Quantum Machine Learning

Quantum Machine Learning
State of the Art and Future Directions

Abstract
Since the training of modern machine learning systems is a computation intensive endeav-
our, the idea of using quantum computing in the machine learning pipeline attracts increas-
ing interest. Indeed, since several quantum machine learning algorithms have already been
implemented on present day noisy intermediate-scale quantum devices, experts expect that
machine learning on reliable, large scale quantum computers will soon become reality. How-
ever, next to potential benefits due to quantum speedup, quantum machine learning may also
entail risks regarding reliability, trustworthiness, safety, and security. In preparation of an in-
depth assessment of these issues, this report therefore surveys the current state of the art in
quantum machine learning. It reviews quantum inspired classical algorithms for classical data,
genuine quantum algorithms for classical data, classical algorithms for quantum data, and
quantum algorithms for quantum data. The particular focus of this review is on methods that
are practically feasible on existing quantum computers.
Authors
Christian Bauckhage, Fraunhofer IAIS
Rainer Bye, Capgemini
Ahmed Iftikhar, Capgemini
Christian Knopf, Capgemini
Merima Mustafic, Capgemini
Nico Piatkowski, Fraunhofer IAIS
Rafet Sifa, Fraunhofer IAIS
Roland Stahl, Capgemini
Eldar Sultanow, Capgemini
Publisher
Federal Office for Information Security
https://www.bsi.bund.de
Federal Office for Information Security

CONTENTS
Contents
1 Introduction 3
2 Quantum Computing in a Nutshell 7

2.1 Adiabatic Quantum Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Quantum Gate Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Further Quantum Computing Paradigms . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.1 Topological Quantum Computing . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 Variational Quantum Computing . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Technical State of the Art and Limitations . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Machine Learning in a Nutshell 22

3.1 The Machine Learning Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Lazy vs. Eager Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 Sources of Uncertainty in Machine Learning . . . . . . . . . . . . . . . . . . . . . . 27
4 Quantum Machine Learning 29

4.1 Quantum Inspired Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.1 Tang’s Quantum Inspired Algorithm for Recommendation Systems . . . . 31
4.1.2 Tensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.1.3 Digital Annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1.4 Quantum Inspired Data Clustering . . . . . . . . . . . . . . . . . . . . . . . 35
4.1.5 Quantum Inspired Gravitational Search . . . . . . . . . . . . . . . . . . . . 36
4.2 Machine Learning for Quantum Computing . . . . . . . . . . . . . . . . . . . . . . 36
4.2.1 Quantum Circuit Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2.2 Quantum State Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2.3 Quantum Noise Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Quantum Enhanced Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3.1 Quantum Algorithms for Linear Algebra . . . . . . . . . . . . . . . . . . . . 41
4.3.2 Quantum Algorithms for Regression . . . . . . . . . . . . . . . . . . . . . . 43
4.3.3 Quantum Algorithms for Clustering . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.4 Quantum Algorithms for Nearest Neighbor Search . . . . . . . . . . . . . . 47
4.3.5 Quantum Algorithms for Classification . . . . . . . . . . . . . . . . . . . . . 48
4.3.6 Quantum Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.7 Quantum Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.8 Quantum Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3.9 Federated and Distributed Quantum Machine Learning . . . . . . . . . . . 59
4.3.10 Variational Quantum Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4 Quantum Machine Learning for Quantum Data . . . . . . . . . . . . . . . . . . . . 64
Federal Office for Information Security 1

CONTENTS
5 Characteristics of QML Methods 66

5.1 Hardware Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.1.1 Quantum Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.1.2 Qubit Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.1.3 Qubit Count . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.1.4 Quantum Circuit Depth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Algorithmic Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2.1 Data Pre- and Post-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2.2 Statistical Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.2.3 Noise Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6 Summary and Outlook 78
Glossary 87
Literature 97
2 Federal Office for Information Security

1 INTRODUCTION
1 Introduction
One of the most striking and likely most disruptive technological developments in the
early 21st century is the maturing of Artificial Intelligence (AI). After decades of worldwide
research which began in the 1950s, (weakly) intelligent technical systems have now become
reality and are increasingly deployed in practice. Prominent examples of recent accomplish-
ments include cognitive systems for robust text- and speech recognition and generation [1, 2,
3], image- and video understanding and synthesis [4, 5, 6], game play [7, 8], recommendation
[9, 10], (medical) diagnostics [11, 12], planning and decision making [13, 14, 15], and convincing
conversational agents [16, 17, 18].
Achievements like these are mainly due to progress in Machine Learning (ML), a branch
of AI concerned with adjusting the parameters of a software system in a training process such
that it can develop cognitive capabilities. In other words, while there were many innovations
in many branches of AI, the recent performance boost is due to systems which learn an in-
tended input-output behavior from analyzing task specific example data [19, 20, 21, 22].
In and off itself, machine learning is not a new idea; it was initially conceptualized by
Turing and had its first boom period in the 1980s after algorithms for training artificial neu-
ral networks became available. Its accelerated improvement over the past decade can be at-
tributed to four technological trends:
1. exponentially growing amounts of (training) data available on the Internet, mainly due
to social media platforms and mobile communication devices, but in no small parts also
due to the digitization of industry and the life- and natural sciences
2. increasing capabilities and decreasing costs of high performance computing hardware,
most notably the availability of GPU computing for serious applications
3. ever more rapidly developed and publicly disseminated open source software and
frameworks for the design, training, and deployment of machine learning systems
4. last but not least, scientific progress in neurocomputing, most notably in the areas of
deep neural networks and deep learning [23]
Indeed, a conclusion to be drawn from the past decade of machine learning research is
that these trends had to come together to enable results such as mentioned above. Although
there exist solutions for very capable learning systems of comparatively moderate sizes [24,
25], the capability of a learning system to exhibit human-like performance seems to be predi-
cated on the fact that its size far exceeds traditional system sizes. In other words, the underly-
ing mathematical models seem to require more flexibility and thus more adjustable parame-
ters than have been considered historically.
For instance, the adjustable parameters of an artificial neural network are its synaptic
weights and modern deep networks often come with hundreds of billions of such weights (see

1 INTRODUCTION
e.g. [18]). In order for such flexible models to reliably learn a desired cognitive skill, classical re-
sults in learning theory suggest that they must be trained with large amounts of representative
training data [26]. Indeed, state of the art systems are commonly trained with up to several
billion data points (see e.g. [18]). Given numbers like these, it is clear that the training of mod-
ern learning systems is a resource intensive endeavor that easily translates to several hundred
GPU years of computation time (see e.g. [18]). Modern machine learning has therefore reached
a point where practical feasibility and success are conditioned on access to high performance
computing hardware.
Against this backdrop, it is not surprising that machine learning researchers are begin-
ning to look at another technology that has recently seen substantial progress, namely quan-
tum information processing (QIP) or quantum computing (QC).
First ideas for quantum information processing were published in the 1970s [27, 28, 29,
30] and the theoretical foundations of quantum computing date back to the 1980s when Rus-
sian and American physicists first conceptualized quantum computers [31, 32, 33, 34]. The ap-
peal of these early contributions was that they showed that quantum bits can carry more in-
formation than classical bits so that quantum computing could be expected to solve certain
difficult problems much faster than classically possible.
It then took about another decade until first quantum algorithms were conceived which,
when running on a quantum computer, would exhibit these hypothesized advantages [35, 36,
37]. However, at the time these algorithms were presented they were mere theoretical exer-
cises in “pen and paper programming” since quantum computers on which they could have
been implemented did not yet exist. Nevertheless, Grover’s amplitude amplification algorithm
convincingly demonstrated that quantum computers can exhaustively search unstructured
lists quadratically faster than digital computers [37, 38] and Shor’s celebrated quantum al-
gorithm for large integer factorization even revealed exponential speedup over classical ap-
proaches [36, 39]. The latter sparked broader attention to the idea of quantum computing, be-
cause Shor’s discovery implied that common data encryption schemes would no longer be
secure if working quantum computers were to become technical reality. In short, these pi-
oneering contributions therefore demonstrated that quantum algorithms running on ideal
quantum computers would indeed exhibit quantum supremacy when dealing with certain dif-
ficult problems. They also made clear that this quantum supremacy would have considerable
practical implications.
First prototypes of genuine quantum computers appeared in the late 1990s. They were
genuine in that they were made up of quantum hardware for physical quantum information
processing as opposed to digital hardware for simulated quantum information processing. For
example, physical implementations of the Deutsch-Jozsa algorithm [35] were reported in 1998
[40, 41] and Shor’s algorithm was first physically implemented by a team at IBM in 2001 [42].
Since then, technical developments progressed quickly and have meanwhile led to commer-
cially accessible quantum computing platforms developed by companies such as IBM, Google,
Microsoft, Rigetti, IonQ, or D-Wave.

1 INTRODUCTION
This state of affairs and the ever intensifying efforts towards marketed solutions con-
firm that quantum computing really is a viable idea; proof of concepts exist, claims as to first
practical observations of quantum supremacy have been made [43], and a growing number
of practitioners see remaining challenges in the development of industrial strength quantum
computers as mere engineering problems rather than as fundamental research problems.
It therefore seems likely that quantum machine learning (QML) will become practical, too,
and that quantum information processing will become applicable at various stages of the ma-
chine learning pipeline.
Indeed, since many of the fundamental problems in machine learning are demanding
optimization or search problems of the kind quantum computing can excel at, research on
QML is noticeably intensifying. The corresponding literature has seen rapid growth and first
textbooks and tutorials have become available [44, 45, 46, 47, 48]. Moreover, more and more
quantum coding challenges are being held that specifically focus on aspects of machine learn-
ing and there are recognizable efforts by industrial players to get machine learners involved in
quantum computing research.
Given these developments, experts predict that quantum machine learning on reliable,
large scale quantum computers will soon facilitate demanding learning tasks and even put
problems into reach which are intractable on digital computers. Should these predictions ma-
terialize, QML technology will further accelerate progress in AI which, in turn, will likely lead
to novel marketable products and commercial solutions impacting economies and societies
[49, 50].
However, next to its potential benefits, QML may also come with certain risks. Since
learning systems now see commercialization and industrial deployment even in sensitive
areas such as autonomous driving or financial services [51, 52, 53, 54], questions as to their
ethics, reliability, trustworthiness, and safety are becoming more urgent than when machine
learning was a mere academic endeavour [55, 56, 57, 58]. In fact, now that machine learning
and quantum computing are beginning to coalesce, these questions are likely to become ever
more important. In other words, given the expected impact of quantum machine learning
on capabilities and utilizability of artificial cognitive systems, it seems appropriate to asses
potential security issues related to quantum machine learning.
In preparation of such an assessment, this present report is intended to take stock and to
give an overview of the current state of the art in quantum machine learning. In section 2,
we first provide a short introduction to quantum computing, its basic principles, and ma-
jor paradigms. Next, in section 3, we clarify the notion of machine learning and elaborate on
common settings and methods.
Section 4 then surveys and reviews the scientific literature on quantum machine learn-
ing. There, we will adhere to the currently established QML taxonomy and consider quantum
inspired classical algorithms for classical data, genuine quantum algorithms for classical data,
classical algorithms for quantum data, and quantum algorithms for quantum data. Our pre-

1 INTRODUCTION
sentation with respect to genuine quantum algorithms will assume a point of view where we
abstract from physical aspects of the hardware of quantum computers. Rather, we will follow
common practice and focus on logical aspects or higher level abstractions that programmers
and software developers are typically exposed to.
Having said this, it is nevertheless important to acknowledge that, as of this writing,
quantum algorithm design is still very much a theoretic endeavor and that many theoretically
valid ideas cannot be implemented on present day quantum computing devices. In other
words, physical or hardware related aspects have to be kept in mind when assessing the
validity of proposed quantum machine learning solutions. To this end, section 5 will review
common assumptions as to data pre- and post-processing, robustness against measurement
noise, numbers of available quantum bits, or required quantum circuit depths and will also
asses how realistic such assumptions are and what they mean for the practical feasibility of
currently proposed QML solutions.
Finally, in section 6, we will summarize key points of this report and provide an outlook
to planned studies on quantum machine learning security.

2 QUANTUM COMPUTING IN A NUTSHELL
2 Quantum Computing in a Nutshell
Quantum computing exploits quantum mechanical phenomena for information pro-

cessing. This offers great computational power but also requires a different kind of thinking
than in the domain of digital computing. The latter may explain why quantum computing has
still not yet received widespread attention in mainstream computer science. As of this writ-
ing, major obstacles for a broader engagement with the topic likely are as follows: 1) quantum
mechanical phenomena such as superposition, entanglement, tunneling, decoherence, or the
uncertainty principle appear to be weird, abstract or unintuitive, and hard to accept for they
cannot be observed in our daily macroscopic environments; 2) the mathematical tools used to
model these phenomena are complex (no pun intended) and again rather abstract; and 3) to
the uninitiated, the mathematical notation used in the quantum computing literature appears
even more abstract and needs getting used to. Acknowledging these difficulties, this section
provides a brief introduction to the basic terminology of- and fundamental concepts behind
quantum computing.
To begin with, we recall that a quantum mechanical system (QM system) is characterized
by a state vector |ψ⟩ in a Hilbert space H over C. The temporal evolution of such a system is
governed by the Schrödinger equation and its observable physical properties (such as position,
momentum, or energy) correspond to Hermitian operators. A measurement of an observable
property of a QM system collapses its state to an eigenvector of the respective operator. This
is to say that a measurement of a quantum variable results in an eigenvalue of the operator
under consideration and the state “jumps” to the corresponding eigenvector. Which of the
eigenvalues and corresponding eigenvectors this will be depends on certain probabilities en-
coded in the state vector. The crucial question as to why this kind of mathematics describes
the behavior of nature on one of its most fundamental levels still awaits conclusive answers;
for now, the empirically undeniable success of this linear algebraic framework has to be at-
tributed to “the unreasonable effectiveness of mathematics in the natural sciences” [59].
Quantum computing (QC) deals with quantum mechanical systems that represent quan-
tum bits or qubits. All throughout our following discussion, we will focus on logical qubits
which are the abstract basic building blocks on which quantum algorithms operate. A physical
qubit, on the other hand, is a physical realization of a logical qubit, namely a physical device
that behaves like a logical qubit and forms a component of the hardware of a quantum com-
puting system. We will neither discuss possible technologies for building such devices, nor will
we discuss the practical advantages and disadvantages of different such technologies. An ex-
tensive overview of the state of the art in this area can be found in a recent study published by
the Federal Office for Information Security [60].
From the point of view of quantum algorithms, our focus on logical qubits makes sense.
It mimics common levels of abstraction in modern software development where program-
mers usually do not have to pay attention to hardware details. In fact, the distinction between
hard- and software aspects of computing was essential for the success of digital computers

once they matured. Current quantum computers, however, have not yet reached this level of
maturity. They are noisy intermediate-scale quantum (NISQ) devices [61] which contain less
than a hundred physical qubits and still suffer from limited coherence times and low fault-
tolerance. From the point of view of existing quantum hardware, logical qubits are therefore
still an idealization since they abstract away technological shortcomings. This may render
some of the mathematically valid ideas for quantum algorithms practically infeasible and, as
of this writing, constitutes yet another difference between quantum- and digital computing.
We will return to this issue in section 5.
On a classical computer, the basic units of information are bits and the mathematics that
describes their behavior is Boolean algebra. On a quantum computer, the basic units of infor-
mation are qubits and the mathematics that models their behavior is complex linear algebra.
While a classical bit is always in one and only one of two possible states (0 or 1), the basic
tenet of quantum information processing is that a qubit exists in a superposition of two basis
states and collapses to either one once measured. Examples of well known physical two-state
quantum systems that exhibit this phenomenon include the polarization of a photon, the spin
of an electron, or the ground- and first excited state of an atom. Yet, it is interesting to note
that quantum mechanical superposition is not exclusive to the subatomic world but may just
as well occur in completely isolated macroscopic systems. This is, for instance, used in quan-
tum information processing devices made of superconducting circuits in which electrical cur-
rents flow in two directions simultaneously.
Superposition is crucial because it implies that the state space of a qubit is not confined
to only two states. It rather consists of infinitely many states which can be represented as two-
dimensional, complex-valued unit vectors that are linear combinations of two distinguished,
linearly independent, orthonormal basis states. Using the Dirac notation, these basis states are
commonly written as |0⟩ and |1⟩ and typically thought of as

1 0
|0⟩ = and |1⟩ = (1)
0 1
In terms of a mathematical equation, the above translates to the statement that (the state vec-
tor of) a qubit can be written as
|ψ⟩ = α0 |0⟩ + α1 |1⟩ (2)
where the coefficients α0 , α1 ∈ C are called the amplitudes of the basis states |0⟩ and |1⟩, re-
spectively. Importantly, these amplitudes have to obey the normalization condition
|α0 |2 + |α1 |2 = 1 (3)
and are interpreted as follows: if a measurement is performed on qubit |ψ⟩, the probability of
finding it in basis state |0⟩ is |α0 |2 whereas the probability of finding it in basis state |1⟩ corre-
sponds to |α1 |2 . In other words, the probability of measuring the qubit in, for instance, state
|0⟩ is given by |⟨0|ψ⟩|2 , because the inner product of |0⟩ and |ψ⟩ evaluates to
⟨0|ψ⟩ = α0 ⟨0|0⟩ + α1 ⟨0|1⟩ = α0 · 1 + α1 · 0 = α0 (4)

Note that measurements performed on |ψ⟩ are irreversible operations because they con-
stitute interactions with the outside world and therefore lead to loss of superposition or quan-
tum decoherence. In other words, once a qubit has collapsed to either one of its basis states, it
henceforth behaves like a classical bit.
Operations on qubits which preserve their quantum mechanical nature are called re-
versible. Mathematically, these correspond to unitary linear operators U ∈ SU2 (C) for which
U U † = U † U = I. Reversible operators can also be written as U = e−iHt/¯h where H is yet an-
other operator called the Hamiltonian. It corresponds to the total energy of a quantum system
in the sense that its spectrum is the set of possible outcomes of measurements of the system’s
total energy.
Another tenet of quantum information processing is that qubits can be combined to form
qubit registers. While a single qubit |ψ⟩ exists in a superposition of 2 states, a quantum register
|ψ⟩ of n qubits exists in a superposition of 2n states. Mathematically, this is to say that
n −1
2X
|ψ⟩ = γi |ψi ⟩ (5)
i=0
P
Here, the γi once again obey i |γi |2 = 1 and the |ψi ⟩ are 2n -dimensional tensor products of
single qubit states. Mathematically, these tensor products are Kronecker products and behave
as follows: If we consider two single qubits represented as

α0 β
|ψ⟩ = α0 |0⟩ + α1 |1⟩ = and |ϕ⟩ = β0 |0⟩ + β1 |1⟩ = 0 (6)
α1 β1
then the state of a system composed of these two qubits is given by the four-dimensional vec-
tor
⎡ ⎤
α0 β 0
α β ⎢ α0 β1 ⎥
|ψ⟩⊗|ϕ⟩ = 0 ⊗ 0 = ⎢ ⎥
⎣α1 β0 ⎦ = α0 β0 |0⟩⊗|0⟩+α0 β1 |0⟩⊗|1⟩+α1 β0 |1⟩⊗|0⟩+α1 β1 |1⟩⊗|1⟩
α1 β1
α1 β1
(7)
This example illustrates that the 2n dimensional state space of an n qubit system is
spanned by a computational basis that consists of 2n basis vectors which are tensor products
of individual basis states. Since such tensor products over basis states occur in many quantum
computing equations, they are typically written more succinctly. For instance, for a quantum
register where n = 3, we would write the 23 = 8 basis states as
|ψ0 ⟩ = |0⟩ ⊗ |0⟩ ⊗ |0⟩ ≡ |000⟩ (8)
|ψ1 ⟩ = |0⟩ ⊗ |0⟩ ⊗ |1⟩ ≡ |001⟩ (9)
..
.
|ψ7 ⟩ = |1⟩ ⊗ |1⟩ ⊗ |1⟩ ≡ |111⟩ (10)

or, even shorter as |i⟩ where i ∈ {0, . . . , 7}. Since the state space of an n qubit system is 2n -
dimensional, operators acting on such states are (products of) 2n × 2n dimensional unitary ma-
trices. Yet another tenet of quantum information processing is that qubits can be entangled.
The physical phenomenon of entanglement is indeed an essential aspect of quantum infor-
mation that has no immediate classical counterpart. To understand its characteristics, we note
that the 2 qubit system in (7) has a complete state that is a tensor product of individual qubit
states. However, systems of qubits can also be in states that can not be expressed in terms of
tensor products. These are called entangled states and a canonical example of such a state is
1 1
|ψ⟩ = √ |00⟩ + √ |11⟩ (11)
2 2
The following simple computations will illustrate why entangled states are special: Con-
sider two systems |ψ1 ⟩ and |ψ2 ⟩ of 2 qubits where |ψ1 ⟩ is in a tensor product state
|ψ1 ⟩ = γ0 |00⟩ + γ1 |01⟩ + γ2 |10⟩ + γ3 |11⟩ (12)
and |ψ2 ⟩ is in the entangled state
1 1
|ψ2 ⟩ = √ |00⟩ + √ |11⟩ (13)
2 2
If we now perform a partial measurement of both systems where only the first qubit of either
system is measured and the second one is left intact, the probability of finding the first qubit of
system |ψ1 ⟩ in, say, state |0⟩ is given by
X
| ⟨0x|ψ1 ⟩ |2 = |γ0 |2 + |γ1 |2 (14)
x∈{0,1}
Moreover, once this partial measurement has been performed, the system will be in the new
(re-normalized) state
P
x∈{0,1} ⟨0x|ψ1 ⟩ |0x⟩ γ0 |00⟩ + γ1 |01⟩
|ψ̂1 ⟩ = qP = p (15)
x∈{0,1} | ⟨0x|ψ1 ⟩ |
2 |γ0 |2 + |γ1 |2
In other words, measuring the first qubit of the first system will cause this system to still be in
a superposition of two computational basis states. In our example, this means that the state
of the second qubit remains undetermined since it could still be found either in state |0⟩ or in
state |1⟩.
If, on the other hand, the same kind of partial measurement is performed on the second
system |ψ2 ⟩, the probability of finding its first qubit in state |0⟩ is given by
X 1 1
| ⟨0x|ψ2 ⟩ |2 = | √ |2 = (16)
2 2
x∈{0,1}

and, upon this measurement, the second system will be in state

P
x∈{0,1} ⟨0x|ψ2 ⟩ |0x⟩
|ψ̂2 ⟩ = qP = |00⟩ (17)
x∈{0,1} | ⟨0x|ψ 2 ⟩ | 2
By the same token, if we had found the first qubit in state |1⟩, the final state of the system
would have been |ψ̂2 ⟩ = |11⟩ and similar outcomes would have been observed if we had mea-
sured the second qubit of the entangled system. That is, whenever two qubits are entangled,
their individual states cannot be measured separately; a measurement of either one of two en-
tangled qubits also determines the state of the other.
Entanglement is of vital importance for quantum information processing because it is
this phenomenon which makes it possible to consider only n (physical) qubits to perform
computation in a 2n dimensional state space. In other words, whereas doubling the “process-
ing power” of a digital computer would require doubling the number of bits, doubling the
“processing power” of a quantum computer only requires adding one extra qubit. Without
entanglement, this exponential increase in performance is provably impossible [62].
Given these prerequisites, we can now discuss the two major paradigms for practical
quantum computing, namely quantum gate computing and adiabatic quantum computing.
The former approaches problem solving by sequencing unitary operators, U1 , U2 , . . . into
quantum circuits and well known examples include Grover’s search algorithm [38] or Shor’s
integer factorization algorithm [39]. The latter considers problem solving as an energy min-
imization task and is concerned with finding ground states of problem specific Hamiltonian
operators H. Adiabatic quantum computing is sometimes falsely said to be less general than
quantum gate computing but both paradigms are provably equivalent with respect to compu-
tational power or expressiveness [63, 64]). Indeed, translating a quantum algorithm from one
formalism to the other requires only polynomial efforts (for instance, in the number of addi-
tionally required qubits or quantum gates) and is therefore possible in principle. Nevertheless,
from the programmer’s point of view, either approach requires its own distinct form of think-
ing and of formalizing problems and we next elaborate on the underlying principles.
2.1 Adiabatic Quantum Computing
Adiabatic quantum computing (AQC), often also called adiabatic quantum optimization
(AQO), relies on yet another quantum mechanical phenomenon. This phenomenon is formal-
ized in the adiabatic theorem [65] which states that, if a quantum system starts in the ground
state of a Hamiltonian operator which then gradually changes over a period of time, the sys-
tem will end up in the ground state of the resulting Hamiltonian. Since Hamiltonians are en-
ergy operators, their ground states correspond to the lowest energy state of the quantum sys-
tem under consideration. The adiabatic theorem therefore characterizes a form of quantum
tunneling which refers to the fact that quantum systems can tunnel through energy barriers.

To harness AQC for problem solving, one prepares a system of qubits in the ground state
of a simple, problem independent Hamiltonian and then adiabatically evolves it towards a
Hamiltonian whose ground state corresponds to a solution to the problem at hand [66]. This
can be done on D-Wave computers [67, 68, 69] which are particularly tailored towards solving
quadratic unconstrained binary optimization problems (QUBOs) of the form
s∗ = argmin s⊺ Q s + q ⊺ s. (18)
s∈{−1,+1}n
Here, the 2n different bipolar vectors s over which to minimize represent possible global states
of a system of n entities each of which might be in one of two local states (+1 or −1). The cou-
pling matrix Q ∈ Rn×n models internal interactions and the vector q ∈ Rn models additional
constraints or external influences.
In physics, QUBOs are known as Ising energy minimization problems [70] and, in ma-
chine learning, they are fundamental to the theory of Hopfield networks [71]. QUBOs pose
discrete- or combinatorial optimization problems which are NP-hard and therefore notori-
ously difficult to solve. They also are surprisingly universal and of considerable practical im-
portance since they often arise in the context of subset selection- or bipartition problems.
Given a set X of n elements xi , subset selection problems ask for a subset X ′ ⊂ X such
that the elements of X ′ meet certain criteria and bipartition problems ask for a partition X =
X1 ∪ X2 such that the elements of either of the disjoint subsets X1 and X2 meet certain criteria.
Now, if a subset selection- or bipartition problem can be (re)written as a QUBO over bipolar
vectors s, the entries of the solution s∗ can be understood as indicator variables. For a subset
selection problem, they define the sought after subset as X ′ = {xi ∈ X | si∗ = +1} and, for a
bipartition problem, we would have X1 = {xi ∈ X | si∗ = +1} and X2 = {xi ∈ X | si∗ = −1},
respectively.
Subset selection- or bipartition problems of this kind abound in data mining, pattern
recognition, machine learning, or artificial intelligence. General use cases can be found in
data base search, the computation of medoids in k-medoids clustering, or the identification of
support vectors in support vector machine training. Prominent practical applications where
QUBOs are central include verification-, planning-, or assignment problems in areas such as
logistics or finance [72, 73, 74, 75].
What renders quantum computing an attractive approach for dealing with difficult QU-
BOs is that seminal work by Farhi et al. [76] provided a simple general recipe for how to solve
them via adiabatic evolution.
Assuming the parameters Q and q of a QUBO to be given, the basic idea is to consider a
time dependent system of n qubits
n −1
2X
|ψ(t)⟩ = αi (t) |ψi ⟩. (19)
i=0

which evolves under a time-dependent Hamiltonian H(t) so that its behavior is governed by
the Schrödinger equation
d
ih̄ |ψ(t)⟩ = H(t) |ψ(t)⟩ (20)
dt
The qubits are prepared in the ground state of a beginning Hamiltonian HB which gradually
changes towards the problem Hamiltonian HP . For the latter, [76] proposes
X
n X
n X
n
HP = Qij σzi σzj + qi σzi (21)
i=1 j=1 i=1
where σzi = I ⊗ . . . ⊗ I ⊗ σz ⊗ I ⊗ . . . ⊗ I denotes the Pauli spin matrix σz acting on the i-th
qubit. For the beginning Hamiltonian, [76] suggests
X
n
HB = − σxi (22)
i=1
where σxi is the Pauli spin matrix σx acting on the i-th qubit. Considering an evolution from
t = 0 to t = T , the Hamiltonian in (20) is assumed to be a convex combination

H(t) = 1 − T · HB + Tt · HP
t
(23)
and can be used to evolve |ψ(t)⟩ from |ψ(0)⟩ to |ψ(T )⟩ where |ψ(0)⟩ is the ground state of
HB .
Finally, at time T , a measurement is performed on the qubits. This causes the whole sys-
tem to collapse to one of its 2n basis states and the probability for this state to be |ψi ⟩ is given
by the amplitude |αi (T )|2 . However, since the adiabatic evolution was steered towards the
problem Hamiltonian HP , basis states |ψi ⟩ that correspond to ground states of HP are more
likely to be found.
The efficiency of this computational paradigm depends on the choice of T which is
known to depend on the minimum energy gap between the ground- and first excited state of
H(t). While exponentially small gaps will render adiabatic quantum computing inefficient,
problems with gaps that scale inverse polynomially can be solved efficiently. In fact, it is
known that energy gaps are inversely proportional to the square root of the number of basis
states that have energies close to global minimum
√ [77]. For problems with a small number of
valid solutions, an appropriate choice is T ∈ O( 2n ) [78]. This, in turn, is to say that adiabatic
quantum computing can solve certain binary optimization problems quadratically faster than
classically possible.
To illustrate the steps involved in setting up practical problems for solution on an adi-
abatic quantum computer and to better explain the process of adiabatic optimization, we
briefly discuss an application example.

The max-sum diversification problem is of importance in location- or portfolio optimiza-

tion and also arises in the context of information retrieval, recommendation, clustering, and
other settings where one has to identify very distinct elements in a given data set. To be more
specific, given a set X = {x1 , . . . , xn } and an appropriate, problem specific distance measure
d(·, ·), max-sum diversification asks for a subset S ∗ ⊂ X of k < n elements of maximum
dispersion and thus consists in solving
X X
S ∗ = argmax d(xi , xj )
S⊂X xi ∈S xj ∈S (24)
s.t. |S| = k.
Collecting the d(xi , xj ) in a distance matrix D ∈ Rn×n and considering a binary indica-
tor vector z ∈ {0, 1}n where zi = 1 if xi ∈ S and zi = 0 otherwise and letting 1 ∈ Rn denote
the vector of all ones, the problem in (24) can also be written as
z ∗ = argmax z ⊺ D z
z∈{0,1}n (25)
s.t. 1⊺ z = k
Furthermore, when introducing a generic Lagrange multiplier λ and using the fact that z ⊺ D z
is quadratic in z, this NP-complete integer programming problem can be written as a QUBO
over z, namely
� 2
z ∗ = argmin −z ⊺ D z + λ 1⊺ z − k ≡ argmin −z ⊺ P z − p⊺ z (26)
z∈{0,1}n z∈{0,1}n
However, the decision variables (18) are bipolar vectors s ∈ {−1, +1}n whereas (26) minimizes
over binary vectors z ∈ {0, 1}n . We therefore emphasize a fact often used when (re)writing
problems for adiabatic quantum computing, namely that it is easy to convert between binary
and bipolar vectors, because z = (s + 1)/2 ⇔ s = 2 · z − 1.
Hence, when plugging z = (s + 1)/2 into (26), it becomes an exercise in algebra [79]
to show that max-sum diversification consists in solving the following energy minimization
problem
s∗ = argmin −s⊺ Q s + q ⊺ s (27)

s∈{−1,+1}n
� �
whose parameters amount to Q = − 14 D − λ 11⊺ and q = − 21 D − λ 11⊺ 1 − λ k 1.
With respect to the original formulation in (24), we note that, once the solution s∗ to (27) has
been found, entries s∗i = +1 indicate which xi ∈ X to select into S ∗ . Figure 1 shows a specific,
rather didactic instance of the problem of max-sum diversification and illustrates the inner
workings of the AQC approach towards solving it. The data in Fig. 1(a) consists of a set X of
n = 12 data points xi ∈ R2 , each of which represents monthly climate conditions (average
temperature and precipitation) in the city of Hamburg. Given this data, the task is to deter-
mine k = 4 months of diverse climatic characteristics.

30 30
25 25
Aug
Jul Aug
Jul
temperature [oC]
temperature [oC]
Jun Jun
20 May Sep 20 May Sep
15 Oct 15 Oct
Apr Apr
10 Mar Nov
10 Mar Nov
5 Feb Jan Dec 5 Feb Jan Dec
0 0
40 50 60 70 80 40 50 60 70 80
precipitation [mm] precipitation [mm]
(a) (b)
|ψi ⟩ |αi |2
|010100100001⟩ 0.742487
|010100010001⟩ 0.104235
|010010100001⟩ 0.020218
|010101000001⟩ 0.010994
|010010010001⟩ 0.002660
|100101010001⟩ 0.002426
|100000001110⟩ 0.002065
|100101100001⟩ 0.001764
|011011000001⟩ 0.001629
|010011000101⟩ 0.001400
(c) (d)
Figure 1: Adiabatic quantum computation for max-sum diversification. (a) A didactic data set
of n = 12 two-dimensional data points representing monthly climate conditions in the city of
Hamburg. (b) Result obtained for k = 4 and Euclidean distances between data points. (c)
Visualization of the evolution of the amplitudes of a corresponding 12 qubit system |ψ(t)⟩.
During its evolution over time t, the system is in a superposition of 212 = 4096 basis states |ψi ⟩
which represent potential solutions to the problem. Initially, each potential solution
(reasonable or not) is equally likely. At the end of the process, one basis state has a noticeably
higher amplitude |αi |2 than the others and is thus more likely to be measured. (d) Ranking of
the ten most likely states for the qubit system to be found in at the end of the adiabatic
evolution; the likeliest final state indexes the months of February, April, July, and December
and thus represents the solution shown in (b).
To set up the QUBO parameters Q and q, the data was normalized to zero mean and unit
variance (a common step in machine learning). After this normalization, Euclidean distances
Dij = ∥xi − xj ∥ between far apart data points are of the order of O(2) so that setting λ = 2 n
will cause neither of the terms in (27) to dominate the minimization problem.

Given Q and q, one can prepare the beginning- and problem Hamiltonian according to
the√above recipe and then let an n = 12 qubit system |ψ(t)⟩ adiabatically evolve over T ∈
O( 212 ) time steps. To visualize the dynamics of this adiabatic quantum search for a solution,
the process was simulated on a digital computer and the quantum computing toolbox QuTiP
[80] was used to solve the Schrödinger equation in (20).
Figure 1(c) shows how the qubit system evolves after it has been prepared in a superposi-
tion of 212 = 4096 basis states |000000000000⟩, |000000000001⟩, . . . , |111111111111⟩ each of
which represents a subset of the given set of data points. In particular, the figure visualizes the
behavior of the amplitudes |αi (t)|2 of the states the system can be measured in. At time t = 0,
all basis states are equally likely but over time their amplitudes begin to diverge; amplitudes
of basis states that correspond to low energy states of the problem Hamiltonian increase while
amplitudes of basis states that could hardly be considered a solution to the problem decrease.
At t = T , certain basis states are therefore more likely to be measured than others and the
table in Fig. 1(d) ranks the ten most likely ones.
The most likely final state in this example is |010100100001⟩ which, when understood as
an indicator vector, indexes the subset consisting of the months of February, April, July, and
December which are indeed diverse with respect to their climatic conditions (see Fig. 1(b)).
Interestingly, the next most likely solution |010100010001⟩ would swap July for August and,
looking at the data in Fig. 1, this behavior of the quantum computing algorithm appears rea-
sonable.
2.2 Quantum Gate Computing
Beside AQC, Quantum Gate Computing (QGC) is the dominant paradigm for quantum
information processing systems. A detailed introduction into that topic can be found in [81].
Whether we are using qubits or bits, we need to manipulate them in order to turn the in-
puts we have into the outputs we need. Thus, QGC borrows many of its defining properties
from classical computers: On a low level, digital computers process information by manipu-
(a)
(b)
b0
b1 |q0 ⟩ = |0⟩ H
b2 |q1 ⟩ = |0⟩
|q2 ⟩ = |0⟩
|q3 ⟩ = |0⟩
Figure 2: Left: Example of a digital subtractor circuit. Right: Example of a quantum gate cir-
cuit for generating a 4-qubit Bell state.

lating sets of bits via atomic logical operations between them. Input bits, being either in the
state 0 or 1, are processed by so-called gates. Each gate performs a fundamental logical oper-
ation, e.g., AND, OR, XOR, NOT. For computations on a few bits, it is useful to represent this pro-
cess in a diagram known as a circuit diagram or just circuit. These have input bits on the left,
output bits on the right, and gates represented by specific symbols in between. An exemplary
circuit that computes the difference of two bits, given a carry bit, is shown in Fig. 2 (a).
The evolution of a closed quantum system is described by a unitary transformation. Thus,
all operations that we perform on qubits must be unitary too. In theory and practice, such op-
erators typically act on only one or two qubits at a time but can be sequenced to form more
complicated qubit transformations. Borrowing terminology from digital computing, quantum
operators acting on qubits are also called quantum gates. An exemplary quantum gate circuit
that creates a Bell state is shown in Fig. 2 (b).
A crucial difference to the classical setting is that, being unitary operators, quantum gates
have to have the same number of input and outputs. Moreover, since the effect of any uni-
tary gate U can be reversed by its Hermitian conjugate U † , a quantum gate’s input can always
be reconstructed from its output. For many classical gates such as the AND gate, this is not the
case. At first sight, this reversibility constraint on quantum gates therefore seems to render
them less expressive or powerful than classical ones. However, by introducing (several) an-
cillary bits, irreversible classical gates can be made reversible and there exist universal sets of
reversible primitives [82]. Reversibility does therefore not restrict computability. Quantum
gates can instead be seen as a generalization of classical reversible gates and the use of ancil-
lary qubits is common practice in quantum computing.
Let us quickly recall the frequently appearing quantum gates and their relation to classi-
cal digital logic. The X-gate constitutes the counterpart of the logical NOT-gate. Applying X to
a qubit switches the amplitudes of |0⟩ and |1⟩.

