcome
by defining a non-unitary walk, one that involves measurements, if it is not reversible.
But the next step, which is to design an algorithm based on this walk, is not clear. There
does not exist a quantum walk algorithm for 3-SAT presently, but a quantum algorithm
based on Grover search and amplitude amplification which works in O(1.153...
n
poly(n))
steps is presented in [8].
16
2.2 Quantum walks–discrete and continuous
Quantum walks as noted earlier, are primarily of two types. Depending on the way
the evolution operator is defined, they can be either discrete-time or continuous-time
quantum walks. These two definitions of quantum walks are not exactly equivalent to
the two types of classical random walks. While they are both based on the classical
definitions, unlike the classical case the discrete quantum walk does not reduce to the
continuous walk when we let the time step between repeated applications of the unitary
tend to zero. This is because discrete-time walks need an extra Hilbert space, called the
“coin” space (from the idea that one flips a coin at each step to determine which way
to walk), and taking the limit where the time step goes to zero does not eliminate this
Hilbert space. Therefore, the properties of discrete and continuous walks are different.
Though there is no obvious reason why one should be preferred, in some cases it has been
shown that coins make these walks faster [7].
2.2.1 Discrete-time walks
A discrete-time quantum walk can broadly be defined as the repeated application of
a unitary evolution operator on a Hilbert space whose size depends on the graph. This
Hilbertspaceusuallyconsistsofthespaceofpossiblepositions(i.e., thevertices)together
withthespaceofpossibledirectionsinwhichtheparticlecanmovealongfromeachvertex
(the coin space). All the concepts which we develop in this thesis such as a formula for
hitting time of a dicrete-time quantum walk, infinite hitting times, effect of symmetry
and quotient graphs are applicable to any undirected graph. However, in most of our
17
simulations we considerd-regular, undirected andd-colorable graphs. This is because the
structure of quantum walks reduces to something more manageable in this case. We will
briefly review the definitions of these graph properties.
A regular graph is one where every vertex is connected to the same numberd of other
vertices. This number is called the the degree of the graph. A graph is undirected if for
everyedgebetweenverticesAandBgoingfromAtoB,anedgegoesfromBtoAaswell.
In this case, we identify the edge from A to B with the edge from B to A, and consider
them a single edge. A regular, undirected graph withN vertices of degreed is considered
d-colorable if the edges incident on every vertex can be numbered 1 through d such that
every edge between two vertices has the same number at either end. Not all d-regular
and undirected graphs can be d-colored. A simple example is the triangle graph where
N = 3 and d = 2. (See Fig. 2.1.) For d-regular, undirected and d-colored graphs, the
Hilbert space of the walk isH
p
⊗H
c
, i.e., the tensor product of the position and direction
(or coin) space. The evolution operator
ˆ
U is given by
ˆ
U =
ˆ
S(
ˆ
I⊗
ˆ
C), where
ˆ
S is called
the shift matrix and
ˆ
C is the coin matrix. The shift matrix encodes the structure of the
graph and is very similar to its adjacency matrix. The vertices, numbered |0i through
|N−1i, are basis states for the vertex Hilbert spaceH
p
and the set of all directions from
each vertex, numbered|1i through|di, are basis states for the coin Hilbert spaceH
c
. In
this basis, the shift matrix for the graph can be given the explicit form:
ˆ
S =
X
v
X
i
|v(i),iihv,i|,
where v(i) is the vertex connected to v along the edge numbered i.
18
The coin matrix
ˆ
C acts only on the coin space, and “flips” the directions before the
shift matrix is applied. Then
ˆ
S moves the particle from its present vertex to the vertex
connected to it along the edge indicated by the coin direction. Though
ˆ
C can be any
unitary matrix, usually coins with some structure are considered. The coins that we used
inourpreviousanalysisaretheGrovercoin
ˆ
GandtheDiscreteFourierTransform(DFT)
coin
ˆ
D. The matrices for these coins are given by:
ˆ
G=2|ΨihΨ|−I =
2
d
−1
2
d
...
2
d
2
d
2
d
−1 ...
2
d
.
.
.
.
.
.
.
.
.
.
.
.
2
d
2
d
...
2
d
−1
, (2.4)
and
ˆ
D =
1
√
d
1 1 1 ... 1
1 ω ω
2
... ω
d−1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1 ω
d−1
ω
2(d−1)
... ω
(d−1)(d−1)
, (2.5)
where|Ψi=
1
√
d
P
i
|ii and ω = exp(2πi/d).
2.2.2 Decoherence in discrete-time walks
Decoherence in any quantum system is a process which destroys quantum superpositions.
Any quantum system will have decoherence (quantum noise). Thus any practical imple-
mentation scheme of quantum walks must deal with decoherence. In general, it can be
19
1
1
2 2
1
1
2 2
2 2
1
1
1
2
Figure2.1: Examplesof2-colorable(square)andnon-2-colorable(triangle)regulargraphs
of degree 2, where the colors are numbered 1 and 2.
thought of as any completely positive and trace preserving (CPTP) map. An arbitrary
CPTP map can be represented using the operator sum (OSR) or Kraus representation.
D(ρ) =
X
i
ˆ
A
i
ρ
ˆ
A
†
i
, (2.6)
where the operators
ˆ
A
i
are the Kraus operators and satisfy the relation:
P
i
ˆ
A
†
i
ˆ
A
i
=
ˆ
I.
Thus in a discrete-time quantum walk with decoherence, we can assume that a step of
the evolution is of the form
E(ρ) =D◦U(ρ), (2.7)
whereU(ρ)=
ˆ
Uρ
ˆ
U
†
as before.
2.2.3 Continuous-time walks
Continuous time quantum walks were defined by Farhi and Gutmann in [27]. For an
undirected graph G(V,E), the unitary evolution operator is defined as
ˆ
U = exp(i
ˆ
Ht),
where
ˆ
H is obtained from the adjacency matrix of the graph. Here again, the vertices of
20
the graph form a basis for the Hilbert space on which
ˆ
U is defined. This gives rise to the
following Schr¨ odinger equation:
i
d
dt
hv|ψ(t)i=hv|
ˆ
H|ψ(t)i. (2.8)
This walk has a structure very similar to that of continuous time Markov chains.
ˆ
H is
defined as
ˆ
H
i,j
=
−γ i6=j if nodes i and j connected
0 i6=j if nodes i and j not connected
d
i
γ i=j
(2.9)
where γ is the jumping rate from a vertex to its neighbor i.e., the transitions between
connected vertices happen with a probability γ per unit time.
But for a regular graph we can take
ˆ
H to be the adjacency matrix because d
i
= d,
where d is the degree of the graph. This means that the Hamiltonian can be written as
ˆ
H =γ(D−A), where D =dI and A is the adjacency matrix of the graph. The matrix
D would lead to a trivial phase factor and can be dropped.
ˆ
H is a symmetric matrix
(and hence
ˆ
U unitary) if the graph is undirected. Therefore, for a regular and undirected
graph, the adjacency matrix
ˆ
H, which acts as the Hamiltonian, is of the form:
H
i,j
=
1 if i and j share an edge,
0 otherwise.
(2.10)
21
As can be seen, this walk has no coin and so the Hilbert space on which
ˆ
U acts is only
the vertex spaceH
p
.
2.2.4 Decoherence in continuous-time walks
Continuous-time quantum walks evolve by an application of a Hamiltonian as defined in
Eq. (2.9). Decoherence in this scenario would give rise to an evolution which can be
described using the Lindblad semigroup master equation as
˙ ρ(t)=[
ˆ
H,ρ(t)]+
X
i
(
ˆ
L
i
ρ(t)
ˆ
L
†
i
+
ˆ
L
†
i
ˆ
L
i
ρ(t)+ρ(t)
ˆ
L
†
i
ˆ
L
i
), (2.11)
where
ˆ
L
i
aretheLindbladoperators. TheLindbladoperatorsrepresenttheenvironmental
interactions which lead to decoherence.
2.3 Cayley graphs
000 001
010
100
101
111
110
011
Figure 2.2: The hypercube in three dimensions as an example of a Cayley graph.
CayleygraphsaredefinedintermsofagroupGandasetS consistingofelementsfrom
Gsuchthattheidentityelemente / ∈S. GivenGandS,theresulting(right)-Cayleygraph
22
Γ(G,S) is one whose vertices are labeled by the group elements, i.e., there is one vertex
foreverygroupelement, andtwoverticesg andhareconnectedbyadirectededgefromg
toh ifg
−1
h∈S, (see [34]). Another way to look at this definition is that from any vertex
g ofaCayleygraph,thereare|S|outgoingedges,onetoeachoftheverticesgs,∀s∈S. A
CayleygraphwillbeconnectedifandonlyifthesetS isageneratingsetforG,anditwill
be undirected ifs
−1
∈S,∀s∈S. The degree of such a graph is|S|, the cardinality of the
generating set. Finally, a d-regular Cayley graph can be d-colored if s
2
= 1,∀s∈S, i.e.,
s
−1
=s. Examples of Cayley graphs on which quantum walks have been studied include
the line Γ(Z,{1,−1}); the cycle Γ(Z
n
,{1,−1}); the hypercube Γ(Z
n
2
,X) where the set
X is the set of canonical generators {(1,0,0,··· ,0),(0,1,0,··· ,0),...,(0,0,0,··· ,1)};
and the graph on the symmetric group Γ(S
n
,Y), whereY is a generating set forS
n
. Let
us look at the hypercube as an example of a Cayley graph where quantum walks have
beenextensivelystudied(seeFig.(2.2)). The hypercube has|Z
n
2
|= 2
n
verticeseachwith
a degree of |X| = n. The vertices can be labeled by an n-bit string from (0,0,··· ,0)
through (1,1,··· ,1). Two vertices are adjacent if they differ only by a single bit. Vertex
~ v is connected ton vertices given by~ v⊕~ s,∀~ s∈X, where⊕ stands for the bit-wise XOR
of the bit strings ~ v and ~ s. One important property of the hypercube is that it can be
n-colored, since ~ s⊕~ s =~ e, ∀~ s ∈ X where ~ e = (0,0,··· ,0) is the identity element. The
unitary evolution operator for a discrete walk on the hypercube becomes
ˆ
U =
ˆ
S(
ˆ
I⊗
ˆ
C),
where
ˆ
S has the form
ˆ
S =
X
~ s
X
~ v
|~ v⊕~ sih~ v|⊗|~ sih~ s|.
23
Since the vertices of the hypercube are bit strings, and adjacent vertices are those that
differ by one bit, the shift matrix of the discrete walk on the hypercube has a natural
form given by
ˆ
S =
ˆ
X⊗
ˆ
I⊗···⊗
ˆ
I⊗|~ s
1
ih~ s
1
|+
ˆ
I⊗
ˆ
X⊗···⊗
ˆ
I⊗
ˆ
|~ s
2
ih~ s
2
|
+...+
ˆ
I⊗
ˆ
I⊗···⊗
ˆ
X⊗|~ s
n
ih~ s
n
|, (2.12)
where
ˆ
X stands for the Pauliσ
x
operator. This structure of
ˆ
S reflects the property of the
hypercube that moving along an edge from~ v corresponds to flipping one bit of~ v. This
structure is also useful in determining its group of symmetries as we shall see later.
24
Chapter 3
Hitting times
3.1 Classical hitting time
Given a regular undirected graph and a particle which starts at some vertex, the classical
random walk is defined as before. At each vertex, the particle moves along any edge
incident on the vertex with some predefined probability. This procedure is then repeated
at the new vertex. The walk continues until the particle arrives at (“hits”) a certain
vertex (called the “final vertex”) for the first time. The hitting time is defined as the
average time until the particle hits the final vertex:
τ(v) =
∞
X
t=0
tp
v
(t), (3.1)
where τ(v) is the hitting time given that the walk starts at vertex v and p
v
(t) is the
probability that the particle hits the final vertex for the first time at time step t (first
crossing probability) given that it was at v at t= 0.
Let us now specialize to the case of the hypercube, where the the final vertex is
assumed to be 11···1. We would like to find the hitting time starting from 00···0.
25
For the classical walk on the hypercube, one can arrive at a recursive relation involving
the hitting time. First, from the symmetry of the hypercube one can conclude that the
hitting time depends only on the Hamming weight of the starting vertex rather than the
vertexitself. TheHammingweightisthenumberof1’sinthestringofbits. AtHamming
weightx, there areC
n
x
=n!/x!(n−x)! vertices. The probability to walk to a vertex with
weight x+1 is (n−x)/n, and the probability to walk to a vertex with weight x−1 is
x/n. So, ifτ(x) denotes the hitting time starting at any vertex with Hamming weightx,
then
τ(x) =
n−x
n
τ(x+1)+
x
n
τ(x−1)+1, (3.2)
with the boundary condition τ(n) =0. This simplifies to
Δ(x) =
n−x−1
x+1
Δ(x+1)−
n
x+1
, (3.3)
where Δ(x)=τ(x)−τ(x+1). Using this recursive formula, we obtain
τ(0)=
n−1
X
x=0
Δ(x) =
n−1
X
x=0
P
x−1
j=0
C
n
x−j
+1
C
n−1
x
, (3.4)
This sum can readily be evaluated for reasonable sizes of n and in fact this sum scales
as ≈ 2
n
. We use this expression to compare the classical hitting time to the quantum
hitting time. We define the hitting time of a quantum walk next.
26
3.2 Hitting time for quantum walks
The hitting time τ
h
of a classical random walk is defined as the average time for the
walk to hit a designated ‘final’ vertex v
f
given that the walk began with some initial
distribution p
i
:
τ
h
=
∞
X
t=0
tp(t), (3.5)
where p(t) is the probability of being in the final vertex for the first time at time step t.
In order to carry this notion of hitting time over to the quantum case, we need to make
the meaning of p(t) more precise. In particular, we need to define clearly what “for the
first time” means for a quantum walk. As described in [47], we do this by performing
a projective measurement of the particle at every step of the walk to see if the particle
has reached the final vertex or not. The measurementM which is used has projectors
ˆ
P
f
and
ˆ
Q
f
=
ˆ
I−
ˆ
P
f
representing the particle being found or not found at the final vertex,
respectively. The projector is defined
ˆ
P
f
= |x
f
ihx
f
|⊗
ˆ
I
c
, where |x
f
i is the final vertex
state and
ˆ
I
c
is the identity operator on the coin space. Using this definition, each step of
the measured walk consists of an application of the unitary evolution operator
ˆ
U followed
by the measurement M.
Now we can use the same expression (3.5) for the hitting time, where the probability
p(t) becomes
p(t)=Tr{
ˆ
P
f
ˆ
U[
ˆ
Q
f
ˆ
U]
t−1
ρ
0
[
ˆ
U
†ˆ
Q
f
]
t−1 ˆ
U
†ˆ
P
f
}. (3.6)
27
To explicitly sum the series in Eq. (3.5) using the expression for p(t) in Eq. (3.6), we
rewrite the expression in terms of superoperators (linear transformations on operators)N
andY, defined by
Nρ=
ˆ
Q
f
ˆ
Uρ
ˆ
U
†ˆ
Q
f
Yρ=
ˆ
P
f
ˆ
Uρ
ˆ
U
†ˆ
P
f
. (3.7)
In terms of N and Y, p(t) = Tr{YN
t−1
ρ
0
}. We introduce a new superoperator O(l)
which depends on a real parameter l:
O(l) =l
∞
X
t=1
(lN)
t−1
, (3.8)
which is a function of the parameter l. The hitting time now becomes
τ
h
=
d
dl
Tr{YO(l)ρ
0
}
l=1
. (3.9)
If the superoperatorI−lN is invertible, then we can replace the sum Eq. (3.8) with
the closed form
O(l) =l(I−lN)
−1
. (3.10)
(The case when I−lN is not invertible is discussed in detail later.) The derivative in
Eq. (3.9) is
dO
dt
(1)=(I−N)
−1
+N(I−N)
−2
= (I−N)
−2
. (3.11)
28
This gives us the following expression for the hitting time:
τ
h
= Tr{Y(I−N)
−2
ρ
0
}. (3.12)
ToevaluateEq. (3.12), wewritethesesuperoperatorsasmatricesusingRoth’slemma
[66]. Asshownin[47],wecanthenvectorizethedensityoperatorsandoperatorsonstates,
and write the action of superoperators as simple matrix multiplication. Any matrix can
be vectorized by turning its rows into columns and stacking them up one by one, so that
a D×D matrix becomes a column vector of size D
2
. For example:
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
→
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
.
Consequently the superoperators become matrices of size D
2
×D
2
. This method of
vectorization takes operators on one Hilbert spaceH to vectors in another Hilbert space
29
H
0
=H⊗H
∗
and so superoperators inH are operators onH
0
. Note that a basis{|u
ij
i}
forH
0
can be obtained from a basis{|v
i
i} forH by defining
|u
ij
i=|v
i
i⊗|v
j
i
∗
. (3.13)
For our superoperatorsN andY we then get
(Nρ)
v
=
h
(
ˆ
Q
f
ˆ
U)⊗(
ˆ
Q
f
ˆ
U)
∗
i
ρ
v
,
(Yρ)
v
=
h
(
ˆ
P
f
ˆ
U)⊗(
ˆ
P
f
ˆ
U)
∗
i
ρ
v
. (3.14)
Let N=(
ˆ
Q
f
ˆ
U)⊗(
ˆ
Q
f
ˆ
U)
∗
and Y = (
ˆ
P
f
ˆ
U)⊗(
ˆ
P
f
ˆ
U)
∗
. The hitting time becomes
τ
h
=I
v
·
Y(I−N)
−2
ρ
v
. (3.15)
Using this vectorization transformation, we treat the superoperators as operators on a
largerHilbertspaceandthuscanfindtheirinverses(whentheyexist). But,theexpression
in Eq. (3.15) is not always well defined because the matrix I−N may not be invertible.
Wewillshowthatwhenitisnotinvertible,itmeansthatthehittingtimebecomesinfinite
for some initial states, and vice versa. This property of quantum walks having infinite
hitting times does not have a classical analogue.
30
3.3 Other definitions of hitting time
Two definitions of the quantum hitting time were given in [38]: one-shot hitting time and
concurrent hitting time. The one-shot hitting time is defined for an unmeasured walk. It
is the time at which the probability of being in the final state is greater than some given
value. More precisely, given some probability p, the one-shot hitting time is defined as
the lowest time τ
sh
(p) such that
|hx
f
|
ˆ
U
τ
sh
(p)
|x
0
i|
2
≥p (3.16)
wherex
f
andx
0
arethefinalandinitialstates,and
ˆ
U istheevolutionoperator(asdefined
above). Essentially this same definition of hitting time was used in the analysis of the
continuous-time walk on the hypercube in [3]. This definition is useful if it is known that
at some time the probability to be in the final state will be higher than some reasonable
value; but for a general graph, this is not guaranteed.
The concurrent hitting time, by contrast, is defined for a measured quantum walk.
Given a probability p, it is the time τ
c
(p) such that the measured walk has a probability
greater than p of stopping at a time less than τ
c
(p). It has been proved that the con-
current hitting time is τ
c
(p) =
π
2
for p = Ω(
1
nlog
2
n
) for a hypercube of dimension n (for
the symmetric initial condition used in this paper). Since we consider only the measured
quantum walk in our analysis, we compare our numerical results to the numerical sim-
ulation of the concurrent hitting time and the bound on it derived in [38]. In the next
section, this is redefined in terms of the residual probability 1−p and plotted against the
hitting time defined in the previous section.
31
If we think of quantum walks as a possible route to new algorithms, then the con-
current hitting time corresponds to the time needed to find a solution with probability
greater than p. The definition of hitting time used here corresponds more to a typical
running time for the algorithm. Both definitions could prove useful for particular pur-
poses. In particular, both definitions give different characterizations of the distribution
p(t)—the probability of first being in the final vertex at time t—in different ways. A
complete understanding of the time needed to find a solution would require knowledge
of the entire distribution p(t) (or τ
c
(p) for all p), which is unlikely to be achievable, in
practice; but τ
h
and τ
c
(p) (for fixed p) give different windows on this function.
