THESIS

Study of Optimization Problems

arXiv:quant-ph/0205020v1 5 May 2002

by Quantum Annealing

Tadashi Kadowaki

Department of Physics, Tokyo Institute of Technology

December, 1998
Acknowledgments

I would like to express my sincerest gratitude to Professor Hidetoshi Nishi-

mori for his guidance, useful discussions and above all his encouragements.
I acknowledge useful discussions with Professor Seiji Miyashita, Dr. Nao-
michi Hatano, Dr. Yukiyasu Ozeki, Professor Kazuyuki Tanaka and Dr.
Jun-ichi Inoue. I also thank all members of the Nishimori group and the
condensed-matter theory group in the Physics Department of Tokyo Institute
of Technology for stimulating discussions.
Numerical calculations were performed on the CRAY C916/12256 in the
Computer Center, Tokyo Institute of Technology.
Most of the work on computers was performed by various free or open
source softwares on Debian GNU/Linux system. I would like to thank people
who support free or open source softwares and the community of Debian
GNU/Linux.
Finally, I thank my family for spiritual and financial supports.

i
Preface

On the occasion to submit my thesis to the preprint server, I would like to

refer to previous and recent studies which are not referred to in the original
thesis.
In the early days of quantum annealing study, Amara et al. used the
imaginary time Schrödinger equation to search the global minimum state of
the classical system[1]. Finnila et al. introduced a controlled and scheduled
quantum effect to the Monte Carlo Simulation[2].
Recently, thermal and quantum annealing in the disordered Ising magnet,
LiHo0.44 Y0.56 F4 , with transverse magnetic field were compared[3,4]. Numer-
ical studies of quantum annealing for the disordered Ising model and the
protein folding problem were performed by a few different groups[5,6].
A quantum computer algorithm is also related with quantum annealing.
The algorithm uses the adiabatic evolution of a time dependent Hamilto-
nian[7,8,9,10].

1 May 2002
TK

Bibliography

[1] P. Amara, D. Hsu and J. E. Straub, Global Energy Minimum Searches

Using an Approximate Solution of the Imaginary Time Schrödinger
Equation, J. Phys. Chem. 97 (1993) 6715.

[2] A. B. Finnila, M. A. Gomez, C. Sebenik, C. Stenson and J. D. Doll,

ii
 iii

Quantum Annealing: A New Method for Minimising Multidimensional

Functions, Chem. Phys. Lett. 219 (1994) 343, chem-ph/9404003
(1994).

[3] J. Brooke, D. Bitko, T. F. Rosenbaum and G. Aeppli, Quantum Anneal-

ing of a Disordered Magnet, Science 284 (1999) 779, cond-mat/0105238
(2001).

[4] J. Brooke, T. F. Rosenbaum and G. Aeppli, Tunable Quantum Tunnel-

ing of Magnetic Domain Walls, Nature 413 (2001) 610, cond-mat/0202361
(2002).

[5] Y. H. Lee and B. J. Berne, Global Optimization: Quantum Thermal

Annealing with Path Integral Monte Carlo, J. Phys. Chem. A 104
(2000) 86.

[6] G. E. Santoro, R. Martoňák, E. Tosatti and R. Car, Theory of Quantum

Annealing of an Ising Spin Glass, Science 295 (2002) 2427.

[7] E. Farhi, J. Goldstone, S. Gutmann and M. Sipser, Quantum Compu-

tation by Adiabatic Evolution, quant-ph/0001106 (2000).

[8] E. Farhi, J. Goldstone, S. Gutmann, J. Lapan, A. Lundgren and D.

Preda, A Quantum Adiabatic Evolution Algorithm Applied to Random
Instance of an NP-Complete Problem, science 292 (2001) 472, quant-
ph/0104129 (2001).

[9] A. M. Childs, E. Farhi and J. Preskill, Robustness of adiabatic quan-

tum computation, Rhys. Rev. A 65 (2002) 012322, quant-ph/0108048
(2001).

[10] E. Farhi, J. Goldstone and S. Gutmann, Quantum Adiabatic Evolution

Algorithms versus Simulated Annealing, quant-ph/0201031 (2002).
Abstract

We introduce quantum fluctuations into the simulated annealing process of

optimization problems, aiming at faster convergence to the optimal state.
Quantum fluctuations cause transitions between states and thus play the
same role as thermal fluctuations in the conventional approach. The idea
is tested by the two models, the transverse Ising model and the traveling
salesman problem.
Adding the transverse field to the Ising model is a simple way to introduce
quantum fluctuations. The strength of the transverse field is controlled as
a function of time similarly to the temperature in the conventional method.
The goal is to find the ground state of the diagonal part of the Hamiltonian
with high accuracy as quickly as possible. We also consider the traveling
salesman problem. This model can be described by the Ising spin, so that
we also apply the same technique to the transverse Ising model.
We solve the time-dependent Schrödinger equation numerically for small
size systems with various types of exchange interactions of the Ising model.
Comparison with the results of the corresponding classical (thermal) method
reveals that the quantum method leads to the ground state with much larger
probability in almost all cases if we use the same annealing schedule of the
control parameters. We check the case of large-size systems by using the
quantum Monte Carlo method. The simulation supports the results of the
small-size systems, while the dynamics of the Schrödinger equation and the
quantum Monte Carlo method are not the same. We find that the simu-
lated annealing by quantum fluctuations has a better performance than the
conventional method for the ground state search of the Ising-spin systems.
The calculation of the traveling salesman problem is also performed as

iv
 v

an application to the general optimization problems. We obtain the same

feature of the fast convergence to the optimal state as the transverse Ising
model by using the quantum fluctuations.
We also find that the relaxation time is quite short for quantum systems
by numerical simulations. We consider this is one of the reasons why the an-
nealing in quantum systems have a better performance of finding the optimal
state in comparison with classical systems.
 vi

List of Papers

1. T. Kadowaki and H. Nishimori, “Quantum annealing in the transverse

Ising model”, Phys. Rev. E 58 (1998) 5355.

2. T. Kadowaki and H. Nishimori, “Monte Carlo study of quantum an-

nealing”, in preparation.

List of Papers Added for Reference

1. T. Kadowaki, Y. Nonomura and H. Nishimori, “Exact Ground-State

Energy of the Ising Spin Glass on Strips”, J. Phys. Soc Jpn. 65 (1996)
1609.

2. M. Yamana, H. Nishimori, T. Kadowaki and D. Sherrington, “High-

Temperature Dynamics of Spin Glasses”, J. Phys. Soc Jpn. 66 (1997)
1962.
Contents

Preface ii

Abstract iv

1 Introduction 1
1.1 The Combinatorial Optimization Problem . . . . . . . . . . . 1
1.2 The NP -hard and The NP -complete Problems . . . . . . . . 2
1.3 Spin Glass Models and Optimization . . . . . . . . . . . . . . 7
1.4 Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Quantum Annealing . . . . . . . . . . . . . . . . . . . . . . . 13

2 Analysis by Differential Equations 18

2.1 The Transverse Ising Model . . . . . . . . . . . . . . . . . . . 19
2.2 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Ferromagnetic Model . . . . . . . . . . . . . . . . . . . 21
2.2.2 Frustrated Model . . . . . . . . . . . . . . . . . . . . . 22
2.2.3 Random Interaction Model . . . . . . . . . . . . . . . . 26
2.3 Solution of the single-spin problem . . . . . . . . . . . . . . . 27
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Monte Carlo Analysis 33

3.1 Monte Carlo Method . . . . . . . . . . . . . . . . . . . . . . . 34
3.2 Quantum Monte Carlo Method . . . . . . . . . . . . . . . . . 35
3.3 Path-Integral Monte Carlo . . . . . . . . . . . . . . . . . . . . 37
3.4 Applying Monte Carlo to Annealing Process . . . . . . . . . . 40
3.5 Results of the Monte Carlo simulation . . . . . . . . . . . . . 47

vii
CONTENTS viii

3.5.1 Dependence on the temperature . . . . . . . . . . . . . 48

3.5.2 Dependence on the number of Trotter slices . . . . . . 49
3.5.3 Results of large size system for the EA model . . . . . 50
3.5.4Difference between the annealing and the quenching
processes . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4 Application to TSP 61
4.1 The Traveling Salesman Problem . . . . . . . . . . . . . . . . 61
4.2 Simulated Annealing on TSP . . . . . . . . . . . . . . . . . . 62
4.3 Quantum Annealing on TSP . . . . . . . . . . . . . . . . . . . 63
4.4 Results of Monte Carlo Simulations . . . . . . . . . . . . . . . 67
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

5 Summary 78

A Single-spin problem 82

Appendix 82
A.1 Case of Γ(t) = −ct (Landau-Zener model) . . . . . . . . . . . 82
A.2 Case of Γ(t) = c/t . . . . . . . . . . . . . . . . . . . . . . . . . 83
√
A.3 Case of Γ(t) = c/ t . . . . . . . . . . . . . . . . . . . . . . . . 85
A.4 Final value dependence on initial condition . . . . . . . . . . . 86

Bibliography 88
Chapter 1

Introduction

1.1 The Combinatorial Optimization Prob-

lem
The combinatorial optimization problem is to find a minimum or maximum
value of a function of very many independent variables, where the variables
take discrete values. This function, usually called the cost function or objec-
tive function, represents a quantitative measure of the “goodness” of some
complex systems. A famous example is the traveling salesman problem [1, 2]
to find the shortest or lowest-cost route to visit given cities. The techniques
to find the shortest route have been studied for a long time, because such a
problem actually happens in various situations, and people have to obtain the
solution. One of the most important application of the optimization prob-
lems today is the circuit design of the computer systems. The cost function
is the wiring length in this case. Long wiring makes delays in electric circuit,
and the speed of computer slows down.
The algorithm to find the shortest tour of the traveling salesman problem
or the wiring length of the circuit design in polynomial time of N, where N
is the size of the problem, is not found yet. When N increases, the complex-
ity of the calculation increases exponentially and goes over the limit of the
computational power. This kind of problems are called the “Nondeterminis-
tic Polynomial-time solvable hard or complete (NP -hard or NP -complete)”
problems. (Details of the definitions are explained in the next section.) Any

1
CHAPTER 1. INTRODUCTION 2

algorithm of solving an NP -hard or NP -complete problem in polynomial

time is not found so far and people consider that there does not exist such
an algorithm. Therefore we consider approximate algorithms by which we
can obtain an approximate solution or an exact solution with some probabil-
ity in polynomial time. In this thesis we focus on how to solve the NP -hard
or NP -complete problems approximately as fast as possible.

1.2 The N P -hard and The N P -complete Prob-

lems
In this section we explain several levels of difficulty of optimization problems.
The NP -hard and NP -complete problems belong to the class quite hard to
solve. The difference between the classes NP -hard and NP -complete is not
in the degree of difficulty but the way to answer the questions; “yes or no”
for NP -complete and in general statements for NP -hard.
First, let us consider the class NP before the definition of the classes
NP -hard and NP -complete. The class NP is a subclass of the solvable class
of decision problems whose answer is expressed by either “yes” or “no”. For
instance, the Hamilton circuit problem is the problem to check the existence
of the Hamilton circuit which is the closed loop constructed by edges and all
the vertexes in a given graph. An example of the Hamilton circuit problem
and its Hamilton circuit are shown in Fig. 1.1.
Roughly speaking, the NP problem is the problem which can be verified
in polynomial time by the people who know the evidence to prove the exis-
tence of the solution. For instance, one can answer “yes” for the Hamilton
circuit problem shown in Fig. 1.1, when the Hamilton circuit (the evidence
in this case) is given as a bold line in Fig. 1.1. The verification time of the
evidence, checking the closed loop which contains all the vertexes, is pro-
portional to N. To be precise, the NP problem is defined as a problem
which can be solved by the nondeterministic Turing machine in polynomial
time of N. However, the definition of the nondeterministic Turing machine
is complicated, and the reader who is not interested in the detail can skip
the following two paragraphs.
CHAPTER 1. INTRODUCTION 3

Figure 1.1: The Hamilton graph is the graph which has a Hamilton circuit.
The bold line is the Hamilton circuit.

Before presenting the definition of the nondeterministic Turing machine,

we explain the definition of the usual (deterministic) Turing machine. The
nondeterministic Turing machine is the extended model of the Turing ma-
chine. The definition of the Turing machine is constructed with

• the tape which includes alphabets (as commands and data).

• the state of the machine.

• the transition function of the state by using the alphabet read from the
tape.

The Turing machine is a simple model of actual computers. For the de-
terministic Turing machine, the transition function is single-valued. Turing
machine reads an alphabet from the tape and modifies its configuration (the
state and alphabets on the tape) from the previous configuration according
to the given alphabet. A unique modified state is given by this transition.
CHAPTER 1. INTRODUCTION 4

On the other hand, the transition function of the nondeterministic Tur-

ing machine is multi-valued. The feature of the nondeterministic Turing
machine quite differs from conventional computers. The configuration of the
nondeterministic Turing machine becomes multiple configurations. This can
be understood loosely in a way that the nondeterministic Turing machine
makes replicas of possible configurations by a single step and the number
of replicas increases exponentially as a function of the step. This feature
is quite useful to solve complicated problems. For finding the ground-state
energy of Ising-spin systems, for example, we have to enumerate all possible
spin configurations, if the ground state is nontrivial. These configurations
are illustrated by the tree structure in Fig. 1.2. The direction of spins is
assigned at each branching point and the level corresponds to the location of
the spin. Each path means the configuration. The nondeterministic Turing
machine can check all paths in parallel. At the first step, the machine makes
a replica and the number of machines is two. Each machine makes a replica
at each step so that the number of machines is 2N at the Nth step. The
calculations in the same level of the tree are performed at a time. By this
procedure, we can enumerate 2N configurations in N steps.
All NP problems can be verified in polynomial time even on a determin-
istic Turing machine, but the difficulty of solving the problems is not the
same. The problems are divided into some subclasses — the subclasses P for
easy problems and NP -complete for difficult problems. The problem which
has the algorithm to be solved in polynomial time by the deterministic Tur-
ing machine belongs to the class of Polynomial-time solvable (P ). Obviously
NP contains P , because the nondeterministic Turing machine includes the
deterministic Turing machine in its definition.
For an example of the P problem, we explain the Hamilton circuit problem
for the graph whose maximum degree of the edge in a vertex is limited to
two belongs to P . The typical graphs are shown in Fig. 1.3. The answers
of the left and the right problems are “yes” and “no” respectively. One can
check the existence of the Hamilton circuit by the following way:
1. Start from a vertex and jump to the neighbor vertex connected by
an edge. (One can not jump to the neighbor which has already been
CHAPTER 1. INTRODUCTION 5

N
N spins 2 configurations

Figure 1.2: The path means the spin configuration. The number of the
leaves (the bottom ends of the tree) equals to the number of configurations
as 2N .

visited once.)

2. Repeat 1 until no vertex to jump is left.

3. If the last vertex is connected to the starting vertex and all the vertexes
have been visited, the answer is “yes”. The answer is “no” other wise.

This procedure needs polynomial time. If the maximum degree of the edge
in a vertex is not limited to two, this procedure does not work.
Let us consider the difficult classes NP -hard and NP -complete. If any
problems in NP can be solved by using a particular algorithm of the problem
repeatedly polynomial times at most, the problem is called “NP -hard”. In
other wards, we can solve any NP -problems by calling the subroutine for a
particular NP -hard problem (maybe once except for other NP -hard prob-
lems). The NP -hard problem is more difficult than any NP problems or as
CHAPTER 1. INTRODUCTION 6

Figure 1.3: The Hamilton circuit problems with the answer “yes” and “no”.
The left graph is the single loop (the Hamilton circuit) and right one is not
the single loop.

difficult as the other NP -hard problems in polynomial order. The algorithm

of the NP -hard problem is the most general algorithm in NP class. The
NP -hard problems by definition are not limited to belong to the class NP .
For instance, the traveling salesman problem is not an NP problem but an
NP -hard problem because the answer is not “yes or no” but the tour length
and route. The subclass “NP -complete” is the set of the NP -hard problems
which belong to the class NP . The NP -complete problems are more difficult
than any other NP problems and as difficult as the NP -hard problems. The
Hamilton circuit problem is known as an NP -complete problem [3].
The difficulty of the problems can be expressed symbolically as P ≤
(NP − P − NP -complete) ≤ NP -complete ≈ NP -hard. It is a trivial
consequence of their definition that if one of the NP -complete problems
can be solved in polynomial time, then all the problems of the NP class
are also polynomial. In this case one would have P = NP . So far no
polynomial algorithm has been found for an NP -complete problem, and the
question of whether P is equal to NP is still open, although the general
belief is that NP -complete problems are not polynomial. The most probable
situation is sketched in Fig. 1.4. The NP problems can be divided into
three subclasses, NP -complete, P and the subclass which does not belong
to the previous two subclasses. The class NP -hard is separated into two
subclasses, NP -complete and the other part by whether the problem is the
CHAPTER 1. INTRODUCTION 7

decision problem or not. Empirically the time it takes to solve an NP -hard

and an NP -complete problems tends to scale exponentially with the size N.
All the problems can be solved by the enumeration method by which all
the configurational space of the problem are enumerated and this calculation
needs the time proportional to eN .