0 1 1 0
X|0⟩ = = = |1⟩
1 0 0 1
Closely related are the Y -gate

0 −i
Y = = −i|0⟩⟨1| + i|1⟩⟨0|
i 0
and the Z-gate
1 0
Z= = |0⟩⟨0| − |1⟩⟨1| .
0 −1
At a first glace, X, Y , and Z have not much in common. However, each of them can be
interpreted as a rotation around the corresponding axis (x-axis, y-axis, z-axis) by π radians in a
conceived 3-dimensional coordinate system. However, applying the Z-gate to one of the basis
states |0⟩, |1⟩ seems to have no effect:

1 0 1 1 0 0
Z|0⟩ = = |0⟩ Z|1⟩ = = −|1⟩
0 −1 0 0 −1 1

Clearly, |0⟩ and |1⟩ are eigenstates of Z. They form the so-called Z-basis. There is, in fact, an
infinite number of bases from which some have specific names in the context of quantum gate
computing. One that appears directly in various quantum algorithms is the X-basis, consti-
tuted by |+⟩ and |−⟩.

1 1 1 1 1 1
|+⟩ = √ (|0⟩ + |1⟩) = √ |−⟩ = √ (|0⟩ − |1⟩) = √
2 2 1 2 2 −1
The specific reason for why this basis appears frequently is the Hadamard gate, also known
as H-gate. It allows us to create a superposition of |0⟩ and |1⟩, effectively realizing a uniform
distribution over {0, 1}.
1 1 1
H=√
2 1 −1
Interpreted as a change of basis, it allows us to move from the Z-basis to the X-basis:
H|0⟩ = |+⟩ H|1⟩ = |−⟩
An important insight for digital circuits is the universality of specific gate sets. There is no
need to have hardware implementations of all possible logical 2-bit gates available—it can be
shown that a single NAND-gate suffices to implement all Boolean operations. Similar relations
can be found for quantum gates as well, e.g. the following equivalence:

1 1 1 1 0 1 1 1
HZH = √ √ =X
2 1 −1 0 −1 2 1 −1
Indeed, there is a finite number of possible logic gates. However, quantum circuits allow
for parametrized gates. These gates require numeric parameters to define their actual function.
They are of utmost importance for quantum machine learning in general, since they allow us
to define families of functions via a single circuit. Such families constitute model classes for
which classical machine learning techniques can be applied to select those models which fit
best to some user specified data. We will revisit this topic in Section 4.
The P -Gate (also known as phase gate) generalizes the Z-gate. It performs a rotation by ϕ
radians around the Z-axis.
1 0
P (ϕ) =
0 exp(iϕ)
Due to closedness, all quantum gates can in fact be absorbed into a single canonical represen-
tation. This U -gate is parameterized and subsumes all possible single qubit gates.
⎡ ⎤
⎢ cos 2 θ
− exp(iλ) sin 2θ ⎥
U (θ, ϕ, λ) = ⎢
⎣
⎥
⎦
exp(iϕ) sin 2θ exp(i(ϕ + λ)) cos 2θ

All gates discussed so far act only on a single qubit. However, multi-qubit operations are re-
quired to realize non-trivial algorithms and to facilitate quantum entanglement between
qubits. The most common way to work with multiple qubits is via controlled not (CNOT) gates.
They are the quantum counterpart to the classical XOR-gate. While single qubit gates can be
fully described by a single matrix, CNOT has two representations,
⎡ ⎤ ⎡ ⎤
1 0 0 0 1 0 0 0
⎢0 0 0 1⎥ ⎢0 1 0 0⎥
CNOT = ⎢ ⎣0 0 1 0⎦
⎥ CNOT = ⎢
⎣0 0 0 1⎦
⎥
0 1 0 0 0 0 1 0
depending on which qubit is the controller and which is the target. The semantic of the CNOT
is as follows: if the control-qubit is in state |0⟩, the target remains unchanged. If the control-
qubit is in state |1⟩, an X-gate (NOT-gate) is applied to the target. This effect can be applied to
create a Bell state over n qubits: One initializes all qubits to |0⟩ and applies a H-gate to one of
them, say, q0 . q0 is then in a uniform superposition of |0⟩ and |1⟩. Then, n − 1 CNOT-gates are
added, connecting q0 with each of the remaining n − 1 qubits. In all cases, q0 is the controller.
Now, measuring any of the n qubits in state |0⟩ (or |1⟩) implies that q0 must be |0⟩ (or |1⟩) too,
and hence all other qubits. In other words, measuring any of the n qubits determines the state
of all remaining qubits. The resulting circuit is depicted in Fig. 2 (b) for n = 4.
As mentioned earlier, the action of any quantum circuit can be written as a product of
unitaries.
U = U1 U2 U3 . . . Ud
Here, d is called the depth of the circuit. Formally, each 1-qubit and 2-qubit operator has to be
extended to an 2n -dimensional operator in order to fit into the above expression.
Measurements are special in that they enforce the quantum state to collapse into a clas-
sical one. The state will lose its probabilistic nature and all subsequent measurements of the
outputs will yield the same result deterministically. A simple but important example of a mea-
surement is that of a qubit in the computational basis. This is a measurement on a single qubit
with two outcomes deﬁned by the two measurement operators M0 = |0⟩⟨0| and M1 = |1⟩⟨1|.
We observe that each measurement operator is Hermitian, and that M02 = M0 as well as
M12 = M1 . Thus, I = M0† M0 + M1† M1 = M0 + M1 . Suppose the state being measured is
|ψ⟩ = α|0⟩ + β|1⟩. Then, the probability of obtaining the measurement outcome 0 is
P(0) = |ψ⟩M0† M0 ⟨ψ| = |ψ⟩M0 ⟨ψ| = |α|2
Similarly, the probability of obtaining the measurement outcome 1 is P(1) = |β|2 .
2.3 Further Quantum Computing Paradigms
The above discussion might lead to a misconception, namely that adiabatic quantum
computing and quantum gate computing are the only “paradigms” in the broad field of quan-
tum information processing. This short section is therefore supposed to help avoid possible

confusion. The following briefly discusses the notions of topological quantum computing and
variational quantum computing. The latter is of considerable interest in the context of quan-
tum machine learning and will be discussed in detail in section 4.
2.3.1 Topological Quantum Computing
Another term ostensibly similar to adiabatic quantum computing or quantum gate com-
puting is topological quantum computing [83]. However, whereas the former two refer to logi-
cal aspects of quantum computing and abstract away from how to physically implement logi-
cal qubits, the latter refers to physical aspects and a physical implementation of qubits.
We therefore only briefly mention that a topological quantum computer is still a mainly
theoretical device that considers two-dimensional quasi-particles called anyons, their braids,
and their characteristics in three-dimensional space-time. If such a quantum computer could
be built, for instance utilizing quantum Hall effects, an expected advantage over currently
common technical realizations based on, say, superconducting circuits or trapped ions, is that
it would be more robust against decoherence.
2.3.2 Variational Quantum Computing
Variational quantum algorithms (VQA) and variational quantum approximate optimization

algorithms (QAOA) combine quantum computations with classical computations in iterative
feedback loops [84, 85]. This is supposed to address limitation of current NISQ devices (limited
numbers of qubits, limited circuit depths, measurement noise) and aims at harnessing the best
of both worlds (classic and quantum).
Either classical optimizers are used to design working quantum circuits, or quantum
computation outcomes are measured and classical analysis is used to update Hamiltonian
operators for the next round of quantum computations. By now, VQAs have been proposed
for many applications and they appear to be the currently best way of exploiting quantum ad-
vantages in the NISQ era [86]. Indeed, one can show that variational quantum computing is
universal and that efficient input/output strategies for looped classic-to-quantum optimiza-
tion are possible [87]. Corresponding methods are also known as hybrid quantum-classical
computing methods and we will return to these ideas in section 4.
2.4 Technical State of the Art and Limitations
As of this writing, most technologies used to physically implement qubits still face is-
sues of stability, decoherence, error tolerance, and scalability. As a consequence, current NISQ
devices therefore still need many physical qubits just for the purpose of error-correction. In

other words, present day logical qubits typically consist of many physical qubits in order to
be able to perform useful computations. This can cause a gap between theory and practice of
quantum algorithm design. For instance, theoretically valid algorithms which assume a large
number of logical qubits to be available may not yet be practical.
Moreover, as of today, quantum computing is still bit level computing. That is, ab-
stract data structures (such as linked lists or binary trees) or control structures (such as
if-then-else-statements or for- or while-loops known from higher level programming
languages or are not yet available to quantum programmers. Even common programming
patterns such as variable assignments (x = y or x = x+1) are not possible on present day
quantum computers because the no-cloning theorem and no-broadcast theorems state that it
is impossible to create independent identical copies of arbitrary quantum states [88].
Although there are efforts towards identifying quantum programming patterns [89] and
although (open source) software development kits for working with quantum computers such
as IBM’s QISKIT [90], Google’s CIRQ [91], Rigetti’s Forest [92] or PennyLane resulting from a
collaboration of several smaller quantum computing companies [93] become increasingly
available, it is important to note that these are tools for very low-level programming. For in-
stance, the software development kits we just mentioned are Python libraries for the mathe-
matical design of quantum circuits. They do allow for running quantum circuits on quantum
computers or simulators but do not yet provide features equivalent to high-level data struc-
tures or control flow structures.
We will return to these issues and their impact on algorithm design for quantum machine
learning in section 5.

3 MACHINE LEARNING IN A NUTSHELL
3 Machine Learning in a Nutshell
Machine learning is the science of fitting parameterized mathematical models to data in

order to realize intelligent systems that can make predictions or perform inference.
While the term data is often understood to mean known facts that can be recorded and
have an implicit meaning, machine learning generally assumes a more technical point of view.
In what follows, we, too, will assume this point of view and understand data to refer to collec-
tions of numeric values that can be obtained from measurements, surveys, interactions with
technical devices, or comparable procedures and can be processed by computers. At first sight,
this focus on numeric values may appear to restrict generality. Note, however, that everything
that is held in the memory of a computer (a text, a piece of music, an image, …) is encoded in
terms of bit patterns and therefore in terms of numbers.
We will further assume that any individual data object can be encoded in terms of a tuple
of m real numbers. Using a general notation, we will write such a tuple as x ∈ Rm and call it a
data point; a collection D = {x1 , . . . , xn } of n data points will be called a data set.
Data can be annotated by metadata which contain additional information. Common ev-
eryday examples are: a time stamp indicating when a picture was taken, a caption describing
the content of a figure, the resource description framework (RDF) tags of a Website, or object
identifiers such as the international standard book number (ISBN) of a book. Again resorting
to an abstract notation, we will express metadata as y ∈ Rl and refer to an annotated data set
D = {(x1 , y1 ) . . . , (xn , yn )} as a labeled data set.
In machine learning, data sets are seen as a source of information about a specific sce-
nario. The fundamental assumption is that data which were gathered in a certain context
or with respect to a certain application domain are not arbitrary. On the contrary, machine
learners always suppose that application specific data sets exhibit certain regularities which, in
general thought, may be intricate or complex and therefore not immediately obvious to hu-
man analysts.
In other words, any machine learning solution for a real world problem implicitly posits
that there exists some generally unknown process or mechanism which can explain variations
in appearance within a given sample of example data. Mathematically, this translates to the
assumption that there exists some function f ∈ F with parameters θ ∈ Θ which models
salient as well as latent characteristics of a set of data. For in stance, in the case of labeled data,
this means
yj = f (xj | θ) + εj (28)
where the noise term εj accounts for possible inaccuracies which may, for instance, be due to
corrupted measurements or faulty annotations. The family of functions F to which the model
f belongs to is called the hypothesis- or model class and Θ denotes the corresponding parame-
ter space.

Other categorical terms which commonly occur in discussion about machine learning are
supervised learning, unsupervised learning, and reinforcement learning. These refer to general
machine learning philosophies which differ with respect to the richness of data available for
training and testing.
Supervised learning deals with labeled data (xj , yj ) and the goal is to train a model such
that it can produce an appropriate output y for a given input x. A variant of this setting is
semi-supervised learning which aims at assigning a label to unlabeled data using the knowl-
edge contained in a small set of labeled data. If a semi-supervised learning system also incor-
porates its own earlier predictions into this assignment processes, it is sometimes referred to
as a corrective learning system.
Unsupervised learning, on the other hand, works with unlabeled data xj and aims at un-
covering latent or inherent structures within a given data set. Examples of structures users
might be interested in are clusters, higher order correlations, or relational dependencies.
Finally, reinforcement learning is a type of supervised learning that receives feedback in
place of a label [94]. Based on such feedback, reinforcement learning methods tune models
that are typically intended for decision making in feedback situations where a current output
may impact the next input. Such a setting can often be formalized in terms of a partially ob-
servable Markov decision process (POMDP) [95]. Roughly speaking, learning in such settings
happens in a (guided) trial and error procedure where the software agent is in some state, de-
cides for an action which leads to a successor state, and only later finds out if a sequence of
actions had a desired effect. This delayed feedback is then used to adjust the action selection
mechanism via dynamic programming approaches and, over time, the agent learns which ac-
tion to perform when in order to achieve certain (long term) goals.
In what follows, we give a more detailed account of the stages involved in training a ma-
chine learning system. We will largely focus on the case of unsupervised learning, yet, every-
thing we discuss also applies to the other learning paradigms.
3.1 The Machine Learning Pipeline
Given the above premises, the process of developing a machine learning solution for a
practical application can be broken down into several distinct phases. It begins with a data
collection campaign in which representative examples for the intended application scenario
are gathered and possibly annotated.
Here, it is of utmost importance to insist on representative data. This means the collected
data points have to cover every foreseeable situation that may occur during the later deploy-
ment of the system. Indeed, a common and easily avoidable cause for insufficient reliability
of machine learning solutions are biased training data which omit or neglect some of the in-
puts a system is expected to handle. In practice, this can have embarrassing consequences (for
instance in incidents where software classified pictures of African Americans to depict goril-

las1 ) to downright catastrophic ramifications (such as when a cars autopilot software failed to
recognize a truck, crashed into it, and caused the death of its driver2 ).
The crucial point is that if a deployed system is confronted with a kind of input which
has been overlooked during its development, its output can be unpredictable. Ideally, the col-
lected data would be balanced and reflect all possible use case situations in about equal parts.
This, however, may not always be possible. For example, in a predictive maintenance scenario
where the task is to evaluate sensor data in order to predict whether a machine will continue
to run smoothly or whether a failure is imminent, there may be much fewer examples of fail-
ure cases than of normal conditions.
Next, there may be a data pre-processing phase in which the collected data might be
brought into a form more amenable for processing or where flawed data points might be
filtered out. For instance, text data might be transformed into an appropriate numeric
representation or measurement noise in sensor data might be smoothed.
After pre-processing, the collected data is split into two disjoint subsets, namely a set of
training data and a set of test data and developers have to decide for a model class. This deci-
sion usually requires some domain expertise and will generally depend on the nature of the
data as well as the intended use case. For instance, when dealing with a time series prediction
scenario, one might opt for a polynomial function, a Markov chain, a Gaussian process, a deci-
sion tree, or a recurrent neural network. Each such general decision is followed by more spe-
cific decisions: Which degree is the polynomial supposed to have? How many states should
the Markov chain contain? How to parameterize the kernel of the Gaussian process? How
deep should the decision tree be? How many neurons should the neural network have and
how should they be interconnected?
As these examples indicate, there often are numerous mathematical models that may
apply to a given problem. Alas, clear cut criteria or simple suggestions for which model to use
when are hard to come by. Rather, substantial research efforts are spent on understanding the
usefulness of different models in different contexts and on developing new models for new
contexts. However, a general recent trend is to avoid overly specialized models but to consider
very flexible and thus widely applicable models such as deep neural networks instead.
In the training phase, the parameters of the chosen model are then adjusted such that the
model matches the training data to the best extend possible. This adjustment happens auto-
matically by means of running learning algorithms which are typically based on (statistical)
optimization techniques. In order for this kind of mathematical learning to be possible at all,
there has to be a criterion for how to measure how well the model and its current choice of pa-
1
see, for example, https://www.forbes.com/sites/mzhang/2015/07/01/google-photos-tags-two-
african-americans-as-gorillas-through-facial-recognition-software/ or https://www.nytimes.
com/2021/09/03/technology/facebook-ai-race-primates.html
2
see, for example, https://arstechnica.com/cars/2019/05/feds-autopilot-was-active-during-
deadly-march-tesla-crash/ or https://www.forbes.com/sites/bradtempleton/2020/06/02/tesla-in-
taiwan-crashes-directly-into-overturned-truck-ignores-pedestrian-with-autopilot-on/

rameters matches the training data. In other words, there has to be an objective for the param-
eter optimization process and these objectives are often formalized in terms of an error- or loss
function. A simple example of such a loss function often used when learning from labeled data
is the mean squared error
1X
n
E(θ) = ∥ yj − f (xj | θ) ∥2 (29)
n
j=1
and the corresponding training objective would be to solve
θ ∗ = argmin E(θ) (30)
θ∈Θ
Note, however, that potential loss functions are as numerous as potential model choices
and that loss functions should, in fact, be chosen with respect to the data and model at hand.
For example, when working with binary labels, one often considers the so called hinge loss or,
when training probabilistic models, one might want to maximize a likelihood. Even other loss
functions are based on divergence measures, entropy criteria, or problem- or model specific
distances. In fact, the investigation of loss functions, their characteristics, and applicability is
an important topic of ongoing machine learning research.
Learning algorithms are numerous, too, and the question of which algorithm to consider
in the training phase should again be decided with respect to the given data, model, and loss
function. Indeed, depending on the chosen model and loss function, it might be trivial to
neigh impossible to optimally solve the learning problem in (30). For instance, if f (x | θ) is
linear in its parameters θ and E(θ) is convex, there will be a closed form solution θ ∗ . Depend-
ing on the numbers of training data points and model parameters, it may still require consid-
erable efforts to practically compute this solution, but one knows that it exists. On the other
hand, if the model f (x | θ) is highly non-linear, the error landscape E(θ) will have numer-
ous local minima and an algorithm that is guaranteed to find the globally optimal solution θ ∗
might not even exist. In situations like these, one often applies optimization techniques such
as gradient descent
θt+1 = θt − ηt · ∇E(θt ) (31)
or other iterative methods. Note, however, that f (x | θ) may be a complicated, compos-
ite function so that the computation of ∇E(θt ) itself may already require a whole chain of
complex operations (such as in case of the backpropagation algorithm for training neural net-
works). Moreover, the choice of meta-parameters of a training algorithm such as the step size
ηt in (31) is typically a non trivial matter and might require careful tuning. All in all, it can thus
be very burdensome to train expressive models with many degrees of freedom or parameters.
In fact, modern models such as deep neural networks with billions of parameters, typically
necessitate the use of high performance computing hardware for training. Improved training
algorithms, their convergence rates, and performance guarantees are therefore yet another
topic of ongoing machine learning research.
In the test phase, the trained model is then validated and evaluated on the test data set.
This requires an appropriate performance measure for which there are again abundantly

many (application dependent) choices. For instance, in the context of a classification task,
one often evaluates the accuracy, i.e. the percentage of correct class predictions, of a trained
classifier. For an information retrieval system, one may be interested in its recall and precision
which measure the percentage of relevant retrieved instances and the percentage of relevant
instances among the retrieved instances. But there also exist highly specialized performance
indicators such as the perplexity of a language model or the BLEU score of a language
translation system.
What is of pivotal importance in this phase is that training data and test data must be
independent. In other words, data points considered during training must not be reconsid-
ered during testing. This is crucial because what really needs to be assessed are the general-
ization capabilities of the trained model. That is, one needs to evaluate how well the function
f (x | θ ∗ ) performs on novel data, i.e. on data it has not seen during training. For reasons we
will explain below, it is actually dangerous to evaluate the trained model on its training data
and, for reasons we explained above, it is again pivotal that the testing data are representative.
Once a trained model has been thoroughly tested and found to perform reliably as well as
accurately, it can finally be deployed in practice and be released into its application phase.
Use cases for this general methodology are manifold but most commonly found in set-
tings where systems need to make data-based predictions, generate data-based suggestions
or decisions, or classify novel observations or measurements. We also note that terms such as
prediction, classification, or decision making can have a rather broad meaning. For instance,
a prediction could be an estimate of tomorrow’s stock market closing price or the statement
that this English sentence translates into that German sentence. A classification could be as
simple as “this picture shows a cat” or as sophisticated as “this picture shows a little boy on a
sunny beach playing with a red ball”. What kind of capabilities a learning system can achieve
largely depends on the nature of the available training data and on the nature of the chosen
model. With respect to the latter, notable strides have been made using modern neural net-
work models and architectures such as variational auto-encoders, generative adversarial net-
works, or transformer networks [96, 97, 2].
3.2 Lazy vs. Eager Machine Learning
Yet another categorization of machine learning techniques contrasts lazy learners with
eager learners. It is worthwhile to briefly mention this distinction as it points to different kinds
of levers one may consider when trying to incorporate quantum computing techniques into
the machine learning pipeline.
A lazy learner is a machine learning system that simply stores data and delays modeling
until asked to make predictions. Such approaches are particularly suitable when dealing with
large and frequently changing databases since only the affected part of the model needs to be
retrained for new data. A simple yet common example of a lazy learner for classification prob-

lems is the k-nearest neighbor method which classifies novel observations based on their dis-
tances to known, i.e. previously learned, prototypes. Given an incoming data point, it searches
a database of prototypes for the k nearest ones, performs a (weighted) majority voting over
their class labels, and assigns the result to the new observation. Here, training is rather simple
and can usually be accomplished quickly as the training problem basically consists in auto-
matically identifying suitable prototypes. Application of such a classifier is usually simple and
not too time consuming as well. However, depending on the size of the prototype database
and the number k of nearest neighbors to consider, the required computations be burden-
some and may necessitate the use of specific data- or index structures to guarantee fast run-
times. A considerable advantage of methods such as a k-nearest neighbor classifier is that they
can easily be (re)trained during the application phase of the system because their prototypes
may be extended or modified.
An eager learner, on the other hand, is an algorithm that trains a model in a dedicated
offline training phase and considers a usually large but static training data set. This can lead to
a very tedious or computationally intensive learning phase, but the application of the trained
model to new examples is very efficient. Most of the well known machine learning models
such as, say, neural networks are trained eagerly. Indeed, since we just mentioned prototype
techniques as a prime example for lazy learning, it is important to note that not all prototype-
based models are lazy. A prominent example of a class of eager prototype-based models are
support vector machines.
3.3 Sources of Uncertainty in Machine Learning
To conclude this short overview of machine learning, we need to point out that the pro-
cess of fitting a parameterized mathematical model to data is inherently statistical. This state-
ment applies to any machine learning paradigm and should be understood as follows: From
an abstract point of view, a fitted model provides a summary or compressed representation of
the information contained in a training sample and there are several uncertainties involved in
its training that may lead to different results in different training runs.
First of all, data collection is an informed but random process since different machine
learning practitioners may collect different data samples when tackling the same problem.
Second of all, modelling is an informed but random process since different practitioners may
opt for different models when tackling the same problem. Third of all, model training may
start with random initializations of the model parameters and thus settle in different local
minima of the chosen loss function.
These sources of uncertainty can lead to undesirable outcomes. For instance, models can
be biased and inappropriate for the data at hand. This is mainly the case when models are too
simple or not flexible enough to capture relevant input output relations. In this case, train-
ing will lead to underfitting. Models can also be too flexible and thus overly sensitive to small
fluctuations in their training data. Training such models on (slightly) different data samples

will likely produce considerably different results and they are said to show high variance. High
variance models tend to learn minute irrelevant details and may therefore suffer from overfit-
ting.
Phenomena like these explain the need for training and testing on independent data sets
because only if training and test data differ can under- or overfitting be identified.
The so called bias-variance dilemma poses a considerable challenge in machine learning
as one typically strives for models that capture patterns in the training data and also general-
ize well to unseen data. Alas, it may not be possible to achieve both these goals simultaneously
since bias and variance tend to be reciprocal. As a rule of thumb, bias can be reduced by focus-
ing on local information as it is done in nearest neighbor models or in radial basis function
models. Variance, on the other hand, can be reduced through averaging over multiple data
points or larger regions in data space as it is done in most other common, typically more com-
plex models.
A common fallacy in this context is to assume that model complexity (measured in terms
of the number of model parameters) causes variance and overfitting. However, this need not
be the case; instead, overfitting in complex models is mainly due to too much freedom in the
choice of their parameters. If this freedom is restricted, for instance by constraining the value
a parameter can assume, overfitting can often be avoided. Imposing restrictions on model pa-
rameters is known as regularization and there exists a host of methods for how to accomplish
this.
Other methods for mitigating variance include data dimensionality reduction, feature
selection, or increasing the size of training data sets. At the same time, adding features or in-
creasing data dimensionality can decrease bias. Moreover, many models and algorithms come
with specific parameters which allow for trading off bias and variance. For instance, choosing
a higher value of k in a k-nearest neighbor model will increase its bias and decrease its vari-
ance whereas a lower value of k will decrease bias and increase variance.
Yet another way of addressing the bias-variance dilemma consist in using ensemble learn-
ing techniques such as boosting or bagging. While boosting algorithms combine many models
of individually high bias into an ensemble of low bias, bagging methods combine individual
models of high variance into an ensemble with low variance. In this context, it is interesting
to observe that recent empirical results indicate that modern neural networks with very wide
layers (which can be seen as ensembles of individually weak learners) do not seem to suffer
from the reciprocal bias-variance characteristics of more traditional models [98, 99].

4 QUANTUM MACHINE LEARNING
4 Quantum Machine Learning
Modern machine learning is undeniably successful but also a very resource intensive en-
deavor. By now, the field has reached a point where practical computational efforts for train-
ing state-of-the-art models can only be dealt with on high performance computing hardware.
Since this trend is likely to continue, it is no surprise that a growing number of machine learn-
ing researchers are beginning to look at the potential benefits of quantum computing.
By now, the scientific literature on quantum machine learning is vast and taking stock of
the state of the field is warranted. In the following survey, we will consider quantum inspired
classical algorithms for classical data, genuine quantum algorithms for classical data, classical
algorithms for quantum data, and quantum algorithms for quantum data.
4.1 Quantum Inspired Machine Learning
Quantum inspired models constitute a broad class of frameworks, methods, and algo-
rithms for classical data processing on classical computers that involve quantum mechanical
concepts, in particular, the mathematics of quantum mechanics and quantum information
processing. Alas, the understanding of what kind of methods and techniques are inspired by
quantum mechanical insights varies widely. Some authors adhere to a narrow interpretation
and only consider methods or algorithms which clearly would not exist without quantum me-
chanics. Others assume a rather broad point of view and consider, say, general optimization
techniques such as simulated- or mean field annealing to be quantum inspired just because
they occur in- or may have originated from the study of certain quantum mechanical systems.
An example of an early contribution that is quantum inspired in the narrow sense can
be found in a book by van Rijsbergen in which he develops a quantum theory of information
retrieval [100]. Modern information retrieval deals with the problem of searching data sets of
unstructured media objects (texts, images, videos, etc.) for items or content related to a user
query. A well known instance of this setting is Web search and Web search engines have be-
come the best known examples of information retrieval systems.
The information retrieval problem is often formalized in terms of linear algebraic vec-
tor space models. Here, n data objects are encoded in terms of m dimensional vectors which
are then gathered in an m × n matrix. A decomposition of this matrix into factor matrices of
lower rank can reveal latent structures in the data and, if queries are encoded in terms of m
dimensional vectors, too, query matching simply becomes the evaluation of inner products.
Based on these ideas and borrowing inspiration from quantum logic, a venerable quan-
tum theoretic approach towards reasoning [101], van Rijsbergen argues that information re-
trieval should ideally be formulated in terms of superpositions in (infinitely dimensional)
Hilbert spaces. There, eigenvectors of quantum density operators would represent basic (la-
tent) concepts to be searched for, and the corresponding eigenvalues would measure overlaps

between concepts and queries. His ideas are convincing and compelling but also of limited
practical value. Suitable quantum hardware to implement them on does not yet exist and digi-
tal emulations that would go beyond toy examples are impossible, because they would require
exponentially large amounts of digital memory to store adequate Hilbert space representa-
tions of media objects.
Examples of a more generous interpretation of quantum inspired algorithms can be
found in a recent contribution by Arrazola et al. [102] who, among others, are concerned
with recommendation systems. Recommendation systems are a topic closely related to
information retrieval, and play an important role in e-commerce. Here, too, popular models
are based on linear algebraic formulations. If user-item preferences are represented in terms
of a typically sparse m × n matrix, the problem of recommending items to users becomes a
problem of predicting missing matrix entries. Just as in the case of vector space information
retrieval, this prediction problem can be tackled using matrix decomposition methods.
While matrix factorization problems frequently occur in computational intelligence ap-
plications, their exact solutions are often hard to come by because the amount of floating
point operations required for modern large scale matrices exceeds what is reasonably possi-
ble on digital computers. Therefore, there exists a vast amount of literature on randomized or
probabilistic algorithms for fast, approximate numerical linear algebra [103, 104, 105, 106, 107,
108].
Indeed, Arrazola and his colleagues consider the FKV sampling algorithm [103], replace
remaining exact procedures (for the computation of eigenvalues) by expected value estima-
tions, and deem this a quantum inspired approach because similar estimators occur in quan-
tum mechanics. However, probabilistic approaches to matrix factorization (with proper renor-
malization) have a venerable history, among others in machine learning [104, 105, 109], where
they have hardly been thought of as inspired by quantum mechanics. In fact, if one were to
subscribe to the broad interpretation of quantum inspired techniques, any statistical inference
involving, say, Monte Carlo sampling, would count as such just because its mathematical ap-
paratus was first developed by the quantum physics community.
To be frank, the use of the term quantum inspired often appears to be a marketing instru-
ment for making certain approaches look more interesting than they turn out to be upon in-
spection. Consider, for instance, methods surveyed by Zhang [110] who benchmarks more
than a hundred quantum inspired evolutionary optimization algorithms. Evolutionary algo-
rithms take inspiration from biology because they work on populations of genotypes (i.e. en-
coded possible solutions to a problem) which express phenotypes (i.e. correspondingly de-
coded solutions). Problem solving or optimization happens in an iterative manner, where
current auspicious genotypes are recombined or mutated into new ones whose phenotypes
then represent new possible solutions. Since better solutions are used to replace worse ones,
the population as a whole becomes better over time. This exploration of the solution space
continues until the population contains enough members which express acceptable solu-
tions. Traditionally, genotypes are often just bit string encodings of more complex objects,

and recombination and mutation happen in a purely random fashion. Obvious improvements
over this baseline include more expressive encodings and more informed or probabilistically
guided updates. Granted, some of the methods surveyed in [110] apply genuine quantum con-
cepts (such as encodings in terms of sets of two-dimensional complex vectors or qubits) but
many others simply rely on probabilistic update mechanisms whose connection to quantum
mechanics has to be considered loose at best.
In what follows, we will therefore attempt to focus on quantum inspired techniques that
are relevant to machine learning and, at the same time, are recognizably rooted in the mathe-
matics of quantum mechanics and quantum computing.
4.1.1 Tang’s Quantum Inspired Algorithm for Recommendation Systems
A broad class of techniques which are genuinely quantum inspired consists of methods
that “dequantify” quantum algorithms. A widely reported example of such a method is due to
Tang [111] who, as a graduate student, famously discovered a classical analogue of a quantum
recommendation system algorithm introduced by Kerenidis and Praksh [112].
The quantum algorithm in [112] repeatedly samples from a low-rank matrix approxi-
mation by means of running quantum phase estimation [113, 114, 115] and quantum pro-
jections (partial measurements). This allows for sampling matrix elements proportional to
their magnitude and thus for sampling matrix elements most relevant to the recommenda-
tion task without having to access every element of the matrix. Kerenidis and Praksh thought
this quantum algorithm to be exponentially faster than classically possible, because a classical
implementation of the procedure seemed to necessitate iterations over all matrix elements.
Crucially, this supposed quantum advantage hinges on the assumption that incoming classical
data can efficiently be encoded in terms of quantum states. However, while the QML litera-
ture often assumes such state preparations to be given and posits the existence of quantum
random access memories (QRAMs) in which these states are available, Kerenidis and Praksh
actually provided a protocol and data structure for this purpose.
Tang’s fundamental insight was that this protocol also allows for encoding classical data
in a representation that satisfies norm constraints required in randomized linear algebra.
There, it is known that norm-based sampling of, say, matrix columns minimizes variance
among all unbiased estimators of factor matrices. Moreover, robust estimates can be obtained
from samples smaller than the size of the matrix to be factored [103, 108]. Tang therefore
swapped quantum state preparation for preparation of a classical sampling procedure and
thus obtained a classical algorithm only polynomially slower than the original quantum
method. In other words, Tang’s work showed that, in this particular matrix completion setting,
potential quantum speedup is not as substantial as it appeared to be at first sight. This led her
to conclude that claims as to speedups of QML algorithms over classical ML algorithms should
take into consideration any state preparation assumptions in a QML model and match them
against sampling assumptions in a corresponding classical ML model.

