Siddhartha Bhattacharyya, Indrajit Pan, Ashish Mani, Sourav De, Elizabeth Behrman,

Susanta Chakraborti (Eds.)

Quantum Machine Learning
De Gruyter Frontiers in
Computational Intelligence

Edited by
Siddhartha Bhattacharyya

Volume 6
Quantum Machine
Learning

Edited by
Siddhartha Bhattacharyya, Indrajit Pan, Ashish Mani,
Sourav De, Elizabeth Behrman, Susanta Chakraborti
Editors
Prof. (Dr.) Siddhartha Bhattacharyya Dr. Sourav De
CHRIST (Deemed to be University) Cooch Behar Government Engineering College
Hosur Road, Bhavani Nagar, S. G. Palya, Post Ghughumari, Harinchawara
Bangalore 736170 Cooch Behar, West Bengal, India
560 029 Karnataka, India [email protected]
[email protected]
Prof. (Dr.) Elizabeth Behrman
Dr. Indrajit Pan Wichita State University
RCC Institute of Information Technology 1845 Fairmount St.
Canal South Road, Beliaghata 67260 Wichita, USA
700 015 Kolkata, India [email protected]
[email protected]
Prof. (Dr.) Susanta Chakraborti
Prof. (Dr.) Ashish Mani Indian Institute of Engineering, Science
Amity University Uttar Pradesh, and Technology
Sector-125, Noida, UP, India Botanic Garden
[email protected] 711103 Shibpur, West Bengal, India
[email protected]

ISBN 978-3-11-067064-6
e-ISBN (PDF) 978-3-11-067070-7
e-ISBN (EPUB) 978-3-11-067072-1
ISSN 2512-8868

Library of Congress Control Number: 2020936493

Bibliographic information published by the Deutsche Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data are available on the Internet at http://dnb.dnb.de.

© 2020 Walter de Gruyter GmbH, Berlin/Boston

Cover image: shulz/E+/getty images
Typesetting: Integra Software Services Pvt. Ltd.
Printing and binding: CPI books GmbH, Leck

www.degruyter.com
Siddhartha Bhattacharyya would like to dedicate this book to the memory of the
deceased Jawans in Pulawama incident.

Indrajit Pan would like to dedicate this book to his research guides Prof. (Dr.) Hafizur
Rahaman and Dr. Tuhina Samanta, who have always mentored him during his doc-
toral research.

Ashish Mani would like to dedicate this book to Most Revered Prof. Prem Saran
Satsangi, Emeritus Chairman, Advisory Committee on Education, Dayalbagh
Educational Institutions, India.

Sourav De would like to dedicate this book to his loving wife Debolina Ghosh and
beloved son Aishik De.

Elizabeth Behrman would like to dedicate this book to EJB, JFB, and JPK; and to the
memory of CFB.

Susanta Chakraborti would like to dedicate this book to his father Late Santosh
Kr. Chakraborty.
Contents
List of Contributors IX

Preface XI

Sandip Dey, Sourav De, and Siddhartha Bhattacharyya

1 Introduction to quantum machine learning 1

Bruce J. MacLennan
2 Topographic representation for quantum machine learning 11

Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

3 Quantum optimization for machine learning 39

Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

4 From classical to quantum machine learning 67

Alokananda Dey, Sandip Dey, Siddhartha Bhattacharyya, Jan Platos,

and Vaclav Snasel
5 Quantum inspired automatic clustering algorithms: A comparative study
of Genetic algorithm and Bat algorithm 89

Siddhartha Bhattacharyya
6 Conclusion 115

Index 117
List of Contributors
Siddhartha Bhattacharyya Sandip Dey
Department of Computer Science and Department of Computer Science
Engineering Sukanta Mahavidyalaya, Sukanta Nagar
CHRIST (Deemed to be University) Dhupguri, Jalpaiguri, West Bengal, India
Bangalore, India [email protected]
[email protected]
Bruce J. MacLennan
Arit Kumar Bishwas University of Tennessee, Knoxville
AIIT, Amity University, Noida Tennessee, USA
Uttar Pradesh, India [email protected]
[email protected]
Ashish Mani
Amlan Chatterjee ASET, Amity University UP
Computer Science Noida, Uttar Pradesh, India
California State University [email protected]
Dominguez Hills, USA
[email protected] Vasile Palade
Faculty of Engineering and Computing
Sourav De Coventry University
Department of Computer Science and Coventry, United Kingdom
Engineering [email protected]
Cooch Behar Government Engineering
College Jan Platos
Cooch Behar, West Bengal, India VSB Technical University of Ostrava
[email protected] Czech Republic
[email protected]
Alokananda Dey
RCC Institute of Information Technology Vaclav Snasel
Kolkata, West Bengal, India VSB Technical University of Ostrava
[email protected] Czech Republic
[email protected]

https://doi.org/10.1515/9783110670707-203
Preface
Imparting intelligence to the machines has always been a challenging thorough-
fare. Over the years, several intelligent tools have been invented or proposed to
deal with the uncertainties encountered by human beings with the advent of the
soft computing paradigm. However, it has been observed that even the soft comput-
ing tools often fall short in offering a reliable and reasonable solution in real time.
Hence, scientists employed hybrid intelligent techniques using the combination of
several soft computing tools to overcome the shortcomings.
Quantum computing has evolved from the studies of Feynman and Deutsche
who evolved efficient searching techniques in the quantum domain. These search-
ing techniques outperform the classical techniques both in terms of time and space.
Inspired by this, researchers are on the spree for conjoining the existing soft com-
puting tools with the quantum computing paradigm to evolve more robust and time
efficient intelligent algorithms. The resultant algorithms are immensely useful for
solving several scientific and engineering problems, which includes data process-
ing and analysis, machine vision, social networks, big data analytics, flow shop
scheduling problems to name a few.
Quantum machine learning is an emerging interdisciplinary research area which
resorts to the principles of quantum physics applied to machine learning. Quantum
machine learning algorithms helps to improve classical methods of machine learning
by taking the advantages offered by quantum computation. Given the inherent paral-
lelism offered due to the features of quantum computing, researchers have evolved
different intelligent tools and techniques which are more robust and efficient in
performance.
Quantum-enhanced machine learning refers to quantum algorithms that solve
tasks in machine learning, thereby improving a classical machine learning method.
Such algorithms typically require one to encode the given classical dataset into a
quantum computer, so as to make it accessible for quantum information processing.
After this, quantum information processing routines can be applied and the result of
the quantum computation is read out by measuring the quantum system. For exam-
ple, the outcome of the measurement of a qubit could reveal the result of a binary
classification task. While many proposals of quantum machine learning algorithms
are still purely theoretical and require a full-scale universal quantum computer to be
tested, others have been implemented on small-scale or special purpose quantum
devices.
This book comprises six well versed chapters from leading quantum machine
learning researchers.
Chapter 1 provides an overview of the basic concepts and principles pertaining to
quantum machine learning. Apart from throwing light on different aspects of quantum
algorithms, the chapter also provides a bird’s eye view on the principles of quantum
reinforcement learning and quantum annealing. The evolution of quantum neural

https://doi.org/10.1515/9783110670707-204
XII Preface

networks with special mention to the pioneering works in this direction is also touched
upon to enlighten the readers.
One of the most common information representations in the brain is the topo-
graphic or computational map, in which neurons are arranged systematically accord-
ing to the values they represent. By representing quantitative relationships spatially,
computational maps enable the brain to compute complex, nonlinear functions to
the accuracy required. Chapter 2 proposes two approaches to quantum computation
for machine learning by means of topographic representation. It shows how to con-
struct unitary operators, implementable on quantum computers, that implement arbi-
trary (including nonlinear) functions via computational maps.
Training in machine learning techniques often requires solving a difficult optimi-
zation problem, which is the most expensive step in the entire mode-building process
and its applications. One of the possible solutions in near future for reducing execu-
tion time of training process in Machine learning techniques is to implement them on
quantum computers instead of classical computers. Chapter 3 discusses a global opti-
mization technique based on Adiabatic Quantum Computation (AQC) to solve mini-
mization of loss function without any restriction on its structure and the underlying
model, which is being learned. Further, it is also shown that in the proposed frame-
work AQC based approach would be superior to circuit-based approach in solving
global optimization problems.
In Chapter 4, the authors discuss the transition from classical machine learning to
quantum machine learning (QML) and explore the recent progress in this domain.
QML is not only associated with the development of high-performance machine learn-
ing algorithms that can run on a quantum computer with significant performance im-
provements but also has a very diverse meaning in other aspects. The chapter tries to
touch those aspects in brief too, but the main focus is on the advancements in the field
of developing machine learning algorithms that will run on a quantum computer.
Chapter 5 is intended to present two automatic clustering techniques of image
datasets, based on quantum inspired framework with two different meta-heuristic
algorithms, viz., Genetic Algorithm (GA) and Bat Algorithm (BA). This work pro-
vides two novel techniques to automatically identify the optimal number of clusters
present in an image dataset and also provides a comparative study between the
Quantum Inspired Genetic Algorithm (QIGA) and Quantum Inspired Bat Algorithm
(QIBA). A comparison is also presented between this quantum inspired algorithms
with their classical counterparts. During the experiment, it is observed that the
quantum inspired techniques outperform over their classical counterparts. The
comparison is prepared based on the mean values of the fitness, standard devia-
tion, standard error of the computed fitness of the cluster validity index and the
optimal computational time. Finally, the superiority of the algorithms is verified in
terms of the p-value which was computed from the statistical superiority test (t-test)
and ranking of the proposed procedures was produced by the Friedman test.
During the computation, the betterment of the fitness was judge by a well-known
 Preface XIII

cluster validity index, named, DB index. The experiments are carried out on four
Berkeley image and two real life grey scale image datasets.
Chapter 6 draws a line of conclusion discussing the achievable from the book.
The chapter also throws light on the future trends of quantum machine learning
involving multilevel quantum systems.
The editors feel that this book would come in good stead to the undergraduate
and postgraduate students of computer science, information science and electron-
ics engineering for a part of their curricula. The editors would also like to take this
opportunity to render their heartfelt gratitude to De Gruyter publishing house for
consenting to publish this book.
Bangalore, Kolkata, New Delhi, Kansas Siddhartha Bhattacharyya
November, 2019 Indrajit Pan
Ashish Mani
Sourav De
Elizabeth Behrman
Susanta Chakrabarti
Sandip Dey, Sourav De, and Siddhartha Bhattacharyya
1 Introduction to quantum machine learning
Abstract: Quantum Machine Learning (QML) is popularly known to be an integrative
approach to learning of the Quantum Physics (QP) and Machine Learning (ML). In
this chapter, an outline of the fundamental ideas and features related to quantum
machine learning is laid out. The different facets of quantum algorithms are dis-
cussed in this chapter. In addition to this, the basic features of quantum reinforce-
ment learning and quantum annealing are also provided in this chapter. Finally, the
chapter deliberates about the advancement of quantum neural networks to through
light in the direction of QML.

Keywords: machine learning, Grover’s Search Algorithm, reinforcement learning,

quantum annealing, quantum neural networks

1.1 Quantum machine learning

Machine learning has become an emerging discipline in the recent technologies. It
can be used in a variety of fields, such as computational biology, computer vision,
computer security and many others. Data analysis (DA) is an equally important part
in the modern industry. The ML and DA analyze data by applying statistical meth-
ods and they provide computers learning capabilities on the basis of the analysis of
observed data. On the basis of learning style, Machine learning algorithms (MAA)
are basically classified into different groups, viz., supervised learning (SL), unsu-
pervised learning (UL) and semi-supervised learning (SSL). The foremost shortcom-
ings of using ML techniques are generally known to be computational time and
storage, especially which involve bulky amounts of data. In addition to these, when
the current deep learning algorithms are used, the training time can turn out to be
even longer. In the subsequent generation, researchers got a viable alternative by
utilizing the power of quantum computers method which can be smarter enough to
reduce the storage and computational time as mentioned above.

Sandip Dey, Department of Computer Science, Sukanta Mahavidyalaya, Jalpaiguri, West Bengal,
India
Sourav De, Department of Computer Science & Engineering, Cooch Behar Government
Engineering College, Cooch Behar, West Bengal, India
Siddhartha Bhattacharyya, Department of Computer Science and Engineering, CHRIST (Deemed
to be University), Bangalore, India

https://doi.org/10.1515/9783110670707-001
2 Sandip Dey, Sourav De, and Siddhartha Bhattacharyya

Many ML problems use linear algebra to figure out matrix operations by ex-
pressing the data in matrices. Quantum computing (QC) is able to make several
linear algebra computations more fast, which unreservedly improve classical ML
tasks [1, 2]. Numerical methods for optimization are well admired field of research,
which intend to improve the computations of the said optimization procedures.
Like classical optimization, quantum optimization, a branch of QC, attempts to
improve the said techniques even further. Two renowned methods of this kind are
Quantum Gradient Descent (QGD) [3] and Quantum Approximate Optimization
Algorithm (QAOA) [4]. These methods are efficiently applied in quantum neural
networks (QNN) like Quantum Boltzman Machines [5].
Of late, the quantum machine learning, a very new field of emerging inter-
disciplinary research area, has come out, with the objective of coalescing the the-
ory of machine learning with the properties of quantum computing. Quantum ma-
chine learning (QML) is intended for implementing machine learning algorithms
(MLLs) in quantum atmosphere (systems) [6], by applying quantum features such
as quantum superposition and quantum entanglement for solving different prob-
lems with high efficacy. Basically, QML is known to be a sub-discipline of quantum
research of information processing, with the objective of introducing quantum al-
gorithms that generally learn from data with the aim of improving existing ap-
proaches in ML. Hence, the purpose is to develop quantum applications of various
MLLs, having the influence of quantum computers as well as the flexibility and
learning ability of MLLs.
Several quantum algorithms have been introduced for a variety of machine
learning models which include neural networks (NN), graphical models, support
vector machines (SVM) so on and so forth. QML explores more basic questions
about the thought of learning from the quantum perspective. In some occasions,
the QML is extensively defined by researchers to apply ML to quantum information.
Besides the implicit methodologies used in QML, there exists few quantum ver-
sions of classical ML algorithms. Quantum Support Vector Machines (QSVM) [1],
which is basically applied for linear classification, is a popular example of this type.
In addition to that, Quantum Principle Component Analysis (QPCA) [7], a popular ap-
proach for dimensionality reduction, Quantum Guassian Mixture Models [8], another
renowned approach for estimating clustering and density. An emerging sub-discipline
of ML is called Deep Learning (DL). These days, quantum computers are employed for
DL applications, which involve significant storage and time. Some popular examples
of these applications are Quantum Boltzmann Machines [5], Quantum Generative
Adversarial Networks [9], Quantum Convolutional Neural Networks [10] and Quantum
Variational Autoencoders [11]. Additionally, a more enlightened field inside ML is
known as Reinforcement Learning (RL). RL can be described as learning as time con-
tinues by exploring the environment.
 1 Introduction to quantum machine learning 3

1.2 Grover search algorithm

Grover’s Search Algorithm (GAA) is a very popular quantum search algorithm [12].
This algorithm finds a group of elements which satisfy a certain condition. For per-
forming this task, a black box, usually known as an oracle, having the capability of
identifying the elements, must satisfy the criteria that are required to find. Let us
suppose that the group has N elements, the oracle is called O(M) times for classical
computations to get all elements that satisfy the above-mentioned criteria. Using
quantum mechanics, this algorithm is able to attain the same result by using
pffiffiffiffiffiffi
Oð MÞ calls to the oracle.
The algorithm has the capability to make several calls to the oracle simultaneously
by exploiting a special feature of quantum computing, called parallel processing. Let
us assume that a search list has N number of elements. GAA uses a Hilbert space of
dimension N to represent them, which can be achieved with n = logM (base 2) qubits.
Each el2N with index y is denoted by an orthonormal vector, called jyi in the qubits’s
state space. The objective is to identify the index z of that particular element satisfying
the given search criteria.
At the outset, the oracle is being employed, in which a unitary operation having
features given by

1. if y = z, Uz jyi = − jyi (1:1a)

2. if y ≠ z, Uz jyi = jyi (1:1b)

The eqs. (1.1a) and (1.1b) can also be expressed as follows:

Uz = I − 2jzihzj (1:2)

where, I denotes the identity operator. The GAA simultaneously uses the last oracle
operator and the Grover diffusion operators, as defined by

Up = 2jpihpj − I (1:3)
P
where, jpi = p1ffiffiNffi Ny= 1 y. The eq. (1.2) and (1.3) are utilized as follows.
At initial, the qubits’ state is being initialized to the jpi state. Thereafter, Uz and Up
are successively applied iteratively for r(N) number of times. Afterwards, the system is
assessed, which provides the eigenvalue ðλz Þ, which may conclude the said index ð zÞ.

1.3 Quantum reinforcement learning

Besides supervised and unsupervised learning, reinforcement learning (RL) is also a
popular category of learning method. In contrast to SL and UL, RL employs a scalar
4 Sandip Dey, Sourav De, and Siddhartha Bhattacharyya

value called reward to assess the input-output pairs and uses the policy of trial-
and error to interact with the environment for learning a mapping from states to
actions. Since the year of 1980, RL has gradually become a significant approach
to ML [5]. So far, it has been extensively applied in artificial intelligence (AI), par-
ticularly in robotics [13–17]. It is due to the fact that it shows an excellent perfor-
mance in on-line adaptation, and in addition, it has a prevailing learning ability
to any complex nonlinear systems.
There may have some complicated problems while dealing with practical appli-
cations, such as the exploration strategy, slow learning speed; in particular han-
dling complex problems. It can be observed, especially whilst the state-action space
grows to be gigantic and the number of parameters that are to be learned grows up
exponentially with the increase of dimension. So far, several methods have been
introduced in the recent years to combat this situation.
Different learning paradigms are united for optimizing RL. Smith [18] pro-
posed a novel model to represent and generalize in model-less RL on the basis of
self-organizing map (SOM) and benchmark Q-learning. In addition, for the sake of
adaptation of fuzzy inference systems with Watkins’ Q-learning for problems hav-
ing large/continuous state-action spaces is also presented [17, 19, 20]. Different RL
methods have been improved in practice [21–24].
Two significant quantum algorithms, called the Shor algorithm [25, 26] and the
Grover algorithm [27, 28], have been introduced. Rigatos and Tzafestas [29] applied
the theory of quantum computing to get the benefit of parallelization of fuzzy-logic
control algorithm (FCA), whose objective is to speed up the fuzzy inference. Quantum-
inspired evolutionary algorithms (QIEA) have been developed to enhance the perfor-
mance of the existing evolutionary algorithms (EA) [30]. Later, Hogg and Portnov [31]
introduced a quantum algorithm to solve combinatorial optimization problem having
over constrained satisfiability and asymmetric traveling salesman. In recent times, the
quantum search method has been applied to dynamic programming [32]. Taking into
account the spirit of computation, Dong et al. [33] have developed the concept of
Quantum Reinforcement Learning (QRL) inspired by the basic concept of quantum
computing, called state superposition principle and parallelism. QRL was developed
to speed-up learning and attaining a trade-off between exploitation and exploration of
RL in the course of simulated experiments.

1.4 Quantum annealing

In statistical mechanics, quantum annealing (QA) that is applied in the field of
quantum-mechanical fluctuations is the quantum version of the simulated anneal-
ing (SA) [34, 35]. It is also known as Quantum Stochastic Optimization algorithm.
Like the SA, quantum annealing is also successfully applied to solve the hard
 1 Introduction to quantum machine learning 5

optimization problem. It works on the principal of the quantum adiabatic evolution in

the active field of the quantum computation and it is a profitable hybridization be-
tween classical and quantum technology. QA is applied in different fields as computer
science [36], machine learning [37, 38], graph theory [39], communications [40–42],
finance [43], aeronautics [44] and any others real world problems.
In simulated annealing [45], a temperature dependent random walk of a statisti-
cal-mechanical system is employed as the cost function of a given optimization prob-
lem. This function determines the potential energy profile of the solution space and
thermal fluctuations avoid that the exploration gets stuck in a local minimum [46]. It
is expected that the system will stay close to thermal equilibrium during time evolu-
tion. It can happen if the rate of decrease of temperature is sufficiently slow, and
thus lead in the end to the zero-temperature equilibrium state, the lowest-energy
state [47]. Due to the general applicability, reasonable performance, and relatively
easy implementation in most cases, SA is employed effectively in many real life appli-
cations. When a problem requires an infinitely long time to derive the exact solution
by keeping the system close to thermal equilibrium, SA is applied to obtain the
approximate solution within a measurable computation time. The quantum fluctua-
tions are induced in QA by introducing artificial degrees of quantum nature, non-
commutative operators. The strength of the quantum fluctuations are controlled to
reach the ground state by slowly reducing the temperature.
Like the finding the minimization of a cost function of a optimization problem, it
can be considered as finding the ground state of a classical Ising Hamiltonian H0 [48].
Different types of practical problems have cost functions with large number of local
minima. Similarly, Ising Hamiltonians are remindful of classical spin glasses [49, 50].
For these types of characteristics, it is very tough to find the global minima for classi-
cal algorithms. This problem can be overcame with the idea to elevate the classical
Ising Hamiltonian H0 to the quantum domain. Based on the adiabatic theorem of
quantum mechanics, the ground state of the classical Ising model can be derived by
formatting the system in the ground state of some initial Hamiltonian H1 [51, 52]. It is
easy to develop both theoretically and experimentally. H1 is selected in such a way it
does not exchange with H0. The Hamiltonian changes gradually from H1 to H0 as the
system parameters modified sufficiently slowly.

1.5 Quantum neural networks

Quantum neural networks (QNNs) are incarnations of neural network models en-
tailing the principles of quantum mechanics. From the computational point of
view [53, 54], one class of quantum neural network combine artificial neural net-
work models and embed the features of quantum computing inside to evolve
more robust and efficient models. The basic philosophy behind these efforts is to
6 Sandip Dey, Sourav De, and Siddhartha Bhattacharyya

circumvent the limitations of classical neural networks in handling big data by resort-
ing to the features of quantum parallelism, interference and entanglement. However,
most of the QNN models try to replace the classical binary or the McCulloch-Pitts
neurons with qubits (also called “qurons”) manifested with quantum mechanical
principles.
In 1995, Subhash Kak [55] and Ron Chrisley [56] put forward the idea of a
quantum neural model by establishing the similarity of the neural activation
function with the quantum mechanical Eigenvalue equation. Ajit Narayanan and
Tammy Menneer introduced a photonic implementation of a quantum neural net-
work using the many-universe theory which collapses into the desired states on
application of quantum measurement [57]. Since then, a lot of efforts have been
invested to find the quantum version of the perceptron. However, the advances
in this direction was impeded due to the fact that the characteristic neural non-
linear activation functions seldom follow the mathematical structure of quantum
theory due to the inherent linear operations in a quantum system. After much
effort, Schuld, Sinayskiy and Petruccione used the quantum phase estimation al-
gorithm [58] to implement the activation function. Apart from this, lots of quan-
tum-inspired models have come up to implement a fuzzy logic based neural
network [59].
Elizabeth Behrman and Jim Steck [60] proposed a novel quantum computing
setup comprising a number of qubits with tunable mutual interactions. In their
model, the interaction strengths are updated using a training set of desired input-
output relations following the classical back-propagation algorithm thereby en-
abling the quantum network to learn an algorithm.
The quantum associative memory was introduced by Dan Ventura and Tony
Martinez in 1999 [61]. The authors proposed an algorithm to emulate an associative
memory for a circuit-based quantum computer. In this algorithm, the memory
states are envisaged as a superposition of quantum states. A quantum search algo-
rithm is then used to retrieve the memory state closest to a given input. This emula-
tion promises an exponential storage capacity of memory states.

1.6 Conclusion
This chapter provides an overview of the basic concepts and principles pertaining
to quantum machine learning. Apart from throwing light on different aspects of
quantum algorithms, the chapter also provides a bird’s eye view on the principles
of quantum reinforcement learning and quantum annealing. The evolution of quan-
tum neural networks with special mention to the pioneering works in this direction
is also touched upon to enlighten the readers.
 1 Introduction to quantum machine learning 7

References
[1] Harrow, A. W., Hassidim, A., & Lloyd, S. Quantum algorithm for solving linear systems of
equations. Physical review letters. 2009, 15(103), 150502.
[2] Rebentrost, P., Mohseni, M., & Lloyd, S. Quantum support vector machine for big data
classification. Physical review letters. 2014, 113(13), p. 130503.
[3] Kerenidis, I., & Prakash, A. Quantum gradient descent for linear systems and least squares.
2017. arXiv:1704.04992 [quant-ph].
[4] Farhi, E., Goldstone, J., & and Gutmann, S. A quantum approximate optimization algorithm.
2014. arXiv:1411.4028 [quant-ph].
[5] Amin, M. H., Andriyash, E., Rolfe, J., Kulchytskyy, B., & Melko, R. Quantum boltzmann
machine. Physical review letters. 2018, 8, 021050.
[6] Schuld, M., Sinayskiy, I., & Petruccione, F. An introduction to quantum machine learning.
Contemporary Physics. 2015, 56(2),172–185.
[7] Lloyd, S., Mohseni, M., & Rebentrost, P. Quantum principal component analysis. Nature
Physics. 2014, 10(9),631–633.
[8] Rahman M., & and Geiger, D. Quantum clustering and gaussian mixtures. 2016.
arXiv:1612.09199v1 [stat.ML].
[9] Lloyd, S., & Weedbrook, C. Quantum generative adversarial learning. Physical review letters,
2018, 040502.
[10] Cong, I., Choi, S., & Lukin, M. D. Quantum convolutional neural networks. 2018.
arXiv:1810.03787v1 [quant-ph].
[11] Khoshaman, A., Vinci, W., Denis, B., Andriyash, E., & Amin, M. H. Quantum variational
autoencoder. 2018. arXiv:1802.05779v1 [quant-ph].
[12] Lavor, C., Manssur, L. R. U., & Portugal, R. Grover’s Algorithm: Quantum Database Search.
2003. arXiv:quant-ph/0301079.
[13] Beom, H. R., & Cho, H. S. A sensor-based navigation for a mobile robot using fuzzy logic and
reinforcement learning. IEEE Transactions on Systems, Man, and Cybernetics. 1995, 25(3),
464–477.
[14] Connolly, C. I. Harmonic functions and collision probabilities. Int. J. Rob. Res. 1997, 16(4),
497–507.
[15] Smart, W. D., & Kaelbling, L. P. Effective reinforcement learning for mobile robots.
Proceedings – IEEE International Conference on Robotics and Automation. 2002, 3404–3410.
[16] Kondo,T., & Ito, K. A reinforcement learning with evolutionary state recruitment strategy for
autonomous mobile robots control. Robotics and Autonomous Systems. 2004, 46(2),111–124.
[17] Tzafestas, S. G., & Rigatos, G. G. Fuzzy reinforcement learning control for compliance tasks of
robotic manipulators. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics).
2002, 32(1),107–113.
[18] Smith, A. J. Applications of the self-organising map to reinforcement learning. Neural Netw.
2002, 15(8/9), 1107–1124.
[19] Glorennec, P. Y, & Jouffe, L. Fuzzy Q-learning. Proceedings of 6th International Fuzzy Systems
Conference. 1997, 659–662.
[20] Er, M. J, & Deng, C. Online tuning of fuzzy inference systems using dynamic fuzzy Q-learning.
IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics). 2004, 34(3),
1478–1489.
[21] Chen, C. L., Li, H. X., & Dong, D. Y. Hybrid control for Robot Navigation: A hierarchical
Q-learning algorithm. IEEE Robotics & Automation Magazine. 2008, 15(2),37–47.
[22] Whiteson, S., & Stone, P. Evolutionary function approximation for reinforcement learning.
Journal of Machine Learning Research. 2006, 7, 877–917.
8 Sandip Dey, Sourav De, and Siddhartha Bhattacharyya

[23] Kaya, M., & Alhajj, R. A novel approach to multiagent reinforcement learning: Utilizing OLAP
mining in the learning process. IEEE Transactions on Systems, Man, and Cybernetics,
Part C (Applications and Reviews). 2005, 35(4),582–590.
[24] Vengerov, D., Bambos, N., & Berenji, H. A fuzzy reinforcement learning approach to power
control in wireless transmitters IEEE Transactions on Systems, Man, and Cybernetics,
Part B (Cybernetics). 2005, 35(4),768–778.
[25] Shor, P. W. Algorithms for quantum computation: Discrete logarithms and factoring.
Proceedings 35th Annual Symposium on Foundations of Computer Science. 1994, 124–134.
[26] Ekert, A., Jozsa, R. Quantum computation and Shor’s factoring algorithm. Rev. Mod. Phys.
1996, 68(3),733–753.
[27] Grover, L. K. A fast quantum mechanical algorithm for database search. Proceedings of the
twenty-eighth annual ACM symposium on Theory of Computing. 1996, 212–219.
[28] Grover, L. K. Quantum mechanics helps in searching for a needle in a haystack. Phys. Rev.
Lett. 1997, 79(2),325–327.
[29] Rigatos, G. G., & Tzafestas, S. G. Parallelization of a fuzzy control algorithm using quantum
computation. IEEE Trans. Fuzzy Syst. 2002, 10(4),451–460.
[30] Sahin, M., Atav, U., & Tomak, M. Quantum genetic algorithm method in self-consistent
electronic structure calculations of a quantum dot with many electrons International Journal
of Modern Physics C. 2005, 16(9),1379–1393.
[31] Hogg, T., & Portnov, D. Quantum optimization. Information Sciences. 2000, 128(3),81–197.
[32] Naguleswaran, S., White, L., & Fuss, I. Quantum search in stochastic planning. Proceedings
of the Sixteenth International Conference on Automated Planning and Scheduling. 2005,
34–45.
[33] Dong, D. Y., Chen, C. L., & Chen, Z. H. Quantum reinforcement learning. Proceedings of the
1st International Conference on Natural Computing. 2005, 3611, 686–689.
[34] Matsuda, Y., Nishimori, H., & Katzgraber, H. G. Ground-state statistics from annealing
algorithms: Quantum vs classical approaches. 2009. arXiv:0808.0365.
[35] Santoro, G. E., Martonak, R., Tosatti, E., & Car, R. Theory of quantum annealing of an Ising
spin glass. Science. 2002, 295(5564),2427–2430.
[36] Choi, V. Adiabatic Quantum Algorithms for the NP-Complete Maximum-Weight Independent
Set. Exact Cover and 3SAT Problems. 2010. http://arxiv.org/abs/1004.2226.
[37] Adachi S. H., & Henderson, M. P. Application of Quantum Annealing to Training of Deep
Neural Networks. 2015. https://arxiv.org/abs/1510.06356.
[38] Amin, M. H., Andriyash, E., Rolfe, J., Kulchytskyy, B., & Melko, R. Quantum Boltzmann
Machine. 2016. https://arxiv.org/abs/1601.02036.
[39] Vinci, W., Markström, K., Boixo, S., Roy, A., Spedalieri, F. M., Warburton, P. A., & Severini,
S. Hearing the Shape of the Ising Model with a Programmable Superconducting-Flux
Annealer. Scientific Reports. 2014, 4(5703). doi:10.1038/srep05703.
[40] Chancellor, N., Szoke, S., Vinci, W., Aeppli, G., & Warburton, P. A. Maximum-Entropy
Inference with a Programmable Annealer. Scientific Reports. 2016, 6(22318).
doi: 10.1038/srep22318.
[41] Otsubo, Y., Inoue, J. I., Nagata, K., & Okada, M. Effect of quantum fluctuation in error-
correcting codes. Physical Review E. 2012, 86(051138). doi:10.1103/PhysRevE.86.051138.
[42] Otsubo, Y., Inoue, J. I., Nagata, K., & Okada, M. Code-division multiple-access multiuser
demodulator by using quantum fluctuations. Physical Review E. 2014, 90(012126).
doi:10.1103/PhysRevE.90.012126.
[43] Marzec, M. Portfolio Optimization: Applications in Quantum Computing. Social Science
Research Network Technical Report. 2014. doi:10.2139/ssrn.2278729.
 1 Introduction to quantum machine learning 9

[44] Coxson, G. E., Hill, C. R., & Russo, J. C. Adiabatic quantum computing for finding low-peak-
sidelobe codes. High Performance Extreme Computing conference, IEEE 7. 2014, 3910–3916.
doi:10.1039/b509983h.
[45] Kirkpatrik, S., Gelatt Jr., C. D., & Vecchi, M. P. Optimization by simulated annealing. Science.
1983, 220, 671–680.
[46] de Falco, D., & Tamascelli, D. An introduction to quantum annealing RAIRO – Theoretical
Informatics and Applications. 2011, 45, 99–116.
[47] Morita, S., & Nishimori, H. Mathematical Foundation of Quantum Annealing. 2011.
https://arxiv.org/abs/0806.1859v1.
[48] Lucas, A. Ising formulations of many NP problems. Front. Physics. 2014, 12(5).
[49] Young, A. P. (Ed). Spin Glasses and Random Fields. Singapore: World Scientific. 1998.
[50] Nishimori, H. Statistical Physics of Spin Glasses and Information Processing: An Introduction.
Oxford University Press, 2001.
[51] Amin, M. H. Physical Review Letters. 102, 220401, ISSN 0031-9007.
[52] Jansen, S., Ruska, M. B., & Seiler, R. Journal of Mathematical Physics. 2007, 48, 102111,
ISSN 00222488.
[53] da Silva, A. J., Ludermir, T. B., & de Oliveira, W. D. Quantum perceptron over a field and
neural network architecture selection in a quantum computer. Neural Networks. 2016, 76,
55–64.
[54] Panella, M., Martinelli, G. Neural networks with quantum architecture and quantum learning.
International Journal of Circuit Theory and Applications. 2011, 39, 61–77.
[55] Kak, S. On quantum neural computing. Advances in Imaging and Electron Physics. 1995, 94(259).
[56] Chrisley, R. Quantum Learning. New directions in cognitive science: Proceedings of the
international symposium, Saariselka, 4–9 August 1995, Lapland, Finland. Pylkkänen, P., &
Pylkkö, P. (Eds). Finnish Association of Artificial Intelligence, Helsinki. 1995, 77–89.
[57] Narayanan, A., & Menneer, T. Quantum artificial neural network architectures and
components. Information Sciences. 2000, 128, 231–255.
[58] Schuld, M., Sinayskiy, I., & Petruccione, F. Simulating a perceptron on a quantum computer.
2014. ArXiv:1412.3635.
[59] Purushothaman, G., & Karayiannis, N. Quantum Neural Networks (QNN’s): Inherently Fuzzy
Feedforward Neural Networks. IEEE Transactions on Neural Networks. 1997, 8(3),679.
[60] Behrman, E. C., Steck, J. E., Kumar, P., & Walsh, K. A. Quantum Algorithm design using
dynamic learning. Quantum Information and Computation. 2008, 8(1&2),12–29.
[61] Ventura, D., & Martinez, T. A quantum associative memory based on Grover’s algorithm.
Proceedings of the International Conference on Artificial Neural Networks and Genetics
Algorithms. 1999, 22–27.
Bruce J. MacLennan
2 Topographic representation for quantum
machine learning
Abstract: One of the most common information representations in the brain is the
topographic or computational map, in which neurons are arranged systematically
according to the values they represent. By representing quantitative relationships
spatially, computational maps enable the brain to compute complex, nonlinear
functions to the accuracy required. This chapter proposes two approaches to
quantum computation for machine learning by means of topographic representa-
tion. It shows how to construct unitary operators, implementable on quantum
computers, that implement arbitrary (including nonlinear) functions via computa-
tional maps.