NP
P
NP-hard
NP-complete
Figure 1.4: The class of NP problems assuming that P 6= NP . Under this
assumption it can be shown that there are NP problems that are neither
NP -complete nor P .

1.3 Spin Glass Models and Their Relation to

Optimization
Combinatorial optimization problems appear in various disciplines. In sta-
tistical mechanics, the energy is equivalent to the cost function of the op-
timization problem. Until the 1970s, almost all the systems to be studied
were homogeneous so that the energy landscape has a simple structure and
the optimal solution (the ground state) is trivial. Although such a system
has many variables, the degrees of freedom can be reduced to small numbers.
For instance, spins turn to the same direction in the ground state of the fer-
romagnetic model and the spin variables are reduced to single-spin variable.
This reduction comes from the homogeneous interactions. In the 1970s the
study of the spin glass started. The spin glass system is a non-homogeneous
system and the ground state is nontrivial. The background of the physics
CHAPTER 1. INTRODUCTION 8

of spin glasses and its relation to the combinatorial optimization problem is

explained in this section.
The study of spin glasses started in 1972 when Cannella and Mydosh [4]
investigated the AC susceptibility of a dilute magnetic alloy Au-Fe and found
out that the AC susceptibility has a nontrivial sharp cusp. In the alloy, the
interaction Jij between two spins of Fe atoms is expressed as the RKKY
interaction [5, 6, 7]
1
Jij ∝ cos(2kF rij ) , (1.1)
rij3

where rij is the distance between two spins and kF is the Fermi wave number.
From this interaction and a random distribution of Fe atoms in space, it is
possible that the interaction is a positive or a negative random variable.
Edwards and Anderson proposed the so-called Edwards-Anderson (EA)
model [8] to introduce the effect of the RKKY interaction in a diluted mag-
netic alloy. The EA model has nearest-neighbor interactions Jij . The value
of the interaction distributes as
"  2 #
 z 1/2 z J0
P (Jij ) = exp − 2 Jij − , (1.2)
2πJ 2 2J z

where z is the number of nearest neighbors. The average of the distribution

√
is J0 /z, and the standard deviation is J/ z. The Hamiltonian is given by
X
H=− Jij Si Sj . (1.3)
n.n.

The EA solution is not solved exactly. Sherrington and Kirkpatrick [9] inves-
tigated an infinite-range model, the so-called SK model. The Hamiltonian is
given by
X X
H=− Jij Si Sj − h Si , (1.4)
<ij> i

where < ij > denotes a summation over all spin pairs. This is the reason
of the name of the infinite-range model. The summation is different from
that in the EA model. If the average of distribution of interactions is equal
CHAPTER 1. INTRODUCTION 9

to zero and the external field is vanishing, a second order phase transition
at a finite temperature Tc (= J/kB ) occurs, and the spin glass phase realizes
below Tc .
The statistical property as the SK model was almost solved by further
studies [10, 11, 12, 13, 14, 15, 16, 17, 18], but it is difficult to analyze the EA
model by the same approaches. The EA model is often studied by computer
simulations, the Monte Carlo simulations or other methods. The great in-
terest of the EA model or the ±J model whose interactions are distributed
binary (at J and −J) is whether the same picture of the SK model — the
picture of the ultrametricity and the structure of the P (q) [13, 18] — exists
or not. One can check this picture from the calculation of the distribution
function P (q) where the overlap q is calculated from the pure states which are
the equilibrium states separated by high energy barriers and can be obtained
by the Monte Carlo simulation. If the simulation starts from a random con-
figuration which is equivalent to the configuration at high temperature, the
system is stuck in a metastable state and it takes long time to escape from
the metastable state because the basins of the pure states and the metastable
states are separated by high energy barriers. To avoid this difficulty of the
simulation, one can start the simulation from the ground state, which is close
to the pure state at low temperature. However, finding the ground state is
also difficult. The reason is the following: Spins on the some plaquettes in the
EA model do not satisfy all the bonds illustrated in Fig.1.5. This situation
occurs, when the product of all the bonds in the plaquette has a minus sign.
The frustrated plaquettes are located at random so that we have to consider
the global structure of the configuration which pays as small penalty of the
energy as possible. Therefore the ground state is not the trivial configura-
tion. The difficulties in studying the EA model are the slow relaxation and
the non-trivial configuration of the ground state. The latter problem is a
typical example of the combinatorial optimization problem.
The SK model is connected with the bipartition problem directly. We
divide the group into two parts. “Likes and dislikes” between the members
is expressed as the numerical value. The task is to find the best partition of
the members. The case of N = 6 is illustrated in Fig. 1.6. The members are
CHAPTER 1. INTRODUCTION 10

?
Figure 1.5: The frustrated plaquette. The solid and the zigzag lines are the
ferromagnetic and antiferromagnetic bonds.

divided into open and closed circles, and the black and the gray lines mean
“likes” and “dislikes” between the members. Assigning +1 to the members
in one part and −1 to the members in the other part, we can regard that the
members are expressed by Ising spins {σi }. The cost function of the “likes
σ σ +1
and dislikes” between two members i and j can be obtained as Jij i 2j . If
the two members i and j are in the same part, “likes and dislikes” is counted
as Jij . If the two members are in the different parts, the value is zero. We
can calculate the total cost of “likes and dislikes” by summing up all the
pairs as:
X σi σj + 1
Total Cost = Jij
2
(ij)
1X
= Jij σi σj + const. (1.5)
2
(ij)

This cost function is exactly the same as the Hamiltonian of the SK model.
The bipartition problem and the finding the ground state of the SK model
are known as the NP -hard problems [19, 20].

1.4 Simulated Annealing

As shown in the previous section, large computational time is needed to solve
an NP -complete or an NP -hard problem exactly. The technique of simu-
lated annealing (SA) was first proposed by Kirkpatrick et al. [21] as a general
CHAPTER 1. INTRODUCTION 11

Figure 1.6: The bipartition problems with N = 6. The open and closed
circles are the divided members of the group. The black and the gray lines
mean “likes” and “dislikes” between the members.

method to solve optimization problems. This method is an approximate al-

gorithm to find the optimal solution in finite time, but the solution converges
to the optimal state in the infinite-time limit. The probability to find the
ground state increase with time and converges to one in the infinite-time
limit. For practical applications, it is important to find a sufficiently good
solution, where “good” means that the cost function is close to the optimal
value.
The idea is to use thermal fluctuations to allow the system to escape
from local minima of the cost function so that the system reaches the global
minimum under an appropriate annealing schedule (the rate of decrease of
temperature). Let us consider a simple example of a system described by
Ising spins {σi ± 1}, binary variables. Spins interact each other as −Jσi σj
P
and the total energy (cost function) of this system is H = − ij Jσi σj , where
ij runs over all possible pairs of sites which interact with each other. If the
two spins σi , σj take the same value, the energy becomes lower. This system
is the simplest model of a ferromagnet. In the optimal configuration, all spins
CHAPTER 1. INTRODUCTION 12

align in the same direction, up or down.

Once we define the energy of the system, we can consider statistical me-
chanics of this system. The spin configuration {σi } is modified by thermal
noise. The probability with which the configuration {σi } appears is propor-
tional to the Boltzmann factor exp(−E({σi })/kB T ), where T is the tempera-
ture and kB is the Boltzmann constant. The probability has to be normalized
and the normalization factor is 1/Z. The inverse of the factor, the so-called
partition function, is expressed as:
X
Z= exp(−E({σi })/kB T ) (1.6)
{σi =±1}

The lower the temperature becomes, the larger the probability of the
spin configuration of the lowest energy becomes. The system freezes to the
ground state as the temperature goes to zero. However, the system does
not always freeze to the ground state in rapid cooling. As illustrated in
Fig. 1.7, the system may be trapped in a local minimum. Therefore, the
temperature should be cooled attentively. At high temperature, the system
changes the configuration almost randomly and searches the global structure
of the energy landscape. The system seeks the area in which the valley of
the global minimum is included when the temperature is decreased. Finally,
the system is stuck at the valley of the global minimum at zero or sufficiently
low temperature (lower than the energy gap of the ground state and the first
excited state). The idea of this cooling is illustrated in Fig. 1.8
By the way, how can we cool the system slowly enough? Geman and
Geman proved a theorem on the annealing schedule for combinatorial op-
timization problems [22]. They showed that any system reaches the global
minimum of the cost function asymptotically if the temperature is decreased
as T = c/ log t or slower, where c is a constant determined by the system
size and other structures of the cost function. When the product of the tem-
perature and the Boltzmann constant kB T become less than the energy gap
between the ground state and the first excited state, the system jumps to
the excited state rarely by thermal fluctuation. Writing such a temperature
as T0 , we obtain the time to find the ground state as t ∼ exp(c/kB T0 ). This
CHAPTER 1. INTRODUCTION 13

Energy

Configuration
Figure 1.7: Cooling from high temperature to low temperature. If the
cooling rate is too fast, the system may be trapped in a local minimum.

bound on the annealing schedule may be the optimal one under generic con-
ditions although faster decrease of the temperature often gives satisfactory
results in practical applications for many systems.

1.5 Quantum Annealing

In SA thermal fluctuations are introduced in the optimization problem so
that transitions between states take place in the process of search for the
global minimum among many states. Thus there seems to be no reasons
to avoid use of other mechanisms for state transitions if these mechanisms
may lead to better convergence properties. One such possibility is the gen-
eralized transition probability of Tsallis [23], which is a generalization of the
conventional Boltzmann-type transition probability appearing in the master
equation and thus used in Monte Carlo simulations. In this method, the
system converges to the optimal state under power-low decrease of the tem-
CHAPTER 1. INTRODUCTION 14

Energy

Configuration
Figure 1.8: The system seek the various levels of the energy-landscape
structure corresponding to the temperature. The system searches the global
(local) structure at high (low) temperatures.

perature [23, 24, 25, 26]. This schedule is faster than the log schedule of the
conventional SA. However, the generalized transition probability does not
satisfy the condition of the detailed balance, the sufficient condition for con-
vergence to the equilibrium state, and the physical meanings of the system
is not trivial.
We seek another possibility of making use of quantum tunneling processes
for state transitions, which we call quantum annealing (QA). In particular
we would like to learn how effectively quantum tunneling processes possibly
lead to the global minimum in comparison to temperature-driven processes
CHAPTER 1. INTRODUCTION 15

used in the conventional method of SA.

In the virtual absence of previous studies along such a line of consid-
eration, it seems better to focus our attention on a specific model system,
rather than to develop a general argument, to gain an insight into the role
of quantum fluctuations in the situation of optimization problem. Quantum
effects have been found to play a very similar role to thermal fluctuations in
the Hopfield model and the image restoration problem in a transverse field
in thermal equilibrium [27, 28]. This observation motivates us to investigate
dynamical properties of the Ising model under quantum fluctuations in the
form of a transverse field. We therefore start the study of the QA by the
transverse Ising model with a variety of exchange interactions. The trans-
verse field controls the rate of transition between states and thus plays the
same role as the temperature does in SA. We assume that the system has no
thermal fluctuations in the QA context and the term “ground state” refers
to the lowest-energy state of the Hamiltonian without the transverse field
term.
Static properties of the transverse Ising model have been investigated
quite extensively for many years [29, 30, 31]. There have however been very
few studies on the dynamical behavior of the Ising model with a transverse
field. We refer to the work by Sato et al. who carried out quantum Monte
Carlo simulations of the two dimensional EA model in an infinitesimal trans-
verse field, showing a reasonably fast approach to the ground state [32]. We
present a new point of view by comparing the efficiency of QA directly with
that of classical SA in reaching the ground state.
In the next chapter, we consider the time-dependent Schrödinger equa-
tion for QA and the classical master equation for SA numerically for small-
size systems with the same exchange interactions under the same annealing
schedules. We explain the model and define the measure of closeness of the
system in QA to the desired ground state. This measure is compared with
the corresponding classical probability that the system is in the ground state.
Calculations of probabilities that the system is in the ground state at each
time for both classical and quantum cases give important implications on
the relative efficiency of the two approaches. The numerical results for QA
CHAPTER 1. INTRODUCTION 16

and SA suggest that QA generally gives a larger probability to find to the

ground state than SA under the same conditions on the annealing schedule
and interactions. We also calculate the analytical solutions for the one-spin
case which turns out to be quite non-trivial. Explicit solutions yield very
useful information to clarify several subtle aspects of the problem.
In Chapter 3, we extend our method by using the Monte Carlo method
to calculate large-size problems. We adopt the master equation as the time-
evolution equation of SA in contrast to the Schrödinger equation of QA. The
original definition of SA is expressed in terms of the Monte Carlo method,
which is equivalent to the master equation. It is natural to consider the
applicability of Monte Carlo method to QA. The time-evolution of the wave
function is expressed as:
Rt
|Ψ(t)i = Te−i 0 dtH(t)
|Ψ(0)i, (1.7)

where the symbol T is the time ordering operator. There are two approaches
to take into account the quantum effects or the time evolution gradually. The
first one is to divide the Hamiltonian into some parts which are easy to diag-
onalize (for instance, into three terms which include σ x , σ y , σ z , respectively)
as

eH = lim (eH0 /m eH1 /m · · · )m . (1.8)

m→∞

This division is the Suzuki-Trotter decomposition. By inserting the complete

P
set i |iihi|(= 1), the system is mapped to a (d + 1)-dimensional classical
system. Once the system is expressed as a classical system, we can perform
the Monte Carlo method, the so-called quantum Monte Carlo method. The
second one is the division of the short-time evolution as
Rt
e−i 0 dtH(t)
= lim e−i∆tH(t) · · · e−i∆tH(0) , (1.9)
∆t→0

the path integral Monte Carlo method. The dynamics of the two approaches
are not exactly the same as the dynamics of the Schrödinger equation. The
performances of the two approaches are the same or better in comparison with
the dynamics of the Schrödinger equation in some sense (a detailed discussion
CHAPTER 1. INTRODUCTION 17

is shown later). We finally adopt the quantum Monte Carlo method for the
calculation of QA, because this method is similar to the original SA and we
can find the optimal state precisely by the time evolution. The results of
the simulations show that QA performs in finding the optimal state better
than SA in spite of the disadvantage in the calculation of the quantum effect
(the calculation on (d + 1)-dimensional systems is larger than d-dimensional
system). We also calculate the difference between the quenched system and
annealed system. The calculation of the quenched system is almost the same
as the study by Sato et al. [32]. In general, quenched systems tend to be
stuck in local minima. We find that the classical quenched system is stuck
more often than the quantum quenched system. We consider this is one of
the reasons why QA works better than SA.
So far, the systems are physical models. The problems of the ground
state finding of physical models occupy the small area of the combinatorial
problems and we have to check applicability to the general problems. One of
the well-known problems is the TSP. We apply QA to the traveling salesman
problem (TSP) in Chapter 4. TSP can be mapped to the Ising spin system
so that we can use the same framework of QA. The modification of the
configuration of TSP in SA is not the same as the modification of the usual
one-spin dynamics of SA. Four spins are changed at a time in TSP. The
quantum-effect term of the Hamiltonian (the cost function) has to be changed
to the product of the four spin operators to fit the dynamics of SA. However,
we adopt the same quantum-effect term, the transverse field, because of the
difficulty of the calculation. In spite of this, QA also works and improves
in comparison to SA. This fact suggests that QA can be applied to various
problems in which we can not define the quantum effect exactly.
Chapter 2

Analysis of Small-size Systems

by Differential Equations

In this chapter, we introduce the quantum annealing (QA) for small-size sys-
tems by using the Schrödinger equation and compare the results of QA and
the simulated annealing (SA). The models (problems) we deal with here are
the models in statistical physics, precisely the Ising model. The transverse
field is introduced as the quantum effect and scheduled as a decreasing func-
tion of the time. This field flips single spin at a time so that the transition
is quite similar to the usual single-spin-flip dynamics of SA.
The analysis is performed numerically and analytically for SA and QA.
The system size of the numerical calculation is up to eight. The calculation
limit of the size is around fourteen by the middle-class supercomputer today.
(for a fourteen-spin system, calculation needs the storage about 2G Byte!)
The models we use as test-bed are ferromagnetic, frustrated and random-
interaction models. We also study for the single-spin quantum system with
longitudinal and transverse field. The exact solutions are obtained for some
schedules. The single-spin system is too small to discuss the property of QA,
but the result provides some information for the results of the numerical
calculations.