Tang’s work therefore provides new directions for classical algorithm research and helps
better understanding for which kind of problems one can or cannot expect exponential quan-
tum speedup. Indeed, in the wake of her discovery, efforts in “dequantifying” quantum algo-
rithms have noticeably increased and combinations of her encoding scheme with randomized
numerical linear algebra are becoming ever more popular. Examples include new methods for
estimating pseudoinverses [116] based on quantum algorithms for the singular value decom-
position [117], sub-linear algorithms for solving linear systems of equations [118] based on the
HHL algorithm [119], and even faster algorithms for recommendation systems and regression
problems [120].
4.1.2 Tensor Networks
Tensor networks are mathematical models originally developed for the study of many-
body quantum systems in condensed matter physics [121, 122]. More recently, they were
found to be helpful tools in quantum information processing and connections to machine
learning models have been established [123, 124].
Tensor networks represent quantum states based on local entanglement structures. This
is particularly useful whenever one is dealing with states that have a tensor product structure.
Consider a system of n qubits whose description would usually require O(2n ) complex coeffi-
cients or, equivalently, a complex-valued tensor Γ with n indices
X
|ψ⟩ = γi1 i2 ...in |i1 ⟩ ⊗ |i2 ⟩ ⊗ · · · ⊗ |in ⟩ (32)
i1 ,i2 ,...,in
However, similar to the matrix factorization techniques discussed above, such tensors can be
modeled in terms of contractions of products of lower order tensors, for example
X
γi 1 i 2 i 3 i 4 = αi1 jk βki2 l δli3 i4 (33)
j,k,l
This allows for graphically describing the tensor as a network of interconnected tensors and,
interestingly, when modeling physical systems, such networks tend to have common basic
building blocks.
Just as the matrix factorization techniques sketched above, low rank tensor decomposi-
tions have a venerable history in machine learning and data science [125]. Among others, they
allow for flexible recommendation, fast video analysis, or robust document understanding
[126, 127, 128]. Since data scientists increasingly deal with multidimensional data of high vol-
ume and higher order latent structures, tensor decompositions and tensor networks together
with corresponding inference- or learning algorithms are attracting more and more interest.
There exist kernelized or non-Euclidean extensions, various cost- or loss functions to guide
the estimation of meaningful factors, and distributed computing approaches. Tensor networks
thus allow for addressing large scale problems and are at the heart of generalized regression

and classification techniques, support tensor machines, higher order canonical correlation
analysis, higher order partial least squares, and generalized eigenvalue decomposition. Last
but not least, they also allow for the optimization of deep neural networks [123, 124].
Indeed, studying the capabilities of deep learning, Lin, Tegmark, and Rolnick recently ar-
gued that tensor network models may provide an explanation as to the good performance of
deep neural networks [129]. Although it has been known for long that mathematical theorems
guarantee that very wide neural networks are universal approximators, i.e. are able to learn
arbitrary functions arbitrarily well, they observe that functions of practical interest can often
be learned more “cheaply” by hierarchical networks with comparatively narrow layers. They
argue that this is due to characteristics such as symmetry, local structure, compositions, and
polynomial combinations of log-probabilities which frequently occur in physics and that deep
networks capture such aspects in a manner similar to tensor network models in quantum in-
formation processing. Indeed, they prove several “no-flattening theorems” which establish
that certain deep neural networks cannot be approximated by shallow ones without loss of ac-
curacy and efficiency. Of particular interest in the context of quantum machine learning is the
authors’ result that a shallow neural network cannot multiply n variables using fewer than 2n
neurons in its single hidden layer.
In an important practical contribution, Stoudenmire and Schwab assumed a tensor net-
work point of view on deep learning [130]. They showed how algorithms for the optimization
of tensor networks can be adapted to supervised learning tasks. In particular, they work with
matrix product states to parameterize deep networks for image classification. On the MNIST
data, a standard benchmark data set in machine learning, they observed less than 1% test error
and thus reached state of the art performance.
Results like this spawned further research into tensor network inspired deep learning.
For instance, Glasser, Pancotti, and Ciracet recently explored the connection to probabilistic
graphical models, a venerable class of machine learning models [131]. They considered gener-
alized tensor networks where information from a tensor might be copied and reused in other
parts of the network, showed how to integrate this idea into common deep learning architec-
ture, and derived an algorithm to train such networks in a supervised manner. This proved to
overcome the limitations of regular tensor networks in higher dimensions while not loosing
computation efficiency. In experimental evaluations with image- and sound data, they found
their method to improve on previously proposed tensor network algorithms. However, an ob-
servation most interesting in the context of quantum machine learning is that their approach
can, in principle, be implemented on quantum computers and may therefore impact future
research on quantum assisted machine learning.
Tensor networks therefore constitute interesting models in the intersection of the disci-
plines of quantum computing and machine learning and there is a quickly growing scientific
community dedicated to the topic. For instance, several workshops on tensor networks for
machine learning have already been held at the NeurIPS conference, a primary venue of ma-
chine learning research. However, so far, corresponding methods have not yet widely caught

on and not yet found their way into mainstream machine learning technology. What is no-
ticeable, though, is that researchers from Google, the Perimeter Institute, and the company X
Development LLC recently released TensorNetwork, an open source library for implementing
tensor network algorithms [132]. This library is supposed to support physicists and machine
learners alike and provides functionalities for efficient high volume sparse data handling and
tensor factorization. Whether or not initiatives like this will boost the use of tensor networks
in theory and practice just as other libraries did in the case of deep learning technology re-
mains to be seen.
4.1.3 Digital Annealing
Adiabatic quantum computers such as produced by D-Wave realize an energy mini-

mization process called quantum annealing and they are specifically tailored towards solving
quadratic unconstrained binary optimization problems (QUBOs) of the general form in
equation (18) in section 2.1. Specific instances of QUBOs occur in the context of verification-,
planning-, or assignment problems in areas such as, say, logistics or finance. While QUBOs
therefore are of considerable practical importance, they are also difficult to solve in general
because they pose combinatorial optimization problems which are NP-hard in general. It
is therefore expected that (adiabatic) quantum computing will have economic impact as it
provides novel approaches towards industrially relevant problems.
Alas, the technical effort required for running present day adiabatic quantum computers
is substantial. While they have the potential to deliver unprecedented computing power, they
must be maintained at temperatures near absolute zero and be protected against magnetic in-
terference, thermal variation, and mechanical vibration in order for their physical realizations
of logical qubits to remain in superposition. Just the amount of energy required to maintain a
reliable cryogenic environment for a current D-Wave 2000Q machine is estimated to exceed
25 kWh [133]. Present day adiabatic quantum computing is thus an expensive endeavor and
may not yet amortize its costs in industrial applications.
However, in machine learning, QUBOs as in (18) are known as Hopfield energy minimiza-
tion problems and play a central role in the theory of Hopfield networks [71]. Hopfield net-
works are a special kind of neural networks inspired by physical representations of spin glass
phenomena [70] and they are of considerable theoretical interest because they provide sim-
ple models for higher cognitive processes such as memory retrieval. Hopfield networks have
a venerable history and are established textbook material [22]. Among others, there exist vari-
ous classical algorithms for Hopfield energy minimization ranging from random updates, over
greedy gradient descent methods to simulated- and mean field annealing.
Against this backdrop, Fujitsu has developed special purpose digital hardware for solving
QUBOs. They refer to their technology as a digital annealer and market it as a solution that
rivals the utility of quantum computers [134].

Fujitsu’s system is based on conventional complementary metal-oxide-semiconductor

(CMOS) technology and von Neumann architectures; their dedicated chip fits onto a single
circuit board, works at room temperatures, and does not require a complex support infras-
tructure. The optimization algorithm implemented on this hardware is based on simulated
annealing which is extended in several directions. For instance, the method employs a paral-
lel trial Monte Carlo procedure which considers several switches of decision variables (sim-
ulated qubits) in parallel. This is supposed to overcome problems with respect to low accep-
tance probabilities in common annealing algorithms. The method also involves a so-called dy-
namic offset mechanism which raises acceptance probabilities after those iterations in which
the Monte Carlo scheme did not make any progress. This is empirically observed to help the
algorithm overcome narrow barriers in the energy landscape of the problem to be solved, and
thus to make faster progress towards solutions. In its current version, the system can solve
QUBOs with up to 1024 variables [135].
A similar digital annealing technology has been independently developed by a team of
researchers at TU Dortmund and Fraunhofer IAIS [133, 136]. Their system is implemented on
very affordable field programmable gate arrays (FPGAs). It can currently solve QUBOs of up to
2048 variables and thus exceeds the number of variables of the D-Wave 2000Q adiabatic quan-
tum computer at a mere 0.006% of its power consumption. Moreover, while D-Wave machines
can currently only solve fully connected problems with up to 119 variables, the FPGA-based
solution supports dense parameter matrices for all 2048 variables.
The system has been shown to successfully solve machine learning problems such as k-
means clustering, maximum a-posterior prediction, or binary support vector machine train-
ing. The optimization algorithm running on this hardware takes parallelism into account and
is based on a customizable (µ + λ) evolutionary algorithm which allows for tuning the max-
imal problem dimension n, the number of parent solutions µ, the number of offspring solu-
tions λ, and the number of bits per coefficient Qij . When working with low-budget FPGAs,
this makes it possible to allocate more FPGA resources either for parallel computation (param-
eters µ and λ) or for the problem size (parameter n and bit depth of Qij ).
4.1.4 Quantum Inspired Data Clustering
The term quantum clustering refers to a class of quantum mechanically inspired density-
based clustering algorithms which assume that clusters are defined in terms of regions of
more densely distributed data points. Specifically, quantum clustering algorithms model a
given set of data in terms of a Gaussian mixture model consisting of one mixture component
per data point. This Gaussian mixture is then considered as a quantum mechanical wave func-
tion for the data set and a quantum potential is constructed such that the data wave function
becomes a stable solution to the time-independent Schrödinger equation. Gradient descent
on this potential causes data points to move towards nearby local minima and data points that
end up close to one another are assumed to belong to the same cluster [137].

More elaborate versions of this approach replace gradient descent by quantum evolution
which can be understood as a kind of non-local gradient descent that, in turn, is capable of
tunneling through potential barriers. While this has considerably higher computational costs,
it allows for interesting data visualizations and the identification of substructures turning the
method into a hierarchical clustering approach [138].
4.1.5 Quantum Inspired Gravitational Search
Gravitational search algorithms (GSAs) are among the newest in the class of swarm op-
timization algorithms which rely on the metaphor of gravitational interaction between data
objects [139]. GSAs have the advantage that they are easy to implement and capable of escap-
ing from local optima of an objective function. They have by now become established tools for
effective and efficient global optimization in solving various kinds of continuous problems.
In particular, the binary version, BGSA, applies to solving binary encoded problems and the
discrete version, DGSA, allows for solving combinatorial problems [139].
Nezamabadi-pour [139] introduced a population based meta-heuristic search algorithm
that is a binary quantum inspired gravitational search algorithm (BQIGSA) combining both
gravitational search and quantum computing. His underlying idea is to solve binary encoded
problems by a quantum bit superposition together with a modified rotation Q-gates strat-
egy. Nezamabadi-pour evaluated the algorithm’s effectiveness by performing experiments
on combinatorial 0–1 knapsack problems or Max-ones functions. He found that BQIGSA can
compete with classical BGSA, conventional genetic algorithms, binary particle swarm opti-
mization including a modified version, a binary differential evolution, a quantum inspired
particle swarm optimization, and three well known quantum inspired evolutionary algo-
rithms.
Lou et al. [140] provide a concrete use case for quantum inspired binary gravitational
search algorithms, namely the problem of predicting failure times of cloud services. They ar-
gue the importance of this case by pointing out that data centers coordinate several hundred
thousand heterogeneous tasks to provide the services’ high reliability. The authors motivate
their research by the great challenge to acquire the optimal parameters for a relevance vector
machine approach towards solving nonlinear predicting problems. In practical evaluations,
they observed similar to better predicting performance of their IQBGSA-RVM algorithm com-
pared to the baselines of chaotic genetic algorithms, binary gravitational search algorithms,
binary particle swarm optimization, quantum inspired binary particle swarm optimization
and standard BQIGSA (which all employ relevance vector machines).
4.2 Machine Learning for Quantum Computing
The notion of machine learning for quantum computing commonly refers either to the
use of classical methods for preparing inputs for quantum processing units and processing

outputs obtained from quantum hardware, or to the use of classical methods for designing
quantum circuits.
4.2.1 Quantum Circuit Design
Quantum circuit design for quantum gate computing requires considerable experience, a
deep understanding of the mathematics of quantum information processing, as well as a good
deal of creativity. Because quantum circuit design can thus be considered a difficult task [89],
the idea of using machine learning or optimization algorithms for this purpose arose early on
and can be traced back to the 1990s.
For instance, Rubinstein [141] is concerned with evolutionary algorithms for quantum
circuit design. He discusses possible encoding schemes and fitness functions and presents
evolved circuits that allow for the production of entangled states. Leier [142] also works with
genetic algorithms for quantum circuit design, yet focuses on how to address exponentially
large search spaces for operators on exponentially large quantum states. His almost twenty
year old findings suggest that quantum circuits can be evolved to a certain extent or that hu-
man experts can manually infer suitable quantum circuits from evolved solutions for small
problem instances. Notably, he observed that the use of evolutionary crossover operators
seemed to deteriorate the quality of the solutions found.
Now that working prototypes of quantum computers have become available, design pro-
cedures similar to the above can be tested in practice. For instance, Franken et al. [143] are con-
cerned with variational quantum eigensolvers (VQEs). These are hybrid algorithms that com-
bine classical and quantum computing steps in order to determine the eigenvalues of large
matrices. Solutions to this general problem are sought after in many areas of science and en-
gineering. For instance, in quantum simulations, the matrix in question often is the Hamilto-
nian of a quantum system and its lowest eigenvalue is of interest as it characterizes the ground
state of the simulated system.
Using the VQE approach, a quantum subroutine is run inside of a classical iterative opti-
mizer. This quantum subroutine prepares a state based on a set of given parameters and per-
forms a series of measurements in the appropriate basis. Measurement results are read into
a classical memory and are then used to classically estimate expected values of the parame-
ters for the next iteration. Franken et al. observe that the efforts for estimating gradients of the
kind of cost function that occur in this process are considerable. As a remedy, they therefore
work with a weight-agnostic evolutionary scheme. They test their approach on real quantum
hardware in the IBM quantum experience environment and use the automatically determined
circuits to solve benchmark problems such as the transverse field Ising Hamiltonian and the
Sherrington-Kirkpatrick spin model [144].
Indeed, the idea of parameterized quantum circuits has lately attracted increasing atten-
tion. It allows for the use of reinforcement learning [145, 146, 147] and other learning algo-

rithms [148, 149] for quantum circuit design. Similarly, machine learning methods are increas-
ingly considered as tools for solving quantum circuit mapping problems, i.e. problems of map-
ping a given general quantum circuit design onto a specific NISQ architecture [150, 151, 152,
153].
4.2.2 Quantum State Preparation
Machine learning approaches are more and more considered for other applications beside
automatic circuit design. For example, they are used as a tool for preparing the initial state and
for modelling the noise characteristics of a quantum computing system.
As will be detail below, the ability to load classical data efficiently into quantum states is
the basis for the realization of many quantum algorithms. However, the best known general
methods require O(2n ) gates to load an exact representation of a generic data structure into
an n-qubit state. Therefore, scaling issues need to be taken into account because scaling can
easily predominate the complexity of a quantum algorithm and thereby impair any potential
quantum advantage.
Grover and Rudolph demonstrate how log-concave and other efficiently integrable
probability distributions can be approximately encoded into an m-qubit register [154]. Even
though the method in Grover and Rudolphs seminal note is presented for the univariate case,
they explain that an extension to multiple dimensions, as it is usually the case in machine
learning, is possible. Nevertheless, the described approach does require that the underlying
distribution can be efficiently integrated by a classical algorithm. This precondition is
frequently violated for most high-dimensional distributions.
Zoufal et al. [155] present a hybrid quantum-classical algorithm for efficient, approximate
quantum state loading as an alternative to the Grover and Rudolph Method. For this purpose,
they describe quantum generative adversarial networks (qGANs) to learn a generative model
that prepares the desired probability distribution. For this a probability distribution is given
implicitly via data samples. The qGAN is composed of two components, the generator, which
is a parametrizable quantum circuit and the discriminator, a classical neural network. Mea-
suring the result of the quantum circuit corresponds to a sample from the distribution that is
induced by the circuit. The parameters of the classical network are tuned to distinguish these
generated samples from the true data samples. Based on this classification, the parameters of
the quantum circuit are tuned to output to improve the quality of the learned distribution.
During training, the qGAN learns a low-dimensional representation of the probability distri-
bution underlying the data samples and loads it into a quantum state. The loading requires
O(poly n) gates and can thus enable the use of quantum algorithms, such as quantum ampli-
tude estimation, which require to load a specific distribution. The idea of qGAN distribution
learning and its loading method have been shown to work in simulations as well as in actual
implementations on superconducting quantum processors.

4.2.3 Quantum Noise Modelling
Harper et al. [156] focuses on the output distribution that is observed through measure-
ments and the noise that affects the accuracy of those measurements, instead of manipulating
the input distribution.
When building large-scale quantum computers, Noise arising from various sources is
the main obstacle. Quantum systems with sufficiently uncorrelated and weak noise are re-
quired for solving large real-world problem instances. Even though there has been substantial
progress in designing hardware specific error correcting codes, as well as improved measure-
ment and qubit hardware, continued progress depends on the ability to characterize quantum
noise faithfully and efficiently with high precision[157].
In [156], the parameters of a classical probabilistic graphical model are tuned such that
the model reproduces the noise distribution of a superconducting quantum processor. The
model puts out an estimate of the effective noise and can detect correlations within arbitrary
sets of qubits. It can also be applied to understand how the noise between pairs of qubits cor-
relates. Visualization can be generated to discover long-range dependencies between the noise
of specific qubits. In fact, the method revealed previously unknown error correlations within
the device that was used for the experimental evaluation. The method is the first implemen-
tation of a provably rigorous and comprehensive diagnostic protocol capable of being run on
real-world quantum processors, according to the authors. It builds the foundation for calibra-
tion in the presence of cross-talk, bespoke quantum error-correcting codes, and customized
fault-tolerance protocols that can greatly reduce the overhead in a quantum computation.
4.3 Quantum Enhanced Machine Learning
From the point of view of machine learning practitioners, the idea of quantum enhanced
machine learning is arguably among the most exciting aspects in the broad field of quantum
machine learning. It deals with the use of quantum computing algorithms for solving com-
putationally demanding learning tasks, i.e. with the processing of classical data on quantum
devices to realize intelligent systems.
The expectation is that quantum speedup will make it possible to considerably accelerate
learning processes or even tackle problems which are still beyond reach even for current su-
per computers. Indeed many common learning tasks involve linear algebra routines on very
large systems of equations or optimization or search problems for which it seems likely that
quantum advantages can be realized.
Since worldwide research on quantum enhanced machine learning has noticeably in-
tensified over the past couple of years, the literature has already become vast, and numerous
quantum methods and algorithms have recently been reported (see, for instance, the intro-
ductory papers by Dunjiko et al. or Biamonte et al. [46, 47] and the references therein). The fol-

lowing survey will thus begin with a broad overview and then review quantum algorithms for
several specific machine learning problems where quantum solutions appear to be auspicious.
Aïmeur et al. [158] were among the first to address quantum information processing (QIP)
for machine learning. They review QIP concepts and investigate novel learning tasks that run
within environments where information is fundamentally quantum mechanical. They illus-
trate their idea by using the case of quantum data set clustering and providing examples of
possible quantum clustering algorithms.
Riste et al. [159] specifically treat problems for which there exists a proven quantum ad-
vantage. That is, they consider rather didactic learning problems whose solution is expensive
on classical computers but can be efficiently obtained on quantum computers. According to
the authors, most such problems involve the repeated use of an oracle.
At this point it seems warranted to clarify the notion of oracles because many quantum
computing algorithms posit their existence. We therefore note that, while it is often difficult
to determine a solution to a given problem, it is often also simple to verify if an alleged solu-
tion really solves the problem at hand. Consider for instance the problem of prime factoriza-
tion. While it requires some effort to determine that 1, 2, 3 and 7 are prime factors of 42, it is
comparatively easy to verify that they are. By the same token, it is also easy to verify that, say,
1, 5 and 13 are not the prime factors of 42. A function that easily accomplishes such a verifica-
tion for a given problem is called an oracle. Just for completeness we also note that for many
if not most computational intelligence problems oracles do not exist. For instance, in chess
it is generally difficult (if not impossible) to determine the single best move for a given game
state. Contrary to prime factorization, it will generally be also at least as difficult to verify if an
alleged best move is indeed optimal.
Riste et al. observe that the costs incurred by an oracle-base quantum algorithm can be
determined by its querying complexity, i.e. by the number of oracle calls needed to find a so-
lution with a given probability. While showcases of oracle-based quantum algorithms have
already been verified experimentally on various platforms with different physical realizations
of logical qubits, the authors note that these showcases typically involved problems that could
be modeled in terms of very few qubits. Based on toy scenarios like these, however, it is hard
to argue for quantum supremacy because classical algorithms, too, could solve problems like
these with a few queries to an oracle. The authors therefore investigated the performance of
classical and quantum approaches in solving a more demanding oracle-based problem known
as learning parity with noise. Using a custom made five-qubit superconducting processor (con-
sisting of superconducting qubits) as well as classical Bayesian computing, they observed a
query complexity in favor of quantum processing which substantially increases depending on
the acceptable error rate and the problem size. Their important major finding is thus that a
significant quantum advantage can already arise even in present day noise-intensive systems.
While general results such as these are encouraging, one has to keep in mind that current
NISQ devices can only realize few logical qubits and still suffer from limited coherence times.
That is, even if the findings in [159] suggest that limited fault tolerance may not be the most

pressing technical challenge in quantum machine learning, certain quantum learning algo-
rithms may still not be practical. O’Quinn and Mao [160] emphasize this and point out that
existing platforms such as IBM Q, IonQ, Rigetti Computing, D Wave, Microsoft Quantum and
Google Quantum do not yet allow for practical implementations of some of the more fanciful
ideas discussed in the quantum machine learning literature.
4.3.1 Quantum Algorithms for Linear Algebra
One of the reasons for why machine learning models often assume input data to be en-
coded in terms of data points x ∈ Rm is rather prosaic and due to technological circum-
stances: as long as the dimension m of a given vector is not too large, conventional digital
computers can easily perform linear algebraic operations. Indeed, there exist venerable soft-
ware libraries such as BLAS and LAPACK which date back to the 1970s, compile on almost any
operating system, and allow for highly efficient linear algebra computations. Moreover, mod-
ern graphics processing units (GPUs) allow for rapid, parallel memory access and are specif-
ically designed to facilitate matrix vector operations. In a sense, numerical computing tech-
nology and machine learning co-evolved and it is therefore not surprising that many machine
learning algorithms involve matrix vector products, matrix inversion, or the decomposition of
matrices into factors of lower rank.
Against this backdrop it seems reasonable to attempt to harness potential quantum ad-
vantages for machine learning sub-routines since quantum computing is all about matrices of
size 2n × 2n acting on 2n dimensional vectors.
Harrow, Hassidim, and Lloyd [119] developed a quantum algorithm for solving systems
of linear equation which can be written in terms of sparse and well conditioned matrices. This
algorithm is considered to be one of the more fundamental, presently known methods with
provable quantum speedup over their classical counterparts. It thus falls within the same cat-
egory as Shor’s factoring algorithm (based on Coppersmith’s quantum Fourier transformation
[161]), Grover’s search algorithm, or quantum random walk algorithms [162, 163] and is by
now simply known as the HHL algorithm.
Harrow, Hassidim, and Lloyd are specifically concerned with solving Ax = b for x where
the n × n matrix A is Hermitian and b is an n-dimensional unit vector. They then suppose an
amplitude encoding |b⟩ of b (for instance using the method in [164]) and a Hamiltonian oper-
ator exp(iAt) and perform quantum phase estimation [113, 114], ancilla bit rotation, inverse
quantum phase estimation, and measurement to obtain an estimate of the eigenbasis of the
coefficient matrix in which it is easy to invert it.
If the coefficient matrix of the considered linear system of n variables is sparse (not
densely populated) and has a low condition number k and—crucially—if users are mainly
interested in the result of a measurement on the solution vector rather than in the solution
vector itself, then HHL has a runtime of O(log n · k 2 ). This constitutes an exponential

speedup over the fastest classical algorithms which generally run in O(n · k). Importantly,
improvements over the originally proposed approach are possible if phase estimation is
replaced by the “linear-combination of unitary method” which approximates matrix inverses
via Fourier- or Chebyshev-series expansions [165]. Moreover, since the HHL algorithms has
been verified experimentally, it is known to work on existing quantum computers, at least for
small problems which meet its prerequisites [166, 167, 168].
The singular value decomposition, also dubbed a “singularly valuable decomposition”
[169], is one of the most important matrix factorization techniques with numerous practical
applications in science and engineering. Assuming that a universal quantum computer with
access to a QRAM was available, Rebentrost et al. [170] present a quantum singular value de-
composition which is exponentially faster than its classical counterparts. Making less wide
reaching assumptions, Gilyen et al. [117] propose a quantum singular value “transformation”
algorithm that is based on the idea of qubitization [171] and computes a bounded polynomial
approximation of the singular value decomposition. Note that the term qubitization refers to
a technique for simulating the time evolution operator exp(iAt) that features prominently
in the HHL algorithm. For this, Low et al. [171] prove that it can be done with low error using
comparatively simple quantum random walk circuits.
Gilyen et al. observe that their model leads to rather simple quantum circuits which tend
to involve only constantly many ancilla qubits. From a general application point of view, their
method allows for estimating Moore-Penrose pseudo-inverses and thus applies to regression
problems. Moreover, the authors also observe that their quantum singular value transforma-
tion leads to a unified framework for several quantum algorithms. Methods that can be seen as
special cases of their approach include fixed-point amplitude amplification, robust oblivious
amplitude amplification, and certain quantum walks. Yet, overall the work in [117] is of mostly
theoretical nature; attempts of a practical implementation are not reported even though the
authors provide a detailed discussion of how to implement quantum linear algebra methods
by representing matrices in terms of unitary circuits and vectors in terms of quantum states.
Clader, Jacobs and Sprouse [172] generalize the HHL algorithm. They propose a state
preparation routine which is capable of initializing generic states and they integrate a quan-
tum compatible preconditioner, which enlarges the problem space for which solutions can be
obtained with an exponential speedup over classical linear systems solvers. The authors ver-
ified their algorithm’s applicability by letting it compute the electromagnetic scattering cross
section of an arbitrary target. As a result, it was found to be exponentially faster than the best
classical algorithm.
Huang, Bharti and Rebentrost [173] deal with near-term quantum algorithms for linear
equation systems of the form Ax = b and examine the use of variational algorithms. As part
of their research, the authors develop a near-term algorithms, which is founded on the clas-
sical combination of quantum states (CQS) method. They conducted experiments of solving
large linear systems by simulating the quantum algorithm on a classical computer and report
the approach to work well.

Subasi, Somma and Orsucci [174] introduce two quantum algorithms based on evolu-
tion randomization, a simple variant of adiabatic quantum computing, to prepare a quantum
state |x⟩, which is proportional to the solution of the linear equation system Ax = b. Both
algorithms are easy to implement and designed using families of Hamiltonians that are linear
combinations of products of A. A quantum oracle is supposed to be given that, for any row of
A, outputs the nonzero matrix elements and their indices. Given this prerequisite, their algo-
rithms exhibits an exponential quantum speedup.
To solve a system of linear equations, Lee, Joo and et al. [175] introduce a hybrid quantum
algorithm built on the HHL algorithm. It reduces the circuit depth from the original algorithm
without accuracy loss of the results. On the contrary, the authors’ experiments (using 4 qubits
by IBM Quantum Experience) produced better results, that is a higher accurate performance
on specific systems of linear equations.
Bravo-Prieto et al. [176] take as a starting point the limitation that existing quantum
solvers of linear equation systems are hardly implementable due to the required circuit depth.
The authors introduce a variational quantum linear solver (VQLS), which is a hybrid algorithm
designed for near-term quantum computers. This solver seeks to variationally prepare |x⟩
such that A|x⟩ ∝ |b⟩ and Bravo-Prieto et al. then derive a termination condition guarantying
the achievement to a desired precision. The authors verified their algorithm using Rigetti’s
quantum computer.
Xu et al. [177] take up the challenge of a quantum algorithm’s high demands on the cir-
cuits depth (leading to a high demand on the quantum device’s universal fault-tolerance) to
solve linear algebra tasks. The authors developed variational algorithms, which are compatible
with noisy intermediate-scale quantum devices. They demonstrate that the linear equation
system’s solutions and matrix-vector multiplications are translatable as the ground states of
the constructed Hamiltonians. They implemented their algorithm using the IBM quantum
cloud services and observed a high solution fidelity of 99.95%.
4.3.2 Quantum Algorithms for Regression
The term regression analysis refers to a broad spectrum of statistical methods for esti-
mating relationships between a dependent variable and one or more independent variables.
While there exist many different algorithms for fitting a regression model to data, the arguably
most well known approach is least squares regression. At its heart, this is a simple linear tech-
nique which nevertheless allows for dealing with non-linear models. The basic setting is as
follows: Given a data matrix X ∈ Rm×n whose columns represent (possibly non-linearly
transformed) data points and a target vector y ∈ Rn , the task is to identify a weight vector
w ∈ Rn such that the loss ∥X ⊺ w − y∥2 is minimal. This is can be understood as an almost
trivial supervised learning problem since its unique closed form solution is known to be given
by w = (XX ⊺ )−1 Xy. Moreover, for the practical computation of this expression there ex-
ist fast and numerically stable gradient descent schemes. Nevertheless, because least squares

regression is such a fundamental technique in data science and machine learning, it has at-
tracted the attention of quantum computing researchers.
Initially, work in this direction was mainly theoretical. For instance, assuming the exis-
tence of a universal quantum computer which when queried about an element in a given row
of a sparse matrix X, produces a quantum state that encodes the column number and, more-
over, is capable of producing copies of an input state |y⟩, Wiebe, Braun, and Lloyd [178] present
an HHL-based algorithm that estimates |w⟩ up to scale.
Schuld, Sinayskiy, and Petruccione [179] address certain shortcomings of this approach.
In particular, they extend it towards dense matrices and improve on the dependency on the
condition number. Their idea is to create amplitude encoded quantum states that represent
the singular value decomposition of the data matrix and then to run phase estimation to in-
vert the unknown singular values. They acknowledge that the problem of preparing quantum
states which encode classical data is non-trivial and assume that states are either given in a
quantum random access memory (QRAM) from which the algorithm could access them or
that there is an oracle that loads data into an entangled register.
Some of the computational steps in [179] apply ideas which were first brought forth by
Wang [180] who presents a quantum algorithm for fitting a linear regression model and fo-
cuses on estimating the quality of the fit and on assessing whether the given data set is suitable
for quantum regression at all. This allows his algorithm to output optimal parameters in clas-
sical form. However, Wang, too assumes availability of a QRAM which might not be techni-
cally feasible in the foreseeable future.
Yu et al [181] introduce an improved quantum algorithm for ridge regression, a regu-
larized version of ordinary least squares regression. Ridge regression allows for more robust
estimates but comes with the need of having to estimate additional hyper-parameters. The
authors develop a quantum algorithm for ridge regression that utilizes the parallel Hamilto-
nian simulation technique for simulating Hermitian matrices in parallel and for implement-
ing a quantum k-fold cross-validation that estimates the ridge regression’s predictive perfor-
mance. Yu et al. compose their algorithm in two stages: Stage 1 searches for a suitable hyper-
parameter α that optimizes the predictive performance of the regression model by running
quantum cross-validation. Stage 2 creates a quantum state which encodes the optimal fitting
parameters of ridge regression with such an α, that is later used to predict new data. Again, a
major (practical) weakness of the proposed method is that the authors quantum encoded data
to be available in a QRAM.
Hou et al. [182] make such an assumption, too. The authors propose a quantum partial
least squares regression algorithm. Partial least squares regression is yet another variant of the
ordinary least squares approach which allows for addressing multiple correlation problems.
The authors develop a method for quantum eigenvector search in order to accelerate the re-
gression parameter selection and propose a density matrix product to avoid multiple QRAM
queries for constructing residual matrices. If a QRAM were available, this algorithm would run
exponentially faster than classical partial least squares estimation techniques.

Liu and Zhang [183] further improve on techniques such as the above. To this end, their
algorithm takes statistic leverage scores of a data matrix into account. They show how to gen-
erate a quantum state proportional to |w⟩ in O(log n) time for sparse and well-conditioned X.
Again, they assume efficient quantum access to classical data was possible.
Gilyen, Song, and Tang [184] rightfully point out that many QML algorithms for low-rank
regression require input to be stored in a QRAM especially those that theoretically achieve
runtimes that are poly-logarithmic in the dimension of the data under consideration. As a
remedy, they describe a classical algorithm for linear regression that borrows from the HHL
algorithm and improves on the method due to Chia et al. [116]. Their stochastic gradient al-
gorithm exploits recent numerical sparsity techniques for fast eigenvector computation and
regression analysis and though not designed as a quantum algorithm point towards possible
genuine quantum solutions.
Finally, Date and Potok [185] recently showed how to perform linear regression on adi-
abatic quantum computers. In order to accomplish this, they cast the regression problem as
a quadratic unconstrained binary optimization problem which they then solve on a D-Wave
2000Q adiabatic quantum computer. While this approach does not suffer from any (as of yet
technically unrealistic) assumptions as to data encoding, it incurs a certain loss in numerical
precision. However, the authors compare their results obtained on a D-Wave machine to those
resulting from classical numerical solution implemented in Python running on desktop com-
puter equipped with a current multi-core Intel processor. Their experiments reveal that the
adiabatic quantum implementation achieves up to 2.8 fold speedup over the digital imple-
mentation but is on par with respect to the regression error metric.
4.3.3 Quantum Algorithms for Clustering
Prototype-based clustering is an unsupervised machine learning problem where mini-

mization of a loss function allows for partitioning a set of n data points into k ≪ n clusters
which are defined in terms of representative elements. The arguably most well known in-
stance of a prototype-based clustering method is k-means clustering where the n data points
xj ∈ Rm are Euclidean vectors and the k prototypes µi ∈ Rm are the centroids of their respec-
tive clusters. These are typically found through an iterative minimization of an variance-based
loss; alas, while this loss is commonly considered to be intuitive, its minimization actually
poses an NP-hard problem even for the simple case where k = 2 [186]. Popular classical k-
means clustering algorithms such as those due to Lloyd, MacQueen, or Hartigan are therefore
but heuristic approaches to the problem and there is no guarantee that they will determine
the optimal solution.
However, for k = 2, k-means clustering becomes a bipartition problem that can be cast as
a QUBO and thus be solved via adiabatic quantum computing [73]. The corresponding Hamil-
tonians are derived from a reformulation of the conventional k-means objective which is
based on Fisher’s analysis of variance and leads to an equivalent criterion that also allows for

kernel k-means clustering via adiabatic quantum computing [187]. In both cases, the adiabatic
quantum optimization procedure performs an exhaustive search over the space of all possible
solutions and does so quadratically faster than classically possible. While the validity of the
ideas in [73, 187] was originally verified in simulation experiments, the approaches have mean-
while been implemented on D-Wave computers and were found to be practically feasible [188,
189, 190].
If the QUBO in [73] is modified appropriately, the approach also allows for the estimation
of k-medoids, i.e. cluster prototypes that approximate cluster means [191] and the resulting
algorithm has been successfully applied to solve cardinality-constrained index-tracking prob-
lems in the financial industry [192].
The work in [73] also shows how to extend adiabatic quantum bipartition clustering to-
wards non-numeric data for which one can define similarity- or distance measures. Given
such measures, relations among non-numeric data points (such as text strings) can be mod-
eled in a graph and the bipartition clustering problem becomes a minimum graph cut prob-
lem. Hamiltonians for this general problem class are algebraically identical to those for k = 2-
means clustering and the approach is practically viable.
A remaining practical challenge with these kind of quantum algorithms for bipartition
clustering is that they require as many logical qubits as there are data points to be clustered.
While this is no different on digital computers, quantum computers that could manipulate as
many logical qubits as modern, large scale data sets would require, do not yet exist.
Aïmeur, Brassard, and Gambs [193, 194] discuss how quantum gate computing can speed
up unsupervised learning algorithms used in data clustering. Specifically, they propose quan-
tum versions for minimum spanning tree clustering, divisive clustering, and k-medians clus-
tering. The basic idea is to consider an oracle which provides knowledge as to distances be-
tween data points and to use this oracle in algorithms based on Grover’s phase amplification.
Tomesh et al. [195] observe that quantum machine learning often assumes classical
data to be encoded in term of quantum states in superposition which, due to the difficulty
of known encoding schemes, can annihilate potential quantum speedup. They therefore
investigate a coreset-based approach towards prototype-based clustering with reduced load-
ing overhead. Coresets are an important concept in machine learning as they can allow for
approximating large sets of data points in terms of small representative sets that can quickly
be determined (often in linear time). Given a coreset for a problem under consideration,
Tomesh et al. apply the quantum approximate optimization algorithm (QAOA) [84] and run
numerical simulations to compare the their approach against classical k-means clustering.
Their results indicate that there exist data sets where QAOA might outperform standard
k-means clustering on a coreset. However, the authors point out that, for their method to
show a quantum advantage over conventional k-means algorithms, problems have to be
rather special in that they have to have appropriate coresets. This is, because their method
crucially hinges on the existing of a characteristic coreset which may not exit in general.