Keywords: computational map, neural network, quantum computation, quantum

machine learning, quantum neural network, topographic map, topographic
representation

2.1 Introduction
This chapter proposes a brain-inspired approach to quantum machine learning
with the goal of circumventing many of the complications of other approaches.
The fact that quantum processes are unitary presents both opportunities and chal-
lenges. A principal opportunity is that a large number of computations can be car-
ried out in parallel in linear superposition, that is, with quantum parallelism.
There are of course many technical challenges in quantum computing, but the
principal theoretical challenge in quantum machine learning is the fact that quan-
tum processes are linear whereas most approaches to machine learning depend
crucially on nonlinear operations. Artificial neural networks, in particular, require
a nonlinear activation function, such as a logistic sigmoid function, for nontrivial
machine learning. Fortunately, the situation is not hopeless, for we know that
nonlinear processes can be embedded in unitary processes, as is familiar from the
circuit model of quantum computation [1].
Despite the complications of quantum machine learning, it presents a tantalizing
approach to implementing large-scale machine learning in a post-Moore’s law techno-
logical era. However, there are many approaches to machine learning and several

Bruce J. MacLennan, University of Tennessee, Knoxville, Tennessee, USA

https://doi.org/10.1515/9783110670707-002
12 Bruce J. MacLennan

approaches to quantum computation (e.g., circuit model, annealing), and it is not obvi-
ous which combinations are most likely of success. Schuld, Sinayskiy, and Petruccione
[2, 3] provide useful recent reviews of some approaches to quantum neural networks
and quantum machine learning more generally. They conclude that none of the pro-
posals for quantum neural networks succeed in combining the advantages of artificial
neural networks and quantum computing. They argue that open quantum systems,
which involve dissipative interactions with the environment, are the most promising
approach. Moreover, few of the proposals include an actual theory of quantum ma-
chine learning.
This chapter explores an approach to the quantum implementation of machine
learning involving nonlinear functions operating on information represented topo-
graphically, as is common in neural cortex. However, there are many approaches to
neurocomputing, that is, to brain-inspired computing, some of which may be more
amenable to quantum implementation than others, and so we must say a few words
about the alternatives. For the brain may be modelled at many different levels, and
models at each of these levels may provide a basis for machine learning. For example,
in recent years spiking neural network models have become popular once again,
largely due to their power advantages when implemented in hardware [e.g., 4, 5]. Such
models mimic the “all or none” action potential generation by biological neurons with-
out addressing the detailed dynamics of the action potential. On the other hand, most
contemporary applications of artificial neural networks, including those used in deep
learning, use a higher level, rate-based model. That is, the real values passed between
the units (neuron analogs) represent the rate of neural spiking rather than individual
spikes. It has been argued that this is the appropriate level for modeling neural infor-
mation processing, since there are many stochastic effects on the generation and recep-
tion of action potentials, and because the fundamental units of information processing
are microcolumns comprising about 100 neurons [[6], chapter 2]. Therefore it is most
fruitful to view neural computation as a species of massively parallel analog computa-
tion. Since quantum computation makes essential use of complex–valued probability
amplitudes, it is also fruitful to treat it as a species of analog computation, and so ana-
log information representation provides one point of contact between quantum compu-
tation and artificial neural networks [7].

2.2 Topographic representation in the brain

Another respect in which information processing in the brain differs from most artifi-
cial neural network models is that biological neural networks are spatially organized,
with connectivity dependent on spatial organization. Although artificial neural net-
works are typically organized in layers, there is generally no spatial relationship
 2 Topographic representation for quantum machine learning 13

among the neurons in each layer;1 the exceptions are convolutional neural networks,
which were in fact inspired by the organization of sensory cortex.
One of the most common spatial information organizations used by the brain is
the topographic representation or computational map. In such representations, distinct
points x in some abstract space X are mapped systematically to physical locations
r = µ(x) in a two-dimensional region of cortex; that is, spatial relationships among the
neurons are correlated with topological relationships in the abstract space. These maps
are especially common in motor areas [8] and sensory areas [[6], ch. 6]. For example,
tonotopic maps have the neurons that respond to different pitches arranged in order
by pitch. Retinotopic maps have a spatial organization that mirrors the organization of
the retina and visual field. Neurons in primary visual cortex that respond to edges are
arranged systematically according to the orientation of the edges. There are many
other examples throughout the brain, and this is perhaps the single most common in-
formation representation used by the brain.
In these topographic maps, a particular value x is represented by an activity
peak in the corresponding cortical location µ(x). The strength of the activity reflects
the value’s degree of presence, importance, or probability. Moreover, multiple si-
multaneous values, with differing relative strengths or probabilities, are repre-
sented by multiple simultaneous activity peaks of differing amplitudes. Therefore,
such cortical maps can represent superpositions of values, each with its own
amplitude.
Topographic maps provide another point of contact between artificial neural
networks and quantum computation, because the computational maps in the brain
are large and dense enough that they can be usefully treated mathematically as
fields, that is, as continuous distributions of continuous quantity [9]. Such represen-
tations are suggestive of quantum mechanical wave functions, which are also con-
tinuous distributions of continuous quantity (the complex probability amplitude).
In both cases these fields are treated mathematically as continuous functions on a
continuous domain, and Hilbert spaces provide the mathematical framework for de-
scribing them [7]. In this chapter we exploit this analogy to implement brain-
inspired approaches to quantum machine learning.
Because of their spatial representation of values, topographic maps can be
used to implement arbitrary functions in the brain, essentially by a kind of table
lookup. Suppose the brain needs to implement a (possibly nonlinear) transfor-
mation, y = f ð xÞ. This can be accomplished by neural connections from locations
r = µ(x) in the input map to corresponding locations s = μ′ðyÞ = μ′½f ðxÞ in the output
map that represents the result space. Thus activity representing x in the input map
will cause corresponding activity representing y in the output map. Moreover, a

1 They are numerically indexed, of course, but interchangeable in terms of their pattern of connec-
tions before learning.
14 Bruce J. MacLennan

superposition of input values will lead to a corresponding superposition of output

values. Therefore, topographic representations allow the computation of nonlinear
functions in linear superposition [9–13], which suggests their usefulness in quan-
tum computation [7]. On the other hand, topographic maps make relatively ineffi-
cient use of representational resources, because every represented value has to
have a location in the map [although this can be mitigated by means of coarse cod-
ing, by which precise values are represented by a population of broadly tuned neu-
rons with overlapping receptive fields [14, 15]]. Therefore, use of topographic
representations will require a reasonably scalable quantum computing technology.
In this chapter we explore two topographic approaches to quantum computation
with a focus on machine learning.

2.3 Topographic basis maps

In the brain, the state of a topographic map is a real-valued function defined over a
(typically two-dimensional) space Ω. To apply these ideas in quantum computation,
we consider a quantum state jψi in which the probability amplitude ψ(r) at location
r ∈ Ω represents the amplitude (presence, importance, etc.) of the value x via the
correspondence r = µ(x). Here r ∈ Ω may be a continuous index representing, for ex-
ample, spatial location, or the index of a discrete quantum state, such as a wave-
length or the state of a qubit register. The states jri form a discrete basis or
continuous pseudo-basis for the input and output quantum states. In the continuous
case, the input value x is represented by a continuous basis state jμðxÞi, which is a
Dirac unit impulse at r = µ(x), that is, jri = δr where δr(s) = δ(s − r). Similarly an out-
 
put value y is represented by the continuous basis state μ′ðyÞi = μ′½f ðxÞi = δμ′ðyÞ . We
call such a representation (whether discrete or continuous) a topographic basis map.
For such a continuous basis we can define a Hilbert-Schmidt linear operator:
ð

Tf = dxμ′½f ðxÞihμðxÞj,
X

where X is the space of input values. (We write T = Tf when f is clear from context.)
This operator has the desired behavior: T jμðxÞi = jμ′½f ðxÞi for all x ∈ X . In this man-
ner the linear operator T computes the nonlinear function f via the computational
maps. We call such an operator a graph kernel because it uses the explicit graph of f
 
[that is, the set of pairs μ′½f ðxÞ, μðxÞ for all x ∈ X ] to do a kind of table lookup.2
Notice that if the input map is a superposition of input values, jψi = ajμðxÞi +
bjμðx′Þi, then the output map will be a superposition of the corresponding results:

2 It is not the same as the graph kernels used in machine learning applied to graph theory.
 2 Topographic representation for quantum machine learning 15

T jψi = ajμ′½f ðxÞi + bjμ′½f ðx′Þi. Therefore, continuous topographic basis maps per-
mit nonlinear functions to be computed in linear superposition (quantum parallel-
ism). This is a step toward quantum computation, but T might not be unitary, and
we have more work to do.
The reader might question the use of a continuous basis. First, note that for sep-
arable Hilbert spaces, the continuous basis can always be replaced by an infinite
discrete basis, for example, by a discrete series of sinusoids or complex exponen-
tials of the appropriate dimension. Second, the infinite discrete basis can be ap-
proximated by a finite discrete basis, for example, by band-limited sinusoids or
complex exponentials. Such an approximation is especially appropriate for neural
network machine learning, which requires only low-precision calculation.

2.3.1 Bijections

We proceed to show several examples of nonlinear computations performed via

quantum computational maps, beginning with a simple case and proceeding to more
complex ones. For simplicity, we will ignore the representation map µ and consider
computations from one quantum state to another. We consider both one-dimensional
continuous domains, Ω = ½xl , xu  and discrete domains, Ω = fx1 , . . . , xn g. Typically
the values would be evenly spaced, for example, Ω = f0, Δx, 2Δx, . . . , ðn − 1ÞΔxg, but
this is not required, and other spacings, such as logarithmic, might be useful.
(Logarithmic maps are found in some sensory regions of the cortex.) In both cases
the vectors fjxij x∈X g are an orthonormal basis (composed of unit vectors in C n for
the discrete case and of delta functions in L2 ðΩÞ for the continuous case). For exam-
ple, in the discrete case, the values x1, . . ., xn might be represented by the composite
state of an N-qubit register, where n = 2N. This might seem to require a large number
of qubits, but even in the absence of coarse coding, seven qubits would be sufficient
to represent values with 1% precision, which is adequate for many machine learning
applications.
We begin with a bijective scalar function f :Ω ! Ω, where Ω = ½ − 1, 1. The hyper-
bolic tangent (appropriately restricted),3 which is a useful sigmoid function for neu-
ral computation, is an example of such a function. The graph kernel to compute the
function topographically is
ð
T = dxjf ðxÞihxj. (2:1)
Ω

Since f is bijective, the adjoint of T is

3 For example, f ðxÞ = ½tanh x=tanh 1j½ − 1, 1 .

16 Bruce J. MacLennan

ð ð
T† = dxjxihf ðxÞj = dyjf − 1 ðyÞihyj. (2:2)
Ω Ω

It is easy then to show that T is unitary and therefore amenable to quantum compu-
tation. In general, we have:

Proposition 1 The graph kernel T of a continuous bijection f :Ω ! Ω′ is unitary.

Proof: Substituting eqs. (2.1) and (2.2), observe:

0 10 1
ð ð
 
T † T = @ f − 1 ðyÞihyjdyA@ f ðxÞihxjdxA
Ω′ Ω
ðð
 −1 
= f ðyÞihyjf ðxÞihx dydx
Ω Ω′
ðð
 −1 
= f ðyÞihyjf ðxÞidyhxdx
Ω Ω′
ð
 
= f − 1 ðf ðxÞÞihxdx
Ω

= IΩ .
Ð
The fourth line follows from the “sifting” property of the Dirac delta: jF ðyÞihyjaidy =
jF ðaÞi. Likewise, Ω′

0 10 1
ð ð
†  −1 
TT = @ jf ðxÞihxjdxA@ f ðyÞihydyA

ðð
Ω Ω′
 
= f ðxÞihxjf − 1 ðyÞihy dxdy
Ω′Ω
ð
 
= jf f − 1 ðyÞ ihyjdy = IΩ′ .
Ω′

Therefore T is unitary.
In the discrete basis case, let yi = f (xi), where f is bijective. The unit vectors jxi i
are a basis for C n , and the unit vectors jyi i are also a (possibly different) basis for
C n . The graph kernel is
X
n X
n
T= jyi ihxi j = jf ðxi Þihxi j, (2:3)
i=1 i=1
 2 Topographic representation for quantum machine learning 17

which is also easily proved to be unitary. More directly, we can observe that T is a
permutation matrix on the basis elements and therefore orthogonal.
Ð
If the input is a weighted superposition of values, jψi = Ω dx pðxÞjxi, then apply-
ing the kernel [eq. (2.1)] will give a corresponding superposition of the outputs:
Ð
Tjψi = Ω dx pðxÞjf ðxÞi. The same applies, of course, in the discrete case. Moreover,
since the graph kernel is unitary, its adjoint computes the inverse function. Therefore,
Ð Ð
if jϕi = Ω′ dy qðyÞjyi, then T † jϕi = Ω′ dy qðyÞjf − 1 ðyÞi. That is, applying the adjoint to
a superposition of outputs will compute a corresponding superposition of inputs.

2.3.2 Non-surjective Injections

As a further step towards the quantum computation of arbitrary functions by means

of computational maps, we consider a relatively simple case: non-surjective injec-
tions (that is, one-to-one non-onto functions). We restrict our attention to finite do-
mains and codomains. Therefore, let Ω = fx1 , . . . , xn g and Ω′ = fy1 , . . . , ym g, where
n < m, and consider an injection f :Ω ! Ω′. Input maps will be in an n-dimensional
Hilbert space HðΩÞ and output maps will be in an m-dimensional Hilbert space
 
H Ω′ . Since n < m, ancillary constants will need to be provided from a space HC ,
and so the complete input space will be in HðΩÞ#HC . Our implementation will also
generate “garbage” output in a space HG , and so the complete output space will
 
be in H Ω′ # HG . The input and output dimensions must be equal, and the simplest
way to accomplish this is to make HC m-dimensional and HG , n-dimensional, so that
our operator is an mn-dimensional Hilbert-space transformation. Let fjw1 i, . . . ,
jwm ig be an orthonormal (ON) basis for HC and let fjv1 i, . . . , jvn ig be an ON basis
 
for HG . (We could in fact use the H Ω′ basis for HC and the HðΩÞ basis for HG ,
but here we develop a more general result.)
Our goal will be to define a unitary U so that U jxijw1 i = jf ðxÞijγi, where jw1 i is
an ancillary constant and jγi is garbage. As we will see, U can be implemented by
an appropriate permutation of the input basis into the output basis, which can be
expressed as the sum of several operators:
def
U = T + S + R + Q. (2:4)

The work of the function f is accomplished by the T component, which maps xj i to
 
jf xj i:
X
n    
def
T = f xj iv1 ihxj jhw1 j. (2:5)
j=1

    
Note that T xj iw1 i = f xj iv1 i and T is a bijection of the n-dimensional subspace
HðΩÞ # Hðfjw1 igÞ. However, T is not unitary since it is not a surjection. See
Figure 2.1.
18 Bruce J. MacLennan

Figure 2.1: Permutation of basis vectors to implement non-surjective injections. After each component
of the kernel, the number of basis vectors that it maps is indicated in parentheses; for example, R
maps ðm − 1Þðn − 1Þ basis vectors. T maps function inputs to outputs, S maps zero amplitudes to non-
range codomain elements, and R and Q bijectively map the remaining basis elements.

The S component ensures that non-range elements of the codomain have prei-
mages in the domain. Therefore, let mnr be the number of codomain elements that
are not in the range of f, that is, mnr = jΩ′ nIm f j = m − n. Call these non-range codo-
main elements fz1 , . . . , zmnr g  Ω′. Then the S component is defined:

def
X
mnr
S= jzi ijv1 ihx1 jhwi + 1 j. (2:6)
i=1

Therefore, S is a bijection of an mnr -dimensional subspace and each non-range ele-

ment jzi ijv1 i has a unique preimage jx1 ijwi + 1 i.
Note that T transforms n basis vectors (those for which the second register is
jw1 i), and S transforms mnr basis vectors (those for which the first register is jx1 i
and the second register is one of jw2 i, . . ., jwmnr + 1 i), for a total of n + mnr = m basis
elements, but the input space has a total of mn basis elements, and the remainder
must be bijectively mapped. Therefore, to complete the unitary operator we add the
following additional components:

def
X
m X
n
R= jyi ijvj ihxj jhwi j, (2:7)
i=2 j=2

def
X
n−1
Q= jy1 ijvj + 1 ihx1 jhwmnr + j j. (2:8)
j=1

R maps ðm − 1Þðn − 1Þ basis elements: those for i ≠ 1 and j ≠ 1, that is, those with nei-
ther jx1 i in the first register nor jw1 i in the second. Q maps the remaining n − 1 basis
 2 Topographic representation for quantum machine learning 19

elements: those which have jx1 i in the first register and jwmnr + 1 i to jwmnr + n i in
the second (recall that m = mnr + n).
Notice that U maps every input basis vector into exactly one output basis vector
and vice versa (see Figure 2.1). Summing the numbers of basis vector dyads for T, S,
R, Q gives:

n + mnr + ðm − 1Þðn − 1Þ + ðn − 1Þ = m + ðm − 1Þðn − 1Þ + n − 1 = mn.

 
Proposition 2 Let Ω and Ω′ be finite sets with n = jΩj, m = Ω′, and m > n. Let
f :Ω ! Ω′ be a non-surjective injection. Let HC and HG be Hilbert spaces of
dimension m and n, respectively (representing ancillary constant inputs and gar-
bage outputs). Let jωi be a fixed basis vector of HC and jυi be a fixed basis vector
   
of HG . Then there is a unitary operator U∈L HðΩÞ # HC , H Ω′ # HG such that
for all x∈Ω,

U ðjxi#jωiÞ = jf ðxÞi#jυi. (2:9)

Proof: The proposition follows from the construction preceding the proposition.
P
If the input to U is a (normalized) superposition, jψi = k pk jxk i, then the output will
P
be a superposition of the corresponding function results: U jψijωi = k pk jf ðxk Þijυi.
The dimension of the input and output spaces of this implementation is mn. A
more resource-efficient but also more complicated implementation operates on a
space of dimension LCMðm, nÞ. The principle is the same: a permutation of the
basis vectors.

2.3.3 Non-injective surjections

Next we consider functions f :Ω ! Ω′ that are surjections but not injections; that is,
f maps onto Ω′ but might not be one-to-one. This includes many useful functions,
such as non-injective squashing functions and Gaussians, but also binary functions
such as addition and multiplication (as explained later).
A non-injective function loses information, and thus for quantum computa-
tion it must be embedded in a larger injective function, which moreover must be
 
unitary. In particular, if f is non-injective (e.g., f ðxÞ = f x′ for some x ≠ x′), then
the corresponding graph kernel will also be non-injective: T jxi = jyi = Tjx′i for
jxi ≠ jx′i. Therefore Tðjxi − jx′iÞ = 0, which implies that jxi − jx′i is in the null space
of T, which we write N ðT Þ. Therefore there is a bijection between the orthogonal
complement of the null space, N ðT Þ? , and the range of the operator, Im T. Hence
we will implement the non-injective operation by decomposing the input jψi into
orthogonal components jψi = jμi + jνi, where jμi∈ N ðT Þ? and jνi∈ N ðT Þ. The jμi
20 Bruce J. MacLennan

component is sufficient to determine the output, so there is a bijection jμi7!T jψi,

and the jνi component preserves the information to differentiate the inputs that
map to this output.
To explain how this separation can be accomplished, we consider the finite-
dimensional case, but it is easily extended. Let Ω = fx1 , . . . , xn g and Ω′ = fy1 , . . . , ym g;
since f is surjective, m ≤ n.
 
The desired operator T is in L H, H′ , where H = HðΩÞ is an n-dimensional
Hilbert space with basis fjx1 i, . . . , jxn ig. The output space H′ is also n-dimensional,
and m of its basis vectors fjy1 , . . . , jym g are used to represent a topographic map of
the function’s codomain (which is also its range), Im f = Ω′ = fy1 , . . . , ym g. Therefore
 
H Ω′ is a subspace of H′. Let fjw1 i, . . . , jwn − m ig be a basis for HðΩ′Þ? , the orthogo-
 
nal complement of H Ω′ in H′. (This subspace will represent “garbage” with no
computational relevance.)
We will define fju1 i, . . . , jum ig to be any ON basis for N ðT Þ? (the row space
of T), where m is the rank of T, and we will define fjv1 i, . . . , jvno ig to be any ON basis
for N ðT Þ, where no = n − m is the nullity of T. These bases will determine the orthogo-
nal components jμi∈ N ðT Þ? and jνi∈ N ðT Þ into which any input is separated.
An example will make this clearer. Suppose Ω = fkΔxj − N < k < N g and
Ω′ = fkΔxj0 ≤ k < N g. Let abs: Ω ! Ω′ be the absolute value function (a noninjective
surjection between these sets). A basis for the nonnull space N ðT Þ? comprises
pffiffiffi
ju0 i = j0i and the vectors juk i = ðj − kΔxi + jkΔxiÞ= 2 (for k = 1, . . . , N − 1). (Note that
j − kΔxi and jkΔxi are orthogonal vectors for k ≠ 0.) These N basis vectors are in a
one-to-one relation with the codomain elements jkΔxi (for k = 0, . . . , N − 1). The nul-
lity is no = ð2N − 1Þ − N = N − 1 and the basis vectors of the null space are:
pffiffiffi
jvk i = ðjkΔxi − j − kΔxiÞ= 2 ðfor k = 1, . . . , N − 1Þ.

Projection onto this space keeps the information necessary to distinguish the spe-
cific preimage that maps to a given output. In this case, it remembers the sign of
pffiffiffi pffiffiffi
the input: note that hvk jkΔxi = + 1= 2 and hvk j − kΔxi = − 1= 2. Therefore, for input
pffiffiffi pffiffiffi
jkΔxi, the orthogonal components are jμi = juk i= 2 and jνi = jvk i= 2; and for input
pffiffiffi pffiffiffi
j − kΔxi, they are jμi = juk i= 2 and jνi = − jvk i= 2. This completes the example and
we return to the construction for an arbitrary non-injective surjection.
For each yi ∈Ω′, let f − 1 fyi g = fxjf ðxÞ = yi g be the inverse image of yi ; these are
disjoint subsets of the domain Ω and correspond to orthogonal subspaces of H. Let
ni = jf − 1 fyi gj be the preimage multiplicity of yi, where n = n1 +    + nm . Because dif-
ferent yi ∈ Im T have different preimage multiplicities, it will be convenient to sepa-
rate f into m constant functions fi : f − 1 fyi g ! fyi g with corresponding graph kernels:
def 1
X 1 X
Ti = pffiffiffiffi jyi ihxj j = jyi i pffiffiffiffi hxj j = jyi ihui j, (2:10)
ni ni
xj ∈f − 1 fyi g xj ∈f − 1 fyi g
 2 Topographic representation for quantum machine learning 21

where we have defined the following normalized basis vectors of N ðT Þ? :

def 1
X
jui i = pffiffiffiffi jxj i, i = 1, . . . , m. (2:11)
ni
xj ∈f − 1 fyi g

The constant maps Ti operate independently on orthogonal subspaces of H.

  − 1=2
Note that hui jxj i ≠ 0 if and only if yi = f xj , and in this case Ti jxj i = ni jyi i.
− 1=2
(The ni factor is required for normalization of the jui i.) Clearly fju1 i, . . . , jum ig is
an ON set, since its elements are normalized linear combinations of disjoint sets of
the basis vectors. Therefore there is a one-to-one correspondence between the out-
put vectors jyi i and the basis vectors jui i of N ðT Þ? .
Next we characterize N ðT Þ and N ðT Þ? . Observe that jψi∈ N ðTi Þ if and only if
0 = Ti jψi = jyi i hui jψi, that is, if and only if hui jψi = 0. Therefore, N ðTi Þ is the n − 1
dimensional subspace orthogonal to jui i and N ðTi Þ? is the one-dimensional sub-
space spanned by {jui ig. Therefore N ðT Þ? is spanned by fju1 i, . . . , jum ig.
The operation T is implemented in two parts, one that handles the components
representing the null space and the other that handles the components representing
its orthogonal complement, the nonnull space. The first part generates “garbage,”
but is required for the operation to be invertible and hence unitary; the second part
does the work of computing f.

We have vj i∈H, the basis vectors for the N ðT Þ subspace of the domain, which has

dimension no = n − m. Let jwj ijj = 1, . . . , no be any basis for HðΩ′Þ? , the orthogonal
   
complement of H Ω′ in H′, which also has dimension n − m. We define N∈L H, H′
to map the null space components down to this no-dimensional subspace:

def
X
no
N = jwj ihvj j. (2:12)
j=1

Since there is a one-one correspondence between basis vectors jui i and output vec-
 
tors jyi i, we implement the function f by an operator M∈L H, H′ defined:

def
X
m
M= jyi ihui j. (2:13)
i=1

This maps the n-dimensional input into an m-dimensional output subspace. As a

result the operator
 
T = M + N∈L H, H′ (2:14)

maps an input jxi to the correct output jf ðxÞi, but with a scale factor and additional
“garbage.” Specifically, for xj ∈ f −1{yi},
   
T xj i = ðM + N Þxj i = Mxj i + N xj i = ni jyi i + jγj i.
− 1=2
(2:15)
22 Bruce J. MacLennan

qffiffiffiffiffiffiffiffi
ni − 1
where jγj i = Njxj i and kjγj ik = ni . Note that the garbage jγj i is superimposed on

the desired output. Subsequent computations operate on the H Ω′ subspace and
ignore the orthogonal subspace, which contains the garbage (which nevertheless
must be retained, since it is entangled with the computational results).
The T operator is just a transformation from the fjx1 i, . . . , jxn ig basis to the
fjy1 i, . . . , jym i, jw1 i, . . . , jwn − m ig basis, and is obviously unitary:

ðM + N Þ† ðM + N Þ = M† M + N † M + M† N + N † N = IN ðT Þ? + 0 + 0 + IN ðT Þ = IH ,
ðM + N ÞðM + N Þ† = MM† + NM† + MN † + NN † = IHðΩ′Þ + 0 + 0 + I ′ ? = IH′ .
HðΩ Þ

Since the T operator is unitary, it can be approximated arbitrarily closely by a com-

bination of H (Hadamard), CNOT (conditional NOT), and T (π/8) gates [[1], §4.5], or
any other universal set of quantum gates.
Unfortunately, the output vectors jyi i have amplitudes that depend on their pre-
  − 1=2
image multiplicities. That is, if yi = f xj , then Tjxj i = ni jyi i + jγj i, and we get dif-
− 1=2  
ferent scale factors ni depending on the preimage multiplicity. For yi = f xj , we
def − 1=2
define si = ni to be this scale factor so that T jxi i = si jyi i + jγj i. We would like to
equalize the differing amplitudes but there does not seem to be a unitary means for
doing so.
It might seem that something like a Grover iteration [16] could be used to rotate
the state vector from si jyi i + jγj i to jyi i, but different si require different numbers of
iterations. Something like Grover’s algorithm with an unknown number of solutions
could be used, but this would require trying multiple rotations. Therefore, it seems
better to accept the unwanted scale factors and work with them. This means that
any jyi i with positive amplitudes are considered outputs from the computation, and
therefore all positive amplitudes are treated the same.
If we ignore the relative magnitudes of positive amplitudes, then a quantum
P
state jψi = j pj jxj i (with pj ≥ 0) can be interpreted as representing the set of all xj
 P
with positive amplitudes: xj ∈Ω j pj > 0 , where we assume of course that j p2j ≤ 1.
Moreover, the sum can be strictly less than one only if there are additional ancillary
states that make up the difference (like jγj i in the previous example). Applying T to
such a state computes a state representing the image of the input set. That is,
P     
Tjψi = j pj sj jf xj i, which represents the set f xj ∈Ωjpj > 0 . If S  Ω is the set
represented by jψi, then T jψi represents its image f [S]. Since zero amplitudes will
always map to zero amplitudes and positive amplitudes will map to positive ampli-
tudes, set membership will be appropriately mapped from the domain to the codo-
main. The preceding construction gives us:

Proposition 3 Let Ω = fx1 , . . . , xn g and Ω′ = fy1 , . . . , ym g be finite sets with n ≥ m.

Suppose fjx1 i, . . . , jxn ig is an ON basis for a Hilbert space H and fjy1 i, . . . , jym ig is
 
an ON basis for a subspace H Ω′ of H. For any surjective function f :Ω↠Ω′ there is
a unitary operator T∈LðH, HÞ such that for any x∈ Ω,
 2 Topographic representation for quantum machine learning 23

1
Tjxi = pffiffiffiffiffi jf ðxÞi + jγi, (2:16)
nx
 qffiffiffiffiffiffiffiffiffi
where nx = jf − 1 ff ðxÞgj, γi∈HðΩ′Þ? , and jjjγijj = nxnx− 1.

Proof: As previously shown [eqs. (2.12)–(2.14)], this operator is given explicitly by

X
m nX
−m
T= jyi ihui j + jwk ihvk j,
i=1 k=1
def P
where jui i = p1ffiffiffi
ni xj ∈f − 1 fyi g jxj i, ni = jf
−1
fyi gj, the jvk i are an ON basis for the or-
thogonal complement of the space spanned by the jui i, and the jwk i are an ON
basis for HðΩ′Þ? .

2.3.4 Arbitrary functions

In the preceding, we have assumed for convenience that the function is either non-
injective or non-surjective, but not both. The solutions are easily extended to arbi-
trary functions, since every function can be factored as a composition of an injec-
tion and a surjection. More directly, we can combine Prop. 3 to implement a
surjection onto its range with Prop. 2 to inject its range into its codomain.
Let the domain Ω = fx1 , . . . , xn g, where n = jΩj, and let fjx1 i, . . . , jxn ig be the
standard basis of HðΩÞ. Let x0 ∉ Ω be an additional value and define the extended
domain Ω = fx0 g ∪ Ω. Then HðΩ Þ has basis fjx0 i, jx1 i, . . . , jxn ig. (For example,
HðΩ Þ may be the state space of N qubits, where n + 1 = 2N.) The additional x0 input
dimension will carry the null space “garbage” from previous computations.
Similarly, let the codomain Ω′ = fy1 , . . . , ym g, where m = j Ω′j, and let fjy1 i, . . . , jym ig
 
be the standard basis of H Ω′ . Let y0 ∉ Ω′ be an additional value and define the
 
extended codomain Ω* = fy0 g ∪ Ω′. Then H Ω* has basis fjy0 i, jy1 i, . . . , jym ig. The
y0 component carries the garbage in the output state.
Let fr1 , . . . , rmr g = Im f and fz1 , . . . , zmnr g = Ω′nImf be the range of f and its
def def

complement, respectively; since every codomain element is either a range element

or not, n = mr + mnr . As before, an n-dimensional input jxj i will be projected into or-
thogonal subspaces (the nonnull and null spaces) of dimension mr and no = n − mr ,
with basis vectors jui i, i = 1, . . . , mr , and jvk i, k = 1, . . . , no , respectively.
An additional input quantum register will be used to provide the constant zero
amplitudes for non-range elements for non-surjective functions. The m + 1 dimen-
 
sional state of this ancillary register will be in, for convenience, HC = H Ω* with
basis fjy0 i, jy1 i, . . . , jym ig. There will also be an additional output quantum register
to hold the null space garbage for non-injective functions. Its n + 1 dimensional
state is in, for convenience, HG = HðΩ Þ with basis fjx0 i, jx1 i, . . . , jxn ig. Note that
both the input and output spaces have dimension ðm + 1Þðn + 1Þ. This is because the
24 Bruce J. MacLennan

ancillary input register is in the same space as the regular output register, and the
ancillary output register is in the same space as the regular input register. This can
be confusing because, as will be seen, we use the extra output vector jy0 i as a con-
stant in the ancillary input register, and the extra input vector jx0 i appears in the
ancillary output register.
   
Our goal is to define unitary U∈L HðΩ Þ # HC , H Ω* # HG so that
    
U sjxj i + tjx0 i # jy0 i = s′jf ðxj iÞ + t′jy0 i # jγi, (2:17)

for scalars s, s′, t, t′ and for jγi ∈ HG . That is, the input register is initialized to the

input xj i with some positive amplitude s, possibly with superimposed garbage with
amplitude t; the ancillary input register is initialized to constant jy0 i. After computa-

tion, the output register will contain the function’s value f ðxj Þi with some positive
amplitude s′; and superimposed garbage with amplitude t′. The ancillary output reg-
ister may also contain garbage. In other words, the argument to f is in the first input
register [corresponding to HðΩ Þ], and its result is in the first output register [corre-
 
sponding to H Ω* ], possibly with garbage in both its input jx0 i component and its
output jy0 i component. The input ancillary register is initialized to a constant jy0 i.
The work of computing f is done by the graph kernel M, which will map the
jui ijy0 i vectors into corresponding ðm + 1Þðn + 1Þ dimensional result vectors jri ijx0 i
 
in the output space H Ω* # HG . (The ancillary jx0 i∈HG output is required so that
the input and output spaces have the same dimension.) To accomplish this map-
ping, define M as follows:

def
X
mr
M = jri , x0 ihui , y0 j. (2:18)
i=1

It maps mr of the basis vectors of HðΩ Þ # HC into mr of the basis vectors of

 
H Ω* # HG (see Figure 2.2). Specifically it is a bijection between the nonnull sub-
 
space of HðΩ Þ # Hðfjy0 igÞ and the output subspace of H Ω* # Hðfjx0 igÞ.
Another component of the transform will map the no null space basis vectors
jvk ijy0 i of the m + 1 dimensional jy0 i subspace of the input space:

def
X
no
N = jy0 , xk ihvk , y0 j. (2:19)
k=1

It maps them to no of the basis vectors jy0 , xk i of the m + 1 dimensional jy0 i subspace
of the output space. M and N together handle non-injective functions by mapping
mr + no = n basis vectors.
For non-surjective functions, zero amplitudes are copied from the ancillary reg-
ister jx0 ijyi i into the appropriate non-range codomain components jzi ijx0 i:

def
X
mnr
S = jzi , x0 ihx0 , yi j. (2:20)
i=1
 2 Topographic representation for quantum machine learning 25

Figure 2.2: Permutation of basis vectors to implement arbitrary function. After each component of
the kernel, the number of basis vectors that it maps is indicated in parentheses. For example,
R maps mn basis vectors. The shapes labeled ui and vk represent projection onto the basis vectors
of the nonnull and null spaces, respectively.