3.4 Results for the Grover coin
We calculated the average hitting time by evaluating the above expression (3.15) in
Matlab. Because of the multiple tensor products, the size of the matrices N and Y
is (2
n
n)
2
×(2
n
n)
2
. Asn increases, the size of these matrices increases exponentially, and
the matrix inversion that Eq. (3.15) demands is not easy to compute. We can always
compute an estimate of the quantum hitting time (strictly, a lower bound), obtained by
iterating the quantum walk for a large number of steps. To get this lower bound, we
define τ
est
() using Eq. (3.5), and summing the series up to a finite number of terms:
τ
est
() =
τc(1−)
X
t=1
tp(t), (3.17)
32
where the concurrent hitting time τ
c
(1−), as defined in [37], is the shortest time T for
which
T
X
t
p(t)≥1−.
Thiscomparisonoftheexactandapproximatevaluesoftheaveragehittingtimeisuseful,
because the matrices in Eq. (3.15) grow faster than those used in summing Eq. (3.5),
which means that for some graphs it may be impractical to calculate the exact value of
τ
h
, but still possible to calculate the lower boundτ
est
(). However, for the quantum walk
on the hypercube with the Grover coin, when the initial state of the walk is a particular
symmetric state, it is possible to reduce the matrix sizes considerably since the walk
remains in a lower dimensional subspace of the original space, as explained below. For
such an initial state we can compute the exact average hitting time (Eq. (3.15)) up to a
large number of dimensions. Figure 3.1 shows both the classical and quantum walks on
the hypercube for dimensions up to 100. The exact average hitting timeτ
h
is plotted as a
dotted line, and the lower boundτ
est
() for= 0.001 is plotted as a solid line. These two
lines almost coincide in the graph, and we can see that when is small enough summing
the series gives a very good estimate of τ
h
. We conjecture that this will remain true for
more general graphs as well. In comparing classical and quantum results, the average
hitting time for the quantum walk is a low order polynomial, whereas the classical walk
grows exponentially with dimension, so there is a very dramatic speed-up in the quantum
case.
The simplification that makes this computation tractable is in the case where the
coin-flip unitary is the Grover coin
ˆ
G defined in (2.4), and the specific starting state is
33
ρ=|0ih0|⊗|φihφ|, where|φi=
1
√
n
P
n
i=1
|ii. Itwas first observedin[57]thatforthewalk
on the hypercube with this initial state and the Grover coin, the state remains always in
a 2n-dimensional subspace, where the walk is on a line with n+1 points. This is rather
similar to the simplification made in the classical case, when we kept track only of the
Hamming weight of the current vertex. With this simplification, the operators in Eq.
(3.12)reduceto (2n)
2
×(2n)
2
matrices obtained in [72], which makes explicit calculations
possible even for high dimensional hypercubes.
We write a set of basis states for this subspace as |R,0i,|L,1i,|R,1i,...,|R,n−
1i,|L,ni, where the first label says whether the state is “right-going” or “left-going”, and
the second label gives the Hamming weight of the state. The initial state is |R,0i, and
the final state is|L,ni. (Note that there are no states|L,0i or|R,ni.) Restricted to this
subspace, the
ˆ
S and
ˆ
C matrices become
S =
n
X
x=0
|R,xihL,x+1|+|L,x+1ihR,x| (3.18)
and
C|L,xi = −cosω
x
|L,xi+sinω
x
|R,xi,
C|R,xi = sinω
x
|L,xi+cosω
x
|R,xi, (3.19)
where cosω
x
= 1−2x/n. We see that for the walk in this subspace, the coin flip is no
longerindependentoftheposition; thisisquiteanalogoustothereductionoftheclassical
walk to the Hamming weight, in which the probabilities favor walking toward x=n/2.
34
How does the average hitting time compare to the concurrent hitting time for the
walk on the hypercube? Figure 3.2 plots the estimate of the average hitting time τ
est
(),
the concurrent hitting time τ
c
(1−), and the bound on τ
c
(1−) (obtained in [38]) as
a function of for dimensions from 10 to 20. (Both the axes are in log scale.) Figure
3.3 plots the same for dimensions from 50 to 60. We can see from these two figures that
bothτ
c
(1−) and the bound on it become less tight for higher dimensions, and both are
much longer than the bound on the average hitting time τ
est
(). This, together with the
comparison ofτ
est
() to the exact valueτ
h
, strongly indicates that the average walk ends
much faster than these bounds might suggest, and also that the estimate τ
est
() is quite
insensitive to the choice of , at least for the hypercube.
0 20 40 60 80 10
10
0
10
5
10
10
10
15
10
20
10
25
10
30
10
35
Dimension of hypercube
Hitting times (Log scale)
!
est
(") for "=0.001
classical hitting time
exact quantum hitting time !
Figure 3.1: Hitting times of Classical and Quantum walks (semi-log scale)
35
10
!3
10
!2
10
10
1
10
2
10
3
10
4
10
5
! (log scale)
Log scale
Bound on "
c
(1!!)
Concurrent hitting time "
c
(1!!)
Hitting time "
est
(!)
Figure 3.2: Comparison of Quantum hitting time, concurrent hitting time and the bound
on the concurrent hitting time for dimensions 10 to 20(log scale)
3.4.1 Hitting time in the presence of decoherence
Thisdecoherencecanbethoughtofasactingonthesystemateverystepofthemeasured
walk. Thus the evolution of the walk can be decomposed into three parts–the unitary
evolution operator followed by the decoherence (CPTP) map and then the partial mea-
surement on the final vertex all taking place in sequence. The combined effect can be
seen as the application of a superoperator at every step and it can be written as
ρ(t+1) = E(ρ(t))
= M◦D◦U(ρ(t))
=
ˆ
M
j
X
i
ˆ
A
i
(
ˆ
Uρ(t)
ˆ
U
†
)
ˆ
A
†
i
!
ˆ
M
†
j
, (3.20)
36
10
!3
10
!2
10
10
2
10
3
10
4
10
5
10
6
! (log scale)
Log scale
Bound on "
c
(1!!)
Concurrent hitting time "
c
(1!!)
Hitting time "
est
(!)
Figure 3.3: Comparison of Quantum hitting time, concurrent hitting time and the bound
on the concurrent hitting time for dimensions 50 to 60(log scale)
where
ˆ
M
j
is the measurement operator with the outcome j ∈{0,1} with
ˆ
M
0
=
ˆ
P
f
and
ˆ
M
1
=
ˆ
Q
f
. The state of the system after n steps can be written as
ρ(t) =E
t
(ρ(0)). (3.21)
The formula for hitting time in Eq. (3.15) assumes a purely unitary evolution and
partial measurement. We now modify this formula to include the effect of decoherence on
thewalk. Inthepresenceofdecoherence, theevolutionwithoutmeasurementisdescribed
by
ρ(t+1) =D◦U(ρ(t)). (3.22)
37
Therefore, we only need to replace the superoperator U to D◦U. This means that the
vectorized quantities Y and N are modified accordingly. They become
Y
D
= (
ˆ
P
f
⊗
ˆ
P
∗
f
)(
X
i
ˆ
A
i
⊗
ˆ
A
∗
i
)(
ˆ
U⊗
ˆ
U
∗
)
N
D
= (
ˆ
Q
f
⊗
ˆ
Q
∗
f
)(
X
i
ˆ
A
i
⊗
ˆ
A
∗
i
)(
ˆ
U⊗
ˆ
U
∗
), (3.23)
and we can write D =
P
i
ˆ
A
i
⊗
ˆ
A
∗
i
as the decoherence superoperator. The structure of
the formula for hitting time given by Eq. (3.15) remains the same i.e.
τ
h
=I
v
·
Y
D
(I−N
D
)
−2
ρ
v
0
. (3.24)
3.4.2 Dephasing in the position and coin space
Complete dephasing has been considered as a kind of decoherence for the discrete-time
quantum walk on the line [41, 14, 15] and the hypercube in [41]. Here we analyze the
effect of this kind of dephasing on the hitting time. Consider the discrete-time quantum
walk on the hypercube. The evolution operator is given by
ˆ
U =
ˆ
S(
ˆ
I ⊗
ˆ
C) where
ˆ
S is
given by Eq. (2.12) and
ˆ
C is the Grover coin given by Eq. (2.4). The decoherence we
consider is the dephasing map given by
D(ρ) =(1−p)ρ+
X
i
p
ˆ
P
i
ρ
ˆ
P
i
, (3.25)
where
ˆ
P
i
is a projector onto a basis element of the form |vi⊗|ci, v corresponds to a
vertex and c to a direction in the coin space. The initial condition chosen is |00...0i⊗
38
(1/
√
n)
P
i
|ii, ie., the
~
0 vertex and the equal superposition of all directions. This initial
condition lies in the subspace that gives fast hitting times for the non-decohering walk.
We analyze it in the presence of decoherence. Fig. (3.4) plots the hitting time against the
decoherence parameter p for dimensions n = 3 to n = 7. Observe that the hitting time
has an initial jump from p = 0 to p = 0.1 after which it increases smoothly. This jump
increases from n = 3 to n = 7 suggesting that for higher dimensions it gets worse. Fig.
(3.5) plots the hitting time for p ranging from 0.01 to 0.1 and we can see that even for
a little decoherence i.e. p = 0.01, the hitting time jumps and then increases uniformly.
Fig. (3.6) plots the hitting time against the decoherence parameter for the initial state
given by|00...0i⊗|1i. We shall show later that this initial state has an infinite hitting
time for the quantum walk with the Grover coin. As can be seen it becomes finite in
the presence of decoherence. This shows how this kind of decoherence can turn purely
quantum behavior into something classical.
3.4.3 Dephasing in the coin space
Now we consider a decoherence process that acts only in the coin space. In particular,
we consider a dephasing map given by
D(ρ)=(1−p)ρ+
X
i
p(
ˆ
I⊗
ˆ
P
i
)ρ(
ˆ
I⊗
ˆ
P
i
), (3.26)
where
ˆ
P
i
are projectors only onto the coin basis states. Fig. (3.7) plots the hitting time
against the parameter p. The behavior for this type of decoherence is qualitatively the
same as before. The hitting time is a little less than in the previous case except atp=0
39
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
10
0
10
1
10
2
10
3
p
!
h
Figure 3.4: Hitting time (log scale) vsp for dimensionsn = 3 ton= 7 with dephasing in
the position and coin space.
and p = 1. At these two points the hitting time is equal to the one in the previous kind
of dephasing precess.
3.4.4 Dephasing in the position space
Finally, we consider a decoherence process that acts only in the position space. The
dephasing map in this case is given by
D(ρ)=(1−p)ρ+
X
i
p(
ˆ
P
i
⊗
ˆ
I)ρ(
ˆ
P
i
⊗
ˆ
I), (3.27)
where
ˆ
P
i
are projectors only onto the position basis states. Fig. (3.8) shows the variation
of hitting time with the decoherence parameter. Note that when there is no dephasing
40
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
3.5
4
4.5
5
5.5
6
p
!
h
Figure 3.5: Hitting time (log scale) vs p for dimensions n = 3 for small values of deco-
herence.
and pure dephasing, i.e. p = 0 and p = 1, then the hitting time is the same whether the
dephasing is in the position space, coin space or both spaces.
3.4.5 Analysis of the slope of hitting time
Here we analyze the slope of the expression for hitting time with respect to the decoher-
ence parameter p in order to determine the behavior of the jump for small values of p.
The hitting time from Eq. (3.15) is
τ
h
=I
v
·
Y(I−N)
−2
ρ
v
0
, (3.28)
41
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
10
0
10
1
10
2
10
3
p
!
h
Figure 3.6: Hitting time (log scale) vs p for dimensions n = 3 for small values of deco-
herence for an initial state which has an infinite hitting time.
where the matrices Y and N are defined as
Y = (
ˆ
P
f
⊗
ˆ
P
∗
f
)((1−p)
ˆ
I⊗
ˆ
I +p
X
i
ˆ
P
i
⊗
ˆ
P
∗
i
)(
ˆ
U⊗
ˆ
U
∗
)
N = (
ˆ
Q
f
⊗
ˆ
Q
∗
f
)((1−p)
ˆ
I⊗
ˆ
I +p
X
i
ˆ
P
i
⊗
ˆ
P
∗
i
)(
ˆ
U⊗
ˆ
U
∗
).
Here we consider complete dephasing i.e.,
ˆ
P
i
is the projector from dephasing in the
position and coin spaces. Differentiating this expression w.r.t p, we get
dτ
h
dp
=I
v
·
dY
dp
(I−N)
−2
ρ
v
0
+2I
v
·
Y(I−N)
−3
dN
dp
ρ
v
0
.
42
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
10
0
10
1
10
2
10
3
p
!
h
Figure 3.7: Hitting time (log scale) vsp for dimensionsn = 3 ton= 7 with dephasing in
the coin space.
The derivative of Y is given by
dY
dp
= (
ˆ
P
f
⊗
ˆ
P
∗
f
)(
dD
dp
)(
ˆ
U⊗
ˆ
U
∗
), (3.29)
where
dD
dp
=−
ˆ
I⊗
ˆ
I +
X
i
ˆ
P
i
⊗
ˆ
P
∗
i
. (3.30)
We can define a similar expression for the derivative ofN. In Fig. (3.9) we plot this slope
for various values of the parameter p. Note that for p = 0 this expression has the same
problem i.e., of non-invertible matrices as the expression for hitting time.
43
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
10
0
10
1
10
2
10
3
p
!
h
Figure 3.8: Hitting time (log scale) vsp for dimensionsn = 3 ton= 7 with dephasing in
the position space.
3.5 Prior work
Kendon [44] has written a good review of the work done so far in analyzing decoherence
in quantum walks on many graphs such as the line, cycle and the hypercube. Here we
review some of the prior work on the hypercube. Decoherence in discrete-time quantum
walks on the hypercube has been studied by Kendon and Tregenna [41], where they
determine its effect on one-shot and concurrent hitting times defined in [38]. It has been
proved that the concurrent hitting time isτ
c
(p
0
) =
π
2
forp
0
= Ω(
1
nlog
2
n
) for a hypercube
of dimension n (for the symmetric initial condition used in this paper). Kendon and
Tregenna observe that for the range 0 ≤ p ≤ 1/n, where n is the dimension of the
hypercube, the quantum speed-up is preserved. However, we notice in our simulations
that for the definition of hitting time considered here (defined in [47]), the situation is
44
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
100
200
300
400
500
600
700
800
900
p
d!
h
/dp
Figure 3.9: Slope of the hitting time vs p for the dimension n= 3 with dephasing in the
position and coin space.
different. The quantum speed-up is not preserved even for small values ofp, which seems
to show that this definition of the hitting time is more sensitive to decoherence as it
captures the time-averaged dynamics.
Decoherence in continuous-time quantum walks on the hypercube has been analyzed
in [3]. The kind of decoherence considered is dephasing in the position basis (recall that
there is no coin space in this kind of walk). The evolution of this walk can be described
by the Hamiltonian
ˆ
H =
ˆ
X⊗
ˆ
I⊗···⊗
ˆ
I +
ˆ
I⊗
ˆ
X⊗···⊗
ˆ
I⊗+...+
ˆ
I⊗
ˆ
I⊗···⊗
ˆ
X, (3.31)
where
ˆ
X is the Pauli σ
x
operator. The hopping rate is chosen as γ = k/n, where n
is the dimension of the hypercube and k is the total energy of the system. In order
45
to incorporate decoherence, a superoperator picture was used. In this picture, the non-
decohering walk has the superoperator
ˆ
U
t
⊗
ˆ
U
†
t
, where
ˆ
U
t
= exp(−it
ˆ
H). In the presence
of decoherence this superoperator becomes
S
t
=
n
Y
j=1
[1 1⊗1 1]⊗···⊗e
ˆ
A
⊗···⊗[1 1⊗1 1], (3.32)
where
ˆ
A=γt[(1 1⊗i
ˆ
X)−(i
ˆ
X⊗1 1)−p(1 1⊗1 1)+p(Π
1
⊗Π
1
)+p(Π
0
⊗Π
0
)], (3.33)
Π
0
= |0ih0| and Π
1
= |1ih1|. It has been observed in [3] that for p < 4k the walk has
quantum behavior and for p>4k it has classical behavior.
3.6 Results for the DFT coin
The hitting time for a discrete-time quantum walk on the hypercube using the DFT
coin, by contrast to the Grover coin case considered above, can actually be infinite. For
n=4 we will demonstrate that for the same initial condition as with the Grover coin, the
hitting time for a quantum walk using the DFT coin is infinity. This is because there
exist eigenvalues of the evolution operator
ˆ
U whose eigenvectors have an overlap with the
initial state, but have no overlap with the final vertex for any state of the coin. In this
subsection we specialize to the DFT coin, but a more detailed and general explanation of
infinite hitting times based on the symmetry of the walk is presented in the next chapter.
46
Supposetherearestates|φiwhichhavenooverlapwiththefinalvertex,hφ|
ˆ
P
f
|φi=0,
and which are eigenstates of the evolution operator
ˆ
U:
ˆ
U|φi = exp(iθ)|φi. If the system
is in the state |φi, clearly there is no probability to ever detect the particle in the final
vertex. Let
ˆ
P be a projector onto all such states |φi. Then
ˆ
P
ˆ
P
f
=
ˆ
P
f
ˆ
P = 0, and
[
ˆ
P,
ˆ
U] = 0, where the commutator is defined as [A,B] = AB−BA. One can write the
initial state as a superposition of vectors in this subspace and its orthogonal complement:
|Ψi=
ˆ
P|Ψi+(
ˆ
I−
ˆ
P)|Ψi. (3.34)
Any state that begins in the subspace selected by
ˆ
P will remain there for all time, and
any state in the orthogonal complement will stay there; this follows from the fact that
the projectors commute with both the unitary transformation
ˆ
U and the measurement
operator
ˆ
P
f
. As one starts the walk, the probability that the particle never reaches the
final state ishΨ|
ˆ
U
†
t
ˆ
P
ˆ
U
t
|Ψi, which ishΨ|
ˆ
P|Ψi.
In order for this probability to be nonzero, there must be eigenstates of the unitary
evolution operator
ˆ
U which have no amplitude for the final vertex. We can readily
demonstrate this for the hypercube with the DFT coin. Consider the 4-dimensional
hypercube. Numericallydiagonalizingtheevolutionoperator
ˆ
U,wefindithasi,−i,1and
−1 among its eigenvalues, each with a degeneracy of 8. Since the subspace corresponding
to the final vertex is 4-dimensional, it is clearly possible to construct a superposition of
eigenvectors of any of these eigenvalues so that it has no overlap with the final vertex
47
in any coin state. For each of the four degenerate eigenvalues we can construct a 4-
dimensional subspace of eigenvectors with no overlap with the final vertex, giving a 16-
dimensional space for all such eigenvectors. By numerically constructing an orthonormal
basis for this space, we can find an expression for the projector
ˆ
P and measure its overlap
with the initial state.
We considered in particular the initial state where the particle was located at the
|00...0i vertex and the coin is in an equal superposition of basis states |0i,...,|n−1i.
For the hypercube with n = 4 and the given initial state, the probability hΨ|
ˆ
P|Ψi is
0.4286, which exactly matches the total probability to never hit the final node after a
large number of iterations in our numerical simulations. Thus, the probability is close to
half that the particle never reaches the final state and the hitting time becomes infinity.
This demonstrates a property of quantum walks not seen in their classical counter-
parts: for certain initial conditions, there is a nonzero probability that the particle never
reaches the finalstate, eventhough the initial and final states of the graph are connected.
For a quantum walk with substantial degeneracy, this phenomenon is likely to be generic.
It might be possible to make the hitting time finite by choosing an appropriate initial
condition—clearly this happens for the Grover coin—but for some coins this may require
aninitialconditionwhichisnotlocalizedononevertex. Fromoursimulations, ititseems
that for higher dimensions the DFT coin behaves similarly to n = 4. For example, for
n =5, oursimulationsshowthattheprobabilitytohitthefinalnodeincreasesslowlybut
does not reach 1 even after many time steps. This could be due to the fact that the final
vertex has no overlap with some eigenvectors of the evolution operator (as for n = 4),
48
and additionally that it overlaps very little for some other eigenvector. This would make
the probability increase slowly but never reach 1.