18
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 19

2.1 The Transverse Ising Model

Let us consider the following Ising model with longitudinal and transverse
fields:
X X X
H(t) = − Jij σiz σjz − h σiz − Γ(t) σix (2.1)
ij i i
X
≡ H0 − Γ(t) σix , (2.2)
i
where the type of interactions will be specified later. The term of longitu-
dinal field was introduced to remove the trivial degeneracy in the exchange
interaction term coming from the overall up-down symmetry that effectively
reduces the available phase space by half. The Γ(t)-term causes quantum
tunneling between various classical states (the eigenstates of the classical
part H0 ). By decreasing the amplitude Γ(t) of the transverse field from a
very large value to zero, we hopefully drive the system into the optimal state,
the ground state of H0 .
The natural dynamics of the present system is provided by the Schrödinger
equation
∂ |ψ(t)i
i = H(t) |ψ(t)i . (2.3)
∂t
We solve this time-dependent Schrödinger equation numerically for small-size
systems. The representation to diagonalize H0 (the z-representation) will be
used hereafter.
The corresponding classical SA process is described by the master equa-
tion
dPi (t) X
= Lij Pj (t), (2.4)
dt j

where Pi (t) represents the probability that the system is in the ith state. We
consider single-spin flip processes with the transition matrix elements given
as

 exp(−Ei /T (t))


 exp(−E /T (t)) + exp(−E /T (t)) (single-spin difference)
Lij = P i j
. (2.5)

− k6=i Lki (i = j)

0 (otherwise)
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 20

In SA, the temperature T (t) is first set to a very large value and then
is gradually decreased to zero. The corresponding process in QA should be
to change Γ(t) from a very large value to zero. The reason is that the high-
temperature state in SA is a mixture of all possible states with almost equal
probabilities, and the corresponding state in QA is the linear combination of
all states with equal amplitude in the z-representation, which is the lowest
eigenstate of the Hamiltonian (2.1) for very large Γ. The low-temperature
state after a successful SA is the ground state of H0 , which should also be the
eigenstate of H(t) as Γ(t) is reduced to zero sufficiently slowly in QA. Another
justification of identification of Γ and T comes from the fact that the T = 0
phase diagram of the Hopfield model in a transverse field has almost the
same structure as the equilibrium phase diagram of the conventional Hopfield
model at finite temperature if we identify the temperature axis of the latter
phase diagram with the Γ axis in the former [27]. We therefore change Γ(t) in
QA and T (t) in SA from infinity to zero with the same functional forms Γ(t) =
√
T (t) = c/t, c/ t, c/ log(t + 1) (t : 0 → ∞) or −ct (t : −∞ → 0). The reason
for choosing these functional forms are that either they allow for analytical
solutions in the single-spin case as shown in Sec. 2.3 or for comparison with
the Geman-Geman bound mentioned in the previous chapter.
To compare the performance of the two methods QA and SA, we calculate
the probabilities PQA (t) = |hg|ψ(t)i|2 for QA and PSA (t) = Pg (t) for SA,
where Pg (t) is the probability to find the system in the ground state at time
t in SA and |gi is the ground-state wave function of H0 . Note that we treat
only small-size systems (the number of spins N = 8) and thus the ground
state can be picked out explicitly. In the ideal situation PQA (t) and PSA (t)
will be very small initially and increase towards 1 as t → ∞.
st st
It is useful to introduce another set of quantities PSA (T ) and PQA (Γ). The
former is the Boltzmann factor of the ground state of H0 at temperature T
while the latter is defined as |hg|ψΓi|2 , where the wave function ψΓ is the
lowest-energy stationary state of the full Hamiltonian (2.1) for a given fixed
value of Γ. In the quasi-static limit, the system follows equilibrium in SA
st st
and thus PSA (t) = PSA (T (t)). Correspondingly for QA, PQA (t) = PQA (Γ(t))
when Γ(t) changes sufficiently slowly. Thus the differences between both
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 21

sides of these two equations give measures how closely the system follows
quasi-static states during dynamical process of annealing.

2.2 Numerical results

We now present numerical results on PSA and PQA for various types of ex-
change interactions and transverse fields. All calculations were performed
with a constant longitudinal field h = 0.1 to remove trivial degeneracy.

2.2.1 Ferromagnetic Model

Let us first discuss the ferromagnetic Ising model with J = const for all pairs
of spins. Figure 2.1 shows the overlaps for the case of Γ(t) = T (t) = 3/ log(t+
1). It is seen that both QA and SA follow stationary (equilibrium) states
during dynamical processes rather accurately. In SA the theorem of Geman
and Geman [22] guarantees that the annealing schedule T (t) = c/ log(1 + t)
assures convergence to the ground state (PSA → 1 in our notation) if c
is adjusted appropriately. Our choice c = 3 is somewhat arbitrary but the
tendency is clear for PSA → 1 as t → ∞, which is also clear from approximate
st
satisfaction of the quasi-equilibrium condition PSA (t) = PSA (T (t)). Although
there are no mathematically rigorous arguments for QA corresponding to
the Geman-Geman bound, the numerical data indicate convergence to the
ground state under the annealing schedule Γ(t) = 3/ log(t + 1) at least for
the ferromagnetic system. It should be remembered that the unit of time
is arbitrary since we have set ~ = 1 in the Schrödinger equation (2.3) and
the unit of time τ = 1 in the master equation (2.4). Thus the fact that the
curves for QA in Fig. 2.1 lie below those for SA at any given time does not
have any positive significance.
If we decrease the transverse field and the temperature faster, Γ(t) =
√
T (t) = 3/ t, there appears a qualitative difference between QA and SA as
shown in Fig. 2.2. The quantum method clearly gives better convergence to
the ground state while the classical counterpart gets stuck in a local minimum
with a non-negligible probability. To see the rate of approach of PQA to 1,
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 22

1
PSA
0.8 st
PSA
PQA
0.6
st
P(t)

PQA
0.4

0.2

0
1 10 100 1000
t
st
Figure 2.1: Time dependence of the overlaps PSA (t), PQA (t), PSA (T (t)) and
st
PQA (Γ(t)) of the ferromagnetic model with Γ(t) = T (t) = 3/ log(t + 1).

we have plotted 1 − PQA in a log-log scale in Fig. 2.3. It is seen that 1 − PQA
behaves as const/t in the time region between 100 and 1000.
By a still faster annealing schedule Γ(t) = T (t) = 3/t, the system becomes
trapped in intermediate states both in QA and SA as seen in Fig. 2.4.

2.2.2 Frustrated Model

We next analyze the interesting case of a frustrated system shown in Fig. 2.5.
The full lines indicate ferromagnetic interactions while the broken line is for
an antiferromagnetic interaction with the same absolute value as the ferro-
magnetic ones. If the temperature is very high in the classical case, the spins
4 and 5 are changing their states very rapidly and hence the effective inter-
action between spins 3 and 6 via spins 4 and 5 will be negligibly small. Thus
the direct antiferromagnetic interactions between spins 3 and 6 is expected to
dominate the correlation of these spins, which is clearly observed in Fig. 2.6
as the negative value of the thermodynamic correlation function hσ3z σ6z ic in
the high-temperature side. At low temperatures, on the other hand, the spins
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 23

0.8

0.6
P(t)

PSA
0.4 st
PSA
PQA
0.2 st
PQA
0
1 10 100 1000
t
Figure 2.2: Time√dependence of the overlaps of the ferromagnetic model with
Γ(t) = T (t) = 3/ t.

1
1-PQA

0.1

1-PQA
1/t
0.01
1 10 100 1000
t
Figure 2.3:
√ Time dependence of 1 − PQA (t) of the ferromagnetic model with
−1
Γ(t) = 3/ t. The dotted line represents t to guide the eye.
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 24

0.8

0.6
P(t)

PSA
0.4 st
PSA
PQA
0.2
st
PQA
0
1 10 100 1000
t
Figure 2.4: Time dependence of the overlaps of the ferromagnetic model with
Γ(t) = T (t) = 3/t.

4 and 5 tend to be fixed in some definite direction and consequently the ef-
fective ferromagnetic interactions between spins 3 and 6 are roughly twice
as large as the direct antiferromagnetic interaction. This argument is justi-
fied by the positive value of the correlation function at low temperatures in
Fig. 2.6. Therefore the spins 3 and 6 must change their relative orientation
at some intermediate temperature. This means that the free-energy land-
scape goes under significant restructuring as the temperature is decreased
and therefore the annealing process should be performed with sufficient care.
If the transverse field in QA plays a similar role to the temperature in
SA, we expect similar dependence of the correlation function hσ3z σ6z iq on the
transverse field Γ. Here the expectation value is evaluated by the stationary
eigenfunction of the full Hamiltonian (2.1) with the lowest eigenvalue at
a given Γ. The broken curve in Fig. 2.6 clearly supports this idea. We
therefore expect that the frustrated system of Fig. 2.5 is a good test ground
for comparison of QA and SA in the situation with a significant change of
spin configurations in the dynamical process of annealing.
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 25

1 4 7

3 6

2 5 8

Figure 2.5: The frustrated model where the solid lines denote ferromagnetic
interactions and the broken line is for an antiferromagnetic interaction.

1
z z
0.8 <σ3 σ6 >c
<σ3zσz6 >q
0.6
<σz3σ6z >

0.4

0.2

-0.2
0 1 2 3 4 5
T, Γ
Figure 2.6: Correlation functions of spins 3 and 6 in Fig. 2.5 for the classical
and quantum cases. In the classical model (full line) the correlation is shown
as a function of temperature while the quantum case (dotted line) is regarded
as a function of the transverse field.
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 26

The results are shown in Fig. 2.7 for the annealing schedule Γ(t) = T (t) =
√
3/ t. The time scale τ is normalized as τ = tTc2 in SA and τ = tΓ2c in QA.
The values, Tc and Γc , are the points where the correlation functions vanish
in Fig. 2.6. Thus both classical and quantum correlation functions vanish at
τ = 1. The tendency is clear that QA is better suited for ground-state search
in the present system.

1
PSA
0.8 st
PSA

0.6
PQA
P(τ)

st
PQA
0.4

0.2

τ
1 10 100 1000

Figure 2.7: Time√dependence of the overlaps of the frustrated model under

Γ(t) = T (t) = 3/ t. Here the time scale τ is normalized by Γc and Tc (the
points where the correlation functions vanish in Fig. 2.6.)

2.2.3 Random Interaction Model

The third and final example is the Sherrington-Kirkpatrick (SK) model of
spin glasses [9]. Interactions exist between all pairs of spins and are chosen
from a Gaussian distribution with vanishing mean and variance 1/N (N = 8
in our case). Figure 2.8 shows a typical result on the time evolution of
√
the probabilities under the annealing schedule Γ(t) = T (t) = 3/ t. We
have checked several realizations of exchange interactions under the same
distribution function and have found that the results are qualitatively the
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 27

same. Figure 2.8 again suggests that QA is better suited than SA for the
present optimization problem.

1
PSA
0.8 st
PSA

0.6
PQA
P(t)

st
PQA
0.4

0.2

0
1 10 100 1000
t
Figure 2.8:
√ Time dependence of the overlaps for the SK model with Γ(t) =
T (t) = 3/ t.

2.3 Solution of the single-spin problem

It is possible to solve the time-dependent Schrödinger equation explicitly
when the problem involves only a single spin and the functional form of the
√
transverse field is Γ(t) = −ct, c/t or c/ t. We note that the single-spin
problem is trivial in SA because there are only two states involved (up and
down) and thus there are no local minima. This does not mean that the
same single-spin problem is also trivial in the quantum mechanical version.
In QA with a single spin, the transition between the two states is caused
by a finite transverse field. The system goes through tunneling processes to
reach the other state, and an appropriate annealing schedule is essential to
reach the ground state. On the other hand, in SA, the transition from the
higher state to the lower state takes place even at T = 0 and thus the system
always reaches the ground state.
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 28

Let us first discuss the case of Γ(t) = −ct with t changing from −∞ to
0. This is the well-known Landau-Zener model and the explicit solution of
the time-dependent Schrödinger equation is available in the literature [33,
34, 35, 36, 37]. With the notation a(t) = h+|ψ(t)i and b(t) = h−|ψ(t)i and
√
the initial condition a(−∞) = b(−∞) = 1/ 2 (the lowest eigenstate), the
solution for b(t) is found to be (see Appendix)
2  
he−πh /8c 2ct + h ih2 + 2c 3/4πi
b(t) = √ − D−λ−1 (−iz) − √ e D−λ−2 (−iz) ,
2 c h 2ch
(2.6)
where D−λ−1 , D−λ−2 represent the parabolic cylinder function (or Weber
function) and z and λ are given as
√
z = 2ce−πi/4 t, (2.7)
ih2
λ = . (2.8)
2c
The final value of b(t) at t = 0 is
√ 2 2
h π2−ih /4c e−πh /8c
b(0) = − √
2 2c
 √ 3/4πi 
1 ce (1 + ih2 /2c)
× + . (2.9)
Γ(1 + ih2 /4c) hΓ(3/2 + ih2 /4c)
The probability to find the system in the ground state at t = 0 is, when
h2 /c ≫ 1,
c2
PQA (0) = |a(0)|2 = 1 − |b(0)|2 ∼ 1 − . (2.10)
16h4
Thus the probability PQA (t) does not approach 1 for finite c.
We next present the solution for Γ(t) = c/t with t changing from 0 to ∞
√
under the initial condition a = b = 1/ 2 (see Appendix):
1
b(t) = √ eiht tic F (1 + ic, 1 + 2ic; −2iht), (2.11)
2
where F is the confluent hypergeometric function. The asymptotic form of
b(t) as t → ∞ is
√
2(2h)−ic Γ(2ic)  −iht−πc/2
b(t) ∼ × e + ceiht+πc/2 (2ht)−1 . (2.12)
Γ(ic)
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 29

The probability to find the system in the target ground state behaves asymp-
totically as

PQA (t) = |a(t)|2 (2.13)

= 1 − |b(t)|2 (2.14)
sinh(πc)
∼ 1−
sinh(2πc)
 
−πc c cos(2ht) c2 eπc
× e + + 22 (2.15)
ht 4h t
−2πc
∼ 1−e , (2.16)

the last approximation being valid for c ≫ 1 after t → ∞. The system does
not reach the ground state as t → ∞ as long as c is finite. Larger c gives
more accurate approach to the ground state, which is reasonable because it
takes longer time to reach a given value of Γ(= c/t) for larger c, implying
slower annealing.
The final example of the solvable model concerns the annealing schedule
√
Γ(t) = c/ t. The solution for b(t) is derived in Appendix under the initial
√
condition a = b = 1/ 2 as
 
1 iht 1 1
b(t) = √ e F − iγ, ; −2iht
2 2 2
 
c (3/4)πi iht 1/2 3
+√ e e (−2iht) F 1 − iγ, ; −2iht , (2.17)
h 2

where γ = c2 /2h. The large-t behavior is found to be

" ( √ (5/4)πi )
√ −πc2 /4h −iht 1 he
b(t) ∼ πe e (2ht)−iγ √ 1
+
2Γ( 2 − iγ) cΓ(−iγ)
( )#
−(1/4)πi
e c
+eiht (2ht)−1/2+iγ √ + √ , (2.18)
2Γ(iγ) 2 hΓ( 21 + iγ)

and the probability PQA (∞) for c2 /h ≫ 1 is obtained as

h2
PQA (∞) = 1 − |b(∞)|2 ∼ 1 − . (2.19)
64c4
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 30

This equation indicates that the single-spin system does not reach the ground
√
state under the present annealing schedule Γ(t) = c/ t for which the nu-
merical data in the previous section suggested an accurate approach. We
therefore conclude that the asymptotic value of PQA (t) in the previous sec-
√
tion may not be exactly equal to 1 for Γ(t) = 3/ t although it is very close
to 1.
√
The annealing schedule Γ(t) = c/ t has a feature which distinguishes
this function from the other ones −ct and c/t. As we saw in the previous
discussion, the final asymptotic value of PQA (t) is not 1 if the initial condition
√
corresponds to the ground state for Γ → ∞, a = b = 1/ 2. However, as
shown in Appendix, by an appropriate choice of the initial condition, it is
√
possible to drive the system to the ground state if Γ(t) = c/ t. This is not
possible for any initial conditions in the case of Γ(t) = −ct or c/t.