This is a mapping of mnr basis vectors between the jx0 i subspaces of the input
and output spaces. The three operators M, N, S handle the mapping of f (and disposal
of the null space).
We have to be careful, however, to handle all the jx0 ijyi i basis vectors since mnr
might be less than n (see Figure 2.2). We map the remaining basis vectors of the jx0 i
subspace that were not used in the S map to components of the jy0 i subspace that
are unfilled by N:
X
mr 
Q =
def y0 , xn + k ihx0 , ym + k j. (2:21)
o nr
k=1

Note that no + mr = n = mnr + mr so that all these vectors are bijectively mapped.
It remains to handle the other basis vectors of the input space in a unitary
way. The preceding maps have either jx0 i or jy0 i, but not both, in the input regis-

ter. The R operator maps the mn basis vectors with neither, xj ijyi i, for i, j ≠ 0, into
their reverses:
X
m X
n  
R =
def yi , xj ihxj , yi . (2:22)
i=1 j=1

The state jx0 ijy0 i remains, and it maps to its reverse:

def
P = jy0 , x0 ihx0 , y0 j. (2:23)
26 Bruce J. MacLennan

In summary, M bijectively maps mr basis vectors, N maps no basis vectors, S

maps mnr, R maps mn, Q maps mr, and P maps one basis vector, which accounts for
all of the (n + 1)(m + 1) basis vectors (see Figure 2.2):

mr + no + mnr + mn + mr + 1 = ðmr + no Þ + ðmnr + mr Þ + mn + 1

= n + m + mn + 1 = ðm + 1Þðn + 1Þ.

The unitary operator to compute f is the sum of these linear maps [eqs. (2.18)–(2.23)]:
def
U = M + N + S + R + Q + P. (2:24)

 
Proposition 4 Suppose f : Ω ! Ω′. Let n = jΩj and m = Ω′. Let HðΩ Þ be an n + 1
 
dimensional Hilbert space with basis fjx0 i, . . . , jxn ig, and let H Ω* be an m + 1
dimensional space with basis fjy0 i, . . . , jym ig. Then there a unitary operator
     
U∈L HðΩ Þ # H Ω* , H Ω* #HðΩ Þ so that for scalars s, t (with jsj2 + jtj2 = 1)
and x∈Ω:

U ½ðsjxi + tjx0 iÞ#jy0 i = ss′jf ðxÞ, x0 i + t′jy0 , γi, (2:25)

 2   2
, where nx = jf − 1 ff ðxÞgj, and ss′ + t′ = 1.
− 1=2
where s′ = nx
def

Proof: By construction [eqs. (2.18) and (2.19)] we know:

 
U xj , y0 i = ðM + N Þxj , y0 i

X 
X 
mr  no
= jri , x0 ihui , y0 j xj , y0 i + jy0 , xk ihvk , y0 j jxj , y0 i
i=1 k=1

    X no 
= f xj , x0 ihui jxj i + y0 i xk ihvk jxj i
k=1
  
= sj f xj , x0 i + jy0 , γi,

 Pno  pffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffi

where γi = 
k = 1 jxk ihvk j xj i and jj jy0 , γijj = nj − 1= nj . Furthermore, by eq. (2.23),

Ujx0 , y0 i = Pjx0 , y0 i = jy0 , x0 i.

Therefore, in the general case where the input register is sjxj i + tjx0 i (with jsj2 +
jtj2 = 1) we have:
  
U sxj i + tjx0 iÞ#jy0 i = sU xj , y0 i + tU jx0 , y0 i
   
= s sj f xj , x0 i + jy0 , γiÞ + tjy0 , x0 i
  
= ssj f xj , x0 i + jy0 iðsjγi + tjx0 iÞ.

❑
 2 Topographic representation for quantum machine learning 27

 
As a consequence, the result that we want is in the first [H Ω* ] quantum register,
but its jy0 i component is garbage and should be ignored in subsequent computa-
tions. Furthermore, the amplitude of the desired result will decrease through succes-
pffiffiffiffi
sive computation stages as a result of attenuation by successive 1= nj factors.

As discussed previously, quantum states with xj i components with positive
amplitudes represent sets of the corresponding xj (j ≠ 0). Applying U to such a state
  
yields a quantum state with positive amplitudes for the corresponding f xj i,
 
which represents the set of corresponding outputs f xj . Thus, U can be applied to
a (crisp) set of values with quantum parallelism.

2.4 Topographic qubit maps

2.4.1 Representation

To further explore quantum computation by topographic maps, in this section we

present an alternative representation of the maps and a circuit-based implementa-
tion of arbitrary functions on a finite domain. In these topographic qubit maps, each
domain value xj ∈Ω = fx1 , . . . , xn g or codomain value yi ∈Ω′ = fy1 , . . . , ym g is as-

signed a separate qubit, whose state, for example, jψj i = pj ′0i + pj j1i, where
 2  2
pj ′ + pj  = 1, represents the activity level of neuron xj by the amplitude pj . This
topographic representation can be contrasted with a binary representation such as
that proposed by Chen, Wang, and Charbon [17]. Our one-out-of-n (or “one-hot”)
representation might seem unrealistically inefficient, but (1) we are assuming a
scalable qubit implementation, which permits arrays of many thousands of qubits,
and (2) neural computations typically require only low precision [in the brain per-
haps as little as one digit: [18]]. Therefore a quantity can be represented by a few
tens of qubits; that is, our m and n will typically be small (m, n < 100).
Like the topographic basis maps, these topographic qubit maps can also be
viewed as representations of subsets of the domain; for S ⊆ Ω:
 X X
Si = 1i + 0i .
j j
xj ∈S xj ∉ S

That is, the xj qubit is in state j1i if xj is in S and is in state j0i if it is not. Therefore

we use the notation  xj i for the topographic map representing just the number xj .
The set of representations of all possible subsets of Ω is then an ON basis for the
2n-dimensional Hilbert space of these qubits. The basis can be written:

fjkijk∈2n g = jSijS  2Ω ,

where on the left 2n is the set of n-bit binary strings, and on the right 2Ω is the
powerset of Ω. Therefore, the sets are basis states and as a consequence
28 Bruce J. MacLennan

topographic qubit maps permit multiple sets to be processed in quantum superposi-

tion. Moreover, because the sets are represented by computational basis vectors,
they can be copied without violating the no-cloning theorem.
By using amplitudes other than 0 and 1, we can represent fuzzy sets. Suppose S
is a fuzzy set with membership function µ: S → [0, 1], and let mj = µ(xj). Then S is
represented by the topographic qubit map
X
n qffiffiffiffiffiffiffiffiffiffiffiffi
jSi = mj j1ij + 1 − m2j j0ij .
j=1

Fuzzy sets cannot, in general, be copied (nor can arbitrary superpositions of crisp
sets).
With the topographic qubit representation, the transformations between compu-
tational maps will be implemented by the quantum circuit model, and so one might
ask whether it would be simpler to implement ordinary binary digital quantum com-
putation. The answer is that computation on topographic maps can be implemented
by a few relatively simple operations (described in the following subsections), so that
computational maps buy a simpler quantum implementation at the cost of greater
representational expense (number of qubits). We expect topographic quantum
computation to be more simply implemented than a full-scale digital quantum arith-
metic-logic unit.

2.4.2 Unary functions

An example will illustrate how to implement an arbitrary finite function f : Ω ! Ω′

by a unitary operator on topographic qubit maps. For any yi not in the range of f,

we set its state ϕi i = j0i supplied as an ancilla. If yi is in the range of f, then it
 
might be the image of a single domain element, xj , that is, yi = f xj , in which case
  
we implement directly ϕi i = ψj i, transferring the state ψj i of input qubit j to out-
 
put qubit i. If there are two domain values mapping into yi , say f xj = yi = f ðxk Þ,
  
then we make ϕi i the logical OR of ψj i and ψk i. This is accomplished with the
two-input OR2 gate:
def
OR2 = CCNOTðX # X # IÞ, (2:26)

where CCNOT is the conditional-conditional-not or Toffoli gate. The result of ORing

the input states is:
 
OR2 jψj i # jψk i # j1i = Xjψj i # Xjψk i # jϕi i. (2:27)

The j1i input is a constant ancilla. The result of the OR is in the third output qubit,
and the first two output qubits, in which the negated inputs remain, are considered
 2 Topographic representation for quantum machine learning 29

     
garbage. If ψj i = p′0i + p1i and ψk i = q′0i + q1i, then OR2 transfers the probabil-
ity amplitudes as follows:
     
OR2 ψj iψk i1i = OR2 p′ 0i + pj1iÞ q′j0i + qj1iÞj1i
 
= p′q′OR2  001i + p′qOR2  011i + pq′OR2 j 101i + pqOR2 j111i
= p′q′j110i + p′qj101i + pq′j011i + pqj001i.

Therefore, the third output qubit is the OR of the first two input qubits with the am-
plitudes shown. If we interpret the squares of the amplitudes as probabilities, then
OR2 computes the correct probabilities for the third output qubit. The first two out-
put qubits are considered garbage but must be retained, for they are entangled with
the third output.
If more than two domain values map into a single codomain value, then we use
the multiple argument ORn, which can be defined recursively in terms of OR2:
h i
ORn jψ1 i    jψn ij1i#ðn − 1Þ = ORn − 1 ðOR2 jψ1 ijψ2 ij1iÞ#jψ3 i    jψn ij1i#ðn − 2Þ
def
ðn > 2Þ.
(2:28)
def
For completeness, we define OR1 = I.
With the preceding motivation, we can give the construction for computing an
arbitrary finite function by topographic qubit maps:

Proposition 5 Suppose f : Ω ! Ω′, where Ω = fx1 , . . . , xn g and Ω′ = fy1 , . . . , yn g. Let

def
mnr = n − jIm f j be the number of codomain elements that are not in the range of f.
def
Let nb be the number of injective domain elements, and let nn = n − nb be the
def
number of non-injective domain elements. Let mn = jIm f j − nb be the number of
non-injective range elements (i.e., those that are the image of two or more domain
elements). Then there is an 2nn + nb + mnr − mn dimensional unitary operator Uf that
computes f by topographic qubit maps as follows:

Uf jfxgi j0i#mnr j1i#ðnn − mn Þ = jff ðxÞgijγi, (2:29)

where jγi is 2(nn − mn) qubits of garbage.

Proof: The inputs are the n elements of the input map, mnr constant j0i ancillae
(for the non-range elements), and nn − mn constant j1i ancillae for the OR2 gates
that map multiple domain elements to the same range element. The latter are re-
quired because each of the non-injective range elements requires a number of OR2
gates that is one less than the number of its preimages; hence the mn non-injective
range elements require nn − mn OR2 gates. Therefore there are

n + mnr + nn − mn = ðnb + nn Þ + mnr + nn − mn = 2nn + nb + mnr − mn

30 Bruce J. MacLennan

input qubits. The mnr constant j0is are passed directly to the output qubits for non-
range codomain elements. The nb qubits for injective inputs are passed to the same
number of output qubits, permuted as required. The outputs of the ORs project onto
the mn qubits that represent range values with more than one preimage. Each OR2
also generates two garbage qubits, for a total of 2(nn − mn). Therefore the total num-
ber of output qubits is

mnr + nb + mn + 2ðnn − mn Þ = 2nn + nb + mnr − mn ;

which is equal to the number of input qubits, as it should be. Next we define Uf
explicitly as the tensor product of three operators:
def
Uf = Ub #Unr #Un . (2:30)

We will use the notation j1iq to represent a j1i state in qubit q, and j0iq to represent
a j0i state in the qubit q.
The Unr operator is an identity operation copying constant j0i ancillae into the co-
domain elements that are not in the range of f. Therefore, let fc1 , . . . , cmnr g  Ω′ − Im f
be this subset, and let jzi i be ancillae qubits to provide constant j0is. Then
Unr ∈LðHmnr , Hmnr Þ is defined:
X
mnr  
Unr =
def 0i h0j + 1i h1j . (2:31)
ci zi ci zi
i=1

That is, the states of the zi input qubits (intended to be j0i) are transferred to the ci
output qubits. This operator can be abbreviated by the following bracket notation:
def
Unr = ½c1 , . . . , cmnr  ½z1 , . . . , zmnr . (2:32)

It is just a permutation of the qubits, which could be implemented by qubit SWAP

operations.
The Ub operator handles the domain elements that are mapped injectively to
their images. Therefore, let fv1 , . . . , vnb g  Im f be the injective domain elements,
and let ui = f (vi) be the corresponding range elements. Then Ub ∈LðHnb , Hnb Þ is a
permutation of this subset of the topographic map elements:
h i h i
def
Ub = u1 , . . . , unb v1 , . . . , vnb . (2:33)

For Un we must OR together the domain elements that correspond to each range
element with more than one preimage. Therefore we define Un as a tensor product
of operators for each such range element:
def  −1

Un = # mi = 1 Vi y i , f
n
fy i g , (2:34)
 2 Topographic representation for quantum machine learning 31

where these yi each have more than one preimage element; for example, f − 1 fyi g =
 
x1 , . . . , xni , where ni = jf − 1fyi gj ≥ 2. The output state ψi i for such a range element

is the OR of the input states ξ j i (j = 1, . . . , ni ) of its preimage elements:
    
ψ iγi = ORn ξ i    ξ i1i#ðni − 1Þ ,
i i 1 ni

where ORni is a cascade of ni − 1 OR2 gates and jγi is 2(ni − 1Þ dimensional garbage.
    
This is accomplished by the operators Vi yi , x1 , . . . , xni ∈L H2ni − 1 , H2ni − 1 :
   defh i  
Vi yi , x1 , . . . , xni = yi , γ1 , . . . , γ2ðn − 1Þ ORni x1 , . . . , xni ; o1 , . . . , oni − 1 ,
i

(2:35)

where o1 , . . . , oni − 1 are the qubits that provide ancillary j1is for the ORs, and the
garbage outputs γ1 , . . . , γ2ðn − 1Þ receive the negated inputs and intermediate OR
i
outputs. The bracket notation identifies the qubits that are the inputs and outputs
of ORni . This completes the construction of Uf [eq. (2.29)].
There are more efficient ways to compute Uf, but the above construction is
easier to understand. This basic approach can be used to approximate a variety of
unary functions useful in neural networks, such as sigmoid functions, including
non-injective, non-surjective squashing functions. However, neural networks also
require non-unary functions such as addition and multiplication, to which we
now turn.

2.4.3 Binary functions

In sensory cortical areas there are many topographic maps that represent two or
more dimensions of a stimulus (e.g., retinal position and edge orientation); local-
ized activity in these maps represent conjunctions of values on these dimensions.
Similarly, quantum computational maps can represent conjunctions of values as in-
puts or outputs of functions.
Suppose we want to compute a function f :Ω × Ω ! Ω′, where Ω = fx1 , . . . , xn g
and Ω′ = fy1 , . . . , yn g. We will represent the input to the function by a two-
dimensional array of qubits for each (xj, xk) pair. (They do not have to be physically
arranged as a two-dimensional array so long as there is a qubit for each pair of val-
ues.) This will require n2 qubits, but we are assuming that n is small because low
precision is adequate for neural networks. Therefore we expect the 2D map to com-
prise typically several thousand qubits. The qubits representing the (xj, xk) pairs are
then mapped to the qubits representing the outputs f (xj, xk) by the method de-
scribed in Prop. 5.
The n2 conjunctions are computed by n2 CCNOT gates, each of which requires a
j0i ancilla and generates two extra qubits (containing the inputs) in addition to the
32 Bruce J. MacLennan

conjunction. However, these extra qubits are passed along the rows and columns to
be used in other conjunctions, and so there are only 2n total garbage qubits. In sum-
mary, there are 2n + n2 input qubits (including n2 ancillae) and n2 + 2n output qubits
(including 2n garbage). That is, if jϕj i is the state of element j of one input map, and

ψ i is the state of element k of the other input map, then the state jχ i of element
k jk
(j, k) of the two-dimensional map is computed by
     
ϕ iψ iχ i = CCNOTϕ iψ i0i. (2:36)
j k jk j k


If ϕj i = p′j 0i + pj1i and jψk i = q′j0i + qj1i, then the result of CCNOT is

jϕj ijψk ijχjk i = p′q′j000i + p′qj100i + pq′j010i + pqj111i.

The qubits are entangled, but the conjunction computes probability-like amplitudes
if we interpret the squares of the amplitudes as probabilities.
Based on the foregoing, we define a unitary operator UOP on a n2 + 2n dimen-
sional Hilbert space that does what amounts to an outer product on two one-
dimensional maps to compute a two-dimensional map:
2
jφijψijχi = UOP jφijψij0i#n , (2:37)

where inputs jφi and jψi are n-dimensional, and output jχi is n2-dimensional and
computed by eq. (2.36).
To illustrate the use of computational maps to implement binary operations,
we will use a simple, useful function, addition. We want to define sum:Ω × Ω ! Ω′
so that sumðx, yÞ = x + y, but we have a problem, since the maximum value of x + y
is greater than the maximums of x and y. Since the range of numbers represented
by our maps is quite limited, this is a more serious problem than overflow in
binary addition. One solution is to make the codomain map large enough; for
example, if Ω = f0, Δx, 2Δx, . . . , ðn − 1ÞΔxg, then let Ω′ = f0, Δx, 2Δx, . . . , 2ðn − 1ÞΔxg.
Generally, however, it is more convenient to have the codomain map be the same
as the domain maps, since this facilitates composing functions. Therefore, another
solution is to scale either the inputs or the output so that we compute, for example,
hsumðx, yÞ = ðx + yÞ=2; this is often useful if we know that we are going to scale the
quantities anyway. A third option is to compose addition with a squashing function,
so that we compute, for example, the truncated tsumðx, yÞ = minðx + y, xn Þ, where
xn = max Ω. This is the solution that we will use for illustration.
If Ω = f0, Δx, . . . , ðn − 1ÞΔxg, then the (j, k) element of the two-dimensional qubit
map will represent the pair of inputs ððj − 1ÞΔx, ðk − 1ÞΔxÞ. This will be mapped to
the sum ðj + k − 2ÞΔx if j + k − 2 < n − 1, and to the maximum value ðn − 1ÞΔx other-
wise. Therefore the constant j + k anti-diagonals above the j + k = n + 1 anti-diagonal
each map to one value, ðj + k − 2ÞΔx, which is the sum of ðj − 1ÞΔx and ðk − 1ÞΔx.
The elements below the j + k = n + 1 anti-diagonal all map to the same ðn − 1ÞΔx
 2 Topographic representation for quantum machine learning 33

value, since the sum is above the maximum representable value in the topographic
map.
Proposition 5 shows how to implement the truncated addition, but it treats it as
a unary function on an n2 -dimensional space, which is wasteful since the intended
output (the sum) is n-dimensional and the remaining n2 − n elements are garbage.
Therefore, we implement a unitary operator that directly maps the input pairs to
the corresponding outputs.
To compute the outer product we require n2 constant j0i ancillae, and this com-
putation also passes the two n-dimensional inputs through as garbage output. The
qubit representing (0, 0) maps bijectively to the output qubit y1 representing 0.
Each of the other n − 1 output qubits yi (i = 2, . . . , n) has two or more domain pairs
mapping to it. As before, let ni be the preimage multiplicity of output i and note
P
that ni= 1 ni = n2 . Each of these non-injective outputs receives its value from an ORni
operation, which requires ni − 1 input j1i ancillae and generates 2ni − 2 qubits of out-
put garbage (ni for the negated inputs and ni − 1 for the intermediate disjunctions).
Therefore, the total number of j1i ancillae is
X
n X
n  
ðni − 1Þ = ni − ðn − 1Þ = n2 − 1 − n + 1 = n2 − n.
i=2 i=2

Moreover, the complete input dimension is 2n + n2 + n2 − n = 2n2 + n.

This is also the complete output dimension, for we have n qubits for the func-
tion value, 2n qubits for the passed-through input arguments (garbage), and the
garbage output from the OR gates, which is:
X
n  
ð2ni − 2Þ = 2 n2 − n .
i=2

That is, the complete output dimension is 3n + 2ðn2 − nÞ = 2n2 + n.

2 2
In summary, there is a unitary operator Utsum ∈L H2n + n , H2n + n so that
2
Utsum jfxgi jfygi j0i#n j1i#nðn − 1Þ = jftsumðx, yÞgi jfxgi jfygi jγi, (2:38)

where the garbage jγi has dimension 2nðn − 1Þ (the passed-through inputs may also
be considered garbage).
Based on this example, we state a more general result.

def
Proposition 6 Suppose f :Ω × Ω ! Ω, and let n = jΩj. Let mnr = n − jIm f j be the
number of codomain elements that are not in the range of f. Then there is a unitary
operator Uf ∈LðH, HÞ, where H is 2n2 + n + 2mnr dimensional Hilbert space, such that:

Uf jfxgi jfygi j0i#ðn nr Þ

j1i#ðn nr Þ
2 +m 2 −n+m
= jff ðx, yÞgi jfxgi jfygi jγi, (2:39)

where the garbage jγi has dimension 2ðn2 − n + mnr Þ (the 2n passed-through inputs
may also be considered garbage).
34 Bruce J. MacLennan

Proof: The operator is constructed very similarly to Utsum , but we also have to con-
sider non-range codomain elements for non-surjective functions, which didn’t
occur in that case. As before, the computation of the outer product will require n2
ancillary j0i inputs and it will generate 2n qubits containing the passed-through in-
puts. We can consider disjoint subsets of the codomain. Codomain elements that
are not in the range of f will need to be sent a j0i state, for which we need an addi-
tional mnr ancillary j0i inputs. Let nb be the number of input pairs that are mapped
one-to-one to the corresponding outputs; they neither require ancillary constants
nor generate garbage. The remaining codomain elements are range elements with
ni ≥ 2; let mn = n − mnr − nb be the number of them. Each of these will receive the OR
of the corresponding (preimage) domain elements. As we saw previously, the ORni
operation requires ni − 1 ancillary j1i qubits and produces 2ni − 2 qubits of garbage.
Therefore, the total number of j1i qubits required for the n2 − nb input pairs map-
ping to the mn non-injectively mapped range elements is:
X
mn X
mn
ðni − 1Þ = ni − mn = n2 − nb − mn = n2 − n + mnr , (2:40)
i=1 i=1

since mn = n − mnr − nb . The garbage generated by the ORs is then

X
mn
 
ð2ni − 2Þ = 2 n2 − n + mnr . (2:41)
i=1

The complete input dimension is 2n (arguments) + (n2 + mnr Þ (for j0i ancillae) +
(n2 − n + mnr ) [for j1i ancillae, eq. (2.40)] = 2n2 + n + 2mnr . The output dimension is 3n
(arguments and result) + 2ðn2 − n + mnr Þ [garbage, eq. (2.41)] = 2n2 + n + 2mnr .

The same approach can be used for operations with more than two arguments, but
the number of qubits increases exponentially with the number of arguments.

2.5 Conversions between representations

Ordinary binary representations can be translated to topographic qubit maps by a
unitary demultiplexer Udemux that operates on an m-qubit binary number jki and
directs a j1i qubit to the kth of n = 2m output qubits (the remainder receiving j0i).
Let jfkgi be the resulting computational qubit map. Then:
    
Udemux ki1i0i#ðn − 1Þ = kifkgi. (2:42)

Udemux operates on an m + n = m + 2m dimensional Hilbert space. A demultiplexer

can be implemented with CSWAP (Fredkin) gates [19].
The opposite translation, from a computational map to a binary representation,
is more complicated. First, we must decide what we want it to do, for in general a
 2 Topographic representation for quantum machine learning 35

topographic qubit map represents multiple values with different amplitudes,

  2  2
jψi = #nj= 1 ðpj ′0i + pj j1iÞ, where pj ′ + pj  = 1. Which xj should it produce? The one
with the maximum amplitude? (And what if several have the same maximum ampli-
 2
tude?) An xj chosen probabilistically based on pj  ? The binary representation of a
P
weighted average, n − 1 nj= 1 pj xj ? A normalized superposition of all the values? The
answer is not apparent, so we leave the question open.4

2.6 Applications to quantum machine learning

Given this general ability to compute non-unitary and even nonlinear functions by
means of topographic qubit maps, it is possible to do the operations useful for ma-
chine learning such as inner products and sigmoid nonlinearities. For example, an
inner product of N-dimensional vectors requires N multiplications and N − 1 addi-
tions. If |Ω| = n, then each multiplication and addition will require approximately
2n2 qubits (Prop. 6), for a total of approximately 2N 2 n2 .
For one layer of a neural network, with say N neurons projecting through an
M × N weight matrix into M neurons, we must do M inner products with the input.
Since crisp sets are represented by computational maps that are basis vectors in the
computational basis, they can be copied. Therefore, the N-dimensional input vector
can be copied M − 1 times to do the M inner products. (This requires M − 1 CNOT
gates and ðM − 1Þn ancillary j0i qubits. Overall, one layer requires about 2MN 2 n2 qu-
bits for the computation (not including ancillary qubits).

2.7 Conclusions
Topographic (computational) maps are widely used in the brain to implement si-
multaneous nonlinear vector transformations on multiple inputs. In this chapter we
have explored two approaches to quantum topographic computing with a focus on
brain-inspired machine learning applications. The first, called a topographic basis
map, assigns locations in the map to state vectors in a continuous or discrete basis
for a quantum Hilbert space. Arbitrary functions can be implemented on such
maps, which can be interpreted as representing crisp sets of inputs, but there is an
unavoidable data-dependent attenuation of the result (relative to a “garbage” state)
that is not easily avoidable. The second approach, called a topographic qubit map,
assigns a separate qubit to each location in the map and uses the relative amplitude

4 It is easy however to produce the binary representation of either the maximum or minimum num-
ber with unit amplitude ( pj = 1, p0j = 0) in a map.
36 Bruce J. MacLennan

of the j1i and j0i states to represent the presence of values in the (crisp or fuzzy) set
represented by the map. Arbitrary functions on these maps are implemented by
well-known quantum logic gates. In particular, computational maps enable the im-
plementation of the functions commonly used in artificial neural networks.

References
[1] Nielsen, M. A. and Chuang, I. L. 2010. Quantum computation and quantum information (10th
Anniversary Edition ed.). Cambridge, UK: Cambridge Univ. Press.
[2] Schuld, M., Sinayskiy, I., and Petruccione, F. 2014. The quest for a quantum neural network.
Quantum Information Processing 13:2567–2586. https://doi.org/10.1007/s11128-014-0809-8
[3] Schuld, M., Sinayskiy, I., and Petruccione, F. 2015. An Introduction to quantum machine
learning. Contemporary Physics 56:172–185. https://doi.org/10.1080/00107514.2014.
964942
[4] Furber, E. F., Lester, D. R., Plana, L. A., Garside, J. D., Painkras, E., Temple, S., and Brown,
A. D. 2013. Overview of the SpiNNaker system architecture. IEEE Transactions on Computers
62:2454–2467. https://doi.org/10.1109/TC.2012.142
[5] Neftci, E., Binas, J., Rutishauser, U., Chicca, E., Indiveri, G., and Douglas, R. J. 2013.
Synthesizing cognition in neuromorphic electronic systems. Proceedings of the National
Academy of Sciences of the United States of America 110:E3468–E3476. http://dx.doi.org/10.
1073/pnas.1212083110
[6] O’Reilly, R. C., Munakata, Y., Frank, M. J., Hazy, T. E., and contributors. 2014. Computational
cognitive neuroscience (2nd ed.). http://ccnbook.colorado.edu/ (accessed January 12, 2019)
[7] MacLennan, B. J. 2017. Field computation: A framework for quantum-inspired computing. In
Quantum Inspired Computational Intelligence: Research and Applications, ed. S. Bhattacharyya,
U. Maulik, and P. Dutta, 85–110. Cambridge, MA: Morgan Kaufmann / Elsevier.
[8] Morasso, P. G., and Sanguineti, V. 1997. Self-organization, computational maps and motor
control. Amsterdam: North-Holland.
[9] MacLennan, B. J. 1999. Field computation in natural and artificial intelligence. Information
Sciences 119:73–89.
[10] Eliasmith, C., Stewart, T. C., Choo, X., Bekolay, T., DeWolf, T., Tang, Y., and Rasmussen,
D. 2012. A large-scale model of the functioning brain. Science 338:1202–1205. http://doi.
org/10.1126/science.1225266
[11] MacLennan, B. J. 1997. Field computation in motor control. In Self-Organization,
Computational Maps and Motor Control, ed. P. G. Morasso and V. Sanguineti, 37–73.
Amsterdam: North-Holland.
[12] MacLennan, B. J. 2009. Field computation in natural and artificial intelligence. In
Encyclopedia of Complexity and System Science, ed. R. Meyers et al., Chapter 6, entry 199,
3334–3360. New York: Springer.
[13] MacLennan, B. J. 2014. The promise of analog computation. International Journal of General
Systems 43:682–696.
[14] Rumelhart, D. E., McClelland, J. L., and PDP Research Group. 1986. Parallel distributed
processing: Explorations in the microstructure of cognition, volume 1: Foundations.
Cambridge, MA: MIT Press.
[15] Sanger, T. D. 1996. Probability density estimation for the interpretation of neural population
codes. Journal of Neurophysiology 76:2790–2793.
 2 Topographic representation for quantum machine learning 37

[16] Høyer, P. 2000. Arbitrary phases in quantum amplitude amplification. Physical Review A
62:052304. https://doi.org/10.1103/PhysRevA.62.052304
[17] Chen, J., Wang, L., and Charbon, E. 2017. A quantum-implementable neural network model.
Quantum Information Processing 16:245. http://doi.org/10.1007/s11128-017-1692-x
[18] McClelland, J. L., Rumelhart, D. E., and PDP Research Group. 1986. Parallel distributed
processing: Explorations in the microstructure of cognition, volume 2: Psychological and
biological models. Cambridge, MA: MIT Press.
[19] Fredkin, E. F. and Toffoli, T. 1982. Conservative logic. International Journal of Theoretical
Physics 21:219–253.
Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee
3 Quantum optimization for machine
learning
Abstract: Machine learning is a branch of Artificial Intelligence that seeks to make ma-
chines learn from data. It is being applied for solving real world problems with huge
amount of data. Though, Machine Learning is receiving wide acceptance, however, ex-
ecution time is one of the major concerns in practical implementations of Machine
Learning techniques. It largely comprises of a set of techniques that trains a model by
reducing the error between the desired or actual outcome and an estimated or pre-
dicted outcome, which is often called as loss function. Thus, training in machine learn-
ing techniques often requires solving a difficult optimization problem, which is the
most expensive step in the entire model-building process and its applications. One of
the possible solutions in near future for reducing execution time of training process in
Machine learning techniques is to implement them on quantum computers instead of
classical computers. It is conjectured that quantum computers may be exponentially
faster than classical computers for solving problems which involve matrix operations.
Some of the machine learning techniques like support vector machines make extensive
use of matrices, which can be made faster by implementing them on quantum com-
puters. However, their efficient implementation is non-trivial and requires existence of
quantum memories. Thus, another possible solution in near term is to use a hybrid of
Classical Quantum approach, where a machine learning model is implemented in clas-
sical computer but the optimization of loss function during training is performed on
quantum computer instead of classical computer. Several Quantum optimization algo-
rithms have been proposed in recent years, which can be classified as gradient based
and gradient free optimization techniques. Gradient based techniques require the nature
of optimization problem being solved to be convex, continuous and differentiable other-
wise if the problem is non-convex then they can find local optima only whereas gradient
free optimization techniques work well even with non-continuous, non-linear and non-
convex optimization problems. This chapter discusses a global optimization technique
based on Adiabatic Quantum Computation (AQC) to solve minimization of loss function
without any restriction on its structure and the underlying model, which is being
learned. Further, it is also shown that in the proposed framework, AQC based approach
would be superior to circuit-based approach in solving global optimization problems.