3.7 Results for a quantum walk on a distorted hypercube
If, as seems likely, the dramatic speed-ups (and slow downs) of quantum walks over their
classical counterparts depend on the symmetry of the graph, it should be instructive to
see the effect of deviations from that symmetry. In this section, we look at results for the
measured walk using the Grover coin on a distorted hypercube. The distorted hypercube
isdefinedbyconstructingtheusualhypercube,andthenswitchingtwooftheconnections.
Pick 4 vertices which form a face–for example, (0...00),(0...01),(0...10),(0...11).
Calling these vertices A,B,C,D for short, we distort the hypercube by connecting A to
D andB toC, and removing the edges betweenA andB and betweenC andD. This is
still a regular graph, and the same quantum walk can be used without having to redefine
the evolution operator. Unlike the usual hypercube, it is no longer a bipartite graph, and
the walk can no longer be reduced to a walk in Hamming weight.
Figure3.10plotsthehittingtimeofaquantumwalkonadistortedhypercubetogether
with that of a classical walk and a quantum walk on regular hypercubes for comparison.
The hitting time for a quantum walk on a distorted hypercube is more than that of a
quantum walk on a regular hypercube, but still is much smaller than the hitting time of
a classical walk. In fact, as the dimension increases one can see that the hitting times of
the quantum walk on the distorted and regular hypercubes converge towards each other.
49
2 4 6 8 10 12 14
10
0
10
1
10
2
10
3
10
4
10
5
Dimension of the hypercube
Hitting time (log scale)
Classical hitting time
Quantum hitting time
(distorted hypercube)
Quantum hitting time
(regular hypercube)
Figure 3.10: Comparison of hitting times on the regular and distorted hypercubes
This presumably reflects the fact that for higher dimensions the symmetry is mostly
unchanged.
3.8 Infinite hitting times
We will now show that it is possible for hitting times to be infinite, and derive a sufficient
condition for the unitary evolution operator to allow infinite hitting times. As before,
we begin by forming the projector
ˆ
P onto the subspace spanned by all eigenstates of
ˆ
U
which have no overlap with the final vertex. This projector is orthogonal to the projector
onto the final vertex,
ˆ
P
ˆ
P
f
=
ˆ
P
f
ˆ
P = 0, and commutes with
ˆ
U, [
ˆ
U,
ˆ
P] = 0. We assume
(for the moment) that this projector is nonzero; later, we will find a sufficient condition
for this to be true, and exhibit quantum walks which satisfy this condition. We can write
50
any initial state as a superposition of a state in the subspace projected onto by
ˆ
P and a
state orthogonal to it, giving the decomposition
|Ψi=
ˆ
P|Ψi+(
ˆ
I−
ˆ
P)|Ψi. (3.35)
Itiseasytoseethatif|Ψiliesentirelyinside
ˆ
P,i.e.,
ˆ
P|Ψi=|Ψi,thenundertheevolution
thesubsequentstateswillneverhaveanycomponentinthefinalstateandtheprobability
defined in eq. (3.6) will be zero. Indeed, since [
ˆ
P,
ˆ
U] =0 and [
ˆ
P,
ˆ
Q
f
] =0,
p(t) = Tr{
ˆ
P
f
ˆ
U[
ˆ
Q
f
ˆ
U]
t−1
ρ
0
[
ˆ
U
†ˆ
Q
f
]
t−1 ˆ
U
†ˆ
P
f
} (3.36)
= Tr{
ˆ
P
f
ˆ
U[
ˆ
Q
f
ˆ
U]
t−1
ˆ
Pρ
0
ˆ
P[
ˆ
U
†ˆ
Q
f
]
t−1 ˆ
U
†ˆ
P
f
} (3.37)
= Tr{
ˆ
P
f
ˆ
P
ˆ
U[
ˆ
Q
f
ˆ
U]
t−1
ρ
0
[
ˆ
U
†ˆ
Q
f
]
t−1 ˆ
U
†ˆ
P
†
ˆ
P
f
} (3.38)
= 0, (3.39)
where ρ
0
= |ΨihΨ|. Therefore, the hitting time for this initial state is infinite. More
generally, if|Ψi has nonzero overlap with
ˆ
P,
ˆ
P|Ψi6= 0, then that component of|Ψi can
never reach the final vertex. The probability of ever hitting the final vertex if one starts
with this initial state is
p=|hΨ|(
ˆ
I−
ˆ
P)|Ψi|
2
<1, (3.40)
and the hitting time is again infinite.
To construct this projector, we look at the spectral decomposition of
ˆ
U. If
ˆ
U has
at least one sufficiently degenerate eigenspace, then we can construct a subspace of this
eigenspace which has a zero overlap with the final state. For instance, consider one such
51
degenerate eigenspace which has a degeneracy of k. Since the vector space at the final
vertex is d dimensional (i.e., it has d coin degrees of freedom), we would be solving the
following d×k system of homogeneous equations:
a
1
v
1
N−d+1
.
.
.
v
1
N
+a
2
v
2
N−d+1
.
.
.
v
2
N
+···+a
k
v
k
N−d+1
.
.
.
v
k
N
= 0. (3.41)
Here,weusealabelingwherethefinalvertexinsomecoinstateoccupiesthelastdentries
of the eigenvectors. The subscript refers to the component of the eigenvector, and the
superscript distinguishes the eigenvectors in the degenerate eigenspace. This system is
under-determined if k >d, and it will always have a nontrivial solution—in fact, it will
have a space of solutions of dimension k−d. Therefore, it is sufficient that there exist
at least one eigenspace of
ˆ
U with dimension greater than the dimension of the coin, in
order to have a nonzero projector
ˆ
P. If there is more than one degenerate eigenvalue
with multiplicity greater than d, the subspace projected onto by
ˆ
P will include all the
eigenvectors of
ˆ
U which have no overlap with the final vertex.
The condition derived above is closely related to the question of invertibility ofI−N
in Eq. (3.15) of the previous section. Here we show that I−N is not invertible if and
only if the projector
ˆ
P is nonzero. Furthermore, in the case whenI−N is not invertible,
the hitting time for a state or density matrix whose support has no overlap with
ˆ
P is
calculated by replacing the inverse of I−N with its pseudo-inverse in Eq. (3.15).
52
Assume that the projector
ˆ
P is nonzero. Then there is at least one eigenvector|vi of
ˆ
U such that
ˆ
P
f
|vi=0. Therefore,
(
ˆ
Q
f
ˆ
U⊗
ˆ
Q
∗
f
ˆ
U
∗
)(|vi⊗|vi
∗
) =(
ˆ
Q
f
⊗
ˆ
Q
∗
f
)(|vi⊗|vi
∗
) =|vi⊗|vi
∗
, (3.42)
since
ˆ
U|vi = exp(iθ)|vi and
ˆ
U
∗
|vi
∗
= exp(−iθ)|vi
∗
, and
ˆ
P
f
|vi = 0 implies that
ˆ
Q
f
|vi =
|vi. Therefore, (I−N)|vi⊗|vi
∗
= 0, I−N has a nonzero nullspace, and hence is not
invertible. This proves the “if” direction.
To prove the “only if” direction, assume that I−N is not invertible. This implies
that there exists a normalized vector|ui∈H⊗H
∗
such that
(
ˆ
I−
ˆ
Q
f
ˆ
U⊗
ˆ
Q
∗
f
ˆ
U
∗
)|ui= 0=⇒
ˆ
Q
f
ˆ
U⊗
ˆ
Q
∗
f
ˆ
U
∗
|ui=|ui. (3.43)
The vector |ui is an eigenvector of
ˆ
Q
f
ˆ
U ⊗
ˆ
Q
∗
f
ˆ
U
∗
with eigenvalue 1. Since
ˆ
Q
f
⊗
ˆ
Q
∗
f
is
a projector, the vector |ui must therefore lie in the eigenspace of eigenvalue 1 of both
ˆ
Q
f
⊗
ˆ
Q
∗
f
and
ˆ
U⊗
ˆ
U
∗
. This can only be true if|ui is of the form:
|ui=
X
i,j
a
ij
|v
i
i⊗|v
j
i
∗
, (3.44)
where the {|v
i
i} are eigenvectors of
ˆ
U, and a
ij
is only nonzero if |v
i
i and |v
j
i lie in
the same eigenspace of
ˆ
U and
ˆ
P
f
|v
i
i =
ˆ
P
f
|v
j
i = 0. Note that the vector |ui need not
correspond to a physical state in the Hilbert space of the walk. But the existence of such
a|ui means that the projector
ˆ
P is nonzero, since there must exist at least one|v
i
i which
has a zero overlap with
ˆ
P
f
. This proves the “only if” direction.
53
We have shown that if I−N is not invertible then the projector
ˆ
P is nonzero and
vice-versa. Now we will see that if a density operatorρ is orthogonal to
ˆ
P,ρ
ˆ
P =
ˆ
Pρ=0,
its corresponding vector ρ
v
lies outside the null space of I−N. If ρ is orthogonal to
ˆ
P, then when written in the eigenbasis {|v
i
i} of
ˆ
U, the diagonal components hv
i
|ρ|v
k
i
are nonzero if and only if
ˆ
P
f
|v
i
i6= 0. This implies that for the corresponding vectorized
quantity ρ
v
, we have
I
v
.(
ˆ
P
f
⊗
ˆ
P
∗
f
)ρ
v
6= 0. (3.45)
For any ρ
v
,
I
v
.(
ˆ
P
f
⊗
ˆ
P
∗
f
)(
ˆ
Q
f
ˆ
U⊗
ˆ
Q
∗
f
ˆ
U
∗
)ρ
v
= 0. (3.46)
Therefore, Nρ
v
6= ρ
v
i.e., ρ
v
does not lie in the null space of I−N. Moreover, if ρ is
orthogonal to
ˆ
P, then so areNρ andYN
t
ρ. This is easy to see, since from Eq. (3.7) we
get
ˆ
PNρ=
ˆ
P(
ˆ
Q
f
ˆ
Uρ
ˆ
U
†
ˆ
Q
f
) =
ˆ
Q
f
ˆ
U
ˆ
Pρ
ˆ
U
†
ˆ
Q
f
= 0, (3.47)
since
ˆ
P
ˆ
Q
f
=
ˆ
P and [
ˆ
U,
ˆ
P]=0. Similarly, we obtain
ˆ
PYρ= 0, since
ˆ
PYρ=
ˆ
P(
ˆ
P
f
ˆ
Uρ
ˆ
U
†
ˆ
P
f
) =0. (3.48)
Therefore, if ρ is orthogonal to
ˆ
P, then all the terms of the type YN
t−1
ρ for all t and
hence all the terms inside the trace of Eq. (3.9) are orthogonal to
ˆ
P, which means that
the vectorized versions of all terms inside the trace in Eq. (3.9) lie outside the null space
of I−N. Thus, for states that do not overlap with
ˆ
P, the hitting time is finite as would
54
be expected, and is given by the same formula Eq. (3.15) with the inverse replaced by a
pseudo-inverse.
We focus on the discrete time quantum walk for the remainder of this section, consid-
ering the walk on the hypercube in particular. It was observed in numerical simulations
mentioned earlier that for a walk on the hypercube with the DFT coin, an initial state
given by
|Ψi=|00···0i⊗
1
√
d
X
i
|ii (3.49)
has an infinite hitting time. The phenomenon of infinite hitting times is not restricted
to the walk with the DFT coin, however. Numerical simulations, followed by analytical
calculations, have shown that it also occurs with the Grover coin, but for different initial
states. In fact, it turns out that for the Grover coin, the symmetric initial state (3.49) is
the only initial state localized at the vertex |00···0i that has a finite hitting time. Any
other superposition of the coin states for that vertex will give an infinite hitting time,
because all such states have a nonzero overlap with
ˆ
P.
Given any vertexv on a graph, it is natural to ask if there exists any superposition of
its coin states which overlaps with
ˆ
P, and for which coin state the overlap is maximum
and for which it is minimum (or zero). We can write the projector
ˆ
P in the form
ˆ
P =
X
i,j,k,l
A
ijkl
|x
i
ihx
j
|⊗|kihl|, (3.50)
55
where{|x
i
i} are the vertices and{|ki} are the directions. Suppose the initial state is
|Ψi=|vi⊗
X
i
α
i
|ii≡|vi⊗|αi. (3.51)
Its overlap with the projector
ˆ
P is given by
hΨ|
ˆ
P|Ψi=
X
k,l
A
vvkl
α
∗
k
α
l
. (3.52)
To find the superposition of coin states such that the overall initial state has the least (or
greatest) overlap with
ˆ
P, define the matrix
ˆ
C
v
=Tr
vertices
n
ˆ
P(|vihv|⊗
ˆ
I
coin
)
o
, (
ˆ
C
v
)
kl
=A
vvkl
. (3.53)
The overlap of the initial state with
ˆ
P can be written in terms of this matrix as
hΨ|
ˆ
P|Ψi=hα|
ˆ
C
v
|αi. (3.54)
The matrix
ˆ
C
v
is Hermitian and positive, and hence has a spectral decomposition into a
complete orthonormal basis of eigenstates with non-negative eigenvalues. Assuming that
{λ
i
,|e
i
i} is the spectral decomposition of
ˆ
C
v
, we can rewrite the overlap as
hΨ|
ˆ
P|Ψi=
X
i
λ
i
|hα|e
i
i|
2
. (3.55)
From the above expression, we see that the overlap is maximum (or minimum) if|αi is in
the direction of the eigenvector with the largest (or smallest) eigenvalue, and zero if|αi is
56
aneigenvectorwithzeroeigenvalue. Therefore, if
ˆ
C
v
doesnothaveazeroeigenvalue(i.e.,
is positive definite), then for that vertex every superposition of coin states will overlap
with
ˆ
P. In other words, the hitting time will be infinity if one starts at that vertex no
matter what coin state one chooses. Numerical calculations for the Grover and the DFT
coins on the hypercube show that for the vertex|00···0i, an equal superposition of coin
states is the only superposition that has a zero overlap with
ˆ
P for the Grover coin, and
no superposition of coin states has a zero overlap for the DFT coin. Moreover, Tr{
ˆ
P} for
the Grover coin on the hypercube forn= 4 is 32, which is fully half the dimension of the
total space (dim=2
4
·4=64). These examples suggest that infinite hitting times may be
a generic phenomenon on graphs with symmetry.
3.9 Discussion
In this chapter, we have examined the definition of hitting time for measured quantum
walksandanalyzedsomeofitsproperties. Weusedthisdefinitiontoobtainanexpression
for hitting time which is valid on any general graph as long as the unitary evolution
operator
ˆ
U ofthewalkisdefined. Wesimulatedthishittingtimeforameasuredquantum
walkusingtheGrovercoinandcomparedittotheclassicalhittingtimeandtothebounds
obtainedonit;thequantumhittingtimeisexponentiallysmallerthantheclassicalhitting
time. We also showed that the bounds on the hitting time obtained in [38] become less
tight as the dimension increases. We then investigated the effect of decoherence on the
hittingtimeofadiscretequantumwalk. Wehaveobservedthatitaffectsthehittingtime
more dramatically than it does for the definitions proposed in earlier work. Classical
57
behavior occurs for small values of the decoherence parameter and the hitting time is
worse than it is for a corresponding classical walk for higher values of this parameter. We
observedthatinthecasewhenthequantumwalkhasinfinitehittingtimes(aphenomenon
whichisconsequenceofthequantumbehaviorofthewalk),decoherencemakesthehitting
time finite and thus make the walk classical. We analyzed the effect of different kinds
of decoherence and noticed that for dephasing (whether in the position or coin space or
both), the hitting time deteriorates in more or less the same way.
These results on fast and infinite hitting times show that simply making a walk quan-
tum does not guarantee a speed-up over the classical case. We demonstrated that the
hitting time for quantum walks can depend sensitively on the initial condition, unlike
classical walks. For certain initial states, the DFT walk can have infinite hitting time,
a phenomenon not possible in classical random walks. This dependence on the initial
state varies with the coin used, since for the same initial state the Grover walk has a
polynomial hitting time. This infinite hitting time is directly related to the degeneracy of
the eigenvalues of the evolution operator. If the evolution operator is highly degenerate,
then it is very likely that there exist initial states which give infinite hitting times.
Thecauseofthespeed-upinquantumhittingtimemaynotbecompletelyclearatthe
moment, but it will be shown in the next two chapters that the symmetry of the graph
plays a major role in both the speed-up and slow down of the quantum walk. For now,
we can see this as an interference effect. In the faster quantum walk, the different paths
leading to the final vertex interfere constructively, enhancing the probability of arrival;
paths which lead to “wrong” vertices interfere destructively, reducing the probability of
meandering around in the graph for long times. Unlike a classical random walk, the
58
quantum walk is sensitive to the presence of a global symmetry which is not apparent
at a purely local level. This phenomenon leads to the speed-up of the continuous-time
quantum walk on the glued-trees graph as well [19].
However, this same reason is undoubtedly the culprit in the slow-down observed for
the DFT walk. The existence of states which never arrive at the final vertex is made
possible by the degeneracy of the evolution operator
ˆ
U—a degeneracy which arises due
to the symmetry of the graph. The existence of states which never arrive at the final
vertex can also be seen as an interference effect, only in this case the interference of paths
which lead to the final vertex is destructive: all amplitude to make a transition to the
final vertex cancels out.
The connection to symmetry is supported by the quantum walk on the distorted
hypercube. We observe that the hitting time is worse than that of the usual hypercube,
but still much smaller than that of a classical walk. The curve of the hitting time on the
distorted hypercube seems to converge slowly to that of a quantum walk on the regular
hypercube. This is probably because the distortion used in our simulations is very mild.
As the dimension grows, and with it the number of edges and vertices, this distortion has
less effect on the overall symmetry.
We hasten to add that symmetry of the graph is not the sole reason for speed-ups
in quantum walks. A polynomial speed-up has been demonstrated in the quantum walk
versions of the search of an unstructured database [72] and the element distinctness
problem [4]. However, the dramatic exponential speed-ups have all been demonstrated in
highly symmetric graphs. In the next two chapters, we explore this idea in more detail
and show how symmetry can have a strong influence on the quantum walk.
59
Chapter 4
Symmetry in quantum walks
Wesawinthepreviouschapterthatdegeneracyoftheevolutionoperatorleadstoinfinite
hitting times. One of the main sources of degeneracy in quantum mechanics is symmetry.
Sincethe
ˆ
S matrixwhichmakesuppartoftheevolutionoperatorencodestheconnections
ofthegraph, itisnaturaltoexpectthatsymmetriesofthegraphwillproducesymmetries
of the evolution operator. Here we analyze this idea.
Is it sufficient to consider only the symmetries of the graph? Apart from the symme-
tries induced from the graph, the evolution operator may have additional symmetries of
its own which lead to additional degeneracy. But such symmetries are difficult to analyze
in generality, and depend on the details of how one defines the quantum walk. First,
there is a choice between the discrete and the continuous walk. Second, for the discrete
walk, if the structure
ˆ
U =
ˆ
S(
ˆ
I⊗
ˆ
C) is used, there is still the freedom to use any unitary
matrix as the coin. In order to generalize our discussion of symmetries to any walk, we
restrict attention to the symmetries induced by the graph alone. This naturally leads us
to the question: “Are there graphs with sufficient symmetry such that, for any walk that
is defined on the graph, the resultant evolution operator will have enough degeneracy to
60
give rise to infinite hitting times?” It turns out that such graphs do exist, and we give
an example of such a class of graphs. We will comment briefly on the effect of additional
symmetries of the coin for the hypercube later in this chapter. We now look at the au-
tomorphism groups of Cayley graphs which provide the examples of graphs with infinite
hitting times.
4.1 Automorphism groups of Cayley graphs
An automorphism of a graph is a permutation of its vertices such that it leaves the graph
unchanged. The set of all such permutations is the automorphism group of the graph.
When the edge labels or colors in the graph are important, as in the case of a discrete
quantumwalk,werestrictourselvestothosepermutationswhichpreservetheedgelabels.
In other words, an edge connecting two vertices has the same label before and after the
permutation. Such automorphisms are called direction-preserving. In general, we could
consider automorphisms where we permute the direction labels along with the vertices
to obtain the same graph with the same coloring. This would form a larger group G of
which the direction-preserving automorphisms are a subgroup H.