2.4 Summary
In this chapter we have started the study of quantum annealing (QA) from
the transverse Ising model obeying the time-dependent Schrödinger equa-
tion. The transverse field term was controlled so that the system approaches
the ground state. The numerical results on the probability to find the sys-
tem in the ground state were compared with the corresponding probability
derived from the numerical solution of the master equation representing the
SA processes. We have found that QA shows convergence to the optimal
(ground) state with larger probability than SA in all cases if the same an-
nealing schedule is used. The system approaches the ground state rather
√
accurately in QA for the annealing schedule Γ = c/ t but not for faster
decrease of the transverse field.
We have also solved the single-spin model exactly for QA in the cases of
√
Γ(t) = −ct, c/t and c/ t. The results showed that the ground state is not
reached perfectly for all these annealing schedules. Therefore the asymptotic
values of PQA (t) in numerical calculations are probably not exactly 1 although
they seem to be quite close to the optimal value 1.
The rate of approach to the asymptotic value close to 1, 1 − PQA (t),
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 31

was found to be proportional to 1/t in Fig. 2.3 for the ferromagnetic model.
On the other hand, the single-spin solution shows the existence of a term
√ √
proportional to 1/ t, see Eq. (2.18). Probably the coefficient of the 1/ t-
term is very small in the situation of Fig. 2.3 and the next-order contribution
dominates in the time region shown in Fig. 2.3.
A simple argument using perturbation theory yields useful information
about the asymptotic form of the probability function if we assume that the
system follows quasi-static states during dynamical processes. The probabil-
ity to find the system in the ground state is expressed using the perturbation
in terms of Γ(≪ 1) as
X 1
PQA (Γ) ∼ 1 − Γ2 (0) (0)
, (2.20)
i6=0 (E0 − Ei )2

(0)
where Ei is the energy of the ith state of the non-perturbed (classical)
(0) √
system and E0 is the ground-state energy. If we set Γ = c/ t, we have
!2
1X c
PQA (Γ) ∼ 1 − (0) (0)
. (2.21)
t i6=0 E0 − Ei

Thus the approach to the asymptotic value is proportional to 1/t as long as

the system stays in quasi-static states. The corresponding probability for SA
is
e−E0 /T X
PSA (T ) ∼ P −Ei /T ∼ 1 − e−(Ei −E0 )/T , (2.22)
i e i6=0

which shows absence of universal (1/t-like) dependence on time.

The present method of QA bears some similarity to the approach by the
generalized transition probability in which the dynamics is described by the
master equation but the transition probability has power-law dependence on
the temperature in contrast to the usual exponential form of the Boltzmann
factor [23]. This power-law dependence on the temperature allows the system
to search for a wider region in the phase space because of larger probabilities
of transition to higher-energy states at a given T (t), which may be the reason
of faster convergence to the optimal states [23, 26]. The transverse field term
CHAPTER 2. ANALYSIS BY DIFFERENTIAL EQUATIONS 32

Γ in our QA represents the rate of transition between states which is larger

than the transition rate in SA (see (2.5)) at a given small value of the control
parameter Γ(t) = T (t). This larger transition probability may lead to a
more active search in wider regions of the phase space, leading to better
convergence similarly to the case of the generalized transition probability.
We have solved the Schrödinger equation and the master equation di-
rectly by numerical methods for the purpose of comparison of QA and SA.
This method faces difficulties for larger N because the number of states
increases exponentially as 2N . The classical SA solves this problem by ex-
ploiting stochastic processes, Monte Carlo simulations, which have the com-
putational complexity growing as a power of N. The corresponding reduction
of the computational complexity is lacking in QA, because the Schrödinger
equation is not replaced directly by the stochastic processes. While the
quantum Monte Carlo is not equivalent to the Schrödinger equation, we will
conduct QA in the quantum Monte Carlo framework in the next chapter.
Another problem is to devise implementations of QA in other optimization
problems such as the traveling salesman problem or the Hamilton problem.
The implementation of QA to the traveling salesman problem is explained
as an example in Chapter 4.
Chapter 3

Monte Carlo Analysis of Larger

Systems

We will apply the Monte Carlo method to QA in this chapter. Two possible
ways to solve QA are considered. Comparison between SA and QA by the
Monte Carlo method is also performed.
We desire to solve SA and QA by another method, because the calcula-
tions of the master equation for SA and the Schrödinger equation for QA be-
come difficult for large-size systems. As shown in the previous chapter, when
we search the ground state by SA and QA, the operation of 2N -dimensional
matrices is needed to solve the differential equations in the Ising-spin sys-
tems. For example, real and complex 1024 ×1024 matrixes are needed for SA
and QA with ten spins, respectively. To avoid such a difficulty of solving the
master equation, the calculation is performed by the Monte Carlo method in
the original definition of SA. Solving SA by the master equation is the equiv-
alent to the Monte Carlo method. From the relation of the master equation
and the Monte Carlo method in SA, we consider QA by the Monte Carlo
method.
Another motivation of considering the Monte Carlo method is the fol-
lowing: In the actual procedure of QA, all the elements of the Hamiltonian
matrix are calculated to solve the the Schrödinger equation. If the matrix is
expressed in the z-representation, the diagonal elements are the classical-term
energy of each configuration. This implies that there is no advantage to the
enumeration method (to enumerate all the possible configurations), because

33
CHAPTER 3. MONTE CARLO ANALYSIS 34

we already calculated all the energy of possible configurations. Therefore, we

have to solve QA by another method which requires less calculations than to
solve the differential equation, like the Monte Carlo method.
We consider the two methods, the path-integral Monte Carlo and the
quantum Monte Carlo methods, for QA. The results of the two methods are
not the same as the result of the Schrödinger equation, but the ground state
can be found by both of the methods. We adopt the quantum Monte Carlo
method for QA and compare the results of SA and QA by the Monte Carlo
simulation.

3.1 Monte Carlo Method

The Monte Carlo method is powerful to analyze various problems of statis-
tical mechanics. In classical statistical mechanics, we obtain the expectation
value of a quantity A as a function of the temperature T = 1/β in the
following form:
P −βE
Ae
hAi = P −βE , (3.1)
e

where the summation runs over all the possible configurations of the system
and E corresponds to the energy of each configuration. The number of all
possible configurations is S N , if the system includes N sites and each site
takes S states independently. The number of the terms in the summation
increases exponentially as a function of the system size N. It is difficult to
calculate the whole summation for large-size systems.
We consider replacing the summation with the weighted sampling. Equa-
tion (3.1) is rewritten as
X
hAi = Pi Ai , (3.2)
i
P
where Pi = e−βEi / j e−βEj . Pi is the Gibbs distribution, the probability of
the configuration i. The sampling configurations are generated to obey the
Gibbs distribution.
CHAPTER 3. MONTE CARLO ANALYSIS 35

The expression of the expectation values by the weighted sampling is

given by
M
1 X
hAisampling = A(i), (3.3)
M i=1
where the summation runs over the sampled configurations. In the large M
limit, hAisampling converges to hAi.
The configuration obeying the Gibbs distribution can be obtained from
the Markov chain which comes from the master equation as
dP (t)
= LP (t), (3.4)
dt
where L is the transition matrix defined in the previous chapter. The matrix
element which is the transition probability from ith state to jth state has
a non-zero value, when the configurations of ith and jth states differ only
by one spin. Therefore, the ith state can be modified to N states. We
create a new configuration from the ith state by the following steps: Choose
a site randomly and determine the direction of the spin on the site by the
transition ratio from the configuration before flip to the configuration after
flip. Next, repeat this procedure N times. These N trials of the spin flip
are called “one Monte Carlo step”. The Markov chain is obtained from the
calculation of this Monte Carlo steps. In principle, every configuration can
change to any configuration in one Monte Carlo step, because all spins can
flip once. However, the distance between two configurations is close, if one
configuration is obtained from the other configuration by one Monte Carlo
step. For the precise calculation, the interval of the sampling has to be long
enough.

3.2 Quantum Monte Carlo Method

In quantum systems, the definition of the expectation value is not the same
as in the classical systems. The expectation value of a quantity A is given
by
TrAe−βH
hAi = , (3.5)
Tre−βH
CHAPTER 3. MONTE CARLO ANALYSIS 36

where H is the Hamiltonian. To calculate the operator e−βH is difficult for

large-size systems because of non-diagonal terms of the Hamiltonian H. The
same procedure as in the classical Monte Carlo method can not be applied
to quantum systems, if we can not diagonalize the Hamiltonian.
The following procedure avoids the above difficulty. Now we consider the
the Ising system with transverse field Γ as a quantum system,
X X
H=− Jij σiz σjz − Γ σix . (3.6)
ij i

This Hamiltonian is not diagonal in the z-representation. By the Trotter

formula, the operator e−βH can be described as a product of many operators
diagonalized in z- or x-representation:

e−βH = lim (eA/M eB/M )M , (3.7)

M →∞
P P
where A = ij Kij σiz σjz (Kij = βJij ) and B = γ i σix (γ = βΓ). This
decomposition is called the Suzuki-Trotter decomposition [38, 39]. Using this
decomposition, we obtain the partition function of the Mth decomposition
ZM as

ZM = Tr(eA/M eB/M )M (3.8)

X
= h{σj1 }|eA/M |{σj1
′ ′
}ih{σj1 }|eB/M |{σj2 }i
{σjk =±1}

×h{σj2 }|eA/M |{σj2

′ ′
}ih{σj2 }|eB/M |{σj3 }i
×······
×h{σjM }|eA/M |{σjM
′ ′
}ih{σjM }|eB/M |{σj1 }i, (3.9)

where |{σjk }i is the Mth direct product of eigenstates {σj } defined as

|{σjk }i = |σj1 i ⊗ |σj2 i ⊗ · · · ⊗ |σjM i. (3.10)

From this decomposition, the partition function is expressed as a trace of

products of diagonalized matrices. Each part of the product is rewritten as
follows:
! N
1 X Y
h{σjk }|eA/M |{σjk
′
}i = exp Kij σik σjk ′
δ(σjk , σjk ). (3.11)
M ij j
CHAPTER 3. MONTE CARLO ANALYSIS 37

N
!
X
h{σjk }|eB/M |{σj,k+1}i = aN
M exp γM σjk σj,k+1 , (3.12)
j=1

where
  1/2 γ
1 2γ 1
aM = sinh , γM = ln coth .
2 M 2 M

The partition function Z is represented as

X
Z = lim aN
M
M
M →∞
{σjk =±1}
(M M
!)
X 1 X X
exp Kij σik σjk + γM σjk σj,k+1
k=1
M ij j=1
X
= lim aN
M
M
M →∞
{σjk =±1}
( M N
!)
X X X
exp βeff Jij σik σjk + ΓM σjk σj,k+1 ,(3.13)
k=1 ij j=1

where βeff = β/M and ΓM = γM /βeff . This is the partition function of a

(d + 1)-dimensional classical spin system at the effective inverse temperature
βeff . The quantum d-dimensional partition function is mapped to a (d + 1)-
dimensional classical partition function. The representation of the classical
(d + 1)-dimension system thus can be obtained by the usual Monte Carlo
method (see Fig. 3.1).
The framework does not change, when the longitudinal field is applied.
P
It produces the term −h i σiz in the Hamiltonian. We can deal with this
P
term similarly to the term − ij Jij σiz σjz .

3.3 Path-Integral Monte Carlo

The time development of the wave function in quantum systems is given as
RT
|Ψ(T )i = Te−i 0 dtH(t)
|Ψ(0)i, (3.14)
CHAPTER 3. MONTE CARLO ANALYSIS 38

real space

J ij
ΓM
Trotter
Figure 3.1: The d-dimensional quantum system is mapped to a (d + 1)-
dimensional classical system at the inverse temperature βeff . The original d-
dimensional space is expressed as the “real space” and the additional space
is expressed as the “Trotter”. Copies of the original system without the
quantum term are placed in the Trotter direction. The real space interactions
Jij are random variables in the real space direction but uniform in the Trotter
direction. The interactions between the Trotter slices, ΓM , are uniform and
depend on the amplitude of the transverse field Γ.

where the symbol T means that the operator acts on the ket vectors with time
ordering. Decomposing the product of the short time (δt = T /M) operators,
we have
T HM −1 T HM −2 T H1 T H0
|Ψ(T )i = lim e−i M e−i M · · · e−i M e−i M |Ψ(0)i, (3.15)
M →∞

where Hj = H(T j/M). By replacing the time with the imaginary time
(iT → β), we obtain
βHM −1 βH0
|Ψ(−iβ)i = lim e− M · · · e− M |Ψ(0)i. (3.16)
M →∞
CHAPTER 3. MONTE CARLO ANALYSIS 39

(a)
real space

J ij
ΓM(β)
(b)
state

β=0 imaginary time β=0

Figure 3.2: (a) A path of the imaginary time evolution is expressed as a
configuration on a (d + 1)-dimensional space. As the transverse field becomes
weak, the interaction between the Trotter slices, ΓM (β), becomes strong.
This fact is illustrated as the thickness of bonds. (b) The initial states |Φ(0)i
are located at each end and evolute toward the center.
CHAPTER 3. MONTE CARLO ANALYSIS 40

The expectation value of A is obtained as,

hA(T )i = hA(−iβ)i (3.17)

= hΨ(−iβ)|A|Ψ(−iβ)i (3.18)
= lim hΨ(0)|e−β/M H0
· · · e−β/M HM −1
A
M →∞
−β/M HM −1
×e · · · e−β/M H0
|Ψ(0)i. (3.19)

If the imaginary time β is a real number, this formulation is similar to

the quantum Monte Carlo method. Dividing the Hamiltonian into two parts,
the terms including σ z or σ x , and inserting the complete sets between each
product, we can map the equation (3.19) to the weighted sampling of the
Monte Carlo simulation. One configuration in the Monte Carlo simulation
corresponds to one possible path of the imaginary time dynamics in the
original system. The summation of the samplings is the discrete version of
the imaginary time path integral. The system simulated in the Monte Carlo
is shown in Fig. 3.2. The initial state |Ψ(0)i is located at the right and
the left sides. The evolved state |Ψ(β)i locates at the center. A path of
the sampling is obtained from the snapshot of the thermal equilibrium state
of a (d + 1)-dimensional space. The horizontal interaction comes from the
transverse field. If the field is scheduled (decreases as a function of time),
the interaction changes stronger toward the center. The same systems are
located in the horizontal direction. These are similar to the Trotter slices in
the quantum Monte Carlo method.