Keywords: Artificial Intelligence, quantum computing, non-convex optimization

Ashish Mani, ASET, Amity University UP, Noida, Uttar Pradesh, India
Siddhartha Bhattacharyya, Department of Computer Science and Engineering, CHRIST (Deemed to
be University), Bangalore, India
Amlan Chatterjee, Computer Science, California State University, Dominguez Hills, USA

https://doi.org/10.1515/9783110670707-003
40 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

3.1 Introduction
Humanity has been enamored with the creation of artificially intelligent machines
since the advent of programmable electronic computing devices in last century. In
1956, the term ‘Artificial Intelligence’ was coined for ‘thinking machines’ in the
workshop held at Dartmouth [1]. Several proposals have been made to make
machines behave intelligently [2], but most of them suffered from non-availability
of technology required to implement them efficiently. Learning is considered as
manifestation of the intelligent behavior, so a machine which can learn a difficult
task will be treated at least artificially intelligent. The discipline which deals with
the learning techniques involving data is called as Machine learning. It tries to
make a computer learn from past experiences represented in from of data and make
reliable decisions from it. These techniques can fine tune their performance as fresh
experiences are gathered. It helps in building a data driven model rather than a
static precise model and is widely used in classification, clustering and pattern rec-
ognition etc. It is being widely used in real world fields like financial services [3],
health services [4], marketing and sales [5], social network analysis [6], and virtual
personal assistants [7] etc.
Some popular Machine Learning algorithms are Artificial Neural Networks
(ANN) [8], Support Vector Machines (SVM) [9], K-means Clustering [10], Regression
[11], Decision trees [12] etc. The generic steps involved in any machine learning proj-
ect are as follows:
Data collection: This step involves collecting data pertaining to the problem to
be solved. The data should be of high quality, which is necessary for a high quality
model. There are several measures available to evaluate the quality of data [13].
Low quality of data will generally result in low quality of model and subsequently
predictions or decisions.
Model selection: This step involves selection of a specific technique like ANN
or SVM etc. Further, it also involves selecting the structure of the model i.e. in
case of ANN, number of hidden layers, no. of neurons in each hidden layer, acti-
vation functions etc.
Training: In this step, the parameters of the model are decided i.e. in case of
ANN and SVM, weights are determined using some loss function i.e., a loss function
which represents the performance of the model on training data set is used to search /
compute the weights.
Validation: In this step, the performance of the model on testing data set is de-
termined and compared with other existing models.
Deployment: Finally the model is deployed so that it can be used for making
predictions or decisions in real situations.
The first step is specific to the problem being solved by using machine learning,
but can be automated in most cases, to make data collection process reliable, quick
 3 Quantum optimization for machine learning 41

and cheap. The second step of Model selection usually requires past experience and
trial and error approach, which can be translated into a heuristic [13]. The selection
of the model is often done more than once and is typically based on the empirical
evidence regarding the performance of the model, which is known after the valida-
tion. The third step of training is usually the most compute intensive approach of
all the other steps and depending on the complexity of model, it can be efficient
also if the underlying process is linear, then its modelling results in well-behaved
loss function i.e., loss function is convex, continuous and differentiable in the re-
gion of interest. However, most loss functions are often not well behaved due to the
underlying nonlinear process which is generating the data and in such cases, the
optimization of loss function becomes a key challenge in successful implementation
of machine learning technique. It has been often experienced that when loss func-
tion is well-behaved, the model is less useful [14] and when loss function is difficult
to optimize, then the resulting model is more powerful [15]. Validation and deploy-
ment are usually computationally less expensive as the trained model takes input
and gives output in the form of predictions or decisions. In this chapter, we will
primarily focus on making training more efficient in machine learning through
quantum optimization.
Optimization is applied routinely in almost all types of machine learning techni-
ques and its applications. Chambers dictionary describes optimization as an act of
‘making the most or best of anything’. Therefore, performing optimization in training
stage of machine learning means finding the most or best suitable parameters of the
model. There are machine learning models like linear SVM whose weights can be de-
termined by matrix inversion techniques [16], which can be implemented efficiently
in case of gate model of quantum computation, thus showing exponential speed-up.
However, such techniques often require existence of quantum RAM [17] and data to
be generated from a linear process with low noise. In real world problems, the as-
sumption of linearity is often violated and thus most popular techniques are non-
linear and convexity cannot be guaranteed. If the optimization problem is convex
then gradient based techniques are guaranteed to find the global optima regardless
of the initialization [18]. So many optimization techniques try to simplify the problem
so that they appear linear or convex [13]. Alternatively, they try to solve the problem
approximately by using techniques like stochastic gradient descent [19], which are
iterative and not the desired single shot techniques. Thus, these techniques are usu-
ally computationally time-consuming as the nature of the problem they are trying to
solve is non-linear, non-differentiable, multi-dimensional, multi-modal, and stochas-
tic. Several quantum enhanced approaches have been proposed to solve computa-
tional issues in machine learning, of which global optimization using adiabatic
quantum computation (AQC) is a promising technique, especially for problems which
are non-linear, non-differentiable, multi-dimensional, and multi-modal. Efforts have
also been made to solve quadratic unconstrained binary optimization problems on
AQC model [20]. In this chapter, we propose an adiabatic quantum computing based
42 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

iterative global optimizer for minimizing the loss function of machine learning tech-
niques, which would be useful when the optimization problem is non-linear, non-
differentiable, multi-dimensional, and multi-modal, where gradient based techniques
like [19] and single shot techniques [21] are not computationally efficient or relevant.
Further, there are some Variational Quantum computing proposals, where they are
focusing on implementing function evaluation on a near term quantum computing
device and implementing a traditional gradient free optimization technique like
Nelder-Meads method on classical computer [19], however, our proposal is different
from such proposals as the entire optimization step will be run on a future Adiabatic
Quantum Computer. This chapter is further organized as follows: Related work about
the development of AQC based algorithms is discussed in Section 3.2. The theoretical
underpinning of Adiabatic Quantum Computing is discussed in Section 3.3. Section
3.4 presents the development of Grover’s Adaptive Search. Problem formulation is de-
scribed in Section 3.5. Section 3.6 presents quantum search based on AQC. Minimum
finding algorithm based on AQC is described in Section 3.7. Section 3.8 discusses the
framework for implementing quantum optimization algorithm based on AQC.
Conclusions and some directions for future work are presented in Section 3.9.

3.2 Related work

Adiabatic quantum algorithms were introduced by Farhi et al. [22] based on the adia-
batic conditions derived from Folk theorem. These algorithms have a Hamiltonian
constructed for the problem being solved which is called as Problem Hamiltonian
(Hfin). The ground state of the Hfin can be expressed as the optimal solution of the
problem. Further, initially the system is prepared in the ground state of an initial
Hamiltonian, which is easy to build. Subsequently, the initial Hamiltonian is changed
progressively towards the Hfin so that the probability of finding the solution is large
at the end of the evolution. However, the change from initial Hamiltonian to final
Hamiltonian i.e. Hfin cannot be instantaneous or even lesser than the minimum time
required according to the adiabatic theorem. If the system is evolved slowly enough
then adiabatic theorem guarantees that system will remain in the ground state of the
instantaneous Hamiltonian, thus the evolution time of varying the Hamiltonian from
initial to problem is a measure of complexity of the AQC based algorithm. Several
constraint satisfaction problems along with optimization problems can be repre-
sented by an aggregated set of local Hamiltonians H = ∑i H i where Hi represents a
local constraint. The ground state of H would satisfy the maximum possible number
of such constraints and represents a desired optimal solution amongst other possible
optimal solutions in case of multimodal problems, i.e. the problems having more
than one optima solutions. The system stays in the instantaneous ground state of the
evolving Hamiltonian as it will remain close to the ground state provided the rate of
 3 Quantum optimization for machine learning 43

change of Hamiltonian from initial to Hfin remains within the bounds specified by the
adiabatic theorem. The bounds on the evolution time is dependent on the energies of
the initial and problem Hamiltonian as well as on the inverse gap of the instanta-
neous Hamiltonians H(t). The energy difference between the ground state of a
Hamiltonian and its first excited state represents the gap of the Hamiltonian. That is
when Hamiltonian is presented as a matrix then the gap is the difference between
Hamiltonian’s smallest and second smallest eigenvalue. Farhi et al. [22] have pro-
posed and investigated AQC based algorithms for NP-complete problems like 3SAT.
However, the key challenge in AQC based algorithm for solving a specific problem
is analytical calculation of the Hamiltonian gap, which is not known for most of the
problems till date. The other approach is to numerically simulate the system but
the computational cost of simulating quantum mechanical system on classical com-
puters is very expensive so the number of qubits in numerical simulations is limited.
However, the numerical simulation performed in [22] on small sized problems did
show that the time required to solve NP-complete problems scaled only polynomially
with the problem size. But later work [23] gave strong negative evidence for the speed
up shown in [22] for AQC based algorithms. It was shown that the performance of
adiabatic algorithms is even sensitive to the choice of initial Hamiltonian and it can
fail if the selection of initial Hamiltonian is not performed according to the struc-
ture of the problem [23]. Further, some investigations were made in [24] by map-
ping the Hfin of certain instances of 3-SAT to an Ising model and it was shown that
the Hamiltonian gap is exponentially small in some problem instances. Further,
some special instances of 3-SAT were also constructed in [25] which were hard for
the AQC based algorithm. Subsequently, it was shown in [26, 27] that local mini-
mum of some optimization problems can cause very small gaps in the spectrum of
the Hamiltonian. These studies had designed hard instances of problems to high-
light the weakness of AQC based algorithms. But in [28], randomly generated in-
stances of the NP-complete problem Exact Cover 3 (EC3), also known as 1-in-3 SAT
were used to test AQC based algorithm and it was shown that failure probability of
AQC based algorithm was high. Altshuler et al. argued that their study provided
strong evidence against adiabatic quantum optimization, i.e. hard instances of NP-
complete problems are difficult for AQC based algorithm but the basis for this conclu-
sion depends on some common properties shared by other NP-complete problems
such as 3-SAT. However, initially Choi [29] and later Dickson and Amin [30] chal-
lenged the argument against AQC in [28]. Choi’s argued that a failure of a specific
instance of AQC based algorithm is not sufficient to conclude that a problem is
hard for adiabatic quantum computation / optimization in general. An AQC based
algorithm comprise of three variable components, namely, initial Hamiltonian,
problem Hamiltonian and evolution path. Thus, there are many combinations
possible of these three components for designing an AQC based algorithm. He fur-
ther argued that if it is to be proved that adiabatic quantum computation is ineffi-
cient even for a particular problem then, it has to be shown that there exists no
44 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

polynomial-time adiabatic quantum algorithms for the problem, which requires testing
all combinations possible for designing AQC based algorithms and not just few instan-
ces. Further, Dickson and Amin [30] analytically showed that there always exist adia-
batic paths for the NP-hard problem of maximum independent set along which no
such crossings of local minima of the final Hamiltonian with its global minimum
occur, thus the Hamiltonian gap need not be vanishingly small and hence designing
of efficient AQC based algorithm is possible. Therefore, in order to prove that adiabatic
quantum optimization fails for any NP-complete problem, one must prove that it is im-
possible to find any such path along with combination of initial and problem / final
Hamiltonian in polynomial time. Therefore, it is still an open problem whether AQC
can solve NP complete problems efficiently. However, it is becoming increasingly clear
that there may not exit one unique path for solving all the NP complete problems.
Quantum unstructured search has been implemented adiabatically by Farhi
et al. [31]. This implementation was based on global evolution model and had a
computational complexity of O(N), where N is the size of the database. Later Roland
and Cerf showed that quadratic speed up is possible in adiabatic version also by
continuously adjusting the rate with which the initial Hamiltonian is switched to
final Hamiltonian by only fulfilling locally, the conditions of adiabaticity [32]. The
adiabatic quantum search has been rigorously studied by Rezakhani et al. [33]
where the results relating the accuracy and the run time of the AQC based algorithm
has been determined. It has also been demonstrated that AQC based algorithm
shows two discernible regimes with respect to error i.e., the error reduces exponen-
tially for short times and then reduces polynomially for longer times, under fairly
general conditions.
In [34], it was demonstrated that a partial adiabatic search algorithm on quan-
tum circuit model could be implemented and the performance of the algorithm was
same as that of the local adiabatic algorithm for the case where a single element is
marked i.e. M = 1 by investigating the minimum energy gap between the first ex-
cited state and the ground state of the system Hamiltonian. Moreover, the algorithm
keeps the advantages of global adiabatic algorithms without losing the speedup of
the local adiabatic search algorithm by evolving globally only for a small interval
during evolution from initial to final / problem Hamiltonian.
In [22], it was shown that quantum circuit models can simulate efficiently an
arbitrary AQC based algorithm. Subsequently, it has also been shown in [35] that
an appropriate AQC based algorithm can efficiently simulate any quantum circuit
model. Further a simple proof of equivalence between adiabatic quantum computa-
tion and quantum computation in the circuit model has been provided by Mizel et al.
[36]. Therefore, AQC and Quantum Circuit models of computation are essentially
equivalent, which implies that a quantum algorithm can be designed in either of the
two models for solving difficult problems. Brady and Van Dam has also investigated
the necessary run-time required for AQC based optimization [37] and showed that
nonadiabatic speedup does not occur for general cases but only for special case.
 3 Quantum optimization for machine learning 45

Recently, Quantum Approximate Optimization Algorithm (QAOA) [38, 39] and the
Variational Quantum Eigensolver [40] have been proposed to address classical com-
binatorial optimization and quantum chemistry problems, respectively.
In this chapter, a framework based on Adiabatic Quantum Computation for
solving global optimization problem has been proposed in Section 3.4., which
would be helpful in training of Machine Learning techniques which models non-
linear processes unlike their computationally efficient counterparts. Further, con-
strained optimization using AQC was discussed in [41].

3.3 Adiabatic quantum computation

Adiabatic quantum computation [42] is an alternative approach to quantum comput-
ing that has been shown to be equivalent to the quantum circuit model, but appears
to have some advantages over the circuit model [31]. It is based on a well-known tech-
nique in quantum mechanics, called adiabatic approximation [43]. It is used for de-
termining approximate solutions of Schrodinger equation subject to the condition
that the Hamiltonian governing the dynamics of the system is changing slowly with
time. This implies that if a quantum system is prepared in its nth eigenstate, then it
will stay in the nth eigenstate throughout the evolution provided the neighborhood
energy gaps are large enough and its Hamiltonian varies slowly enough.
Let a quantum system in state |φ〉 evolve according to the Schrodinger’s equation:

d
ih jΦðtÞi = H ðtÞjΦðtÞi (3:1)
dt

where H(t) is the Hamiltonian of the quantum system and ћ is reduced Planck’s
constant [31]. Let Hini be the Hamiltonian of the system in its initial state i.e. at t = 0
and Hfin be the Hamiltonian of the system in its final state i.e. at t = T, where

HðtÞ = ð1 − sðtÞÞ Hini + sðtÞHfin (3:2)

where s(t) is a continuous function with s(0) = 0 and s(T) = 1. The instantaneous
eigenstates associated with Hi are denoted by jψi i. In general, let eigenstates of H(t)
be jψk ðtÞi such that:

HðtÞjψk ðtÞi = Ek ðtÞjψk ðtÞi (3:3)

where Ek(t) are the corresponding eigenvalues and k labels the eigenstates (k = 0
indicates the ground state).
The minimum gap between the lowest two eigenvalues is defined as follows:

gmin = min ½E1 ðtÞ − E0 ðtÞ (3:4)

0≤t≤T
46 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

The maximum value of the matrix element of dH/dt between the two corresponding
eigenstates is given by the following equation:
  
 dH 
 
Dmax = max   (3:5)
0 ≤ t ≤ T  dt 1, 0 

  dH 
 
where dH dt 1, 0 = hψ1 ðtÞ dt ψ0 ðtÞi
The adiabatic theorem states that if the system at time t = 0 is in its ground
state jψini i = jψ0 ð0Þi and let it evolve under the Hamiltonian H(t), then
 
hψ ðT ÞjΦðT Þi ≥ 1 − ε2 (3:6)
0

provided:
Dmax
2
≤ε (3:7)
gmin

where 0 < ε  1,
and
 
hψ ðtÞdH ψ ðtÞi
 1 0 
 dt
 < ε, t ϵ ½0, T . (3:8)
 ½E1 ðtÞ − E0 ðtÞ2 

where ε  1.
The above conditions for adiabatic computation are based on Folk theorem,
which has come under criticism recently [44]. These conditions do not guarantee
general applicability of Adiabatic approximation. A number of attempts have been
made to rigorously study Adiabatic Theorem and provide new conditions for ge-
neric applicability of adiabatic approximation [45]. Tong et al. have also proposed a
set of conditions, which are sufficient for verifying the validity of adiabatic approxi-
mation [46]. These are three conditions:

 
hψn ðtÞjψ′m ðtÞi
 
I.  E ðtÞ − E ðtÞ   1, t ϵ ½0, T  (3:9)
n m

ðT 

 hψn ðtÞjψ′m ðtÞi ′
 
II.  E ðtÞ − E ðtÞ dt  1 (3:10)
n m
0

ðT 

 hψn ðtÞjψ′m ðtÞi  
 ′ ′ 
III.  E ðtÞ − E ðtÞ  hψn ðtÞjψl ðtÞi dt  1 (3:11)
n m
0

where m ≠ l, n ≠ m, T is total evolution time, jψk ðtÞi and Ek(t) are eigenvectors and
corresponding eigenvalues of H at time t, k = 1,. . .,N (N is the size of state space of
the system, which includes l, m and n). ψ′m is first derivative of ψm.
 3 Quantum optimization for machine learning 47

A rigorous study has been made of applicability of adiabatic approximation on

the basis of these three conditions in case of Grover’s Quantum Search under AQC
framework in [46]. It has been found that the results obtained originally by Folk
theorem is still valid. As this chapter, primarily uses Grover’s search, so for rest of
the chapter, the results of Folk theorem will be applied.
Grover’s Quantum Search algorithm [47] and its variants can be used as a subrou-
tine in designing algorithms for solving optimization problems. The variants of Grover’s
search algorithm, which do not require prior knowledge of the number of target entities
like Boyer’s proposal [48] and fixed point schemes [49], are especially suited for such
problems. Durr and Høyer proposed an algorithm in [50] for finding minimum value in
a database by employing the algorithm proposed in [48] as a subroutine.
The quantum algorithm for locating minima searches for the minimum entry
in an un-arranged table ‘D’ of K items, where every item has a value belonging to
an ordered set. The pseudo code for quantum minimum searching algorithm is as
follows:
i. Randomly select a threshold index Jin ∈ [0,K−1].
ii. Repeat till the total running time is more than 22.5√K + 1.4 lg2 K.
a. Apply the Boyer et al. quantum searching algorithm [48].
b. Observe the first register: let D[Jin′] be the outcome.
c. If D[Jin′] < D [Jin], then set threshold index Jin to Jin′.
iii. Return Jin.

The quantum minimum searching algorithm finds the index of the minimum value
and it forms the inspiration for further developing optimization algorithms and heu-
ristics, so in a way, was a path breaking research.

3.3.1 Oracle

A classical oracle of a function f(x) can be implemented reversibly as shown in the

Figure 3.1. Similarly regular quantum oracle can be implemented as shown in
Figure 3.2 [51]. However, it differs from the reversible oracle as it may be given
input in some superposition of basis states and produces output that may be en-
tangled and in some superposition of basis states.

3.4 Grover’s adaptive search

Baritompa et al. generalized the quantum minimum search algorithm for solving
finite global optimization problems, which have been formulated as follows [52]:
48 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

Classical
a a
Reversible

Oracle
b b⊕ f(a)
Query

Figure 3.1: Reversible implementation of Classical Oracle.

| a〉 Quantum | a〉

Oracle

| b〉 Query | b ⊕ f(a)〉

Figure 3.2: Quantum Oracle.

Optimize fðαÞ, α 2 A (3:12)

where f is a real valued function on a finite set A.

Let the cardinality of set A be N. Let a1 < a2 < a3 . . . < aL be ‘L’ distinct values of
the objective function f(α) for all α in A. Therefore, L should be less than or equal to
N i.e. if L would be equal to N then for every value of α in set A, there would be a
distinct objective function value of f. Let there be a uniform probability measure λ on
A, such that ϒ be the range measure given by a stochastic vector (ϒ1, ϒ2, . . ., ϒL) in-
P
duced by f, so that ϒj = |f−1(aj)|/N for j = 1, 2, . . ., L. Let Pj = ji = 0 ¡j , the probability
that a random point has value of ‘aj’ or less, i.e. PL = 1. Therefore, there is an improv-
ing region corresponding to all values of objective function f(α), which belongs to a
sub-set of A, where f(α) has strictly better value [52]. The measure of improving region
under λ is denoted as improving fraction P. So if we choose a particular value for
objective function, say ‘θ’, then the improving region would comprise of set of all the
point whose objective function values are better than ‘θ’.
The algorithm proposed by Baritompa et al. is known as Grover Adaptive search
(GAS), which takes a sequence of rotation counts and an objective function value
threshold as the input. Its pseudo code is as follows [52]:
i. Select α1 in A, and set θ1 = f(α1).
ii. Set count nc = 1.
 3 Quantum optimization for machine learning 49

iii. While (termination condition is not met)

{
a. Perform a Grover search of rnc rotations on f with threshold Vnc, and de-
note the outputs by α and θ.
b. If v < Vnc, set θnc+1 = α and Vnc +1 = v; otherwise, set θnc+1 = θnc and Vnc +1 = Vnc.
c. nc = nc + 1.
}

GAS has been studied using an adaptive search framework developed in [53–55]
which has been applied for theoretically analyses of convergence properties of opti-
mization methods that use stochastic global search. One of the underlying assump-
tions of the adaptive algorithms is that search can find better (“improving”) solutions
(at some cost). GAS is based on an oracular function that can find better solutions
than the current one with a finite probability in every iteration, if one exists. This
oracle takes an input string α ∈ A, computes f(α), compares f(α) with current
threshold and outputs zero or one depending on the result of comparison and the
type of optimization problem being solved i.e. whether it is a minimization or
maximization problem. Such oracles can be implemented using reversible / quan-
tum circuits as shown in Section 3.3.1.
Thus, GAS generates a series of domain solutions, which are uniformly distrib-
uted in the improving region of the previous solution. Further, according Theorem 1
and Corollary 6 given in [54], such a series converges to a global optimal solution
rapidly; e.g., a unique optimal solution in a domain of size K can be located after
1 + ln(K) such improvements, in expectation [54].

Theorem 1 [54]. The expected number of iterations, I, required for solving the finite
optimization problem described by eq. (3.12) is:
XN
¡i
(a) I = 1 + for strong Pure Adaptive Search.
P
i=2 i
XN
¡i
(b) I = 1 + for weak Pure Adaptive Search.
P
i=2 i−1

where Strong Pure Adaptive Search implies that each iteration of the search algo-
rithm should sample objective function value from a strictly improving region. The
weak Adaptive search implies that each iteration of the search algorithm should
sample objective function value that is either equal or better.

Corollary 6 [54]. The expected number of iterations, I, for solving finite global opti-
mization described by eq. (3.12), and given a uniform distribution on the objective
function values, is:
50 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

(a) bounded above by 1 + logN for strong Pure Adaptive Search.

(b) bounded above by 2 + log(N−1) for weak Pure Adaptive Search.

GAS corresponds to the weaker version of Pure Adaptive search as it cannot be

guaranteed that after application of Grover Search, an strictly better solution
would be found. It can be argued that Grover Search can return an objective func-
tion value, which is less than the threshold and so it does not corresponds to Pure
Adaptive Search Framework, but, in such a case, the inferior value would be dis-
carded and the current threshold would be treated as the output of the algorithm
for the particular iteration. Thus, the overall GAS corresponds to weaker version
of Pure Adaptive Search.
An important design decision of Grover’s Adaptive Search is determination of the
rotation count sequence. It is possible to determine the optimal rotation count ‘r’ for
every iteration of the GAS provided the information regarding improving fraction ‘P’
of the domain is available. Thus, if the optimal rotation count is used for every run of
Grover’s Search, then a better (improved) solution will almost certainly be found in
every iteration of the GAS. However, in real world problems, the improving fraction P
is often unknown. An estimate of the improving fraction P can always be made by
using all the information available at the end of an iteration and considering the na-
ture of the problem. That is a Bayesian approach can be applied by tracking the im-
proving fraction at every iteration through the sequence of posterior distributions,
which will be subsequently used for choosing rotation count at the particular itera-
tion. However, this kind of approach might be very complicated and cumbersome
[56]. Thus, other methods have been devised to strike a balance between ease of im-
plementation and optimality of rotation count selection, which includes method
based on Boyer et al. [48] used by Durr and Hoyer illustrated in [50].
Another method proposed by Baritompa et al. is based on maximizing the bene-
fit to cost ratio denoted by parameter b in the heuristic [52]. The benefit has been
taken to be the expected decrease in the improving fraction of the domain, which is
difficult to ascertain without knowing the nature of the problem being solved. The
cost is dependent on the rotation count. The pseudo code for heuristic is as follows:
I. Initialize u0 ðΩÞ = Ω // u0 is a polynomial
II. For i = 1, 2, . . . , do: // Generating rotation count r for ith iteration
ð1
A. Eui = 1 − ui ðΩÞdΩ
0
B. β = 0
C. For r = 0, 1, . . . , until Eui =ðr + 1Þ < 2β, do:
ð1
gr ðPÞ
(i) vðΩÞ = ui ðΩÞ + Ω dui ðPÞ ==gr ðPÞ ¼ sin2 ðð2r þ 1Þsin−1 √PÞ
Ω P
ð1
(ii) Ev = 1 − vðΩÞdΩ
0
 3 Quantum optimization for machine learning 51

ðEui − Ev Þ
(iii) b=
r+1
(iv) If b > β, then
a) ȓ = r
b) β = b
c) ui + 1 ðΩÞ = vðΩÞ
D. Output ȓ

The sequence of rotation counts produced by the heuristic is independent of the

particular optimization problem and is as follows for first 33 entries [52]:

0, 0, 0, 1, 1, 0, 1, 1, 2, 1, 2, 3, 1, 4, 5, 1, 6, 2, 7, 9, 11, 13, 16, 5, 20, 24, 28, 34, 2, 41,

49, 4, 60, . . . .

Recently, the above heuristic method has been further studied and improved by Liu
and Koehler [56]. One improvement has been to simplify the calculations involved
in computing distributions by employing Chebyshev polynomials, which consider-
ably reduced the computation time. The other improvement has been to convert the
above heuristic into a dynamic one, which takes into account whether the Grover
search in a particular iteration was successful or not. This has been made possible
by using Bayes Law to update the distribution after one complete round of Grover
Search depending on success or failure of finding a better solution. The heuristic
for computing dynamic rotation schedule is as follows:
I. Initialize u0 ðΩÞ = Ω
II. For i = 1, 2, . . . , do: // Generating rotation count r for ith iteration
ð1
A. Eui = 1 − ui ðΩÞdΩ
B. β = 0 0
C. For r = 0, 1, . . . , until Eui =ðr + 1Þ < 2β, do:
ð1
(i) b = Pgr ðPÞdui ðPÞ
0
(ii) If b > β, then
_
a) r = r
b) β = b

D. Output ȓ

E. If (Search_outcome = Success)
52 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

Ðy Ð1
0 gr ðPÞdui ðPÞ + y Py gr ðPÞdui ðPÞ
ui + 1 ðΩÞ = Ð1
0 gr ðPÞduc ðPÞ

Else
Ðy
ui ðΩÞ − 0 gr ðPÞdui ðPÞ
ui + 1 ðyΩÞ = Ð1
1 − 0 gr ðPÞduc ðPÞ

Liu and Koehler have reported that formal derivation of the expected time required
by the proposals, made by Durr and Hoyer [50] (based on Boyer et al. algorithm
[48]), Baritompa et al. [52] and their own improved and dynamic version [56] of
quantum algorithms for global optimization, is hard to derive. This is because the
basic subroutine employed for searching needs knowledge of the distribution func-
tion of the improving fraction. Therefore, numerical simulations have been used to
compare the methods. However, such simulation studies are computationally very
expensive as they require digital computer to simulate a quantum computer. Thus,
the performance of such algorithms could be investigated on only problems of small
size, which have not much practical significance.
This chapter identifies that basic problem in the above global optimization al-
gorithms is due to the dependence of basic search subroutine on the improving
fraction to determine the rotation count in a particular iteration. This problem can
be solved by changing quantum computation model from quantum circuit model to
Adiabatic Quantum Computation (AQC) model.
AQC provides a distinct advantage over quantum circuit model on Grover’s
Search i.e. it eliminates the problem of overshooting. This is possible because the so-
lution is encoded as the ground state of the Problem Hamiltonian, therefore, even if
the system is evolved for time period longer than required, the system remains in the
ground state of the problem Hamiltonian. This feature of AQC makes it possible to
formally derive the expected running time of Global Optimization Algorithm based
on AQC.

3.5 Problem formulation

Optimization problems that are regularly solved in machine learning and engi-
neering domain are mostly in continuous domain. Such problems are classified as
unconstrained and constrained global optimization problems. The unconstrained
global optimization problems (UCOP) in continuous variables are generally formu-
lated as follows:
 3 Quantum optimization for machine learning 53

Optimize f(x) where x = (x1, x2, . . .., xE) ε RE ;

Such that:

xil < xi < xiu ; where xi is the ith variable with xil and xiu as its lower and upper limits.
(3:13)

The objective function f(x) as well as the inequality and the equality constraints are
often nonlinear, non-convex, non-differentiable, multimodal and of high dimen-
sion. Thus, optimization techniques, which are efficient by being deterministic, like
gradient based methods and exhaustive search, are often rendered ineffective for
solving such ill-defined problems [57]. Calculus based methods typically perform
local optimization by using search space domain information like local gradient
[58]. Enumerative Strategies suffer from the curse of dimensionality i.e. the runtime
complexity increases exponential with the dimension of the problem [59].
It has been shown in Section 3.4 that Quantum global optimization algorithms
can be used for solving finite search domain problems. The search domain of any
real variables problem is infinite. In order to overcome this limitation of quantum
algorithms, the real variables search space in [60] has been discretized. That is a
variable is represented by a finite number of qubits, which determines the resolu-
tion and the level of discretization. For example, the decision variables, xi, in con-
tinuous domain are bounded by xil and xiu. The search space for variable xi can be
discretized by taking ni qubits, so that there will be 2ni distinct values of xi which
can be selected between the range of [xil, xiu]. The maximum discretization error,
Δei, is as follows:

Δei = ðxiu − xil Þ=2ðni + 1Þ (3:14)

This discretization error is present even in computation performed with classical

digital computers. The total number of qubits, nq, required to map the discretized
solution space is as follows:
X
nq ¼ n
i i
(3:15)

The larger the number of qubits more is the accuracy, however, it also increases the
running time of the quantum algorithm. Therefore, there is a trade-off involved be-
tween accuracy and computation time. An output or auxiliary qubit is also required
which evolves according to the result of the computation of the objective function
implemented by a reversible circuit [52].
54 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

3.6 AQC based quantum search

Adiabatic Quantum Computation is analog analogue of discrete gate-based quan-
tum computing, which is a generic framework that depends on Hamiltonian and
Adiabatic theorem [31]. A Problem Hamiltonian (Hfin) is designed for the problem
being solved such that it is a time independent Hamiltonian, and the ground state
of Hfin, jψfin i, represents a solution of the problem. The quantum system is evolved
by a time dependent Hamiltonian, H(t). H(0) corresponds to a time independent
Hamiltonian known as initial Hamiltonian (Hini), whose ground state is easy to
construct. Hamiltonian H(t) is time-varying and governs the evolution of the sys-
tem and it varies according to the bounds given by adiabatic theorem, from Hini to
Hfin, so that the quantum system stays near the ground state of the instantaneous
Hamiltonian H(t) till it reaches Hfin. Thus, Adiabatic Quantum Algorithm has fol-
lowing main steps:
1) Determine the number of qubits, nq, required to represent the problem in the
quantum system.
2) Find an easily constructible initial ground state and construct its corresponding
Hini.
3) Design the Hfin according to the problem being solved.
4) Construct the time dependent Hamiltonian H(t), which will govern the evolu-
tion of the system for time T:

HðtÞ = Hini *ð1 − t=TÞ + Hfin *ðt=TÞ,

or,
^ ðsÞ = Hini *ð1 − sÞ + Hfin *ðsÞ,
H (3:16)

where s = t/T and s(0) = 0 and s(T) = 1.

5) Estimate total evolution time T.
6) Evolve the Quantum System according to Schrodinger’s Equation for time T:

djψðtÞi
ih = H ðtÞjψðtÞi (3:17)
dt
where ħ = 1.
7) Perform Measurement on the Quantum system after time T to obtain the solu-
tion of the problem.

Search for a marked element in an unstructured database with Ne elements was also
attempted by Farhi et al. in [31] by using global version of the Adiabatic Theorem for
their quantum algorithm. This resulted in a running time of the order of N i.e. O(N),
which was no better than a classical algorithm. Roland and Cerf improved the perfor-
mance of Adiabatic Quantum Algorithm for this problem by using local adiabatic
 3 Quantum optimization for machine learning 55

evolution approach [32]. Thus, Adiabatic version of Grover’s search also has the same
complexity as that of Circuit model. This chapter employs local version of Adiabatic
evolution proposed by Roland and Cerf in [32] for Grover’s Search to design a quantum
algorithm for solving optimization problems in continuous domain.
The algorithm starts with initial state as |φ0〉, which is in equal superposition of
all the basis states, Ne (j represents index of elements in unsorted database):

1 Ne X−1
jϕ0 i = pffiffiffiffi jji (3:18)
Ne j = 0

The beginning Hamiltonian, Hini, of the system is chosen as:

Hini = I − jφ0 ihφ0 j, (3:19)

where I is the Identity matrix and |φ0〉 is its ground state with zero energy. Therefore,
the Hini, for quantum search algorithm is independent of the problem. Let us assume
that there exist a problem Hamiltonian Hfin, which can be represented as follows:
X
Hfin = I − jϕM ihϕM j, (3:20)
M2me

where me is the ensemble of solutions and |φM〉 represents the ground state, and is
a solution. Hfin can be applied without explicitly knowing |φM〉, by using an oracle
to tell whether the solution has been found or not. It is implemented as follows:

Hfin jφi i = fpbðφi Þjφi i (3:21)

where function fpb(φi) has been implemented as follows:

Ifðjφi i 2 fjφM igÞ

fpbðφi Þ = 0

else

fpbðφi Þ = 1

That is, if |φi〉 is the marked state, then fpb(φi) is equal to zero, corresponding to
the ground state, i.e. lowest energy and if |φi〉 is not the marked state, then fpb(φi)
is equal to one, i.e. having higher energy level than the ground state. The fpb(φi)
can be reversibly implemented by using the same technique as described for oracles
in Section 3.3.1.
The time dependent Hamiltonian that evolves the system under initial
Hamiltonian to final Hamiltonian is as follows:

HðtÞ = ð1 − sðtÞÞ Hini + sðtÞHfin (3:22)

56 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

where s(t) is a continuous function with s(0) = 0 and s(T) = 1. Let eigenstates of H(t)
be jψk ðtÞi such that:

fpbðqÞ = 0;
(3:23)
fpbðqÞ = 1;

where Ek(t) are the corresponding eigenvalues and k labels the eigenstates (k = 0
indicates the ground state).
The gap between the lowest two eigenvalues, E0 and E1, is described by the fol-
lowing equation [32]:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi


me
g ð sÞ = 1 − 4 1 − sðtÞð1 − sðtÞÞ (3:24)
Ne
p
gmin = ðme =Ne Þ and occurs at s = 1=2.