Since the vertex Hilbert spaceH
v
has its basis elements in one-to-one correspondence
with the vertices of the graph, and the coin Hilbert space has a basis in correspondence
with the direction labels, the automorphisms (which are just permutations of vertices
and directions) are permutation matrices. In fact, these are all the permutation matrices
on H
v
⊗H
c
that leave
ˆ
S unchanged, i.e., {all
ˆ
P|
ˆ
P
ˆ
S
ˆ
P
†
=
ˆ
S, where
ˆ
P is a permutation
matrix}. Inthisrepresentation, anydirection-preservingautomorphismhasthestructure
61
ˆ
P
v
⊗
ˆ
I
c
, where
ˆ
P
v
actssolelyonH
v
and
ˆ
I
c
onH
c
. Suchautomorphismsbecomeimportant
if we wish to consider the symmetries of
ˆ
U ≡
ˆ
S(
ˆ
I⊗
ˆ
C). Clearly, any automorphism of
this type is a symmetry of
ˆ
U, since
(
ˆ
P
v
⊗
ˆ
I
c
)
h
ˆ
S(
ˆ
I⊗
ˆ
C)
i
(
ˆ
P
v
⊗
ˆ
I
c
)
†
=
h
(
ˆ
P
v
⊗
ˆ
I
c
)
ˆ
S(
ˆ
P
v
⊗
ˆ
I
c
)
†
i
(
ˆ
I⊗
ˆ
C) =
ˆ
S(
ˆ
I⊗
ˆ
C). (4.1)
Elements of G in general do not act trivially on the coin space. Because of this, they
need not be symmetries of
ˆ
U unless the coin flip operator
ˆ
C respects these symmetries.
To illustrate all this, consider the example of a hypercube in 2 dimensions (i.e., a
square). The vertex labels are {(00),(01),(10),(11)} (which also form a basis for H
v
);
the edges connecting (00) to (01) and (10) to (11) are both labeled 1, and the edges
connecting (00) to (10) and (01) to (11) are both labeled 2. Thus, the transformation
(00)↔(01)and(10)↔(11), orthetransformation(00)↔(10)and(01)↔(11), orboth
together, are automorphisms of this graph which need no permutation of the directions.
Together with the identity automorphism (which permutes nothing), these permutations
form the direction-preserving subgroup H. In a matrix representation on the Hilbert
spaceH
v
⊗H
c
, they are,
0 1 0 0
1 0 0 0
0 0 0 1
0 0 1 0
⊗
ˆ
I
c
,
0 0 1 0
0 0 0 1
1 0 0 0
0 1 0 0
⊗
ˆ
I
c
,
0 0 0 1
0 0 1 0
0 1 0 0
1 0 0 0
⊗
ˆ
I
c
,
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
⊗
ˆ
I
c
62
where
ˆ
I
c
is the 2×2 identity matrix acting on the coin space. These permutations can
be easily seen to be H ={
ˆ
I⊗
ˆ
X⊗
ˆ
I,
ˆ
X⊗
ˆ
I⊗
ˆ
I,
ˆ
X⊗
ˆ
X⊗
ˆ
I,
ˆ
I⊗
ˆ
I⊗
ˆ
I}. Just as in the
representation of
ˆ
S matrix in terms of the Pauli
ˆ
X operators given by Eq. (2.12), this
group denotes a bit flip in the first, second or both bits of each vertex, together with the
identity, which gives no flip. (See Fig. 4.1.)
00 01
10 11
2 2
1
1 01 00
11 10
2 2
1
1
10 11
00 01
2 2
1
1 11 10
01 00
2 2
1
1
(00) (01)
(10) (11)
(00) (10)
(01) (11)
(00) (01)
(10) (11)
(00) (10)
(01) (11)
00 01
10 11
2 2
1
1 01 00
11 10
2 2
1
1
10 11
00 01
2 2
1
1 11 10
01 00
2 2
1
1
(00) (01)
(10) (11)
(00) (10)
(01) (11)
(00) (01)
(10) (11)
(00) (10)
(01) (11)
Figure 4.1: The direction-preserving automorphism group of the n=2 hypercube.
The permutation (10) ↔ (01), reflecting along the diagonal while keeping (00) and
(11) fixed, will be an automorphism only if we interchange the directions 1 ↔ 2. Sim-
ilarly, the permutations (00) ↔ (11), (00) → (01) → (11) → (10) and (00) → (10) →
(11)→(01) are automorphisms when we interchange the two directions. If we view these
63
00 01
10 11
2 2
1
1 11 01
10 00
1 1
2
2 00 10
01 11
1 1
2
2
01 11
00 10
1 1
2
2
(01)
00 01
10 11
2 2
1
1
(01) (10) (00) (11) (00)
10 00
11 01
1 1
2
2
00 01
10 11
2 2
1
1 01
10
1 1
2
2 00 10
01 11
1 1
2
2
01 11
00 10
1 1
2
2
(01)
00 01
10 11
2 2
1
1
(01) (10) (00) (11) (00)
10 00
11 01
1 1
2
2
(00) (01) (11) (10) (00) (01) (11) (10) (00) (10) (11) (01) (00) (10) (11) (01)
Figure 4.2: Automorphisms which interchange directions for the n=2 hypercube.
permutations along with those obtained above, we obtain a new group G for which H is
a subgroup. In a matrix representation, the new automorphisms are,
0 1 0 0
0 0 0 1
1 0 0 0
0 0 1 0
⊗
ˆ
X
c
,
1 0 0 0
0 0 1 0
0 1 0 0
0 0 0 1
⊗
ˆ
X
c
,
0 0 1 0
1 0 0 0
0 0 0 1
0 1 0 0
⊗
ˆ
X
c
,
0 0 0 1
0 1 0 0
0 0 1 0
1 0 0 0
⊗
ˆ
X
c
,
where
ˆ
X
c
acts on the coin space and corresponds to an interchange of the two directions.
(See Fig. 4.2.) These four elements of G need not be symmetries of
ˆ
U, since the coin
need not be symmetric under conjugation with
ˆ
X
c
. However, for the hypercube, if we
use the Grover diffusion matrix as the coin, then the automorphism group G is indeed
its group of symmetries, since the Grover coin is symmetric under any permutation of
its basis elements. The symmetry group of the evolution operator would be H if the
DFT coin is used, since the DFT does not have permutation symmetry. It is important
64
to note that the symmetry group defined above is not the only thing that influences the
degeneracy of the evolution operator. Degeneracy of the coin flip operator
ˆ
C can also
induce degeneracy in the evolution operator, and the coin may be degenerate even if it
does not have permutation symmetry (like the DFT coin). We discuss this in more detail
later after describing the relationship between symmetry and degeneracy.
It can be shown that the direction-preserving automorphism groupH for any Cayley
graph is isomorphic to the group on which the graph is defined. This is because any
direction-preserving automorphism of a Cayley graph is a left translation by a group
element, andconverselyalllefttranslationsaredirection-preservingautomorphisms. The
first part of the statement is easy to see. Consider any left translation L
a
: G → G
which has the action L
a
(g) = ag, for all g ∈ G. Now, given vertices g and h in G,
they are connected by an edge from g to h if g
−1
h = s, where s ∈ S. Clearly, after
the transformation we still have (ag)
−1
(ah) = g
−1
h = s and hence this automorphism
preserves the direction labels. Since the group elements are basis states of the vertex
Hilbert space, a left translation by a group element corresponds to a permutation matrix
on this Hilbertspace. The fact that every direction-preserving automorphism of a Cayley
graph is a left translation by a group element becomes important in the discussion of
regular and irreducible representations later in this chapter.
Finally, let us explicitly construct the representation of the automorphism group for
the hypercube, which has H
∼
=Z
n
2
and G
∼
=H·S
n
. In terms of the Pauli operators the
representation ofH is{
ˆ
I
ˆ
I
ˆ
I···
ˆ
I⊗
ˆ
I
c
,
ˆ
X
ˆ
I
ˆ
I···
ˆ
I⊗
ˆ
I
c
,
ˆ
X
ˆ
X
ˆ
I···
ˆ
I⊗
ˆ
I
c
,...,
ˆ
X
ˆ
X
ˆ
X···
ˆ
X⊗
ˆ
I
c
},
where the tensor product symbol has been dropped in the vertex space, and
ˆ
I
c
is the
identity operator in the coin space. In fact, the representation of H for any Cayley
65
graph will be of the form
ˆ
P ⊗
ˆ
I, where
ˆ
P is a permutation matrix on the vertex space
and
ˆ
I is the identity on the coin space. The group G for the hypercube will become
H·S
n
={h·π|h∈H,π∈S
n
}, where S
n
is the permutation group on n elements which
is assumed to act on the direction labels.
We now briefly review linear representations of finite groups, and describe how the
symmetry group of
ˆ
U affects its degeneracy by determining the dimensions of the irre-
ducible representations of the group. (For further details, see for example [76] or [71].)
4.2 Representations of finite groups
A linear representation of a finite group G on a finite-dimensional vector space V is a
map σ : G → GL(V) , such that σ(gh) = σ(g)σ(h). GL(V) is the space of invertible
linear maps of V onto itself. If V is d-dimensional, and we choose a basis of d vectors
in V, then σ(g) for g ∈ G becomes a d×d invertible matrix. The trace of this matrix
is called the character of the representation. Therefore, characters are maps χ :G→C
with χ(g) = Tr(σ(g)). Two representations σ
1
and σ
2
of the group G on vector spaces
V
1
and V
2
, respectively, are considered equivalent if there exists an invertible linear map
τ : V
1
→ V
2
such that τ ◦σ
1
(g) = σ
2
(g)◦τ, for all g ∈ G. The characters of equivalent
representations are equal, a fact which follows from the cyclic property of the trace
operator.
Assume that the vector space V has an inner product defined on it. Since the group
G is assumed finite, it can be shown that any representation σ is equivalent to a unitary
representation—there exists a basis forV in whichσ(g) is a unitary matrix for allg∈G
66
(see [71]). A vector space W is said to be invariant or stable under the action of G if
x ∈ W ⇒ σ(g)x ∈ W for all g ∈ G. If the vector space V has a subspace W which is
invariant under the action of G, then it can be shown that its orthogonal complement
W
⊥
is also invariant under G, and so V can be decomposed
V =W ⊕W
⊥
. (4.2)
Thismeansthatthe representationσ as a matrix onV can be writtenina blockdiagonal
form consisting of two blocks as
σ =
σ|
W
0
0 σ|
W
⊥
, (4.3)
where σ|
W
and σ|
W
⊥ are the restrictions of σ to the subspaces W and W
⊥
.
The linear map σ :G→GL(V) is called an irreducible representation (irrep), if it is
a representation and no non-trivial subspace ofV is stable under the action ofG. Equiv-
alently, it is an irreducible representation if it is not a direct sum of two representations.
Any representation on V can, by an appropriate choice of basis, be written as a repre-
sentation in block diagonal form, where each block corresponds to an irrep. SoV can be
decomposed in the following way:
V =W
1
⊕W
2
⊕···⊕W
k
, (4.4)
67
where each of the W
i
is stable under the action of σ(G). This decomposition is unique,
up to reordering of the spaces and an overall equivalence transformation. By an abuse
of notation, we will use the same labels (V or W
i
) to refer both to the vector space and
to the group representation on that space. Each component W
i
of the decomposition in
(4.4) is isomorphic to an irrep, and the number of such W
i
isomorphic to a given irrep
does not depend on the details of the decomposition.
If we define an inner product for characters,
(φ|χ) =(1/|G|)
X
g∈G
φ(g)χ(g)
∗
, (4.5)
it can be shown that the number of times an irrep with character χ
i
occurs in a repre-
sentation with a character χ is given by (χ|χ
i
) [71].
We now define a regular representation of a group G. Suppose |G| = n, and let V
be an n-dimensional vector space. Let {e
t
} define a basis for V, which is labeled by
the group elements of t ∈ G. A regular representation of G is a map σ : G → V such
that σ(g)e
t
= e
gt
. It can easily be seen that the action of G on the basis vectors is a
left translation, and in matrix form the representations of the group elements will be
permutation matrices. An important property of the regular representation is that its
decomposition into irreps contains all the irreps of the group in it, and each irrep has a
multiplicity equal to its dimension [76]. Therefore, we can write
V =n
1
W
1
⊕n
2
W
2
⊕···⊕n
k
W
k
, (4.6)
68
where we use the notationnW to meann copies of the spaceW,nW =W⊕W⊕···W.
The{W
i
} are all the inequivalent irreps of the group, and n
i
is the dimension of W
i
.
It will be useful to note that all Abelian groups have one-dimensional irreps. The
converse of this statement is also true: all groups which have only one-dimensional irreps
are Abelian. Finally, supposeσ
1
:G→V
1
andσ
2
:G→V
2
are two linear representations
ofG onV
1
andV
2
, thenσ
1
⊗σ
2
:G→V
1
⊗V
2
is a representation ofG onV
1
⊗V
2
. (The
tensor product W
1
i
⊗W
2
j
of two irreps W
1
i
and W
2
j
of V
1
and V
2
, respectively, need not
be an irrep of V
1
⊗V
2
, however.)
Nowconsidertheunitaryoperator
ˆ
U onafinitedimensionalvectorspaceV whichhas
a group of symmetriesG. This means that the matricesσ(g) representating the elements
ofg∈G onV all commute with
ˆ
U: [σ(g),
ˆ
U] =0. Since
ˆ
U is unitary, we can decompose
V into a direct sum of eigenspaces of
ˆ
U:
V =U
1
⊕U
2
⊕···⊕U
m
. (4.7)
We can also decompose V into a direct sum of irreps of G:
V =W
1
⊕W
2
⊕···⊕W
k
. (4.8)
It can be shown using Schur’s lemma (see [76]) that since G is the group of symmetries
of
ˆ
U, each irrep of G must lie entirely inside some eigenspace of
ˆ
U. Therefore, for some
i and j, if W
i
⊂ U
j
then the degeneracy of U
j
is at least equal to the dimension of W
i
.
69
Using this fact, we show now that if a graph has sufficient symmetry (in a particular
sense), then it will lead to quantum walks with infinite hitting times.
4.3 Discrete-time walks on Cayley graphs
It was observed earlier that the direction-preserving automorphism group of a Cayley
graph is the group of left translations of the group elements. Since every group element
correspondstoavertexofaCayleygraphandeveryvertexcorrespondstoabasiselement
on the vertex Hilbert space (H
v
) of the walk, the automorphism group of a Cayley graph
is a regular representation of G on H
v
. Every direction-preserving automorphism of a
Cayley graph will induce a representation on this Hilbert space which looks likeσ(g)⊗
ˆ
I
c
,
where σ(g) is the regular representation of G on H
v
and
ˆ
I
c
is the identity on the coin
space. (Note that σ(g)⊗
ˆ
I
c
is not a regular representation of G.) In order to prove that
this walkhas an infinite hitting time for certain initial states, we need to show that
ˆ
U has
at least one degenerate eigenspace whose dimension is greater than the dimension of the
coin. Since every irrep lies completely inside an eigenspace, if one of the irreps occurring
in σ⊗
ˆ
I has a dimension greater than the dimension of the coin, then we can say that
the eigenspace containing that irrep has a degeneracy greater than the dimension of the
coin. We now show that every irrep of G occurs in σ⊗
ˆ
I. We have,
Tr(σ⊗
ˆ
I) =dχ, (4.9)
70
where χ = Tr(σ) and d is the dimension of the coin. If χ
i
is any irreducible character of
G, then
(Tr(σ⊗
ˆ
I)|χ
i
) =d(χ|χ
i
). (4.10)
Sinceχ is the character of the regular representation ofG, (χ|χ
i
)6= 0 for any irreducible
characterχ
i
. Therefore,allirrepsofGoccurinσ⊗
ˆ
I,andifanyirrepofGhasadimension
greater than the dimension of the coin then there is an eigenspace of
ˆ
U whose dimension
is greater than the dimension of the coin. So for Cayley graphs, a sufficient condition for
any discrete-time quantum walk to have infinite hitting times for some initial conditions
is that the group used to define the graph have an irrep with dimension greater than the
degree of the graph.
For a regular graph which is not a Cayley graph, this can be modified as follows. A
discrete time walk defined on a graph will have an infinite hitting time for certain initial
statesifatleastoneirrepoccurringintheinducedrepresentation(ontheHilbertspaceof
the walk) of the direction-preserving automorphism group of the graph has a dimension
greater than the degree of the final vertex. This is a somewhat more difficult to evaluate,
since unlike a Cayley graph, the induced representation of the symmetry group for a
general graph is not guaranteed to include every irrep. But in principle it is not difficult
to check.
For an example of a graph with infinite hitting times, consider the Cayley graph on
the symmetric group S
n
: Γ(S
n
,X), where X is a generating set for S
n
. In order to use
the form
ˆ
U =
ˆ
S(
ˆ
I⊗
ˆ
C), the graph needs to be|X|-colored and so we chose a generating
set whose elements x are such that x
2
=e, where e is the identity element. For S
n
, such
71
a generating set is any set ofn−1 transpositions, e.g.,{(1,2),(2,3),...,(n−1,n). These
formthebasisforthecoinspace, andsothedimensionofthecoinisthecardinalityofthe
generating set: |X| = n−1. Therefore, a symmetric group S
n
which has an irrep with
a dimension greater thann−1 will have infinite hitting times for some initial conditions
for any coin matrix. It turns out that for n≥ 5 any symmetric group S
n
possesses this
property [36], and so the corresponding Cayley graph Γ(S
n
,X) will have infinite hitting
times.
All this indicates that it is not so much the size of the symmetry group that matters
for infinite hitting times, but rather the kind of group, or more precisely the size and
number of irreps of the group occurring in the induced representation. Consider the
hypercube, which has the group H
∼
= Z
n
2
as its symmetry group (if the coin has no
permutation symmetry like the DFT coin). This group is Abelian and hence has only
one-dimensional irreps. So one would expect the unitary evolution operator having this
as its symmetry group to have very little or no degeneracy. But when the DFT coin is
used the evolution operator has sufficient degeneracy to have infinite hitting times. This
is because it is the degeneracy of the DFT coin, rather than the symmetry group, that
makes the evolution operator degenerate. This is supported by numerical evidence which
shows that when a randomly generated non-degenerate or slightly degenerate unitary
coin is used instead of the DFT, the unitary evolution operator has a very small or no
degeneracy. The degeneracy of the evolution operator with an Abelian symmetry group
seems to come only from the coin.
72
4.4 Continuous-time walks on Cayley graphs
In the prior discussion of infinite hitting times we have used the definition (3.5) for the
hitting time, which is only well-defined for discrete time quantum walks. We have not
described a suitable measurement process to define the hitting time for a continuous time
walk; nor is it obvious how to do so in the quantum case, where the presence or absence
of measurements has a profound effect on the dynamics. Any notion of hitting time for
the continuous case, however, must include a measurement performed on the final vertex
at some time which will verify if the particle has arrived there or not. This leaves an
ambiguity in the definition of hitting time for finite hitting times, but the notion of an
infinite hitting time still has an intuitive definition: a continuous time quantum walk has
infinite hitting time if, for any set of measurements on the final vertex at any sequence
of times, there is always a bounded, nonzero probability that the particle will never be
found at the final vertex.
Continuous time quantum walks do not have a coin matrix, and their evolution op-
erator for undirected graphs is
ˆ
U(t) = exp(i
ˆ
Ht) where
ˆ
H is the adjacency matrix of the
graph. Sincethereisnocoin,thedegreeoffreedomatanygivenvertexisonedimensional.
Any eigenspace of
ˆ
U with a degeneracy greater than one can therefore contribute to the
projector
ˆ
P having zero overlap with the final vertex and commuting with
ˆ
U. Whenever
ameasurementonthefinalvertexisperformed,themeasurementoperatorswillcommute
with
ˆ
P since
ˆ
P|x
f
ihx
f
|=0, where|x
f
i is the final vertex state. A nonzero
ˆ
P necessarily
means an infinite hitting time for initial states that overlap with it. Therefore, only if
ˆ
U
is completely non-degenerate, and none of its eigenvectors have a zero overlap with the
73
final state, will there be finite hitting times for all initial states. This is a lesser degree
of degeneracy than is needed in the discrete time case, and we therefore expect infinite
hittingtimestobeevenmorecommonincontinuoustimewalksthandiscretetimewalks.