3.4 Applying Monte Carlo Method to An-

nealing Process
In statistical mechanics, Monte Carlo simulation and the master equation of
the classical system give the same result except for statistical errors. The
results in the dynamics are also the same. (For the dynamics of the SK model,
see [40, 41, 42, 43]) From this equivalence, we regard the Monte Carlo step
as the time of the master equation. Figure 3.3 shows the probability to find
the ground state of the SK model with N = 8 for the master equation and
 CHAPTER 3. MONTE CARLO ANALYSIS 41

1
Master eq.
Monte Carlo
0.8

0.6
P(t)

0.4

0.2

0
1 10 100 1000
t

Figure 3.3: The results of the master equation and the Monte Carlo simu-
lation. Two curves show good agreement.

the Monte Carlo simulation. The probability for the Monte Carlo simulation
is obtained from the ratio of the number of the ground state configurations
for each time in one hundred independent runs. The temperature schedule
√
of both cases is T = 3/ t. Two curves show good agreement.
On the other hand, there is no reason that the quantum Monte Carlo
simulation is equivalent to the solution of the Schrödinger equation. However,
we regard that the Monte Carlo step in the quantum Monte Carlo simulation
is also equivalent or similar to the time in the Schrödinger equation. How
different between the two results is shown later.
Besides the above approach, the path-integral Monte Carlo also describes
the dynamics of the quantum system. The imaginary time path integral
of the quantum system and the partition function at the temperature 1/β,
where β is the imaginary time, have the same expression. The time develop-
ing operator e−βH can be decomposed like the Suzuki-Trotter decomposition
in the quantum Monte Carlo method. Therefore, the imaginary time path
 CHAPTER 3. MONTE CARLO ANALYSIS 42

1
Schrödinger eq.
Quantum Monte Carlo
Imaginary Time Schrödinger eq.
0.8 Path Integral Monte Carlo

0.6
P(t)

0.4

0.2

0
1 10 100 1000
t

Figure 3.4: The results for the Schrödinger equation, the quantum Monte
Carlo method, the imaginary time Schrödinger equation and the path-
integral Monte Carlo method. Imaginary time Schrödinger equation means
the results of imaginary
p time dynamics of the Schrödinger equation. The
schedule of Γ(t) is 3/ (t).
1

0.8

0.6
P(t)

0.4

0.2 Schrödinger eq.

Quantum Monte Carlo
Imaginary Time Schrödinger eq.
Path Integral Monte Carlo
0
1 10 100 1000
t

Figure 3.5: The results for the Schrödinger equation, the quantum Monte
Carlo method, the imaginary time Schrödinger equation and the path-
integral Monte Carlo method. The schedule of Γ(t) is 3/t. Except for the
result for the Schrödinger equation, the probabilities converge to one.
CHAPTER 3. MONTE CARLO ANALYSIS 43

integral is evaluated numerically.

In this formulation, the path integral is replaced with the Monte Carlo
simulation. However, we have three difficulties about this simulation. The
Hamiltonian with the imaginary time H(−iβj/M) is no longer a Hermitian
matrix, because the amplitude of the transverse field Γ(t) is a function of
the time and generally becomes a complex number when β is a real number.
We can avoid this difficulty by changing the definition of the time in Γ(t)
√ √
as t → tt∗ , where t∗ is the complex of t conjugate of t, because tt∗
is a real number for any complex number t. Secondly, the simulation is
performed by replacing imaginary numbers with real numbers, so that we
have to evaluate A(T ) from A(β) by the process of analytic continuation. The
analytic continuation for the function whose values are given as a numerical
number is a great problem itself and needs large computational power like
the simulated annealing. Finally, the computational time of the path-integral
Monte Carlo method may be longer than the quantum Monte Carlo method.
The imaginary time evolution is mapped to the spatial axis, so that we
have to wait until the system becomes in equilibrium. We can decrease the
transverse field in the quantum Monte Carlo method without care that the
system is trapped at a local minimum or not. In the path-integral Monte
Carlo, we care about that the system is trapped at a local minimum or not,
because the expectation value is obtained from the average of various paths
whose distribution obeys the Gibbs distribution of the (d + 1)-dimensional
system.
If we accept that the imaginary time expectation value A(β) is regarded
as a real time one A(T ), the final problem is still a difficult problem. It takes
long time to wait until the system becomes in equilibrium, if the landscape
of the system is complicated.
We compared the above two approaches, the quantum and the path-
integral Monte Carlo methods, with the results of the direct solutions of the
Schrödinger equation. The system is the SK model with N = 8 and the
√
schedule is Γ = 3/ t. The probability to find the ground state is plotted
in Fig. 3.4. The quantum Monte Carlo is performed with M = 1000. We
wait 5000 Monte Carlo steps for the initial relaxation and take average over
CHAPTER 3. MONTE CARLO ANALYSIS 44

the next 5000 steps in the path-integral Monte Carlo simulation. We put
the inverse temperature β = 1000 for each calculation. The result of the
imaginary time dynamics of the Schrödinger equation is also plotted.
In Fig. 3.5, we plot the solution whose schedule is Γ = 3/t. In this
schedule, the solution of the Schrödinger equation does not converge to the
ground state with probability one, but the others converge to one. The curve
of the path-integral Monte Carlo agrees with the curve of the imaginary time
Schrödinger equation.
From these results, we can conclude that the quantum Monte Carlo
and the path-integral Monte Carlo simulations are not the same as the
Schrödinger equation. We regard that the assumptions of the two Monte
Carlo simulations are not correct. The assumptions are the following: For
the quantum Monte Carlo simulation, we assume that the Monte Carlo step
equals to the real time. For the path-integral Monte Carlo simulation, we
assume that the imaginary time dynamics equals to the real time one.
However, the goal is to find the ground state as fast as possible. The two
Monte Carlo methods find the ground state configuration with probability
one, even if the schedule is too fast to solve for QA by the Schrödinger
equation. For this purpose, we can adopt one of the Monte Carlo methods.
We consider that the quantum Monte Carlo method has the advantage. The
quantum Monte Carlo method is similar to SA, because the control parameter
(T for SA and Γ for QA) decreases as a function of the Monte Carlo step.
Thus, the longer we perform the simulation, the larger the probability to
find the ground state becomes. On the other hand, a long simulation makes
the statistical errors small for the path-integral Monte Carlo method, but
the probability does not increase. For these reasons hereafter we adopt the
quantum Monte Carlo method for QA of the large-size systems.
We compared the probabilities to find the ground state in the Monte Carlo
√
version of SA and QA for the SK model with N = 51 and T = Γ = 3/ t
in Fig. 3.6. In this calculation, we perform 100 independent runs of long-
time (t = 100000) SA with the sufficiently slow schedule T = 3/ ln(t +
1) and determine the ground state energy beforehand. Using this ground
state energy, we can calculate the probability to find the ground state. The
 CHAPTER 3. MONTE CARLO ANALYSIS 45

1
SA
QA
0.8

0.6
P(t)

0.4

0.2

0
10 100 1000 10000
t

Figure 3.6: Overlaps with the√ground state are plotted. The system is the
SK model with N = 51 and 3/ t schedule.

probability to find the ground state is the ratio of the number of the ground
state configurations for each time in ten thousand independent runs of SA
and Trotter slices of QA. The probability in QA is larger than SA. This result
means that the QA in the quantum Monte Carlo method also improves the
performance in finding ground state. The quantity 1 − P (t) is plotted in
Fig. 3.7 under the log-log scale. The curve converges to zero asymptotically
as 1/t. This implies that the system almost follows the stationary state and
the annealing schedule is sufficiently slow. Moreover, we apply fast schedule
to the same system and the probability converges to one asymptotically as
1/t2 . In this fast schedule, system also follow the stationary state.
We also plot the average energy of SA and QA under the same condi-
tion in Fig. 3.9. As we need the ground state of the classical term of the
Hamiltonian, we neglect the contribution of the transverse field term in the
calculation of the energy. We take the average energy which does not contain
the interactions between the Trotter slices for each snapshot of the Trotter
 CHAPTER 3. MONTE CARLO ANALYSIS 46

1
QA
1/t
1−P(t)

0.1

0.01
10 100 1000 10000
t

Figure 3.7: The quantity 1 − P (t) is plotted. The conditions are the same
as in Fig. 3.6.

1
QA
1/t2
1−P(t)

0.1

0.01
10 100 1000 10000
t

Figure 3.8: The quantity 1 − P (t) is plotted. The schedule is Γ = 3/t.

CHAPTER 3. MONTE CARLO ANALYSIS 47

slices in QA. The average of the energy per spin is calculated by M inde-
pendent runs for SA and M Trotter slices for QA, respectively. In spite of
the large difference between SA and QA in P (t), the energy of QA is lower
than SA, but the difference is not so large. This can be understood that if
the system is trapped in a local minimum, the trapped system also lowers
the average energy but does not count in P (t).
Calculating the probability to find the ground state is a useful index to
check which method has a good performance. However it is too difficult to
perform this calculation for large-size systems, because the calculation needs
the ground state configuration, but we do not know it. In this case, we regard
that the method whose average energy is lower has a better performance. If
the average energy of QA is lower than SA, QA may find the ground state
more efficiently than SA.
On the other hand, another index can be considered. That is the lowest
energy in the M independent runs of SA or in M Trotter slices of QA. How-
ever, comparing the performance by the average energy gives more precise
conclusion than by the lowest energy, because the lowest energy depends on
the initial state more than the average energy. If the initial state is in the
basin of the ground state, the system goes to the ground state with large
probability. We will calculate SA and QA by the Monte Carlo method for
the large-size systems up to N = 10000 and compare their performance by
the average energy.

3.5 Results of the Monte Carlo simulation

Two additional parameters which do not appear in the Schrödinger equation
are needed, when the calculation of QA is performed by the quantum Monte
Carlo simulation. The first one is the (inverse) temperature T (= 1/β), and
second one is the Trotter number M. The original dynamics, the Schrödinger
equation, does not contain thermal fluctuations, so that we have to take the
limit β → ∞. The Trotter number M should also be infinity to take into
account the quantum effect correctly.
Each Trotter slice is the classical system with the effective inverse tem-
 CHAPTER 3. MONTE CARLO ANALYSIS 48

−0.4
SA
QA
−0.45

−0.5

−0.55
<E>

−0.6

−0.65

−0.7

−0.75
10 100 1000 10000
t

Figure 3.9: The average energy is√plotted for SA and QA. The system is
the SK model with N = 51 and 3/ t schedule.

perature βeff (= β/M) except for the interactions between Trotter slices. We
consider the parameters βeff and M instead of the parameters β and M, be-
cause it is natural to regard that we simulate the classical (d+1)-dimensional
system at the effective inverse temperature. The infinity limit of the Trotter
number also means the infinity limit of the inverse temperature, because the
ratio β/M is fixed.

3.5.1 Dependence on the temperature

First, we consider the effective inverse temperature βeff . At the initial time,
the strength of the interaction between Trotter slices is small and each slice
is almost an independent classical system. At the end, the strength goes to
infinity. The M spins in the Trotter direction take the same value because of
the interactions of the infinite strength and the freedom of this direction van-
ish. The (d + 1)-dimensional system which is equivalent to the d-dimensional
quantum system reduced to the d-dimensional classical system. This can
CHAPTER 3. MONTE CARLO ANALYSIS 49

be understood naturally, because the transverse field Γ(t) vanishes and the
Hamiltonian goes to the classical Hamiltonian.
The spins in the Trotter direction take same value when the interactions
between the Trotter slices is stronger than the thermal fluctuation. If the
temperature is high, we need long time until the interaction becomes large
enough. From this, the temperature should be low to save the time for
the calculation. On the other hand, the temperature is required to be high
enough for searching the configuration space widely at the initial time. If
the temperature is low, the system can not visit various configurations and
trapped at a local minimum.
To satisfy the above two conflicting conditions, we have to choose mod-
erate values of temperature. The system can seek all possible configurations
when the temperature is above the critical temperature of the order-disorder
transition of the classical system. Thus, we choose βeff = 1 for the SK model.
The dependence of the probability on βeff is shown in Fig. 3.10. The
condition is βeff = 1/2, 1, 2, 5, N = 8 and M = 1000 for the SK model in the
√
3/ t schedule. The simulation with βeff = 1 has a best performance among
the four.

3.5.2 Dependence on the number of Trotter slices

Secondly, we consider the Trotter number M. A simulation with large num-
ber of Trotter slices gives precise results, but the computational power for
the calculation becomes large. The computational time for the quantum
Monte Carlo simulation is M times longer than the conventional simulated
annealing. We have to estimate the reasonable Trotter number of the correct
simulation.
We check the dependence of the simulation on the Trotter number. The
probabilities to find the ground state versus the transverse field are plotted
for the cases of M = 10, 20, 50, 100, 200, 500, 1000 in Fig. 3.11. The condition
√
is Γ = 2/ t, N = 625 and βeff = 1 for the two-dimensional Edward-Anderson
model. We find that M = 100 is large enough.
From the above two results, we regard that the conditions of βeff = 1 and
 CHAPTER 3. MONTE CARLO ANALYSIS 50

0.8

0.6
P(t)

0.4

0.2 β/M=0.5
β/M=1
β/M=2
β/M=5
0
10 100 1000
t

Figure 3.10: The probability dependence on the ratio β/M. The simulation
with βeff = 1 has a best performance among the four.

M = 100 give a reasonable solution. Hereafter we will use this condition.

3.5.3 Results of large size system for the EA model

Next, we consider systems of larger size than the previous calculations to
demonstrate that our method can be applied to the actual problems. The
model is the two-dimensional Edward-Anderson (EA) model. The calculation
per Monte Carlo step for the EA model is less than the SK model. The spin
flip is determined by the calculation of the energy gap of the spin flip. We
have to sum up N −1 interactions for the SK model, because all spins interact
with each other in such an infinite range model. For the two-dimensional EA
model on the square lattice the number of the interactions is four. Therefore,
we can accelerate the summation in the EA model (N − 1)/4 times faster
than the SK model. The acceleration affects both of the methods SA and
QA.
We choose the standard condition of the calculation of this model as
 CHAPTER 3. MONTE CARLO ANALYSIS 51

0.3
M=1000
0.28 M=500
M=200
0.26 M=100
M=50
Magnetization

M=20
0.24 M=10
0.22

0.2

0.18

0.16

0.14
0.1 1
Γ

Figure 3.11: The magnetization dependence on the Trotter number M.

Curves of over M = 100 are large enough.

follows; N = 625(252) with periodic boundary conditions, βeff = 1, M = 100

√
and the schedule T = Γ = c/ t. The critical temperature Tc is considered
zero in this model. This is not the same value of the SK model. However,
the zero temperature dynamics only seeks the configuration whose energy is
lower than the present configuration and tends to be trapped at local minima,
so that we use the same condition as βeff = 1.
We check how the system follows the stationary (equilibrium) state which
means equilibrium state at fixed T for SA and ground state at fixed Γ for
QA. For small-size systems, we can calculate the probabilities to find the
ground state PSA (t) and PQA (t) and compare with the probabilities in the
st st
stationary state PSA (T (t)) and PQA (Γ(t)). If the two curves, the results of
the dynamics and the statics, are the same, the system follows the stationary
state. For large-size systems we cannot calculate the probability to find the
ground state, because the calculation of the probability needs the ground
state configuration beforehand and that is difficult. Nevertheless, we can
CHAPTER 3. MONTE CARLO ANALYSIS 52

check whether the system follows the stationary state or not by another
quantity. In our case, a longitudinal magnetic field is applied to the system,
and magnetic order grows as a consequence of this field. If the system is in
the stationary state, the magnetization is characterized by the temperature
for SA or the transverse field for QA and does not depend on the speed of
√
the schedule. The functional form of the schedule is c/ t, thus the speed of
our schedule is controlled by c (slow for large c and fast for small c). The
system can not follow the stationary state for the rapid schedules, the cases
of small c. We check the dependence of magnetization on c.
The magnetization versus the temperature for SA and the transverse
field for QA is plotted in Figs. 3.12 and 3.13. For SA, the final value of the
magnetization (the left end of the curves) depends on c. This means that the
schedule is too fast to follow the stationary state. This result implies that
the system tends to be trapped at local minima in this schedule up to c = 50.
For QA, the final value of the magnetization does not depend on c beyond
√
c = 10. Thus, we can regard that the schedule Γ = c/ t with c = 10 is slow
enough to search the ground state of the two-dimensional EA model. The
conclusion of these two results for SA and QA is that SA does not follow the
√
stationary state and QA does in the c/ t schedule. This is consistent with
small-size analysis by differential equations (the master equation for SA and
the Schrödinger equation for QA).
We average out the energy for SA with M(= 100) independent runs and
QA with M Trotter slices. The results on the average energy are plotted in
Fig. 3.14. Only in the middle region of the time, the energy in QA is higher
than in SA, but finally the energy in QA goes lower than in SA.
As shown in the previous chapter, the performance in finding the ground
state is improved by the quantum effect. However, the quantum Monte
Carlo uses the Suzuki-Trotter decomposition by definition. The calculation
of the quantum Monte Carlo needs Trotter-number M times longer than the
calculation of the classical system. Therefore, the comparison between SA
and QA is still incomplete. The time scales of the two computations are not
the same. One Monte Carlo step in QA takes about M times evaluations of
spin flips than in SA. (To be precise, the calculation in QA is M(z + 2)/z
 CHAPTER 3. MONTE CARLO ANALYSIS 53