4.3.4 Quantum Algorithms for Nearest Neighbor Search
In what follows, the term nearest neighbor search is used to refer to the following gen-
eral setting: Given an unstructured set {xi }ni=1 ⊂ Rm of prototypes and an input data points
x ∈ Rm , determine which prototype is closest to the input, i.e. solve x∗ = argmini ∥xi − x∥.
Given the above discussion, it is clear that this problem occurs in the context of assigning data
to clusters. Moreover, if the prototypes are labeled, nearest neighbor search also allows for
nearest neighbor classification in that input data points x can be automatically labeled using
the label of their respective closest prototype x∗ .
Classically, naïve nearest neighbor search requires efforts of O(n). However, there exist
specific data structures such as k D-trees which allow for nearest neighbor search in O(log n).
This gain comes at the cost of having to pre-process the prototypes in order to structure them
correspondingly. This typically requires efforts of O(n · log n) which amortize if the proto-
types have to be searched repeatedly, for instance, in data base search scenarios. While nearest
neighbor search is thus not too demanding a problem for classical computers, it is interest-
√
ing to observe that quantum computing can, in principle, accomplish search in O( n) even
without data pre-processing.
Wiebe et al. [196] stress the importance of nearest neighbor search in clustering and clas-
sification and present quantum nearest neighbor approaches for binary classification and k-
means clustering. In essence, their approach allows for quantum computation of Euclidean
distances between sparse data points and assumes the availability of quantum oracles that
can efficiently look up the j-th value of prototype xi and identify the l-th non-zero value in
prototype xi . They furthermore assume that is is possible to prepare a quantum state that en-
codes ∥xi − x∥ and present a quantum circuit for this purpose. They then apply a classical
quantum minimum search algorithm due to Dürr and Høyer [197] which itself is a variant of
Grover’s amplitude amplification procedure for searching with oracles [37, 38]. They prove
that the number of oracle queries of their method scales as O(n · log n) which is thus as fast as
informed classical search. In numerical simulation experiments on common machine learn-
ing benchmark data they find that their technique is robust to noise arising from coherent
amplitude estimation, performs well, and asymptotically outperforms classical sampling tech-
niques for nearest neighbor search.
Llyod et al. [198] are concerned with the problem of k-nearest neighbor search. Whereas
naïve classical solutions to this problem would require efforts of O(kn), the quantum algo-
rithm proposed in [198] has a runtime complexity of only O(log kn) and thus achieves expo-
nential speedup. However, the authors once again assume that classical data has efficiently
been loaded into a QRAM so that their algorithm can access it in a quantum parallel manner.
They then proceed to prepare states that represent sub-norms of the input data on which they
run adiabatic quantum optimization to identify and select the k nearest prototypes to a given
query.
In a recent contribution, Basheer et al. [199] integrate ideas from [196] and [198] with re-

cent techniques for preparing data into amplitudes of quantum states [200] and present spe-
cific quantum circuits for k√-nearest neighbor search in classification. Their algorithm has a
runtime complexity of O( kn) but, crucially, does not make any assumptions as to the avail-
ability of QRAMs. In numerical experiments, in which the authors simulate classification
problems with few classes using only few qubits, they observe the approach to work well.
They also point out that their methods seamlessly applies to downstream processing in
quantum computing settings. That is, the proposed algorithm can be directly used on quan-
tum data and thus circumvent the need of reconstructing quantum states using measure-
ments on an ensemble of identical states in physics applications.
Concerned with a much simplified nearest neighbor setting, Schuld et al. [201] present a
distance-based classifier which they realize in terms of a simple quantum interference circuit.
Excluding a simple state preparation procedure, the circuit consist of only one Hadamard gate
two single-qubit measurements units. It computes distances between input data points and
two prototypes in quantum parallel and has been implemented in the IBM quantum experi-
ence environment.
4.3.5 Quantum Algorithms for Classification
Pattern recognition or classification is the problem of assigning labels (signifying classes

or categories) to observations of objects (represented in terms of data points) and a function
or an algorithm that accomplishes classification, especially in a specific scenario, is called a
classifier.
If a classifier can classify objects from two classes (say pictures of cats and dogs), it is said
to be a binary classifier. A classifier that can classify multiple classes (say pictures of cats, dogs,
mice, horses, elephants, …) is called a multi-class classifier. Since any multi-class classification
problem can be addressed by several one-versus-all classifiers, binary classification is the more
fundamental task and the problem of training a robust and reliable binary classifier using la-
beled training data is one the most important problems in machine learning.
There exist numerous possible models that allow for binary classification. Some of these
are probabilistic (naïve Bayes classifiers or Bayesian networks), some operate on categorical
data (decision trees), and some are geometric or linear algebraic in nature (least squares classi-
fiers, linear discriminant classifiers, support vector machines, etc.).
A simple binary linear classifier for input x ∈ Rm is a threshold function of the following
form
y = f (x | θ) = sign(x⊺ w − b) (34)
whose parameters θ = {w, b} are a projection- or weight vector w ∈ Rm and a threshold-

or bias value b ∈ R. While the internal computations x⊺ w − b of this model are indeed lin-
ear, the function sign is said to be a non-linear activation function which produces outputs in

{−1, +1} and thus allows for binary decision making. It is also common to consider activa-
tion functions such as the hyperbolic tangent which produces outputs in (−1, +1), the logistic
function with outputs in (0, 1), or rectified linear units with outputs in [0, ∞). The choice of
activation can impact the ease of training but the fundamental problem in each case is to de-
termine suitable w and b. Depending on the choice of learning algorithm (least squares train-
ing, linear discriminant training, support vector training, etc.) this can be more or less chal-
lenging but is well understood in general and computationally not too demanding.
Binary linear classifiers learn linear decision boundaries which divide the underlying data
space into two disjoint half-spaces, namely {x | f (x) ≥ 0} and {x | f (x) < 0}. However, it
is interesting to note that many linear classification models allow for invoking the kernel trick
[202]. This way, data can implicitly be processed in very high to infinitely dimensional kernel
Hilbert spaces. This then makes it possible to use simple, linear models in order to learn highly
non-linear decision boundaries.
Since linear classifiers (kernelized or not) rely on linear algebraic computations, they of-
ten allow for corresponding quantum computing algorithms. Early attempts in this direction
were made by Schuld et al. [203, 204]. In [203] they argue that quantum computing can outper-
form classical techniques in the case of ambiguous input patterns. They develop a quantum
nearest neighbor classifier that involves Trugenberger’s quantum algorithm for measuring
Hamming distances in quantum associative memories [205, 206]. In simulation experiments
on standard benchmark data, they find their method to work well. Its runtime behavior is
similar to classical implementations but the authors remark that, if there was an efficient O(n)
approach for constructing the required quantum superposition states or if a QRAM would ex-
ist, their quantum algorithm would be independent of the number of training data, a feat that
seems impossible for the classical counterpart.
In [204], Schuld et al. present a quantum perceptron, i.e. a quantum circuit that realizes
the binary linear classification
� function
in (34). They propose quantum phase encoding of
φ = x⊺ w as |φ⟩ = √12 |0⟩ + ei φ |1⟩ and then use the phase estimation algorithm [113,
114] to determine the corresponding output |y⟩. In terms of runtime or ease of implementa-
tion this does not seem to offer any benefits over the classical approach; however, the authors
point out that their quantum perceptron yields outputs which contain quantum information.
This could be used for superposition-based learning procedures in which training data enter
the algorithm as a superposition of data points which would then allow for quantum parallel
processing.
Wiebe et al. [207] ask if quantum computation could lead to non-trivial improvements
in the computational and statistical complexity of perceptrons and present algorithms which
answer both these questions affirmatively. To this end, they consider the notion of the ver-
sion space of a set of training data for a binary classification problem. This term refers to the
set of all hyper-planes that perfectly separate the two classes under consideration. Seen from
the version space perspective, the problem of classifier training can be cast as a search prob-
lem which, in turn, can be solved using Grover’s quantum search algorithm. Based on these

ideas, Wiebe et al. develop two algorithms with different speedups for perceptron training.
√
The first trains in O( n) where n denotes the number of training data points; the second
� −1/2
trains in O ρ where ρ denotes the optimal margin between the two classes under consid-
eration. As of this writing, it is difficult to imagine how this could be achieved classically. The
work in [207] therefore establishes that quantum computing can speed up perceptron train-
ing and thus improve on one of the most fundamental tasks in machine learning. The authors
thus conclude that the quantum computing point of view can lead to a deeper general under-
standing of machine learning and may lead to models for which there is no classical analogue.
Tacchino et al. [208] have recently shown how to practically implement a simple binary-
valued perceptron on an existing quantum device, namely a 5-qubits IBM quantum computer
based on superconducting technology. In practical experiments, they demonstrate success-
ful classification of 4 bits strings using 2 qubits and of 16 bits strings using 4 qubits. This hints
at an exponential storage advantages over classical methods but the authors concede that
generic, i.e. non binary patterns, may need quantum states or unitary transformations whose
preparation requires exponentially many quantum gates. This, in turn, would eliminate the
quantum advantage of their method. They also acknowledge that current NISQ era devices do
not allow for arbitrary controlled operations but require them to be broken down into oper-
ations involving only single- and two-qubit gates which significantly increases circuit depth
and thus the risk of decohenrence. As a possible future generalization away from binary in-
puts, the authors point to phase encoding similar to the approach in [204].
In another recent contribution, Schuld et al. [209] observe that quantum computing bears
certain similarities to the idea of working with kernel methods in machine learning in that
both, quantum computers and kernel machines, process information in (possibly infinitely
high-dimensional) Hilbert spaces. They argue that this can lead to new ideas for the design of
quantum machine learning algorithms and, in particular, interpret the problem of encoding
classical data in a quantum state as the problem of computing a non-linear kernel function
that maps the data into quantum Hilbert space.
To this end, they associate a quantum Hilbert space with a machine learning feature
space and derive a kernel that corresponds to the inner product of quantum states. They
discuss how this allows for the encoding of computational basis states, amplitude encoding
of data points, and representations of product states and use these insights to devise a pa-
rameterized circuit model of a quantum support vector machine. The circuit is composed
of displacement- and phase gates whose parameters are determined via classical stochastic
gradient optimization. Schuld et al. thus consider a variational quantum training algorithm
that combines quantum computations with classical computations in an iterative feedback
loop. Experimental verification with a simple mini benchmark data set of two-dimensional
data points from two classes suggests that the idea works well in practice.
Similarly, Grant et al. [210] are concerned with the capabilities of present day NISQ de-
vices and point out that hybrid quantum-classical algorithms where quantum computers
run model circuits and classical computers perform statistical loss minimization to train the

model circuit are currently the best strategy for quantum (assisted) machine learning. Build-
ing on concepts first proposed in [211], they therefore propose hierarchical quantum circuits
for binary classification. In particular, they consider parameterized circuits of tensor network
topology. In experiments with tree tensor networks and multi-scale entanglement renormal-
ization (MERA) networks, they determine their parameters classically and then deploy them
on an ibmqx4 machine available within the IBM quantum experience environment. The data
they consider are small subsets of standard machine learning benchmarks and their results
suggest that the method is robust to noise and can achieve high accuracy. Importantly, the
authors stress that their method allows for classifying classical- as well as quantum data.
Finally, Date et al. [190] recently discussed how to cast the problem of training linear sup-
port vector machines for binary classification as a quadratic unconstrained binary optimiza-
tion problem. While naïve classical algorithms would require efforts of O(n3 ) for support vec-
tor machine training with n data points, setting up the QUBO requires only O(n2 ) computa-
tions and its solution could subsequently happen on an adiabatic quantum computer. This is
not demonstrated practically but, in line with their linear regression results in [185], should be
entirely possible. The authors’ mainly theoretical work thus suggests that well established ma-
chine learning baseline models can be trained on adiabatic quantum computers from where
the resulting model parameters can easily be read into digital memories for further processing.
4.3.6 Quantum Boosting
Boosting is a machine learning technique tailored towards classifier training that is rather
easy to implement and known to yield well working systems. There are several variants and
flavors, but the original idea of adaptive boosting (AdaBoost) [212] is arguably still the most
popular approach. Assuming a set of individually weak classifiers, AdaBoost evaluates an ex-
ponential loss to determine which (re-)weighted sum of their outputs provide a strong classi-
fier. The resulting boosted classifier is also known as an ensemble classifier or, if the individual
weak learners are non-linear threshold units such as in (34), a shallow neural network. Shortly
after its inception in the 1990s, classifier boosting led to breakthroughs in rapid and reliable
computer vision [213] and correspondingly trained systems were, for instance, implemented
in consumer cameras to detect smiling faces. Alas, the good performance of boosted classi-
fiers comes at the price of typically extensive training times, as it requires numerous rounds
of training to select a high-performance ensemble among a very large number of individual
learners. Yet, since boosting is essentially a subset selection problem, it has been identified as a
setting for potential quantum speedup early on.
Neven et al. [214] propose adiabatic quantum optimization for boosting. They transform
the originally continuous weight optimization problem into an optimization problem over
discrete variables of low bit depth and consider an adapted quadratic loss function. This allows
them to express boosted classifier training as a QUBO. In experiments with heuristic surro-
gates for quantum hardware as well as with an early D-Wave machine, their QBoost approach

compares favorably to classical AdaBoost. It reduces training efforts and leads to better gener-
alization and faster runtimes because the resulting ensembles are typically found to be small.
However, the authors concede that these advantages are mainly due to their bit-
constrained learning model and not a quantum effect per se. The notable result is thus
that a restricted (and hence implicitly regularized) model that was specifically designed for
implementation on quantum hardware can outperform unrestricted conventional models.
Yet, since solving the underlying discrete optimization problem might be too demanding
for conventional computers if the problem size is large, the work in [214] establishes that
quantum computing enables approaches towards machine learning that would be classically
infeasible.
Pudenz and Lidar [215], too, consider quantum adiabatic evolution for classifier training
in a manner similar to that of Neven et al. However, they also apply adiabatic quantum com-
puting in the application phase of their system and adiabatically evolve strong classifiers iden-
tified during training on a superposition of inputs in order to be able to identify likely anoma-
lous elements in their data space in a quantum parallel manner. The practical application they
consider is the problem of verification and validation of classical software where program-
ming errors (bugs) are the anomalies to be detected. Extensive simulation experiments indi-
cate that their methods works well.
Schuld and Petruccione [216] consider the problem of creating a classifier ensemble to be
a problem of quantum state preparation. A quantum parallel evaluation of individual quan-
tum classifiers on such states allows for estimating their combined or aggregated decision in
terms of a single qubit measurement. The authors argue that their framework allows for expo-
nentially large ensembles and thus for advantages over classical approaches.
As a quantum gate model of a weak classifier, they consider their earlier proposal of a
quantum perceptron [204] and, based thereupon, describe a protocol for preparing states that
represent classifier ensembles. Crucially, this requires estimates of the accuracy of the individ-
ual classifiers which can then be used in a weighting scheme. While the authors are more con-
cerned with the feasibility of quantum ensembles rather than with their potential advantages,
they do point to several possible applications in quantum physics and remark that their ideas
suggest approaches towards optimization-free quantum machine learning. In other words, a
potential benefit of the proposed approach is that measuring large enough ensembles of indi-
vidually weak classifiers can lead to accurate decisions without that the ensemble would have
to be tuned in a training process.
4.3.7 Quantum Support Vector Machines
Kernel machines are an important class of classical machine learning models. The
key idea observation is that many (linear) machine learning methods require access to the
data only through inner products between feature representations of two data points, i.e.

|q0 ⟩ = |0⟩ H H H H
|q1 ⟩ = |0⟩ H H † H † H
Uϕ(x) Uϕ(x) Uϕ(y) Uϕ(y)
|q2 ⟩ = |0⟩ H H H H
|q3 ⟩ = |0⟩ H H H H
Figure 3: Variational 4-qubit circuit for quantum kernel machines. Data points x and y are
both mapped into the 24 -dimensional Hilbert space via parameters of unitary gates Uϕ(x) and
Uϕ(y) .
⟨ϕ(x)|ϕ(y)⟩. The mapping ϕ : X → Rd to the d-dimensional feature space has to be specified

by the user and is crucial for the quality of the resulting model. When the dimension d is
large, the explicit computation of ϕ can be circumvented via a so-called kernel function
k(x, y) = ⟨ϕ(x)|ϕ(y)⟩.
The most prominent incarnation of this type of method is the support vector machine
(SVM) [217]. Classically, the choice of the feature map ϕ(·) or, more commonly, the kernel
function k(·, ·) are either domain specific (for instance, string kernels, graph kernels, tree ker-
nels, image kernels, document embedding kernels, time series kernels, …) or generic (for in-
stance, radial basis function kernels, polynomial kernels, Fisher kernels, …).
In contrast, the feature map of quantum kernel machines is hardware specific. Instead of
relying on problem specific knowledge to construct the feature map or the kernel, the intrinsi-
cally 2n -dimensional Hilbert space of an n-qubit register is utilized to realize the feature map
[218].
A schematic representation of a corresponding quantum gate circuit is shown in Fig. 3.
There, a maximal superposition is prepared and then passed through n-qubit unitaries
which create the feature space transformation of the data. It is important to understand
that data does not enter the circuit in the discrete qubit state space. Instead, data is passed
in form of parameters of universal unitary gates Uϕ(x) and Uϕ(y) representing (parts of)
the high-dimensional feature map. Each classical n-bit binary string is interpreted as one
feature and the corresponding probability amplitudes of the qubit state as feature val-
ues. The actual kernel value k(x, y) is then given by estimating the transition amplitude
†
| ⟨ϕ(x), ϕ(y)⟩ |2 = |⟨0n |Uϕ(y) Uϕ(x) |0n ⟩|2 . Clearly, the specific choice of Uϕ(·) is not fixed and
can be tuned for the application at hand. Obtaining the full kernel matrix for n data points
requires n(n + 1)/2 runs of that circuit. The resulting quantum kernel matrix is then ready
to be used in any kind of kernel machine, for instance, in support vector machines, kernel
regression, or kernelized k-means clustering.
In addition to quantum k-nearest neighbors algorithms, Wittek [219] introduces the con-
cepts of quantum support vector machine algorithms based on a slightly extended formula-
tion of traditional SVMs using least-squares optimization. He indicates that quantum SVMs
have a high generalization performance, even though the extension to general multi-class

problems is expensive.
Wang et al. [220] take advantage of the fact that the least squares support vector machine
algorithm shows high efficiency in processing a small quantity of data for the problem of
trend recognition. To optimize the parameters of the least squares support vector machine
and to establish the curve fitting model, the authors use quantum particle swarm optimization
to realize a faster global parameter search. To corroborate the practicality of their ideas, the
authors perform on a mathematical error analysis.
Rebentrost, Mohseni, and Lloyd [221] demonstrate that support vector machines can
be implemented on a quantum computer such that they achieve logarithmic complexity in
the data dimension and the number of training examples. This constitutes an exponential
speedup over classical sampling algorithms which run in polynomial time. The main com-
ponent of the quantum algorithm developed by the authors is a non-sparse matrix exponen-
tiation technique which inverts the training data inner-product (kernel) matrix. However,
the latter is based on quantum regression methods such as those discussed above. Since such
methods assume that quantum states can either be prepared efficiently or are available for
convenient lookup in a quantum random access memory, the ideas in [221] may be of limited
practical feasibility on current NISQ era devices.
Havenstein, Thomas and Chandrasekaran [222] compared the runtime and accuracy of a
classical SVM against two QSVMs, a kernel-based QSVM and a variational QSVM. The authors
have determined that, for binary classification problems, a QSVM does not provide any sub-
stantial improvement over a classical SVM. However, in some multiclass classification cases
the variational QSVMs exhibit a higher accuracy than classical SVMs. The QSVM multiclass
classifiers are therefore promising as the number of available qubits increases. A similar con-
clusion has been reached by Zahorodko et al. [223] who expect that quantum-enhanced ma-
chine learning is suitable for classifying high-dimensional data sets in cases where especially
binary classification can be processed by single-qubit systems in an efficient way.
4.3.8 Quantum Neural Networks
Artificial neural networks are machine learning models that mimic the way information
is processed in biological brains. They are composed of small interconnected computational
units called neurons which receive weighted input from other neurons. The weighted input
received by a neuron is then typically summed up and subjected to a non-linear activation
function. In other words, individual neurons can typically be thought of as a kind of binary
linear classifier similar to the one in (34).
An important result due to Hornik [224] is that large enough neural networks are uni-
versal approximators. This mathematical statement is to be understood in the sense that they
can approximate any Borel measurable function arbitrarily well . Their general universal ap-
proximation characteristic makes neural networks powerful tools for a wide range of artifi-

cial intelligence problems. Because of this, neural networks had their first boom period in the
1980s when the backpropagation algorithm for their training became available [225]. They
then faded out of popularity because, back then, computers were not yet powerful enough to
perform the extensive computations required in supervised neural network training. Over the
past decade they returned to the limelight and are currently the most common tool in ma-
chine learning.
As the task of training large neural networks is computationally burdensome, quantum
neural networks have become a popular topic among quantum computing researchers. In-
deed, parameterized quantum circuits, i.e. quantum circuits whose gates or unitary operators
come with tunable phase parameters, bear a certain resemblance to layered neural networks.
Moreover, similar to the universal approximation theorem for neural networks, large or com-
plex enough quantum circuits are known to be able to represent any target function with ar-
bitrary precision [129]. Lin, Tegmark, and Rolnick also argue that data sets arising from mea-
surements of physical systems will exhibit symmetry and locality so that there should exist
less than exponentially large models that lead to useful results. Against this backdrop, varia-
tional quantum circuit models are intended to approximate solutions to a task at hand while
also restricting the number of quantum operators and quantum circuit depth.
The proven universal approximation capability of classical neural networks crucially de-
pends on the fact that their synaptic summations are subjected to non-linear activation func-
tions. However, this poses a challenge for quantum neural networks. For instance, when data
is encoded in the amplitudes of a quantum state, one cannot apply an arbitrary non-linear
function without distorting it. Allcock et al. [226] therefore suggest algorithms for training and
evaluating feed forward neural networks which are based on canonical classical feed forward
and backpropagation procedures. Their algorithms rely on a quantum subroutine for approx-
imating inner products between vectors which would yield training times quadratically faster
in the size of the network than the classical counterparts. In addition, the authors assume that
their approach is intrinsically resilient to overfitting because it quantum mimics classical reg-
ularization techniques. However, although the hybrid quantum-classical training procedure
in [226] closely follows classical neural network training, it may not yet be practical. This is
because Allcock et al. assume all sequential steps during their training procedure to be com-
puted classically while quantum operations are only used for estimating the inner products in
these steps. This requires their in- and outputs to be read from- and written to a QRAM which
is impossible on current quantum computers.
Other works approach the aforementioned challenge with measurements. Beer et al.
[227] propose a quantum analogue of a classical neuron from which they construct quantum
feed forward neural networks capable of universal quantum computation. The basic idea is to
consider a quantum perceptron to be an arbitrary unitary operator that maps m input qubits
to n output qubits. Inputs are initialised in a possibly unknown mixed state and outputs in a
fiducial product state, that is, in quantum state one can reliably reproduce with low variabil-
ity. Using fidelity as a cost function, the authors propose a training procedure where percep-
tron unitaries are updated using phase shifts proportional to the loss incurred for a given pair

of network in- and outputs. Regarding this procedure, they note that, if their quantum neu-
rons are organized in layered architecture, updates can be accomplished layer by layer without
having to access the full quantum circuit. They argue that this first of all reduces memory re-
quirements and second of all decouples memory requirements from network depth. In other
words, training efforts scale with the width of the proposed quantum neural network.
Running simulation experiments with quantum neural networks of moderate sizes, Beer
et al. find their approach to generalize well to previously unseen testing data and also to be ro-
bust against noisy training data. A drawback with respect to present day quantum computers,
however, is the assumption that quantum operators acting on arbitrarily many qubits were
available. Given the present state of the art, this is technically not yet possible on existing de-
vices.
Concerned with implementations on NISQ era quantum computers, Mitarai et al. [228]
propose hybrid quantum-classical algorithm for quantum circuit learning. Their framework
realizes task specific learning via iterative parameter tuning in circuits of a fixed depth. A the-
oretical analysis backed by simulation experiments reveals that their scheme can approximate
nonlinear functions and the authors expect that the approach should be implementable on
existing quantum computing devices.
Verdon et al. [229], too, focus on quantum neural network realizations for near-term
quantum computers. In particular, they point out that it generally appears to be difficult to
determine the initial parameters of a quantum circuit such that subsequent hybrid quantum-
classical optimization will quickly converge towards local minima of the loss function under
consideration. They therefore consider the use of classical neural networks which they train
to generate initial parameters for the quantum learning process. They empirically find that
this appears to allow for considerably faster variational quantum optimization in that the total
number of optimization iterations required to reach a given accuracy is reduced considerably.
The authors experimentally verify this for the quantum approximate optimization algorithm
(QAOA) and for the variational quantum eigensolver (VQE) which they apply to problems such
solving a graph max-cut task, the Sherrington-Kirkpatrick Ising model, or the Hubbard model.
In fact, they observe that optimization strategies learned by the classical neural network seem
to generalize across problem sizes. This leads them to expect that it may be possible to train
their system on small problems that can be classically simulated and then transfer it to larger
problems which are classically intractable. This way, the number of costly iterations in varia-
tional quantum-classical optimization for such settings could be reduced.
Indeed, the issue of how to initialize parameterized quantum circuits for quantum neural
network training with hybrid quantum-classical gradient descent algorithms is critical. Mc-
Clean et al. [230] observe that random initializations are popular and commonly considered
by many researchers as they are simple as well as hardware efficient. However, they point to
apparent inherent limitations of this approach when dealing with more than a few qubits. The
authors demonstrate analytically as well as numerically that descent based training of even
reasonably parameterized quantum circuits suffers from vanishing gradients and that there

exist “barren plateaus in quantum neural network training landscapes”. This is to say that,
during gradient based parameter optimization, expected values of observable required for the
process tend to concentrate which causes their gradients to tend to zero. In other words, Mc-
Clean et al. show that there is a high probability for gradient-based variational quantum neu-
ral network training to get stuck in local minima and thus to yield sub-optimal parameteriza-
tions. They argue that this is largely due to technical limitations of current quantum circuits
and has to be taken into consideration when designing larger quantum neural networks.
Concerned with a practical implementation of parameterized quantum circuits on a
trapped ion quantum computer, Zhu et al. [231] consider parameter optimization by means
of particle swarm optimization and Bayesian optimization. Their work contains several note-
worthy innovations. First of all they consider the problem of training generative model cir-
cuits in the spirit of generative neural networks. These are neural networks which learn to
produce new, plausible data points. Often, this happens in encoder-decoder architectures
where training data are first compressed into low-dimensional representations and then de-
compressed into their original form. Known applications range from automatic text gener-
ation, over deep faked images or videos to the synthesis of plausible particle trajectories in
high-energy physics simulations.
Zhu et al. argue that generative models can be expected to benefit considerably from po-
tential quantum speedup and, second of all, describe a quantum circuit that is able to learn to
reproduce patterns in the bars-and-stripes data set, a very simple, low-dimensional, yet com-
mon benchmark for experiments with novel techniques that is attributed to MacKay [20]. Tai-
lored towards this data, they propose a quantum circuit design consisting of parameterized
rotation and entangling operators which they adjust in a hybrid quantum-classical manner. In
the classical parameter optimization steps, they work with particle swarm- and Bayesian opti-
mization methods both of which are well established tools in classical machine learning. They
implement their approach on a custom made quantum computing platform presented in [232]
and observe that the convergence of their quantum circuit towards an optimized model criti-
cally depends on the optimization strategy. Their practical results suggest that it is practically
possible to successfully train high-dimensional universal quantum circuits and that quantum
neural networks may thus play a role in future generative modeling applications.
Leyton-Ortega et al. [233], too, are concerned with generative modeling for the canonical
bars-and-stripes data and further investigate the observation that the performance of hybrid
quantum-classical algorithms seems to depend on the choice of classical optimizer and on
the circuit model. They argue that any conclusive (empirical) statement to resolve this issue
requires to conduct extensive experiments with different optimization algorithms and circuit
designs. To accomplish this in practice, they work with Rigetti’s quantum cloud service and
investigate the practical performance of data-driven quantum circuit training for different
classical solvers and different circuit designs. For the latter, they consider different entangling
qubit connectivity graphs and varying circuit depths. Their extensive experiments reveal that
gradient-free optimization algorithms lead to much better results than gradient-based solvers.
In particular, the former seem to be much less affected by the noise characteristics of existing

quantum computing systems.

Another example of a real world implementation of quantum neurons can be found in
work by Tacchino et al. [208]. They introduce a quantum neuron design where m = 2n dimen-
sional data- and weight vectors are encoded using only n qubits which makes use of the expo-
nential storage advantage of quantum computers. To generate entangled states they consider
hyper-graph states [234] and prepare them using several controlled Z gates. Non-linear out-
puts are realized by means of quantum measurements of ancilla qubits. The authors demon-
strate the practical feasibility of their ideas through an implementation of the n = 2 case on
an IBM Q5 quantum computer and find that their quantum neuron can successfully learn a
simple bars-and-stripes pattern completion task.
Zhao et al. [168] observe that Bayesian methods such as Gaussian processes regression
provide proven and successful machine learning models for reasoning under uncertainty
and that they have recently been adapted to deep learning problems. They also observe that
there exist proposals for quantum Gaussian process regression based on the HHL algorithm
[235] and discuss how this allows for deep network training without backpropagation. Hav-
ing established this connection, the authors propose a hybrid quantum-classical algorithm for
Bayesian deep learning. They show how non-linear kernel matrices can be approximated in
terms of polynomial series and then easily lend themselves to be used as quantum density op-
erators. Using the HHL algorithm, whose very specific prerequisites are met by Gaussian pro-
cess kernel matrices, their quantum subroutine achieves polynomially faster matrix inversion
than classical methods and the authors simulate their algorithm on a Rigetti quantum vir-
tual machine. For a considerably reduced setting with 2 × 2 Gaussian process kernel matrices,
they also present implementations on the Rigetti 8Q-Agave and IBM Q5 quantum comput-
ers. Interestingly, they find that the probability of successful training is much higher on the
IBM architecture which they attribute to its longer coherence times. In particular, the authors
observe the probability of their protocol to train successfully to amount to 89% which is an
encouraging result with respect to future efforts involving larger problem sizes.
Helmholtz machines are a class of neural networks that specifically allow for genera-
tive modelling [236]. They consist of two sub-networks, a bottom-up recognition network
that maps input data to a distribution over hidden variables and a top-down generative net-
work that generates novel data from an instantiation of the values of the hidden variables. The
training of a Helmholtz machine usually happens in an unsupervised manner using an algo-
rithm known as the “wake-sleep algorithm” [237].
Van Dam et al. [238] present a hybrid quantum-classical approach for Helmholtz ma-
chines. Their corresponding parameterized shallow quantum circuit models can be trained
in a gradient-free manner using an optimization scheme based on the wake-sleep algorithm.
The authors implement their system on the Quantum Inspire simulator and evaluate its prac-
tical performance on a reduced bars-and-stripes data set consisting of binary images of size
2 × 2 pixels. For this data they consider Helmholtz machines with four visible neurons and
three hidden ones. They observe that their hybrid quantum-classical algorithm learns better

network parameters than a corresponding purely classical implementation. They emphasize

that their method can efficiently approximate the underlying probability distributions with
only polynomially many evaluations of the quantum circuit and also expect their approach to
be implementable on real quantum computing devices.
Boltzmann machines are yet another type of neural networks, in particular, a probabilis-
tic extension of Hopfield networks. As such it is rather straightforward to conceive of quan-
tum Boltzmann machines; this basically requires the definition of a suitable energy function
whose minimizers constitute appropriate network parameters. Again due to their close con-
nection to Hopfield networks, training can be realized via adiabatic quantum optimization
[239] which, in turn, can be implemented on D-Wave computers [240].
Chen et al. [241] point out the the internal probabilistic states of a Boltzmann machine
can be translated into quantum states which then leads to a purely quantum mechanical in-
terpretation of Boltzmann machines called Born machines. These Born machines are yet an-
other kind of generative model and can be used to represent distributions of classical data
in terms of quantum states in superposition. In this regard, Liu and Wang [242] remark that
quantum sampling should be of lower computational complexity than sampling in corre-
sponding classical implementations. They realize this by means of projective measurements of
qubits and propose a gradient-based quantum-classical algorithm for optimization of a quan-
tum circuit such that the likelihood of generated samples can be estimated. In simulation ex-
periments with bars-and-stripes data they observe their approach can learn data distributions,
especially when deeper circuits are being used.
Coyle et al. [243] consider a restricted subset of Born machines, namely those whose
Hamiltonians correspond to those encountered in Ising models. For these, they show that
there is no efficient classical sampling scheme for the underlying quantum circuits and there-
for propose a quantum-classical algorithm based on gradient descent. An innovative aspect
of this approach is that the authors consider loss functions such as the Sinkhorn divergence
or the Stein discrepancy. In experiments with practical implementation on Rigetti’s forest
platform (using a simulator as well as the Aspen quantum processing unit), the authors find
that their procedure works well and reliably. As a potential future practical application of
their generative modeling method, the authors point to the problem of quantum circuit
compilation.
4.3.9 Federated and Distributed Quantum Machine Learning
In today’s world, where cloud computing has established itself as a highly salable com-
modity tool for the of massive data sets, and where a great variety of data is available too, new
concepts such as federated learning [244] promise further accelerated development. In feder-
ated learning, a central node holds the global model, it receives the trained parameters from
client devices and then aggregates them to generate an updated and improved global model
that is shared to all client nodes [244].

Chen and Yoo [244] deal with this concept in the context of quantum machine learning
and discuss distributed training across several quantum computers. The authors expect a sub-
stantial reduction of training time and an advantage from the data security and data privacy
perspective since the training is performed where the data is located. They demonstrate this
federated training approach in experiments with hybrid quantum-classical machine learning
models (which could potentially be translated to pure quantum machine learning models).
In particular, in simulation experiments, they consider a quantum neural network combined
with a classical pre-trained convolutional model on an image analysis problem. They observe
that their federated training protocol does not sacrifice performance for accuracy (at least in
the considered setting) and conclude that federated quantum machine learning might help to
preserve privacy and to distribute computational efforts across arrays of NISQ devices.
Sheng and Zhou [245] introduced a protocol for distributed secure quantum machine
learning. Their overall goal is to augment classical clients with basic added quantum comput-
ing technology to delegate a remote quantum machine learning tasks to a quantum server in
a manner such that privacy data is preserved preserved. They present a distributed quantum
machine learning protocol for the specific scenario where clients together with a remote
server can assign two-dimensional data points to different clusters. Their protocol is secure
in that it does not leak information relevant to this task and, similar to quantum information
transmission, would immediately realize eavesdropping attempts. The authors suggest that
this protocol could be extended to higher dimensional data and to bid data settings, but, for
now, their work is rather conceptual and an interesting first step into a new direction.
4.3.10 Variational Quantum Algorithms
Above, we already saw several hybrid- or variational quantum computing approaches to-
wards quantum machine learning. Examples included variational approaches towards quan-
tum circuit design [143], quantum state preparation and loading [155], quantum linear system
solving [175, 176], quantum support vector machine training [209], and ideas for the realiza-
tion of quantum neural networks [210, 228, 229, 230, 231, 238]. Indeed, given the technical ca-
pabilities of present day NISQ devices, hybrid quantum-classical solutions, where quantum
computers run model quantum circuits and classical computers repeatedly perform (statis-
tical) optimization to adapt those models to a given task the task, currently appear to be the
best strategy for quantum (assisted) machine learning in a manner that allows for harnessing
quantum supremacy [246].
What all present such approaches have in common is that they involve parameterized
quantum circuits. While these circuits may or may not have been specifically designed for a
task at hand, they typically consist of tunable gates in sequence or in parallel. Most commonly,
parameterized gates perform qubit rotations and the problem is to adjust their parameters
such that the circuit as a whole performs the desired computation with high probability. In
order to convey a better understanding of this general idea, we next discuss two prominent

examples of variational quantum computing algorithms.

A well known and fundamental example of a variational quantum computing algorithm
is the variational quantum eigensolver (VQE) introduced by Peruzzo et al. [85]. Motivated by
the problem of having to estimate the ground state energy of certain chemical molecules, they
consider the general problem of estimating the bottom eigenvalues of Hamiltonian operators.
They observe that traditional quantum algorithms for eigenvalue computation relied on the
phase estimation methods [115]. While these methods promise exponential speedup, they also
generally require exponentially large numbers of quantum gates so that real world implemen-
tations on existing quantum computers are only possible for problems of small size [247].
Peruzzo et al. therefore propose a different approach and begin by observing the classical
result that, for any Hermitian operator H, the Rayleigh quotient
⟨ψ|H|ψ⟩
r= (35)
⟨ψ|ψ⟩
is minimal exactly if |ψ⟩ is the bottom eigenvector of H. Even more, if |ψ⟩ is the bottom eigen-
vector of H, then it is known that r corresponds to the sought after minimum eigenvalue.
Based on this observation, the authors propose to consider a random, parameterized
state |ψ(θ)⟩ and to systematically vary θ to determine the minimizer of (35). In order to effi-
ciently evaluate r on a quantum computer, they represent H as a linear combination of tensor
products of Pauli operators and note that expected values of such operators can be estimated
by means of local qubit measurements only. A considerable advantage their approach has
over phase estimation techniques is that it requires only shallow quantum circuits as opposed
to deep ones. It will therefore be less affected by short coherence times on existing quan-
tum devices. A drawback is that the variational approach require polynomially many execu-
tions of the circuits as opposed to just in the case of phase estimation approaches. For each
of these calls, |ψ(θ)⟩ needs to be prepared and, in order to do this efficiently, the authors pro-
pose to let |ψ(θ)⟩ = U (θ)|ϕ⟩. Here, |ϕ⟩ is an appropriate, easily prepared reference state (for
their molecular energy application, the authors consider the Hartree-Fock ground state) and
U (θ) = exp(−iR(θ)) where R is a rotation that acts on each qubit individually and can be
decomposed into products of two simple Pauli operators. A quantum circuit which imple-
ments these computations can then be called in each iteration t = 1, 2, . . . , T of an outer
loop executed on a classical computer which updates θt+1 = θt − dθt where, for instance,
dθt = r(θt + ∆θ) − r(θt − ∆θ) is an approximation of the gradient ∇r(θt ) of the Rayleigh
quotient r seen as a function of θ. Note, however, that gradient-free optimization schemes
are possible as well. In fact, the latter do not seem to suffer from the phenomenon of “barren
plataeus” [230] but often lead to better solutions than gradient-based schemes [143].
Peruzzo et al. demonstrate the practical feasibility of their method in experiments with
an implementation an a custom made photonic quantum computer. In particular, they suc-
cessfully show that ground state energies of the 4 × 4 Hamiltonian of a helium hydride ion can
be calculated reliably.

The quantum approximate optimization algorithm (QAOA) due to Farhi et al. [84] is an-
other variational method that was specifically developed to determine approximate solu-
tions to combinatorial optimization problems. In the above discussion of adiabatic quantum
computing, we already saw that potential solutions to such problems can often be encoded in
P variables zj which have to meet m conditions ck (z) ∈ {0, 1} and thus maxi-
terms of n binary
mize c(z) = k ck (z). Translating this setting into a formulation involving an n-qubit system
|s⟩ in a 2n dimensional state space, Farhi et al. propose to consider an operator
Y
m
U (C, γ) = exp(−iγC) = exp(−iγCk ) (36)
k=1
where C denotes the overall problem Hamiltonian. They also define an operator
Y
n
U (B, β) = exp(−iβB) = exp(−iβσjx ) (37)
j=1
P
where B = j σjx and σjx denotes the Pauli spin matrix σ x acting on the jth qubit. The next
crucial idea is to introduce a total of 2p angles γ1 , . . . , γp , β1 , . . . , βp and to consider the system
|γ, β⟩ = U (B, βp ) U (C, γp ) · · · U (B, β1 ) U (C, γ1 ) |s⟩ (38)
which, as the authors emphasize, can be computed by a quantum circuit of depth (m + 1) · p.