Making the connection to symmetry again, for a continuous walk to have infinite
hitting times, a sufficient condition is that at least one irrep with dimension greater
than one occurs in the induced representation of the automorphism group of the graph.
Consider once more the example of Cayley graphs. As discussed in the case of discrete
time quantum walks above, the induced representation of the automorphism groupG on
theHilbertspaceH
v
istheregularrepresentation. Sincethereisnocoin,H
v
istheHilbert
space of the walk. All the irreps of the group appear in this representation because it is
regular. Therefore, if any of the irreps of G has a dimension greater than one, then the
walk will have infinite hitting times for certain starting states. Since only Abelian groups
have all their irreps of dimension one, any Cayley graph defined on a non-Abelian group
will have infinite hitting times for the continuous walk.
4.5 Discussion
In this chapter, we have examined the role of symmetry in shaping the properties of the
quantum walk. The role of symmetry is not restricted to producing infinite hitting times.
Symmetry can also be related to the exponentially fast hitting times observed in [38, 47].
Foroneofthecasesexaminedin[47]—thediscretewalkonthehypercubewiththeGrover
coin—weobserveanexponentiallysmallerhittingtimethantheclassicalwalkonthesame
graph. But this happens only for the symmetric initial state|Ψi=|00···0i⊗
1
√
d
P
i
|ii.
74
Other superpositions of coin states do not have this speed-up, but rather lead to infinite
hitting times, because their overlap with
ˆ
P is nonzero.
We noted that the group of symmetries of then-dimensional hypercube isG=H·S
n
when one takes into account the direction labels, where H is the normal subgroup of
direction-preserving automorphisms, and S
n
is the permutation group on n elements (in
this case, the different graph directions), and is also a subgroup of G. It was observed
in [72] that |Ψi is the simultaneous eigenstate of eigenvalue 1 of the subgroup S
n
(more
precisely, the simultaneous eigenstate of the representation operators of the subgroup).
Since every element of S
n
commutes with
ˆ
U for the walk with the Grover coin, a state
that begins in an eigenspace of the permutation group will remain in the same eigenspace
at all times. That is,
U
t
|Ψi=U
t
σ(g)|Ψi=σ(g)U
t
|Ψi, (4.11)
whereσ is the representation ofG on the Hilbert space of
ˆ
U andg∈S
n
. Thus,
ˆ
U
t
|Ψi is
an eigenstate of σ(g) with eigenvalue 1. This eigenspace has dimension 2n. It turns out
that the final vertex with an equal superposition of coin states|11···1i⊗
1
√
d
P
i
|ii also
lies in this eigenspace. Since the walk never escapes this subspace to explore other parts
of Hilbert space, it leads to an exponentially fast hitting time. For a measured walk, if
the symmetry subgroup commutes with the measurement operators (as is true in this
case), then the same argument holds. This shows that those symmetries of the graph
which are passed on to the evolution operator can create subspaces to which the walk
may be confined. If the final vertex has no overlap with such a subspace for any coin
state, then a walk starting in that subspace will have an infinite hitting time. Otherwise,
75
the hitting time will be finite, or even exponentially small depending on the dimension of
this subspace relative to the full Hilbert space.
Weshouldalsopointoutthattheconditionsderivedaboveforinfinitehittingtimesare
sufficient for a particular graph to have infinite hitting times; but they are not necessary
conditions. For example, the symmetry group of the hypercube is Abelian, and hence
does not imply that the evolution operator
ˆ
U must be degenerate. Nevertheless, infinite
hitting times are observed for quantum walks on the hypercube, due to the fact that the
choicesofcoinflipoperator(theGrovercoinortheDFT)bothhavetheirownsymmetries,
which increases the total degeneracy of the evolution operator. Infinite hitting times are
therefore likely to be even more common than the conditions derived here would suggest.
One can make very plausible intuitive arguments that both infinite hitting times
and exponentially fast hitting times are related to symmetry. This makes it seem likely
that the ideal problem to be solved by a quantum walk would be a problem with global
symmetry, but in which this symmetry is not apparent at the local scale. In the next
chapter a framework which explains fast hitting times observed in quantum walks, is
presented. Using symmetry of the graph again, we show that the quantum walk that
possesses these symmetries can be confined to a smaller graph called a quotient graph. If
this new graph is exponentially smaller than the original graph, then it can lead to fast
hitting times.
76
Chapter 5
Quotient graphs
Theideaofrestrictingasearchtoaninvariantsubspaceofthefullsearchspacehasproved
very fruitful in Grover’s search algorithm. In both the quantum walk-based algorithm on
the hypercube in [72] and the “glued-trees” graph in [19], the quantum algorithm works
very fast by searching a smaller space, where it is known that the solution lies in this
space. Inthischapter,weexplorethisconceptforquantumwalksonmoregeneralgraphs.
Using symmetry arguments, we show that it is possible to find invariant subspaces of the
total Hilbert space on which the walk is defined. The automorphism group of the graph
produces a group of symmetries of the evolution operator for the walk. This group of
symmetries in turn determines the invariant subspace of the walk. If the initial state is
in this subspace, the quantum walk effectively evolves on a different graph—a quotient
graph, which can in some cases be much smaller than the original graph. Here, we give
a general construction of quotient graphs, given the original graph and a subgroup of its
automorphism group. We determine the structure of the quantum walk on the quotient
graph. We then apply the analysis of hitting times developed in [47] and [48] to quotient
77
graphs, and investigate the possibility of both infinite hitting times and reduced hitting
times on quotient graphs.
5.1 Action of an automorphism group
Consider any undirected graph Γ and letV(Γ) andE(Γ) denote its vertex and edge sets.
Let the graph be colored, not necessarily consistently i.e., the edge between vertices v
i
and v
j
may be colored with a color c
k
in the direction from v
i
→v
j
and a color c
l
from
v
j
→ v
i
. This creates the Hilbert space of positions and colors (or directions) and the
total spaceH is spanned by basis vectors|v
1
,c
1
i,...|v
n
,c
m
i. Let this set of basis vectors
beX. The set of colors at each vertex is not the same for all the vertices since the graph
may be irregular. We assume that vertices having the same degree have the same set of
colors. Denote byC
v
the set of colors used to color edges going from the vertexv. Thus,
the shift matrix for this graph is,
S =
X
v
i
∈V(Γ),k∈Cv
i
|v
j
,c
l
ihv
i
,c
k
|. (5.1)
This matrix encodes the structure of the graph Γ which includes edge colors. An auto-
morphism of the graph Γ, as defined above is a permutation matrix which preserves S
under conjugation i.e., a matrixP such thatPSP
†
=S. The set of automorphisms form
a group which we denote by G.
Now consider a subgroup (not necessarily proper)H of this automorphism group. We
would like to know what kind of action this subgroup has on the graph and hence on the
Hilbert space. First, we define what is meant by the term action [67].
78
Definition 5.1. If X is a set and G is a group, then X is a G-set if there is a function
α :G×X →X (called a left action), denoted by α : (g,x)→gx, such that :
• 1x=x, for all x∈X; and
• g(hx)=(gh)x, for all g,h∈G and x∈X .
Definition 5.2. If X is a G-set and x∈X, then the G-orbit (or just orbit) of x is
O(x) ={gx:g∈G}⊂X. (5.2)
The set of orbits of a G-set X form a partition and the orbits correspond to the
equivalence classes under the equivalence relation x ≡ y defined by y = gx for some
g∈G. WecandefinetheactionofthesubgroupH ofthepermutationgrouponthesetof
basis elementsX of the Hilbert spaceH as the multiplication of its matrix representation
σ(H) (in the basis given by the vectors X) with a basis vector. This is a well-defined
action since σ(1)|xi =|xi and σ(g)(σ(h)|xi) = (σ(g)σ(h))|xi =σ(gh)|xi. Therefore, the
set X is partitioned into orbits under the action of H.
Since H is a subgroup of the automorphism group, these orbits can be related to the
graph Γ through the following results.
Theorem 5.3. If |v,c
i
i and |v,c
j
i are in different orbits, then the set of all the vertices
in the orbits of |v,c
i
i and |v,c
j
i are the same.
Proof. If the graph Γ is irregular (regular graphs are just a special case), then clearly
any automorphism takes a given vertex to another vertex of the same degree. Thus,
79
automorphisms permute vertices of a certain degree among themselves. Therefore, on
the Hilbert spaceH, the matrix representation of any automorphism can be written as
P =
M
d∈D
P
d
, (5.3)
where the set D contains all the different degrees in the graph. Consider the subspace
of vertices of a given degree d which can be written as H
V
d
⊗H
C
d
. Now, if any given
permutation takes|v
1
,c
i
i to|v
2
,c
j
i, then it takes all the basis vectors associated withv
1
to those of v
2
. Thus, the set of all vertices that lie in the orbit of |v
1
,c
i
i must be the
same as the set of vertices that lie in the orbit of |v
1
,c
j
i (if|v
1
,c
i
i and|v
1
,c
j
i lie in the
same orbit, then this is trivially true). Sincev
1
is arbitrary, the set of vertices in the two
orbits must be the same.
By an abuse of language, say that a vector |v
1
,c
1
i is “connected” to |v
2
,c
2
i if the
edge colored c
1
from vertex v
1
on the graph is connected to vertex v
2
along the color c
2
(i.e., the term|v
2
,c
2
ihv
1
,c
1
| occurs in S).
Theorem 5.4. If |v
1
,c
1
i and |v
2
,c
2
i are “connected,” |v
1
,c
1
i lies in the orbit O
1
and
|v
2
,c
2
i lies in orbit O
2
(not necessarily distinct from O
1
), then each of the remaining
vectors of O
1
are “connected” to some vector of O
2
.
Proof. If|v
1
,c
1
i and|v
2
,c
2
i are connected, then there is a term of the type|v
2
,c
2
ihv
1
,c
1
|
in S. When we conjugate by some automorphism h∈H i.e., perform σ(h)Sσ(h)
T
, then
this term transforms to σ(h)|v
2
,c
2
ihv
1
,c
1
|σ(h)
T
and this must be a term in S because
σ(h)Sσ(h)
T
=S. This means that the vector that|v
1
,c
1
i gets taken to, is connected to
80
the vector that |v
2
,c
2
i gets taken to by σ(h). Since this is true for all h ∈ H, all the
vectors in the orbit of|v
1
,c
1
i are “connected” to some term in the orbit of|v
2
,c
2
i.
The above result applies equally well to any vector in O
2
. Therefore, one can think
of the orbitsO
1
andO
2
as being “connected”.
5.2 Quotient graphs and quantum walks
Based on the action on a graph Γ of the subgroupH of its automorphism group, consider
the following construction of a graph—a quotient graph. The set of vertices occuring in
an orbitO is a single vertexv
O
on the new graph and the number of orbits that have the
same set of vertices is the degree of this new vertex. Thus, for a given vertexv
O
, the set
of directions are the various orbits which correspond to the same vertex set. If an orbit
O
1
is “connected” to O
2
, then the vertices v
O
1
and v
O
2
are connected in the quotient
graph. If O
1
and O
2
are identical, this corresponds to a self loop for v
O
1
. This means
that there can be self loops in the quotient graph even if there are none in the original
graph. We denote the quotient graph obtained by the action of the subgroup H on Γ as
Γ/H or Γ
H
.
Now consider a basis vector|xi≡|v,ci and itsH-orbitO
x
={σ(h)|xi:h∈H}. The
vector|˜ xi≡
1
√
|Ox|
P
h∈H
σ(h)|xiisaneigenvectorofeigenvalue1ofallthematricesσ(h)
for h∈H since
σ(h)|˜ xi =
1
p
|O
x
|
X
h
0
∈H
σ(h)σ(h
0
)|xi=
1
p
|O
x
|
X
h
0
∈H
σ(hh
0
)|xi
=
1
p
|O
x
|
X
h
00
∈H
σ(h
00
)|xi=|˜ xi. (5.4)
81
Similarly, the vector |˜ yi formed from a vector |yi of another orbit is also an eigenvector
of eigenvalue 1. Each of these vectors{|˜ xi} are orthonormal, since they are formed from
orbits and distinct orbits do not intersect and they span the simulataneous eigenspace of
eigenvalue 1 of the matricesσ(H). We denote the Hilbert space spanned by these vectors
by H/H or H
H
. Note that the vectors |xi are just representatives, and any vector in
its orbit could be used to generate|˜ xi. Since the{|˜ xi} are in one to one correspondence
with the orbits, we let|O
x
i denote a vector inH and|˜ xi denote the corresponding basis
vector inH
H
.
Each basis vector in this space corresponds to a vertex and direction on the quotient
graph, just as the basis vectors ofH, namely{|v,ci} represent a vertex and direction on
Γ. Suppose that a vertex ˜ v on Γ
H
comes from the set of vertices in the orbit O
1
and
that the orbits O
2
,...,O
k
are all the orbits with the same set of vertices. Since each of
these orbits is “connected” to some other orbit (either in this set or outside), the degree
of ˜ v is k. Therefore, all the basis vectors |O
1
i,...,|O
k
i can be associated with ˜ v and
the edges along which they are “connected” to other orbits, as the different directions.
An alternate labelling of these vectors could be |˜ v,c
1
i,...,|˜ v,c
k
i and likewise for each
vertex. Note that this does not produce any natural coloring scheme induced from Γ, on
the edges of Γ
H
.
We now show that any discrete quantum walk on Γ induces a discrete quantum walk
on Γ
H
as long as
ˆ
U respects H i.e., σ(h)
ˆ
Uσ(h)
†
=
ˆ
U ∀h ∈ H. Let us define a discrete
quantumwalkastheapplicationofanyunitary
ˆ
U whichtakesaparticleonagivenvertex
82
v to some superposition of vertices that it is connected to and the directions of only those
edges which connect them to v. On a basis state it acts as
ˆ
U|v,c
i
i=
X
j
a
j
|v(c
j
),c
0
j
i, (5.5)
where |v(c
j
),c
0
j
i and |v,c
j
i are “connected”,
P
j
|a
j
|
2
= 1, ∀j and the sum runs over all
the colors of the edges on the side of v. Given this definition for a walk, we have the
following results.
Theorem 5.5. Let H be a subgroup of the automorphism group of Γ and let
ˆ
U be a
discrete quantum walk defined on Γ such that
ˆ
U respects the symmetries of the subgroup
i.e., [
ˆ
U,σ(h)] = 0, ∀h ∈ H. If the initial state lies in the subspace spanned by all the
orbits {|O
i
i} under the action of H, then the walk is contained in the subspace.
Proof. We have
ˆ
U
t
|O
i
i=
ˆ
U
t
σ(h)|O
i
i=σ(h)
ˆ
U
t
|O
i
i. (5.6)
This shows that since |O
i
i lies in the eigenspace of eigenvalue 1,
ˆ
U
t
|O
i
i also lies in the
same space, which is spanned by{|O
i
i}.
Theorem 5.6. Let H be a subgroup of the automorphism group of Γ and let
ˆ
U be a
discrete quantum walk defined on Γ such that
ˆ
U respects the symmetries of the subgroup
i.e., [
ˆ
U,σ(h)] = 0, ∀h ∈ H. If the initial state lies in the subspace spanned by all the
H-orbits {|O
i
i}, then
ˆ
U induces a walk on Γ
H
in the Hilbert space H
H
.
83
Proof. In order to show that
ˆ
U induces a walk on Γ
H
, we need to show that its action is
similar to Eq. (5.5):
ˆ
U|Oi=
X
j
b
j
|O
j
i, (5.7)
where|O
j
i arethe orbitsconnected to|Oi. But this follows from the fact that if the walk
moves the particle from a vector to vectors “connected” to it, then it does the same for
superpostions of vectors or the orbit states|Oi.
We can derive the structure of this induced walk from the original walk by making
use of its action on the orbit states. The induced walk on the subspace H
H
becomes
ˆ
U
H
=
P
˜ x,˜ y
hO
y
|
ˆ
U|O
x
i|˜ yih˜ x|. This defines a unitary operator inH
H
because,
ˆ
U
†
H
ˆ
U
H
=
X
˜ x,˜ y,˜ y
0
hO
y
0|
ˆ
U
†
|O
x
ihO
x
|
ˆ
U|O
y
i|˜ y
0
ih˜ y|
=
X
˜ x,˜ y,˜ y
0
hO
y
0|
ˆ
U
†
P
H
ˆ
U|O
y
i|˜ y
0
ih˜ y|
= I
H
, (5.8)
since
ˆ
U commutes with P
H
, where P
H
is the projector ontoH
H
. Now consider the shift
matrix of the walk. Its action on H
H
is given by
ˆ
S
H
=
P
˜ x,˜ y
hO
y
|
ˆ
S|O
x
i|˜ yih˜ x|. The
expression hO
y
|
ˆ
S|O
x
i is non-zero if and only if the two orbits are “connected”. If two
orbits are connected then they must be a superposition of the same number of vectors
i.e., |O
x
| =|O
y
| and each vector in the superposition in |O
x
i is connected to one vector
in the superposition in|O
y
i. Therefore,
hO
y
|
ˆ
S|O
x
i=|O
x
|/
q
|O
x
||O
y
|.
84
Thus,
ˆ
S
H
=
X
˜ x,˜ y
|˜ yih˜ x|. (5.9)
This means that the action of
ˆ
S
H
is very similar to the action of
ˆ
S in that it takes the
walker from any vertex to the vertex it is connected to in the quotient graph. The action
of the coin which was
ˆ
I⊗
ˆ
C on the original graph becomes
ˆ
C
H
on the quotient graph so
that
ˆ
U
H
=
ˆ
S
H
ˆ
C
H
. Moreover,
ˆ
C
H
can be decomposed as follows,
ˆ
C
H
=
ˆ
C
1
⊕
ˆ
C
2
⊕···⊕
ˆ
C
N
, (5.10)
where N is the total number of vertices of the quotient graph and each
ˆ
C
i
acts only on
the basis vectors associated with the vertex v
i
of the quotient graph and each
ˆ
C
i
has a
dimension d
i
which corresponds to the degree of the v
i
. In the following examples, such
a decomposition is provided along with a list of the basis vectors on the quotient graph
such that
ˆ
C
1
acts on the first d
1
basis vectors,
ˆ
C
2
acts on the next d
2
vectors etc.
5.3 Examples of quotient graphs
In this section, we illustrate the above abstract formalism with some examples. In all
of the examples we use the following notation to describe the subgroups used to find
quotient graphs. The elements of the subgroups denote permutations of directions, but
it is to be understood that this has to be done along with an appropriate permutation
of vertices, which makes it an automorphism of the graph. Although such a permutation
of vertices need not exist for every permutation of directions, they exist for the examples
85
that we consider here. Moreover, this permutation of vertices can be specified simply:
permute the generators which are in one-to-one correspondence with the directions in the
same way as the directions and this induces a permutation of vertices.
For example, let (1,2) be a group element. This is the automorphism obtained by
interchanging directions 1 and 2 and interchanging generators t
1
and t
2
so that vertices
such as t
1
t
2
go to t
2
t
1
etc. We do not consider direction preserving automorphisms in
the following examples, since they tend to give rise to quotient graphs with self loops.
Finally, we use cycle notation to denote permutations, i.e., (1,2,3) means 1 goes to 2, 2
goes to 3 and 3 goes to 1.
Example 1. As the first example, consider the Cayley graph Γ(S
3
,{(1,2),(2,3)}), and
let {t
1
,t
2
} = {(1,2),(2,3)}. The basis vectors of the Hilbert space of the walk are
{|e,1i,|e,2i,|t
1
,1i,...,|t
1
t
2
t
1
,2i}. TheautomorphismgroupofthisgraphisAut(Γ(S
3
,T))'
R(S
3
)Z
2
. Consider the subgroup H =Z
2
which corresponds to interchanging the direc-
tions 1 and 2. The orbits under the action of this subgroup are
|O
1
i = (1/
√
2)(|e,1i+|e,2i),
|O
2
i = (1/
√
2)(|t
1
,1i+|t
2
,2i),
|O
3
i = (1/
√
2)(|t
1
,2i+|t
2
,1i),
|O
4
i = (1/
√
2)(|t
1
t
2
,2i+|t
2
t
1
,1i),
|O
5
i = (1/
√
2)(|t
1
t
2
,1i+|t
2
t
1
,2i),
|O
6
i = (1/
√
2)(|t
1
t
2
t
1
,1i+|t
1
t
2
t
1
,2i).