The matrix element of dH/dt in eq. (3.5) can be alternately expressed as follows:

h dH
dt
i 1, 0
=
ds
dt
h ddsH~ i
1, 0
(3:25)

Using eqs. (3.5) and (3.25), for all times t, we get new condition
 
ds
  ≤ ε  g ð sÞ 
2
 dt   dH~  (3:26)
h ds i1, 0 

~
Using eq. (3.26) and bounding jhddsH i1, 0 j ≤ 1, the Hamiltonian H(t) is evolved at a fol-
lowing rate:

 
ds me
= εg2 ðsÞ = ε 1 − 4 1 − s ð 1 − sÞ (3:27)
dt Ne

After integration, we get:

( )" ( sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi! ) sffiffiffiffiffiffiffiffiffiffi!#
1 Ne −1 Ne − me −1 Ne −
t= pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi tan ð2s − 1Þ + tan (3:28)
2ε me ðNe − me me me

And from the above eq. (3.28), s(t) can be obtained as follows:
(
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi " (pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi) sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi!# )
1 me me ðNe − me Þ Ne − me
sðtÞ = tan 2ε t − tan − 1 +1 (3:29)
2 Ne − me Ne me
 3 Quantum optimization for machine learning 57

Therefore, the total computation time T can be calculated by equating s = 1 in

eq. (3.28), we get:
( ) sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi!
1 Ne Ne − me
T= pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi tan − 1 (3:30)
ε me ðNe − me Þ me
( sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi!)
−1 Ne − me π
max tan ≤ ∀ N e > me (3:31)
me 2
( ) ( sffiffiffiffiffiffi!)
Ne pffiffiffi Ne Ne
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ≤ 2 ∀ me ≤ (3:32)
me ðNe − me Þ me 2

Using above three equations i.e. from (3.30) to (3.32), we can conclude that
pffiffiffi sffiffiffiffiffiffi!
π 2 Ne Ne
T≤ ∀ me ≤ (3:33)
2ε me 2

The above result is very useful from the point of quantum search based on AQC for
solving optimization problems, which have been attempted in [52]. Further, the re-
sult in [32] was derived for cases in which me ≪ Ne. Thus, the bounds provided here
are applicable to more generic cases.
One observation, which shows that AQC is a better model than Quantum Circuit
model for implementing canonical Grover’s search is that, it does not overshoot the
marked element if the system is evolved for time longer than required i.e. if the rate
of evolution given by eq. (3.27) is slower than necessary, there is no overshooting
involved as the system remains in the eigenstate of the final Hamiltonian. That is if
the system is evolved at gmin = 1/√Ne instead of gmin = √(me/Ne), where me > 1, then
it just takes more time to evolve than necessary, but the solution is still reached,
unlike Grover’s search on circuit model in which if more rotations than required are
given to the state vector, then it overshoots the target state. However, in such cases,
the overall computational complexity is O(√Ne) instead of O(√(Ne/me)), which when me
is large would be a substantial slow down.

3.7 AQC based minimum searching algorithm

The algorithm presented by Durr and Hoyer [50] on quantum circuit model can be
modified to fit into AQC framework as follows:
a. Set me = Ne/2 and termination counter I = 0.
b. Select the initial threshold, Yth, by randomly taking a sample.
c. Initialize Hini as given in eq. (3.19).
58 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

d. Let Hfin be as given by eq. (3.20).

where function fpb (φi) has been implemented as follows:

If ðD½φi  < YthÞ == D is the table

fpbðφi Þ = 0

else

fpbðφi Þ = 1

e. Employ Hamiltonian Ĥ(s) = (1-s(t))*Hini + s(t)*Hfin, where s(t) is as given eq.

(3.29).
f. Perform evolution of the system for time T as per eq. (3.1). The time T is given
by eq. (3.33).
g. Measure the qubits in |Ψ(T)〉 to find the solution of the problem and present re-
sult by classical evaluation as objective function value, y.
h. If D[y] is less than threshold Yth, then

Yth = D½y
i. If I > Ttot, then Stop
j. I=I+T
k. me = me−1
l. Go to c.

Theorem 1: The above algorithm finds the minimum in a database in O(√Ne) time
with error probability ε2 in expectation.

Proof: Let us rank all the elements in the database with rank 1 assigned to the ele-
ment with minimum value and rank Ne assigned to the element with maximum
value. All other elements are also ranked according to their position, if there are
several elements with same value, then rank them sequentially as they appear with-
out breaking order with elements having different values.
The above algorithm would move through a succession of points before finding the
minima, therefore, assume that it starts with the element as threshold, whose rank is
Ne/2, (which can be found by either randomly choosing the threshold by sampling the
database a few times) then the total running time, Ttot, of the algorithm is as follows;
X
N=2
Ttot = fPðNe , rn ÞT ðNe , rn − 1Þg (3:34)
r=1

where P(Ne,rn) is the probability of the element with rank rn ever being selected
and T(Ne,rn−1) is the expected time required to evolve the system for finding an
element with a better rank than rn.
 3 Quantum optimization for machine learning 59

The P(Ne,rn) has been shown to be independent of the database size in lemma 1
of [50] and is dependent only on its rank rn.
PðNe , rn Þ = 1=rn (3:35)

The T(Ne, rn−1) can be found by inserting the value of rn in eq. (3.33):
pffiffiffi sffiffiffiffiffiffiffiffiffiffiffi!
π 2 Ne
T ðNe , rn − 1Þ = (3:36)
2ε rn − 1

Therefore,
Ne " ( pffiffiffi sffiffiffiffiffiffiffiffiffiffiffi!)#
pffiffiffi Ne 


X 2
1 π 2 Ne π 2 pffiffiffiffi X 2
1 1
Ttot = = Ne pffiffi (3:37)
r
r =2 n
2ε rn − 1 2ε rn = 1
rn + 1 rn
n

pffiffiffi
5π 2 pffiffiffiffi
Ttot = Ne (3:38)
4ε

Thus, the AQC based minimum finding algorithm can be analyzed analytically and
has a better running time in expectation.

3.8 Adiabatic quantum optimization algorithm

The Adiabatic quantum optimization algorithm has been developed for solving un-
constrained optimization problems, which have been described in Section 3.5. The
AQC based minimum finding algorithm has been used as the underlying framework
for developing the optimization algorithm.
The unconstrained optimization problems in continuous domain can be con-
verted into a finite global optimization problem as shown in Section 3.5. Each do-
main point can be ranked on the basis of its objective function value i.e. the point
with best objective function value can be assigned RANK 1 and the point with worst
objective function value can be assigned RANK Ne. The rest of points can similarly be
assigned a suitable rank. If two or more points have same objective function value
then they would be assigned different ranks, randomly chosen within the range, while
maintaining their relative position with rest of the points. Thus, the points are now
analogues to the index of the database, so, the optimization algorithm can be designed
on the same framework as the minimum finding algorithm described in Section 3.5.
Adiabatic quantum optimization algorithm for solving Unconstrained Global
Optimization problem is as follows:
a. Set me = Ne/2 and termination counter I = 0.
b. Compute Yth by taking a random sample of domain point x.
60 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

c. Construct Beginning Hamiltonian Hini, which would be same for all the UCOP
problems.
X
Hini = Hini i , where

Hini i = I − jψi ihψi j, where

p p
jψi i = ð1= 2Þ*ð0i + ð1= 2ÞÞ

d. Make Problem Hamiltonian Hfin, which is specific to the particular UCOP, how-
ever the structure would remain same irrespective of the UCOP.
Hfin|q1q2q3. . .qn〉 = fpb (q1q2q3. . .qn) |q1q2q3. . .qn〉

where fpb(q1q2q3. . .qn) is a Boolean function, which models the problem and
can be constructed reversibly as follows:
If (f(q) ≥ Yth) . . . . . . for Minimization

fpbðqÞ = 0;

else

fpbðqÞ = 1;

where |q1q2q3. . .qn〉 are n qubits representing the decision variables of the UCOP
being solved given eq. (3.13) and q = q1q2q3. . .qn. The fpb(q) would be imple-
mented in an Oracle and so its outcome would be reflected in the state of ancilla
qubit, qo. Yth is the threshold value computed from the objective function of the
UCOP being solved. The first value of Yth may be assigned randomly or seeded
by meta-heuristics techniques [57].
e. Employ Hamiltonian Ĥ(s) = (1-s(t))*Hini + s(t)*Hfin, where s(t) is as given eq. (3.29).
f. Perform evolution of the system for time T as per eq. (3.1). The time T is given
by eq. (3.33).
g. Measure the qubits in |Ψ(T)〉 to find the solution of the problem and present re-
sult by classical evaluation as objective function value, y.
h. If y is less than threshold Yth, then

Yth = y – Δy where Δy ≥ 0 for minimization problem

i. If I > Ttot, then Stop, where Ttot is given by eq. (3.38)

j. I=I+T
k. me = me−1
l. Go to c.
 3 Quantum optimization for machine learning 61

3.8.1 Discussions

The nucleus of the proposed algorithm is a search routine designed on AQC frame-
work to find optimal value. It adapts the evolution rate to local adiabatic condition
based on local adiabatic evolution, which has O(√Ne) run-time complexity and is
asymptotically optimal as shown in [32].
The termination of the run of algorithm for solving a particular instance of
the problem is also supported by the laws of Quantum mechanics (i.e. Adiabatic
Theorem) that is if no improvement is made even after considering me = 1, then
the search can be terminated with conclusion that no better points exists with
probability 1-ε2, where ε ≪ 1. This consideration makes the proposed algorithm a
global optimizer.
The working of the algorithm is further illustrated with the help of an example.
Let us Minimize f(x) = x2 such that −1.0 < x < 1.0
The variable x is discretized between [−1.0, 1.0] by using 10 qubits i.e. n = 10.
The total number of domain points in search space, N = 2n = 210 = 1024.
For first iteration: Let ε = 0.707, x = 0.5; Yth = (0.5)2 = 0.25, me = 512, there-
fore T = 2.22 time units. If the run is successful, a new threshold would be found
or else with the earlier threshold, and after reducing me by 1, the second itera-
tion would be executed, so on and so forth. Subsequently, say, after 256 itera-
tions, lf |x| > 0.25, then it would have a high chance of success run as T for which
it would be evolved is larger than required, whereas if |x| < 0.25, then T is less
than required but in next few iterations, the T would become equal or larger
(as me is reduced) than required for a success run. Moreover, when |x| < 0.25, the
performance of the algorithm is better than expected. Therefore, if the algorithm
fails, it is because its performance is better than expected otherwise it is almost
guaranteed to produce a good result. Thus, after running the algorithm for Ttot
time, it can be re-run for me = 1 and if no improvement is found in threshold
then it can be said that global optimal has been found with failure probability at
most ε2.
The actual running time of the algorithm can be improved further by doing
some additional processing like counting the successful and unsuccessful iterations
and for updating the value of me as the eq. (3.38) gives an average case result.
However, the analytical analysis of such adaptive schemes would be difficult and is
left for future studies.
AQC optimization algorithm is better than the Quantum circuit implementa-
tion of Grover Adaptive Search as the rotation count is no longer required to be
guessed or randomly / heuristically assigned. Further, even if the quantum sys-
tem were evolved for time duration larger than the minimum time, the system
would only get closer to the ground state and would not overshoot the marked
state containing the solution.
62 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

3.9 Conclusions
Machine learning is a branch of Artificial intelligence, which tries to develop techni-
ques that have ability to learn from experiences. These experiences are articulated as
data generated from underlying models. The goal of any machine learning technique is
to abstract the underlying model reliably for specific tasks like prediction and decision
making. Optimization problems are ubiquitous in machine learning and other domains
like engineering and a number of specialized techniques for efficiently solving specific
problems have been developed. However, general purpose techniques like enumerative
strategies have run time complexity of O(N), where N is the number of domain points.
Grover’s quantum search algorithm provides quadratic speed up over its classical
counter parts. Baritompa et al. integrated ideas from Grover’s search into the frame-
work of Pure adaptive search and called the new framework as Grover Adaptive search.
It was further improved by Liu and Koehler. However, GAS was essentially an optimi-
zation heuristic due to which analytical analysis was difficult.
A novel quantum optimization algorithm based on Adiabatic Quantum
Computation framework, which employs local evolution instead of global evolution,
has been proposed. It is also inspired by the application of Grover’s Search algorithm
for finding minimum solution in a database, however, the proposed algorithm solves
real domain optimization problems, which are routinely solved in Machine learning
algorithms. Further, the proposed algorithm has been analytically analyzed and has
been shown to have the same signature of quadratic speed up of Grover’s search al-
gorithm. Baritompa et al. had made a similar effort on gate-based quantum computa-
tion model for solving unconstrained global optimization problem. However, the
proposed AQC based algorithm is conceptually better than gate-based model as there
is no fear of overshooting the optima, which is a major concern in gate-based models,
when a priori knowledge of the number of optimal solutions is not available,
whereas the optimal solutions are encoded as the ground state of the problem
Hamiltonian. The problem of overshooting in gate-based model could further in-
crease convergence time in real world implementations.
In future, we would like to develop a simulator and architecture of a hybrid
Classical and Adiabatic Quantum Computer on which we can demonstrate operations
of our proposal of Quantum Optimization Algorithm for Machine Learning techniques,
which would require addressing several challenging issues like implementation of loss
function on a quantum computer and weights updating in learning mechanism.

References
[1] Kline, R. R. Cybernetics, Automata Studies and the Dartmouth Conference on Artificial
Intelligence. IEEE Annals of the History of Computing, October–December, 2011, IEEE
Computer Society.
 3 Quantum optimization for machine learning 63

[2] Russell, S., & Norvig, P. Artificial Intelligence: A Modern Approach, 3rd ed. Prentice Hall Press
Upper Saddle River, NJ, USA, 2009.
[3] Columbus, L. Why AI Is The Future of Financial Services. August, 2019. https://www.forbes.
com/sites/louiscolumbus/2019/08/15/why-ai-is-the-future-of-financial-services
/#452381ba3847
[4] Insight Team. AI And Healthcare: A Giant Opportunity. Februrary, 2019. https://www.forbes.
com/sites/insights-intelai/2019/02/11/ai-and-healthcare-a-giant-opportunity/#35e679924c68
[5] Grühn, B., Kässer, M., Köstring, J., Padhi, A., & Tschiesner, A. Winning tomorrow’s car buyers
using artificial intelligence in marketing and sales. February, 2019. https://www.mckinsey.
com/industries/automotive-and-assembly/our-insights/winning-tomorrows-car-buyers-
using-artificial-intelligence-in-marketing-and-sales
[6] Teich, D. A. Artificial Intelligence Can Address Social Media Risks And Rewards. March, 2019.
https://www.forbes.com/sites/davidteich/2019/03/27/artificial-intelligence-can-address-
social-media-risks-and-rewards/#295c4dc9686c
[7] Hoy, M. B. Alexa, Siri, Cortana, and More: An Introduction to Voice Assistants. Medical
Reference Services Quarterly. 2018, 37(1),81–88. doi:10.1080/02763869.2018.1404391.
PMID 29327988
[8] Zhang, G., Patuwo, B. E., & Hu, M. Y. Forecasting with artificial neural networks: The state
of the art. International Journal of Forecasting, 1998, 14(1),35–62, ISSN 0169-2070.
https://doi.org/10.1016/S0169-2070(97)00044-7
[9] Chang, C., & Ling, C. LIBSVM: A library for support vector machines. ACM Transactions on
Intelligent Systems and Technology. 2011, 2(3). 10.1145/1961189.1961199
[10] Jain, A. K. Data clustering: 50 years beyond K-means. Pattern Recognition Letters. 2010, 31
(8),651–666, ISSN 0167-8655. https://doi.org/10.1016/j.patrec.2009.09.011
[11] Liu, H., Miao, E. M., Wei, X. Y., & Zhuang, X. D. Robust modeling method for thermal
error of CNC machine tools based on ridge regression algorithm. International
Journal of Machine Tools and Manufacture. 2017, 113, 35–48, ISSN 0890-6955.
https://doi.org/10.1016/j.ijmachtools.2016.11.001
[12] Garg, P., Neider, D., Madhusudan, P., & Roth, D. Learning invariants using decision trees and
implication counterexamples. Proceedings of the 43rd Annual ACM SIGPLAN-SIGACT
Symposium on Principles of Programming Languages (POPL ‘16). ACM, New York, NY, USA,
2016, 499–512. https://doi.org/10.1145/2837614.2837664
[13] Tangirala, A. K. Principles of System Identification: Theory and Practice 1st Edition. CRC
Press, 2014.
[14] LeCun, Y., Bengio, Y., & and Hinton, G. Deep learning. Nature. 2015, 521, 436–444.
[15] Poplin, R., Varadarajan, A. V., Blumer, K., Liu, Y., McConnell, M. V., Corrado, G. S., Peng, L.,
& Webster, D. R. Prediction of cardiovascular risk factors from retinal fundus photographs via
deep learning. Nature Biomedical Engineering. 2018, 2, 158–164.
[16] Rebentrost, P., Mohseni, M., & Lloyd, S. Quantum support vector machine for big data
classification. Phys. Rev. Lett. 2014, 113, 130503.
[17] Giovannetti, V., Lloyd, S., & Maccone, L. Quantum Random Access Memory. Phys. Rev. Lett.
2008, 100, 160501.
[18] Optimization for Machine Learning. Sra, S., Nowozin, S., & Wright, S. J. (Eds). The MIT Press,
2011.
[19] Rebentrost, P., Schuld, M., Wossnig, L., Petruccione, F., & Lloyd, S. Quantum gradient
descent and Newton’s method for constrained polynomial optimization. New J. Phys. 2019,
21, 073023.
64 Ashish Mani, Siddhartha Bhattacharyya, and Amlan Chatterjee

[20] Klymko, C., Sullivan, B. D., & Humble, T. S. Adiabatic quantum programming:
minor embedding with hard faults. Quantum Inf Process. 2014, 13(709).
https://doi.org/10.1007/s11128-013-0683-9
[21] Harrow, A. W., Hassidim, A., & Lloyd, S. Quantum algorithm for linear systems of equations.
Phys. Rev. Lett. 2009, 103, 150502.
[22] Farhi, E., Goldstone, J., Gutmann, S., Lapan, J., Lundgren, A., & Preda, D. A quantum adiabatic
evolution algorithm applied to random instances of an NP-complete problem. Science. 2001,
292(5516),472–476.
[23] Farhi, E., Goldstone, J., Gutmann, S., & Nagaj, D. How to Make the Quantum Adiabatic
Algorithm Fail. International Journal of Quantum Information. 2008, 6(503).
[24] Reichardt, B. The quantum adiabatic optimization algorithm and local minima. Proceedings
of the 36th Annual Symposium on the Theory of Computing (IEEE). Computer Society Press,
New York, 2004, 279–287.
[25] van Dam, W., Mosca, M., & Vazirani, U. How powerful is adiabatic quantum computation?
Proceedings of 42nd Annual IEEE Symposium on Foundations of Computer Science. 2001,
279–287.
[26] Amin, M. H. S. Effect of Local Minima on Adiabatic Quantum Optimization. Physical Review
Letters. 2008, 100, 130503.
[27] Farhi, E., Goldstone, J., Gutmann, S., Meyer, H. B., & Shor, P. Quantum Adiabatic Algorithms,
Small Gaps, and Different Paths. 2009. arXiv:0909.4766v2 [quant-ph]
[28] Altshuler, B., Krovi, H., & Roland, J. Anderson localization makes adiabatic quantum optimization
fail. Proceedings of National Academy of Science. USA, 2010, 107(28),12446–12450.
[29] Choi, V. Different Adiabatic Quantum Optimization Algorithms for the NP-Complete Exact
Cover and 3SAT Problems. 2011. arXiv:1010.1221v3
[30] Dickson, N. G., Amin, M. H. S. Does Adiabatic Quantum Optimization Truly Fail for NP-
complete problems? 2011. arXiv:1010.0669
[31] Farhi, E., Goldstone, J., Gutmann, S., & Sipser, M., Quantum Computation by Adiabatic
Evolution. 2000. arΧiv:quant-ph/0001106v1
[32] Roland, J. & Cerf, N. Quantum search by local adiabatic evolution, Physical Review A. 2002,
65(4), 042308.
[33] Rezkhani, A. T., Pimachev, A. K., & Lidar, D. A. Accuracy versus run time in an adiabatic
quantum search. Physical Review A. 2010, 82, 052305.
[34] Zhang, Y.Y & Lu, S.F. Quantum search by partial adiabatic evolution. Phys. Rev. A. 2010, 82,
034304.
[35] Aharonov, D., Dam, W. van Kempe, J., Landau, Z., Lloyd, S., & Regev O. Adiabatic quantum
computation is equivalent to standard quantum computation. Proceedings of 45th Annual
IEEE Symposium on Foundations of Computer Science. 2004, 42–51.
[36] Mizel, A., Lidar, D. A., & Mitchell, M. W. Simple proof of equivalence between adiabatic
quantum computation and the circuit model. Phys. Rev. Lett. 2007, 99, 070502.
[37] Brady, L. T. & van Dam, W. Necessary adiabatic run times in quantum optimization. Phys.
Rev. 2017, A 95, 032335.
[38] Guerreschi, G. G. & Matsuura, A. Y. QAOA for Max-Cut requires hundreds of qubits for
quantum speed-up. Scientific Reports. 2019, 9, 6903.
[39] Farhi, E., Goldstone, J., & Gutmann, S. A Quantum Approximate Optimization Algorithm. 2014.
http://arxiv.org/abs/1411.4028v1
[40] McClean, J. R., Romero, J., Babbush, R., & and Aspuru-Guzik, A. The theory of variational
hybrid quantum-classical algorithms. New J. Phys. 2016, 18, 023023.
[41] Mani, A. & and Patvardhan, C. Some Investigations for Solving Constrained Optimization
Problems using Adiabatic Quantum Computation. 2011 Eighth International Conference on
 3 Quantum optimization for machine learning 65

Information Technology: New Generations. Las Vegas, NV, 2011, 1101–1102. doi:10.1109/
ITNG.2011.218
[42] Albash, T. & Lidar, D. A. Adiabatic quantum computation. Rev. Mod. Phys. 2018, 90, 015002.
[43] Kato, T. On the Adiabatic Theorem of Quantum Mechanics. Journal of the Physical Society of
Japan. 1950, 5(6),435–439.
[44] Marzlin, K. P. & Sanders, B. C. Inconsistency in the application of the adiabatic theorem.
Phys. Rev. Lett. 2004, 93, 160408.
[45] Amin, M. H. S. Consistency of the Adiabatic Theorem. Physical Review Letters. 2009, 102,
220401.
[46] Tong, D. M., Singh, K., Kwek, L. C., & Oh, C. H. Sufficiency Criterion for the Validity of the
Adiabatic Approximation. Phys. Rev. Lett. 2007, 98, 150402.
[47] Grover, L. K. Quantum Mechanics helps in searching for a needle in a haystack. Phys. Rev.
Lett. 1997, 78(2),325–328.
[48] Boyer, M., Brassard, G., Hoyer, P., & Tapp, A. Tight bounds on quantum searching. Fortschr.
Phys. 1998, 46, 493–506.
[49] Tulsi, T. A., Grover, L. K., & Patel, A. A new algorithm for fixed point quantum search.
Quantum information & computation. 2006, 6, 483–494. arXiv: quantph/0505007
[50] Dürr, C. & Hoyer, P. A Quantum Algorithm for Finding the Minimum. http://lanl.arxiv.org/
abd/quant-ph/9607014
[51] Kenigsberg, D. Grover’s Quantum Search Algorithm and Mixed States. M.Sc. Thesis. Israel
Institute of Technology, October, 2001.
[52] Baritompa, W. P., Bulger, D. W., & Wood, G. R. Grover’s Quantum Algorithm applied to Global
Optimization. SIAM J. OPTIM. 2005, 15(4),1170–1184.
[53] Zabinsky, Z. B. & Smith, R. L. Pure adaptive search in global optimization. Math.
Programming. 1992, 53, 323–338.
[54] Zabinsky, Z. B, Wood, G. R., Steel, M. A., & Baritompa, W. P. Pure adaptive search for finite
global optimization. Math. Programming. 1995, 69, 443–448.
[55] Bulger, D. W., Alexander, D. L. J., Baritompa, W. P., Wood, G. R., & Zabinsky, Z. B. Expected
hitting time for backtracking adaptive search. Optimization. 2004, 53, 189–202.
[56] Liu, Y., Koehler, G. J. Using modifications to Grover’s Search algorithm for quantum global
optimization. European Journal of Operational Research. 2010, 207(2),620–632.
[57] Michalewicz, Z. Genetic Algorithms + Data Structures = Evolution Programs. 3rd ed., Berlin,
Springer-Verlag, 1996.
[58] Goldberg, D. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-
Wesley Publishing Company, Inc., NY, USA, 1989.
[59] Bellman, R. Adaptive Control Processes: A Guided Tour, Princeton University Press, 1961.
[60] Protopopescu, V., Barhen, J. Quantum Algorithms for Continuous Global Optimization.
Proceedings of the 34th Workshop on Optimization and Control with Applications, Erice,
Sicily, July 9–17, 2001.
Arit Kumar Bishwas, Ashish Mani, and Vasile Palade
4 From classical to quantum machine
learning
Abstract: In recent years, Machine Learning (ML) has started to be ubiquitously ap-
plied to practically most of the human activity domains. Although traditional, or clas-
sical machine learning (CML) approaches are useful in solving many complex tasks,
there are still many challenges that such approaches are facing. One issue is the limi-
tation in the processing speed of current silicon technology based computers, which
made researchers look to the underlying quantum theory principles and try to run
complex quantum computing experiments. In future, in order to develop more com-
plex artificial intelligence systems, we will have to process huge amounts of data at
high speed, and the existing classical computing will probably not serve well this
purpose, due to silicon technology limitations. A different technology that can handle
huge volumes of data and at high speed is needed. In recent years, the progress in
quantum computing research seems to provide hopeful answers to overcome the
speed processing barrier. This is very important for the training of many computa-
tional intensive machine learning models. The latest advancements in quantum
technology appear to be promising, which can boost the field of machine learning
overall. In this chapter, we discuss the transition from classical machine learning
to quantum machine learning (QML) and explore the recent progress in this do-
main. QML is not only associated with the development of high-performance ma-
chine learning algorithms that can run on a quantum computer with significant
performance improvements but also has a very diverse meaning in other aspects.
The chapter tried to touch those aspects in brief too, but the main focus is on the
advancements in the field of developing machine learning algorithms that will
run on a quantum computer.

Keywords: quantum algorithm, quantum random access memory (QRAM), quantum

deep learning, quantum computation

4.1 Introduction
Machine learning is an important and popular field of study these days when
Artificial Intelligence and intelligent algorithms have become embedded in our ev-
eryday life. Machine learning models are trained with large amounts of data using

Arit Kumar Bishwas, AIIT, Amity University, Noida, Uttar Pradesh, India
Ashish Mani, EEE, ASET, Amity University, Noida, Uttar Pradesh, India
Vasile Palade, Faculty of Engineering and Computing, Coventry University, Coventry, UK

https://doi.org/10.1515/9783110670707-004
68 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

some mathematical and statistical based algorithms. Machine learning algorithms

are categorized into two broad classes – supervised and unsupervised learning al-
gorithms. In supervised learning, data are provided in the form of labelled data. In
the case of unsupervised learning, the datasets are not labelled. In recent years,
machine learning has demonstrated great visibility and publicity in applications to
many important domains of real life, such as medical image processing, online
product recommendation systems, financial modelling and prediction, risk evalua-
tion, human language processing, etc. The recent advancements in deep learning
research opened even more possibilities for solving complex problems. In machine
learning, data play a vital role. Some problems can be solved well with less data,
but deep learning requires huge volumes of data. Apart from the issue of availabil-
ity of big datasets, another crucial factor is the processing and learning speed of the
system. Deep learning algorithms are computationally very intensive. Although by
using GPUs and high-performance computing systems, this problem can be some-
how alleviated, it still takes a long time to perform machine learning (and espe-
cially deep learning) on big data. The rule of thumb is that, in order to train with
more data, we would require more processing power. To develop more complex fu-
ture artificial intelligence systems that process a huge amount of data, the existing
classical computing technologies will probably not be sufficient, due to the limita-
tions related to the silicon-based technology. We need a different technology,
which can handle the processing speed issue and which can process huge volumes
of data. Classical computers are now part of our lives and help us in solving com-
plex problems, but still, there are great challenges in solving some of the problems
(for example optimization problems) whose computational complexities increases
at a very high scale with a small increase in inputs. These kind of problems are very
difficult or not possible to solve with the classical computers. Quantum computing is
the answer to overcome these processing issues. Quantum computer’s computational
power scales significantly as the system size grows due to its unique properties
which are not found in classical systems, and two of the most important properties
are – superposition & entanglement. Section 3 discusses these properties in detail.
The term quantum machine learning (QML) mostly refers to classical machine
learning algorithms [1, 2] that can be run on a quantum computer. This involves
mostly researchers working on developing quantum versions of classical machine
learning algorithms, concerned mainly with the speed-up factors for the training pro-
cedure. Apart from this, the term QML also has a general connotation. Developing a
fully functional commercial quantum computer is a very active field, but it is still not
clear if we can have one to work in the near future. In recent time, researchers fo-
cused on developing small quantum subroutines/devices which can combine with
classical machine learning approaches to get the quantum advantages to some ex-
tent [3]. These quantum subroutines/devices can be designed to solve some of the
complex tasks which are computationally difficult in the classical domain [4–6].
 4 From classical to quantum machine learning 69

Some recent works have also exhibited that by using quantum algorithms we
can analyze quantum states [7]. The term QML is also associated with analyzing the
data generated from quantum experiments. Applying machine learning in analyz-
ing quantum phase transitions is a very classic example of this [8–11].
Another active research field in QML is to understand the structural and opera-
tional similarities between certain physical systems and deep learning. Deep leaning
uses some mathematical and numerical techniques that are inspired by quantum
formulation, and in the same way, deep learning explores some structural and pro-
cedural equivalences with quantum systems [12–14].
Big companies like D-Waves [15], IBM [16], and Google [17] demonstrated
some progress in building a quantum computer. Here, the D-Wave’s model is
based on the adiabatic quantum process [18] and the IBM [16] is working on
mostly non-adiabatic models. IBM released the IBM-Q quantum computer to be
accessed and used freely (with the restriction of 5 qubits system access at the
time). Apart from the progress in quantum hardware, there is very active research
going on in the area of developing quantum algorithms. Some of the most notable
quantum algorithms are Grover’s search [10], Shor’s algorithm [19], HLL [20],
etc., which demonstrated significant speed-up gains insolving computational
problems. In recent years, the advancement in quantum technologies has in-
volved some state-of-the-art machine learning approaches [21–25]. These quan-
tum machine learning algorithms are exponentially faster than their classical
counterparts.
In this chapter, our focus will be on a category of classical machine learning
algorithms that can run on a quantum computer. At first, we discuss quantum com-
puting basics in brief, including some well-known techniques that are useful in the
advancement of quantum machine learning algorithms, followed by the recent ad-
vancements in the field of quantum machine learning algorithms.

4.2 Overview of quantum computing

4.2.1 Hilbert space & Dirac notation

A Hilbert space C n is an abstract vector space which is very useful for representing
a quantum system. The Dirac notation [26] is a very convenient way of defining
70 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

vectors in the Hilbert space, named after the Nobel laureate Paul Dirac. In Dirac no-
tation, a vector V is represented as:
2 3
v0
6 7
6 v1 7
6 7
V = 6 . 7 = jVi (4:1)
6 .. 7
4 5
vn

Here, jVi is known as “ket V”. In the same context, the dual vector of jVi is known
as “bra V” and represented as:

hVj = ½v0 v1 . . . vn  = V T (4:2)

We define the inner product of vectors U and V in Dirac notation as:

2 3
v0
6 7
6 v1 7
6 7
hUjVi = ½u0 u1 . . . un 6 . 7 = U T V (4:3)
6 .. 7
4 5
vn

Also, the length of the vector in the Hilbert space, known as “Norm”, is defined as:
pffiffiffiffiffiffiffiffiffiffiffiffi
kjVik = hVjVi (4:4)

Similarly, the outer product can be defined as follows with the Dirac notation:
2 3
v0 2 3
6 7 v0 u0    v0 um
6 v1 7 6 .
6 7 .. .. 7
jVihU j = 6 . 7½u0 u1 . . . um  = 6
4 .. .
7
. 5 (4:5)
6 .. 7
4 5
vn u0    vn um
vn

4.2.2 Quantum bits and gates

In classical computation, the fundamental unit for the building block of any com-
putation is the bit. In the same context, the fundamental component of any quan-
tum computation is known as a qubit. In classical computation, a base-2 number
known as a bit can be either 0 or 1, but a qubit can also be in both states, 0 and 1,
at the same time. Mathematically we represent a qubit as a two-dimensional state
space in C 2 with orthogonal basis vectors j0i & j1i as,
 4 From classical to quantum machine learning 71

jψi = a0 j0i + a1 j1i;ja0 j2 + ja1 j2 = 1 (4:6)

The states j0i & j1i are in superposition (this will be covered in detail in the next
section) and the complex scalar amplitudes a0 & a1 are the probabilities of obtain-
ing the values j0i & j1i upon measurement, respectively.
A single qubit is of no significant use, but a bunch of qubits can involve much
more practical usability. Like a classical computer, the quantum register is the col-
lection of qubits. In the quantum register, a qubit in a register can be in a superpo-
sition of j0i & j1i, and upon measurement collapse to a bit. So, if there is a n qubits
register, then there will be 2n possible states in superposition. For example, suppose
we have a 3 qubits register, then there will be 23 = 8 states in superposition as shown:
2X
3 −1

jψ3 i = ai jii = a0 j000i + a1 j001i + a2 j010i + a3 j011i + a4 j100i + a5 j101i

i=0 (4:7)
+ a6 j110i + a7 j111i

Like classical computing, we have quantum gates too, which are used to perform
some kind of operations on qubits and registers. The followings are some of the
most important quantum gates:
a. Hadamard gate ðH Þ – This is a very important quantum gate. The Hadamard gate
transforms a qubit into a superposition of qubits as shown:
j0i + j1i
Hj0i = pffiffiffi (4:8)
2

j0i − j1i
Hj1i = pffiffiffi (4:9)
2

b. Pauli-X quantum gate ð XÞ – Pauli–X is quantum equivalent to classical NOT

gate. It swaps the amplitudes of the j0i & j1i.
c. Pauli-Y quantum gate ðY Þ – It swaps the amplitudes of j0i & j1i, multiplies each
amplitude by i, and then negates the amplitude of j1i.
d. Pauli-Z quantum gate ðZ Þ – When applying a Pauli-Z quantum gate to a qubit, it
negates the amplitude of the qubit if it is j1i, otherwise, it performs no action if
the qubit is j0i.

4.2.3 Superposition

The Quantum superposition [27] is a phenomenon in the quantum world in which

quantum particles appear to exist in multiple states simultaneously. It is the fun-
damental principle of quantum mechanics. When the particle is measured in
72 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

superposition, the superposition collapses and results in correspond to only one of

the possible states.
Thus, at any instance, a qubit can be in a superposition of both states j0i and j1i
simultaneously, such as:
!
a
jWi = aj0i + bj1i = (4:10)
b

where the coefficients a and b are complex numbers and the amplitudes of the com-
ponents j0i and j1i respectively, possessing the following property:

jaj2 + jbj2 = 1 (4:11)

4.2.4 Entanglement

Quantum entanglement [28] is a very important property of quantum mechanics, in

which qubits interact with each other instantaneously irrespective of the distances
between them through an unknown way of communication that is not limited to
the speed of light. Correlated particles remain entangled as long as they are isolated
irrespective of the distance between them.
If we cannot factor a multi-qubit state into the direct product of a definite state
for each qubit individually, then the states are entangled. Hence, a pair of qubits
A and B are entangled if and only if we cannot write the joint state jψAB i as the prod-
uct of a state for qubit A and a state for qubit B, i.e., if and only if jψAB i = jψA i # jψB i
for any choice of states jψA i and jψB i.
We are considering two non-interacting systems A and B, with respect to finite
dimensional Hilbert spaces HA and HB .
HA # HB is the tensor product of the composite system in the Hilbert space. The
state of the composite system is jψA i # jψB i, where the first system is in state jψA i and
the second in the state jψB i.
The above represented states of the composite system are separable states. But,
all states are not, in general, the separable states.
By fixing a basis jiA i for HA and a basis j jB i for HB , the most general state
in HA # HB is of the form:
X
jψAB i = cij jiA i # jjB i (4:12)
i, j

This state is separable if

cij = cAi cBj (4:13)

 4 From classical to quantum machine learning 73

yielding:
   
 X   X 
ψ i = A
i  ψ i = B
 A ci Ai and  B ci jB i (4:14)
i, j j

It is inseparable if

cij ≠ cAi cBj (4:15)

If a state is not separable, it is known as entangled state.