The expected value of the Hamiltonian C under this “angle” state is Ep (γ, β) = ⟨γ, β|C|γ, β⟩
and the maximum value it can attain is called
Mp = max Ep (γ, β) (39)

γ,β
Crucially, the author prove that, if p → ∞, then Mp → maxz c(z). This then suggests a
variational algorithm for approximately solving the given combinatorial problem: Choose
p, initialize γ1 , . . . , γp , β1 , . . . , βp , run a quantum circuit to compute |γ, β⟩ or to measure
Ep (γ, β), use this observation to classically adjust the current angle parameters, and repeat
until Ep (γ, β) does not improve anymore.
The reason why this idea works is that (38) can be seen as an approximation of an adia-
batic quantum computing procedure with Hamiltonian H(t) = (1 − t/T ) · B + t/T · C [84].
Farhi et al. also consider several special cases where the problem Hamiltonian is of special
structure and thus allows for particularly simple optimization routines. But, in general, the
method is very versatile and can, in principle, be used to approximately solve any optimiza-
tion problem that can be cast as a QUBO. The quality of the solution will depend on the choice
of the parameter p. With respect to implementations on current NISQ devices, this allows for
trading off circuit depth against approximation quality. Better approximations require larger
values of p which lead to deeper quantum circuits which, in turn, may suffer from decohen-
rence. Smaller values of p entail shallower circuits which may be run in practice but also pro-
duce sub-optimal solutions.

Yet, with respect to quantum advantages that can be achieved by running QAOA, the jury
is still out. According to a calculation by Dalzell et al. [248], QAOA would outperform any ex-
isting classical supercomputer for problems involving 420 qubits and 500 constraints. As prac-
tically relevant combinatorial optimization problems go, this is a rather moderate size, how-
ever, quantum computers that could work on problem sizes like this do not yet exist. More-
over, Akshay et al. [249] identified certain inherent limitations of the procedure. In particular,
they found that its capabilities depends on the ratio of the number m of problem constraints
and the number n of problem variables. In combinatorial optimization, it is known that prob-
lems where α = m/n < 1 tend to have several to many solutions whereas problems with
α = m/n > 1 have only few solutions if at all. For the latter kind of problems, Akshay et
al. found that, even for large p, QAOA does not seem to work well. The authors refer to this as
a (solution) reachability deficiency of QAOA and concede that further research is required to
better understand the phenomenon.
Overall, however, variational- or hybrid quantum-classical approaches can considerably
reduce quantum computing resources (circuit depth, coherence time, qubit counts) required
to run quantum machine learning methods in a stable manner. Importantly, while the state
spaces considered in this context are exponentially large, parameterized circuits typically have
fixed structures (mainly laid out by human experts) and the number of their parameters only
scales polynomially with the number of qubits. In fact, a common current design principle
for parameterized circuits is to consider NISQ hardware efficient layouts, i.e. circuits of low
qubit connectivity and with comparatively simple gates. Here, tensor network designs are of
increasing interest but another strategy is to apply variational algorithms to design simple
circuits on which to run variational algorithms.
The appeal of variational quantum algorithms of parameterized quantum circuits to
those who develop quantum neural networks can likely be attributed to the fact that both,
deep neural networks and parameterized quantum circuits, consist of basic computational
units arranged in layered structures. Moreover, both involve the use of classical optimization
algorithms to adjust their parameters to a given, specific problem. Some authors therefore
even refer to parameterized quantum circuits as quantum neural networks per se. However,
there are differences which make this terminology questionable.
First of all, the basic computational units in a quantum circuit are unitary operators
rather than non-linear function as in the case of neural networks. Above, we already discussed
that the capabilities of neural networks crucially depend on the fact that the activation
functions of neurons are non-linear. Non-linear computations in a quantum circuit can be
realized in a manner that preserves coherence, e.g. through entanglement, or in a manner
that destroys coherence, e.g. through qubit measurements. In particular in combination with
ancialla qubits, these operations allow for modelling the behavior of artificial neurons.
Second of all, internal states of a quantum circuit cannot be read out in a manner that
would preserve their quantum nature. Contrary to neural networks, it is therefore impossi-
ble to compute local gradients within a quantum circuit. Consequently, it is impossible to ap-

ply methods like the backpropagation algorithm to train quantum circuits. As of this writing,
there thus do not exist any workable proposals for how to adjust the parameters of a parame-
terized circuit without truly external supervision through an independent (classical) optimiza-
tion algorithm.
4.4 Quantum Machine Learning for Quantum Data
If the prospects of quantum enhanced machine learning for classical data are most excit-
ing for machine learning researchers, then the prospects of quantum machine learning and
quantum analytics for quantum data are likely most exciting for physicists and chemists. This
is because the simulation of physical or chemical systems or processes on quantum computers
is widely seen as another promising application of quantum computing [250]. Such quantum
computing simulations produce quantum data whose subsequent analysis would either re-
quire to measure and read them into classical computers or, preferably, could itself involve
quantum algorithms. Quantum machine learning for quantum data thus refers to the idea of
using quantum machine learning methods to process quantum mechanical data on quantum
devices.
For example, Grant et al. [210] note that their hierarchical quantum circuits for binary
classification can be applied to the problem of quantum state classification. Basheer et al. [199],
too, point out that their quantum circuits for k-nearest neighbor classification can be directly
used on quantum data. Chen et al. [251] particularly focus on the question of how to classify
quantum data and consider the specific problem of distinguishing between pure and mixed
states. To accomplish this, they train a parameterized quantum circuit and report that their
system performs close to optimal on the training data and, more importantly, also general-
izes well to previously unseen inputs. Since they are dealing with the classification of genuine
quantum mechanical data, they conclude that their approach is among the first to successfully
solve a learning problem for which there is no classical counterpart.
The term quantum state tomography refers to the general problem of predicting (prob-
abilities of) outcomes of measurements of hidden quantum states. Traditional approaches
to this estimation problem would require exponentially many measurements but Aaronson
[252] observes that the problem could also be thought of a a statistical learning problem. He
then proves that estimates can be obtained from only linearly many measurements. Together
with a team of coworkers he meanwhile verified this theoretical expectation in practice [253].
Working with a custom made photonic quantum computer, they experimentally demonstrate
this linear scaling behavior for their system in optical systems and conclude that computa-
tional learning theory provides a useful tool for the analysis of information that is of gen-
uinely quantum mechanical nature.
Legeza and Solyom [254] investigate quantum data compression for finite quantum sys-
tems for cases with dependent density matrices. The authors demonstrated a connection be-
tween the entropy of the left or right block and dimension of the Hilbert space of that block.

Romero et al. [255] are concerned with the problem of autoencoding quantum data. They
propose a simple parameterized quantum circuit that can act as an autoencoder neural net-
work, i.e. can compress and decompress its input data. In simulation experiments, they train
their model in a quantum variational manner so that it learns to compress a set of quantum
state data for which classical compression algorithms supposedly do not work. In particu-
lar, they consider the problem of compressing Fermionic wave functions, i.e. eigenstates of
a number operator, of a system of n quantum particles, compare their results to analytically
computed predictions, and find it to work well. Ding et al. [256], too, are concerned with au-
toencoding quantum states and propose a corresponding quantum circuit composed of quan-
tum adders. In an experimental implementation in the Rigetti cloud, they consider states of
three superposed qubits and find their method to produce results that agree well with theoret-
ical expectations.
In classical machine learning, autoencoders are often used as generative models and have
been extended towards more capable architectures such as generative adversarial networks
(GANs). In this spirit, Benedetti et al. [257] propose a system composed of two parameterized
quantum circuits for generative modeling. They then use their system to sample unknown
pure quantum states. Numerical simulation experiments suggest that the approach works well
and the authors foresee applications in quantum state tomography. Similarly, Hu et al. [258]
demonstrate that quantum state synthesis based on generative adversarial learning is practi-
cally possible. They implement their approach on a superconducting quantum circuit and find
that it can be trained to generate high fidelity quantum output from quantum input
In their article “quantum data processing and error correction”, Schumacher and Nielsen
[259] deal with the behavior of noisy quantum information channels. The authors introduce
the concept of coherent information which specifies a (per quantum information processing
non-increasing) quantity measuring the amount of quantum information conveyed in the
noisy channel. This quantity indicates the necessary and sufficient condition for the existence
of perfect quantum error correction. Quantum error correction, in turn, is of importance in
the NISQ era since existing devices suffer from noise, unreliable operations, and decoherence
so that methods for generating fault-tolerant result are sought after.

5 CHARACTERISTICS OF QML METHODS
5 Characteristics of Quantum Machine Learning Methods
The ever faster growing literature on quantum machine learning and the many success
stories reported therein seem to suggest that quantum computing for computational intelli-
gence is just about to break into the mainstream. However, as of this writing, the content of
many reports, including many of those surveyed above, has still to be taken with a grain of salt.
The two major caveats are that reported results might be exaggerated or that reported results
abstract away from technical reality and assume the existence of universal quantum comput-
ers whose hypothetical capabilities far exceed what is possible with current NISQ era devices.
As a testimony to the former, we refer to Aaronson’s article “Read the Fine Print” [260] in
which he critiques what he calls the “quantum machine learning mini-revolution”. In particu-
lar, he presents a critical analysis of what the HHL algorithm [119] for linear system solving is
capable of and contrasts that with how it is commonly perceived by enthusiasts in the public
and the scientific community.
Aaronson observes that HHL is often seen as a quantum algorithm for solving general
linear systems Ax = b in a time exponentially faster than classically possible. That is, if
A ∈ Cn×n and b ∈ Cn , then HHL is often seen as a method for finding x ∈ Cn in a time
proportional to log n rather than to n2 . He then points to the fine print and emphasizes that
HHL is actually a method for creating a quantum state |x⟩ which approximately represents
the sought after solution x. He further emphasizes that the quantum speedup of the method
hinges on the assumption that a state |b⟩ representing b can be prepared efficiently. This en-
coding problem may be easy in certain special cases, for example, if b is a unit vector whose en-
tries have about the same magnitude. In general, however, state encoding will require efforts
at least polynomial in n. Moreover, once |b⟩ has been prepared, it must be coherently subjected
to the operator exp(−iAt) for a period of time proportional to ks/ϵ · log n. Here, k and s mea-
sure condition number and sparseness of matrix A and ϵ indicates the approximation error
users are willing to tolerate. Aaronson finally stresses that not every matrix is well conditioned
and sparse enough to meet the implicit requirements under which the HHL algorithm is guar-
anteed to run efficiently.
Indeed, the general utility of HHL has been scrutinized before. For instance, Childs [261]
observes that producing a quantum state |x⟩ = A−1 |b⟩ is not tantamount to solving the origi-
nal linear system. Rather, he points out the decoding problem and notes that obtaining x from
|x⟩ requires O(n) quantum measurements. This leads him to conclude that, when it comes
to really solving a system of linear equations, present quantum algorithms do not yet have an
exponential edge over classical approaches.
Both Aaronson and Childs note that the inventors of the HHL algorithm acknowledge
all of this. Indeed, the above shortcomings and restrictions might be the reason why Harrow,
Hassidim, and Lloyd [119] remark that there are situations in which we may not be interested
in the actual solution x but only in some summary statistic of it, such as, say, the sum of its
entries. They also stress that, in situations like these, quantum algorithms may work much

faster than their classical counterparts and move on to argue that any possible quantum com-
putation could be understood as an instance of linear equation solving. This argument, in
turn, would either imply that classical computers should be able to efficiently simulate quan-
tum computations (which they are not) or that quantum computers have an inherent linear
equation solving advantage over classical computers. As the latter has not yet been conclu-
sively demonstrated in general, Childs is critical about hardness results such as this one, and
questions their role as a means of arguing for quantum supremacy in applied settings.
Overall, Aaronson concedes that the HHL algorithm and its decedents constitute theo-
retical as well as practical progress in quantum computing. At the same time, he emphasizes
the need to carefully examine conditions or assumptions required for a quantum algorithm to
achieve exponential speedup. Especially if effort required for in- and output or pre- and post-
processing of data are factored in, many quantum computing solutions turn out to be only
polynomially faster than classical algorithms. As Tang’s former supervisor, Aaronson certainly
knows what he is talking about. We therefore recall that Tang famously found a novel classi-
cal algorithm which—after appropriate pre-processing steps—works only polynomially slower
than its quantum counterpart. From this, she concluded that state preparation assumptions in
a quantum computing solution need to be matched against classical pre-processing assump-
tions before any claims can be made about significant quantum speedup.
To add further credence to this conclusion, we mention another example from our own
experience. Recall that reports such as [209] or [226] describe classifiers which involve quan-
tum subroutines for computing approximate inner products between quantum state vectors.
However, if “approximate” inner products between high dimensional vectors are all that is
asked for, we note that there exist classical methods which—after appropriate pre-processing
steps—can compute them in O(1), i.e. in a time independent of the data dimensionality [262].
Especially in cases where inner products need to be evaluated repeatedly, e.g. in nearest neigh-
bor searches, the costs for pre-processing quickly amortize and potential quantum advantages
may become less significant. Granted, techniques such as in [262] appear not to be widely
known but they exist and should be taken into account by quantum machine learning re-
searchers.
The second caveat we mentioned in the introductory paragraph has to do with the com-
mon assumption that universal quantum computers are taken as a given when developing
quantum computing algorithms. Yet, given the current state of technical developments, this
is a far reaching assumption as many of the expected capabilities of universal quantum com-
puters can not be realized by current NISQ era devices. Hence, to asses the actual technical
feasibility of proposed quantum machine learning approaches, we have to check them against
a catalogue of criteria. Important such criteria will be discussed in the following.

5.1 Hardware Requirements
When quantum algorithms are designed, properties of the hardware at hand are usually
not taken into account. This is not surprising, since abstracting the actual hardware away is
at the core of computer science. Fundamental results about complexity, data structures, and
algorithms rely on theoretical models of computation, e.g., Turing machines and other types
of automata. This is clearly sufficient when the correctness of an algorithm, its asymptotic
runtime, or bounds on its memory consumption are being analyzed. In these scenarios, it
is not necessary to bother about the clock rate of a processor, the amount of available main
memory, or the data width of some bus architecture. In fact, theoretical machines possess an
infinite amount of memory. Nevertheless, when algorithms have to run on real computers,
all these constraints materialize and have to be considered for implementation. Algorithms
can be equivalent under a theoretical model while their actual number of consumed clock cy-
cles on a specific processor can differ by a large amount. Hence, one of the algorithms would
consume far more energy and they stop being equivalent in appliances where a steady energy
supply cannot be guaranteed, e.g., in the context of autonomous systems. Moreover, without
specific tuning of an algorithm, state of the art GPUs cannot unfold their potential and many
of their computational capabilities will lie idle [263]. Some of these practical considerations
have found their way back into the theoretical community. One striking example are cache-
oblivious algorithms [264], where the existence of a memory hierarchy is introduced to the
theoretical analysis. Moreover, the explicit consideration of hardware properties also found
its way to machine learning, e.g., machine learning models which are designed to rely only on
integer arithmetic [265].
While the explicit incorporation of computational architectures has led to various suc-
cesses in classical computer science, similar ideas are rather unknown in the quantum ma-
chine learning literature: When QML methods are derived, there is no treatment of particu-
larities of the underlying quantum compute architecture. Authors do neither discuss whether,
e.g., the usage of a Toffoli gate in a quantum circuit is problematic, nor is it discussed if using
O(log n) ancilla qubits leads to a practical relevant method. Especially because practical quan-
tum computing hardware is still in its infancy and resources are scarce, real world constraints
must be considered in order to understand if a specific QML algorithm is viable.
5.1.1 Quantum Memory
Some low-end classical processors like microcontrollers are not equipped with a float-
ing point unit or units for processing wide single instruction multiple data vector operations.
When an algorithm designer assumes the availability of such functional units, the algorithms
are unlikely to run satisfactorily on other processors. The same happens in the design of quan-
tum algorithms: Various QML methods require the availability of some sort of quantum mem-
ory. Given the current technical state of the art, the existence of an exact quantum memory is
at least questionable due to the no-cloning theorem [88]. The theorem asserts that there is no

unitary U for which two non-orthogonal quantum states |ψ⟩ and |ϕ⟩ exist, such that
U (|ψ⟩ ⊗ |x⟩) = |ψ⟩ ⊗ |ψ⟩ and U (|ϕ⟩ ⊗ |x⟩) = |ϕ⟩ ⊗ |ϕ⟩
whereas |x⟩ is some arbitrary state that we want to overwrite with the content from |ψ⟩ or |ϕ⟩.
In short, we will not be able to construct a unitary operator that allows us to copy the content
of arbitrary qubit registers |ψ⟩ and |ϕ⟩ to a target register |x⟩. It is important to understand
that the no-cloning theorem prohibits the existence of a quantum computing system in which
|x⟩ is a working register and |ϕ⟩ and |ψ⟩ are states which are stored in some type of quantum
memory—making it effectively impossible to load data from the memory or to write data into
the memory.
Nevertheless, a whole line of QML research relies on this concept, including quantum rec-
ommendation systems [112], quantum deep convolutional neural networks [266], quantum
gradient descent, methods for solving linear systems, and methods for ordinary least squares
regression [267]. The primary building block of these approaches is the QRAM [268]. While
being frequently cited, fundamental issues like the no-cloning theorem are not discussed in
the paper that introduces QRAM. Instead, the authors refer to “quantum memory elements”
without describing how they can be implemented. Moreover, a three-level quantum system
is required to realize the routing in the so-called bucket-brigade memory architecture. While
this is not a problem per se, commercial quantum computers with hardware for ternary quan-
tum states are not available. Hence, rendering algorithms which rely on these concepts as not
viable. Finally, 2n memory slots are assumed to be available, where n is the number of bits in
the address register. Clearly, in the light of currently available hardware and under considera-
tion of vendor roadmaps, the availability of a large number of qubits is not given at the time of
writing.
Physical implementations of superconducting qubits reside on the chip at fixed locations
and are connected via a well-defined pattern, the so-called connectivity structure. Any spe-
cific structure originates mainly from practical considerations. In IBM Q Systems, the pre-
dominant connectivity is known as Heavy Hexagon structure, shown in Fig. 4 (a). These low-
degree graphs are designed to minimize the possibility of frequency collisions and optimize
the hardware performance within superconducting qubit architectures. The larger the num-
ber of neighbors of a qubit, the more frequencies are required to realize two qubit gates using
cross-resonance interaction. A specialized family of error correcting codes for this structure
exhibits outstanding properties with respect to decoding and frequency collisions [269].
Beside its favorable properties for error correction, the connectivity of a quantum gate
processor has direct impact on the depth of actual quantum circuits. To see this, one has to
consider the transpilation of user specified circuits. During transpilation an input circuit is
compiled to a sequence of native gates such that all operations agree with the connectivity
structure and noise properties of a specific quantum processor. For any non-trivial circuit, a
series of transformations must be conducted to make it compatible with a given target de-
vice, and optimize them to reduce the effects of noise on the resulting outcomes. Due to high-
dimensional noise distributions and limited gate sets, rewriting quantum circuits to match

hardware constraints and optimizing for performance can be far from trivial. Quantum cir-
cuit compilation can have a non-linear control flow and exhibits complex branching patterns.
Most importantly, it encompasses the decomposition of gates involving three or more qubits
into 2-qubit gates. Clearly, the heavy hexagon structure is pairwise and hence contains no
connection between three or more qubits. As a direct result, an apparently “shallow” quantum
gate circuit, consisting of a single unitary operation among 10 qubits, can thus eventually ex-
hibit a high depth. On the other hand, a high depth requires long decoherence and dissipation
times of the system, and thus, might not be viable.
5.1.2 Qubit Connectivity
To understand the impact on QML methods, let us consider quantum kernel machines
from Sec. 4.3.7. The power of the quantum kernel comes from the fact that data is mapped
into a feature space whose dimension is exponential in the number of data dimensions. It is
important to understand that the dimensions of the feature map shall not be independent of
each other. In other words, all qubits that underlie the feature space representation shall be
entangled with each other. For quantum kernel machines [218], this is realized by considering
the following unitary: X Y
Uϕ(x) = exp − i ϕs (x) σzv
S⊂[n] v∈S
Here, ϕs (x) denotes a data dependent angle. The expression S ⊂ [n] denotes all possible sub-
sets of n qubits. We hence conjecture that Uϕ(x) creates an entanglement between all qubits.
However, we know that the process of transpilation will break this down into 2-qubit gates,
and thus, a very deep and most likely not viable circuit. The authors of [218] are, of course,
aware of this particularity and propose to restrict the subsets S to those of size two, hence, tak-
ing the actual hardware structure into account. While this solution is perfectly reasonable,
it greatly limits the expressiveness of the resulting kernel, and thus the accuracy of any QML
method that relies on this kernel construction.
27 33
20 28
26 32
0 1 2 3 4 5 6 7 8 9 25
21 29
34 35 24
10 11 12
13 14 15 16 17 18 19 20 21 22 23
16 17 18
24 19
25 26 27
24 25 26 22 30
4 12 16 17 18 19
36 37
27 28 29 30 31 32 33 34 35 36 37 20
23
5 31
13 21
8 9 10 11
38 39 40
12 22
41 42 43 44 45 46 47 48 49 50 51 0 1 28 39 10 11 0 1 2 3 13 23
28 29 30 31
4 14
52 53 54
6 14 5 15
55 56 57 58 59 60 61 62 63 64
6 38
7 15 7 39
(a) Heavy hexagon, 65-qubit. (b) Chimera (C2), 32-qubit. (c) Pegasus (P2), 40-qubit.
Figure 4: Qubit connectivity overview.

In case of the D-Wave quantum annealer, the Chimera and Pegasus structures define the
qubit neighborhoods. They arise from considerations about a high inter-qubit connectivity.
Formal definitions in terms of edge generating functions of the graphs shown in Figs. 4 (b) and
(c) together with a brief discussion of their graph theoretic properties are provided in [270].
While connectivity seems to be a rather technical property of a quantum annealer, it has an
immediate impact on the problems that can actually be solved on that chip. More precisely,
the structure imposes constraints on the sparsity pattern of the QUBO matrix that encodes the
problem that has to be solved. In Chimera, the qubits can couple to up to δ = 6 other qubits.
In Pegasus, each qubits can couple to up to δ = 15 other qubits. This implies that each row of
the QUBO matrix may only contain δ non-zero entries. Moreover, those entries must reside
at fixed positions within each row. Revisiting the insights from the previous section, many
QML methods that are based on quantum annealing require a completely dense QUBO matrix,
e.g. the clustering method presented in [73]. There, the dimension of the QUBO problem is
equal to the number of data points that have to be clustered and the QUBO matrix itself stems
from the kernel matrix between all pairs of data points, and hence is dense.
Limited connectivity can be mitigated by introducing so-called chains into the prob-
lem’s Hamiltonian. That is, the optimization problem is rephrased such that multiple qubits
are enforced to take on the same value in the system’s ground state, thus effectively increasing
the fan-out of that qubit while reducing the total number of available qubits. To understand
the consequences, consider the D-Wave Advantage 4.1 Quantum Annealer with 5627 qubits.
When a complete dense QUBO matrix is required, chains of 17 qubits are introduced which
results in an effective dimension of 177. The maximum dimension of the problem is hence
reduced by orders of magnitude, severely reducing the number of data points that can be pro-
cessed by the aforementioned kernel clustering approach.
5.1.3 Qubit Count
The number of qubits in real world systems is ever increasing. At the time of writing, at
most 127 qubits are available for quantum gate devices, e.g., the IBM Eagle processor, and at
most 5627 qubits for quantum annealing, e.g., D-Wave Advantage 4.1.
As mentioned in the previous section, the problem structure has a direct impact on the
number of qubits that are actually available to a quantum annealer. When the problem di-
mension exceeds the number of available qubits, the problem is too large to be embedded in
its entirety on the quantum processor. Instead, the problem is sequentially decomposed into
smaller problems, each of which can fit on the quantum processor, with variables selected by
highest impact on the objective function. Clearly, variables cannot be selected solely by maxi-
mal impact on the objective function, since those qubits may not be connected at all. In order
to represent a local structure of the problem, traversal techniques such as breadth-first (BFS)
or priority-first selection (PFS) can capture features that represent local structures within a
problem. Instead of selecting the working set of qubits via their impact, cutset conditioning

can be applied. That is, the values of a subset B of variables (the “cutset”) are fixed, effectively
partitioning the set of variables [n] = A ∪ B with A ∩ B = ∅. Thus, the QUBO problem
argmins∈{−1,+1}n Q(s) = s⊺ Q s + q ⊺ s becomes
argmin Q(sA , sB )
sA ∈{−1,+1}n
Finding the optimal setting for sA while values in sB are fixed breaks the problem into sep-
arate components that can be solved independently. When A is chosen carefully, each inde-
pendent problem is small enough to be solved on the quantum annealer. This procedure has
to be repeated multiple times, choosing different subsets A. The general procedure is outlined
in [271].
QUBO problems can indeed be solved on quantum gate computers, too. Based on the
Variational Quantum Eigensolver [85], the Quantum Approximate Optimization Algorithm
[84] provides specialized variational circuits, designed for addressing combinatorial optimiza-
tion problems. The underlying variational forms have later been improved, yielding the Quan-
tum Alternating Operator Ansatz [272]. As opposed to the quantum annealers with fixed con-
nectivity, QAOAs dimension does not degrade as a function of the QUBO density. However,
the number of qubits in contemporary quantum gate systems is lower than in quantum an-
nealers. Thus, the same splitting techniques can be applied. Hence, larger optimization prob-
lems can be addressed even in case of a relatively low number of qubits.
For non-iterative or non-variational algorithms, it is in general not possible to split any
quantum gate computation into smaller parts—mostly because entanglement cannot be main-
tained among the sub-problems. A reasonable direction can be the minimization of required
qubits. However, the number of required “data” qubits or “input” qubits is problem specific.
Methods for embedding high-dimensional points into a lower-dimensional manifold are
available aplenty, but due to their highly approximate nature, it is unclear if the resulting em-
bedded points still contain enough information. Assuming that the data qubits are fixed, the
other major source of qubit utilization are so-called ancilla qubits. In a quantum computation,
there is no way to deterministically construct a prescribed state unless one is given access to
qubits whose original state is known in advance.
As an example of this concept, consider a matrix M whose singular values are upper
bounded by 1. Let us denote the singular value decomposition of M by RDV . Then, the ma-
trix √
√ M R I − D 2V
U=
R I − D2 V −M
is unitary. This can be verified by observing that U U † = I. Note, however, that the dimension
of the matrix has been doubled. Interpreting M as an operator that shall be applied on some
state |ψ⟩, we have to extend the state by one additional qubit |0⟩a , such that U (|ψ⟩ ⊗ |0⟩a )
applies M to |ψ⟩. A generalization of this idea is the linear combination of unitaries [273].
Unitaries are not closed under summation. However, a sum of m unitary matrices can be ex-
tended to an unitary operator by using log m ancilla qubits.

Ancilla qubits appear frequently in quantum machine learning [113, 114, 234]. However,
care must be taken when the number of ancillas is provided in an asymptotic manner, e.g.,
[117, 274]. Given n input qubits, log n ancilla qubits can be reasonable. Nevertheless, O(log n)
may contain large constants, such that the number of required ancillas actually exceeds the
number of available qubits by a large margin.
5.1.4 Quantum Circuit Depth
With the currently available NISQ technology, the result of a quantum computation on a
physical device may deviate significantly from the expected outcome. A fundamental reason
is that the quantum computer, despite all technical efforts, is not perfectly isolated and inter-
acts (weakly) with its environment. In particular, there are two major effects of the environ-
ment that can contribute to computational errors: dissipation and decoherence in the sense of
dephasing [275, 276]. Dissipation describes the decay of qubit states of higher energy due to an
energy exchange with the environment. Decoherence, on the other hand, represents a loss of
quantum superpositions as a consequence of environmental interactions. Typically, decoher-
ence is more dominating than dissipation. In current quantum gate devices, these effects are
captured by the quantities T 1 and T 2. Both can be interpreted as decay constants. That is, the
probability that a qubit will stay in state |1⟩ after time t is proportional to exp(−t/T 1). Sim-
ilarly, T 2 indicates the decay constant for which an initial state |+⟩ will evolve into an equal
classical probabilistic mixture of the |+⟩ and |−⟩ states, so that one can no longer confidently
predict the state. That is, it’s the autocorrelation time after which the initial and final states
become uncorrelated.
Decoherence effectively limits the length of the longest computation and thus the al-
lowed depth of a quantum circuit. In other words, the more operations have to be performed
on a qubit register, i.e. the deeper a quantum circuit has to be, the likelier it is that decoher-
ence will happen. As explained in Section 5.1.2, transpilation will replace any unitary operator
which acts on 3 or more qubits into a series of one-qubit and two-qubit gates. This implies
that a single high-order operator might result in a circuit that is actually too deep to be exe-
cuted by a system. Despite decoherence and dephasing constants T 1 and T 2 being available
for most quantum processors, they are not utilized to adapt quantum machine learning meth-
ods to the underlying hardware and to account for the fact that the maximum computation
time is limited.
Quantum devices based on adiabatic evolution are affected by decoherence as well [277].
The direct effect is a shorter usable annealing time and thus the time until a solution to the
user specified optimization problem must be delivered. To see why this is significant, one
has to recall that simple optimization problems can already be solved by classical computers.
Thus, we are mostly interested in quantum solvers for hard optimization problems. However,
hard problems, e.g., measured via the Hamiltonian’s spectral gap, require a longer computa-
tion time—which is prohibited by the upper bound on the annealing time.

5.2 Algorithmic Requirements
We have discussed a series of pitfalls that arise in the design of quantum machine learn-
ing algorithms due to particularities of the quantum hardware. Of course, issues can also man-
ifest solely from algorithmic aspects. In classical computing, efficient data structures and data
types with a sufficient precision must be used. Components like pseudo random number gen-
erators and algebra sub-routines must be reliable and numerically stable. Parallel code has to
be checked for race conditions and other problems that might emerge due to concurrency.
In quantum machine learning, similar issues can show up. In what follows, we discuss
pre- and post-processing of inputs and outputs, stochasticity of the quantum computation,
and noise sensitivity.
5.2.1 Data Pre- and Post-Processing
The question of how to encode data for training or inference with quantum machine
learning methods so as to benefit from quantum advantages still awaits a final answer—if pos-
sible at all. Expected advantages, i.e. quantum speedup, may not emerge in practice if data pre-
processing would require exponential efforts.
The root of the pre-processing problem lies in the preparation of states. To see this, con-
sider the n-qubit state
n −1
2X
√
|ψ⟩ = pj |j⟩
j=0
where |j⟩ denotes the computational basis state that corresponds to the binary encoding of
the integer j, and pi is an arbitrary but fixed, classical probability mass function over {0, 1}n .
Clearly, various types of data can be encoded into that representation: e.g., pj could represent
1. (normalized) intensity values of pixels in an 2n/2 × 2n/2 image,
2. term frequencies over some text corpus with a vocabulary of size 2n ,
3. or the probability of observing a traffic jam on a specific set of street segments.
Related encodings indeed appear in the literature [278, 279, 280]. In that representation, a
greyscale image in 4K resolution (3840 x 2160 pixel) would occupy ⌈log2 (3840 × 2160)⌉ = 23
qubits. It is, however, important to understand that preparing the state |ψ⟩ requires resources
that are proportional to the dimension of the Hilbert space of the underlying qubit register. In
the above example n = 23 was small enough such that processing 2n values is feasible. If n is
large, we cannot simply prepare |ψ⟩. Even if a quantum algorithm would yield an exponential
improvement over the best known classical algorithm given input state |ψ⟩, it would not be
possible to gain any advantage due to the complexity of preparing the initial state.

By assuming efficient integrability of an input probability distribution pi , Grover and

Rudolph explain how to efficiently generate the quantum state in a coherent fashion such that
subsequent processing can be conducted on a quantum device [154]. However, their assump-
tions are not satisfied in most setups.
Recently, a classical model of computation was introduced which assumes that we can ef-
ficiently sample instances from a data set, where the probability to draw a specific data point
is proportional to its ℓ2 -norm. It can be shown that this is a natural analog to quantum al-
gorithms that assume efficient state preparation of classical data. Based on this reasoning,
classical versions of quantum algorithms for principal component analysis [281] and nearest-
centroid clustering [198] can be devised. The runtimes of these classical algorithms are only
polynomially slower which suggest that the exponential speedups of their quantum counter-
parts are an artifact of state preparation assumptions [282].
The same issues indeed arise when we want to read the result of a quantum computation.
When the full Hilbert space representation of the state encodes our desired outcome, an expo-
nential number of results has to be measured. Thus, even if the quantum computation itself
delivers an exponential speedup, this advantage will vanish when all probability amplitudes of
the final state have to be estimated.
5.2.2 Statistical Error
Statistical learning theory (SLT) is an integral part of machine learning. It may be the core
distinguishing property between machine learning and plain numerical optimization. The
theory itself is universal and independent of the underlying compute architecture. Through
the eyes of statistical learning theory, there is nothing special about quantum machine learn-
ing. QML has to obey the laws and limitations predicted by SLT. At a first glance, this does not
sound harmful. Nevertheless, some QML algorithms claim speed-ups over classical methods
which can easily be falsified by basic insights from SLT. In what follows, we explain the basic
reasoning behind this.
Assume that we try to find a function f such that the expected loss ℓ, also known as risk, is
minimized:
R(f ) = EP [ℓ(Y, f (X))]
Here, X is a random variable that represents the observed data or features, and Y is a random
variable representing the class label (or regression target). The pairs (X, Y ) follow the distri-
bution P. In practice, P is indeed unknown, but a data set D that contains samples from P is
available—allowing for the computation of the empirical risk:
1 X
R̃(f ) = ℓ(y, f (x))
|D|
(x,y)∈D
Let now H denote a machine learning method, e.g., H may contain all decision trees, all sup-
port vector machines with the same fixed kernel, or all neural networks with the same fixed

network structure. Under rather mild assumptions, H contains a minimizer of R̃ [283]. Let us
denote the empirical risk minimizer by f̃ = arg minf ∈H R̃(f ). The generalization error (also
known as estimation error) can be expressed as distance between both minimizers in terms of
R:
s
c(H) − log δ
R(f̃ ) − inf R(f ) ≤ 2 sup |R̃(f ) − R(f )| ≤ Θ
f ∈H f ∈H |D|
where Ω denotes asymptotic equivalence and δ ∈ (0, 1) is a probability. The last inequality
holds with probability 1−δ over the choice of the data set. Moreover, (H) denotes a measure of
complexity or expressivness of H (e.g., VC dimension or Rademacher complexity [283]).
p When
H and δ are treated as fixed constants, the estimation error is upper bounded by Θ( 1/|D|).
Hence, in order make use of all available
p data, any additional error that is introduced by a QML
method shall not be larger than O( 1/|D|)—no matter if it is generated by an approximate
optimization, measurement noise, or any other type of algorithmic error. When we denote the
learned function that incurs an additional algorithmic error by f̃ ε , we have
p
R(f̃ ε ) − R(f̃ ) = Θ( 1/|D|) + (R̃(f̃ ε ) − R̃(f̃ )) .
| {z }
Algorithmic error
While this insight might not sound significant, it can be used to prove that quantum
learning algorithms must have at least polynomial runtime in the dimension of the training
data, and therefore cannot achieve exponential speedups over classical polynomial time ma-
chine learning algorithms—even when special data storage like QRAM is assumed [284].
To understand the impact, consider quantum least squares (QLS) regression [267]. The
QLS algorithm is approximate. Its output is a state |ŵ⟩ while the state that contains the true
regression weights are |w⟩. For the error, we have ||ŵ⟩ − |w⟩| ≤ γ. According to [179], the run-
time is O(κc γ −β polylog(|D|)) for some c, β > 0 and data condition κ. According to the rea-
soning above, the algorithmic error γ should not exceed the statistical error, since otherwise
we would waste parts of the available data. We thus set γ = |D|−1/2 . In this case, the runtime
becomes O(κc |D|β/2 polylog(|D|)), and thus, polynomial in the number of data points—an ex-
ponential speedup over classical least squares cannot be justified.
Statistical effects also affect the accuracy of the readout. A specific number of readouts
is required to obtain the result of the quantum computation with a desired accuracy. Using
techniques developed within the field of quantum metrology, it is often possible to achieve a
precision that scales as 1/m, where m is the number of readouts [285].
In the quantum regression setting, this implies that from state |ŵ⟩, we can obtain ŵ by
sampling the circuit m times such that ∥|p
ŵ⟩ − ŵ∥ ≤ Ω(1/m). By the same reasoning as above,
this means that we have to obtain m = |D| shots for the readout in order to not discard
information from our data set D. No matter how fast our regression algorithm is, the number
of readouts will be a polynomial in the number of data points.