86
The original and the quotient graph in this case are shown in Fig. 5.1. The unitary
describing the quantum walk on Γ is given by
ˆ
U =
ˆ
S(
ˆ
I⊗
ˆ
C) where
ˆ
S =
ˆ
S
0
+
ˆ
S
0†
and
ˆ
S
0
= |e,1iht
1
,1|+|e,2iht
2
,2|+|t
1
,2iht
1
t
2
,2|+|t
2
,1iht
2
t
1
,1|+|t
1
t
2
,1iht
1
t
2
t
1
,1|
+ |t
2
t
1
,2iht
1
t
2
t
1
,2|.
This becomes
ˆ
S
H
on the quotient graph and is given by
ˆ
S
H
=
ˆ
S
0
H
+
ˆ
S
0†
H
and
ˆ
S
0
H
=|O
1
ihO
2
|+|O
3
ihO
4
|+|O
5
ihO
6
|. (5.11)
Thiscanalsobewrittenbygivingnewlabelstotheverticesanddirectionsofthequotient
graph:
ˆ
S
0
H
=|v
1
,Rihv
2
,L|+|v
2
,Rihv
3
,L|+|v
3
,Rihv
4
,L|, (5.12)
where we have relabeled|O
1
i through|O
6
i as|v
1
,Ri through|v
4
,Li. Note that there is
no |v
1
,Li and|v
4
,Ri which exactly corresponds to the way these vertices are connected
in the quotient graph. Now, if we take the coin to be C = σ
x
, the Pauli X operator
(which is also the Grover coin in two dimensions), then on the quotient graph the coin
flip matrix
ˆ
F
H
=(
ˆ
I⊗
ˆ
C)
H
becomes
ˆ
F
H
=|v
1
,Rihv
1
,R|+|v
4
,Lihv
4
,L|+
ˆ
F
0
H
+
ˆ
F
0†
H
, (5.13)
87
where
ˆ
F
0
H
=|v
2
,Lihv
2
,R|+|v
3
,Lihv
3
,R|. It can also be written as
ˆ
F
H
=
ˆ
1⊕
ˆ
C
0
⊕
ˆ
C
0
⊕
ˆ
1, (5.14)
where
ˆ
C
0
=
ˆ
X, the Pauli σ
x
operator. Thus, the walk becomes
ˆ
U
H
=
ˆ
S
H
ˆ
F
H
i.e.,
ˆ
U
H
=
0 0 1 0 0 0
1 0 0 0 0 0
0 0 0 0 1 0
0 1 0 0 0 0
0 0 0 0 0 1
0 0 0 1 0 0
.
e
t
2
t
1
t
2
t
2
t
1
t
1
t
2
t
1
e
t
1
t
2
t
1
t
2
t
2
t
1
t
1
t
2
t
1
Quotient graph under H
t
1
Figure 5.1: The graph Γ(S
3
,{(1,2),(2,3)}) and its quotient graph.
88
Example 2. NowconsidertheCayleygraphΓ(S
3
,{(1,2),(2,3),(1,3)})where{(1,2),(2,3),(1,3)}
= {t
1
,t
2
,t
3
}. A subgroup of its automorphism group is S
3
which consists of all per-
mutations of the three directions. Consider a subgroup of this consisting of H
1
=
{e,(1,2,3),(1,3,2)}. Under the action of this subgroup, the orbits are
|O
1
i = (1/
√
3)(|e,1i+|e,2i+|e,3i),
|O
2
i = (1/
√
3)(|t
1
,1i+|t
2
,2i+|t
3
,3i),
|O
3
i = (1/
√
3)(|t
1
,3i+|t
2
,1i+|t
3
,2i),
|O
4
i = (1/
√
3)(|t
1
,2i+|t
2
,3i+|t
3
,1i),
|O
5
i = (1/
√
3)(|t
1
t
2
,1i+|t
1
t
2
,2i+|t
1
t
2
,3i),
|O
6
i = (1/
√
3)(|t
2
t
1
,1i+|t
2
t
1
,2i+|t
2
t
1
,3i).
The shift matrix for this walk becomes
ˆ
S
H
1
=
ˆ
S
0
H
1
+
ˆ
S
0†
H
1
, where
ˆ
S
0
H
1
=|O
1
ihO
2
|+|O
6
ihO
3
|+|O
5
ihO
4
|. (5.15)
We can relabel the quotient graph as shown in Fig (5.2). The matrix
ˆ
S
0
H
1
becomes
ˆ
S
0
H
1
=|v
1
,1ihv
2
,1|+|v
3
,1ihv
2
,2|+|v
4
,1ihv
2
,3|. (5.16)
89
If we choose the Grover coin for the walk, the walk on the quotient graph becomes
ˆ
U
H
1
=
0 −
1
3
2
3
2
3
0 0
1 0 0 0 0 0
0 0 0 0 0 1
0 0 0 0 1 0
0
2
3
2
3
−
1
3
0 0
0
2
3
−
1
3
2
3
0 0
(5.17)
Fig. 5.2 also shows the quotient graphs for the above Cayley graph with subgroupsH
2
'
S
3
and H
3
'{e,(2,3)}. The basis states of the quotient Hilbert spaceH
H
2
are
|O
1
i = (|e,1i+|e,2i+|e,3i)/
√
3,
|O
2
i = (|t
1
,1i+|t
2
,2i+|t
3
,3i)/
√
3,
|O
3
i = (|t
1
,2i+|t
1
,3i+|t
2
,1i+|t
2
,3i
+ |t
3
,1i+|t
3
,3i)/
√
6,
|O
4
i = (|t
1
t
2
,1i+|t
1
t
2
,2i+|t
1
t
2
,3i+|t
2
t
1
,1i
+ |t
2
t
1
,2i+|t
2
t
1
,3i)/
√
6,
90
and the basis states ofH
H
3
are
|O
1
i = |e,1i,
|O
2
i = (|e,2i+|e,3i)/
√
2,
|O
3
i = |t
1
,1i,
|O
4
i = (|t
1
,2i+|t
1
,3i)/
√
2,
|O
5
i = (|t
2
,2i+|t
3
,3i)/
√
2,
|O
6
i = (|t
2
,3i+|t
3
,2i)/
√
2,
|O
7
i = (|t
2
,1i+|t
3
,1i)/
√
2,
|O
8
i = (|t
1
t
2
,3i+|t
2
t
1
,2i)/
√
2,
|O
9
i = (|t
1
t
2
,1i+|t
2
t
1
,1i)/
√
2,
|O
10
i = (|t
1
t
2
,2i+|t
2
t
1
,3i)/
√
2.
The unitary corresponding to the walk on the quotient graph of H
2
is
ˆ
U
H
2
=
0 −1/3 2
√
2/3 0
1 0 0 0
0 0 0 1
0 −2
√
2/3 1/3 0
, (5.18)
and the one on the quotient graph of H
3
is
ˆ
U
H
3
=
ˆ
S
H
3
·
ˆ
C
H
3
. (5.19)
91
e
t
1
t
2
t
3
t
1
t
2
t
2
t
1
H
1
H
3
H
2
e
t
1
t
2
t
3
t
1
t
2
t
2
t
1
e
t
1
t
2
t
3
t
1
t
2
t
2
t
1
e
t
2
t
3
t
1
t
1
t
2
t
2
t
1
Figure 5.2: The graph Γ(S
3
,{(1,2),(1,3),(2,3)}) and its quotient graphs.
The matrices
ˆ
S
H
3
and
ˆ
C
H
3
are given by,
ˆ
S
H
3
=
ˆ
S
0
+
ˆ
S
0†
and
ˆ
C
H
3
=
ˆ
C
0
⊕
ˆ
C
0
⊕
ˆ
C
00
⊕
ˆ
C
00
where,
ˆ
S
0
= (|O
1
ihO
3
|+|O
2
ihO
5
|+|O
4
ihO
10
|+|O
6
ihO
9
|
+ |O
7
ihO
8
|,
ˆ
C
0
=
−1/3 2
√
2/3
2
√
2/3 1/3
92
and
ˆ
C
00
=
−1/3 2/3 2/3
2/3 −1/3 2/3
2/3 2/3 −1/3
.
Example 3. In this example, we determine the quotient graph of Γ(S
4
,T) for T =
{(1,2),(1,3),(1,4)} = {t
1
,t
2
,t
3
} under the subgroup H ' S
3
which corresponds to all
possible permutations of the directions at each vertex. The original and the quotient
graphs are shown in Fig (5.3), where “t” has been dropped in the vertex labels. There
are 14 orbits under the action of this subgroup. These are
|O
1
i = (|e,1i+|e,2i+|e,3i)/
√
3,
|O
2
i = (|t
1
,1i+|t
2
,2i+|t
3
,3i)/
√
3,
|O
3
i = (|t
1
,2i+|t
2
,1i+|t
3
,2i
+ |t
2
,3i+|t
1
,3i+|t
3
,1i)/
√
6,
|O
4
i = (|t
1
t
2
,1i+|t
2
t
1
,2i+|t
3
t
2
,3i
+ |t
2
t
3
,2i+|t
3
t
1
,3i+|t
1
t
3
,1i)/
√
6,
|O
5
i = (|t
1
t
2
,2i+|t
2
t
1
,1i+|t
3
t
2
,2i
+ |t
2
t
3
,3i+|t
3
t
1
,1i+|t
1
t
3
,3i)/
√
6,
93
|O
6
i = (|t
1
t
2
,3i+|t
2
t
1
,3i+|t
3
t
2
,1i
+ |t
2
t
3
,1i+|t
3
t
1
,2i+|t
1
t
3
,2i)/
√
6,
|O
7
i = (|t
1
t
2
t
1
,1i+|t
1
t
2
t
1
,2i+|t
3
t
2
t
3
,3i
+ |t
3
t
2
t
3
,2i+|t
1
t
3
t
1
,3i+|t
1
t
3
t
1
,1i)/
√
6,
|O
8
i = (|t
1
t
2
t
1
,3i+|t
3
t
2
t
3
,1i+|t
1
t
3
t
1
,2i)/
√
3,
|O
9
i = (|t
1
t
2
t
3
,1i+|t
2
t
1
t
3
,2i+|t
3
t
2
t
1
,3i
+ |t
1
t
3
t
2
,1i+|t
2
t
3
t
1
,2i+|t
3
t
1
t
2
,3i)/
√
6,
|O
10
i = (|t
1
t
2
t
3
,2i+|t
1
t
3
t
2
,3i+|t
2
t
1
t
3
,1i
+ |t
2
t
3
t
1
,3i+|t
3
t
1
t
2
,1i+|t
3
t
2
t
1
,2i)/
√
6,
|O
11
i = (|t
1
t
2
t
3
,3i+|t
1
t
3
t
2
,2i+|t
2
t
1
t
3
,3i
+ |t
2
t
3
t
1
,1i+|t
3
t
1
t
2
,2i+|t
3
t
2
t
1
,1i)/
√
6,
|O
12
i = (|t
3
t
1
t
2
t
1
,1i+|t
2
t
3
t
2
t
1
,3i+|t
3
t
1
t
2
t
1
,2i
+ |t
2
t
3
t
2
t
1
,2i+|t
1
t
3
t
1
t
2
,1i+|t
1
t
3
t
1
t
2
,3i)/
√
6,
|O
13
i = (|t
1
,1i+|t
2
,2i+|t
3
,3i)/
√
3,
|O
14
i = (|t
1
t
3
t
2
t
1
,1i+|t
1
t
3
t
2
t
1
,2i+|t
1
t
3
t
2
t
1
,3i
+ |t
2
t
3
t
1
t
2
,1i+|t
2
t
3
t
1
t
2
,2i+|t
2
t
3
t
1
t
2
,3i)/
√
6.
The unitary walk on the quotient graph can be written as
ˆ
U
H
=
ˆ
S
H
·
ˆ
C
H
. (5.20)
94
Thematrices
ˆ
S
H
and
ˆ
C
H
aregivenby,
ˆ
S
H
=
ˆ
S
0
+
ˆ
S
0†
and
ˆ
C
H
= 1⊕
ˆ
C
0
⊕
ˆ
C
00
⊕
ˆ
C
0
⊕
ˆ
C
00
⊕
ˆ
C
0
⊕1
where,
ˆ
S
0
= (|O
1
ihO
2
|+|O
3
ihO
4
|+|O
5
ihO
7
|+|O
6
ihO
9
|
+ |O
8
ihO
12
|+|O
10
ihO
13
|+|O
11
ihO
14
|),
ˆ
C
0
=
−1/3 2
√
2/3
2
√
2/3 1/3
and
ˆ
C
00
=
−1/3 2/3 2/3
2/3 −1/3 2/3
2/3 2/3 −1/3
.
Example 4. Consider the hypercube. The automorphism group of the hypercube is
Aut(Γ(Z
n
2
,Y)) ' Z
n
2
S
n
. We focus on the subgroup H
1
= S
n
and look at the resulting
quotient graph. We consider the case when n = 3, but the procedure for a general n is
very similar. The subgroupH
1
consists of all possible permutations ofn directions. The
orbits under the action of this subgroup are given by
|O
1
i = (|000,1i+|000,2i+|000,3i)/
√
3,
|O
2
i = (|001,1i+|010,2i+|100,3i)/
√
3,
|O
3
i = (|001i(|2i+|3i)+|010i(|1i+|3i)
+ |100i(|1i+|2i))/
√
6,
95
e
1
2
3
21
31
12
32
13
23
121
321
131
231
312
232
132
213
123
3121
2321
1312
1321
2312
e 1,2,3
21,31,12
32,13,23
121,323,131
123,213,321
132,231,312
3121,2321
1312
1321,2312
Figure 5.3: The graph Γ(S
4
,{(1,2),(1,3),(1,4)}) and it quotient graph under the sub-
group H.
|O
4
i = (|011i(|1i+|2i)+|101i(|1i+|3i)
+ |110i(|2i+|3i))/
√
6,
|O
5
i = (|011,3i+|101,2i+|110,1i)/
√
3,
|O
6
i = (|111,1i+|111,2i+|111,3i)/
√
3.
The graph becomes a line as shown in Fig. 5.4 and all the vertices of a certain Hamming
weight collapse to a point. This fact had first been observed in [57]. In [72], this idea
96
was used to construct a search algorithm on the hypercube. As observed in [72], the
states on the line can be relabeled|0,Ri,|1,Li,|1,Ri,|2,Li,|2,Ri,|3,Li. For the general
hypercube of dimension n, these states generalize to
|x,Ri =
s
1
(n−x)
n
x
X
|~ x|=x
X
x
d
=0
|~ x,di,
|x,Li =
s
1
(x)
n
x
X
|~ x|=x
X
x
d
=1
|~ x,di, (5.21)
where|~ x| is the Hamming weight of~ x.
000 001
010
100
101
111
110
011
000
100
010
001
110
101
011
111
H
1
H
2
000
001
111
110
011
101
100
010
Figure 5.4: The n= 3 hypercube and its quotient graphs.
97
Under the action of a different subgroup H
2
= S
n−1
consisting of permutations of
n−1 directions and the corresponding permutations of vertices, the quotient graph is
shown in Fig. (5.4). The basis states ofH
H
2
when n = 3, are
|O
1
i = |000,1i,
|O
2
i = (|000,2i+|000,3i)/
√
2,
|O
3
i = |001,1i,
|O
4
i = (|001,2i+|001,3i)/
√
2,
|O
5
i = (|010,2i+|100,3i)/
√
2,
|O
6
i = (|010,1i+|100,1i)/
√
2,
|O
7
i = (|010,3i+|100,2i)/
√
2,
|O
8
i = (|011,2i+|101,3i)/
√
2,
|O
9
i = (|011,1i+|101,1i)/
√
2,
|O
10
i = (|011,3i+|101,2i)/
√
2,
|O
11
i = (|110,2i+|110,3i)/
√
2,
|O
12
i = |110,1i,
|O
13
i = (|111,2i+|111,3i)/
√
2,
|O
14
i = |111,1i.
Foranygeneraln,thegraphisstillplanarasshowninFig.5.5andtherewillbe6n−4
basisstates. Theycanbelabeledas|x0,Li,|x0,Ri,|x0,Di,|x1,Li,|x1,Ri,|x1,Ui,where
xistheHammingweightofthelastn−1bits(whichfallundertheactionofthesubgroup
98
x0,x=0
x1,x=0
x0,x=1
x1,x=1
x0,x=2
x1,x=2
x0,x=n-1
x1,x=n-1
Figure 5.5: The quotient graph of a general hypercube under the group S
n−1
.
S
n−1
) and the bit next to it is the first bit. L,R,U and D stand for left, right, up and
down respectively. They are given by
|x0,Ri =
s
1
(n−1−x)
n−1
x
X
|~ x|=x
X
x
d
=0
|~ x0,di,
|x0,Li =
s
1
(x)
n−1
x
X
|~ x|=x
X
x
d
=1
|~ x0,di
|x0,Di =
s
1
n−1
x
X
|~ x|=x
|~ x0,1i
|x1,Ri =
s
1
(n−1−x)
n−1
x
X
|~ x|=x
X
x
d
=0
|~ x1,di,
|x1,Li =
s
1
(x)
n−1
x
X
|~ x|=x
X
x
d
=1
|~ x1,di
|x1,Ui =
s
1
n−1
x
X
|~ x|=x
|~ x1,1i. (5.22)
99
Note that the states |x0,Ui and |x1,Di do not exist. Moreover, |x0,Li and |x1,Li do
not exist when x = 0 and |x0,Ri and |x1,Ri do not exist when x =n−1. The unitary
matrices describing the walk on these graphs are
ˆ
U
H
1
=
0 −1/3 2
√
2/3 0 0 0
1 0 0 0 0 0
0 0 0 −1/3 2
√
2/3 0
0 2
√
2/3 1/3 0 0 0
0 0 0 0 0 1
0 0 0 2
√
2/3 1/3 0
, (5.23)
and
ˆ
U
H
2
=
ˆ
S
H
2
·
ˆ
C
H
2
. (5.24)
The matrices
ˆ
S
H
2
and
ˆ
C
H
2
are given by,
ˆ
S
H
2
=
ˆ
S
0
+
ˆ
S
0†
and
ˆ
C
H
2
=
ˆ
C
0
⊕
ˆ
C
0
⊕
ˆ
C
00
⊕
ˆ
C
00
⊕
ˆ
C
0
⊕
ˆ
C
0
where,
ˆ
S
0
= (|O
1
ihO
3
|+|O
2
ihO
5
|+|O
4
ihO
8
|+|O
6
ihO
9
|
+ |O
7
ihO
11
|+|O
10
ihO
13
|+|O
12
ihO
14
|),
ˆ
C
0
=
−1/3 2
√
2/3
2
√
2/3 1/3
100
and
ˆ
C
00
=
−1/3 2/3 2/3
2/3 −1/3 2/3
2/3 2/3 −1/3
.