For example, the following composite state is an entangled state:
1
pffiffiffi ðj0A i # j1B i − j1A i # j0B iÞ, (4:16)
2

where j0A i & j1A i are two basis vectors of HA and j0B & j1B are two basis vectors
of HB .

4.2.5 Quantum adiabatic model

If a quantum system is set in the ground state of a Hamiltonian H, the system re-
mains in this state. According to the adiabatic theorem [18], the system will still stay
close to the ground state as long as the variation is very slow enough. Suppose that
jE0 ; ti & jE1 ; ti are the ground and first excited states of the Hamiltonian H ðtÞ, where
E0 ðtÞ and E1 ðtÞ are the energies corresponding to the ground and first excited states,
respectively. We then define the minimum gap between these two eigenvalues as:

gmin = min ½E1 ðtÞ − E0 ðtÞ (4:17)

0≤t≤T

So, when we prepared the system at time t = 0 in its ground state jE0 ; 0 and allow it
to evolve under the Hamiltonian H ðtÞ, we find:

jhE0 ;TjψðT Þij2 ≥ 1 − ε2 (4:18)

  
       
max  dH  dH dH 
0 ≤ t ≤ T dt 1, 0  
where 1  ε ≥ , and = E ;
1 t E ; t
dt 
2 0
gmin dt 1, 0

4.2.6 SWAP-test

A SWAP-test estimates jhψ1 jψ2 ij2 , when given two input quantum states jψ1 i and

jψ2 i, with precision O p1ffiffiffi
M
, where M is the number of times the measurement has
been done.
74 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

4.2.7 Quantum algorithms

Quantum algorithms are the special type of algorithms that are written for running
on a quantum computer. These algorithms are based on the principles of quantum
mechanics. This is one very active research area, and recent research works show
that many of the quantum versions of classical algorithms are way more efficient in
terms of computational complexity, with some of them being exponentially faster
than the classical counterparts. Some of the most known quantum algorithms are
Grover’s search algorithm [10], Shore’s algorithm [19], quantum support vector ma-
chine [21, 23, 24], and the HHL algorithm [20].

4.3 Essential foundations for building a quantum

machine learning framework
In this section, we discuss the essential foundations to develop quantum machine
learning algorithms. These foundational concepts are very important. No single con-
cept can be used to develop a quantum machine algorithm, and different concepts
are used to cultivate different algorithms. In this section, we explain some of the
very key foundational concepts which provide support in building quantum ma-
chine learning algorithms. We have discussed how different approaches used these
concepts to articulate a variety of quantum machine algorithms in Section 4.4 & 4.5.

4.3.1 Grover’s search algorithm

Grover’s search algorithm [10] is a quantum search algorithm which p performs

ffiffiffiffi the
search to find a unique item from an unordered set of N items with Oð N Þ run time
complexity. The best counter classical search algorithm performs with OðN Þ run-
time complexity. The algorithm is as follows [29]:

Algorithm 1: Grover’s Search Algorithm

1. Initialize quantum state j0i#n
2. Apply the Hadamard transformation to all the qubits

1 2X
n−1
jψi = H #n j0i#n = pffiffiffiffiffi jyi
2n y = 0
 4 From classical to quantum machine learning 75

π pffiffiffiffinffi
3. Apply the Grover iteration function R ≈ 2 times
4

½ð2jψihψj − I ÞOR jψi ≈ jy0

4. Perform the measurement for output, y0

4.3.2 Quantum random access memory

Quantum random access memory (QRAM) [30] is the random access memory in a
quantum paradigm. It encompasses the output and the address registers which are
composed of qubits. The QRAM permits to access the data quantum mechanically
and access the memory in coherent quantum superposition. The address register em-
braces a superposition of addresses. The data register DR stores the output of the
QRAM as a superposition of data, which is associated with the address register RADR ,
X X  
ψ jji
j j RADR
! ψ ji D i
j j ADR j DR
(4:19)
P
where RADR holds a superposition of addresses j ψj jjiRADR , and Dj is the jth memory
cell content. Reconstructing any quantum state from QRAM takes Oðlog N Þ steps,
where N is the complex vector dimension.

4.3.3 Quantum dot product

We discuss a quantum-based process for dot product evaluation [31] of two training
inputs (normalized), k~ Xi i & k~
Xj i (in the quantum arrangement), in a linear kernel.
For calculating a dot product of k~ Xi i & k~
Xj i, at first, we generate two quantum
states kψi and kϕi using an ancilla variable. Next, we estimate the squared norms
sum of the two training inputs, say parameter Z = j~ Xi j2 + j~
Xj j2 . At last, we execute a
swap test to implement a projective measurement on the ancilla alone.
Initially, by querying the QRAM, we construct a quantum state jψi:
1 
jψi = pffiffiffi j0ij~
Xi i + j1ij~
Xj i (4:20)
2

Let us assume another quantum state:


1
jξi = pffiffiffi ðj0i − j1iÞ # j0i (4:21)
2
76 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

We put on a unitary transformation [32] e − iHt to the state jξi, where

Xi kj0ih0j + k~
H = ððk~ Xj kj1ih1jÞ # σx Þ is a Hamiltonian. This outcomes the following
state,
 
1
pffiffiffi ðcosðk~Xi ktÞj0i − cosðk~ Xj ktÞj1iÞ # j0i
2
  (4:22)
i ~ ~
− p ffiffi
ffi ðsinðkX i ktÞj0i − sinðkXj kÞj1iÞ # j1i
2

With an appropriate choice of t, we now measure the ancilla bit, where ~

Xi t, ~
Xj t  1,
which outcomes in the state:
1
jϕi = qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðk~
Xi kj0i − k~
Xj kj1iÞ, (4:23)
ðk~ Xi k + k~
2 2
Xj k Þ

with probability 21 ½ðk~

X i k + k~
2 2 2
Xj k Þt .
This sanctions the estimation of the squared norms sum of j~ Xi i & j~ Xj i. By ap-
~ 2
~ 2
plying quantum counting, we can estimate ðkX i k + kXj k Þ and generate the quan-
tum state jϕi with accuracy ϵ. The complexity will be O ðϵ − 1 Þ. Using an ancilla
alone, we then execute a swap test with states jψi & jϕi. If jψi & jϕi are equal then
the measurement gives us a zero. Thus, the overall complexity in evaluating a sin-
gle dot product of the training instances, considering the QRAM access [30], esti-
mating ðk~Xi k + k~
2 2
Xj k Þ, and fabricating the quantum state jϕi is O ðϵ − 1 log NÞ.

4.3.4 The HHL algorithm

Harrow, Hassidim, and Lloyd [20] developed the HHL algorithm which quantum me-
chanically inverts a system of linear equations. This algorithm is one of the most sig-
nificant and fundamental quantum algorithms, and the base for constructing many
QML algorithms. The algorithm pursues to solve the system of equations A~ x =~b in a
quantum computer. Quantum mechanically it is represented as Ajxi = jbi, where A is
P
a Hermitian matrix. ~ b is expressed as a quantum state jbi = n bn jEn i over log2 N qu-
bits where En is an eigenvector of A with eigenvalue λn ≥ Λ, and ~ x is stated as jxi. The
algorithm aims to solve jxi = A − 1 jbi, where A − 1 is the inverse of the A. We then com-
pute λn by applying the phase estimation. By uncomputing the phase-estimation we

rffiffiffiffiffiffiffiffiffiffiffi 
P 2
get n bn jEn i λΛn j1i + 1 − Λ2 j0i , where λΛn is the angle of arcsin through which an
λn

ancillary qubit is rotated. To succeed, we apply Oðjj Ajj=ΛÞ number of times the state
preparation circuit after applying amplitude amplification.
As compared to the classical counterpart, the algorithm claims exponential
speed-up gain under certain problem statements. The algorithm takes OððlogNÞ2 Þ
 4 From classical to quantum machine learning 77

quantum states to output, whereas it takes OðNlogN Þ steps to output in a classical

computer.

4.3.5 Quantum principal component analysis

Principal component analysis (PCA) [33] is a very important technique for di-
mensionality reduction. In a more formal way, we can define that PCA operates by
P T
diagonalizing the covariance matrix C = j ~ v~
vj of the data, where T is the transpose
operation. The covariance matrix encapsulates the correlations between data’s dif-
ferent components. The covariance matrix in the form of eigenvectors and eigenval-
P
ues can be represented as C = k ek~ ck† , where ek are the eigenvalues correspond
ck~
to the eigenvectors ~
ck of C. The few eigenvalues which are large as compared to the
rests, are known as the principal components. Each of the principal components
are then considered as the new feature vectors. The classical algorithm for PCA
works in Oðd2 Þ runtime complexity, where d is the dimension of the Hilbert space.
In quantum PCA [33], a randomly chosen data vector is mapped into a quantum
P 
state using QRAM [30]. The resulting state has a density matrix ρ = N1 j vj ihvj ,
where N is the cardinality of the data vector set. Using density matrix exponentia-
tion [34] with the quantum phase estimation algorithm [35], and repeating the data
sampling, allow us to decompose the data vectors into principal components. The
quantum PCA algorithm works in OððlogdÞ2 Þ runtime complexity.

4.4 Quantum machine learning approaches

4.4.1 HHL or quantum amplitude based approaches

The objective of amplitude encoding based machine learning algorithms [9, 11, 20,
33, 36–39, 40] is to articulate quantum machine learning algorithms whose resources
grow in polynomial runtime with the n qubits, and expands to a logarithmic growth
with amplitudes and thus the input dimension. Many QML algorithms based on this
approach uses the speedup advantages by solving a system of linear equations quan-
tum mechanically [20]. Quantum matrix inversion is one of the main core processes
in solving the system of linear equations exponentially faster than the classical
counterpart. Some of the notable quantum machine learning algorithms based on
amplitude encoding approach are the quantum support vector machine [21–23]
and least-square linear regression [37, 38].
78 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

4.4.2 Grover’s search-based approaches

Machine learning algorithms based on Grover’s search algorithms, which use the tech-
nique of amplitude amplification to unfold an unstructured search problem with a qua-
dratic runtime speedup, apply Grover’s search subroutine to solve a larger problem.
Quantum K-Means [25], K-Median [41], and k-nearest neighbors [42] are some of the
machine learning algorithms which use Grover’s search algorithms as a subroutine to
solve the machine learning task. Another recent use of the Grove’s search subroutine is
in training the perceptron [43].

4.4.3 Quantum annealing based approaches

Very recently, some experiments have been done based on quantum annealing
[44]. In this work, the researchers used a quantum annealer to classify and rank the
transcription factors binding to study how the proteins (in a gene) identify and spe-
cifically bind to their DNA targets. The researchers used the D-Wave system to im-
plement the quantum annealing algorithm. Here, the machine learning techniques
have been the effective tools to disclose interaction mechanism at the genetic level.

4.5 Recent progress in quantum machine learning

algorithms
The research in developing quantum machine learning algorithms has exponentially
grown in recent years, and day by day new interesting quantum machine learning
algorithms are being published. Here, we are covering the most ground-breaking and
popular algorithms which demonstrated significant state-of-the-art performances as
compared to their classical counterparts.

4.5.1 Quantum binary least square support vector machine

The quantum binary least square support vector machine is a break-through in the
field of QML, where the authors, Rebentrost et al. [21], demonstrated an exponential
speedup gain in support vector machine formulation.
In the least square SVM, the optimization problem is transformed into a system
of linear equations which encompasses the kernel matrix for solving the optimiza-
tion problem. We solve the system of linear equations in a much faster way by solv-
ing the least squares formulation and speeding up the kernel matrix calculations.
 4 From classical to quantum machine learning 79

The following areas where we take the quantum advantages to improve the perfor-
mance of the SVM are:
a. Data preparation with quantum random access memory (QRAM) – One of the
challenging tasks in quantum machine learning is to prepare the data set for
feeding into the quantum systems for processing. QRAM helps in preparing the
classical data set into the quantum form. QRAM takes Oðlog2 dÞ steps to query
the information from memory for reconstructing a state, where d is the dimen-
sion of the feature vector.
b. Kernel matrix calculation – The dot product evaluation plays the main role in
the kernel matrix calculation. So, if we get the speed up advantages in the dot
product evaluation in the quantum formulation, it will imply overall speedup
gain in the kernel matrix calculation. In [35], authors have discussed the faster
way to do dot product calculation with the help of quantum properties. Also,
the exponentiation of the inverse of the normalized kernel matrix K − 1 is per-
formed in a quantum paradigm. As we have discussed earlier, QRAM takes only
Oðlog2 dÞ steps to query for reconstructing a state, thus a single dot product with
QRAM takes O ðϵ − 1 log2 dÞ steps, where ϵ is the accuracy.
c. Least square formulation – The speedup gain is possible during the training
phase because of the quantum mechanical implementation of the exponentially
faster eigenvector formulation in the non-sparse density matrices [35], the
sparse matrix simulation [45], and the matrix inversion algorithm [35].

The least square SVM is represented as the following formulation:

" #
0 1T
^ ~
Fðjb, αiÞ = j~ ^
yi, F = , (4:24)
1 K + γ − 1I
^ = F is the normalized F, and trF is a trace of F. The target is to solve the
where F trF
^ jb, ~
optimization problem, Fð yi, for the parameters b and ~
αiÞ = j~ α. Here, K is the ker-
nel matrix.

4.5.2 Quantum multiclass support vector machine

The SVM is a binary class classification algorithm. To handle multiclass classifica-

tion, we use one-against-all and all-pair algorithms. The quantum multiclass support
vector machine (SVM) is the generalized version of the quantum binary SVM. In [23],
the authors discussed two techniques: the quantum version of the one-against-all as
well as the quantum version of all-pair algorithms. In one-against-all, we first con-
struct and train k binary classifiers and then each of these classifiers classify a given
unknown input. The class for which the corresponding classifier’s probability confi-
dence score is the highest, is selected as the predicted class. In all-pair, we construct
80 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

kðk − 1Þ=2 classifiers. Each classifier is trained with a dataset which contains one pri-
mary class data and the rest of the data for another class, i.e., we make pairwise data-
sets and train the classifiers, resulting in kðk − 1Þ=2 total classifiers. When an unknown
input is to get predicted, these all kðk − 1Þ=2 classifiers predict the class for the given
input and by applying voting mechanism we select the class to be predicted. The fol-
lowing algorithms show the quantum version of these two approaches, for detail ex-
planations please refer to [22, 23].

Algorithm 2: Quantum One-Against-All Algorithm

1. initialize the index Iindex = any randon element, 1 ≤ Iindex ≤ k
2. initialize jVqn i as the vector of all classified class probabilities in QRAM
3. while ðPm < Pthreshold Þ
P
4. initialize the memory as jΨi = p1ffiffi r jrijIindex i
k
5. GROVER-QUANTUM-SEARCH ðjΨi, jVqn i, Iindex Þ
6. if ðjVqn i½Iindex new  > jVqn i½Iindex Þ
7. Iindex = Iindex new
8. return Iindex

Algorithm 3: Quantum All-Pair Algorithm

1. initialize class = any random element, 1 ≤ class index ≤ k
2. initialize jVq i as the vector of all classified classes
3. initialize frequency estimate sclass index with any very small value
4. INITIAL-FREQUENCY-COUNT ( class index, sclassindex Þ
5. while ðtotal running time < O ðlogkÞÞ
P
6. initialize the memory jCi i = j βi, j jji
7. initialize the memory as
X
k−1
jψc i = jiijCi ijclass indexijsclass index i
i=0
 
8. GROVER-QUANTUM-SEARCH (ψc i, Vq i, class index, sclass index Þ
9. MEASURE-REGISTER ðjψc iÞ
 ε

10. if sclass index new > sclass index + 2k
11. class index = class indexnew
12. sclass index = sclass index new
13. return jVq i½class index
 4 From classical to quantum machine learning 81

4.5.3 Quantum K-means clustering

Clustering is a very important problem in machine learning and more generally falls
under unsupervised learning. Clustering algorithm groups the unlabeled dataset into
multiple categories based on their similar features. The K-Means clustering algorithm
is a very popular and simple clustering algorithm. K-Means groups the unlabelled da-
taset into k categories, however, we have to specify K in advance. In [25], authors
show the formulation of a K-Means clustering algorithm in the quantum paradigm.
In the quantum K-Means, the closest centroids are calculated in the same way as in
the classical version, but by using the Grover’s search [10]. The quantum version of k-
Means has been claimed to be exponentially faster in runtime complexity as com-
pared to the classical counterpart. The quantum K-Means is based on an adiabatic
quantum process.
Suppose the output of the quantum K-Means clustering algorithm is in the fol-
lowing quantum state:
X X
jξi = M − 1=2 jcl ijli = M − 1=2 jcijli (4:25)
l c, lϵc

We construct the above output quantum state with O ðϵ − 1 K logðKMN ÞÞ runtime com-
plexity, where ϵ is the accuracy, and with O ðϵ − 1 log ðKMN ÞÞ runtime complexity
when the clusters are well separated. The adiabatic gap in this case is Oð1Þ.
Let us first select K objects with labels ic as initial seeds. We begin with the
state:
 X  X #D
ðMKÞ − 1=2
jc′ijli K − 1=2
jcijic i (4:26)
c′l c

xi ′ j2 in the c′l
xl − ~
Here, D copies of the state permit us to evaluate the distances j~
c
superposition component. By applying the adiabatic algorithm with the initial
Hamiltonian
X
H0 = 1 − jφihφj, jφi = K − 0.5 jc′i (4:27)
c′

The initial clustering is:

X
jψ1 i = M − 1=2 jcijli (4:28)
c, lϵc

We evolve the initial Hamiltonian H0 slow enough to the Hamiltonian:

X  2 
 
H2 = xl − ~
~ xi ′  c′ihc′j#jlihl (4:29)
c
c′l
82 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

P
We can construct the separate cluster jφc1 = M − 0.5 lϵc jli, assuming we have D cop-
ies of jψ1 i. We can now estimate the number of quantum states Mc in the cluster c.
With D copies of jψ1 i together with the clusters jφc1 i, we evaluate the average dis-
tance between xl and the individual cluster’s mean
 
 X 2
~ −1
xK  = j~
~ xl − ~
2
xl − ðMc Þ xc j , (4:30)
Kϵc′
 2
− i~ x  δt  
to each component c′li of the superposition using
x −~
by applying a phase e l c′
the following Hamiltonian,
Xn 2 o  
HF = ~ xc′  c′ihc′j # jlihl # I D
xl − ~ (4:31)
c′l

With the initial Hamiltonian H0 = 1 − jφihφj at the ground state, we initiate the adi-
P   P
abatic evolution on the state fðMKÞ − 0.5 g c′, l c′lijΩ1 i#D , where jφi = K − 0.5 c′ jc′i
is the superposition of the cluster centroids. We get the final state:
( )
X 
ðM Þ − 0.5 c′ili jΩ1 i#D = jΩ2 ijΩ1 i#D (4:32)
c′lϵc′

By repeating the final state D times, we construct D copies of jΩ2 and keep on iterat-
ing the cluster assignment procedure until it gets converged to jξi.

4.5.4 Quantum deep learning

Deep learning (DL) revolutionized the world of machine learning with its perfor-
mance for a large domain of applications. DL is an efficient and effective tool for ma-
chine learning modeling. With the development of GPUs, a very wide area opened for
developing highly complex machine learning models using deep learning, which was
not possible earlier without large availability of high-performance computing. Deep
learning requires a large amount of training data. Processing such a huge volume
of datasets is a computationally complex task, this inspired the exploration of the
quantum advantages for DL algorithms. Recent progress in developing program-
mable photonic circuits and quantum annealers [46–48] open up the possibilities
of developing a quantum framework for deep learning algorithms.
The latest research on Boltzmann machine formulation in a quantum paradigm
[49] demonstrated very strong footsteps towards quantum deep learning. The current
results are focusing on developing quantum deep learning algorithms which do not
require a fully functional quantum computer. These quantum DL algorithms can run
with quantum annealers, which are special-purpose quantum information processors
and they are very easy to construct as compared to larger quantum computers.
 4 From classical to quantum machine learning 83

4.6 Recent progress in the implementation

of quantum machine learning
We discuss here some of the significant developments on implementing quantum
machine learning algorithms.
In 2003 [50], a group of researches discussed that image data can be very effi-
ciently encoded in quantum states. This may reduce the resources to some order of
magnitude to perform some specific and challenging image processing tasks. During
the year of 2009 [51], D-Wave demonstrated some remarkable results with their adia-
batic D-Wave quantum system. Using regularized boosting with a non-convex objec-
tive function, researchers conducted experiments in the D-Wave system to identify
cars in images. In the same year (2009), another research group showed mapping an
input data and memorized the data with a Hamiltonian using a quantum Hopfield
network with adiabatic quantum computation [52].
Since 2013 [53, 54], NASA, Google research, and the University Space Research
Association are working on exploring the D-Wave quantum computer for solving
complex optimization problems, which may have an impact in solving some of the
machine learning algorithms with the help of advantages offered by quantum prop-
erties. In 2015, researchers demonstrated the classification between the digit ‘6’ and
‘9’ by implementing a quantum support vector machine algorithm with an NMR
(Nuclear magnetic resonance) technology on a liquid-state quantum computer [55].
Another important result came in the same year 2015 when a research team imple-
mented an all-optical linear classifier using non-linear photonics [56].
Recently, IBM launched an online framework for access to their quantum
computer (named IBM Q), and the system supports around 32 + qubits to work
with. They allow free access up to 5 qubits to work on the real quantum computer.
We can run small specific quantum routines (e.g., Grover’s search, Hadamard
gate, quantum SWAP test, etc.) into this system, which can help in solving com-
plex machine learning use cases by using the outputs of these specific small quan-
tum subroutines [16]. A fully functional quantum neural network is possible by
implementing nonlinear interactions in the quantum dynamics, recently claimed
by a research team in 2016 [57].
Very recently, in 2017 [58], the researchers demonstrated that their model is
able to generate handwritten digits. The model is a probabilistic generative model
based on arbitrary pairwise connectivity. Also in 2017, researchers encoded image
information with the quantum image in a quantum system and discussed some in-
teresting challenges on this topic [59]. In 2017, a very interesting research work by
Yao et.al. discussed the implementation of quantum image processing and its appli-
cations for detecting edges [60].
In July 2018, Google announced the first public alpha of Cirq, an open source
framework at the First International Workshop on Quantum Software and Quantum
84 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

Machine Learning (QSML). This will allow developers to build quantum algorithms,
which may include some simple quantum machine learning algorithms with small
scale data sets, without requiring a quantum physics background [61]. In 2020,
Google announced TensorFlow Quantum which is a quantum machine learning li-
brary for rapid prototyping of hybrid quantum-classical ML models [62].

4.7 Conclusion
Quantum machine learning is a fast emerging field of research that opens new hori-
zons for artificial intelligence based research and applications. We are seeing active
research developments in this area, and with the advancement of quantum com-
puters and quantum devices, it may solve some of the most challenging problems
that are difficult to solve with classical machine learning like clustering and classifi-
cation training with big data in reasonable time period. Recent advancements in
deep learning helped us train complex machine learning models with the help of
GPUs and high-performance computing, which demonstrated remarkable speed ups
and improved accuracies. More complex machine learning problems require high
volumes of data and tremendous computational power, and, due to the limitations
of silicon-based chip technology, we have significant constraints on the upper
bound of the computational power. Quantum based systems promise significant im-
provements in speed up factors as compared to the classical systems. Some of the
quantum machine learning algorithms, such as the quantum binary/multiclass
SVM, quantum K-Means clustering, and the quantum Boltzmann machine, have re-
vealed substantial performance improvements with big data. However, we have im-
portant challenges that need to be addressed in order to make effective and efficient
quantum machine learning algorithms, such as: how to transform classical big data
into quantum form in a very efficient way; the hardware where such quantum algo-
rithms will be run is still at initial stages of development, so proper testing and vali-
dation of these theoretical quantum machine learning models is still an issue. Also,
the QML opens new doors for exploring other domains like understanding the proc-
essing of neural network based algorithms like deep learning, understanding the
quantum experiments data analysis with the classical machine learning. Small
quantum devices are easier to develop as compared to fully functional quantum
computers. It is believed these small specific quantum devices can speed up the
merging of quantum terminologies and methods with the machine learning ones
and can solve many domain-specific complex problems in a hybrid framework,
where such small quantum devices are used to solve a small complex task in a big
complex problem domain.
 4 From classical to quantum machine learning 85

References
[1] Alpaydin, E. Introduction to Machine Learning. Adaptive Computation and Machine Learning.
MIT Press, Cambridge, MA, 2004.
[2] Wikipedia, cite_note 5-2.
[3] Schuld, M., Sinayskiy, I., & Petruccione, F. An in-troduction to quantum machine learning.
Contemporary Physics. 2014, 56(2),172–185.
[4] Benedetti, M., Realpe-Gómez, J., Biswas, R., & Perdomo-Ortiz, A. Quantum-Assisted Learning
of Hard-ware-Embedded Probabilistic Graphical Models. Physical Review X. ISSN 2160-3308.
2017, 7(4),041052. arXiv:1609.02542. doi:10.1103/PhysRevX.7.041052
[5] Farhi, E. & Neven, H. Classification with Quantum Neural Networks on Near Term Processors.
2018-02-16. arXiv:1802.06002.
[6] Schuld, M., Bocharov, A., Svore, K., & Wiebe, N. Circuit-centric quantum classifiers. 2018-04-
02. arXiv:1804.00633.
[7] Yu, S., Albarran-Arriagada, F., Retamal, J. C., Wang, Y. -T., Liu, W., Ke, Z.-J., Meng, Y., Li, Z.-P.,
and Tang, J.-S. Reconstruction of a Photonic Qubit State with Quantum Rein-forcement
Learning. 2018-08-28.
[8] Broecker, P., Assaad, F. F., & Trebst, S. Quantum phase recognition via unsupervised
machine learning. 2017-07-03. arXiv:1707.00663
[9] Huembeli, P., Dauphin, A., & Wittek, P. Identifying Quantum Phase Transitions with
Adversarial Neural Networks. Physical Review B. 2018, 97(13),134109. arXiv:1710.08382.
doi:10.1103/PhysRevB.97.134109. ISSN 2469-9950
[10] Grover, L. K. A fast quantum mechanical algorithm for database search. Proceedings, 28th
Annual ACM Symposium on the Theory of Computing. May 1996, 212.
[11] Melnikov, A. A., Nautrup, H. P., Krenn, M., Dunjko, V., Tiersch, M., Zeilinger, A., & Briegel,
H. J. Active learning machine learns to create new quantum experiments. Proceedings of the
National Academy of Sciences. ISSN 0027-8424. PMC 5819408. PMID 29348200. 2018. 115(6),
1221–1226. arXiv:1706.00868. doi:10.1073/pnas.1714936115
[12] Huggins, W., Patel, P., Whaley, K. B., & Stoudenmire, E. M. Towards Quantum Machine
Learning with Tensor Networks. 2018. arXiv:1803.11537
[13] Carleo, G., Nomura, Y., & Imada, M. Constructing exact representations of quantum many-
body systems with deep neural networks. 2018-02-26. arXiv:1802.09558
[14] Bény, C. Deep learning and the renormalization group. 2013-01-14. arXiv:1301.3124
[15] https://www.dwavesys.com/quantum-computing
[16] https://quantumexperience.ng.bluemix.net/qx/experience
[17] https://ai.google/research/teams/applied-science/quantum-ai/
[18] Adiabatic Quantum Computation. Adiabatic quantum computation, Wikipedia, the free
encyclopedia, 2016-06-20.
[19] Shor, P. W. AT&T Research. arxiv.org/abs/quant-ph/9508027
[20] Harrow, A. W., Hassidim, A., & Lloyd, S. Quantum Algorithm for Linear Systems of Equations.
Phys. Rev. Lett. 2009, 103, 150502.
[21] Rebentrost, P., Mohseni, M., Lloyd, S. Quantum support vector machine for big data
classification. 2014. arXiv:1307.0471v3
[22] Bishwas, A. K. Big Data Classification with Quantum Multiclass SVM and Quantum One-
Against-All Approach. Conference Proceedings to IEEE 2nd International Conference on
Contemporary Computing and Informatics (IC3I). 2016. DOI: 10.1109/IC3I.2016.7918805
[23] Bishwas, A. K. An All-Pair Approach for Big Data Multiclass Classification with Quantum SVM.
Springer Nature’s Quantum Information Processing Journal. September 2018.
doi.org/10.1007/s11128-018-2046-z
86 Arit Kumar Bishwas, Ashish Mani, and Vasile Palade

[24] Bishwas, A. K. Quantum Supervised Clustering Algorithm for Big Data. IEEE 2018 3rd
International Conference for Convergence in Technology (I2CT). April 2018.
DOI: 10.1109/I2CT.2018.8529799
[25] Lloyd, S., Mohseni, M., Rebentrost, P. Quantum algorithms for supervised and unsupervised
machine learning. 2013. arxiv.org/abs/1307.0411
[26] https://en.wikipedia.org/wiki/Bra%E2%80%93ket_notation
[27] https://en.wikipedia.org/wiki/Quantum_superposition
[28] https://en.wikipedia.org/wiki/Quantum_entanglement
[29] https://people.cs.umass.edu/~strubell/doc/quantum_tutorial.pdf
[30] Giovannetti, V., Lloyd, S., Maccone, L. Quantum random access memory. Phys. Rev. Lett.
2008, 100, 160501. arXiv: 0708.1879
[31] https://peterwittek.com/understanding-quantum-svms.html
[32] Schuster, D. I., Sears, A. P., Ginossar, E., DiCarlo, F., Frunzio, L., Morton, J. J. L., Wu, H.,
Briggs, G. A. D., Buckley, B. B., Awschalom, D. D., & Schoelkopf, R. J. High-Cooperativity
Coupling of Electron-Spin Ensembles to Superconducting Cavities. Phys. Rev. Lett. 2019, 105,
140501.
[33] Lloyd, S., Mohseni, M., Rebentrost, P. Quantum principal component analysis. 2013. arxiv.
org/abs/1307.0401
[34] Lloyd, S., Mohseni, M., Rebentrost, P. Quantum principal component analysis. Nat. Phys.
2014, 10(631).
[35] Harrow, A. W., Hassidim, A., & Lloyd, S. Quantum Algorithm for Linear Systems of Equations.
Phys. Rev. Lett. 2009, 103, 150502.
[36] Krenn, M. Automated Search for new Quantum Experiments. Physical Review Letters. PMID
26991161. 2016, 116(9),090405. arXiv:1509.02749. Bibcode:2016PhRvL.116i0405K.
doi:10.1103/PhysRevLett.116.090405
[37] Wiebe, N., Braun, D., & Lloyd, S. Quantum Algorithm for Data Fitting. Physical Review Letters.
PMID 23006156. 2012, 109(5), 050505. arXiv:1204.5242. Bibcode:2012PhRvL.109e0505W.
doi:10.1103/PhysRevLett.109.050505
[38] Schuld, M., Sinayskiy, I., & Petruccione, F. Prediction by linear regression on a quantum
computer. Physical Review A. 2016, 94(2), 022342. arXiv:1601.07823. Bib-code:
2016PhRvA.94b2342S. doi:10.1103/PhysRevA.94.02234
[39] Zhao, Z., Fitzsimons, J. K., & Fitzsimons, J. F. Quantum assisted Gaussian process regression.
2015. arXiv:1512.03929
[40] Berry, D. W., Childs, A. M., & Kothari, R. Hamiltonian simulation with nearly optimal
dependence on all parameters. 56th Anual Symposium on Foundations of Computer Science.
IEEE. 2015, 792–809. arXiv:1501.01715. doi:10.1109/FOCS.2015.54
[41] Aïmeur, E., Brassard, G., & Gambs, S. Quantum speed-up for unsupervised learning. Machine
Learning. ISSN 0885-6125. 2013, 90(2),261–287. doi:10.1007/s10994-012-5316-5
[42] Wiebe, N., Kapoor, A., & and Svore, K. Quantum Algorithms for Nearest-Neighbor Methods for
Supervised and Unsupervised Learning. Quantum Information & Computation. 2014, 15(3),
0318–0358. arXiv:1401.2142
[43] Wiebe, N., Kapoor, A., & Svore, K. M. Quantum Perceptron Models. Advances in Neural
Information Processing Systems. 2016, 29, 3999–4007. arXiv:1602.04799
[44] https://www.nature.com/articles/s41534-018-0060-8
[45] Berry, D. W., Ahokas, G., Cleve, R., Sanders, B. C. Efficient quantum algorithms for simulating
sparse Hamiltonians. Communications in Mathematical Physics. 2007, 270, 359–371.
[46] Adachi, S. H., Henderson, M. P. Application of quantum annealing to the training of deep
neural networks. 2015. arXiv:1510.06356 [quant-ph]
 4 From classical to quantum machine learning 87

[47] Denil, M., de Freitas, N. Toward the implementation of a quantum rbm. NIPS Deep Learning
and Unsupervised Feature Learning Workshop. 2011, 5.
[48] Dumoulin V., Goodfellow, I. J., Courville, A., Bengio, Y. On the challenges of physical
implementations of RBMs. 2013. arXiv:1312.5258 [stat.ML]
[49] Amin, M. H., Andriyash, E., Rolfe, J., Kulchytskyy, B., Melko, R. Quantum Boltzmann machine.
2016. arXiv:1601.02036 [quant-ph]
[50] Wadcock, J., Allen, E., Day, M., Frick, S., Hinchliff, J., Johnson, M., Morley-Short, S., Pallister,
S., Price, A., & Stanisic, S. Advances in quantum machine learning. 2015. arXiv:1512.02900
[51] Neven, H., Denchev, V., Drew-Brook, M., Zhang, J., & Macready, W. NIPS 2009 Demonstration:
Binary Classification using Hardware Implementation of Quantum Annealing. 2010.
[52] Neigovzen, R., Neves, J. L., Sollacher, R., & Glaser, S. J. Quantum pattern recognition with
liquid-state nuclear magnetic resonance. Physical Review A. 2009, 79(4), 042321.
arXiv:0802.1592. doi:10.1103/PhysRevA.79.042321
[53] Google Quantum A.I. Lab Team. Google Plus. 31 January 2017.
[54] NASA Quantum Artificial Intelligence Laboratory, NASA. 31 January 2017.
[55] Li, Z., Liu, X., Xu, N., & Du, J. Experimental Realization of a Quantum Support Vector
Machine. Physical Review Letters. PMID 25910101. 2015, 114(14), 140504. arXiv:1410.1054,
doi:10.1103/PhysRevLett.114.140504
[56] Tezak, N. & Mabuchi, H. A coherent perceptron for all-optical learning. EPJ Quantum
Technology. 2015, 2. arXiv:1501.01608. doi:10.1140/epjqt/s40507-015-0023-3
[57] Li, Y., Holloway, G., Benjamin, S., Briggs, G. Baugh, J., & Mol, J. A simple and robust
quantum memristor. Physical Review B. 2017, 96(7),075446. arXiv:1612.08409.
doi:10.1103/PhysRevB.96.075446
[58] Benedetti, M., Realpe-Gómez, J., Biswas, R., & Perdomo-Ortiz, A. Quantum-assisted learning
of graphical models with arbitrary pairwise connectivity. Physical Review X. 2017, 7(4),
041052. arXiv:1609.02542. doi:10.1103/PhysRevX.7.041052
[59] Yan, F., Iliyasu, A., & Le, P. Quantum image processing: A review of advances in its security
technologies. International Journal of Quantum Information. 2017, 15(3), 1730001.
[60] Yao, X. W., Wang, H., Liao, Z., Chen, M. C., Pan, J., Li, J., Zhang, K., Lin, X., Wang, Z., Luo, Z.,
Zheng, W., Li, J., Zhao, M., Peng, X., & Suter, D. Quantum Image Processing and Its
Application to Edge Detection: Theory and Experiment. Physical Review X. 2017, 7(3), 031041.
arXiv:1801.01465
[61] https://ai.googleblog.com/2018/07/announcing-cirq-open-source-framework.html
[62] https://www.tensorflow.org/quantum
Alokananda Dey, Sandip Dey, Siddhartha Bhattacharyya,
Jan Platos, and Vaclav Snasel
5 Quantum inspired automatic clustering
algorithms: A comparative study of Genetic
algorithm and Bat algorithm
Abstract: This article is intendant to present two automatic clustering techniques of
image datasets, based on quantum inspired framework with two different meta-
heuristic algorithms, viz., Genetic Algorithm (GA) and Bat Algorithm (BA). This work
provides two novel techniques to automatically find out the optimum clusters present
in images and also provides a comparative study between the Quantum Inspired
Genetic Algorithm (QIGA) and Quantum Inspired Bat Algorithm (QIBA). A comparison
is also presented between these quantum inspired algorithms with their analogous
classical counterparts. During the experiment, it was perceived that the quantum in-
spired techniques beat their classical techniques. The comparison was prepared based
on the mean values of the fitness, standard deviation, standard error of the computed
fitness of the cluster validity index and the optimal computational time. Finally, the
supremacy of the algorithms was verified in terms of the p-value which was computed
by t-test (statistical superiority test) and ranking of the proposed procedures was pro-
duced by the Friedman test. During the computation, the betterment of the fitness
was judge by a well-known cluster validity index, named, DB index. The experiments
were carried out on four Berkeley image and two real life grey scale images.