0.25
C=2
C=5
C=10
0.2 C=20
C=50
Magnetization

0.15

0.1

0.05

0
0.1 1 10
T

Figure 3.12: The dependence of magnetization on c for the conventional

simulated annealing (SA)
0.28
C=1
0.26 C=2
C=5
0.24 C=10
0.22 C=20
Magnetization

C=50
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.1 1 10
Γ

Figure 3.13: The dependence of magnetization on c for quantum annealing

(QA)
 CHAPTER 3. MONTE CARLO ANALYSIS 54

−0.1
SA
QA
−0.2

−0.3
<E>

−0.4

−0.5

−0.6

−0.7
10 100 1000 10000 100000
t

Figure 3.14: √Average energy for SA and QA are plotted. The schedule is
T = Γ = 10/ t and the system size is N = 625.

times longer than the calculation in SA, where z is the number of nearest
neighbors and “2” is the number of the interactions in the Trotter direction.
It is exactly “M times” for the infinite range model in the thermodynamic
limit.) Taking into account this disadvantage gives a reasonable comparison
between SA and QA. We provide a reasonable definition of the time t′ of
which the quantity is plotted and compared as a function as,

t′ = t (for SA) t′ = Mt (for QA), (3.20)

The Trotter number is 100 in our case, so that the time is t′ = 100t for QA.
The average energy for SA and QA is plotted in Fig. 3.15 under the
condition of such a scaling of the time. QA improves the performance in
finding the ground state. In the calculation, we adopt that the schedule in
SA is 100 times slower than QA, because SA could not follow the stationary
√
state in the c/ t schedule (see Fig. 3.12). The total Monte Carlo step of
SA is 100 times longer than QA, because we have to use same amount of
the computational power to compare the two methods (QA needs 100 times
CHAPTER 3. MONTE CARLO ANALYSIS 55

calculations than SA). The values of the temperature and the transverse
field at the end of the simulations are the same. We also plot the cases of
various schedules for the temperature and the transverse field in Fig. 3.16
and Fig. 3.17. QA’s performance is also good in the various schedules. For
another system size, N = 10000, the average energy in QA is also lower than
SA as seen in Fig. 3.18.
At the end of simulations, we cannot reach the zero temperature or zero
transverse field. The values are quite small (∼ 0.03) but not equal to zero.
For example, the strength of the interacrion between the Trotter slices is 1.73,
√
when Γ = 0.1/ 10. It is not so strong compared with the effective inverse
temperature βeff = 1. The thermal fluctuation still remains. To remove
this fluctuation, we execute zero temperature dynamics after the annealing.
Then, almost all the Trotter slices (more than 90% for the situation of the
calculation for Fig. 3.15) converge to the same configuration in QA, while
the independent runs in SA do not.

3.5.4 Difference between the annealing and the quench-

ing processes
Let us take another approach to investigate the property of QA. How effi-
ciently does the annealing process find the ground state in comparison with
the quenching process? We calculate the average energy of the two processes,
the annealing and the quenching, for SA and QA. The temperature and the
√
transverse field are quenched at T = Γ = 0.1/ 10 = 0.0316 · · · , and sched-
√
uled from infinity to zero by a function 10/ t. The quenching process of the
quantum system is almost equivalent to the case of the work by Sato et al. [32]
They considered the dynamics of the quenched system with Γ → ∞. The re-
sults are shown in Fig. 3.19. The temperature T and the transverse field Γ for
annealed and quenched systems are the same at the end of the simulations. If
the systems stay in the stationary state, the curves arrive at the same point
at the end. The average values of energy of SA, QA and Γ-quenched systems
converge to almost the same value, but the T -quenched system is trapped
in a high-energy state. Of course the energy of the Γ-quenched system is
CHAPTER 3. MONTE CARLO ANALYSIS 56

higher than that of QA (see Fig. 3.20. It is an enlargement of Fig. 3.19).

The behavior of the two quenched systems are quite different, though both
of the controlled parameters T and Γ are quenched. The quantum system
relaxes faster than the classical system, when the transverse field takes the
same value as the temperature of the classical system. This is consistent
with the result that QA reaches the stationary state in magnetization, but
SA does not when the coefficient c changes from 2 to 50.

3.6 Summary
We applied the Monte Carlo method to QA. There are two ideas to translate
the time in the Schrödinger equation to the Monte Carlo method. One is the
path-integral Monte Carlo which regards the Trotter direction as the time
(more precise, the imaginary time). The other is the quantum Monte Carlo
with “the Monte Carlo step”-dependent inter-Trotter interaction. We regard
the Monte Carlo step as the time in the Schrödinger equation.
The path-integral Monte Carlo is performed in the imaginary time dy-
namics. This results in the different behavior from the solution of the
Schrödinger equation. For the quantum Monte Carlo, the result also differs
from the result of the Schrödinger equation. Two reasons are mentioned.
The energy dissipation occurs in the Monte Carlo simulation, while it does
not occur in the original Schrödinger equation. Beside this, the time in the
path-integral and the quantum Monte Carlo methods are not same as the
time in the Schrödinger equation. However, these methods have good per-
formance and the ground state is found, even if the schedule is too fast to
find the ground state for the Schrödinger equation.
We adopt the quantum Monte Carlo for large-size QA to reduce the com-
putational time. We have to wait until the system is in equilibrium in the
path-integral Monte Carlo and summing up sampling paths. This takes large
computational time.
The Trotter number we adopt is M = 100, which is large enough for
our calculations (two-dimensional EA model with N = 625). The results are
compared with SA. We found that QA improves a performance in finding the
CHAPTER 3. MONTE CARLO ANALYSIS 57

ground state even if we take into account that the QA need M times more
calculations than SA. Moreover, almost all the configurations of the Trotter
slices converge to the same configuration.
We also checked the difference between annealed and quenched systems.
The relaxation time of the classical quenched system is quite long. On the
other hand, the quantum quenched system does not so differ from the an-
nealed system QA. This implies that the quantum effect accelerates relax-
ations. Moreover, it is remarkable that the average energy of the quantum
quenched system is lower than SA. The short relaxation time makes the de-
pendence on the schedule small. This can be seen in Figs. 3.12 and 3.13.
A rapid decrease in the characteristic relaxation times for the random Ising
magnet, LiHo0.167 Y0.833 F4 , with the transverse field is observed [44]. We
consider that the results may come from the same mechanism.
From the application side, the optimal schedule is not known in general.
If the dependence of the performance to find the ground state on the anneal-
ing speed is small, we can decrease the rate not to reach the ground state
or the optimal configuration. This is an attractive property for the actual
applications.
 CHAPTER 3. MONTE CARLO ANALYSIS 58

−0.45
SA C=2
SA C=20
QA C=2
−0.5

−0.55
<E>

−0.6

−0.65

−0.7
1000 10000 100000
t’

Figure 3.15: Time evolution of the average energy for QA and SA. The
system size is N = 625. One hundred Trotter slices are averaged for QA
and 10 runs are averaged for SA. For “SA C = 2” and “QA C = 2” the
schedules of the temperature and the transverse field are the same function.
“SA C = 20” and “QA C = 2” take √ the same value of the temperature
and the transverse field T = Γ = 0.1/ 10 = 0.0316 · · · at the end of the
simulations.
−0.45
SA T=0.4079/log(t+1)
SA T=2/log(t+1)
SA T=5/log(t+1)
−0.5 QA C=2

−0.55
<E>

−0.6

−0.65

−0.7
1000 10000 100000
t’

Figure 3.16: Time evolution of the average energy for QA and SA. The
system size is N = 625. One hundred Trotter slices are averaged for QA and
 CHAPTER 3. MONTE CARLO ANALYSIS 59

−0.65
SA C=50
QA C=5
−0.655

−0.66

−0.665
<E>

−0.67

−0.675

−0.68

−0.685
10000 100000 1e+06
t’

Figure 3.17: Time evolution of the average energy √ for QA and √SA. The
system size is N = 625 and the schedule is T = 50/ t and Γ = 5/ t. One
hundred Trotter slices are averaged for QA and 10 runs are averaged for SA.
The temperature and the transverse field at the end of the simulations are
the same value T = Γ = 0.05.
−0.62
SA C=20
−0.625 QA C=2
−0.63
−0.635
−0.64
<E>

−0.645
−0.65
−0.655
−0.66
−0.665
10000 100000
t’

Figure 3.18: Time evolition of the average energy√for QA and SA. √ The
system size is N = 10000 and the schedule is T = 20/ t and Γ = 2/ t. One
hundred Trotter slices are averaged for QA and 10 runs are averaged for SA.
The temperature and the transverse
√ field at the end of the simulations are
 CHAPTER 3. MONTE CARLO ANALYSIS 60

−0.1
SA
SA(quench)
−0.2 QA
QA(quench)
−0.3
<E>

−0.4

−0.5

−0.6

−0.7
10 100 1000 10000 100000
t

Figure 3.19: Solutions for annealing and √ quenching processes are plotted.
The annealing
√ schedule is T = Γ = 10/ t and the quenched value is T =
Γ = 0.1/ 10 = 0.0316 · · · which is the final value of the schedule.
−0.66
QA
QA(quench)
−0.665

−0.67
<E>

−0.675

−0.68

−0.685

−0.69
1000 10000 100000
t

Figure 3.20: Solutions for annealing and quench process for QA are plotted.
This is the enlargement of Fig. 3.19.
Chapter 4

Application to the Traveling

Salesman Problem

All the problems we calculated up to the previous chapter were on the Ising-
spin systems. However, the combinatorial optimization problem is not lim-
ited to the problems based on physics. One of the major problems is the
traveling salesman problem (TSP) [1, 2]. We will apply our method to the
TSP in this chapter. The TSP is a hard problem to solve by a computer,
say an NP -hard problem. Therefore, comparison between SA and QA in the
TSP may provide a good measure of their performances in the problems of
other areas not limited to physics. In other words, we will check generality
of QA for combinatorial problems.

4.1 The Traveling Salesman Problem

We have N points or cities in a space with distances dij between them. The
task is to find the minimum-length closed tour that visits each city once
and returns to its starting point. Figure 4.1 shows an example. Practical
applications include scheduling of truck deliveries, airline crews, and the
movements of an automatic drill-press or robot arm. In many applications
the “distances” dij are abstract quantities not related to a Euclidean dis-
tance between points in a real space; they may not even satisfy the triangle
inequality dij ≤ dik + dkj . For the example of truck deliveries, the cost to
transport freight from city i to j is affected by many factors, amount of the

61
CHAPTER 4. APPLICATION TO TSP 62

(a) (b)
Figure 4.1: The traveling salesman problem, showing a good (a) and a bad
(b) solution to the same problem.

freight to city j, labor, gas, highway charge, the weather, etc. The length of
the tour is defined by
N
X
L= di(m),i(m+1) , (4.1)
m=1

where i(m) means the city of the mth stop, the tour is closed as i(N + 1) =
i(1) and N is the number of cities.
One can think of a simple approach, listing all possible tours of the prob-
lem to search the optimal tour. The number of all the possible tours is
(N − 1)!/2. This number increases rapidly and becomes over the limit of the
computer power. Thus, a faster algorithm is needed to solve the actual prob-
lems. Some algorithms are optimized for TSP [1, 2] and others are general
algorithms which can be applied to TSP [21]. The simulated annealing and
the quantum annealing belong to the latter.

4.2 Simulated Annealing on TSP

The simulated annealing is based on the Monte Carlo simulation, so that we
have to determine the dynamics, how to rearrange of the tour. We choose
the simplest rearrangement of the tour like one-spin flip in the Ising-spin
system. That is to exchange two stops of cities i and j, which is illustrated
in Fig. 4.2.
CHAPTER 4. APPLICATION TO TSP 63

This rearrangement of the tour is realized in the following way: (1) Choose
two cities i and j. (2) Exchange stops of cities i and j with the probability

exp(−βLB )
ωA→B = , (4.2)
exp(−βLA ) + exp(−βLB )

where A is the route before exchange and B is the route after exchange.
One Monte Carlo step consists of N(N − 1)/2 exchange trials of all pairs.
We fix the starting point and reduce the trials to (N − 1)(N − 2)/2, because
the tour is closed. The reversibility of the tour still remains in our calculation.
√
The temperature T is scheduled in a form c/ t, where t is the Monte Carlo
step.

i i

j j

Figure 4.2: The simplest rearrangement of the tour. This rearrangement is

expressed in four spin-flip dynamics.

4.3 Quantum Annealing on TSP

We have to introduce a quantum fluctuation to the TSP, but that is not
trivial. The system is not the physical system. We cannot easily imagine
how to introduce a quantum fluctuation. A straightforward way is that we
express TSP as a binary system namely, the Ising system, so that we can
introduce the quantum effect. The idea to express TSP as a binary system
was proposed by Hopfield and Tank [45, 46].
CHAPTER 4. APPLICATION TO TSP 64

We choose stochastic binary units nia to represent possible solutions:

nia = 1 if and only if city i is the ath stop on the tour and nia = 0 for
otherwise. There are N 2 units altogether. The spin representation can be
obtained by the transformation nia → (σia + 1)/2. The total length of the
tour is
1X
L = dij nia (nj,a+1 + nj,a−1 ) (4.3)
2 ij,a
1X
= dij σia (σj,a+1 + σj,a−1 ) + const. , (4.4)
8 ij,a

and there are two constraints:

X
nia = 1 (for every city i), (4.5)
a
X
nia = 1 (for every stop a). (4.6)
i

The first constraint says that each city appears only once on the tour, while
the second says that each stop on the tour is at just one city. The tour is
expressed as a configuration of N × N units as illustrated in Fig. 4.3
The simplest rearrangement of the route of the traveling salesman is ex-
changing the stop of cities i and j. In the spin expression, four spins are
flipped at a time. The salesman stops at cities i and j for the ath and the
bth stops respectively. On-route spins take σia = σjb = 1, and off-route spins
take σib = σja = −1. One up spin and one down spin in the same row are
flipped at a time, so that the summation which runs in the horizontal direc-
tion does not change. In the vertical direction, the same situation occurs.
The constraints (4.5) and (4.6) are kept in the four-spin flip. The Monte
Carlo simulation in the spin expression is performed by this procedure.
We fix σ11 = 1 which means that the tour starts from city 1. Possible ex-
changes with any two stops of cities are the combination number of choosing
two from N − 1 as (N − 1)(N − 2)/2. One Monte Carlo step must contain
(N − 1)(N − 2)/2 spin-flip trials. This number of trials is suitable that the
system has N 2 spins and O(N 2 ) trials are performed in one Monte Carlo
step like the usual one-spin flip procedure.
CHAPTER 4. APPLICATION TO TSP 65

Stop
1 2 3 4 (1)
1

2 d13
City

d32 d41
3 d24

4
Figure 4.3: The illustrated tour is 3 → 2 → 4 → 1 with N = 4 TSP.
Solid and open circles denote units that are “1” and “0” respectively. The
condition is open boundary.

Let us consider to introduce quantum effects to the classical system. The

dynamics of SA was described by one spin-flip procedure in the preceding
chapters, and we chose the quantum effect which flips one spin at a time.
P
That is the transverse field term, −Γ i σix . The linear combination of the
off-diagonal operator σix has a non-zero value, when a bra state differ from
a ket state only by one spin. If the two spin-flip dynamics is applied to SA,
P
we have to choose the square term of σix as −Γ ij σix σjx . As known in the
Suzuki-Trotter decomposition of the XY or Heisenberg models, the two-spin
interaction is mapped to the four-spin interaction in the classical Ising model
of the (d + 1)-dimensional system [47].
On the other hand, the traveling salesman problem mapped to the Ising
system has to flip four spins at a time. The quantum fluctuation term should
P
be the four-spin interaction term as −Γ ijkl σix σjx σkx σlx . We have an eight-
spin interaction term by the Suzuki-Trotter decomposition of the Hamilto-
CHAPTER 4. APPLICATION TO TSP 66

nian with this term. Figure 4.4 shows the interactions in a decomposed
classical system of one-, two- and four-spin interactions of the quantum sys-
tem. Spins which contact through the gray domain in the figure interact
with each other. The computational cost is expected to increase when the
number of spins which interact each other increases. Moreover, various types
of spin-flip dynamics have to be introduced to recover ergodicity for such a
system.

i i j i j k l

Suzuki-Trotter decomposition

i i j i j k l

Figure 4.4: By the Suzuki-Trotter decomposition one-, two- and four-spin

interactions transform to interactions which contain spins twice as many —
two, four and eight spins.