Example 5. While we have shown how to construct quotient graphs for discrete-time
walks on Cayley graphs, the idea of a quotient graph is more general. In this example,
we consider the “glued trees” graph shown in Fig. 5.6. This graph is not regular and
hence not a Cayley graph. It is undirected, and so we can easily define a continuous
walk on it. Because the continuous walk does not have a coin space, we need not consider
permutationsofdirectionsintheautomorphisms. Quantumwalksonthisgraphwerefirst
analyzed in [18], and it has been shown that quantum walks move exponentially faster
on this graph from “entrance” to “exit” than classical walks. The main reason for this
exponentialspeedupisthatthequantumwalkmovesinasuperpositionofallthevertices
in a given column. It can be seen that in any given column, the vertices which branch
out from the same vertex in the previous column can be interchanged as long as the
corresponding interchange on the other side of the central column takes place. Therefore,
the automorphism group of this graph is Z
k
2
, where k is one half of the total number of
vertices on one side of the central column. Under the action of these automorphisms, the
vertices in each column form a single orbit, and hence collapse to a single point in the
quotient graph. There are 2n+1 orbits under the action of this subgroup, where the
columns j are such that 0≤j≤2n. The orbits can be written as
|O
j
i= 2
−min[j,2n−j]/2
X
v∈ column j
|vi. (5.25)
101
The Hamiltonian for the quantum walk on the quotient graph becomes [40]
h
˜
j|H|
˜
j±1i = −
√
2γ
h
˜
j|H|
˜
ji =
2γ j = 0,n,2n
3γ otherwise,
(5.26)
with all other matrix elements zero. This is also shown in Fig. 5.6 where the γ has been
dropped for brevity.
A
A
A
A
A
A
A
A
@
@
@
@
@
@
@
@
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
A
A
A
A
A
A
A
A
@
@
@
@
@
@
@
@
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
v v
v v
v v
v
v
v
v
v v
v v
v v
v v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v
v v v v v v v v v
−
√
2 −
√
2 −
√
2 −
√
2 −
√
2 −
√
2 −
√
2 −
√
2
2 3 3 3 2 3 3 3 2
Figure 5.6: The glued trees graph and its quotient graph.
102
5.4 Automorphism group of the quotient graph
In this section we determine the automorphisms of the quotient graph which are induced
from the automorphism group of the original graph and the subgroup used to obtain
the quotient graph. In dealing with the automorphisms of Γ we used permutations of
vertices and edges, and this in turn corresponds to permutations of basis vectors which
preserves the shift matrix. On the quotient graph, we define those permutations of orbits
which preserve the
ˆ
S
H
matrix as automorphisms, since there is no natural choice of edge
colors. These permutations of orbits which preserve the new shift matrix also preserve
the quotient graph.
LetG
1
be a set of automorphisms of Γ which are of the following type. If they take a
basis vector belonging to aH-orbitO
1
to a basis vector belonging toO
2
, then they take
every basis vector inO
1
to some basis vector inO
2
. Clearly, all the automorphisms inH
are of this type, under the special case when O
1
=O
2
. It is also easy to verify that G
1
is a subgroup of G, and that H is a subgroup of G
1
. We now show that H is a normal
subgroup of G
1
i.e., ghg
−1
∈H,∀g∈G
1
and∀h∈H.
Theorem 5.7. Given the group G
1
defined as above we have,
1. The subgroup H is a normal subgroup of G
1
.
2. G
1
is the largest subgroup of G such that H is a normal subgroup of G
1
—that is,
for any g∈G, if ghg
−1
∈H ∀h∈H then g∈G
1
.
Proof. We show this by considering the action of all of these group elements on the set
of basis vectors.
103
1. Let x be any basis element belonging to some H-orbit O
1
, and let g take every
element in O
2
to some element in O
1
. Then ghg
−1
x = ghy, where y ∈ O
2
. Now,
hy =z where z ∈O
2
since these orbits are formed under the action of H. Hence,
gz =x
0
, where x
0
∈O
1
. But every x
0
∈O
1
can be written as h
0
x for some h
0
∈H.
Thus, ghg
−1
∈H.
2. Consider some basis element x∈O
3
and let gx∈O
4
. Since ghx =h
0
gx, g(hx) =
h
0
y =y
0
, where y,y
0
∈O
4
. Therefore, g takes hx∈O
3
to y
0
∈O
4
, but since h∈H
is arbitrary, g takes every element of O
3
to some element of O
4
. It follows that
g∈G
1
.
Since H is normal in G
1
, the quotient set G
1
/H, i.e., the set of all cosets gH =
{gh|g∈H}, is a group. This group has a natural representation in the Hilbert spaceH
H
as a permutation matrix in the basis where each orbit is a basis vector.
Theorem 5.8. G
1
/H ⊂ Aut(Γ
H
).
Proof. Consideranyautomorphismg∈G
1
andletσ(g)beitsrepresentationinH. Then,
we have σ(g)Sσ(g
−1
)=S. The projection of this intoH
H
is given by
P
H
σ(g)Sσ(g
−1
)P
H
=P
H
SP
H
=S
H
. (5.27)
104
The representationσ(g) commutes withP
H
, since it permutes all the vectors in an orbit
to vectors in another orbit. Therefore,
P
H
σ(g)P
H
SP
H
σ(g
−1
)P
H
=S
H
. (5.28)
But as a representation, P
H
σ(g)P
H
= σ(gH). This means that the representation of
gH ∈G
1
/H inH
H
is a group of symmetries of S
H
and therefore G
1
/H ⊂Aut(Γ).
We see that the quotient graph is obtained from Γ modulo the symmetries in H.
5.5 Hitting time on quotient graphs
In this subsection, we address the question of when quantum walks on quotient graphs
haveinfinitehittingtimes. Itispossiblethatforsomesubgroups,thewalkonthequotient
graph does not have infinite hitting times even if the walk on the original graph does.
In order to carry over the discussion of hitting times to quotient graphs, we must keep
in mind that the evolution operator is now followed by a measurement. To remain on
the quotient graph (i.e., in the subspace given by
ˆ
P
H
), the measurement operators must
commute with the symmetry operators σ(h), h∈H.
If this condition is satisfied, then we can obtain a condition to check whether the
quotient graph has initial states with infinite hitting times: if the subspace of those
initial states with infinite hitting times on the original graph whose projector is
ˆ
P, has no
nontrivial intersection with the subspace whose projector is
ˆ
P
H
i.e.,
ˆ
P∩
ˆ
P
H
=∅, then the
walk on the quotient graph does not have infinite hitting times. If there is a nontrivial
intersection, thenitdoes. (Herewehaveusedtheprojectorsontothesubspacestodenote
105
the spaces themselves.) This condition can also be verified by obtaining the restriction
of the evolution operator and the measurement operators onto the quotient graph. By
diagonalizing the new unitary evolution operator and constructing the projector of states
˜
P which have no overlap with the new final vertex state. The subspace of these states is
exactly the intersection
ˆ
P∩
ˆ
P
H
. We will examine this condition for some of the examples
considered above.
In the first example, we choose the final vertex to be t
1
t
2
t
1
. The measurement op-
erators are
ˆ
P
f
=|t
1
t
2
t
1
iht
1
t
2
t
1
|⊗
ˆ
I and
ˆ
I−
ˆ
P
f
. This measurement commutes with the
subgroup chosen, and the quotient graph does not have infinite hitting times. This is
because the original graph does not have infinite hitting times either i.e.,
ˆ
P =∅.
For the second example, for the graph Γ(S
3
,{(1,2),(2,3),(1,3)}), we used three dif-
ferent subgroups and form their quotient graphs. In order to determine whether the
quotient graph has infinite hitting times for various subgroups, we must choose different
final vertices for the different subgroups since the measurement must commute with the
symmetries. Therefore, for H
1
we choose t
1
t
2
as the final vertex, and the measurement
operators are |t
1
t
2
iht
1
t
2
|⊗
ˆ
I and its orthogonal complement. This measurement com-
mutes with the subgroup H
1
. For this final vertex and measurement, the original graph
has infinite hitting times i.e.,
ˆ
P 6=∅ and the quotient graph also has infinite hitting times
i.e.,
ˆ
P∩
ˆ
P
H
1
6=∅. In fact, using theC-matrix defined above in Eq. (3.53), we find that if
the initial vertex is the identity|ei, then there is no superposition of coin states that has
a finite hitting time, because
ˆ
C
v
does not have a zero eigenvalue for v =e.
For the subgroup H
2
, we choose the final vertices to be t
1
t
2
and t
2
t
1
. Therefore,
the measurement on the original graph must be a projective measurement with outcomes
106
ˆ
P
f
=(|(t
1
t
2
iht
1
t
2
|+|t
2
t
1
iht
2
t
1
|)⊗
ˆ
I anditsorthogonalcomplement. Forthismeasurement
and final vertices, the original graph has
ˆ
P =∅. Therefore, the quotient graph also does
not have infinite hitting times. For the subgroup H
3
, the measurement operators are
ˆ
P
f
=(|(t
1
t
2
iht
1
t
2
|+|t
2
t
1
iht
2
t
1
|)⊗
ˆ
I anditsorthogonalcomplement. Forthismeasurement,
neither the original graph nor the quotient graph have infinite hitting times.
In example 3, for the Cayley graph Γ(S
4
,{(1,2),(1,3),(1,4)}), we choose the final
vertices to be|t
1
t
3
t
2
t
1
i and|t
2
t
3
t
1
t
2
i. In this case, we find that while the original graph
has infinite hitting times, the quotient graph does not. In fact, on the original graph,
the equal superposition of all coin states at the vertex|ei is the only superposition which
does not have an infinite hitting time (i.e., the
ˆ
C
v
matrix has only one zero eigenvalue
with the equal superposition of coin states as its eigenvector). It is precisely this vector
which is included in the subspace of the quotient graph. This is not a coincidence—in
both cases, it is picked out by the symmetries of the graph.
In example 4, for the two different subgroups of the automorphism group considered
for the hypercube, we find that the behavior of hitting times is very different. For the
subgroup H
1
, the quotient graph becomes a line with the vertex
~
0 = 00...0 on one end
and the vertex
~
1 = 11...1 on the other. If one designates the final vertex to be
~
1 by
choosingthemeasurementoperatorstobe
ˆ
P
f
=|
~
1ih
~
1|⊗
ˆ
I anditsorthogonalcomplement,
then we find that this quotient graph does not have infinite hitting times for any initial
state:
ˆ
P∩
ˆ
P
H
1
=∅. In fact, if the initial state is|00...0i⊗
1
d
P
i
|ii, then the hitting time
is polynomial in d, the dimension of the hypercube [38, 47]. Using the C-matrix for the
original graph, we find that if the initial vertex is 00...0, then the equal superposition
107
of all directions is the only zero eigenvector of C
v
, which means that it is the only coin
state that does not have an infinite hitting time.
On the other hand, choosing 11...10 as the final vertex and using the subgroup H
2
,
wefindthatthequotientgraphshowninFig.5.4doeshaveinfinitehittingtimesforsome
initial states i.e.,
ˆ
P∩
ˆ
P
H
2
6=∅. Using theC-matrix again, we find that if the initial vertex
is 00...0, then the equal superposition of all directions once again is the only coin state
that has no infinite hitting times. For every other superposition of coin states for that
vertex (i.e., every other|αi in Eq. (3.51)) C
v
has a nonzero eigenvalue.
5.6 Discussion
We have investigated the behavior of quantum walks on undirected graphs by making
use of the automorphism group of the graph. Automorphisms of the graph may become
symmetriesofthediscretequantumwalk,dependingonthesymmetriesofthecoinmatrix.
Quantum walks which respect the symmetries of some subgroupH of this automorphism
group have an invariant subspace in the total Hilbert space. We showed that the walk
restricted to this subspace can be seen as a (different) quantum walk on a quotient graph,
andthatthisgraphcanbeconstructedfromtheoriginalgraphgiventhesubgroupH. The
dynamics of the new walk can also be derived from the original walk and the subgroup.
The quotient graph is obtained from the original graph by identifying vertices and edges
which form an orbit under the action of H; this means that the quotient graph and the
new quantum walk both have no symmetries coming from H. The new quantum walk
108
only has the remaining automorphisms as its possible symmetries, and so it has, in a
sense, “used up” the ones in H.
To discuss hitting times, we use the measured walk defined in [38] and [47], which
consists of the application of a unitary operator followed by a projective measurement at
each time step. For the walk on the quotient graph to be preserved, the choice of mea-
surement must commute with the symmetries in H. This restriction is very important:
even if the walk and initial state both have a larger group of symmetries, the walk will
be on a quotient graph corresponding to a smaller subgroup H if the measurement does
not commute with the remaining elements of the larger group.
Forinstance,inExample2,usingthesubgroupH
1
,weobtainedawalkonitsquotient
graph. Suppose the measurement is a projective measurement of the vertex t
1
t
2
. The
initialstate|ei⊗(|1i+|2i+|3i)/
p
(3), andthewalkwiththeGrovercoin
ˆ
U bothhaveall
the symmetries of H
2
. But the measurement, which commutes with all the elements of
H
1
, does not commute with all those inH
2
, and the effective walk will be on the quotient
graph corresponding to H
1
.
The remaining symmetries of the evolution operator can lead to degeneracy in its
eigenspectrum [48], and may result in infinite hitting times on the quotient graph. In
general, we found a condition to determine whether the walk on the quotient graph will
have infinite hitting times: given the original graph, the quantum walk and any subgroup
H, one can determine the projector onto states with infinite hitting time
ˆ
P, and the
invariant subspace of the quotient graph
ˆ
P
H
. If
ˆ
P∩
ˆ
P
H
=∅ then the quotient graph does
not have infinite hitting times.
109
Even when the hitting time is not infinite for an initial state on the quotient graph, it
is possible that it could be extremely long. It would be useful to have a criterion to pick
out subgroups of the automorphism group whose quotient graphs have exponentially fast
hitting times. For example, in the case of the hypercube, using the subgroup H
1
(whose
quotient graph is a line) turns out to give very fast hitting times. But on a general
undirected graph it is not easy to determine whether there is a subgroup whose quotient
graph gives fast hitting times.
To investigate this, we need to make the notion of “fast” more precise. One way to
define “fast” for a parametrized class of graphs (such as the hypercube, where the pa-
rameter is the dimension) is to say that the hitting time must beO(logN) (exponentially
smaller),whereN isthenumberofverticesofthegraph. Usingthisnotion,wecanexpect
fast hitting times to exist in graphs which have quotient graphs with an exponentially
smaller number of vertices. While this is not necessarily a sufficient condition for fast
hitting times, it is interesting to observe that both the quantum walk search algorithm
on the hypercube and the glued trees graph are examples of symmetric graphs where
the quotient graph is exponentially smaller than the original graph. In the case of the
hypercube, the hitting time for the effective walk on the quotient graph is exponentially
smaller than the number of vertices in the original graph, and exponentially smaller than
the classical hitting time; the same is true of the continuous-time walk on the glued-trees
graph. It is interesting to note that in both of these cases the quotient graph is a finite
line. Thisseemstosuggestthatifwecanidentifygraphswhichhavethelineasaquotient
graph, they may be fruitful ground to look for more examples of walks with fast hitting
times. Thisremainsverymuchanopenquestion,butitisourbeliefthatgraphsymmetry
110
is of vital importance in the understanding of hitting times for quantum walks, and that
understanding the structure of quotient graphs is the key to further progress.
111
Chapter 6
Quantum walks in decoherence-free subspaces
We have seen in Chapter 3, that decoherence has a detrimental effect on hitting times
of quantum walks, especially fast hitting times. This is because the examples of fast
hitting times that have been found in the literature are due to the symmetry of the
graph. This would not be the case in the presence of decoherence since it would break
this symmetry and force the walk to leave this subspace. Infinite hitting times are also
due to symmetry and its effect of confining the walk to a subspace. We have examined
the effect of decoherence on infinite hitting times and we have seen that it makes them
finite by breaking the symmetry. However, if the decoherence has symmetries of its
own, then it can have invariant subspaces and the evolution in these subspaces can be
purely unitary. In this chapter, we explore this possibility. We derive conditions on the
decoherencesuchthatithasenoughsymmetrytopreservetheevolutioninsomesubspace.
This can be applied to preserve subspaces that lead to fast hitting times. The concept
of decoherence-free subspaces (DFS) has been the subject of intense research in the last
few years. Decoherence-free subspaces are a result of the symmetry associated with the
decoherence operators and as such are subspaces in which the evolution is purely unitary.
112
Refs. [51,52]presentoverviewsoftheprogressinthisfield. Theconditionsunderwhicha
given subspace becomes a DFS of the evolution is given for various different formulations
of the evolution. Here, we first give the basic definitions and conditions of decoherence-
free subspaces. Then we review the general theory of DFS using representation theory of
complex algebras and apply it to quantum walks to arrive at a general condition for the
subspaceH
H
to be a DFS (recall from Chapter 5, that the subspaceH
H
is the subspace
of the quotient graph that gives fast hitting times for some graphs.) We also provide
examples of decoherence which have the subspaceH
H
inside a decoherence-free subspace
for both the discrete-time and continuous-time walks on the hypercube.
6.1 Decoherence-free subspaces and quantum walks
In order to apply the DFS formalism to the discrete-time quantum walk, consider an
evolution given by the OSR
D(ρ) =
X
i
ˆ
A
i
ρ
ˆ
A
†
i
, (6.1)
where the Kraus operators satisfy the relation
P
i
ˆ
A
†
i
ˆ
A
i
=
ˆ
I.
Definition 6.1. A system with a Hilbert space H is said to have a decoherence free
subspaceH
0
if every pure state in this subspace is invariant under the OSR i.e.,∀|ji∈H
0
we have
X
i
ˆ
A
i
|jihj|
ˆ
A
†
i
=|jihj|. (6.2)
We now state the main theorem which gives the necessary and sufficient condition for
a subspace to be a DFS.
113
Theorem 6.2. A subspace
˜
H is a decoherence-free subspace iff the Kraus operators act
proportional to the identity on the subspace, i.e., ∀ |ji∈H
0
ˆ
A
i
|ji=c
i
|ji. (6.3)
The above theorem states that for H
0
to be a DFS of the given OSR, the entire
subspaceH
0
must lie inside a single degenerate eigenspace of each of the Kraus operators.
This can be put in the language of the representation theory of the algebras, generated
by the operators
ˆ
A
i
. We state the important results in this direction. First, it can be
shown that the Kraus operators generate a complex, associative and †-closed algebra A
(†-closed means that if M ∈A, then M
†
∈A.) We also assume that the identity lies in
this algebra. We now have the following results.
Theorem 6.3. If A is a complex, associative, †-closed, algebra such that I ∈A, then A
has the following decomposition:
A
∼
=
M
J∈J
I
n
J
⊗M(d
J
,C). (6.4)
This states that A reduces to a direct sum of irreducible complex matrix algebras
M(d
J
,C) of dimension d
J
with a multiplicity n
J
. The Hilbert space on which they
operate can be decomposed similarly as
H =
M
J∈J
H
n
J
⊗H
d
J
. (6.5)
114
Fromthetheoryofdecoherence-freesubspaces[51],thefollowingresultgivesthecondition
for a subspace to be a DFS.
Theorem6.4. ThenecessaryandsufficientconditionforasubspaceH
0
tobeadecoherence-
free subspace of the dynamics is that it must lie in the degeneracy of a single one-
dimensional irreducible representation.
This theorem states that in order to be a DFS, H
0
must lie in the I
n
K
part of the
decomposition for which d
K
= 1 i.e., using the decomposition in Eq. (6.5), H
0
⊂ H
n
K
.
When d
K
6= 1, we obtain the more general decoherence-free subsystems. To apply this
to quantum walks, we are interested only in decoherence-free subspaces and so we only
need one-dimensional irreducible subalgebras. Finally, to perform computations which
preservetheDFS,theunitaryoperationsmustcommutewiththealgebrainsidetheDFS.
This means that the general form for the operations that preserve this DFS is
ˆ
U ∈M(n
K
,C)⊕M(d−n
K
,C), (6.6)
where d is the total dimension of the Hilbert space.