5.2.3 Noise Robustness
NISQ devices suffer from various types of noise, e.g., measurement noise or gate errors.
These uncertainties arise in addition to the general probabilistic nature of quantum computa-
tion. One might wonder why we should even care about additional stochastic effects, instead
of treating them as part of the probabilistic computation itself. Correction for measurement
noise exists such that there is no direct practical problem with it. However, the error induced
by noisy quantum gates is notoriously harder to characterize. One reason is crosstalk in super-
conducting quantum processors, e.g., performing operations on two pairs of qubits simulta-
neously can induce an error if the qubits are physically close on the chip. While crosstalk and
other sources of noise are hardware related problems, it is up to the algorithm whether slight
variations during the computation will lead to a completely different result.
As an example, consider adiabatic quantum computation. AQC is a stochastic procedure—
there is no guarantee that the outcome, measured after a finite amount of time, is the ground
state of the target Hamiltonian. It can well be that the adiabatic evolution ends up in an
excited (sub-optimal) state. The conditions under which the adiabatic process stays in the
ground state until the end of the evolution are well understood [63, 286]. The key property is
the Hamiltonian’s spectral gap, that is, the difference between the ground state energy and the
energy of the first excited state. Moreover, there is sufficient theoretical evidence that spectral
gaps of random problem instances are likrly to become super-exponentially small [287, 288].
Thus, the time required for adiabatic evolution to remain in the ground state is longer than
the time required for a classical brute force search through the full state space. It turns out that
exponentially small gaps appear close to the end of the adiabatic evolution for large random
instances of NP-complete problems. This implies that, unfortunately, adiabatic quantum
optimization fails: The system gets trapped in one of the numerous local minima. Since the
adiabatic evolution of NISQ devices is subject to minor perturbations, variations can affect
the spectral gap and hence alter the success probability of the computation. Thus, QUBO
formulations should take this problem into account. Nevertheless, experimental results show
that the QAOA algorithm seems to be more robust against exponentially small spectral gaps
[289].
Another line of research addresses the formal robustness verification of quantum ma-
chine learning algorithms against unknown quantum noise. As mentioned above, noise may
arise from various sources, rendering a full characterization of the noise distribution almost
infeasible. In [290], the authors relate noise robustness to adversarial examples and present
an analytical robustness bound to assess the robustness of quantum classification algorithms.
Their results are, however, not evaluated on real quantum computing devices. At the time of
writing, evaluating the noise robustness of quantum machine learning algorithms is still an
open problem.

6 SUMMARY AND OUTLOOK
6 Summary and Outlook
After decades of worldwide research, artificial intelligence has finally made considerable
strides with respect to capabilities and applicability. As a consequence, weakly intelligent cog-
nitive systems are now increasingly deployed in practice and appear in more and more areas
of our daily- and professional lives. An interesting general observation in this context is that
accelerated progress and noteworthy recent accomplishments in this area have mainly been
driven by modern large scale machine learning.
Machine learning is a branch of artificial intelligence that deals with adjusting the param-
eters of software agents in a training process such that they can develop cognitive capabilities
and problem solving skills. Put differently, the recent performance boost in artificial intel-
ligence is mainly due to systems which analyze large amounts of task specific training data
in order to learn an intended input-output behavior. The tacit assumption behind this idea
is that any real world data must have been produced by some generally unknown process or
mechanism and that this process or mechanism can be modeled mathematically. This trans-
lates to the assumption that there exists a suitable parameterized function whose parameters
can be automatically adjusted such that the input-output behavior of that function reflects or
mimics the characteristics of the given training data.
There exist numerous possible machine learning models (mathematical formalizations
of a given application scenario) as well as numerous learning algorithms (mechanisms to fit
a given model to a given set of training data). Traditionally, models and algorithms for their
training were chosen with respect to the task at hand, often specifically tailored towards a spe-
cific setting. However, the dominating trend over the past decade and arguably the main rea-
son behind the success of modern machine learning has been to work with domain agnostic
models and training algorithms. In particular, deep learning has proven to be tremendously
successful in a wide range of computational intelligence problems.
Deep learning involves deep (and wide) artificial neural networks which are composed of
millions of artificial neurons which interact over even more artificial synapses. Since such net-
works thus come with billions of adjustable parameters (synaptic weights and activation func-
tion bias values), they provide very flexible general models which can be trained with (variants
of) the backpropagation algorithm to learn a desired cognitive skill. However, in order for this
to happen reliably, deep neural networks need to be trained with vast amounts of representa-
tive training data. Due to the number of data to be processed and the number of parameters to
be adjusted, the training of modern deep networks is a formidable task that requires consid-
erable computational resources in order to happen within reasonable time. Indeed, breaking
down the efforts involved in training state of the art systems such as, say, OpenAI’s GPT3 for
text analysis- and synthesis reveals training times of several hundred GPU years which, to be
possible within a matter of days, necessitates the use of dedicated compute clusters. In other
words, state of the art machine learning has reached a point where its practical feasibility and
success are conditioned on access to high performance computing hardware.

Given this state of affairs regarding the computational needs of current machine learn-
ing systems, it is not surprising to find that more and more researchers are beginning to look
at quantum computing as a tool to be deployed at different stages of the machine learning
pipeline. The basic observation is that quantum computing, too, has made considerable strides
over the past decade and is now becoming practical. While quantum computing will not elim-
inate the dependency on special purpose hardware, it promises significantly faster compu-
tations across a wide range of scientific or industrial applications. With respect to machine
learning, the ambition is to harness potential quantum speedup especially for the training of
ever more complex systems.
The advantages of quantum computing over classical digital computing are rooted in the
fact that it exploits quantum mechanical phenomena for information processing.
While digital computers operate on bits which are in one and only one of two possi-
ble states, quantum computers operate on qubits. These are logical interpretations of phys-
ical two-state quantum systems which exist in a superposition of two basis states. The state
space of a qubit thus consists of infinitely many states so that a qubit can carry more informa-
tion than a classical bit. While the mathematics that describes the behavior of classical bits is
Boolean algebra, the mathematics that describes the behavior of qubits is complex linear al-
gebra. Qubits can be represented as two-dimensional, complex-valued unit vectors that are
formed as a linear combination of two distinguished, linearly independent, orthonormal ba-
sis vectors. The squares of the norm of the coefficients of a qubit state vector are called am-
plitudes and have an important probabilistic interpretation. If a measurement is performed
on a qubit, it will decohere, i.e. loose the property of superposition, and collapse to either one
of its basis states. The probability of the measured state to be the first basis state corresponds
to the squared norm of the first coefficient and the probability of measuring the qubit in the
second state corresponds to the squared norm of the second coefficient. Computations in-
volving operations on- and subsequent measurements of qubits are therefore probabilistic
rather than deterministic as in the case of classical bits. However, just as classical bits can be
combined into bit registers, quantum bits can also be combined into qubit registers. Adding a
single qubit to a register increases the dimension of the register’s state space by a factor of two
and so a quantum register of n qubits exists in a superposition of 2n basis states. Another cru-
cial difference to classical computing is that qubits can be entangled. Whenever two or more
qubits are entangled, their individual states cannot be measured separately but a measure-
ment of an entangled qubit will also determine the (combined) state of the others.
The interplay of these phenomena gives rise to the potential supremacy of quantum
computing over digital computing: on a quantum computer it is possible to work with only
n qubits in order to perform computations in an 2n dimensional space. This is of considerable
interest for combinatorial optimization which deals with exponential search spaces and plays
a crucial role in a branch of artificial intelligence known as problem solving, as well as in gen-
eral parameter selection problems in certain machine learning techniques.
Adiabatic quantum computers such as produced by D-Wave are especially tailored to-

wards solving an important class of combinatorial optimization problems, namely QUBOs,

which occur in machine learning contexts such as data clustering, classifier boosting, or sup-
port vector machine training. The basic idea in adiabatic quantum computing is to express a
given problem as an energy minimization problem and to devise an energy function which is
known to attain its minima at the unknown solutions to the problem. The energy of a quan-
tum system is characterized in terms of a Hamiltonian operator whose eigenvalue spectrum
is the set of all possible outcomes when measuring the system’s energy; its eigenstates re-
flect to which states of the system these energies correspond to. The lowest energy state of
the quantum system under consideration is also called the ground state of the corresponding
Hamiltonian and, if the energy function for a given problem can be translated into a Hamil-
tonian operator, its ground state can be found via adiabatic quantum computing. Adiabatic
quantum computing exploits a phenomenon summarized in the adiabatic theorem. It states
that, if a quantum system starts in the ground state of a Hamiltonian which then gradually
changes towards another Hamiltonian, the system will end up in the ground state of the re-
sulting Hamiltonian. On an adiabatic quantum computer one thus operates with two Hamil-
tonians, a problem independent one and the problem Hamiltonian and gradually changes the
former into the latter. This general idea therefore exploits a form of quantum tunneling, i.e. a
quantum mechanical principle which states that quantum systems can tunnel through energy
barriers. The latter constitutes an advantage over classical energy-based optimization where
optimizers may get trapped in local minima and thus may fail to find optimal solutions to a
problem. Moreover, for many problems of practical importance, adiabatic quantum comput-
ers can search the solution state space polynomially faster than classically possible. While this
may not sound like much, for large problems it can make the difference between tractable and
intractable.
At first sight, the way of thinking required for adiabatic quantum computing may look
abstract and unusual. However, problem solving by energy minimization is a well established
general paradigm in classical machine learning. In fact, the well understood Hopfield net-
works, a venerable type of neural networks, can be seen as a classical or digital analogue to
adiabatic quantum computing. This is to say that anything that can be accomplished by run-
ning a classical Hopfield network on a digital computer can, in principle, also be accomplished
on an adiabatic quantum computer. In this sense, the conceptual gap between the classical
paradigm of Hopfield neural networks and adiabatic quantum computing is therefore rather
narrow.
Quantum gate computing manipulates qubits in a manner that more closely resembles
classical digital computing. On the hardware level, digital computers process information by
manipulating sets of bits via atomic logical operations which are implemented in terms of so-
called gates. In quantum gate computing, corresponding operations on qubits are realized via
the application of quantum mechanical operators. Mathematically, these are unitary transfor-
mations so that any operation on qubits must be unitary too. Such operators typically act on
one or two qubits at a time but can be sequenced or executed in parallel to form more compli-
cated operations on sets of qubits. In analogy to classical computing, complex computational

units that are composed of individual quantum gates are called quantum circuits. From a logi-
cal point of view, the fundamental problem in quantum gate computing is therefore to devise
quantum circuits that show an intended input/output behavior. This is generally a non-trivial
problem and it is interesting to note that classical machine learning is increasingly seen as tool
for quantum circuit design.
What makes quantum gate computing attractive from the point of view of machine
learning is that, mathematically, quantum gate computing is nothing but applied complex
linear algebra. Due to mechanisms such as superposition or entanglement, however, the
linear algebraic operations that occur in quantum gate computing implicitly act on state
spaces that are exponentially larger than those considered in classical computing. Since many
common machine learning tasks involve linear algebraic operations on large amounts of
high dimensional data vectors, it seems auspicious to try to encode such data in qubit state
spaces and to try to leverage quantum advantages when computing with such states. Here,
the general expectation is that quantum speedup will considerably accelerate corresponding
learning processes or even allow to tackle hitherto intractable problems.
Indeed, in the words of Aaronson, there has recently been a “quantum machine learning
mini-revolution” and the number of scientific reports on quantum circuits for certain general
problems in machine learning has grown considerably over the past decades. Problems which
have been considered in this context include basic linear algebra routines, regression, or clas-
sification. With respect to the latter, there exist different paradigms so that it is no surprise to
find numerous proposals of quantum circuits for nearest neighbor classifiers, basic linear clas-
sifiers (perceptrons), ensemble classifiers, support vector machines, and neural networks.
Present day ideas for how to realize quantum machine learning routines on quantum
gate computers commonly involve variational quantum computing algorithms or hybrid
quantum-classical methods. These consider parameterized quantum circuits which may be
designed manually or automatically but in either case consist of tunable quantum gates. The
problem solved by hybrid quantum-classical methods therefore is the problem of adjusting
the parameters of individual gates such that the circuit as a whole performs the desired com-
putation with high probability. The basic structure of variational quantum algorithms con-
sist of an outer loop run on a digital computer which manages the current estimates of the
sought after parameters. In each iteration, these parameters are used to setup computations
on a quantum computer whose outcomes are then measured. Given these measurement re-
sults, classical optimization techniques are then used to estimate improved parameters and
the whole processes is iterated until the quantum circuit shows the desired input/output be-
havior within a given tolerance.
Given the technical capabilities of existing, present day quantum computers, variational-
or hybrid quantum-classical algorithms are appealing because they have been found to
considerably reduce the quantum computing resources (circuit depth, coherence time,
qubit counts) that are required to successfully run a quantum machine learning method.
Moreover, researchers interested in developing quantum neural networks often even consider

parameterized quantum circuits as a quantum computing analogue of classical deep neural

networks. This is likely because deep neural networks as well as parameterized quantum
circuits consist of basic computational units arranged in layered structures and because both
involve classical optimization algorithms for parameter adjustment. However, one has to
be careful with such analogies since there also are crucial differences. First of all, quantum
gates in a quantum circuit realize unitary operators rather than non-linear functions as
in the case of neural networks. Non-linear computations are vital for the general problem
solving capabilities of neural networks but they can not be realized through unitary operators
alone. Non-linear quantum computations either involve the generous use of ancilla qubits
or state measurements. Second of all, internal states of a quantum circuit cannot be read
out without destroying their quantum coherence, i.e. without loosing quantum aspects such
as superposition. Classical neural network training based on error backpropagation does
therefore not apply to parameterized quantum circuits and the use of variational- or hybrid
quantum-classical algorithms for circuit optimization currently constitutes the only viable
approach to this problem.
The ever faster growing literature on quantum machine learning and the many success
stories reported therein seem to suggest that the application of quantum computing algo-
rithms to computational intelligence problems could soon break into the mainstream. How-
ever, in the present era of noisy intermediate-scale quantum (NISQ) computing, overly enthu-
siastic expectation still need to be reigned in. This has mainly to do with the technical capa-
bilities of present day quantum computers which often differ from the assumptions made by
quantum algorithm developers.
First of all, quantum algorithm design considers properties of logical qubits rather than
those of physical qubits. The former are the basic building blocks on which quantum algo-
rithms operate, the latter are physical devices inside of a quantum computer which behave
like two-state quantum systems. From the point of view of algorithm design, the focus on log-
ical qubits is reasonable and mimics levels of abstraction in modern software development
where most programmers need not worry about hardware details. Current quantum comput-
ers, however, have not yet reached the same level of maturity as digital computers but still ex-
hibit certain limitations. As of now, pure logical qubits are therefore an idealization since they
abstract away shortcomings of physical qubits in present day NISQ devices. These typically
contain less than a hundred qubits, often suffer from limited coherence times, and are sus-
ceptible to low fault-tolerance due to internal fluctuations or measurement noise. Indeed, it
still is challenging to technically create and manipulate quantum states and to maintain their
quantum mechanical properties over longer periods of time. It is rather marvelous that this
has become possible at all and experts expect that continuing technological progress will lead
to better and more powerful devices. Lack of fault tolerance is a critical issue and a techno-
logical milestone will be the implementation of quantum error correction mechanisms sim-
ilar to those used in digital computing. Here, it is interesting to note that errors often result
from the application of less robust quantum gates. As of this writing, only rather small quan-
tum circuits work really stable. Since adiabatic quantum computers do not involve quantum

gate computations, they can more reliably manipulate larger (physical) qubit systems. At the
same time, their use is currently restricted to rather specific energy minimization problems so
that present day devices are not as universal as quantum gate computers. Even though both
paradigms are theoretically equivalent, the emulation of quantum circuits on an adiabatic
quantum computer would require qubit connectivity structures that have not yet been real-
ized.
While presently realizable qubit counts, qubit connectivity, dimensionality of quantum
gates, and quantum circuit depths impose practical restrictions on the kind of quantum algo-
rithms that can run successfully on current NISQ devices, there are further limitations regard-
ing the feasibility of certain logical quantum computing concepts. We therefore emphasize the
following additional points.
Second of all, it is often not immediately clear how to encode classical data such that it
can be processed on quantum computers; neither may it be obvious how to decode measured
qubit states into classical representations that would allow for meaningful downstream pro-
cessing. This input-output problem is often abstracted away by quantum algorithm designers
which, in turn, may have dire consequences with respect to claims about quantum speedup.
For instance, a quantum algorithm operating on a quantum encoding of classical data might
be exponentially faster than its classical counterpart operating on the classical data. However,
if the effort for preparing quantum states that encode the classical data is itself exponential,
then the apparent quantum advantage vanishes. Here, it is noticeable that designers of quan-
tum algorithms often simply hypothesize universal quantum computers. Among others, these
machines are supposed to come with quantum random access memories (QRAMs) which, sim-
ilar to their digital counterparts, can hold data for processing. However, given present day
technologies, such QRAMs are not yet possible and it is even questionable if they will ever ex-
ist in their truest sense. This is due to another quantum mechanical principle, the no-cloning
theorem, which states that it is impossible to create independent identical copies of arbitrary
quantum states. To the best of our current knowledge, and unless reliable quantum error cor-
rection becomes available, a QRAM would thus only allow for approximate repeated access to
quantum states for processing. Quantum state decoding, too, can diminish quantum advan-
tages. Even if a quantum algorithm can produce a quantum state representation for a sought
after solution much faster than classically possible, the effort for measuring and reading the
resulting state into classical memory could still be so substantial that any advantage disap-
pears. Before claims as to the superiority of quantum algorithms can be made, it is therefore
pivotal to factor in efforts for state preparation or measurements and to match these against
pre- and post-processing efforts of efficient classical algorithms.
Third of all, as of now quantum computing is essentially bit level computing. That is,
present day quantum algorithm design deals with the design of problem specific quantum
circuits or problem specific energy functions. Contrary to classical computing, there are no
abstract data structures, such as linked lists, binary trees, or heaps. Neither are there or con-
trol structures such as if-then-else-statements or for- or while-loops known from higher
level programming languages and common classical programming patterns such as the use of

variables, too, are not immediately possible on present day quantum computers. This is chal-
lenging in so far as it suggests that certain existing machine learning algorithms build around
such constructs may, for the foreseeable future, not be realized on quantum computers. While
there are increased efforts towards quantum compilers which automatically translate higher
level programs into quantum circuits, these are still in their infancy. Also, while high level
application programmings interfaces (APIs) for quantum computing are increasingly avail-
able (e.g. QISKIT, CIRQ, Forest, PennyLane), it is important to note that these are mainly tools
for setting up quantum computing processes. That is, users of these tools still have to think
on the linear algebraic or “state and operator” level of quantum computing and use the API
only to implement vectors, matrices and their interplay. Another caveat with respect to these
APIs is that they often allow for efficient digital simulations of quantum information process-
ing. While this is undoubtedly helpful in the design stage of algorithm design, it can also be
misleading as the simulated quantum processors typically are universal quantum processors.
This, in turn, means that a quantum machine learning algorithm that works on a simulated
quantum computer may not work on a currently existing physical quantum computer. A fi-
nal caveat with regard to quantum algorithms for machine learning is that quantum compu-
tations which involve measurement steps are inherently probabilistic. This is of course well
known to quantum computing practitioners but may be overlooked by novices. Any quantum
computation must therefore be repeated several times and any result obtained this way has to
be interpreted in terms of expectations rather than in terms of deterministic outcomes.
Fourth of all, while machine learning is a comparatively new scientific discipline, quan-
tum machine learning is even newer. This has consequences for the best practices and stan-
dards in the field. To clarify what this may mean, we note that, in its first couple of decades,
classical machine learning has gone through a verifiability and reproducibility crisis. Practical
results tended to be reported, without disclosing implementation details, data collections or
processing protocols, or experimental procedures in detail. Regarding the validity of claimed
capabilities of a reported method, such omissions can make a considerable difference. It is, for
instance, pivotal that training and testing of a machine learning system happen on indepen-
dent data sets, because a low error on the training data does not imply that the trained system
can generalize well. On the contrary, an exceedingly good performance on the training data
is often a symptom of overfitting. Over time, these issues have been recognized by the ma-
chine learning community and, as a consequence, present practice for scientific publication is
to require authors to provide their code, data, and experimental protocols. In the new field of
quantum machine learning, this is not yet the case. Indeed, reading the corresponding scien-
tific literature, it is noticeable that crucial details as to how a practical result has been obtained
are often missing. In many reports on practical quantum machine learning, it often not even
recognizable if, say, rigorous evaluation practices have been followed so that one may argue
that present day quantum machine learning is experiencing a reproducibility crisis period.
For now, this may be acceptable as the field is in its nascent phase; it is, however, important to
keep in mind that the good performance of some of the currently reported methods may not
scale or generalize to large or different application settings.

While all these caveats may dampen enthusiasm for the prospect of quantum machine
learning, recent technological progress has been substantial enough to merit serious engage-
ment with the topic. In other words, although quantum computers and quantum machine
learning algorithms are not yet mature enough to impact the way machine learning happens
in practice, it currently seems reasonable to expect that—due to considerable investments by
institutional and industrial stakeholders—the underlying technology will continue to develop
and improve ever more quickly. This, in turn, may soon lead to practically viable solutions and
cause hitherto unexpected developments and disruptions. A foresight process which fathoms
potential benefits and risks of quantum machine learning therefore seems appropriate.
In particular, potential risks related to the use of quantum machine learning have not yet
received the same amount of attention as its benefits and corresponding reports are scarce. In
other words, while ethics, reliability, trustworthiness, and safety of classical machine learn-
ing have by now been recognized as important topics, quantum machine learning has not yet
been scrutinized in these regards. However, given the expected impact of quantum machine
learning on capabilities and utilizability of artificial cognitive systems, it seems appropriate to
asses potential security issues related to quantum machine learning. In upcoming reports, we
will therefore investigate questions pertaining to the reliability or vulnerability of quantum
machine learning systems. Further considerations include whether or not quantum machine
learning allows for new kinds of attacks on- or new defense mechanisms for critical digital in-
frastructures. The general goal will be to assess quantum machine learning from the point of
view of cybersecurity and to determine what kind of measures, if any, will be required in this
regard.

GLOSSARY
Glossary
Adiabatic quantum A quantum computing paradigm that is based on the adiabatic

computing (AQC) theorem and particularly tailored towards solving QUBOs; in
adiabatic quantum computing, a qubit system gradually evolves
from the ground state of a problem independent beginning
Hamiltonian to the ground state of a Hamiltonian which models
a given problem; the method thus uses quantum annealing or
quantum tunneling for problem solving.
Adiabatic quantum A synonym for adiabatic quantum computing.

optimization (AQO)
Adiabatic theorem A mathematical formulation of the quantum mechanical prin-

ciple that, if a quantum system starts in the ground state of a
Hamiltonian operator which then gradually changes for a period
of time, the system will end up in the ground state of the resulting
Hamiltonian.
Application phase (of The final stage in the practical development of a machine learn-
an ML system) ing system; once the system has been trained and tested and was
found to perform robustly and reliably, it can be deployed in
practical applications.
Artificial Intelligence A branch of computer science concerned with the design, devel-
(AI) opment and deployment of technical systems or software agents
that have cognitive capabilities such as text- or image under-
standing, planning, and decision making.
Backpropagation A very widely used algorithm for the training of (feed-forward)

neural networks; given an appropriate loss function, the algo-
rithm essentially applies the chain rule of differentiation to adjust
the parameters of all neurons in all layers of a neural net; put
differently, the method performs gradient descend in a highly
non-linear error landscape; there exist numerous variants of the
original version of the algorithm, for instance, to automatically
determine optimal step sizes for the descent procedure.

GLOSSARY
Basis states States of quantum mechanical systems are described as linear

combinations over basis states; for instance, the state of a qubit
is represented as a two-dimensional, complex-valued unit vector
that is a linear combination of two distinguished, linearly inde-
pendent, orthonormal basis states.
Beginning Hamilto- A problem independent Hamiltonian operator used in adiabatic

nian quantum computing; the beginning Hamiltonian is usually cho-
sen such that its ground state can be easily prepared and used as
the initial state for the adiabatic problem solving process.
Computational basis A basis system used to express the state of a qubit system; the
state of a system of qubits corresponds to a vector in a complex
Hilbert space and the basis used to (numerically) express the en-
tries of this vector can be chosen arbitrarily; the computational
basis is often the most “natural” choice, i.e. a basis that suits the
problem at hand or in which state vectors have a simple form.
Decoding problem The problem of measuring or reading a quantum state which re-
sults from a quantum computation; quantum computing results
usually need to be transferred to digital computers in order to
enable further processing or analysis; if this transfer cannot be
done efficiently, advantages due to quantum speedup can be lost.
Decoherence A term referring to the loss of quantum coherence; if a quantum

systems decoheres, it loses its quantum properties; decoherence
of qubit happens through measurement or (unwanted) inter-
actions with the outside world; qubit systems have thus to be
kept as isolated as possible when performing quantum computa-
tions; if a qubit decoheres, it collapses to one of its basis states and
henceforth behaves like a classical bit.
Digital annealer A specialized digital computing device tailored towards solving

QUBOs; digital annealers classically emulate the workings of
adiabatic quantum computers; this requires efficient (hardware)
implementations of classical optimization algorithms; compared
to present day adiabatic quantum computers, digital annealers are
much cheaper and much more energy efficient.
Eager learner A machine learning system that trains a model in an offline train-
ing phase and considers a complete existing training data set for
re-training and rather than only newly added data.

GLOSSARY
Encoding problem The problem of preparing quantum states that represent classi-
cal data; if classical data cannot efficiently be encoded in terms
of quantum states, computational advantages due to quantum
speedup can be lost.
Ensemble learning Ensemble Learning methods combine different, typically sim-

ple (machine learning) models into a single stronger model that
achieves better results than the individual member of the ensem-
ble. Ensemble Learning allows, for example, for an average value
from the results of these different models.
Entangled state In quantum computing, this term refers to a state in which a sys-
tem of qubits can exists such that their individual states cannot
be measured separately; rather, a measurement of an entangled
qubit will also determine the (combined) state of the others; en-
tanglement is a quantum mechanical phenomenon for which
there is no classical analogue and which is essential for the inner
workings of quantum computers.
Federated learning In federated learning, each participant retains their own data and
contributes to a common learning process in which the system as
a whole benefits from the capabilities of each contributor.
Generalization The capability of a trained machine learning system to generalize

the predictive performance it acquired from analyzing training
data to previously unseen test data.
Ground state A term used to refer to the lowest energy state a quantum me-
chanical system can be in; mathematically the lowest energy of
a quantum system corresponds to the bottom eigenvalue of its
Hamiltonian operator and the system’s ground state is given by
the corresponding eigenvector.
Hamiltonian operator A quantum mechanical operator acting on the state vectors of

quantum mechanical systems; Hamiltonians feature prominently
in Schrödinger equations; their eigenvalues represent possible en-
ergy levels the corresponding quantum system can be measured
in.
Kernel machine A machine learning term used to refer to a wide class of machine
learning models; kernel machines are machine learning models
that make use of the kernel trick when processing data.

GLOSSARY
Kernel trick A computational method used in kernel machines which allows

for applying inherently linear methods to non-linear machine
learning problems. Invoking the kernel trick involves two steps:
1) to rewrite a given machine learning method such that any oc-
currence of data vectors is in form of inner products and 2) to
replace any computation of an inner product by the computation
of an appropriate kernel function.
Lazy learner A machine learning system that simply stores data and delays
modeling until asked to make predictions; a simple example for
this paradigm is a nearest neighbor classifier.
Logical qubit A mathematical idealization of a physical two-state quantum

system; logical qubits exist in a superposition of two basis states
and can therefore represent more information than classical bits;
they form the basic units in quantum computing.
Loss function A criterion to measure how well a machine learning model and its
current choice of parameters represent a given set of data; qual-
ity measurement is done by assigning to each prediction of the
model the loss that occurs when a prediction deviates from the
correct prediction; the more predictions deviate from the ground
truth training data, the higher the output of the loss function.
Machine Learning A branch of artificial intelligence that deals with adjusting the
(ML) parameters of technical systems or software agents in a training
process such that they can develop cognitive capabilities and
problem solving skills.
Model class A machine learning term used to refer to a parameterized family

of functions; once a model class has been chosen for a machine
learning problem, the task is to determine that model within
the class that provides the best fit to a given set of training data;
adjusting the model to the data at hand happens through ma-
chine learning algorithms which tune its parameters such that the
model is in good agreement with the data; often, this happens by
means of minimizing a loss function.
No-cloning theorem A quantum mechanical principle that states that it is impossi-

ble to create independent identical copies of arbitrary quantum
states.

GLOSSARY
Noisy intermediate- A term used to describe existing, present day quantum computers
scale quantum device which contain less than a hundred physical qubits and still suffer
(NISQ) from limited coherence times and low fault-tolerance.
Oracle A kind of function posited or required in many quantum comput-

ing algorithms; oracle functions can quickly verify if a supposed
solution to a problem really is a solution; for instance, while the
problem of finding the prime factors of 42 is somewhat difficult,
it is easy to verify that 2, 3, and 7 are the solution; by the same
token, it is also easy to verify that, say, 5 and 13 are not.
Overfitting A type of machine learning modelling error that occurs when a

model corresponds too closely to a given data set and is therefore
likely to produce very different results for (slightly) different data
samples.
Parameterized quan- A kind of quantum gate that requires numeric parameters to

tum gate define its actual function; parameterized quantum gates are of
pivotal importance for variational quantum algorithms and for
quantum machine learning; they allow for representing machine
learning models via a single quantum circuit and for the use of
optimization techniques to adjust the input/output behavior of
the circuit such is agrees with given task specific data.
Phase gate A single qubit quantum gate that modifies or shifts the phase of
the quantum state of a qubit.
Physical qubit A physical realization of a logical qubit, namely a device or appa-

ratus that behaves like a two-state quantum system and forms a
component of the hardware of a quantum computing system.
Problem Hamiltonian A Hamiltonian operator used in adiabatic quantum comput-

ing; the problem Hamiltonian is designed in such a manner that
its ground states represent the sought after solution to a given
(combinatorial optimization) problem; in adiabatic quantum
computing, a qubits system is prepared in the ground state of a
problem independent beginning Hamiltonian which then grad-
ually changes towards the problem Hamiltonian; this will cause
the qubit system to end up in a ground state of the latter and thus
solve the problem at hand. The design of problem Hamiltonians
requires experience with (re)formulating QUBOs.

GLOSSARY
Quadratic uncon- A kind of combinatorial optimization problem where the decision

strained binary op- variables can only assume two values (typically either 0 and 1 or
timization (QUBO) −1 and +1); many practically important subset selection or set
bi-partition problem can be written as QUBOs but their solution
is classically difficult as they are generally NP-hard.
Quantum approxi- A variational quantum algorithm in which parameterized quan-

mate optimization tum gates are optimized to approximate adiabatic quantum
algorithm (QAOA) computations.
Quantum bit (qubit) A two-state quantum system; a qubit exists in a superposition of

two basis states and can therefore represent more information
than a classical bit; qubits are the basic unit of information in
quantum computing or quantum information processing and are
modeled in terms of vectors in a complex Hilbert space; one often
distinguishes between logical qubits (which represent the math-
ematical essence of two-state quantum systems) and physical
qubits (which are physical instances or realizations of two-state
quantum systems); in other words, while there exists different
physical manifestations of qubits, their essential (mathematical)
properties are the same.
Quantum computing A computational paradigm that harnesses the principles of quan-

tum mechanics for information processing; exploiting quantum
mechanical phenomena such as superposition or entanglement
offers great computational power but also requires a different
kind of algorithmic thinking than in classical digital computing.
Quantum gate com- A quantum computing paradigm sometimes also referred to as

puting the quantum circuit model of computation; in quantum gate
computing, operations on qubits are realized via the application
of quantum mechanical operators which are physically imple-
mented as quantum gates; in analogy to classical computing,
complex computational units which are composed of individual
quantum gates are called quantum circuits.
Quantum gate A basic computational unit within a quantum circuit; quantum

gates are used to perform quantum mechanical operations on
individual qubits or on systems of qubits.

GLOSSARY
Quantum information A scientific discipline that deals with theory and practice of sys-
processing tems which process quantum information; an important aspect
of quantum information processing is quantum computing; how-
ever, quantum information processing is sometimes seen to be
more comprehensive than quantum computing because it also
involves topics such as quantum communication or quantum
sensing.
Quantum inspired A term used mainly to describe classical algorithms whose de-
computing sign or operating principles are inspired by- or try to emulate
quantum mechanical phenomena; for instance, a digital annealer
digitally emulates the inner workings of an adiabatic quantum
computer.
Quantum machine A term mainly used to refer to the idea of incorporating quantum
learning computing routines at different stages of the machine learning
pipeline; in a much broader sense, the term is also used to refer to
quantum inspired classical algorithms for classical data analysis,
genuine quantum algorithms for classical data analysis, classical
algorithms for quantum data analysis, and quantum algorithms
for quantum data analysis.
Quantum supremacy A term used to describe the fact that there exist problems which a
quantum computer can solve but a classical computer can not in
feasible time; in order to exhibit quantum supremacy with respect
to a given problem, a quantum algorithm has to provide super-
polynomial speedup over the best possible classical algorithm.
Quantum tunneling A quantum mechanical principle that describes the fact that
quantum systems can tunnel through energy barriers; quantum
annealing or adiabatic quantum computing exploit this phe-
nomenon.
Regularization A mathematical technique or concept designed to prevent ma-

chine learning algorithms from overfitting by imposing con-
straints on model parameters that reduce the variance of a model.
Reinforcement learn- A machine learning paradigm tailored to the problem of learning

ing “what to do when” in order to achieve a long term goal; rein-
forcement learning typically considers states an agent can be in,
actions an agent can perform to transit to another state, and de-
layed rewards an agent has to maximize.

GLOSSARY
Superposition A quantum mechanical principle which states that the sum (su-
perposition) of two or more quantum states is another valid
quantum state; superposition is often interpreted in the sense
that a quantum mechanical system can be in many different
states simultaneously. Upon measurement, the system will col-
lapse to one of the superposed states.
Supervised learning A machine learning paradigm where machine learning mod-

els are trained on annotated training data; the goal is to adjust
the model parameters such that the model is able to map the
given input data to the given output data with high fidelity; im-
portantly, models trained in a supervised manner have to be
evaluated on independent test data in order to verify that they
can generalize to situations not contained in the training data; a
common example of a supervised learning problem is classifier
training.
Test phase (of an ML A stage in the practical development of a machine learning sys-
system) tem; given problem specific test data and a trained model, the
model is evaluated on the test data using an appropriate perfor-
mance measure; it is pivotal that the test data are independent of
the data used to train the model, otherwise issues such as overfit-
ting cannot be detected.
Topological quantum A still mainly theoretical concept for how to realize physical
computing (TQC) qubits based on the idea of working with quasi-particles called
anyons.
Training phase (of an A stage in the practical development of a machine learning sys-
ML system) tem; given problem specific training data and a parameterized
mathematical model, the model parameters are adjusted auto-
matically such that the model matches the training data to the
best extend possible; often, optimal model parameters are esti-
mated by means of minimizing a loss function.
Two-state quantum A quantum mechanical system whose state space consists

system of infinitely many states which can be represented as two-
dimensional, complex-valued unit vectors that are linear com-
binations of two distinguished, linearly independent, orthonor-
mal basis states. Well known real world examples of two-state
quantum systems include the polarization of a photon (with basis
states vertical or horizontal) or the spin of an electron (with basis
states up or down).

GLOSSARY
Unsupervised learning A machine learning paradigm that aims to detect and uncover
latent or inherent structures within a given data set of unlabeled
data; a common example of an unsupervised learning problem is
data clustering.
Variational quantum An algorithm that combines quantum computations with clas-

algorithm sical computations in an iterative feedback loop; variational
quantum algorithms or hybrid quantum-classical algorithms are
used to adjust the parameters of a parameterized quantum circuit
such that it shows an intended input/output behavior within a
given tolerance; a variational quantum algorithm involves an
outer loop run on a digital computer which manages the current
parameter estimates, in each iteration, these parameters are used
to setup computations on a quantum computer whose outcomes
are measured, measurements are then used in classical optimiza-
tion techniques to estimate improved parameters and the whole
processes iterates until convergence.
Variational quantum A variational quantum algorithm for estimating the bottom

eigensolver (VQE) eigenvalues of Hamiltonian operators.