To avoid such difficulties, we adopt the same quantum fluctuation we

P
used before as −Γ i σix . The same approach as in the previous chapter can
be applied. The system is mapped to a (d + 1)-dimensional system by the
Suzuki-Trotter decomposition. However, the spin-flip procedure of QA for
CHAPTER 4. APPLICATION TO TSP 67

TSP differs from the procedure of QA in the previous chapter. The four
spins in a Trotter slice are flipped at a time to keep the constraints, though
one spin was flipped so far. The spins are flipped by obeying the transition
probability which is obtained from the energy gap between the states before
flip and after flip. We computed the energy gap from the classical energy
term of the Hamiltonian (the tour length) and the interactions between the
Trotter slices for the four spins. That is shown in Fig. 4.5. We can perform
QA for the traveling salesman problem in this procedure.

σja

σjb
City

Trotter σia

σib
Stop

Figure 4.5: The four spins (big circles) are flipped at a time. The transition
probability is computed from the length of the tour and the two-spin inter-
actions of the nearest neighbors in the Trotter directions. (the interactions
between big and small circles)

4.4 Results of Monte Carlo Simulations

We perform SA and QA for the four cases of TSP whose optimal tours are
shown in Fig. 4.6. The first problem “random” is the case in which the cities
are located at random. Cities are located at random but clumped with dense
CHAPTER 4. APPLICATION TO TSP 68

and spares regions for the second problem “semi-random”. The optimal route
of the third problem has an “H” shape. (we call this problem “H-character”
hereafter.) The fourth problem is “ulysses16” of TSPLIB [48]. The number
of cities is N = 16 for all the problems.
√
The cities are located on a square with the side length N to make the
length of the tour extensive for “random”, “semi-random” and “H-character”.
For “ulysses16”, we re-scale dij and set the average to 2.2. The average,
the dispersion and the ratio of the dispersion and the average are shown in
Table 4.1.

Figure 4.6: Optimal tours of the four problems.

Simulations are performed with 500 Trotter slices for QA and 100 inde-
pendent runs for SA. We observe two quantities, the probability to find the
minimum-length P (t) and the average of length hLi. The probability P (t) is
obtained from the ratio of the number of the ground state configurations for
each time in 100 independent runs for SA and 500 Trotter slice for QA. The
√
probabilities of SA and QA for the four problems with the 10/ t scheduling
are plotted in Fig. 4.7 and the energy values are in Fig. 4.8. The plots are
CHAPTER 4. APPLICATION TO TSP 69

random semi-random H-character ulysses16

average (a) 2.195 2.642 2.155 2.200
dispersion (b) 0.973 1.152 0.867 1.556
b/a 0.443 0.436 0.402 0.707

Table 4.1: The average, the dispersion and the ratio of the dispersion and
the average for the four problems.

located in the same order as Fig. 4.6. In this schedule, QA has a better
performance than SA for all the four problems. Moreover, we also calculated
√ √
for the case of 5/ t and 2/ t schedules (see Fig. 4.9–4.12). Results show
that QA is better than SA in both of the average length and the probability
for all problems. The probabilities of SA for the “H-character” problem are
obviously saturated and fall short of probability one, and the probability of
QA almost reaches one. This is similar to the situation in the differential
equations for small-size systems for SA and QA.
Next, we investigate the asymptotic property of finding the optimal state.
The quantity 1−P (t) is plotted in Fig. 4.13 under the log-log scale. The curve
is proportional to 1/t in the asymptotic region. This implies that the system
almost follows the stationary state and the annealing schedule is sufficiently
slow. The reason is discussed in Sec. 2.4.
Lastly, we compare the annealed and quenched systems of SA and QA
to make the effect of the annealing clear. Annealing is introduced to avoid
slow relaxation in low or zero temperature in the system whose landscape is
complicated. If the annealing works, the length for the annealed system is
lower than the quenched system. The quenching parameter is the tempera-
ture T for SA and the transverse field Γ for QA. These values are quenched
√
to 0.1/ 10 = 0.0316 · · · . In the annealed system, the temperature and the
transverse field decrease from infinity to the same value as in the quenched
system. We calculate for the two problems, “random” and “ulysses16”. The
results of the probability and the average energy are shown in Fig. 4.14
and 4.15. We can list the four system in the order of the performance to
find the ground state: (1) the quantum system with annealing “QA”, (2)
 CHAPTER 4. APPLICATION TO TSP 70

1 1
SA SA
QA QA
0.8 0.8

0.6 0.6
P(t)

P(t)
0.4 0.4

0.2 0.2

0 0
10 100 1000 10000 100000 10 100 1000 10000 100000
t t
1 1
SA SA
QA QA
0.8 0.8

0.6 0.6
P(t)

P(t)
0.4 0.4

0.2 0.2

0 0
10 100 1000 10000 100000 10 100 1000 10000 100000
t t

Figure
√ 4.7: Probability to find the minimum-length. The scheduling is
10/ t.

32 40
SA SA
30 QA QA
28 35

26
30
24
<L>

<L>

22
25
20
18 20
16
14 15
10 100 1000 10000 100000 10 100 1000 10000 100000
t t
32 12000
SA SA
30 QA 11500 QA
11000
28
10500
26
10000
24
<L>

<L>

9500
22 9000
8500
20
8000
18
7500
16 7000
14 6500
10 100 1000 10000 100000 10 100 1000 10000 100000
t t

√
Figure 4.8: Average of the length. The scheduling is 10/ t.
 CHAPTER 4. APPLICATION TO TSP 71

1 1
SA SA
QA QA
0.8 0.8

0.6 0.6
P(t)

P(t)
0.4 0.4

0.2 0.2

0 0
10 100 1000 10000 10 100 1000 10000
t t
1 1
SA SA
QA QA
0.8 0.8

0.6 0.6
P(t)

P(t)
0.4 0.4

0.2 0.2

0 0
10 100 1000 10000 10 100 1000 10000
t t

Figure
√ 4.9: Probability to find the minimum-length. The scheduling is
5/ t.

30 32
SA SA
28 QA 30 QA

26 28

24 26
<L>

<L>

22 24

20 22

18 20

16 18

14 16
10 100 1000 10000 10 100 1000 10000
t t
30 11000
SA SA
28 QA 10500 QA

26 10000
9500
24
9000
<L>

<L>

22
8500
20
8000
18 7500
16 7000
14 6500
10 100 1000 10000 10 100 1000 10000
t t

√
Figure 4.10: Average of the length. The scheduling is 5/ t.
 CHAPTER 4. APPLICATION TO TSP 72

1 1
SA SA
QA QA
0.8 0.8

0.6 0.6
P(t)

P(t)
0.4 0.4

0.2 0.2

0 0
10 100 1000 10 100 1000
t t
1 1
SA SA
QA QA
0.8 0.8

0.6 0.6
P(t)

P(t)
0.4 0.4

0.2 0.2

0 0
10 100 1000 10 100 1000 10000 100000
t t

Figure
√ 4.11: Probability to find the minimum-length. The scheduling is
2/ t.

24 25
SA SA
23 QA 24 QA
22 23
21
22
20
<L>

<L>

21
19
20
18
17 19

16 18

15 17
10 100 1000 10 100 1000
t t
26 9500
SA SA
25 QA QA
24 9000
23
8500
22
<L>

21
<L>

8000
20
19
7500
18
17 7000
16
15 6500
10 100 1000 10 100 1000 10000 100000
t t

√
Figure 4.12: Average of the length. The scheduling is 2/ t.
 CHAPTER 4. APPLICATION TO TSP 73

1
QA
1/t
1−P(t)

0.1

0.01
100 1000 10000
t

Figure 4.13: The quantity

√ 1 − P (t) is plotted. The transverse field is
scheduled as Γ = 10/ t.

the quantum system without annealing “QA(quench)”, (3) the classical sys-
tem with annealing “SA” and (4) the classical system without annealing
“SA(quench)”. We can find that both of the annealed systems show better
performance than the quenched systems. It is notable that the probability
for the quenched quantum system “QA(quench)” is larger than SA. In spite
of annealing the temperature, SA finds the ground state less often than the
quenched quantum system. From these results, we obtain the following con-
clusions. The relaxation of the quantum systems is faster than the classical
systems whatever we control the parameters, annealing or quenching. The
annealing process accelerate the relaxations for both of the quantum and the
classical systems.
CHAPTER 4. APPLICATION TO TSP 74

4.5 Summary
We have applied QA to TSP by mapping the problem to the Ising system.
SA has a four-spin flip dynamics in the spin representation of the TSP. QA
should contain the quantum effect which flips four spins at a time. Because
of the difficulty of the calculation, we have adopted the one-spin dynamics
by applying the transverse field.
Numerical results show that QA has better performance than SA in the
probability to find the minimum-length of the closed tour and the average
of the tour length and obeys the 1/t-law asymptotically. We only calculated
the case in which the Trotter number is 500 for QA. The dependence on the
Trotter number was not studied, and we did not estimate how many Trotter
slices the quantum Monte Carlo simulation need. We also did not check
the actual performance of QA. QA needs the computational power Trotter-
number times as large as the calculation of SA, but we did not take into
account this disadvantage. We only calculated the quantities as a function
of the Monte Carlo steps, not the real computational time. In other words,
we cannot conclude which methods find the minimum-length tour of TSP
in small computer power, SA or QA, and only conclude so far that QA
can be applied to TSP and QA improves in terms of the Monte Carlo step
not the real computational time. These are the future problems. Another
future problem is to study the effect of the four-spin interaction term as a
quantum effect. It is not clear that the multi-spin interaction makes the
system converge to the optimal state faster or not.
If QA for TSP accelerates the search for the optimal solution actually,
QA may be applied to various problems which can be mapped to the Ising-
spin systems. We should have used a four-spin transition term as a quantum
effect in QA for TSP to have correspondence with the four-spin flip dynamics
P
of SA, but we adopt a one-spin transition term −Γ i σix . The results show
that QA finds the optimal solution and the one-spin transition term works.
This term forces neighboring Trotter slices to be in the same configuration
in the representation of Suzuki-Trotter decomposition. Assuming that this
effect is the essence of QA, we can import such an effect to non-Ising systems.
CHAPTER 4. APPLICATION TO TSP 75

For instance, it is possible to extend our method to the problems represented

in terms of Potts spins.
 CHAPTER 4. APPLICATION TO TSP 76

1
SA
SA(quench)
QA
0.8 QA(quench)

0.6
P(t)

0.4

0.2

0
10 100 1000 10000 100000
t
1
SA
SA(quench)
QA
0.8 QA(quench)

0.6
P(t)

0.4

0.2

0
10 100 1000 10000 100000
t

Figure 4.14: Difference in the probability between annealed and quenched

systems. Top and bottom figures are the results of “random” and “ulysses16”
respectively.
 CHAPTER 4. APPLICATION TO TSP 77

32
SA
30 SA(quench)
QA
28 QA(quench)
26
24
<L>

22
20
18
16
14
10 100 1000 10000 100000
t
12000
SA
11500 SA(quench)
11000 QA
QA(quench)
10500
10000
<L>

9500
9000
8500
8000
7500
7000
6500
10 100 1000 10000 100000
t

Figure 4.15: Difference in the average energy between annealed and

quenched systems. Top and bottom figures are the results of “random” and
“ulysses16” respectively.
Chapter 5

Summary

We have proposed a new idea on the general method to solve combinato-

rial optimization problems. The idea is quantum annealing (QA) in which
quantum tunneling effects cause transitions between states, in contrast to the
usual thermal transitions in simulated annealing (SA). First, we studied QA
by the time-dependent Schrödinger equation for small-size systems. Next, we
extended QA to the stochastic process, the quantum Monte Carlo method.
The two dynamics, the Schrödinger equation and the quantum Monte Carlo
dynamics, are not equivalent to each other. However, for the purpose of find-
ing the ground state, both of them work and the performance is improved
in comparison with SA. Finally, we presented the possibility of the imple-
mentation of QA for the general combinatorial optimization problems. The
test-bed was the traveling salesman problem (TSP). The chance of finding
the optimal solution by QA was larger than by SA.
The idea of QA is based on SA which was proposed by Kirkpatrick et
al. to find the optimal state of the optimization problems by introducing the
annealing process [21]. Annealing is performed by decreasing the temperature
of the system from a high temperature to a low temperature. For instance,
let us consider the silicon oxide SiO2 . If the decreasing rate is slow enough,
the system is in crystalline state. If the rate is fast, the glass state appears.
SA imitates this process by the Monte Carlo simulation. We consider QA by
replacing thermal fluctuations in SA with quantum fluctuations.
First, we adopt the transverse field as the quantum effect of the Ising

78
CHAPTER 5. SUMMARY 79

spin systems. The task is to find the ground state of the system without
the transverse field. The transverse field term is controlled and vanishes
in the infinite-lime limit so that the system approaches the ground state.
The dynamics of this system is not described by the Monte Carlo simulation
used for SA. The natural dynamics of the present system is provided by
the Schrödinger equation. We checked the performance of QA in finding
the ground state in comparison with SA by numerical calculations. In these
calculations, we applied the same function to the time-dependent schedules
of the temperature T and the amplitude of the transverse field Γ. For the
√
c/ t schedule, the behaviors of QA and SA were quite different. By QA we
find the ground state with probability one, while we do not find the ground
state by SA. We also found the remarkable property that the probability to
find the ground state converges to one asymptotically as 1/t. This property
suggests that the system follows the stationary (or almost stationary) state,
because the probability to find the system in the ground state is expressed
using the perturbation in terms of Γ(≪ 1) as
X 1
PQA (Γ) ∼ 1 − Γ2 (0) (0)
,
i6=0 (E0 − Ei )2

(0)
where Ei is the energy of the ith state of the non-perturbed (classical)
(0) √
system and E0 is the ground-state energy. When we put Γ = c/ t, the
1/t-law appears.
Secondly, we have considered the calculation for the large-size systems.
The calculation of the Schrödinger equation faces difficulties for larger N
because the number of states increases exponentially as 2N and the size of
the storage for the Hamiltonian reaches the storage limit of the computer.
The Monte Carlo method can be used to avoid this difficulty. There are two
ideas to replace the dynamics of the Schrödinger equation with the Monte
Carlo method. One is the path-integral Monte Carlo [49, 50] which regards
the Trotter direction as the time (more preciously, the imaginary time). The
other is the quantum Monte Carlo with “the Monte Carlo step”-dependent
interaction between Trotter slices. We regard the Monte Carlo step as the
time in the Schrödinger equation. From the similarity to SA, we adopted the
CHAPTER 5. SUMMARY 80

quantum Monte Carlo for large-size systems of QA.