Keywords: automatic clustering, quantum computing, meta-heuristic algorithm, ge-

netic algorithm, bat algorithm, DB Index, statistical test (t-test), Friedman test

5.1 Introduction
Clustering [1, 2] is a process of separating a heterogeneous dataset of different types
of objects into some meaningful sets of homogeneous objects. Mathematically clus-
tering can be described as follows. A dataset DS of K number of points can be repre-
sented as follows.
DS = fdp1 , dp2 , dp3 , . . . , dpK g, in which, each dpi represents a single point of
the specific dataset DS, where i = f1, 2, . . . , K g. Now, in order to achieve a successful

Alokananda Dey, RCC Institute of Information Technology, Kolkata, West Bengal, India.
Siddhartha Bhattacharyya, CHRIST (Deemed to be University), Bangalore, Karnataka, India.
Sandip Dey, Sukanta Mahavidyalaya, Jalpaiguri, West Bengal, India.
Jan Platos, Vaclav Snasel, VSB Technical University of Ostrava, Czech Republic.

https://doi.org/10.1515/9783110670707-005
90 Alokananda Dey et al.


cluster CN of the dataset DS into p number of partition namely, CN 1 , CN 2 , . . . , CN p
the following properties must be hold.

CN i ≠ ∅, ∀i = f1, 2, . . . , pg.
CN i ∩ CN j = ∅, ∀i, j = f1, 2, . . . , pg and i ≠j.
Xp
C = DS.
i = 1 Ni

Clustering can be achieved in several ways and researchers are involved to design
several methods for successful clustering, viz., K-means [3], Partitioning methods [4],
Hierarchical clustering [5], Model-based clustering [6], Fuzzy clustering [7], Density-
based clustering [8], etc. It is seen that among these methods, the K-means [3] out
performs over others. In case of K-means algorithm an apriori knowledge is always
required about the existing number of clusters within a dataset. On the other hand,
an automatic clustering technique can be developed to discover the optimum number
of clusters present in the dataset, in spite of knowing very less information or without
having any information about the inspected dataset, which may yields usefulness in
order to solve the purpose of segmentation as well as classification problem, which
influenced the researchers to explore this field of clustering with automation.
Nature inspired meta-heuristic algorithms are very much useful for providing
an optimal solution from a simple or complex optimization problems within a small-
est time stamp. Some renowned nature inspired population based meta-heuristic
algorithms are genetic algorithm (GA) [9], particle swarm optimization (PSO) [10],
differential evolution (DE) [11], bacterial foraging optimization (BFO) [12], ant colony
optimization (ACO) [13], bat algorithm [14, 15] etc. Some nature inspired meta-
heuristic algorithms used for automatic clustering techniques are available in the
literature [16–20]. Though, these algorithms can be able to identify a solution from
a population within a minimum time frame still they may suffer from early conver-
gence. In order to resolve the premature convergence problem, researchers are very
much interested to associate the nature inspired meta-heuristic algorithms with the
silent features of quantum computing, which yields the quantum inspired techniques
to flourish day by day. Various kinds of computational problems can be solved by
using the effects of different quantum mechanical phenomena which are presented
in [21]. Several quantum inspired techniques associated with various meta-heuristic
algorithms are invented till now and the efficiency of the quantum inspired techni-
ques over their corresponding classical counterparts is established based on conver-
gence, superior value of fitness and other related parameters [23–33]. These type of
quantum inspired algorithms are developed for solving various kinds of optimization
problems like mathematical function optimization [34], combinational optimization
problem (knapsack problem) [35], multilevel image thresh holding [23–27], image
analysis [22], automatic clustering [28–32] and task scheduling in distributed system
[36], to name a few.
 5 Quantum inspired automatic clustering algorithms 91

In this article, two different quantum inspired meta-heuristic algorithms, viz.,

QIGA and QIBA are proposed for automatic clustering technique of image datasets
and also a comparison was produced between two different classical algorithms,
viz., the classical genetic algorithm and the classical bat algorithm with their corre-
sponding quantum influenced counterparts. The computed fitness values of the
mean, standard deviation of that fitness, standard error of that fitness and the mini-
mum computational time were the yardstick for adjudging the efficacy of these
quantum inspired techniques over their analogous classical counterparts. Though,
the experimental results are in favor of the quantum inspired techniques, still a sta-
tistical superiority test, viz., the unpaired t-test [37] was conducted and the corre-
sponding evaluated p value establishes the efficacy of these quantum inspired
techniques. Finally, Friedman test [38–39] was conducted among all the proposed
procedure to judge the superiority of a procedure over another. The cluster validity
index namely, DB index [40] was used as an objective function to compute the va-
lidity of the clustering.
The rest of the article is organized as follows. The overview of the genetic algo-
rithm and the overview of the bat algorithm are summarized in Section 5.2 and 5.3 re-
spectively. The fundamental idea of Quantum Computing is discussed in Section 5.4.
The Section 5.5 presents the overview of the Cluster Validity Index (DB-Index). The
steps involved in the proposed Quantum Inspired Genetic Algorithm and Quantum
Inspired Bat Algorithm are presented in Section 5.6 and 5.7 respectively. The experi-
mental outcomes and analysis are discussed in Section 5.8. Finally, a conclusion is
drawn in Section 5.9.

5.2 Overview of Genetic algorithm

The well-known nature inspired meta-heuristic algorithm named, Genetic Algorithm
(GA) is a kind of evolutionary algorithm based on the concept of Darwin’s theory of
evolution. John Holland is known as the inventor of GA, who proposed this algorithm
in the year 1975 [9].
GA is a very useful population based adaptive searching process, in which each
individual in a population is called a chromosome. In order to achieve the optimal
solution GA basically explores the entire search space. During the search process the
current population evolves to a same sized next population by applying three genetic
operations, such as selection, crossover and mutation. In the successive iteration, the
best chromosomes are selected among all the participating chromosomes from the
previous pool and the selection is carried over depending upon their computed fitness
value. Several methods can be used for the purpose of selection, such as Boltzmann
selection, Roulette Wheel selection, Rank selection, Tournament selection etc. The
crossover operation is used to produce a new pair of chromosomes by combining the
92 Alokananda Dey et al.

portions of two or more randomly selected parent chromosomes which lead to pro-
duce a population by replacing the parent chromosomes by the new one. During the
crossover operation one or multiple points can be selected as the partitioning points
from each of the individual in the population. In order to ensure the genetic diversity
in the population mutation operation takes place in which a randomly selected posi-
tion of a randomly selected chromosome is altered depending upon a predefined
value called mutation probability. This kind of alteration may be performed on a sin-
gle point or multiple points of one or more randomly selected individuals in the popu-
lation. In order to achieve faster convergence it is recommended to perform elitism at
each generation. GA was already successfully applied in many applications like data
clustering [17, 18], portfolio management [41], image processing [42] and many more.
Basic steps involved in Genetic algorithm are presented as follows.
1. Population of chromosomes is initialized randomly.
2. Fitness of each of the chromosome is evaluated.
3. Selection of chromosomes for next generation is done.
4. Do the following steps until the termination condition are satisfied.
5. Crossover is accomplished between a pair of randomly selected parent chro-
mosomes to get child chromosomes, depending upon a predefined crossover
probability.
6. Mutation is applied on the randomly selected offspring/ chromosome to intro-
duce diversity in the population, depending upon a predefined mutation
probability.
7. Selection is carried out for generating the next generation population.
8. End Loop.

5.3 Overview of Bat algorithm

In the year 2010, an efficient nature inspired population based meta-heuristic algo-
rithm, viz., Bat algorithm and this algorithm was initially developed by Yang [14].
This algorithm is capable to identify the optimal solution or near optimal solution
from a given problem, which can be an optimization problem of simple or complex
type. Bats generally use echolocation for sensing the prey and recognizing their
roosting crevices at the dark. They can also avoid their upcoming obstacles by
using echolocation. Generally, bats can be able to produce a very loud sound pulse
and listen for the bounced back echo from their surroundings. Depending upon the
loudness of the response and the time delay from the emission to the reflection they
can understand the shape, size and the velocity of the prey and can navigate ac-
cordingly. It is known that the high frequency emitted sound pulses influence them
to give detailed information about their close environment so that they are able to
detect the position of the prey precisely and allow them to move shorter distances.
 5 Quantum inspired automatic clustering algorithms 93

But, if the emitted sound pulses have low frequency then they travel longer dis-
tance and provide rough information about their surroundings. Also the higher val-
ues of the loudness of the echolocation indicate that they are nearer to the prey and
produce sound pulse very quickly. The following process describes the basic steps
involved in the hunting strategy of the bats.
1. All bats are able to detect the distance between the prey/food and the obstacles
of their surroundings by using the echolocation characteristics.
2. Bats are able to fly randomly from a position Pi with the velocity Vi, a fixed fre-
quency Fmin , with varying wavelength λ and loudness L0 to search the prey.
3. Their loudness can be changed from a large number L0 to a minimum constant
number Lmin and also the frequency can vary from Fmin to Fmax .

Let us considered in a search space, n numbers of bats are flying randomly. In a

time stamp t, they are initiated their search from the position Pi having initial veloc-
ity Vi, frequency Fmin , pulse rate ri and loudness L0. At time stamp t + 1, they may
change their next position and velocity by using the following equations.

Fi = Fmin + ðFmin − Fmax Þβ (5:1)

 
Vit + 1 = Vit + Pit + 1 − Pbest Fi (5:2)
Pit + 1 = Pit + Vit + 1 (5:3)

where, β 2 ½0, 1. The minimum and the maximum frequencies are Fmin = 0 and
Fmax = 2 respectively. Here, Pbest signifies the global best bat’s location among the
positions of the n numbers of bats in the current population. After evaluating the
fitness of all solutions in the population, the position of the best solution among
them is identified as the global best position. Now, it remains to check whether the
bats are near or far away from the prey by using their pulse rates. A low pulse rate
ri with high probability (rand1 > ri) indicates that the bats will fly near to the current
superior bat before carrying out a short random fly there. Otherwise, it is considered
that the bats are nearer to the prey and they will do a random fly around their cur-
rent position. Here, rand1 is a uniformly distributed random number between [0, 1].
The following equation is used to represent the new positions of the bats.

Pnew = Pold + εLt (5:4)

where, ε 2 ½ − 1, 1 and at the time stamp t, the average loudness of n bats is com-
P
puted by Lt = n1 ni= 1 Lti . After the random fly, if a bat is able to achieve a better posi-
tion than the current global best position and its emitted sound is loud enough and
also better than a random number (rand2 < Li ), then it is considered that the bat will
fly to this new position and the current global best position value will be updated
with the newly generated position value. Here rand2 is a random number which
is distributed uniformly between [0, 1]. Now, the pulse rate (ri) of the bats are
94 Alokananda Dey et al.

increased and the loudness (Li ) of the bats are decreased simultaneously by the fol-
lowing equations.
 
rit + 1 = ri0 1 − e − γt (5:5)
Lti + 1 = αLti (5:6)

where, ri0 2 ½0, 1 represents the initial pulse rate in which γ ð0 < γ < 1Þ and
α ð0 < α < 1Þ are constants. Now, the current Pbest is selected by computing the fit-
ness of all the bats.
Basic steps involved in Bat algorithm are presented as follows.
1. Bat population P is initialized randomly.
2. Pulse frequency (F Þ, velocity (V), pulse rate (ri ) and loudness (L0 ) are initial-
ized for each of the individual.
3. Fitness fn of each individual is computed.
4. Do the following steps until the iteration number ( Jn ) reach to the specified
maximum iteration number (MaxJ).
5. New bat population P n is generated by adjusting the frequency (F n ) and veloc-
ity (V n ) of each individual. (eqs. (5.1)–(5.3)).
6. if ðrand > ri Þthen
7. Among all the best solutions, a single solution is selected and a local solution
is produced around this selected solution (best solution). (eq. (5.4))
8. end if
9. By producing a random fly, a new solution is generated.
10. if ðrand < L0i and fn ðP i Þ < fn ðP best ÞÞthen
11. The new solution is accepted.
12. Pulse rate (ri ) is increased. (eq. (5.5))
13. Loudness (L0i ) is decreased. (eq. (5.6))
14. end if
15. The current best bat is identified.
16. end loop

5.4 Fundamental idea of quantum computing

A classical computer is capable to hold a single binary bit 0 or 1 at a time. Unlike
classical computer, a quantum computer is capable to store a unit of information
which is called quantum bit or qubit, in which a single qubit is able to store the
numbers zero (0) and one (1) at the same moment, which referred as the superposi-
tion state in quantum computing [24–26]. It is known that the qubits are able to ac-
complish numerous processes simultaneously which yields fastest execution of
process. In a two-dimensional Hilbert space, a qubit are represented by a unit vector
 5 Quantum inspired automatic clustering algorithms 95

and are considered as a two state quantum-mechanical phenomena. In QC, the

superpositions of the basis states are represented by the following equation.
X
n
jΨi = Ci jV i i
i=1

jΨi = C1 jV 1 i + C2 jV 2 i + . . . + Cn jV n i (5:7)

where, V i is the ith quantum states and Ci 2 C. For a two states quantum bit or
qubit, eq. (5.7) can be denoted by a linear superposition as jΨi = C1 j0i + C2 j1i, in
which, the state j0i is acknowledged as “ground state” and j1i is acknowledged as
“excited state”, and Ci belongs to the set of complex number which should satisfy
the following normalization condition.
X
n
jCi j2 = 1 (5:8)
i=1

In QC, there exist two striking features, viz., Coherence and Decoherence. Coherence
is represented by the linear superposition ðjΨiÞ of the basis states whereas, decoher-
ence is occurred when a forceful destruction is happened on the previously defined
linear superposition. Quantum computing generally uses quantum logic gates as hard-
ware devices.
A predefined unitary operator is used to update the qubits within a fixed time
period. Mathematically, the unitary U operator quantum gates hold the relation
U † = U − 1 and UU † = U † U = I.
In quantum computer a n qubit quantum register is able to store 2n states simul-
taneously. As for example the eight states |000〉, |001〉, |010〉, |011〉, |100〉, |101〉, |110〉
and |111〉 require three quantum registers in which the three qubit system can be
 
α1 α2 α3
represented as β1 β2 β3
where α2i + β2i = 1, i = 1, 2, 3.
But in case of classical computer in order to store these eight states eight regis-
ters are required.
There are several quantum gates exists, such as Rotation Gate, Pauli – X Gate
or NOT Gate, C-NOT Gate, Toffoli Gate, Fredkin Gate and Hadamard Gate [24–26]
etc. and they can be used in several ways in different types of problems.
Rotation gate strategy can be used to update the values of ðαi , βi Þ of the ith
qubit by the following equation.
! ! !
α′i cos θi − sin θi αi
= (5:9)
β′i sin θi cos θi βi

Here, the rotation angle for each of the qubit is represented by θi , which is chosen
according to specific problem. If αi and βi represent the probability amplitude of a
qubit before rotation then after rotation they become α′i and β′i respectively and can
be achieved by the following way.
96 Alokananda Dey et al.

! !
α′i cos ðθ + δÞi
= (5:10)
β′i sin ðθ + δÞi



αi cos δi
where =
βi sin δi

5.5 Cluster validity index

In the year 1979, David L. Davies and Donald W. Bouldin introduced an efficient
cluster validity index, named, Davies-Bouldin (DB) Index [40], which was used
in this article as an objective function to compute the fitness of each individual.
The optimum number of clusters within a given dataset is determined by the cal-
culated value of the DB index [40], where, minimum value of DB index [40] is
taken for consideration. It can be considered as a function of the ratio between
the sum of within-cluster scatter and the cluster separation. Let us consider the
cluster similarity measure is CMij between two different clusters Cli and Clj . It is
given by:

dMi + dMj
CMij = (5:11)
dci cj

Here dMi be the dispersion measure of ith cluster and is denoted as

 2 21
1 Xtn oi 
 ðiÞ 
dMi = k=1
Dp k − Cci 
tn oi

such that, tn oi denotes the total number of objects and Cci denotes the ith cluster
ðiÞ
center in cluster Cli and ∀Dpk 2 Cli .
The distance between the clusters Cli and Clj and can be defined as
 
dci cj = Cci − Ccj .

The following equation is used to define the Davies – Bouldin (DB) index [40].
1 Xnoc
DB I = H
i=1 i
(5:12)
noc
 
Where Hi = max j = 1, 2, ..., noc , i≠j CMij , i = 1, 2, . . . , noc
 5 Quantum inspired automatic clustering algorithms 97

5.6 Quantum inspired Genetic algorithm

5.6.1 Proposed methodology for QIGA

In this section, the operational strategy of quantum inspired GA algorithm is dis-

cussed. In order to recognize the optimum number of cluster presents within im-
ages, the proposed method was experimented on image dataset.
At first, a population ðPcÞ was initialized with N number of chromosomes. Each
individual chromosome was of length Ln from the population was considered as a
solution of the problem. (Line 1).
Now, the population ðPcÞ was encoded to produce the quantum states Qs . The
encoding process is discussed as follows. 
Initially for each of the chromosome of the population a state θsi, j 0 ≤ θsi, j ≤ 2π
was created randomly. Using the value of θi, j the quantum state Qs was created. For
a particular tth generation the quantum state Qs ðtÞ was represented as follows.
Qs ðtÞ = fqst
1 , q2 , . . . , qN g,
st st

where, qi ði = 1, 2, . . . , N Þ denotes the arbitrary elements of Qs ðtÞ and can be de-

st

noted as
" st #
αi, 1 j αsti, 2 j. . .jαi, m j
st
qi =
st
; here m = Ln
βst st st
i, 1 j βi, 2 j . . . j βi, m j

st st st st
and αsti, j = cos θi, j and βi, j = sin θi, j was computed from the random value of θi, j ,
where i = f1, 2, . . . , N g and j = f1, 2, . . . , Lng.(Line 2).
Now, in order to generate the active cluster points a selection procedure was ap-
plied on the population and the selection process was guided by the state Qs ðtÞ to pro-

duce a binary solution BcðtÞ. The binary solution BcðtÞ = bct1 , bct2 , . . . , bctN was
produced by observing the state Qs ðtÞ in the initial generation t = 1.
Each of the binary solution bcti was generated as follows.
(
1, If βst
i, j > αi, j
st
bcti, j = , s.t.i = f1, 2, . . . , N g and j = f1, 2, . . . , Lng.
0, Otherwise

The criteria bcti, j = 1 was set to identify the activated cluster points from the popu-
lation PcðtÞ for the tth generation, where i = f1, 2, . . . , N g and j = f1, 2, . . . , Lng.
(Line 3–8).
Now, the fitness of each individual chromosome from the quantum population
was computed by an objective function (DB index [40] – eq. (5.12)).
The best chromosomes from the quantum population were chosen for generat-
ing the next generation population. (Line 9).
Now, a quantum rotation gate Gsti, j ðϕÞ was used for updating each quantum
state Qs ðtÞ, which can be denoted as follows.
98 Alokananda Dey et al.

" #
cos ϕst st
i, j − sinϕi, j
Gsti, j ðϕÞ = ,
sin ϕst st
i, j cos ϕi, j

where, the rotation angle ϕsti, j was chosen randomly between (−0.5, 0.5).
The updated quantum state Qnew ðtÞ was identified by the eq. (5.10).(Line 11).
In case of quantum inspired genetic algorithm, the probability amplitude ðα, βÞ
of qubits were used to encode the chromosomes. After that quantum rotation gate was
used to update ðα, βÞ, which lead to realize the genetic operation for generating the
offspring instead of determining them from their parents as the chromosomes were in
superposition or entangled state. In general, rotation influences the search space to
extend as it is able to produce 2n number of states from n qubit chromosome. The dif-
ferent values of ðα, βÞ were responsible for generating the states (Lines 12–18).
A predefined mutation probability was used to perform the quantum mutation
operation. The probability amplitude α would be converted to β and β would be
converted to α for the greater value of the predefined mutation probability from a
random number between (0,1). The following quantum NOT gate strategy was used
to alter the probability amplitude ðα, βÞ to ðβ, αÞ.
    
0 1 αi, j β
= i, j
1 0 βi, j αi, j
    
0 1 cos θi, j sin θi, j
= (Lines 19–26).
1 0 sin θi, j cos θi, j

The generation number t was updated by t = t + 1 and perform the above mentioned
procedure for a predefined number of generation or until it met the stopping crite-
ria. (Lines 10–27).
The best chromosome was identified and preserved along with its fitness and
number of cluster.(Line 28).
The working principle for QIGA is presented as follows.

Algorithm: Quantum Inspired Genetic Algorithm for automatic clustering

Input Parameter:
Maximum number of generation: MxG
Size of population: N
Mutation probability: μ
Output parameter:
Optimum cluster number: O N C
Optimum fitness value: O F T

1. At the beginning, a population Pc was created randomly from an image dataset

with N number of chromosomes where each chromosome was considered as a
solution of the problem. During this process, the intensity value of each pixel of
 5 Quantum inspired automatic clustering algorithms 99

the input image was normalized. The length of each chromosome was taken as
Ln.The maximum pixel intensity value of the input image first passed through
the square root function then was assigned to Ln. The image was also normal-
ized between 0 & 1 and named it as Inormal .
2. Now quantum state QP ðtÞ was created and each chromosome of Pc was encoded
to produce Pα and Pβ using the information of the quantum state.
3. for i = 1 to N do
4. for j = 1 to Ln do
5. N Ci numbes of active cluster centers were created from Pci by using the condi-
β
tion Pij > Pijα and stored the cluster centers in CPi .
6. Now, the fitness F T i was computed from CPi with the help of a cluster validity
index, named, DB − Index. [eq. (5.12)]
7. end for
8. end for
9. The best chromosome was identified and preserved along with its fitness value
F T 2 F T i , ∀i 2 N and its corresponding number of cluster point N C .
10. for t = 1 to MxG do
11. The quantum state QN ðtÞ was created by applying the rotation gate on QP ðtÞ.
12. Continued step – 2 for the newly created quantum state QN ðtÞ.
13. Continued steps- (3–8) for identifying the best chromosome.
14. If the best chromosome was identified in the population for the quantum state
QN ðtÞ then
15. The value of the quantum state was updated by QN ðtÞ
16. Otherwise
17. The previous value of the quantum state QP ðtÞ was preserved
18. end if
19. for i = 1 to N do
20. If the current chromosome was not the best chromosome then
21. j = randð0, LnÞ
22. If μ > randð0, 1Þ then
β
23. Exchanged Pijα and Pij
24. end if
25. end if
26. end for
27. end for
28. Finally, the algorithm produced and reported the optimum number of cluster
O N C along with its associated value of fitness O F T .
100 Alokananda Dey et al.

5.7 Quantum inspired Bat algorithm

5.7.1 Proposed methodology for QIBA

Now a day, the bat algorithm is widely applied in several types of complex optimi-
zation problems as it follows an easy mechanism for implementation and its quick
convergence rate which sometimes leads a premature convergence. The premature
convergence can be occurred for several reasons like the bats may be trapped into
local optima due to the decreasing diversity of the bats or the most promising bat
falls into the local optima then it influences all other bats to follow the same trajec-
tory as other bats are instructed to follow the current optimal solution and more-
over in bat algorithm there is no specified ways to get rid of the local optima. In
order to solve this problem a quantum inspired bat algorithm was proposed in this
article. The diversity of the bat population was achieved through quantum rotation
gates which helped the proposed algorithm to avoid premature convergence with-
out affecting the convergence speed though the efficiency of the proposed algo-
rithm in terms of computational time was guaranteed [43].
The anticipated algorithm is established on the framework of bat algorithm
with the flavor of c quantum computing. In order to control the exploitation and
exploration, the parameters BR and BA are used and they are updated for guiding
the local and global search process. Though, the new velocity and positions of each
of the candidates are updated through the following equation.
8
>
< Vit , if ðδi , δbest Þ = ð0, 0Þ
 
Vit + 1 = Vit + Pit + 1 − Pbest , if ðδi , δbest Þ = ð1, 1Þ (5:13)
>
: t  t+1 
Vi + Pi − Pbest rand, if ðδi , δbest Þ = fð1, 0Þorð0, 1Þg

Pit + 1 = Pit + Vit + 1 (5:14)

Where δi and δbest represents the quantum state of the current bat and the local
best bat respectively and rand is a random number between 0 and 1. If the state of
both the current bat and the best bat are 0 it indicates that this current bat will not
influence the search process during that point of time so there is no need to change
the velocity of that position. If the state of both the current bat and the best bat are
1 it indicates that this current bat will very much influence the search process as it
is close to the current optimal solution so it will change its current velocity using
the information of the current optimal position. And if the state value of the current
bat is just opposite of the state value of the best bat then it indicates that it is far
away from the current optimal solution so it will like to make a random fly near the
current best solution and update its velocity by using the information of the current
optimal solution along with a random value between 0 and 1. Finally the new posi-
tion of the current bat was computed by using the newly generated velocity of it.
 5 Quantum inspired automatic clustering algorithms 101

During the search process the positions of each bat were updated as follows.
(
Pold + εALt if rand > pτ
Pnew = (5:15)
Pold + εðAL − Pold Þ otherwise
t

P
Where, ε 2 ½ − 1, 1 and at the time stamp t, ALt = n1 ni= 1 ALti represents the average
loudness of n bats and pτ represents the pulse rate.
If the pulse rate pτ is low than a probability (rand > pτ) then it indicates that
the bats will fly near to the current superior bat before carrying out a short random
fly there. Otherwise, it is considered that the bats are nearer to the prey and they
will do a random fly around their current position. Here rand is a uniformly distrib-
uted random number which is chosen between [0, 1] during the computation.
The pulse rate ( pτi ) is increased and the loudness ( ALi ) of the bats is decreased
by the following equations.
 
pτti + 1 = pτ0i 1 − e − γt (5:16)
ALit + 1 = α ALti (5:17)

Where, the constant term γ and the preliminary pulse rate pτ0i both were chosen
between 0 and 1. Subsequently, the fitness values of every bats were estimated and
the current Pbest was selected. Here α was considered as a contraction – expansion
and can be defined as follows.
αi − α,
α = αi − t (5:18)
T

Here αi defines the initial value and α, defines the last value of α. The highest num-
ber of iteration is denoted by T . During the execution αi = 1 and α, = 0.5 is set for
increasing the performance of the algorithm [43, 44].
This section describes the process to identify the optimum number of clusters
belong to the dataset of an image by the quantum inspired BAT algorithm.
Initially, the bat population P BAT was initialized and considering the size of
the population was Psize, in which each of the solution was of length BLn and each
of the solution from the population was considered as a solution of the problem.
The initialization was carried out with the normalized pixel intensity value of the
input image, which was chosen randomly. The maximum intensity value of the
input image first passed through the square root function then it was considered as
the length of each particle ðBLn). Another variable Inormal was used to store the nor-
malized value (between 0 and 1) of each pixel intensity of the input image (Line 1).
Now, quantum state QBAT was created for each of the elements of P BAT. And
P BAT was encoded by P BAT α and P BAT β the eq. (5.8) should be satisfied. The
encoding process is shown as follows.
 
Initially, θi, j 0 ≤ θi, j ≤ 2π was created randomly for each of the particle of the
solution of the population, which was used to produce the quantum state namely,
102 Alokananda Dey et al.

QBAT . For a particular tth generation the quantum state QBAT ðtÞ was represented as
follows.
QBAT ðtÞ = fqBATt
1 , qBATt
2 , . . . , qBATt
N }, where qBATt
i ði = 1, 2, . . . , PsizeÞ denotes the
BAT
arbitrary elements of Q ðtÞ and can be represented as

"  BATt   BATt  #

αBATt α  . . .α 
i, 1 i, 2 i, m
qBATt =     , ðst. m = BLnÞ
i
βBATt βBATt  . . . βBATt 
i, 1 i, 2 i, m

st
i, j and βi, j were computed by
here, each value of αst
! !
αBATt
i, j cos θBATt
i, j
= ; where i = f1, 2, . . . , P sizeg and j = f1, 2, . . . , BLng
βBATt
i, j sin θBATt
i, j

β
and the corresponding values of them are preserved in P BATi,α j and P BATi, j re-
spectively. (Line 2).
The frequency ðBF i Þ and velocity (BV i ), Pulse Rate (BRi ) and Loudness (BAi ) were
initialized for every quantum solutions inside the quantum population (Lines 3–4).
Now, in order to generate the active cluster points a selection procedure
was applied on the population and the selection process was guided by the
quantum state QBAT ðtÞ to produce a binary solution BsðtÞ. The binary solution

BsðtÞ = bst1 , bst2 , . . . , bstN was produced by perceiving the state QBAT ðtÞ in the initial
generation t = 1.
Each of the binary solution bcti was generated as follows.
(
βt
1, If P BATij > P BATijαt
bsi, j =
t
, st.i = f1, 2, . . . , P sizeg and j = f1, 2, . . . , BLng
0, Otherwise

The criteria bsti, j = 1 was set to identify the activated cluster points from the popu-
lation P BAT ðtÞ (Figure 5.1) for the tth generation, where i = f1, 2, . . . , Psizeg
and j = f1, 2, . . . , BLng. It was considered that the excited states are influenced to

Cluster Centroid

α1 β1 α2 β2 α3 β3 ... αn βn

𝛿1 𝛿2 𝛿3 ... 𝛿n

Activation Thresholds

Figure 5.1: Cluster Centroid Representation Scheme.

 5 Quantum inspired automatic clustering algorithms 103

make the active cluster centroids from each of the solutions of the population and
are denoted by N Ci for each of the solution of the population. The N Ci number of
activated unique cluster points was stored in BCPi .
Now, from BCPi , the fitness ðF T i Þ of each solution was evaluated with the help
of a cluster validity index namely, DB − Index. [eq. (5.12)]. (Lines 5–10).
Now, the best fitness value O F T was identified from all computed fitness val-
ues (O F T i Þ and saved along with the respective number of cluster points O N C .
(Line 11).
α
Using rotation gate (eq. (5.10)) strategy another set of populations P BATnew
β α β
and P BATnew were created. Update the values of P BAT and P BAT by using
α
their earlier values or by using their recently created values of P BATnew and
β
P BATnew on the basis of the best value of fitness. Now, if the recently created val-
ues were not able to make better fitness values then the current set of values were
rejected and the old set of values were preserved in the population otherwise the
new set of values were accept. Now, it was possible to create a novel quantum pop-
ulation as it produces the diversity in the solution space (Line 12).
The rotation gate updating process was applied which yielded the best fitness
value ðBF T Þ and the respective cluster points were preserved in BCP and Bnc con-
served the total number of cluster points (Line 13).
During each execution of the main loop, the quantum positions and velocities
were updated by the eqs. (5.13) and (5.14) (Line 16).
The rotation gate strategy was responsible for the improvement of the best
quantum solutions (Line 17).
The best solution was selected from the computed fitness of the updated posi-
tions and a local solution was generated by using the eq. (5.15). (Lines 18–20).
Next, the new quantum solutions were estimated and by applying rotation
β
gate strategy each value of the state QBAT ðtÞJ , P BATJα and P BATJ were updated.
(Lines 21–22).
Now the new solution was accepted and the pulse rate of each bat was increased
and the loudness of each bat was decreased by the eqs. (5.16) and (5.17). (Lines 23–26).
The best quantum solutions was determined and saved (Line 27).
In order to achieve the best solution the main loop was continued for a prede-
fined number of times (Lines 14–29).
Finally, the best quantum solution was determined and the optimum number
of cluster was identified along with its associated value of fitness. (Line 30).
The working principle for QIBA is presented as follows.