LITERATURE
Literature
[1] G. Hinton et al. “Deep Neural Networks for Acoustic Modeling in Speech Recognition:
The Shared Views of Four Research Groups”. In: IEEE Signal Processing Magazine 29.6
(2012). DOI: 10.1109/MSP.2012.2205597. URL: http://cs224d.stanford.edu/
papers/maas_paper.pdf.
[2] J. Devlin et al. “BERT: Pre-training of Deep Bidirectional Transformers for Language
Understanding”. In: arXiv:1810.04805v2 [cs.CL] (2018). URL: https://arxiv.org/abs/
1810.04805v2.
[3] A. van den Oord et al. “Parallel WaveNet: Fast High-Fidelity Speech Synthesis”. In: Proc.
Int. Conf. on Machine Learning (ICML). 2018. URL: https://arxiv.org/abs/1711.
10433v1.
[4] A. Krizhevsky, I. Sutskever, and G. Hinton. “ImageNet Classification with Deep Convo-
lutional Neural Networks”. In: Proc. Advances in Neural Information Processing Systems
(NeurIPS). 2012. DOI: 10.1145/3065386. URL: https://proceedings.neurips.cc/
paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf.
[5] J. Ng et al. “Beyond Short Snippets: Deep Networks for Video Classification”. In: Proc.
Conf. on Computer Vision and Pattern Recognition (CVPR). IEEE, 2015. DOI: 10.1109/
CVPR.2015.7299101. URL: https://arxiv.org/abs/1503.08909v2.
[6] T. Wang et al. “Video-to-Video Synthesis”. In: Proc. Advances in Neural Information Pro-
cessing Systems (NeurIPS). 2018. DOI: 10 . 5555 / 3326943 . 3327049. URL: https : / /
arxiv.org/abs/1808.06601v2.
[7] D. Silver et al. “Mastering the Game of GO with Deep Neural Networks and Tree
Search”. In: Nature 529.7587 (2016). DOI: 10 . 1038 / nature16961. URL: http : / /
airesearch.com/wp-content/uploads/2016/01/deepmind-mastering-go.pdf.
[8] M. Moracik et al. “DeepStack: Expert-level Artificial Intelligence in Haeds-up No-limit
Poker”. In: Science 356.6337 (2017). DOI: 10 . 1126 / science . aam6960. URL: https :
//arxiv.org/abs/1701.01724.
[9] H. Xue et al. “Deep Matrix Factorization Models for Recommender Systems”. In: Proc.
Int. Joint Conf. on Artificial Intelligence (IJCAI). 2017. DOI: 10.5555/3172077.3172336.
URL: https://www.ijcai.org/proceedings/2017/0447.pdf.
[10] F. Fusco et al. “RecoNet: An Interpretable Neural Architecture for Recommender Sys-
tems”. In: Proc. Int. Joint Conf. on Artificial Intelligence (IJCAI). 2019. DOI: 10 . 24963 /
ijcai.2019/325. URL: https://www.ijcai.org/proceedings/2019/0325.pdf.
[11] D. Ciresan et al. “Mitosis Detection in Breast Cancer Histology Images with Deep Neu-
ral Networks”. In: Proc. Int. Conf. on Medical Image Computing and Computer-Assisted
Intervention (MICCAI). 2013. DOI: 10.1007/978- 3- 642- 40763- 5_51. URL: https:
//link.springer.com/content/pdf/10.1007%5C%2F978-3-642-40763-5_51.pdf.

LITERATURE
[12] A. Estava et al. “Dermatologist-level Classification of Skin Cancer with Deep Neural
Networks”. In: Nature 542.7639 (2017). DOI: 10 . 1038 / nature21056. URL: https :
//www.ncbi.nlm.nih.gov/pmc/articles/PMC8382232/pdf/nihms-1724608.pdf.
[13] N. Ernest et al. “Genetic Fuzzy based Artificial Intelligence for Unmanned Combat
Aerial Vehicle Control in Simulated Air Combat Missions”. In: J. of Defense Management
6.1 (2016). DOI: :10.4172/2167-0374.1000144. URL: https://www.longdom.org/
open-access/genetic-fuzzy-based-artificial-intelligence-for-unmanned-
combat-aerialvehicle-control-in-simulated-air-combat-missions-2167-
0374-1000144.pdf.
[14] M. Schwarz et al. “NimbRo Picking: Versatile Part Handling for Warehouse Automa-
tion”. In: Proc. Int. Conf. on Robotics and Automation (ICRA). IEEE, 2017. DOI: 10.1109/
ICRA.2017.7989348. URL: http://www.is.uni- bonn.de/papers/ICRA_2017_
Schwarz.pdf.
[15] Y. Shrestha, V. Krishna, and G. von Krogh. “Augmenting Organizational Decision-
Making with Deep Learning Algorithms: Principles, Promises, and Challenges”. In: J.
of Business Research 123 (2021). DOI: 10 . 1016 / j . jbusres . 2020 . 09 . 068. URL:
https://arxiv.org/abs/2011.02834v1.
[16] D. Lukovnikov et al. “Neural Network-based Question Answering over Knowledge
Graphs on Word and Character Level”. In: Proc. Conf. on Word Wide Web (WWW). ACM,
2017. DOI: 10.1145/3038912.3052675. URL: https://jens-lehmann.org/files/
2017/www_nn_factoid_qa.pdf.
[17] Y. Leviathan and Y. Matias. Google Duplex: An AI System for Accomplishing Real-World
Tasks Over the Phone. Google AI blog. 2018. URL: https://ai.googleblog.com/2018/
05/duplex-ai-system-for-natural-conversation.html.
[18] T. Brown et al. “Language Models are Few-Shot Learners”. In: Proc. Advances in Neural
Information Processing Systems (NeurIPS). 2020. URL: https://arxiv.org/abs/2005.
14165v4.
[19] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning: Data Min-
ing, Inference, and Prediction. 2nd edition. Springer, 2009. DOI: 10.1007/978-0-387-
84858-7. URL: https://web.stanford.edu/~hastie/Papers/ESLII.pdf.
[20] D. MacKay. Information Theory, Inference, & Learning Algorithms. Cambridge University
Press, 2003. URL: https://www.inference.org.uk/itprnn/book.pdf.
[21] C. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. URL: https://
link.springer.com/book/9780387310732.
[22] S. Haykin. Neural Networks and Learning Machines. 3rd. Pearson, 2009. URL: http://
dai.fmph.uniba.sk/courses/NN/haykin.neural-networks.3ed.2009.pdf.
[23] Y. Bengio. “Learning Deep Architectures for AI”. In: Foundations and Trends in Machine
Learning 2.1 (2009). DOI: 10.1561/2200000006. URL: https://www.iro.umontreal.
ca/~lisa/pointeurs/TR1312.pdf.

LITERATURE
[24] R. Ramamurthy et al. “Leveraging Domain Knowledge for Reinforcement Learning Us-
ing MMC Architectures”. In: Proc. Int. Conf. on Artificial Neural Networks (ICANN). 2019.
[25] M. Lechner et al. “Neural Circuit Policies Enabling Auditable Autonomy”. In: Nature
Machine Intelligence 2 (2020). DOI: 10.1038/s42256-020-00237-3.
[26] V. Vapnik. The Nature of Statistical Learning Theory. Springer, 2000. DOI: 10.1007/978-
1 - 4757 - 3264 - 1. URL: https : / / statisticalsupportandresearch . files .
wordpress. com / 2017 / 05 / vladimir - vapnik- the - nature - of - statistical-
learning-springer-2010.pdf.
[27] J. Park. “The Concept of Transition in Quantum Mechanics”. In: Foundations of Physics
1.1 (1970). DOI: 10.1007/BF00708652.
[28] C. Bennet. “Logical Reversibility of Computation”. In: IBM J. of Research and Develop-
ment 17.6 (1973). DOI: 10.1147/rd.176.0525.
[29] A. Holevo. “Bounds for the Quantity of Information Transmitted by a Quantum Com-
munication Channel”. In: Problemy Peredachi Informatsii 9.3 (1973). URL: http://mi.
mathnet.ru/eng/ppi/v9/i3/p3.
[30] R. Ingarden. “Quantum Information Theory”. In: Reports on Mathematical Physics 10.1
(1976). DOI: 10.1016/0034-4877(76)90005-7.
[31] P. Benioff. “The Computer as a Physical System: A Microscopic Quantum Mechanical
Hamiltonian Model of Computers as Represented by Turing Machines”. In: J. of Statisti-
cal Physics 22.5 (1980). DOI: 10.1007/BF01011339.
[32] Y. Manin. Computable and Noncomputable (in Russian). Kibernetika. Moscow: Sovet-
skoye Radio, 1980.
[33] R. Feynman. “Simulating Physics with Computers”. In: Int. J. of Theoretical Physics 10
(1982). DOI: 10.1007/BF02650179.
[34] R. Feynman. “Quantum Mechanical Computers”. In: Foundations of Physics 16.6 (1986).
DOI: 10.1007/BF01886518.
[35] D. Deutsch and R. Jozsa. “Rapid Solutions of Problems by Quantum Computation”. In:
Proc. of the Royal Society London A 439 (1992). DOI: 10.1098/rspa.1992.0167.
[36] P. Shor. “Algorithms for Quantum Computation: Discrete Logarithms and Factoring”.
In: Proc. Annual Symp. on Foundations of Computer Science. IEEE, 1994. DOI: 10.1109/
SFCS.1994.365700.
[37] L. Grover. “A Fast Quantum Mechanical Algorithm for Database Search”. In: Proc. Symp.
on Theory of Computing. ACM, 1996. DOI: 10.1145/237814.237866.
[38] L. Grover. “From Schrödinger’s Equation to the Quantum Search Algorithm”. In: Quan-
tum Information Processing 56.2 (2001). DOI: 10.1007/s12043-001-0128-3.
[39] P. Shor. “Polynomial-time Algorithms for Prime Factorization and Discrete Logarithm
Problems”. In: SIAM J. on Computing 26.5 (1997). DOI: 10.1137/S0036144598347011.

LITERATURE
[40] M. Mosca. “Quantum Computer Algorithms”. PhD thesis. University of Oxford, 1999.
URL: https://www2.karlin.mff.cuni.cz/~holub/soubory/moscathesis.pdf.
[41] I. Chuang, N. Gershenfeld, and M. Kubinec. “Experimental Implementation of Fast
Quantum Searching”. In: Physical Review Letters 80.15 (1998). DOI: 10 . 1103 /
PhysRevLett.80.3408.
[42] L. Vandersypen et al. “Experimental Realization of Shor’s Quantum Factoring Algo-
rithm Using Nuclear Magnetic Resonance”. In: Nature 414.6866 (2001). DOI: 10.1038/
414883a.
[43] F. Arute et al. “Quantum Supremacy Using a Programmable Superconducting Proces-
sor”. In: Nature 574.7779 (2019). DOI: 10.1038/s41586-019-1666-5.
[44] P. Wittek. Quantum Machine Learning: What Quantum Computing Means to Data Min-
ing. Academic Press, 2014. DOI: 10.1016/C2013-0-19170-2.
[45] M. Schuld and F. Petruccione. Machine Learning with Quantum Computers. Springer,
2021. DOI: 10.1007/978-3-030-83098-4.
[46] V. Dunjiko, J. Taylor, and H. Briegel. “Quantum-Enhanced Machine Learning”. In: Phys-
ical Review Letters 117.13 (2016). DOI: 10 . 1103 / PhysRevLett . 117 . 130501. URL:
https://arxiv.org/abs/1610.08251.
[47] J. Biamonte et al. “Quantum Machine Learning”. In: Nature 549.7671 (2017). DOI: 10 .
1038/nature23474.
[48] J. Abhijith et al. “Quantum Algorithm Implementations for Beginners”. In: arXiv:1804.
03719 [cs.ET] (2020). URL: https://arxiv.org/abs/1804.03719v2.
[49] M. Cusumano. “The Business of Quantum Computing”. In: Communications of the ACM
61.10 (2018). DOI: 10.1145/3267352. URL: https://cacm.acm.org/magazines/
2018/10/231363-the-business-of-quantum-computing/fulltext.
[50] C. Bauckhage et al. Quantum Machine Learning – An Analysis of Expertise, Research, and
Applications. Fraunhofer Big Data and Artificial Intelligence Alliance. 2020.
[51] M. Bojarski et al. “The NVIDIA PilotNet Experiments”. In: arXiv:2010.08776 [cs.CV]
(2020). URL: https://arxiv.org/abs/2010.08776v1.
[52] E. Brito et al. “A Hybrid AI Tool to Extract Key Performance Indicators from Financial
Reports for Benchmarking”. In: Proc. Symp. on Document Engineering (DocEng). ACM,
2019. DOI: 10.1145/3342558.3345420. URL: https://www.researchgate.net/
profile / Eduardo - Brito / publication / 335944246 _ A _ Hybrid _ AI _ Tool _ to _
Extract _ Key _ Performance _ Indicators _ from _ Financial _ Reports _ for _
Benchmarking / links / 5dd299ab4585156b351d2e02 / A - Hybrid - AI - Tool - to -
Extract - Key - Performance - Indicators - from - Financial - Reports - for -
Benchmarking.pdf.

LITERATURE
[53] R. Sifa et al. “Towards Automated Auditing with Machine Learning”. In: Proc. Symp. on
Document Engineering (DocEng). ACM, 2019. DOI: 10.1145/3342558.3345421. URL:
https : / / www . researchgate . net / profile / Birgit - Kirsch / publication /
335943303 _ Towards _ Automated _ Auditing _ with _ Machine _ Learning / links /
5d96fe5d458515c1d391daef / Towards - Automated - Auditing - with - Machine -
Learning.pdf.
[54] R. Ramamurthy et al. “ALiBERT: Improved Automated List Inspection (ALi) with
BERT”. In: Proc. Symp. on Document Engineering (DocEng). ACM, 2021. DOI: 10.1145/
3469096 . 3474928. URL: https : / / www . researchgate . net / profile / Birgit -
Kirsch/publication/335943303_Towards_Automated_Auditing_with_Machine_
Learning/links/5d96fe5d458515c1d391daef/Towards- Automated- Auditing-
with-Machine-Learning.pdf.
[55] E. Commission. White Paper on Artificial Intelligence – A European Approach to Excel-
lence and Trust. COM(2020) 65 final. 2020. URL: https : / / ec . europa . eu / info /
publications/whitepaper- artificial- intelligence- european- approach-
excellence-and-trust_en.
[56] K. Morik et al. “Yes We Care! – Certification for Machine Learning Methods through the
Care Label Framework”. In: arXiv:2105.10197 [cs.LG] (2021). URL: https://arxiv.org/
abs/2105.10197v1.
[57] K. Beckh et al. “Explainable Machine Learning with Prior Knowledge: An Overview”. In:
arXiv:2105.10172 [cs.LG] (2021). URL: https://arxiv.org/abs/2105.10172v1.
[58] A. Holzinger et al. “Digital Transformation for Sustainable Development Goals (SDGs)-
A Security, Safety and Privacy Perspective on AI”. In: Proc. Cross-Domain Conf. for Ma-
chine Learning and Knowledge Extraction (CD-MAKE). 2021. DOI: 10.1007/978-3-030-
84060- 0_1. URL: http://eprints.cs.univie.ac.at/7013/1/AI- Digital-
Transformation-Sustainability-2021.pdf.
[59] E. Wigner. “The Unreasonable Effectiveness of Mathematics in the Natural Sciences”.
In: Communications in Pure and Applied Mathematics 13.1 (1960). DOI: 10.1002/cpa.
3160130102. URL: https://www.maths.ed.ac.uk/~v1ranick/papers/wigner.pdf.
[60] F. Wilhelm et al. Status of Quantum Computer Development. Whitepaper Federal Office
for Information Security. 2020. URL: https : / / www . bsi . bund . de / EN / Topics /
Cryptography/QuantumComputing/quantum_computing.html.
[61] J. Preskill. “Quantum Computing in the NISQ Era and Beyond”. In: Quantum 2 (2018).
DOI: 10.22331/q-2018-08-06-79.
[62] G. Vidal. “Efficient Classical Simulation of Slightly Entangled Quantum Computations”.
In: Physical Review Letters 91.14 (2003). DOI: 10.1103/PhysRevLett.91.147902.
[63] D. Aharanov et al. “Adiabatic Quantum Computation Is Equivalent to Standard Quan-
tum Computation”. In: SIAM Review 50.4 (2008). DOI: 10 . 1137 / 080734479. URL:
https://arxiv.org/abs/quant-ph/0405098v2.

LITERATURE
[64] A. Mizel, D. Lidar, and M. Mitchell. “Simple Proof of Equivalence between Adiabatic
Quantum Computation and the Circuit Model”. In: Physical Review Letters 99.7 (2007).
DOI: 10.1103/PhysRevLett.127.139901. URL: https://arxiv.org/abs/quant-
ph/0609067v2.
[65] M. Born and V. Fock. “Beweis des Adiabatensatzes”. In: Zeitschrift für Physik 51.3–4
(1928). DOI: 10.1007/BF01343193.
[66] T. Albash and D. Lidar. “Adiabatic Quantum Computation”. In: Reviews of Modern
Physics 90.1 (2018). DOI: 10.1103/RevModPhys.90.015002. URL: https://arxiv.
org/abs/1611.04471v2.
[67] Z. Bian et al. The Ising Model: Teaching an Old Problem New Tricks. Tech. rep. D-Wave
Systems, 2010. URL: https : / / www . dwavesys . com / media / vbklsvbh /
weightedmaxsat_v2.pdf.
[68] M. Johnson et al. “Quantum Annealing with Manufactured Spins”. In: Nature 473.7346
(2011). DOI: 10.1038/nature10012.
[69] T. Lanting et al. “Entanglement in a Quantum Annealing Processor”. In: Physical Review
X 4.021041 (2014). DOI: 10.1103/PhysRevX.4.021041. URL: https://arxiv.org/
abs/1401.3500v1.
[70] E. Ising. “Beitrag zur Theorie des Ferromagnetismus”. In: Zeitschrift für Physik 31.1
(1925). DOI: 10.1007/BF02980577.
[71] J. Hopfield. “Neural Networks and Physical Systems with Collective Computational
Abilities”. In: PNAS 79.8 (1982). DOI: 10 . 1073 / pnas . 79 . 8 . 2554. URL: https :
//www.pnas.org/content/pnas/79/8/2554.full.pdf.
[72] A. Lucas. “Ising Formulations of Many NP Problems”. In: Frontiers in Physics 2 (2014).
DOI: 10.3389/fphy.2014.00005. URL: https://arxiv.org/abs/1302.5843v3.
[73] C. Bauckhage et al. “Ising Models for Binary Clustering via Adiabatic Quantum Com-
puting”. In: Proc. Int. Conf. on Energy Minimization Methods in Computer Vision and Pat-
tern Recognition (EMMCVPR). Springer, 2017. DOI: 10.1007/978-3-319-78199-0_1.
[74] H. Ushijima-Mwesigwa, C. Negre, and S. Mniszewski. “Graph Partitioning Using Quan-
tum Annealing on the D-Wave System”. In: Int. Workshop on Post Moore’s Era Super-
computing. New York: ACM, 2017. DOI: 10 . 1145 / 3149526 . 3149531. URL: https :
//arxiv.org/abs/1705.03082v1.
[75] C. Bauckhage, R. Sanchez, and R. Sifa. “Problem Solving with Hopfield Networks and
Adiabatic Quantum Computing”. In: Proc. Int. Joint Conf. on Neural Networks. IEEE,
2020. DOI: 10.1109/IJCNN48605.2020.9206916.
[76] E. Farhi et al. “Quantum Computation by Adiabatic Evolution”. In: arXiv:quant-
ph/0001106 (2000). URL: https://arxiv.org/abs/quant-ph/0001106v1.
[77] M. Amin. “Effect of Local Minima on Adiabatic Quantum Optimization”. In: Physical
Review Letters 100.13 (2008). DOI: 10.1103/PhysRevLett.100.130503. URL: https:

LITERATURE
[78] J. Roland and N. Cerf. “Quantum Search by Local Adiabatic Evolution”. In: Physical Re-
view A 65.4 (2002). DOI: 10.1103/PhysRevA.65.042308. URL: https://arxiv.org/
abs/quant-ph/0107015v1.
[79] C. Bauckhage, R. Sifa, and S. Wrobel. “Adiabatic Quantum Computing for Max-Sum
Diversification”. In: Proc. SIAM Int. Conf. on Data Mining (SDM). SIAM, 2020. DOI: 10.
1137/1.9781611976236.39. URL: https://epubs.siam.org/doi/pdf/10.1137/1.
9781611976236.39.
[80] J. Johansson, P. Nation, and F. Nori. “QuTiP 2: A Python Framework for the Dynamics
of Open Quantum Systems”. In: Computer Physics Communications 184.4 (2013). DOI:
10.1016/j.cpc.2012.11.019. URL: https://arxiv.org/abs/1211.6518v1.
[81] M. A. Nielsen and I. L. Chuang. Quantum Computation and Quantum Information.
Cambridge University Press, 2016. ISBN: 978-1-10-700217-3. URL: https : / / www .
cambridge.org/de/academic/subjects/physics/quantum- physics- quantum-
information-and-quantum-computation/quantum-computation-and-quantum-
information-10th-anniversary-edition?format=HB.
[82] M. Saeedi and I. Markov. “Synthesis and Optimization of Reversible Circuits – A Sur-
vey”. In: ACM Computing Surveys 45.2 (2013). DOI: 10.1145/2431211.2431220.
[83] M. Freedman et al. “Topological Quantum Computation”. In: Bulletin of the American
Mathematical Society 40.1 (2003). DOI: 10.1090/S0273-0979-02-00964-3.
[84] E. Farhi, J. Goldstone, and S. Gutmann. “A Quantum Approximate Optimization Algo-
rithm”. In: arXiv:1411.4028 [quant-ph] (2014). URL: https://arxiv.org/abs/1411.
4028v1.
[85] A. Peruzzo et al. “A Variational Eigenvalue Solver on a Photonic Quantum Processor”.
In: Nature Communications 5.1 (2014). DOI: 10.1038/ncomms5213.
[86] M. Cerezo et al. “Variational Quantum Algorithms”. In: Nature Review Physics 3 (2021).
DOI: 10.1038/s42254-021-00348-9.
[87] J. Biamonte. “Universal Variational Quantum Computation”. In: Physical Review A 103.3
(2021). DOI: 10.1103/PhysRevA.103.L030401. URL: https://link.aps.org/doi/
10.1103/PhysRevA.103.L030401.
[88] W. Wootters and W. Zurek. “A Single Quantum Cannot be Cloned”. In: Nature 299.5886
(1982). DOI: 10.1038/299802a0.
[89] F. Leymann. “Towards a Pattern Language for Quantum Algorithms”. In: Proc. Int.
Workshop on Quantum Technology and Optimization Problems (QTOP). Springer, 2019.
DOI: 10.1007/978-3-030-14082-3_19.
[90] The Qiskit Contributors. Qiskit: An Open-source Framework for Quantum Computing.
2021. DOI: https:doi.org/10.5281/zenodo.2573505.
[91] The Cirq Contributers. Cirq, a Python Framework for Creating, Editing, and Invoking
Noisy Intermediate Scale Quantum (NISQ) Circuits. 2021. DOI: 10 . 5281 / zenodo .
4750446.

LITERATURE
[92] J. Otterbach et al. “Unsupervised Machine Learning on a Hybrid Quantum Computer”.

In: arXiv:1712.05771 [quant-ph] (2017).
[93] V. Bergholm et al. “PennyLane: Automatic Differentiation of Hybrid Quantum-classical
Computations”. In: arXiv:1811.04968 [quant-ph] (2018).
[94] R. Sutton and A. Barto. Reinforcement Learning. 2nd edition. Bradford Books, 2018.
[95] M. Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming.
Wiley-Interscience, 2005.
[96] D. Kingma and M. Welling. “Auto-Encoding Variational Bayes”. In: arXiv:1312.6114
[stat.ML] (2014). URL: https://arxiv.org/abs/1312.6114.
[97] I. Goodfellow et al. “Generative Adversarial Nets”. In: Proc. Advances in Neural Informa-
tion Processing Systems (NeurIPS). 2014. URL: https://proceedings.neurips.cc/
paper/2014/file/5ca3e9b122f61f8f06494c97b1afccf3-Paper.pdf.
[98] B. Neyshabur, R. Tomioka, and N. Srebro. “In Search of the Real Inductive Bias: On the
Role of Implicit Regularization in Deep Learning”. In: arXiv:1412.6614 [cs.LG] (2014).
URL: https://arxiv.org/abs/1412.6614v4.
[99] B. Neal et al. “A Modern Take on the Bias-Variance Tradeoff in Neural Networks”. In:
arXiv:1810.08591 [cs.LG] (2018). URL: https://arxiv.org/abs/1810.08591v4.
[100] K. van Rijsbergen. The Geometry of Information Retrieval. Cambridge University Press,
2004. DOI: 10.1017/CBO9780511543333.
[101] G. Birkhoff and J. von Neumann. “The Logic of Quantum Mechanics”. In: Annals of
Mathematics 37.4 (1936). DOI: 10.2307/1968621.
[102] J. Arrazola et al. “Quantum-inspired Algorithms in Practice”. In: Quantum 4 (2020). DOI:
10.22331/q-2020-08-13-307.
[103] A. Frieze, R. Kannan, and S. Vempala. “Fast Monte-Carlo Algorithms for Finding Low-
rank Approximations”. In: J. of the ACM 51.6 (2004). DOI: 10.1145/1039488.1039494.
URL: https://www.math.cmu.edu/~af1p/Texfiles/SVD.pdf.
[104] P. Drineas, R. Kannan, and M. Mahoney. “Fast Monte Carlo Algorithms III: Comput-
ing a Compressed Approixmate Matrix Decomposition”. In: SIAM J. on Computing 36.1
(2006). DOI: 10.1137/S0097539704442702. URL: https://www.stat.berkeley.edu/
~mmahoney/pubs/matrix3_SICOMP.pdf.
[105] D. Achlioptas and F. McSherry. “Fast Computation of Low-rank Matrix Approxima-
tions”. In: J. of the ACM 54.9 (2007). DOI: 10 . 1145 / 1219092 . 1219097. URL: https :
//www.cs.princeton.edu/courses/archive/spr04/cos598B/bib/Mcsherry-
svd.pdf.
[106] M. Mahoney and P. Drineas. “CUR Matrix Decompositions for Improved Data Analysis”.
In: PNAS 106.3 (2009). DOI: 10.1073/pnas.0803205106. URL: https://www.pnas.
org/content/pnas/106/3/697.full.pdf.

LITERATURE
[107] C. Thurau, K. Kersting, and C. Bauckhage. “Deterministic CUR for Improved Large-
Scale Data Analysis: An Empirical Study”. In: Proc. Int. Conf. on Data Mining. SIAM,
2012. DOI: 10 . 1137 / 1 . 9781611972825 . 59. URL: https : / / epubs . siam . org /
doi/pdf/10.1137/1.9781611972825.59.
[108] R. Kannan and S. Vempala. “Randomized Algorithms in Numerical Linear Algebra”. In:
Acta Numerica 26 (2017). DOI: 10.1017/S0962492917000058. URL: https://www.cc.
gatech.edu/~vempala/papers/acta_survey.pdf.
[109] R. Salakhutdinov and A. Mnih. “Probabilistic Matrix Factorization”. In: Proc. Ad-
vances in Neural Information Processing Systems (NeurIPS). 2007. DOI: 10 . 5555 /
2981562 . 2981720. URL: https : / / papers . nips . cc / paper / 2007 / file /
d7322ed717dedf1eb4e6e52a37ea7bcd-Paper.pdf.
[110] G. Zhang. “Quantum-inspired Evolutionary Algorithms: A Survey and Empirical
Study”. In: J. of Heuristics 17 (2011). DOI: 10.1007/s10732-010-9136-0.
[111] E. Tang. “A Quantum-Inspired Classical Algorithm for Recommendation Systems”. In:
Proc. Symp. on the Theory of Computing (STOC). ACM, 2019. DOI: 10.1145/3313276.
3316310.
[112] I. Kerenidis and A. Prakash. “Quantum Recommendation Systems”. In: Proc. Innova-
tions in Theoretical Computer Science Conf. (ITCS). 2017. DOI: 10.4230/LIPIcs.ITCS.
2017.49. URL: https://arxiv.org/abs/1603.08675v3.
[113] A. Kitaev. “Quantum Measurements and the Abelian Stabilizer Problem”. In:
arXiv:quant-ph/9511026 (1995). URL: https://arxiv.org/abs/quant-ph/9511026.
[114] R. Cleve et al. “Quantum Algorithms Revisited”. In: Proc. of the Royal Society of Lon-
don. Series A: Mathematical, Physical and Engineering Sciences 454.1969 (1998). DOI:
10 . 1098 / rspa . 1998 . 0164. URL: https : / / royalsocietypublishing . org /
doi/abs/10.1098/rspa.1998.0164.
[115] D. S. Abrams and S. Lloyd. “Quantum Algorithm Providing Exponential Speed Increase
for Finding Eigenvalues and Eigenvectors”. In: Physical Review Letters 83.24 (1999). DOI:
10.1103/PhysRevLett.83.5162. URL: https://link.aps.org/doi/10.1103/
[116] N. Chia et al. “Quantum-inspired Algorithms for Solving Low-rank Linear Equation
Systems with Logarithmic Dependence on the Dimension”. In: Int. Symp. on Algorithms
and Computation (ISAAC). 2020. DOI: 10.4230/LIPIcs.ISAAC.2020.47.
[117] A. Gilyen et al. “Quantum Singular Value Transformation and Beyond: Exponential
Improvements for Quantum Matrix Arithmetics”. In: Proc. Symp. on the Theory of Com-
puting (STOC). ACM, 2019. DOI: 10.1145/3313276.3316366.
[118] N. Chia, H. Lin, and C. Wang. “Quantum-inspired Sublinear Classical Algorithms for
Solving Low-rank Linear Systems”. In: arXiv:1811.04852 [cs.DS] (2018).

LITERATURE
[119] A. W. Harrow, A. Hassidim, and S. Lloyd. “Quantum Algorithm for Linear Systems of
Equations”. In: Physical Review Letters 103 (15 2009). DOI: 10.1103/PhysRevLett.103.
150502. URL: https://arxiv.org/abs/0811.3171v3.
[120] N. Chepurko et al. “Quantum-Inspired Algorithms from Randomized Numerical Linear
Algebra”. In: arXiv:2011.04125 [cs.DS] (2020).
[121] R. Orus. “A Practical Introduction to Tensor Networks: Matrix Product States and Pro-
jected Entangled Pair States”. In: Annals of Physics 349 (2014). DOI: 10.1016/j.aop.
2014.06.013. URL: https://arxiv.org/abs/1306.2164.
[122] J. Biamonte and V. Bergholm. “Tensor Networks in a Nutshell”. In: arXiv:1708.00006
[quant-ph] (2017).
[123] A. Cichocki et al. “Tensor Networks for Dimensionality Reduction and Large-scale Op-
timization: Part 1 Low-Rank Tensor Decompositions”. In: Foundations and Trends in
Machine Learning 9.4–5 (2016). DOI: 10.1561/2200000059.
[124] A. Cichocki et al. “Tensor Networks for Dimensionality Reduction and Large-scale Op-
timization: Part 2 Applications and Future Perspectives”. In: Foundations and Trends in
Machine Learning 9.6 (2017). DOI: 10.1561/2200000067.
[125] L. De Lathauwer, B. De Moor, and J. Vanderwalle. “A Multilinear Singular Value Decom-
position”. In: SIAM J. on Matrix Analysis and Applications 21.4 (2000). DOI: /10.1137/
S0895479896305696.
[126] R. Sifa et al. “Matrix and Tensor Factorization Based Game Content Recommender Sys-
tems: A Bottom-Up Architecture and a Comparative Online Evaluation”. In: Proc. Conf.
on Artificial Intelligence and Interactive Digital Entertainment (AIIDE). AAAI, 2018.
[127] C. Bauckhage et al. “Fast Learning for Customizable Head Pose Recognition in Robotic
Wheelchair Control”. In: Proc. Int. Conf. Automatic Face and Gesture Recognition (FG).
IEEE, 2006.
[128] L. Hillebrand et al. “Interpretable Topic Extraction and Word Embedding Learning Us-
ing Non-Negative Tensor DEDICOM”. In: Machine Learning and Knowledge Extraction
3.1 (2021). DOI: 10.3390/make3010007.
[129] H. Lin, M. Tegmark, and D. Rolnick. “Why Does Deep and Cheap Learning Work So
Well?” In: J. of Statistical Physics 168 (2017). DOI: 10.1007/s10955-017-1836-5.
[130] E. Stoudenmire and D. Schwab. “Supervised Learning with Tensor Networks”. In: Proc.
Advances in Neural Information Processing Systems (NeurIPS). 2016.
[131] I. Glasser, N. Pancotti, and J. Cirac. “From Probabilistic Graphical Models to Generalized
Tensor Networks for Supervised Learning”. In: IEEE Access 8 (2020). DOI: 10 . 1109 /
ACCESS.2020.2986279.
[132] C. Roberts et al. “TensorNetwork: A Library for Physics and Machine Learning”. In:
arXiv:1905.01330 [physics.comp-ph] (2019).

LITERATURE
[133] S. Mücke, N. Piatkowski, and K. Morik. “Hardware Acceleration of Machine Learning

Beyond Linear Algebra”. In: Proc. European Conf. on Machine Learning and Principles
and Practice of Knowledge Discovery in Databases (ECML/PKDD). 2019. DOI: 10.1007/
978-3-030-43823-4_29.
[134] J. Boyd. “Fujitsu’s CMOS Digital Annealer Produces Quantum Computer Speeds”. In:
IEEE Spectrum May (2018).
[135] M. Aramon et al. “Physics-Inspired Optimization for Quadratic Unconstrained Prob-
lems Using a Digital Annealer”. In: Frontiers in Physics 7 (2019). DOI: 10.3389/fphy.
2019.00048. URL: https://www.frontiersin.org/article/10.3389/fphy.2019.
00048.
[136] S. Mücke, N. Piatkowski, and K. Morik. “Learning Bit by Bit: Extracting the Essence of
Machine Learning”. In: Proc. Conf. Learning, Knowledge, Data, Analytics (KDML-LWDA).
2019. URL: http://ceur-ws.org/Vol-2454/paper_51.pdf.
[137] D. Horn and A. Gottlieb. “Algorithm for Data Clustering in Pattern Recognition Prob-
lems Based on Quantum Mechanics”. In: Physical Review Letters 88.1 (2001). DOI: 10 .
1103/PhysRevLett.88.018702. URL: https://link.aps.org/doi/10.1103/
[138] M. Weinstein and D. Horn. “Dynamic Quantum Clustering: A Method for Visual Ex-
ploration of Structures in Data”. In: Physical Review E 80.6 (2009). DOI: 10 . 1103 /
PhysRevE.80.066117. URL: https://link.aps.org/doi/10.1103/PhysRevE.80.
066117.
[139] H. Nezamabadi-pour. “A Quantum-inspired Gravitational Search Algorithm for Binary
Encoded Optimization Problems”. In: Engineering Applications of Artificial Intelligence 4
(2015). DOI: 10.1016/j.engappai.2015.01.002.
[140] J. Lou et al. “Failure Prediction by Relevance Vector Regression with Improved
Quantum-inspired Gravitational Search”. In: J. of Network and Computer Applications
103 (2018). DOI: 10.1016/j.jnca.2017.11.013.
[141] B. Rubinstein. “Evolving Quantum Circuits Using Genetic Programming”. In: Proc.
Congress on Evolutionary Computation (CEC). IEEE, 2001. DOI: 10.1109/CEC.2001.
934383.
[142] A. Leier. “Evolution of Quantum Algorithms using Genetic Programming”. PhD thesis.
University of Dortmund, 2004.
[143] L. Franken et al. “Gradient-free Quantum Optimization on NISQ Devices”. In:
arXiv:2012.13453 [quant-ph] (2020).
[144] D. Panchenko. The Sherrington-Kirkpatrick Model. Springer, 2013. DOI: 10.1007/978-
1-4614-6289-7.
[145] K. McKiernan et al. “Automated Quantum Programming via Reinforcement Learning
for Combinatorial Optimization”. In: arXiv:1908.08054 [quant-ph] (2019).

LITERATURE
[146] K. Guy and G. Perdue. Using Reinforcement Learning to Optimize Quantum Circuits in
the Presence of Noise. Tech. rep. Batavia, IL, USA: Fermi National Accelerator Lab, 2020.
[147] T. Fösel et al. “Quantum Circuit Optimization with Deep Reinforcement Learning”. In:
arXiv:2103.07585 [quant-ph] (2021).
[148] L. Cincio et al. “Machine Learning of Noise-Resilient Quantum Circuits”. In: Physical
Review X Quantum 2 (1 2021). DOI: 10 . 1103 / PRXQuantum . 2 . 010324. URL: https :
//link.aps.org/doi/10.1103/PRXQuantum.2.010324.
[149] M. Pirhooshyaran and T. Terlaky. “Quantum Circuit Design Search”. In: Quantum Ma-
chine Intelligence 3.25 (2021). DOI: 10.1007/s42484-021-00051-z.
[150] A. Zulehner, A. Paler, and R. Wille. “An Efficient Methodology for Mapping Quantum
Circuits to the IBM QX Architectures”. In: IEEE Trans. on Computer-Aided Design of Inte-
grated Circuits and Systems 38.7 (2019). DOI: 10.1109/TCAD.2018.2846658.
[151] G. Li, Y. Ding, and Y. Xie. “Tackling the Qubit Mapping Problem for NISQ-Era Quantum
Devices”. In: Proc. Int. Conf. on Architectural Support for Programming Languages and
Operating Systems (ASPLOS). ACM, 2019. DOI: 10 . 1145 / 3297858 . 3304023. URL:
https://doi.org/10.1145/3297858.3304023.
[152] G. Acampora and R. Schiattarella. “Deep Neural Networks for Quantum Circuit Map-
ping”. In: Neural Computing and Applications 33 (2021). DOI: 10.1007/s00521- 021-
06009-3.
[153] J. Kusyk, S. Saeed, and M. Uyar. “Survey on Quantum Circuit Compilation for Noisy
Intermediate-Scale Quantum Computers: Artificial Intelligence to Heuristics”. In: IEEE
Trans. on Quantum Engineering 2 (2021). DOI: 10.1109/TQE.2021.3068355.
[154] L. K. Grover and T. Rudolph. “Creating superpositions that correspond to efficiently
integrable probability distributions”. In: CoRR abs/quant-ph/0208112v1 (2002). arXiv:
0208112v1. URL: https://arxiv.org/abs/quant-ph/0208112v1.
[155] C. Zoufal, A. Lucchi, and S. Woerner. “Quantum Generative Adversarial Networks for
learning and loading random distributions”. In: npj Quantum Information 5.1 (2019).
DOI: 10.1038/s41534-019-0223-2. URL: https://arxiv.org/abs/1904.00043v2.
[156] R. Harper, S. T. Flammia, and J. J. Wallman. “Efficient Learning of Quantum Noise”. In:
Nature Physics 16.12 (2020). DOI: 10 . 1038 / s41567 - 020 - 0992 - 8. URL: https : / /
arxiv.org/abs/1907.13022v2.
[157] J. M. Martinis. “Qubit metrology for building a fault-tolerant quantum computer”. In:
npj Quantum Information 1.1 (2015). ISSN: 2056-6387. DOI: 10.1038/npjqi.2015.5.
[158] E. Aïmeur, G. Brassard, and S. Gambs. “Machine Learning in a Quantum World”. In:
Proc. Canadian Conf. on Artificial Intelligence. 2006. DOI: 10.1007/11766247_37.
[159] D. Riste et al. “Demonstration of Quantum Advantage in Machine Learning”. In: npj
Quantum Information 3.16 (2017). DOI: 10.1038/s41534-017-0017-3.
[160] W. O’Quinn and S. Mao. “Quantum Machine Learning: Recent Advances and Outlook”.
In: IEEE Wireless Communications 27.3 (2020). DOI: 10.1109/MWC.001.1900341.