We found that QA improves a performance in finding the ground state
even if we take into account that QA needs M times more calculations than
SA. The Trotter number we adopt is M = 100, which is large enough for
our calculations (for the two-dimensional EA model with N = 625). The
1/t convergence of the probability to find the ground state still appears in
the Monte Carlo calculation of QA. We also found that the relaxation of
the quantum system is faster than the classical system, if the amplitude of
the transverse field Γ in the quantum system and the temperature in the
classical system have the same value. This implies that the quantum effect
accelerates relaxations. We consider that this fact is the reason for the better
performance of QA than SA.
Lastly, we have applied QA to TSP by expressing the problem in terms of
the Ising system. Numerical results show that QA has a better performance
than SA in the probability to find the minimum-length of the closed tour and
the average of the tour length at the same Monte Carlo step. Moreover, the
1/t convergence of the probability to find the optimal tour still appears in
TSP by QA. These results imply that the QA works also in TSP. However,
we did not check the actual performance of QA. QA needs the computational
power Trotter-number times in comparison with SA, but we did not take into
account this disadvantage. We only calculated the quantities as a function
of the Monte Carlo step, not the real computational time. In other words,
we can not conclude which method finds the minimum-length tour of TSP
with smaller computer power, SA or QA, and only conclude so far that QA
can be applied to TSP and QA improves in terms of Monte Carlo steps, not
in the real computational time.
The findings in our study are summarized briefly as follows: QA provides
a general and faster algorithm to find the ground state of the Ising system
whose energy landscape is complicated. If we adopt the schedule of the
√
transverse field as Γ = c/ t, the asymptotic form of the probability to miss
the ground state is proportional to 1/t. The relaxation of the quantum
system is faster than the classical system under the condition of the control
parameter Γ = T . From the application side, the optimal schedule is not
CHAPTER 5. SUMMARY 81

known in general. If the relaxation is fast, we do not have to take into

account the possibility to be stuck in local minima more seriously in QA
than SA. QA may also work for the combinatorial optimization problems
which can be mapped to the Ising spin system.
The future problems are the followings: The analytical foundations of
the better performance of QA than SA and the critical rate of the decreasing
the transverse field are needed. (The same method of the analysis of SA
by Geman and Geman [22] may be applied to the analysis of the Trotter
decomposed system of QA.) The application to the study of the ground
state property of the Ising spin systems, especially the three-dimensional EA
model, is attractive. Whether the same picture as the SK model, the Parisi
picture [13, 18], can be applied to the three-dimensional EA model or not is
an open problem. For some problems in SA the multi-spin flip is performed
and the corresponding interaction term representing quantum effect in QA
is the multi-spin interaction term. For instance, SA for TSP needs four-
spin flip, but we adopt only a single-spin interaction term (the transverse
field). To study the effect of the multi-spin interaction term is needed. It is
not clear whether the multi-spin interaction makes the system converge to
the optimal state faster or not. On the other hand, the transverse field is
mapped to the interaction between the Trotter slices and the strength of the
interaction is controlled by the strength of the field. This interaction forces
neighbor Trotter slices to be in the same configuration in the representation
of Suzuki-Trotter decomposition. Assuming that this effect is the essence of
QA, we can import such an effect to a wide class of non-Ising systems. For
example, it is possible to extend our method to the problems represented
in terms of Potts spins. This is one possible way to apply QA to general
problems which can not be expressed in terms of the Ising spin.
Appendix A

Single-spin problem

In this Appendix we explain some technical aspects to derive the exact so-
lution of the time-dependent Schrödinger equation for the transverse Ising
√
model with a single spin. The three cases of Γ(t) = −ct, c/t and c/ t will
be discussed.

A.1 Case of Γ(t) = −ct (Landau-Zener model)

[33, 34, 35, 36, 37]
Let us express the solution of the Schrödinger equation at time t by the
parameters a = h+|ψ(t)i and b = h−|ψ(t)i. The Schrödinger equation (2.3)
with H = −hσ z −Γσ x is expressed as a set of first order differential equations
for a and b. It is convenient to change the variables as
1 1
ã = √ (a + b), b̃ = √ (a − b), (A.1)
2 2
by which the Schrödinger equation is now

d2 b̃(t)
+ (−ic + h2 + c2 t2 )b̃(t) = 0. (A.2)
dt2
By using the notation
√
z = 2ce−πi/4 t, (A.3)
ih2
λ = , (A.4)
2c

82
APPENDIX A. SINGLE-SPIN PROBLEM 83

we find

d2 b̃(t) 1 1
2
+ (λ + − z 2 )b̃(t) = 0. (A.5)
dz 2 4
√
The initial state is specified as a = b = 1/ 2 or b̃ = 0 as t → −∞. The
solution of (A.5) satisfying this condition is the parabolic cylinder function
D−λ−1 (−iz) [51]. Thus, we obtain the solution as
!
1 db̃(t)
ã(t) = −ctb̃(t) − i , (A.6)
h dt
b̃(t) = C1 D−λ−1 (iz), (A.7)

where C1 is a constant. To fix C1 , we use the condition

2
2C1 ceπh /8c
|ã(−∞)| = √ = 1. (A.8)
h 2c
Then we have
h 2
C1 = √ e−πh /8c . (A.9)
2c
The wave function of this system is given in Eq. (2.6).

A.2 Case of Γ(t) = c/t

We next consider the case of Γ(t) = c/t. By eliminating a from the Schrödinger
equation, we obtain

d2 b(t) 1 dΓ(t) db(t)

2
−
dt Γ(t) dt dt
 
2 2 ih dΓ(t)
+ h + Γ (t) − b(t) = 0. (A.10)
Γ(t) dt

Substituting Γ(t) = c/t, we have

 
d2 b(t) 1 db(t) 2 ih c2
− + h + + 2 b(t) = 0. (A.11)
dt2 t dt t t
APPENDIX A. SINGLE-SPIN PROBLEM 84

The solutions of this equation are expressed by the confluent P function [51]
 
 ∞
 z }| { 0 

P̃ ih 1 ic t 

 −ih 0 −ic 
 
 z ∞

}| {
0 

iht ic
= e t P̃ 0 1 + ic 0 −2iht  , (A.12)

 1 
ic −2ic

the right-hand side of which has two independent expressions in terms of the
confluent hypergeometric function

f (t) = eiht tic F (1 + ic, 1 + 2ic; −2iht), (A.13)

g(t) = eiht tic (−2iht)−2ic
×F (1 − ic, 1 − 2ic; −2iht). (A.14)

The general solution is b(t) = C1 f (t) + C2 g(t). Using the initial condition
1
b(0) = C1 + C2 = √ , (A.15)
2
1
a(0) = C1 − C2 = √ , (A.16)
2
we find
1
b(t) = √ eiht tic F (1 + ic, 1 + 2ic; −2iht). (A.17)
2

The asymptotic forms of b(t) and |b(t)|2 are then given as

√
2(2h)−ic Γ(2ic)
b(t) ∼
Γ(ic)
 −iht−πc/2
× e + ceiht+πc/2 (2ht)−1 , (A.18)

 
2 sinh(πc) −πc c cos(2ht) c2 eπc
|b(t)| ∼ e + + 22 . (A.19)
sinh(2πc) ht 4h t
APPENDIX A. SINGLE-SPIN PROBLEM 85
√
A.3 Case of Γ(t) = c/ t
√
The final solvable model has Γ(t) = c/ t. The Schrödinger equation (A.10)
is then expressed as
 
d2 b(t) 1 db(t) 2 2c2 + ih
− + h + b(t) = 0. (A.20)
dt2 2t dt 2t
The solution is the confluent P function [51]
 

 z ∞ 0 

}| {
P̃ 1
ih 2 − iγ 0 t 

 1 
−ih iγ 2
 

 z ∞ 0 

iht
}| {
= e P̃ 0 1
− iγ 0 −2iht  , (A.21)
 2
 1 
1 iγ 2

where γ = c2 /2h. The two independent solutions are thus [51]

1 1
f (t) = eiht F ( − iγ, ; −2iht), (A.22)
2 2
3
g(t) = eiht (−2iht)1/2 F (1 − iγ, ; −2iht). (A.23)
2
The general solution of (A.20) is therefore the linear combination of the above
two functions

b(t) = C1 f (t) + C2 g(t). (A.24)

The constants C1 and C2 are fixed by the requirement

1
b(0) = C1 = √ , (A.25)
2
√
h 1
a(0) = √ e(5/4)πi C2 = √ . (A.26)
2c 2
Substituting C1 and C2 into Eq. (A.24), we find
 
1 iht 1 1
b(t) = √ e F − iγ, ; −2iht
2 2 2
 
c (3/4)πi iht 1/2 3
+√ e e (−2iht) F 1 − iγ, ; −2iht . (A.27)
h 2
APPENDIX A. SINGLE-SPIN PROBLEM 86

The asymptotic form is

" (
(5/4)πi
√
)
√ −πc2 /4h 1 he
b(t) ∼ πe e−iht (2ht)−iγ √ +
2Γ( 12 − iγ) cΓ(−iγ)
( )#
−(1/4)πi
e c
+eiht (2ht)−1/2+iγ √ + √ . (A.28)
2Γ(iγ) 2 hΓ( 21 + iγ)

The probability |b(∞)|2 that the system remains in the excited state can be
calculated as the asymptotic form of (A.28) with the condition c2 /h ≫ 1
 2
2 πe−γπ  1 γ −1/2 e(5/4)πi 
|b(∞)| = + (A.29)
2  Γ( 21 − iγ) Γ(−iγ) 
  iγ 2
e−γπ  1/2 1 
+ γ −1/2 e(5/4)πi (−iγ)iγ+1/2  (A.30)

∼ e − iγ
4  2 
 
e−γπ  iγ i 
2 1 h2
∼ (−iγ) = = . (A.31)
4  8γ  256γ 64c4

A.4 Dependence of the final value on the ini-

tial condition
We show that we can choose the initial condition so that the final state is
√
the ground state when Γ = c/ t. This is not possible for Γ = −ct or c/t.
From Eq. (A.24), the asymptotic form of the solution as t → ∞ is
√ 2 2
πe−πc /4h−iht (2ht)−ic /2h
b(t) ∼ C1
Γ(1/2 − ic2 /2h)
√ 2 2
i πhe−πc /4h−iht (2ht)−ic /2h
+C2 . (A.32)
c2 Γ(−ic2 /2h)

The coefficients C1 and C2 are fixed under the conditions b(∞) = 0 and
|a(0)|2 + |b(0)|2 = 1 as
 −1/2
sinh(πc2 /h)
C1 = 1+ , (A.33)
2 sinh2 (πc2 /2h)
ic2 Γ(−ic2 /2h)
C2 = C1 . (A.34)
hΓ(1/2 − ic2 /2h)
APPENDIX A. SINGLE-SPIN PROBLEM 87

This solution is not the ground state of the Hamiltonian H(0).

The reason why one cannot obtain such a solution for the other schedules
(Γ = c/t, −ct) is the following: The general solution for Γ = c/t also has two
coefficients, and the initial state is represented as the linear combination of
two terms whose phases are indefinite:
 
a(0) = C1 tic t→0 − C2 t−ic t→0 , (A.35)
 
b(0) = C1 tic t→0 + C2 t−ic t→0 . (A.36)
√
The lowest-energy state at t = 0 corresponds to a(0) = b(0) = 1/ 2 (times
an arbitrary phase factor), which is realized by choosing C2 = 0 in Eqs.
(A.35) and (A.36). The indefiniteness of tic as t → 0 is irrelevant because
this is only the overall phase factor. Such a situation does not happen for
other values of a(0) and b(0), leading to a serious difficulty to determine the
wave function at t = 0. Thus we cannot choose an initial condition other
√
than a(0) = b(0) = 1/ 2. A similar fact exists in the case of Γ = −ct.
Bibliography

[1] E. L. Lawler and J. K. Lenstra and A. H. G. Rinnooy Kan and D. B.

Shmoys, The Traveling Salesman Problem: A guided Tuor of Combina-
torial Optimization (John Wiley & Sons, 1985).

[2] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide

to the Theory of NP-Completeness (Freeman, San Fransisco, 1979).

[3] A. V. Aho, J. E. Hopcroft and J. D. Ullman, The design and analysis

of computer algorithms (Addison-Wesley, 1974).

[4] V. Cannella and J. A. Mydosh, Phys. Rev. B 6 (1972) 4220.

[5] M. A. Ruderman and C. Kittel, Phys. Rev. 96 (1954) 99.

[6] T. Kasuya, Progr. Theoret. Phys. 16 (1954) 45.

[7] K. Yoshida, Phys. Rev 106 (1957) 893.

[8] S. F. Edwards and P. W. Anderson, J. Phys. F 5 (1975) 965.

[9] D. Sherrington and S. Kirkpatrick, Phys. Rev. Lett. 35 (1975) 1792.

[10] J. R. L. de Almeida and D. J. Thouless, J. Phys. A 11 (1978) 983.

[11] G. Parisi, Phys. Lett. 73 A (1979) 203.

[12] G. Parisi, J. Phys. A 13 (1980) L115, 1101 and 1187.

[13] G. Parisi, Phys. Rev. Lett. 50 (1983) 1946.

[14] D. J. Thouless, P. W. Anderson and R. G. Palmer, Phil. Mag. 35 (1977)

593.

88
BIBLIOGRAPHY 89

[15] F. Tanaka and S. F. Edowards, J. Phys. F 10 (1980) 2769.

[16] N. D. Mackenzie and A. P. Yong, J. Phys. C 16 (1983) 5321.

[17] K. Nemoto and H. Takayama, J. Phys. C 18 (1985) L529.

[18] C. De Dominicis and A. P. Yong, J. Phys. A 16 (1983) 2063.

[19] J. Hertz, A. Krogh and R. G. Palmer, Introduction to the Theory of

Neural Computation (Addison-Wesley, 1995).

[20] S. Shinomoto, Statisical mechanics in information theory (Maruzen,

1992) (in Japanese)

[21] S. Kirkpatrick, C. D. Gelett and M. P. Vecchi, Science 220 (1983) 671.

[22] S. Geman and D. Geman, IEEE Trans. PAMI 6 (1984) 721.

[23] C. Tsallis and D. A. Stariolo, Physica A 233 (1996) 395.

[24] T. J. P. Penna, Phys. Rev. E 51 (1995) R1.

[25] I. Andricioaei and J. E. Straub, Phys. Rev. E 53 (1996) R3055.

[26] H. Nishimori and J. Inoue, J. Phys. A 31 (1998) 5661.

[27] H. Nishimori and Y. Nonomura, J. Phys. Soc. Jpn. 65 (1996) 3780.

[28] K. Tanaka and T Horoguchi, IEICE A J80-A (1997) 2117 (in Japanese)

[29] B. K. Chakrabrati, A. Dutta and P. Sen, Quantum Ising Phases and

Transitions in Transverse Ising Models (Springer, Berlin, 1996).

[30] D. Lancaster and F. Ritort, J. Phys. A 30 (1997) L41.

[31] J. V. Alvarez and F. Ritort, J. Phys A 29 (1996) 7355.

[32] T. Sato, N. Hatano, M. Suzuki and H. Takayama (unpublished).

[33] S. Miyashita, Prog. Theor. Phys. 69 (1983) 714.

[34] C. Zener, Proc. R. Soc. London, Ser A 137 (1932) 696.

BIBLIOGRAPHY 90

[35] S. Miyashita, J. Phys. Soc. Jpn. 64 (1995) 3207.

[36] S. Miyashita, J. Phys. Soc. Jpn. 65 (1996) 2734.

[37] H. De Raedt, S. Miyashita, K. Saito, D. Gracı́a-Pablos and N. Gracı́a,

Phys. Rev. B 56 (1997) 11761.

[38] H. F. Trotter, Proc. Am. Math. Soc. 10 (1959) 545.

[39] M. Suzuki, Commum. Math. Phys. 51 (1976) 183.

[40] A. C. C. Coolen, S. N. Laughton and D. Sherrington, Phys. Rev. B 53

(1996) 8184.

[41] S. N. Laughton, A. C. C. Coolen and D. Sherrington, J. Phys. A 29

(1996) 763.

[42] H. Nishimori and M. Ymamna, J. Phys. Soc. Jpn.65 (1996) 3.

[43] M. Yamana, H. Nishimori, T. Kadowaki and D. Sherrington, J. Phys.

Soc. Jpn. 66 (1997) 1962.

[44] Wenhao Wu, et. al., Phys. Rev. Lett. 67 (1991) 2076.

[45] J. J. Hopfield and D. W. Tank, Biological Cybernetics 52(1985) 141.

[46] J. J. Hopfield and D. W. Tank, Science 233 (1986) 625.

[47] M. Suzuki, Porg. Theor. Phys. 56 (1976) 1454.

[48] TSPLIB is a library of sample instances for the TSP (and re-
lated problems) from various sources and of various types.
TSPLIB is available on the Internet. http://www.iwr.uni-
heidelberg.de/iwr/comopt/soft/TSPLIB95/TSPLIB.html.

[49] M. Creutz and B. Freedman, Ann. Phys 132 (1981) 427.

[50] J. Bonc̆a and J. E. Gubernatis, Phys. Rev. E 53 (1996) 6504.

[51] A. Erdélyi et al., Higher Transcendental Functions (MaGraw-Hill, New

York, 1953).