We now apply this to quantum walks with decoherence and find conditions under
whichthedecoherencepreservesthesubspaceofinterest–H
H
. Firstly,usingthecondition
of degeneracy in Eq. (6.3), we need that
ˆ
A
i
|O
j
i=c
i
|O
j
i (6.7)
115
whereO
j
is an orbit under the action of H and is a basis vector inH
H
. In order to find
a condition in terms of the algebra generated by the symmetry operators, σ(h), we first
note that for the symmetry operators, ifH
0
is a degenerate eigenspace i.e.,
σ(g)|ji=c
g
|ji, (6.8)
forall|ji∈H
0
, thenthis subspace is preservedby the unitary evolution operator
ˆ
U. This
can be seen easily, for we have
σ(g)
ˆ
U
t
|ji=
ˆ
U
t
σ(g)|ji=c
g
ˆ
U
t
|ji. (6.9)
Now, the complex algebra generated by the group of symmetry operators (denote it by
A
H
) is associative,†-closed and contains the identity. Using the above general theory, we
can decompose this into a sum of irreducible subalgebras as
A
H
∼
=
M
J∈J
0
I
n
J
⊗M(d
J
,C). (6.10)
The condition Eq.(6.8) is equivalent to the condition that the subspaceH
H
must belong
to the degeneracy of a one-dimensional irreducible subalgebra in order to be preserved by
the evolution and the evolution operator
ˆ
U must lie in the algebra that commutes with
A
H
in the DFS, i.e., if for some K
0
∈J
0
for which d
0
K
= 1, the subspace H
H
lies in its
116
degeneracy, then
ˆ
U must lie in the algebra M(n
0
K
,C)⊕M(d−n
0
K
,C). Combining the
these conditions, we have that
A
∼
= I
n
K
⊗C⊕
M
J∈J−K
I
n
J
⊗M(d
J
,C)
A
H
∼
= I
n
0
K
⊗C⊕
M
J∈J
0
−K
0
I
n
J
⊗M(d
J
,C)
ˆ
U ∈ M(n
0
K
,C)⊕M(d−n
0
K
,C)
= M(n
K
,C)⊕M(d−n
K
,C), (6.11)
where we needn
0
K
=n
K
. Now ifH
H
⊂H
n
K
, then the quantum walk which starts inside
H
H
will be unperturbed by decoherence.
As an example, consider the discrete-time quantum walk on the hypercube. Consider
a decoherence which has the following Kraus operators
ˆ
A
i
=κ
i
ˆ
S
ii+1
⊗(|iihi+1|+|i+1ihi|), (6.12)
wherei∈{1,2,...,n−1}.
ˆ
S
ij
is the swap operation on the qubits numberedi andj. It
is given by
ˆ
S
ij
=
ˆ
I⊗···⊗
ˆ
I
| {z }
i−1
⊗
ˆ
S
2
⊗
ˆ
I⊗···⊗
ˆ
I
| {z }
n−i−1
, (6.13)
where
ˆ
S
2
=|0ih0|⊗|0ih0|+|1ih1|⊗|1ih1|+|0ih1|⊗|1ih0|+|1ih0|⊗|0ih1|. (6.14)
117
The constants κ
i
need to satisfy
n−1
X
i=1
|κ
i
|
2
= 1 (6.15)
inordertohave
P
i
ˆ
A
†
i
ˆ
A
i
=
ˆ
I. Thiskindofdecoherencemightoccurinimplementationsof
quantumwalksonthehypercubeusingnspins. Thedecoherenceislocalinthesensethat
it only affects adjacent spins. This decoherence has the subspace H
H
with basis vectors
given above as a DFS. These basis vectors and Kraus operators satisfy the condition in
Eq. (6.7) since we have
ˆ
A
i
|O
j
i=κ
i
(
ˆ
S
ii+1
⊗(|iihi+1|+|i+1ihi|))|O
j
i=κ
i
|O
j
i. (6.16)
Each|O
j
i is an eigenvector of eigenvalue 1 of the swap operators.
Continuous-time quantum walks evolve by an application of a Hamiltonian as defined
in Eq. (2.9). Decoherence in this scenario would give rise to an evolution which can be
described using the Lindblad semigroup master equation as
˙ ρ(t)=[
ˆ
H,ρ(t)]+
X
i
a
i
(
ˆ
L
i
ρ(t)
ˆ
L
†
i
+
ˆ
L
†
i
ˆ
L
i
ρ(t)+ρ(t)
ˆ
L
†
i
ˆ
L
i
). (6.17)
The necessary and sufficient conditions for the existence of decoherence-free subspace for
such an evolution is given in the following theorem [52].
Theorem 6.5. A subspace H
0
is a decoherence-free subspace of the above evolution if
and only if the Lindblad operators are proportional to the identity on the subspace i.e., ∀
|ji∈H
0
ˆ
L
i
|ji=d
i
|ji. (6.18)
118
An example of a decoherence model which has as its DFS the subspaceH
H
is where
the Lindblad operators are
ˆ
L
i
=κ
i
ˆ
S
ij
, (6.19)
where
ˆ
S
ij
has the same meaning as before. Indeed, as before one can derive a sufficient
condition on the Lindblad operators which is that they must lie in the algebra generated
by the group H. If this is true then the subspaceH
H
will be a DFS.
The decoherence operators (Kraus or Lindblad) need not lie in the algebra generated
by the symmetry operators as in the examples. The above examples consider a kind
of decoherence that might occur in a physical implementation of a quantum walk on
the hypercube. But in general, the decoherence operators need not lie in the algebra
generated by the symmetry operators. The decoherence operators just need to commute
with the symmetry operators inside the DFS.
6.2 Discussion
Quantum walks can have invariant subspaces if the walk has sufficient symmetries. A
walk can be endowed with enough symmetry such that there exists a suitable invariant
subspace, which can be used to design an algorithm. The quantum-walk based search
algorithms exploit such invariant subspaces. If such a walk is implemented using any
quantum system, it will have decoherence affecting it. If the decoherence has symmetries
i.e., if the decoherence operators have degeneracies, then we can have decoherence-free
subspaces. We have determined conditions which the decoherence operators must satisfy
in order to have a certain subspace of interest (from an algorithmic point of view) lies
119
in a DFS of the dynamics. We presented examples of such a decoherence for a quantum
walk on the hypercube.
In general, it is difficult to know the symmetries of the decoherence operators without
looking at a specific implementation of quantum walks. Unfortunately, there are not
many implementations of quantum walks and the few that exist are all for the walk on a
line. Here we present the general conditions that the decoherence must satisfy in order
to have the invariant subspace H
H
as a DFS. For a specific implementation, one can
check if the decoherence satisfies these conditions to know if that scheme can successfully
implement a quantum walk algorithm which exploits invariant subspaces by preserving
these subspaces even in the presence of decoherence.
120
References
[1] Y. Aharanov, L. Davidovich, N. Zagury, Phys. Rev. A. 48(2), 1687-1690 (1992).
[2] D. Aharanov, A. Ambainis. J. Kempe and U. Vazirani, in Proc. 33rd Annual ACM
Symposium on Theory of Computing (STOC 2001), 50–59 (Assoc. for Comp. Ma-
chinery, New York, 2001).
[3] G. Alagi´ c and A. Russell, Decoherence in Quantum Walks on the Hypercube, Phys.
Rev. A 72, 062304 (2005).
[4] A. Ambainis, Proceedings of FOCS’04, pages 22-31.
[5] A. Ambainis, Quantum walks and their algorithmic applications, Intl. J. Quantum
Information 1, 507 (2003).
[6] A. Ambainis, SIAM J. comput. 34, 1324–1330 (2005).
[7] A. Ambainis, J. Kempe and A. Rivosh, Coins make quantum walks faster, in Proc.
16th ACM-SIAM Symposium on Discrete Algorithms (SODA 2005), 1099–1108
(SIAM, 2005). quant-ph/0402107.
[8] A. Ambainis, “Quantum search algorithms”, quant-ph/050412.
[9] D. Bacon, Decoherence, control and symmetry in quantum computers, PhD thesis.
[10] E. Bach, S. Coppersmith, M. Goldschen, R. Joynt and J. Watrous, One-dimensional
quantum walks with absorbing boundaries, JournalofComputerandSystemSciences
69, 562-592 (2004).
[11] P. Benioff, The computer as a physical system: A macroscopic quantum mechanical
Hamiltonian model of computers as represented by Turing machines, J. Stat. Phys.
22(5), pp. 563-591(1980).
[12] C. H. Bennett and G. Brassard, Quantum cryptography: Public key distribution
and coin tossing, in Proceedings of IEEE International Conference on Computers,
SystemsandSignalProcessing,pp.175-179,IEEE,NewYork1984.Bangalore,India,
December 1984.
[13] C. H. Bennett, G. Brassard, C. Crepeau, R. Jozsa, A. Peres and W. Wootters,
Teleporting an unknown quantum state via dual classical and EPR channels, Phys.
Rev. Lett. 70, pp. 1895-1899, 1993.
121
[14] T.A. Brun, H.A. Carteret and A. Ambainis, The quantum to classical transition for
random walks, Phys. Rev. Lett. 91, 130602 (2003).
[15] T.A. Brun, H.A. Carteret and A. Ambainis, Quantum random walks with decoherent
coins, Phys. Rev. A 67, 032304 (2003).
[16] T.A. Brun, H.A. Carteret and A. Ambainis, Quantum walks driven by many coins,
Phys. Rev. A 67, 052317 (2003).
[17] H. Buhrman and R. Spalek, Proc. SODA, 2006.
[18] A.M. Childs, E. Farhi and S. Gutmann, Quantum Information Processing 1, 35
(2002).
[19] A.M. Childs, R. Cleve, E. Deotto, E. Farhi , S. Gutmann and D.A. Spielman, Ex-
ponential algorithmic speedup by quantum walk , in Proc. 35th ACM Symposium on
Theory ofComputing (STOC 2003), 59–68(Assoc.forComp.Machinery, NewYork,
2003).
[20] A. Childs, J. Goldstone, Spatial search by quantum walk, Phys. Rev. A 70, 022314
(2004).
[21] M. Christandl, N. Datta, A. Ekert and A. Landahl, Phys. Rev. Lett. 92, 187902
(2004).
[22] M. Christandl, N. Datta, T. Dorlas, A. Ekert, A. Kay and A. Landahl, Phys. Rev.
A 71, 032312 (2005).
[23] D. Deutsch, Quantum theory, the Church-Turing principle, and the universal quan-
tum computer, Proc. Roy. Soc. London A 400 (1985), pp. 97-117.
[24] D.Deutsch,Quantumcomputationalnetworks,Proc.Roy.Soc.LondonA425(1989).
[25] D. DeutschandR. Jozsa, Rapid solution of problems by quantum computation, Proc.
Roy. Soc. London A 439 (1992), pp. 553-558.
[26] W. D¨ ur, R. Raussendorf, V. M Kendon and H.-J. Briegel, Phys. Rev. A 66, 052319
(2002).
[27] E. Farhi and S. Gutmann, Phys. Rev. A 58, 915 (1998).
[28] D. Feder, Phys. Rev. Lett. 97, 180502 (2006).
[29] L. Fedichkin, D. Solenov and C. Tamon, Quantum Inf. Comp. 6 (3), pp. 263-273,
(2006).
[30] R.P.Feynman, R.B.LeightonandM.Sands, Feynman lectures on physics, Addison
Wesley, 1964.
[31] R. P. Feynman, Simulating physics with computers, Int. J. Theor. Phys. 21 (1982).
[32] R. P. Feynman, Quantum mechanical computers, Optics News11 (1985), pp. 11-20.
122
[33] H. Gerhardt and J. Watrous, Continuous-time walks on the symmetric group, in
Proc.7thInternationalWorkshoponRandomizationandApproximationTechniques
in Computer Science (RANDOM 2003), edited by S. Arora, K. Jansen, J.D.P. Rolim
and A. Sahai, 290–301 (Springer, Berlin, 2003).
[34] J.L.GrossandT.W.Tucker, TopologicalGraphTheory(WileyInter-science1987).
[35] L. Grover, A fast quantum mechanical algorithm for database search, Proc. 28th
STOC, pp. 212-219, Philadelphia, Pennsylvania, 1996.
[36] G. D. James and A. Kerber, The representation theory of the symmetric group,
Encyclopedia of Mathematics 16, Addison-Wesley, 1981.
[37] J. Kempe, Quantum random walks-an introductory overview, Contemp. Phys. 44,
307–327 (2003).
[38] J. Kempe, Quantum random walks hit exponentially faster, in Proc. of 7th Intern.
Workshop on Randomization and Approximation Techniques in Comp. Sc. (RAN-
DOM 2003), edited by S. Arora, K. Jansen, J.D.P. Rolim and A. Sahai, 354–369
(Springer, Berlin, 2003).
[39] J.Kempe,D.Bacon,D.A.LidarandK.B.Whaley,Phys.Rev.A63,042307(2000).
[40] J. P. Keating, N. Linden, J. C. F. Matthews, and A. Winter, e-print quant-
ph/0606205.
[41] V. Kendon and B. Tregenna, Phys. Rev. A 67, 042315 (2003).
[42] V. Kendon, Quantum walks on general gaphs, Int. J. Quantum Inf.4(5), 791 (2006).
[43] V. Kendon and B. Sanders, Phys. Rev. A 71, 022307 (2004).
[44] V. Kendon, e-print quant-ph/0606016.
[45] N. Konno, Quantum Information Processing 1 (5), pp. 345-354 (2002).
[46] N. Konno, T. Namiki and T. Soshi, quant-ph/0205065.
[47] H. Krovi and T. A. Brun, Hitting times for quantum walks on the hypercube, Phys.
Rev. A 73, 032341 (2006).
[48] H. Krovi and T. A. Brun, Quantum walks with infinite hitting times, Phys. Rev. A
74, 042334 (2006).
[49] H. Krovi and T. A. Brun, Quantum walks on quotient graphs, Phys. Rev. A 75,
062332 (2007).
[50] H. Krovi and T. A. Brun, Quantum walks in decoherence-free subspaces, in prepara-
tion.
[51] J. Kempe, D. Bacon, D.A. Lidar, and K.B. Whaley, Theory of Decoherence-Free
Universal Fault-Tolerant Quantum Computation, Phys. Rev. A 63, 042307 (2001).
123
[52] D.A. Lidar and K.B. Whaley, Decoherence-Free Subspaces and Subsystems in Irre-
versibleQuantumDynamics,F.BenattiandR.Floreanini(Eds.),p.83-120(Springer
Lecture Notes in Physics, Vol. 622, Berlin, 2003).
[53] C. Lomont, The Hidden Subgroup Problem - Review and Open Problems, quant-
ph/0411037.
[54] F. Maginez, M. Santha, M. Szegedy, Proc. SODA, 2005.
[55] F. Maginez and A. Nayak, Proc. ICALP, 2005.
[56] Yu. Manin, Computable and uncomputable, Sovetskoye Radio (1980).
[57] C. Moore and A. Russell, in Proc. 6th Intl. Workshop on Randomization and Ap-
proximation Techniques in Computer Science (RANDOM 2002), edited by J. D. P.
Rolim and S. Vadhan, 164–178 (Springer, Berlin, 2002).
[58] A. Montanaro, Quantum walks on directed graphs, quant-ph/0504116
[59] R.MotwaniandP.Raghavan,RandomizedAlgorithms(CambridgeUniversityPress,
Cambridge, 1995).
[60] A.NayakandA.Vishwanath,Quantumwalkontheline,DIMACSTechnicalReport
2000-43, e-print quant-ph/0010117.
[61] M.NeilsenandI.Chuang, QuantumComputationandQuantumInformation(Cam-
bridge University Press, Cambridge, 2000).
[62] J. Preskill and P. Shor, Phys.Rev.Lett. 85 (2000) pp. 441-444.
[63] P. Richter, New J. Phys. 9 (2007) pp. 72.
[64] P. Richter, quant-ph/0609204.
[65] D. Rolf, 3-SAT ∈ RTIME(1.32971
n
), Diploma thesis, Humboldt-universit¨ at zu
Berlin, 2003.
[66] W.E. Roth, Bull. Amer. Math. Soc. 40, 461 (1934).
[67] J. Rotman, Introduction to the theory of groups, Springer-Verlag, New York 1995.
[68] M. R¨ otteler and T. Beth, Polynomial-time solution to the hidden subgroup problem
for a class of non-Abelian groups, quant-ph/9812070.
[69] C. A. Ryan, M. Laforest, J. C. Boileau and R. Laflamme, Phys. Rev. A 72, 062317
(2005).
[70] B. C. Sanders, S. D. Bartlett, B. Tregenna and P. L. Knight, Phys. Rev. A 67,
042305 (2003).
[71] J.-P. Serre, Linear Representations of Finite Groups (Springer-Verlag, New York,
1977).
124
[72] N. Shenvi, J. Kempe and K. Birgitta Whaley, A quantum random walk search algo-
rithm, Phys. Rev. A 67, 052307 (2003).
[73] U. Sh¨ oning, 40
th
Annual Symposium on FOCS, IEEE, Los Alamitos, CA, pp. 17-19.
[74] P.W. Shor, in Proceedings of the 35th Annual Symposium on the Theory of Com-
puter Science, edited by S. Goldwasser, 124 (IEEE Computer Society Press, Los
Alamitos, CA, 1994).
[75] D. Simon, On the power of quantum computation, in Proceedings of 35
th
Annual
Symposium on Foundations of Computer Science, pp. 116-123, IEEE Press, Los
Alamitos, CA, 1994.
[76] S. Sternberg, Group Theory and Physics (Cambridge University Press, Cambridge,
1994).
[77] F.W. Strauch, Connecting the discrete and continuous-time quantum walks, Phys.
Rev. A 74, 030301(R) (2006).
[78] B. C. Travaglione and G. J. Milburn, Phys. Rev. A 65, 032310.
[79] B.Tregenna, W.Flanagan, R.Maile, V.Kendon, Controlling discrete quantum walk:
coins and initial states, New J. Phys. 5, 83.
125
Abstract (if available)
Abstract
A discrete-time quantum walk on a graph is the repeated application of a unitary evolution operator to a Hilbert space corresponding to the graph. Hitting times for discrete quantum walks on graphs give an average time before the walk reaches an ending condition. We derive an expression for hitting time using superoperators, and numerically evaluate it for the walk on the hypercube for various coins and decoherence models. The hitting time for a classical random walk on a connected graph will always be finite. We show that, by contrast, quantum walks can have infinite hitting times for some initial states. We seek criteria to determine if a given walk on a graph will have infinite hitting times, and find a sufficient condition, which for discrete time quantum walks is that the degeneracy of the evolution operator be greater than the degree of the graph. The phenomenon of infinite hitting times is in general a consequence of the symmetry of the graph and its automorphism group.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Topics in quantum information -- Continuous quantum measurements and quantum walks
PDF
Quantum coding with entanglement
PDF
Imposing classical symmetries on quantum operators with applications to optimization
PDF
Quantum computation and optimized error correction
PDF
Quantum steganography and quantum error-correction
PDF
Dynamical error suppression for quantum information
PDF
Topics in quantum information and the theory of open quantum systems
PDF
Entanglement-assisted coding theory
PDF
Applications and error correction for adiabatic quantum optimization
PDF
Topics in modeling, analysis and simulation of near-term quantum physical systems with continuous monitoring
PDF
Open quantum systems and error correction
PDF
Quantum error correction and fault-tolerant quantum computation
PDF
Quantum computation in wireless networks
PDF
Applications of quantum error-correcting codes to quantum information processing
PDF
Theory and simulation of Hamiltonian open quantum systems
PDF
Quantum and classical steganography in optical systems
PDF
Coherence generation, incompatibility, and symmetry in quantum processes
PDF
Quantum information techniques in condensed matter: quantum equilibration, entanglement typicality, detection of topological order
PDF
Open-system modeling of quantum annealing: theory and applications
PDF
Lower overhead fault-tolerant building blocks for noisy quantum computers
Asset Metadata
Creator
Krovi, Hari
(author)
Core Title
Symmetry in quantum walks
School
Viterbi School of Engineering
Degree
Doctor of Philosophy
Degree Program
Electrical Engineering
Publication Date
07/20/2007
Defense Date
07/02/2007
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
algorithm,computing,OAI-PMH Harvest,quantum,symmetry,walk
Language
English
Advisor
Brun, Todd A. (
committee chair
), Haas, Stephan (
committee member
), Lidar, Daniel (
committee member
)
Creator Email
krovi@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-m633
Unique identifier
UC1459899
Identifier
etd-Krovi-20070720 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-518045 (legacy record id),usctheses-m633 (legacy record id)
Legacy Identifier
etd-Krovi-20070720.pdf
Dmrecord
518045
Document Type
Dissertation
Rights
Krovi, Hari
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Repository Name
Libraries, University of Southern California
Repository Location
Los Angeles, California
Repository Email
cisadmin@lib.usc.edu
Tags
algorithm
computing
quantum
symmetry
walk