LITERATURE
[161] D. Coppersmith. “An Approximate Fourier Transform Useful in Quantum Factoring”.

In: arXiv:quant-ph/0201067 (2002). URL: https : / / arxiv . org / abs / quant - ph /
0201067.
[162] A. Ambainis. “Quantum Walk Algorithm for Element Distinctness”. In: Symp. on Foun-
dations of Computer Science(FOCS). IEEE, 2004. DOI: 10.1109/FOCS.2004.54.
[163] M. Szegedy. “Quantum Speed-up of Markov Chain Based Algorithms”. In: Symp. on
Foundations of Computer Science (FOCS). IEEE, 2004. DOI: 10.1109/FOCS.2004.53.
[164] L. Grover and T. Rudolph. “Creating Superpositions that Correspond to Efficiently In-
tegrable Probability Distributions”. In: arXiv:quant-ph/0208112 (2002). URL: https :
//arxiv.org/abs/quant-ph/0208112.
[165] A. M. Childs, R. Kothari, and R. D. Somma. “Quantum Algorithm for Systems of Linear
Equations with Exponentially Improved Dependence on Precision”. In: SIAM Journal on
Computing 46.6 (2017). DOI: 10.1137/16M1087072.
[166] X.-D. Cai et al. “Experimental Quantum Computing to Solve Systems of Linear Equa-
tions”. In: Physical Review Letters 110.23 (2013). DOI: 10 . 1103 / PhysRevLett . 110 .
230501. URL: https://link.aps.org/doi/10.1103/PhysRevLett.110.230501.
[167] J. Pan et al. “Experimental Realization of Quantum Algorithm for Solving Linear Sys-
tems of Equations”. In: Physical Review A 89.2 (2014). DOI: 10 . 1103 / PhysRevA . 89 .
022313. URL: https://link.aps.org/doi/10.1103/PhysRevA.89.022313.
[168] Z. Zhao et al. “Bayesian Deep Learning on a Quantum Computer”. In: Quantum Ma-
chine Intelligence 1 (2019). DOI: 10.1007/s42484-019-00004-7.
[169] G. Golub and J. van Loan. Matrix Computations. Johns Hopkins University Press, 1996.
[170] P. Rebentrost et al. “Quantum Singular-Value Decomposition of Nonsparse Low-rank
Matrices”. In: Phys. Rev. A 97.1 (2018). DOI: 10 . 1103 / PhysRevA . 97 . 012327. URL:
https://link.aps.org/doi/10.1103/PhysRevA.97.012327.
[171] G. Low and I. Chuang. “Hamiltonian Simulation by Qubitization”. In: Quantum 3 (2019).
DOI: 10.22331/q-2019-07-12-163.
[172] B. D. Clader, B. C. Jacobs, and C. R. Sprouse. “Preconditioned Quantum Linear System
Algorithm”. In: Physical Review Letters 110 (2013). DOI: 10.1103/PhysRevLett.110.
[173] H.-Y. Huang, K. Bharti, and P. Rebentrost. “Near-term quantum algorithms for linear
systems of equations”. In: arXiv.org [quant-ph] (2019). URL: https://arxiv.org/abs/
1909.07344v2.
[174] Y. Subas, R. D. Somma, and D. Orsucci. “Quantum Algorithms for Systems of Linear
Equations Inspired by Adiabatic Quantum Computing”. In: Physical Review Letters 122
(6 2019). DOI: 10.1103/PhysRevLett.122.060504. URL: https://arxiv.org/abs/
1805.10549v2.

LITERATURE
[175] Y. Lee, J. Joo, and S. Lee. “Hybrid quantum linear equation algorithm and its experimen-
tal test on IBM Quantum Experience”. In: Scientific Reports 9 (2019). ISSN: 2045-2322.
DOI: 10.1038/s41598-019-41324-9. URL: https://www.nature.com/articles/
s41598-019-41324-9.pdf.
[176] C. Bravo-Prieto et al. “Variational Quantum Linear Solver”. In: arXiv.org [quant-ph]
(2020). URL: https://arxiv.org/abs/1909.05820v2.
[177] X. Xu et al. “Variational algorithms for linear algebra”. In: Science Bulletin 66.21 (2021).
ISSN: 2095-9273. DOI: 10.1016/j.scib.2021.06.023. URL: https://arxiv.org/
abs/1909.03898v2.
[178] N. Wiebe, D. Braun, and S. Lloyd. “Quantum Algorithm for Data Fitting”. In: Physical
Review Letters 109.5 (2012). DOI: 10.1103/PhysRevLett.109.050505. URL: https:
//link.aps.org/doi/10.1103/PhysRevLett.109.050505.
[179] M. Schuld, I. Sinayskiy, and F. Petruccione. “Prediction by Linear Regression on a Quan-
tum Computer”. In: Physical Review A 94 (2016). DOI: 10.1103/PhysRevA.94.022342.
[180] G. Wang. “Quantum Algorithm for Linear Regression”. In: Physical Review A 96 (2017).
DOI: 10.1103/PhysRevA.96.012335. URL: https://arxiv.org/abs/1402.0660v6.
[181] C.-H. Yu, F. Gao, and Q.-Y. Wen. “An Improved Quantum Algorithm for Ridge Regres-
sion”. In: IEEE Trans. on Knowledge and Data Engineering 33.3 (2021). DOI: 10 . 1088 /
1674-1056/ac1b84.
[182] Y.-Y. Hou et al. “Quantum Partial Least Squares Regression Algorithm for Multiple Cor-
relation Problem”. In: Chinese Physics B (2021). DOI: 10.1088/1674-1056/ac1b84.
[183] Y. Liu and S. Zhang. “Fast Quantum Algorithms for Least Squares Regression and Statis-
tic Leverage Scores”. In: Theoretical Computer Science 657 (2016). DOI: 10.1016/j.tcs.
2016.05.044. URL: http://www.cse.cuhk.edu.hk/~syzhang/papers/QLS.pdf.
[184] A. Gilyen, Z. Song, and E. Tang. “An Improved Quantum-inspired Algorithm for Linear
Regression”. In: arXiv:2009.07268 [cs.DS] (2020). URL: https://arxiv.org/abs/2009.
07268.
[185] P. Date and T. Potok. “Adiabatic Quantum Linear Regression”. In: Scientific Reports 11
(2021). DOI: 110.1038/s41598- 021- 01445- 6. URL: https://www.nature.com/
articles/s41598-021-01445-6.pdf.
[186] D. Aloise et al. “NP-Hardness of Euclidean Sum-of-Squares Clustering”. In: Machine
Learning 75.2 (2009). DOI: 10 . 1007 / s10994 - 009 - 5103 - 0. URL: https : / / link .
springer.com/content/pdf/10.1007/s10994-009-5103-0.pdf.
[187] C. Bauckhage et al. “Adiabatic Quantum Computing for Kernel k=2 Means Clustering”.
In: Proc. Conf. Learning, Knowledge, Data, Analytics (KDML-LWDA). 2018. DOI: http :
/ / publica . fraunhofer . de / documents / N - 520827 . html. URL: http : / / ceur -
ws.org/Vol-2191/paper3.pdf.

LITERATURE
[188] M. Jünger et al. “Performance of a Quantum Annealer for Ising Ground State Compu-
tations on Chimera Graphs”. In: arXiv:1904.11965 [cs.DS] (2019). URL: https://arxiv.
org/abs/1904.11965v1.
[189] D. Arthur and P. Date. “Balanced k-Means Clustering on an Adiabatic Quantum Com-
puter”. In: Quantum Information Processing 20 (2021). DOI: 10 . 1007 / s11128 - 021 -
03240-8. URL: https://arxiv.org/abs/2008.04419v1.
[190] P. Date, D. Arthur, and L. Pusey-Nazzaro. “QUBO Formulations for Training Machine
Learning Models”. In: Scientific Reports 11 (2021). DOI: 10.1038/s41598-021-89461-4.
URL: https://www.nature.com/articles/s41598-021-89461-4.pdf.
[191] C. Bauckhage et al. “A QUBO Formulation of the k-Medoids Problem”. In: Proc. Conf.
Learning, Knowledge, Data, Analytics (KDML-LWDA). 2019. DOI: http : / / publica .
fraunhofer.de/dokumente/N- 562211.html. URL: http://ceur- ws.org/Vol-
2454/paper_39.pdf.
[192] S. W. Hong et al. “Market Graph Clustering via QUBO and Digital Annealing”. In: J.
of Risk and Financial Management 14.1 (2021). DOI: 10 . 3390 / jrfm14010034. URL:
https://www.mdpi.com/1911-8074/14/1/34.
[193] E. Aïmeur, G. Brassard, and S. Gambs. “Quantum Clustering Algorithms”. In: Proc. Int.
Conf. on Machine Learning (ICML). 2007. DOI: 10 . 1145 / 1273496 . 1273497. URL:
https://icml.cc/imls/conferences/2007/proceedings/papers/518.pdf.
[194] E. Aïmeur, G. Brassard, and S. Gambs. “Quantum Speed-up for Unsupervised Learning”.
In: Machine Learning 90.2 (2013). DOI: 10.1007/s10994-012-5316-5.
[195] T. Tomesh et al. “Coreset Clustering on Small Quantum Computers”. In: Electronics
10.14 (2021). DOI: 10 . 3390 / electronics10141690. URL: https : / / www . mdpi .
com/2079-9292/10/14/1690.
[196] N. Wiebe, A. Kapoor, and K. Svore. “Quantum Algorithms for Nearest-Neighbor Meth-
ods for Supervised and Unsupervised Learning”. In: Quantum Information & Computa-
tion 15.3–4 (2015). DOI: 10.5555/2871393.2871400. URL: https://arxiv.org/abs/
1401.2142v2.
[197] C. Dürr and P. Høyer. “A Quantum Algorithm for Finding the Minimum”. In:
arXiv:quant-ph/9607014 (1999).
[198] S. Lloyd, M. Mohseni, and P. Rebentrost. “Quantum Algorithms for Supervised and
Unsupervised Machine Learning”. In: arXiv:1307.0411 [quant-ph] (2013). URL: https :
[199] A. Basheer, A. A‫ﮔ‬am, and S. Goyal. “Quantum k Nearest Neighbor Algorithm”. In:
arXiv:2003.09187 [quant-ph] (2020). URL: https://arxiv.org/abs/2003.09187.
[200] K. Mitarai, M. Kitagawa, and K. Fujii. “Quantum Analog-Digital Conversion”. In: Physi-
cal Review A 99.1 (2019). DOI: 10.1103/PhysRevA.99.012301. URL: https://link.
aps.org/doi/10.1103/PhysRevA.99.012301.

LITERATURE
[201] M. Schuld, M. Fingerhuth, and F. Petruccione. “Implementing a Distance-based Clas-

sifier with a Quantum Interference Circuit”. In: Europhysics Letters 119.6 (2017). DOI:
10.1209/0295-5075/119/60002.
[202] B. Schölkopf and A. Smola. Learning with Kernels: Support Vector Machines, Regulariza-
tion, Optimization, and Beyond. MIT Press, 2002. URL: https://ieeexplore.ieee.
org/servlet/opac?bknumber=6267332.
[203] M. Schuld, I. Sinayskiy, and F. Petruccione. “Quantum Computing for Pattern Classifi-
cation”. In: Proc. Pacific Rim Int. Conf. on Artificial Intelligence (PRICAI). Springer, 2014.
DOI: 10.1007/978-3-319-13560-1_17.
[204] M. Schuld, I. Sinayskiy, and F. Petruccione. “Simulating a Perceptron on a Quantum
Computer”. In: Physics Letters A 379.7 (2015). DOI: https://doi.org/10.1016/j.
physleta . 2014 . 11 . 061. URL: https : / / www . sciencedirect . com / science /
article/pii/S037596011401278X.
[205] C. A. Trugenberger. “Probabilistic Quantum Memories”. In: Physical Review Letters 87.6
(2001). DOI: 10.1103/PhysRevLett.87.067901. URL: https://link.aps.org/doi/
10.1103/PhysRevLett.87.067901.
[206] C. A. Trugenberger. “Quantum Pattern Recognition”. In: Quantum Information Process-
ing 1.6 (2002). DOI: 10.1023/A:1024022632303.
[207] N. Wiebe, A. Kapoor, and K. Svore. “Quantum Perceptron Models”. In: Proc. Advances
in Neural Information Processing Systems (NeurIPS). 2016. DOI: 10 . 5555 / 3157382 .
[208] F. Tacchino et al. “An Artificial Neuron Implemented on an Actual Quantum Processor”.
In: npj Quantum Information 5 (2019). DOI: 10.1038/s41534-019-0140-4.
[209] M. Schuld and N. Killoran. “Quantum Machine Learning in Feature Hilbert Spaces”. In:
Physical Review Letters 122.4 (2019). DOI: 10.1103/PhysRevLett.122.040504. URL:
https://link.aps.org/doi/10.1103/PhysRevLett.122.040504.
[210] E. Grant et al. “Hierarchical Quantum Classifiers”. In: njp Quantum Information 4 (2018).
DOI: 10.1038/s41534-018-0116-9.
[211] W. Huggins et al. “Towards Quantum Machine Learning with Tensor Networks”. In:
Quantum Science and Technology 4.2 (2019). DOI: 10.1088/2058-9565/aaea94.
[212] Y. Freund and R. Schapire. “A Decision-Theoretic Generalization of On-Line Learning
and an Application to Boosting”. In: J. of Computer and System Sciences 55.1 (1997). DOI:
10 . 1006 / jcss . 1997 . 1504. URL: https : / / www . cis . upenn . edu / ~mkearns /
teaching/COLT/adaboost.pdf.
[213] P. Viola and M. Jones. “Robust Real-Time Face Detection”. In: Int. J. of Computer Vision
57 (2004). DOI: 10 . 1023 / B : VISI . 0000013087 . 49260 . fb. URL: https : / / www .
researchgate . net / profile / Michael - Jones - 66 / publication / 220660094 _
Robust_Real-Time_Face_Detection/links/02bfe50d33d12e86ed000000/Robust-
Real-Time-Face-Detection.pdf.

LITERATURE
[214] H. Neven et al. “QBoost: Large Scale Classifier Training with Adiabatic Quantum Op-
timization”. In: Proc. Asian Conf. on Machine Learning (ACML). 2012. URL: http : / /
proceedings.mlr.press/v25/neven12/neven12.pdf.
[215] K. Pudenz and D. Lidar. “Quantum Adiabatic Machine Learning”. In: Quantum Informa-
tion Processing 12.5 (2013). DOI: 10.1007/s11128-012-0506-4.
[216] M. Schuld and F. Petruccione. “Quantum Ensembles of Quantum Classifiers”. In: Scien-
tific Reports 8.2772 (2018). DOI: 10.1038/s41598-018-20403-3.
[217] B. E. Boser, I. Guyon, and V. Vapnik. “A Training Algorithm for Optimal Margin Classi-
fiers”. In: Proc. Conf. on Computational Learning Theory (COLT). 1992. DOI: 10 . 1145 /
130385 . 130401. URL: https : / / www . researchgate . net / profile / Bernhard -
Boser/publication/2376111_A_Training_Algorithm_for_Optimal_Margin_
Classifier/links/560eccc208ae0fc513ee8fc9/A- Training- Algorithm- for-
Optimal-Margin-Classifier.pdf.
[218] V. Havlicek et al. “Supervised Learning with Quantum-enhanced Feature Spaces”. In:
Nature 567.7747 (2019). DOI: 10.1038/s41586-019-0980-2.
[219] P. Wittek. “Supervised Learning and Support Vector Machines”. In: Quantum Machine
Learning. Academic Press, 2014. DOI: 10.1016/B978-0-12-800953-6.00015-3. URL:
https://www.sciencedirect.com/science/article/pii/B9780128009536000153.
[220] H. K. Wang et al. “Application of the Least Squares Support Vector Machine Based on
Quantum Particle Swarm Optimization for Data Fitting of Small Samples”. In: Mechan-
ical Science and Engineering IV. Vol. 472. 2014. DOI: 10.4028/www.scientific.net/
AMM.472.485.
[221] P. Rebentrost, M. Mohseni, and S. Lloyd. “Quantum Support Vector Machine for
Big Data Classification”. In: Physical Review Letters 113 (2014). DOI: 10 . 1103 /
physrevlett.113.130503.
[222] C. Havenstein, D. Thomas, and S. Chandrasekaran. “Comparisons of Performance be-
tween Quantum and Classical Machine Learning”. In: SMU Data Science Review 1.4
(2018).
[223] P. V. Zahorodko et al. “Comparisons of performance between quantum-enhanced and
classical machine learning algorithms on the IBM Quantum Experience”. In: Journal of
Physics: Conference Series 1840 (2021). DOI: 10.1088/1742-6596/1840/1/012021.
[224] K. Hornik, M. Stinchcombe, and H. White. “Multilayer Feedforward Networks are Uni-
versal Approximators”. In: Neural Networks 2 (1989).
[225] D. Rumelhart, G. Hinton, and R. Williams. “Learning Representations by Back-
propagating Errors”. In: Nature 323.6088 (1986). DOI: 10.1038/323533a0.
[226] J. Allcock et al. “Quantum Algorithms for Feedforward Neural Networks”. In: ACM
Trans. on Quantum Computing (2018). DOI: 10.1145/3411466. URL: https://arxiv.
org/abs/1812.03089v2.

LITERATURE
[227] K. Beer et al. “Training deep quantum neural networks”. In: Nature Communications 11
(2020). DOI: 10 . 1038 / s41467 - 020 - 14454 - 2. URL: https : / / www . nature . com /
articles/s41467-020-14454-2.pdf.
[228] K. Mitarai et al. “Quantum Circuit Learning”. In: Physical Review A 98.3 (2018). DOI: 10.
1103/PhysRevA.98.032309. URL: https://link.aps.org/doi/10.1103/PhysRevA.
98.032309.
[229] G. Verdon et al. “Learning to learn with quantum neural networks via classical neural
networks”. In: arXiv:1907.05415 [quant-ph] (2019). URL: https : / / arxiv . org / abs /
1907.05415v1.
[230] J. R. McClean et al. “Barren Plateaus in Quantum Neural Network Training Landscapes”.
In: Nature Communications 9 (2018). DOI: 10.1038/s41467-018-07090-4.
[231] D. Zhu et al. “Training of Quantum Circuits on a Hybrid Quantum Computer”. In: Sci-
ence Advances 5.10 (2019). DOI: 10.1126/sciadv.aaw9918.
[232] S. Debnath et al. “Demonstration of a Small Programmable Quantum Computer with
Atomic Qubits”. In: Nature 4.536 (2016). DOI: 10.1038/nature18648.
[233] V. Leyton-Ortega, A. Perdomo-Ortiz, and O. Perdomo. “Robust Implementation of
Generative Modeling with Parametrized Quantum Circuits”. In: Quantum Machine In-
telligence 3 (2020). DOI: 10.1007/s42484-021-00040-2.
[234] M. Rossi et al. “Quantum Hypergraph States”. In: New J. of Physics 15.11 (2013). DOI: 10.
1088 / 1367 - 2630 / 15 / 11 / 113022. URL: https : / / doi . org / 10 . 1088 / 1367 -
2630/15/11/113022.
[235] Z. Zhao, J. K. Fitzsimons, and J. F. Fitzsimons. “Quantum-assisted Gaussian Process Re-
gression”. In: Phys. Rev. A 99.5 (2019). DOI: 10 . 1103 / PhysRevA . 99 . 052331. URL:
https://link.aps.org/doi/10.1103/PhysRevA.99.052331.
[236] P. Dayan et al. “The Helmholtz Machine”. In: Neural Computation 7.5 (1995). DOI: 10.
1162/neco.1995.7.5.889. URL: https://doi.org/10.1162/neco.1995.7.5.889.
[237] G. Hinton et al. “The “Wake-Sleep” Algorithm for Unsupervised Neural Networks”. In:
Science 268.5214 (1995). DOI: 10 . 1126 / science . 7761831. URL: https : / / www .
science.org/doi/abs/10.1126/science.7761831.
[238] T. van Dam et al. “Hybrid Helmholtz Machines: A Gate-based Quantum Circuit Imple-
mentation”. In: Quantum Information Processing 19 (2020). DOI: 10 . 1007 / s11128 -
020-02660-2.
[239] M. H. Amin et al. “Quantum Boltzmann Machine”. In: Phys. Rev. X 8.2 (2018). DOI: 10.
1103/PhysRevX.8.021050. URL: https://link.aps.org/doi/10.1103/PhysRevX.
8.021050.
[240] V. Dixit et al. “Training Restricted Boltzmann Machines With a D-Wave Quantum An-
nealer”. In: Frontiers in Physics 9 (2021). DOI: 10 . 3389 / fphy . 2021 . 589626. URL:
https://www.frontiersin.org/article/10.3389/fphy.2021.589626.

LITERATURE
[241] S. Cheng, J. Chen, and L. Wang. “Information Perspective to Probabilistic Modeling:

Boltzmann Machines versus Born Machines”. In: Entropy 20 (2018). DOI: 10 . 3390 /
e20080583.
[242] J.-G. Liu and L. Wang. “Differentiable Learning of Quantum Circuit Born Machines”. In:
Physical Review A 98.6 (2018). DOI: 10 . 1103 / PhysRevA . 98 . 062324. URL: https :
//link.aps.org/doi/10.1103/PhysRevA.98.062324.
[243] B. Coyle et al. “The Born supremacy: quantum advantage and training of an Ising Born
machine”. In: njp Quantum Information 6 (2019). DOI: 10.1038/s41534-020-00288-9.
[244] S. Y.-C. Chen and S. Yoo. “Federated Quantum Machine Learning”. In: Entropy (2021).
DOI: 10.3390/e23040460. URL: https://arxiv.org/abs/2103.12010v1.
[245] Y.-B. Sheng and L. Zhou. “Distributed secure quantum machine learning”. In: Science
Bulletin 62 (2017). DOI: 10.1016/j.scib.2017.06.007.
[246] M. Benedetti et al. “Parameterized Quantum Circuits as Machine Learning Models”. In:
Quantum Science and Technology 4.4 (2019). DOI: 10.1088/2058-9565/ab4eb5.
[247] J. D. Whitfield, J. Biamonte, and A. Aspuru-Guzik. “Simulation of Electronic Structure
Hamiltonians Using Quantum Computers”. In: Molecular Physics 109.5 (2011). DOI: 10.
1080/00268976.2011.552441.
[248] A. Dalzell et al. “How Many Qubits Are Needed for Quantum Computational
Supremacy?” In: Quantum 4 (2020). DOI: 10.22331/q-2020-05-11-264.
[249] V. Akshay et al. “Reachability Deficits in Quantum Approximate Optimization”. In:
Physical Review Letters 124.9 (2020). DOI: 10.1103/PhysRevLett.124.090504. URL:
https://link.aps.org/doi/10.1103/PhysRevLett.124.090504.
[250] J. Cirac. “Quantum Computing and Simulation: Where We Stand and What Awaits Us”.
In: Nanophotonics 10.1 (2021). DOI: doi:10.1515/nanoph- 2020- 0351. URL: https:
//doi.org/10.1515/nanoph-2020-0351.
[251] H. Chen et al. “Universal Discriminative Quantum Neural Networks”. In: Quantum Ma-
chine Intelligence 3.1 (2020). DOI: 10.1007/s42484-020-00025-7.
[252] S. Aaronson. “The Learnability of Quantum States”. In: Proc. of the Royal Society A:
Mathematical, Physical and Engineering Sciences 463.2088 (2007). DOI: 10.1098/rspa.
2007.0113. URL: https://royalsocietypublishing.org/doi/abs/10.1098/rspa.
2007.0113.
[253] A. Rocchetto et al. “Experimental Learning of Quantum States”. In: Science Advances 5.3
(2019). DOI: 10.1126/sciadv.aau1946. URL: https://www.science.org/doi/abs/
10.1126/sciadv.aau1946.
[254] Ö. Legeza and J. Sólyom. “Quantum data compression, quantum information genera-
tion, and the density-matrix renormalization-group method”. In: Physical Review B 70
(2004). DOI: 10.1103/PhysRevB.70.205118.

LITERATURE
[255] J. Romero, J. P. Olson, and A. Aspuru-Guzik. “Quantum Autoencoders for Efficient

Compression of Quantum Data”. In: Quantum Science and Technology 2.4 (2017). DOI:
10.1088/2058-9565/aa8072. URL: https://doi.org/10.1088/2058-9565/aa8072.
[256] Y. Ding et al. “Experimental Implementation of a Quantum Autoencoder via Quantum
Adders”. In: Advanced Quantum Technologies 2.7–8 (2019). DOI: https://doi.org/10.
1002/qute.201800065. URL: https://onlinelibrary.wiley.com/doi/abs/10.
1002/qute.201800065.
[257] M. Benedetti et al. “Adversarial Quantum Circuit Learning for Pure State Approxima-
tion”. In: New Journal of Physics 21.4 (2019). DOI: 10.1088/1367- 2630/ab14b5. URL:
https://doi.org/10.1088/1367-2630/ab14b5.
[258] L. Hu et al. “Quantum Generative Adversarial Learning in a Superconducting Quantum
Circuit”. In: Science Advances 5.1 (2019). DOI: 10.1126/sciadv.aav2761. URL: https:
//www.science.org/doi/abs/10.1126/sciadv.aav2761.
[259] B. Schumacher and M. A. Nielsen. “Quantum Data Processing and Error Correction”. In:
Physical Review A 54.4 (1996).
[260] S. Aaronson. “Read the Fine Print”. In: Nature Physics 11.4 (2015). DOI: 10 . 1038 /
nphys3272.
[261] A. Childs. “Equation Solving by Simulation”. In: Nature Physics 5.12 (2009). DOI: 10 .
1038/nphys1473.
[262] O. Egecioglu, H. Ferhatosmanoglu, and U. Ogras. “Dimensionality Reduction and Sim-
ilarity Computation by Inner-product Approximations”. In: IEEE Trans. on Knowledge
and Data Engineering 16.6 (2004). DOI: 10.1109/TKDE.2004.9.
[263] Tuning CUDA Applications for Ampere. Application Note. 2788 San Tomas Expressway,
Santa Clara, CA 95051: NVIDIA Corporation, 2021. URL: https://docs.nvidia.com/
cuda/pdf/Ampere_Tuning_Guide.pdf.
[264] M. Frigo et al. “Cache-Oblivious Algorithms”. In: Foundations of Computer Science. 1999.
DOI: 10.1109/SFFCS.1999.814600.
[265] N. Piatkowski, S. Lee, and K. Morik. “Integer undirected graphical models for resource-
constrained systems”. In: Neurocomputing 173 (2016). DOI: 10.1016/j.neucom.2015.
01.091.
[266] I. Kerenidis, J. Landman, and A. Prakash. “Quantum Algorithms for Deep Convolutional
Neural Networks”. In: Proc. Int. Conf. on Learning Representations (ICLR). 2020. URL:
https://arxiv.org/abs/1911.01117v1.
[267] I. Kerenidis and A. Prakash. “Quantum Gradient Descent for Linear Systems and Least
Squares”. In: Physical Review A 101.2 (2020). DOI: 10.1103/physreva.101.022316.
[268] V. Giovannetti, S. Lloyd, and L. Maccone. “Architectures for a Quantum Random Access
Memory”. In: Physical Review A 78.5 (2008). DOI: 10.1103/physreva.78.052310.

LITERATURE
[269] C. Chamberland et al. “Topological and Subsystem Codes on Low-Degree Graphs with
Flag Qubits”. In: Physical Review X 10.1 (2020). DOI: 10.1103/PhysRevX.10.011022.
URL: https://arxiv.org/abs/1907.09528.
[270] N. Dattani, S. Szalay, and N. Chancellor. “Pegasus: The Second Connectivity Graph for
Large-scale Quantum Annealing Hardware”. In: arXiv:1901.07636 [quant-ph] (2019).
[271] R. Dechter and J. Pearl. “The cycle-cutset method for improving search performance in
AI applications”. In: Conference on Artificial Intelligence Applications. IEEE, 1987.
[272] S. Hadfield et al. “From the Quantum Approximate Optimization Algorithm to a Quan-
tum Alternating Operator Ansatz”. In: Algorithms 12.2 (2019). ISSN: 1999-4893. DOI:
10.3390/a12020034.
[273] G. H. Low and I. L. Chuang. “Hamiltonian Simulation by Qubitization”. In: Quantum 3
(2019). ISSN: 2521-327X. DOI: 10.22331/q-2019-07-12-163.
[274] S. Bravyi et al. “On the complexity of quantum partition functions”. In: (2021). arXiv:
2110.15466 [quant-ph].
[275] W. H. Zurek. “Decoherence and the Transition from Quantum to Classical — Revisited”.
In: Quantum Decoherence: Poincaré Seminar 2005. Ed. by B. Duplantier, J.-M. Raimond,
and V. Rivasseau. Birkhäuser Basel, 2007. ISBN: 978-3-7643-7808-0. DOI: 10 . 1007 /
978-3-7643-7808-0_1.
[276] B. Vacchini. “Quantum Noise from Reduced Dynamics”. In: Fluctuation and Noise Let-
ters 15.3 (2016). ISSN: 1793-6780. DOI: 10.1142/s0219477516400034.
[277] A. M. Childs, E. Farhi, and J. Preskill. “Robustness of adiabatic quantum computation”.
In: Physical Review A 65.1 (2001). ISSN: 1094-1622. DOI: 10 . 1103 / physreva . 65 .
012322.
[278] P. Q. Le, F. Dong, and K. Hirota. “A flexible representation of quantum images for poly-
nomial preparation, image compression, and processing operations”. In: Quantum In-
formation Processing 10.1 (2011). DOI: 10.1007/s11128-010-0177-y.
[279] Y. Zhang et al. “NEQR: a novel enhanced quantum representation of digital images”. In:
Quantum Information Processing 12.8 (2013). DOI: 10.1007/s11128-013-0567-z.
[280] F. A. González and J. C. Caicedo. “Quantum Latent Semantic Analysis”. In: Advances in
Information Retrieval Theory. 2011. DOI: 10.1007/978-3-642-23318-0\_7.
[281] S. Lloyd, M. Mohseni, and P. Rebentrost. “Quantum principal component analysis”. In:
Nature Physics 10.9 (2014). DOI: 10.1038/nphys3029.
[282] E. Tang. “Quantum Principal Component Analysis Only Achieves an Exponential
Speedup Because of Its State Preparation Assumptions”. In: Phys. Rev. Lett. 127 (6 2021).
DOI: 10.1103/PhysRevLett.127.060503.

LITERATURE
[283] S. Shalev-Shwartz and S. Ben-David. Understanding Machine Learning - From Theory to

Algorithms. Cambridge University Press, 2014. ISBN: 978-1-10-705713-5. URL: http :
//www.cambridge.org/de/academic/subjects/computer- science/pattern-
recognition - and - machine - learning / understanding - machine - learning -
theory-algorithms.
[284] C. Ciliberto et al. “Statistical limits of supervised quantum learning”. In: Physical Review
A 102.4 (2020). ISSN: 2469-9934. DOI: 10.1103/physreva.102.042414.
[285] V. Giovannetti, S. Lloyd, and L. Maccone. “Quantum-Enhanced Measurements: Beating
the Standard Quantum Limit”. In: Science 306.5700 (2004). DOI: 10 . 1126 / science .
1104149.
[286] M. B. Hastings. “The Power of Adiabatic Quantum Computation with No Sign Prob-
lem”. In: (2020). arXiv: 2005.03791 [quant-ph].
[287] B. Altshuler, H. Krovi, and J. Roland. “Adiabatic quantum optimization fails for random
instances of NP-complete problems”. In: (2009). arXiv: 0908.2782 [quant-ph].
[288] B. Altshuler, H. Krovi, and J. Roland. “Anderson localization makes adiabatic quantum
optimization fail”. In: Proceedings of the National Academy of Sciences 107.28 (2010).
DOI: 10.1073/pnas.1002116107.
[289] L. Zhou et al. “Quantum Approximate Optimization Algorithm: Performance, Mech-
anism, and Implementation on Near-Term Devices”. In: Physical Review X 10.2 (2020).
ISSN: 2160-3308. DOI: 10.1103/physrevx.10.021067.
[290] J. Guan, W. Fang, and M. Ying. “Robustness Verification of Quantum Classifiers”. In: Lec-
ture Notes in Computer Science (2021). ISSN: 1611-3349. DOI: 10.1007/978- 3- 030-
81685-8_7.

Quantum Machine Learning

Uploaded by

Quantum Machine Learning

Uploaded by

Quantum Machine Learning

State of the Art and Future Directions

Federal Office for Information Security

2 Quantum Computing in a Nutshell 7

3 Machine Learning in a Nutshell 22

4 Quantum Machine Learning 29

Federal Office for Information Security 1

5 Characteristics of QML Methods 66

6 Summary and Outlook 78

2 Federal Office for Information Security

Federal Office for Information Security 3

4 Federal Office for Information Security

Federal Office for Information Security 5

6 Federal Office for Information Security

2 Quantum Computing in a Nutshell

Quantum computing exploits quantum mechanical phenomena for information pro-

Federal Office for Information Security 7

8 Federal Office for Information Security

Federal Office for Information Security 9

|ψ1 ⟩ = γ0 |00⟩ + γ1 |01⟩ + γ2 |10⟩ + γ3 |11⟩ (12)

and |ψ2 ⟩ is in the entangled state

10 Federal Office for Information Security

and, upon this measurement, the second system will be in state

2.1 Adiabatic Quantum Computing

Federal Office for Information Security 11

12 Federal Office for Information Security

Federal Office for Information Security 13

The max-sum diversification problem is of importance in location- or portfolio optimiza-

s∗ = argmin −s⊺ Q s + q ⊺ s (27)

14 Federal Office for Information Security

Federal Office for Information Security 15

2.2 Quantum Gate Computing

16 Federal Office for Information Security

Federal Office for Information Security 17

H|0⟩ = |+⟩ H|1⟩ = |−⟩

18 Federal Office for Information Security

2.3 Further Quantum Computing Paradigms

Federal Office for Information Security 19

2.3.1 Topological Quantum Computing

2.3.2 Variational Quantum Computing

Variational quantum algorithms (VQA) and variational quantum approximate optimization

2.4 Technical State of the Art and Limitations

20 Federal Office for Information Security

Federal Office for Information Security 21

3 Machine Learning in a Nutshell

Machine learning is the science of fitting parameterized mathematical models to data in

22 Federal Office for Information Security

3.1 The Machine Learning Pipeline

Federal Office for Information Security 23

24 Federal Office for Information Security

Federal Office for Information Security 25

3.2 Lazy vs. Eager Machine Learning

26 Federal Office for Information Security

3.3 Sources of Uncertainty in Machine Learning

Federal Office for Information Security 27

28 Federal Office for Information Security

4 Quantum Machine Learning

4.1 Quantum Inspired Machine Learning

Federal Office for Information Security 29

30 Federal Office for Information Security

4.1.1 Tang’s Quantum Inspired Algorithm for Recommendation Systems

Federal Office for Information Security 31

4.1.2 Tensor Networks

32 Federal Office for Information Security

Federal Office for Information Security 33

4.1.3 Digital Annealing

Adiabatic quantum computers such as produced by D-Wave realize an energy mini-

34 Federal Office for Information Security

Fujitsu’s system is based on conventional complementary metal-oxide-semiconductor

4.1.4 Quantum Inspired Data Clustering

Federal Office for Information Security 35

4.1.5 Quantum Inspired Gravitational Search

4.2 Machine Learning for Quantum Computing

36 Federal Office for Information Security

4.2.1 Quantum Circuit Design

Federal Office for Information Security 37