Algorithm: Quantum Inspired Bat Algorithm for automatic clustering

Input parameter:
Maximum number of Iteration: MI
Size of population: Psize
104 Alokananda Dey et al.

Output parameter:
Optimum cluster number: O N C
Optimum fitness value: O F T

1. A population P BAT having Psize number of particles (bat) was initialized.

2. Quantum state QBAT ðtÞ was created for each of the elements of P BAT. And in
order to encode P BAT the information of the quantum state was used and pro-
duce P BAT α and P BAT β .
3. The frequency ðBF Þi and velocity ðBV i Þ were initialized for every solution of
the population.
4. The loudness ðBAi Þ and the pulse rates ðBRi Þ were also initialized.
5. for i = 1 to Psize do
6. for j = 1 to BLn do
7. N Ci numbers of cluster points from P BATi were created by satisfying the con-
β
dition P BATij > P BATijα and the cluster points were stored in BCPi .
8. Now, the fitness F T i was evaluated from BCPi by using the cluster validity
index (DB − IndexÞ.
9. end for.
10. end for.
11. The best solution in terms of the best fitness value O F T was identified.
12. Now, the quantum state QBAT ðtÞ was updated by using quantum rotation gate
to generate new set of values of P BAT α and P BAT β .
13. The steps 5 to 12 were continued until a predefined condition was satisfied in
order to identify the best fitness value BF T .
14. for I = 1 to MI do
15. for J = 1 to Psize do
16. The velocity and location of every particles in P BATJ were modified by using
the eqs. (5.13) and (5.14).
β
17. Each value of the state QBAT ðtÞJ , P BATJα and P BATJ were updated using the
rotation gate strategy in order to achieve a better solution from the population.
 
18. if Rand ð0, 1Þ > BRij then
19. The best solution was selected after calculating the fitness of all bats from
their updated positions and a native solution was generated with the help of
the eq. (5.15).
20. end if .
21. Now, new solution was evaluated.
β
22. Each value of the state QBAT ðtÞJ , P BATJα and P BATJ were updated by using the
rotation gate strategy in order to achieve a better solution from the population.
23. if ðRand ð0, 1Þ < BAi &BF T new < BF T Best Þ then
24. New solution was accepted.
25. BRi was increased and BAi was decreased by using the eqs. (5.16) and (5.17).
26. end if .
 5 Quantum inspired automatic clustering algorithms 105

27. The best solution was chosen.

28. end for.
29. end for
30. Finally, the algorithm produced and reported the optimal number of cluster
points O N C with its associated fitness value O F T .

5.8 Experiment and result analysis

The aim of this article is to propose two different quantum inspired algorithms com-
bined with the working procedure of two dissimilar meta heuristic algorithms
namely, Genetic algorithm and Bat algorithm, which were accomplished to recog-
nize the optimum number of clusters present within a given image. In order to ac-
complish the experiment, the DB – Index [40] was used as an objective function.
The optimal result was realized for the minimum value obtained from the computa-
tion. The following subsections presents the results obtained from the experiment.

5.8.1 Representation scheme of cluster centroids

The cluster centroids within a given image dataset were used to automatically recog-
nize the optimal number of cluster points present in that dataset. During the execu-
tion of the process, the following equation was used to choose the cluster centroids.
(
1 if βi > αi
δi = (5:19)
0 Otherwise

where, δi ð1 < i < nÞ was the activation threshold. By observing the value of δi , the
cluster centroids were chosen from a single solution. Here αi and βi were the com-
plex coefficients and δi was considered similar to the state of the quantum bit. A
value of 1 for δi represents the activated cluster center implies that the quantum bit
is in excited state otherwise it is in the ground state. The following Figure 5.1 is
showing the Cluster Centroid representation scheme.

5.8.2 Experimental result analysis

In this article two different population based quantum inspired meta-heuristic algo-
rithms, viz., quantum inspired genetic algorithm and quantum inspired bat algo-
rithm are presented. The experiment was performed over four Berkeley images and
two real life images. The performance of any meta-heuristic algorithm is very much
dependent upon the settings of the parameter i.e., an appropriate selection of
106 Alokananda Dey et al.

parameters can influence the convergence speed of the meta-heuristic algorithm to

a great level.
The parameters for QIGA were chosen as Number of generations was consid-
ered for the range of values 100 to 500; Population size was taken from 25 to 50.
Mutation probability was chosen as 0.01. And for CGA the number of generations
was considered from 100 to 500. Population size was taken from 25 to 50. Crossover
and Mutation probability was chosen as 0.9 and 0.01 respectively. The parameters
for QIBA and CBA were chosen as number of iterations from 100 to 300, Population
size was taken from 25 to 50, γ = α = 0.5, initially pulse rate could be denoted as
ri0 2 ½0, 1, the wavelength range was ½λmin , λmax  = ½0, 1, and the range of fre-
quency was ½Fmin , Fmax  = ½0, 1 and the loudness L0 = 0.5.
During the Experiment it was seen that the quantum inspired methods can pro-
duce a good result without having a large population or large number of iteration
than their corresponding classical methods. Due to the probability amplitude ðα, βÞ
of qubits a single solution may generate several solutions which may lead to in-
crease the size of a small population. So in case of quantum inspired methods the
search can start with a small population. For each of the specific settings of param-
eters each procedure was run for 30 times. The mean value of the fitness was com-
puted from several runs. The computed valued from several runs was presented in
the following tables.
The optimal number of clusters along with their associated optimal fitness
value for all the implemented algorithms like QIGA, CGA, QIBA and CBA were pre-
sented in Table 5.1. From Table 5.1 it can be stated that the computed fitness value
from the method QIBA was providing the best fitness value (minimum value)
among all other implemented methods for every image datasets and it was also
seen that the better fitness values were achieved from quantum inspired methods
than their classical counterparts. In Table 5.2 the optimal execution time for all the

Table 5.1: Optimal value of the cluster number (O N C ), Optimal fitness value (O F T ) for QIGA,
CGA, QIBA and CBA.

Dataset QIGA CGA QIBA CBA

O NC O FT O NC O FT O NC O FT O NC O FT

#  .  .  .  .

#  .  .  .  .

#  .  .  .  .

#  .  .  .  .

couple  .  .  .  .

clown  .  .  .  .

 5 Quantum inspired automatic clustering algorithms 107

Table 5.2: Optimal execution time (O T E ) in second for QIGA, CGA, QIBA and CBA.

Dataset QIGA ðO T E Þ CGA ðO T E Þ QIBA ðO T E Þ CBA ðO T E Þ

# . . . .

# . . . .

# . . . .

# . . . .

couple . . . .

clown . . . .

implemented methods was presented which shows that the quantum inspired
methods were taking less time than their classical counterparts and the processing
time taken by QIBA was the minimal among other methods. Table 5.3 provides the
computed mean value, standard deviation and standard error from QIGA and CGA.
Here it is seen that QIGA is providing more promising results than CGA in terms of
mean, standard deviation and standard error. Similarly QIBA is providing more
promising results than CBA which is presented in Table 5.4. From these two tables
it is seen that QIBA has the minimum error among all other. The statistical superior-
ity test named unpaired t- test was performed between QIGA and CGA and between
QIBA and CBA which is presented in Table 5.5. From Table 5.5 it is seen that out of
six results three results are extremely significant and remaining three results are
very significant for the comparison between QIGA and CGA. And for QIBA and CBA,
out of six results four results are extremely significant, one result is very significant
and one result is significant. This analysis indicates that the quantum inspired

Table 5.3: Mean value of fitness ðF μ Þ, Standard deviation of fitness ðF σ Þ and Standard error of
fitness (S E ) for QIGA, CGA.

Dataset QIGA CGA

Fμ Fσ SE Fμ Fσ SE

# . . . . . .

# . . . . . .

# . . . . . .

# . . . . . .

couple . . . . . .

clown . . . . . .

108 Alokananda Dey et al.

Table 5.4: Mean value of fitness ðF μ Þ, Standard deviation of fitness ðF σ Þ and Standard error of
fitness (S E ) for QIBA and CBA.

Dataset QIBA CBA

Fμ Fσ SE Fμ Fσ SE

# . . . . . .

# . . . . . .

# . . . . . .

# . . . . . .

couple . . . . . .

clown . . . . . .

Table 5.5: Result of unpaired t − test (P − value) between QIGA and CGA and between QIBA and CBA.

Dataset QIGA & CGA QIBA & CBA

P − value Significant Level P − value Significant Level

# <.  <. 

# .  . 

# .  . 

# <.  <. 

couple .  <. 

clown <.  <. 

Level of Significance:  – Extremely Significant,  – Very Significant,  – Significant

methods are much more superior to their classical counterparts. Finally, the result
of Friedman test for QIGA, CGA, QIBA and CBA is presented in Table 5.6. As per the
ranking QIBA is being the most preferable method as it has achieved the first rank.

5.8.3 Simulation of work

Python environment was used to carry out the experimental process. The experi-
ment was done on four Berkeley images, two of dimensions 80 × 120 and other two
of dimensions 120 × 80 and two real life images of dimensions 512 × 512. The devel-
opment environment was consisting of Windows 7 with the configuration facility of
DELL Intel(R) Core(TM) i3, 2.00 GHz and 4.00 GB RAM.
 5 Quantum inspired automatic clustering algorithms 109

The efficiency of the suggested methods over their classical counterparts was
proved depending upon computed mean fitness value, standard deviation, stan-
dard error and smallest computation time. Moreover, the competence of the quan-
tum inspired algorithms over their classical algorithms was justified by performing
the t – test (statistical superiority test). The optimal number of cluster (O N C Þ and
its associated optimal fitness value (O F T ) for all the algorithms (QIGA, CGA,
QIBA and CBA) were reported in Table 5.1. The optimal execution time (O T E )
in second for all the algorithms (QIGA, CGA, QIBA and CBA) were shown in
Table 5.2. The mean fitness value ðF μ Þ, standard deviation (F σ ) and standard error
(S E ) of the algorithms QIGA and CGA and were demonstrated in Table 5.3 and the
same things for the algorithms QIBA and CBA were demonstrated in Table 5.4.
Finally, the result of t − test [37] was presented in Table 5.5 which was performed
with 95% confidence level. During the test, the P − value is checked. If it is less
than 0.05 then null hypotheses is considered for rejection against the alternative
hypothesis. Finally, Friedman test [38, 39] was performed among the participating
procedure and was enlisted in Table 5.6. The obtained rank from the proposed al-
gorithms QIGA, CGA, QIBA and CBA are 2.66, 3.91, 1.08 and 2.33 respectively.
According to the rank value, QIBA can be considered as the superior algorithm as
it has the minimum rank value.

Table 5.6: Result of Friedman test for QIGA, CGA, QIBA and CBA. The rank for each
procedure is shown within the parenthesis.

Dataset QIGA CGA QIBA CBA

# . () . () . () . ()

# . (.) . () . () . (.)

# . () . () . () . ()

# . (.) . () . (.) . ()

couple . (.) . () . () . (.)

clown . (.) . (.) . () . ()

Average Rank . . . .

5.8.4 Dataset used

The following images (Figure 5.2) were used during the experiment and during the
accomplishment of the process the normalized values between (0,1) were consid-
ered for the images.
110 Alokananda Dey et al.

(a) (b) (c) (d)

(e) (f)

Figure 5.2: Test Images: (a) #86000 of size 80 × 120, (b) #92059 of size 80 × 120, (c) #94079 of
size 120 × 80, (d) #97017 of size 120 × 80, (e) couple of size 512 × 512, (f) clown of size 512 × 512.

5.9 Conclusion
This article is envisaged an automatic clustering technique to recognize the optimal
number of cluster points present in an image dataset, which shows a new way to
combine the features of quantum computing with two different meta-heuristic algo-
rithms, viz., Genetic algorithm and Bat algorithm. These proposed techniques were
found superior over their classical counterparts as they out performed over corre-
sponding classical counterparts for automatically identifying the optimal number of
cluster points present in an image dataset within a significantly small time frame.
Also the effectiveness of the quantum inspired algorithms were judged and presented
depending upon the computed value of optimal fitness, mean fitness value, standard
deviation of the fitness, standard error of the fitness and more over the computed
p − value from the statistical superiority test t − test. Finally, Friedman test were per-
formed to identify the superior technique between the quantum inspired versions
with their classical version. Moreover, it was seen that the quantum inspired Bat
 5 Quantum inspired automatic clustering algorithms 111

algorithm was the best performing procedure among the others. All the computed re-
sults were presented in the previous tables in order to proof the effectiveness of the
quantum inspired methods over their classical counterparts as well as prove the su-
periority of QIBA. So far, only gray level images were used for experimental purpose
and the proposed techniques were developed to satisfy only one objective at a time.
Currently authors are involved to explore the realm of quantum computing to
resolve the concerned multi objective optimization problems effectively and effi-
ciently within a smaller time frame and also trying to take care of the true colour
images for this direction of the research.

References
[1] Jain, A., & Dubes, R. Algorithms for Clustering Data. Prentice Hall, 1988.
[2] Jain, A. K., Murty, M. N., & Flynn, P. J. Data clustering: A review. ACM Computing Surveys.
1999, 31(3),264–323.
[3] Yadav, J. & Sharma, M. A Review of K-mean Algorithm. International Journal of Engineering
Trends and Technology (IJETT). July 2013, 4(7).
[4] Saket, S., Pandya, J, S. An Overview of Partitioning Algorithms in Clustering Techniques.
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET).
June 2016, 5(6).
[5] Pestunov, I., Rylov, S. A., & Berikov, V. Hierarchical clustering algorithms for segmentation of
multispectral images. Optoelectronics, Instrumentation and Data Processing. 2015, 51,
329–338.
[6 Wehrens, R., Buydens, L., Fraley, C., & Raftery, A. Model-Based Clustering for Image
Segmentation and Large Datasets via Sampling. Journal of Classification. 2004, 21(2),
231–253.
[7] Chen, M. & Ludwig, S. A. Color Image Segmentation Using Fuzzy C-Regression Model.
Hindawi Advances in Fuzzy Systems. 2017, 4582948, 15.
[8] Ye, Q., Gao, W, & Zeng, W. Color image segmentation using density-based clustering. 2003
International Conference on Multimedia and Expo. ICME ‘03. Proceedings. Baltimore, MD,
USA, 2003, 03TH8698, II–401.
[9] Holland, J. Adaptation in Natural and Artificial Systems: An Introductory Analysis with
Applications to Biology, Control and Artificial Intelligence. The University of Michigan Press,
August 1975, 53.
[10] Hui-liana, F. & and Xian-lib, L. Discrete particle swarm optimization for TSP based on
neighborhood. Application Research of Computers. 2011, 2, 030.
[11] Storn, R. & Price, K. Differential evolution – a simple and efficient heuristic for global
optimization over continuous spaces. Technical Report TR-95-012, ICSI, 1995.
[12] Passino, K. M. Bacterial foraging optimization. Innovations and Devel-opments of Swarm
Intelligence Applications. 2012, 219–233.
[13] Odili, J. B., Deng, H., Yang, C., Sun, Y., et al. Application of Ant Colony Optimization to
Solving the Traveling Salesman’s Problem. Science Journal of Electrical & Electronic
Engineering. 2013, 2013.
[14] Yang, X.-S. A new metaheuristic bat-inspired algorithm. Studies in Computational
Intelligence. 2010, 284, 65–74.
112 Alokananda Dey et al.

[15] Alihodzic, A. & Tuba, M. Improved hybridized bat algorithm for global numerical
optimization. Proceedings of the 16th IEEE International Conference on Computer Modelling
and Simulation (UKSim-AMSS ‘14). March 2014, 57–62.
[16] Liuabcd, Y., Wud, X., Shenb, Y. Automatic clustering using genetic algorithms. Applied
Mathematics and Computation, 15 October 2011, 218 (4), 1267–1279.
[17] Raposo, C., Antunes, C. H., Barreto, J. P. Automatic Clustering Using a Genetic Algorithm with
New Solution Encoding and Operators. Automatic Clustering Using a Genetic Algorithm with
New Solution Encoding and Operators. Murgante, B. et al. (Eds.). Computational Science and
Its Applications – ICCSA 2014. Lecture Notes in Computer Science, Springer, 2014, 8580.
[18] Garai, G., Chaudhuri, B. B. A novel genetic algorithm for automatic clustering. Journal Pattern
Recognition Letters Archive. Elsevier Science Inc. New York, NY, USA, 19 January 2004, 25(2),
173–187.
[19] Das, S., Ajith, A., & Konar, A. Automatic Clustering Using an Improved Differential Evolution
Algorithm. Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions,
2008, 38, 218–237. 10.1109/TSMCA.2007.909595
[20] Abubaker, A., Baharum, A., Alrefaei, M. Automatic Clustering Using Multi-objective Particle
Swarm and Simulated Annealing. 1 July 2015, 10(7), e0130995. doi: 10.1371/journal.
pone.0130995
[21] McMohan, D. Quantum Computing Explained. John Wiley & Sons, Inc., Hoboken, New Jersey,
2008.
[22] Dey, S., Bhattacharyya, S., & Maullik, U. Quantum Behaved Swarm Intelligent Techniques for
Image Analysis: A Detailed Survey. Handbook of Research on Swarm Intelligence in
Engineering. 2015.
[23] Dey, S., Saha, I., Bhattacharyya, S., & Maulik, U. Multi-level thresholding using quantum
inspired meta-heuristics. Knowledge-Based Systems. 2014, 67, 373–400.
[24] Dey, S., Bhattacharyya, S., & Maullik, U. Optimum Gray Level Image Thresholding using a
Quantum Inspired Genetic Algorithm. Advanced Research on Hybrid Intelligent Techniques
and Applications. 2015, 201.
[25] Dey, S., Bhattacharyya, S., & Maulik, U. Quantum inspired genetic algorithm and particle
swarm optimization using chaotic map model based interference for gray level image
thresholding. Swarm and Evolutionary Computation. 2014, 15, 38–57.
[26] Dey, S., Bhattacharyya, S., & Maullik, U. Efficient quantum inspired meta-heuristics for multi-
level true color image thresholding. Applied Soft Computing. 2017, 56, 472–513.
[27] Dey, S., Bhattacharyya, S., & Maullik, U. New quantum inspired meta-heuristic techniques for
multi-level color image thresholding. Applied Soft Computing. 2016, 46, 677–702.
[28] Dey, S., Bhattacharyya, S., & Maullik, U. Quantum inspired automatic clustering for multi-
level image thresholding. Proceedings of International Conference On Computational
Intelligence and Communication Networks (ICCICN 2014), RCCIIT, Kolkata, India. 2014,
247–251.
[29] Dey, S., Bhattacharyya, S., & Maullik, U. Quantum inspired automatic clustering for multi-level
image thresholding. Proceedings of International Conference On Computational Intelligence
and Communication Networks (ICCICN 2014), RCCIIT, Kolkata, India. 2014, 242–246
[30] Dey, S., Bhattacharyya, S., Snasel, V., Dey, A., & Sarkar, S. PSO and DE based novel quantum
inspired automatic clustering techniques. 2017, 285–290. 10.1109/ICRCICN.2017.8234522
[31] Dey, A., Dey, S., Bhattacharyya, S., Snasel, V., & Hassanien, A. E. Simulated Annealing Based
Quantum Inspired Automatic Clustering Technique. Hassanien, A., Tolba, M., Elhoseny, M.,
Mostafa, M. (Eds.). The International Conference on Advanced Machine Learning
Technologies and Applications (AMLTA2018). Advances in Intelligent Systems and
Computing. Springer, Cham, 2018, 723.
 5 Quantum inspired automatic clustering algorithms 113

[32] Dey, S., Bhattacharyya, S., & Maullik, U. Quantum-Inspired Automatic Clustering Technique
Using Ant Colony Optimization Algorithm. Quantum-Inspired Intelligent Systems for
Multimedia Data Analysis. 2018. 10.4018/978-1-5225-5219-2.ch002
[33] Narayanan, A. & Moore, M. Quantum-inspired genetic algorithms. Proc. IEEE Evolutionary
Computation. 1996, 61–66.
[34] Yang, Y., Kuo, S., Lin, F., Liu, I., & Chou, Y. Improved Quantum-Inspired Tabu Search
Algorithm for Solving Function Optimization Problem. 2013 IEEE International Conference on
Systems, Man, and Cybernetics, Manchester. 2013, 823–828.
[35] Han, K.-H., & Kim, J.-H. Quantum-inspired evolutionary algorithm for a class of combinatorial
optimization. IEEE Transactions on Evolutionary Computation. December 2002, 6(6),580–593.
[36] Alam, T., Gandhi, T., & Nitin, N. Quantum Genetic Algorithm with Rotation Angle Refinement
for Dependent Task Scheduling on Distributed Systems. 2017. 10.1109/IC3.2017.8284295
[37] Flury, B. A First Course in Multivariate Statistics. Springer-Verlag, Berlin, Germany, 1997.
[38] Friedman, M. The use of ranks to avoid the assumption of normality implicit in the analysis of
variance. Journal of the American Statistical Association. 1937, 32, 674–701.
[39] Friedman, M. A comparison of alternative tests of significance for the problem of m rankings.
Annals of Mathematical Statistics. 1940, 11, 86–92.
[40] Maulik, U., & Bandyopadhyay, S. Performance evaluation of some clustering algorithms and
validity indices. IEEE PAMI. 2002, 24, 1650–1654.
[41] Aranha, C. Portfolio management by genetic algorithm with error modeling. Information
Sciences. 2007, 459–465.
[42] Paulinas, M., & Usinskas, A. A survey of genetic algorithms applications for image
enhancement and segmentation. Information Technology and Control. 2007, 36, 278–284.
[43] Dey, A., Bhattacharyya, S., Dey, S., Platos, J., & Snasel, V. Quantum-Inspired Bat
Optimization Algorithm for Automatic Clustering of Grayscale Images. Recent Trends in
Signal and Image Processing. 2019, 89–101. 10.1007/978-981-13-6783-0_9
[44] Sun, J., Fang, W., Wu, X., Palade, V., & Xu, W. Quantum-Behaved Particle Swarm
Optimization: Analysis of Individual Particle Behavior and Parameter Selection. Evolutionary
Computation. September 2012, 20(3),349–393.
Siddhartha Bhattacharyya
6 Conclusion
Machine learning is a branch of Artificial intelligence, which tries to develop techni-
ques that have ability to learn from experiences. These experiences are articulated as
data generated from underlying models. The goal of any machine learning technique
is to abstract the underlying model reliably for specific tasks like prediction and deci-
sion making. Enticed by the versatility of quantum mechanics, scientists have em-
barked on evolving machine learning algorithms based on the principles of quantum
mechanics. As a result, different incarnations of the classical machine learning algo-
rithms have come up enveloped within the quantum computing mechanisms. This
book intends to throw a light on some of the recent advancements in this direction.
The book first introduces to the readers the basics of quantum machine learn-
ing with reference to the well-known quantum algorithms in the form of Grover’s
search algorithm [1], quantum reinforcement learning [2–6] and quantum annealing
[7, 8]. A bird’s eye view on the evolution of quantum neural networks [9–17] is also
presented. The information processing capabilities of quantum algorithms depends
on the quantum implementation of machine learning algorithms involving nonlin-
ear functions operating on information represented topographically, as is common
in neural cortex. Several approaches to quantum computation for machine learning
by means of topographic representation have been proposed. These approaches
show ways to construct unitary operators, implementable on quantum computers,
that implement arbitrary (including nonlinear) functions via computational maps.
Optimization problems are ubiquitous in machine learning and other domains
like engineering and a number of specialized techniques for efficiently solving spe-
cific problems have been developed. A novel quantum optimization algorithm
based on Adiabatic Quantum Computation through local evolution and inspired by
the ideas from Grover’s Search algorithm [1] can be evolved for finding minimum
and employed in solving unconstrained optimization problems.
As a transition from the classical paradigm to quantum paradigm, several re-
searchers are involved in developing quantum inspired algorithms to bridge the gap
between the two paradigms. Notable among them include the quantum inspired
metaheuristic algorithms capable of delivering outputs in real time. Different incarna-
tions of the quantum inspired metaheuristic algorithms have come up, thanks to the
relentless efforts on the part of the scientists. These algorithms are efficient enough
to obtain stable outputs by emulating the quantum mechanical principles on classi-
cal computers. In most of the cases, it is found that the quantum inspired versions of
the metaheuristic algorithms outperform their classical counterparts.
Although most of the quantum machine learning algorithms are based on the
bilevel qubits, of late, much work is being carried out using multilevel quantum
systems (using qutrits or qudits), which promise more efficiency.

https://doi.org/10.1515/9783110670707-006
116 Siddhartha Bhattacharyya

References
[1] Lavor, C., Manssur, L.R.U., & Portugal, R. Grover’s Algorithm: Quantum Database Search.
arXiv:quant-ph/0301079.
[2] Beom, H. R., & Cho, H. S. A sensor-based navigation for a mobile robot using fuzzy logic
and reinforcement learning. IEEE Transactions on Systems, Man, and Cybernetics. 1995, 25
(3), 464–477.
[3] Connolly, C. I. Harmonic functions and collision probabilities. The International Journal
of Robotics Research. 1997,16(4), 497–507.
[4] Smart, W. D., & Kaelbling, L. P. Effective reinforcement learning for mobile robots.
Proceedings – IEEE International Conference on Robotics and Automation. 2002, 3404–3410.
[5] Kondo, T., & and Ito, K. A reinforcement learning with evolutionary state recruitment strategy
for autonomous mobile robots control. Robotics and Autonomous Systems. 2004, 46(2),
111–124.
[6] Tzafestas, S. G., & Rigatos, G. G. Fuzzy reinforcement learning control for compliance tasks
of robotic manipulators. IEEE Transactions on Systems, Man, and Cybernetics. 2002, 32(1),
107–113.
[7] Matsuda, Y., Nishimori, H. & Katzgraber, H. G. Ground-state statistics from annealing
algorithms: Quantum vs classical approaches. 2009. arXiv:0808.0365.
[8] Santoro, G. E., Martonak, R., Tosatti, E. & Car, R. Theory of quantum annealing of an Ising
spin glass. Science. 2002, 295(5564),2427–2430.
[9] da Silva, A. J., Ludermir, T. B., & de Oliveira, W. R. Quantum perceptron over a field
and neural network architecture selection in a quantum computer. Neural Networks.
2016, 76, 55–64.
[10] Panella, M., & Martinelli, G. Neural networks with quantum architecture and quantum
learning. International Journal of Circuit Theory and Applications. 2011, 39, 61–77.
[11] Kak, S. On quantum neural computing. Advances in Imaging and Electron Physics. 1995,
94, 259.
[12] Chrisley, R. Quantum Learning. New directions in cognitive science: Proceedings of the
international symposium, Saariselka, 4–9 August 1995, Lapland, Finland. Pylkkänen, P. &
Pylkkö, P. (Eds.). Finnish Association of Artificial Intelligence, Helsinki, 1995, 77–89.
[13] Narayanan, A., & Menneer, T. Quantum artificial neural network architectures and
components. Information Sciences. 2000, 128, 231–255.
[14] Schuld, M., Sinayskiy, I., & Petruccione, F. Simulating a perceptron on a quantum computer.
2014. ArXiv:1412.3635.
[15] Purushothaman, G., & Karayiannis, N. Quantum Neural Networks (QNN’s): Inherently Fuzzy
Feedforward Neural Networks. IEEE Transactions on Neural Networks. 1997, 8(3), 679.
[16] Behrman, E. C., Steck, J. E., Kumar, P., & Walsh, K. A. Quantum Algorithm design using
dynamic learning. Quantum Information and Computation. 2008, 8(1&2), 12–29.
[17] Ventura, D., & Martinez, T. A quantum associative memory based on Grover’s algorithm.
Proceedings of the International Conference on Artificial Neural Networks and Genetics
Algorithms. 1999, 22–27.
Index
adiabatic 69, 73, 81, 82, 83 Fuzzy clustering 90
Amplitude 12, 13, 14, 18, 22, 23, 24, 27, 28, 29, fuzzy set 28, 36
32, 35, 71, 72, 77, 95, 98, 106
analog computation 12 Gates 22, 29, 31, 33, 34, 35, 36, 70, 95
Analysis 1, 61, 84, 90, 91, 105–106 Generations 12, 92, 97, 98, 106
ancilla 17 Genetic Algorithm 90, 91–92, 98, 105, 110
Automatic clustering 90, 91, 110 graph kernel 14, 16
Grover iteration 22
Bat Algorithm 90, 91, 92–94, 100, 103, 105, 110 Grover’s algorithm 22
Berkeley image 105, 108 Grover’s search 69, 74, 78, 81, 83
brain-inspired computation 11, 12, 35
Hadamard gate 71, 83
CCNOT 28, 31, See also Toffoli gate Hamiltonian 73, 76, 81, 82, 83
Chromosomes 91, 92, 97, 98 Heterogeneous 89
Classical counterparts 69, 74, 76, 77, 78, 81, HHL algorithm 74, 76
90, 91, 106, 107, 108, 109, 110, 111, 115 Hilbert space 13, 15, 69, 70, 72, 77
Cluster centroids 82, 103, 105 Hilbert-Schmidt linear operator 14
Cluster Validity Index 91, 96–106 Homogeneous 89
CNOT 22
coarse coding 14 Image dataset 91, 97, 105, 106, 110
Coherence 95 Intensity 98, 101
computational map 11, 28, 31, 35, See also
topographic map K-Means 78
Convergence 49, 62, 90, 92, 100, 106
crisp set 22, 27, 28, 35 least square SVM 78, 79
Crossover 91, 106 Loudness 92, 93, 101, 103
CSWAP 34, See also Fredkin gate
Machine learning 39
DB index 91 Mean 91, 106, 107
Decoherence 95 Meta-heuristic 90, 91, 92, 110
Deep Learning 2, 12, 68, 69, 82, 84 Methods 49, 50, 52, 53, 90, 91, 106, 109, 111
demultiplexer 34 Mutation 91, 98, 106
Dirac delta 14, 15, 16
Dirac notation 69, 70 Nature inspired 90, 91, 92
Diversity 92, 100 neural network, See also quantum neural
network, spiking neural network
Echolocation 92 Normalized 19, 20, 21, 35, 75, 79, 99, 101
Entangled 22, 29, 32, 47, 72–73, 98
Evolutionary 91 Optimal number 105, 106, 109, 110

field computation 13 Pauli 71

Fitness 90, 91, 93, 96, 97, 98, 101, 103, Population 14, 90, 91, 92, 93, 97, 98, 100, 101,
106, 110 102, 103, 105, 106
Fredkin gate 34, See also CSWAP precision 15, 27
Frequency 92 preimage 18
Friedman test 91, 108, 109, 110 preimage multiplicity 20

https://doi.org/10.1515/9783110670707-007
118 Index

Prey 92, 93, 101 Randomly 43, 58, 59, 61, 77, 92, 93, 101
Principal component analysis 77 reinforcement learning 2, 3, 115
Probability 13, 42, 43, 48, 92, 93, 101 Results 41, 44, 82, 83, 91, 107
Pulses 92, 93, 94 Rotation angle 95, 98
Rotation gate 95, 97, 100, 103
Quantum algorithms 74
quantum annealing 78 Searching 47, 52, 91
Quantum Approximate Optimization Selection 40, 41, 43, 91, 97, 102
Algorithm 45 set 22, 27, 28, 36, See also crisp set,
Quantum Boltzmann Machines 2, 84 fuzzy set
quantum computation 28 Shor algorithm 4
Quantum computing 3, 4, 5, 6, 11, 14, spiking neural network 12
41, 45, 54, 68, 69–74, 90, 94–96, 100, Standard deviation 91, 107, 109, 110
110, 115 Standard error 91, 109, 110
Quantum Deep Learning 82 Statistical 68
Quantum Generative Adversarial Networks 2 superposition 11, 14, 15, 17, 19, 28
Quantum Gradient Descent Algorithm 2 supervised learning 68
Quantum inspired 90, 91, 97, 98, 100, 101, Support Vector Machines 2
106, 107, 110, 115
quantum machine learning 68, 69, 74, 77, 78, Table 13, 47
79, 83, 84 Time 41, 43, 45, 58
Quantum matrix inversion 77 Toffoli gate 28, See also CCNOT
quantum multiclass support vector machine 79 topographic basis map 14, 35
quantum neural network 12, 35 topographic map 11, 14, 28, 31, 35, See also
Quantum NOT gate 98 topographic basis map, topographic qubit
quantum parallelism 11, 15, See also map, See also computational map
superposition topographic qubit map 27, 35
Quantum random access memory 67, 75 topographic representation 11, 13, See also
quantum register 71 topographic map
Quantum state 6, 14, 15, 27, 69, 73, 75, 76, 77,
81, 83, 97, 100, 104 unsupervised learning 1, 68, 81
Quantum Support Vector Machines 2, 74,
77, 83 vector space 69
Qubit 3, 6, 14, 27, 28, 29, 30, 31, 32, 34, 35,
36, 53, 60, 69, 70, 71, 72, 74, 76, 77, 83, wave function 13
94, 95, 106
De Gruyter Frontiers in Computational Intelligence
Already published in the series

Volume 5: Machine Learning Applications

S. Bhattacharyya, R. Das, S. Nandy (Eds.)
ISBN 978-3-11-060853-3, e-ISBN (PDF) 978-3-11-061098-7,
e-ISBN (EPUB) 978-3-11-060866-3

Volume 4: Intelligent Decision Support Systems

S. Bhattacharyya, S. Borra, M. Bouhlel, N. Dey (Eds.)
ISBN 978-3-11-061868-6, e-ISBN (PDF) 978-3-11-062110-5,
e-ISBN (EPUB) 978-3-11-061871-6

Volume 3: Big Data Security

I. Banerjee, S. Bhattacharyya, S. Gupta (Eds.)
ISBN 978-3-11-060588-4, e-ISBN (PDF) 978-3-11-060605-8,
e-ISBN (EPUB) 978-3-11-060596-9

Volume 2: Intelligent Multimedia Data Analysis

S. Bhattacharyya, I. Pan, A. Das, S. Gupta (Eds.)
ISBN 978-3-11-055031-3, e-ISBN (PDF) 978-3-11-055207-2,
e-ISBN (EPUB) 978-3-11-055033-7

Volume 1: Machine Learning for Big Data Analysis

S. Bhattacharyya, H. Baumik, A. Mukherjee, S. De (Eds.)
ISBN 978-3-11-055032-0, e-ISBN (PDF) 978-3-11-055143-3,
e-ISBN (EPUB) 978-3-11-055077-1

www.degruyter.com