Journal of Physics A:

Mathematical and Theoretical

TOPICAL REVIEW • OPEN ACCESS You may also like

- Elementary proof of QAOA convergence
A practitioner’s guide to quantum algorithms for Lennart Binkowski, Gereon Koßmann,
Timo Ziegler et al.
optimisation problems - Domain wall encoding of discrete variables
for quantum annealing and QAOA
Nicholas Chancellor
To cite this article: Benjamin C B Symons et al 2023 J. Phys. A: Math. Theor. 56 453001
- Filtering variational quantum algorithms for
combinatorial optimization
David Amaro, Carlo Modica, Matthias
Rosenkranz et al.
View the article online for updates and enhancements.

This content was downloaded from IP address 201.92.52.244 on 06/12/2024 at 01:29

 Journal of Physics A: Mathematical and Theoretical

J. Phys. A: Math. Theor. 56 (2023) 453001 (30pp) https://doi.org/10.1088/1751-8121/ad00f0

Topical Review

A practitioner’s guide to quantum

algorithms for optimisation problems
Benjamin C B Symons, David Galvin, Emre Sahin,
Vassil Alexandrov and Stefano Mensa∗
The Hartree Centre STFC, Sci-Tech Daresbury, Warrington WA4 4AD, United
Kingdom

E-mail: [email protected]

Received 4 July 2023; revised 14 September 2023

Accepted for publication 6 October 2023
Published 18 October 2023

Abstract
Quantum computing is gaining popularity across a wide range of scientific dis-
ciplines due to its potential to solve long-standing computational problems that
are considered intractable with classical computers. One promising area where
quantum computing has potential is in the speed-up of NP-hard optimisation
problems that are common in industrial areas such as logistics and finance.
Newcomers to the field of quantum computing who are interested in using this
technology to solve optimisation problems do not have an easily accessible
source of information on the current capabilities of quantum computers and
algorithms. This paper aims to provide a comprehensive overview of the theory
of quantum optimisation techniques and their practical application, focusing
on their near-term potential for noisy intermediate scale quantum devices. The
paper starts by drawing parallels between classical and quantum optimisation
problems, highlighting their conceptual similarities and differences. Two main
paradigms for quantum hardware are then discussed: analogue and gate-based
quantum computers. While analog devices such as quantum annealers are
effective for some optimisation problems, they have limitations and cannot
be used for universal quantum computation. In contrast, gate-based quantum
computers offer the potential for universal quantum computation, but they face
challenges with hardware limitations and accurate gate implementation. The
paper provides a detailed mathematical discussion with references to key works
in the field, as well as a more practical discussion with relevant examples. The

∗
Author to whom any correspondence should be addressed.

Original Content from this work may be used under the terms of the Creative Commons Attribution
4.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the
title of the work, journal citation and DOI.

1751-8121/23/453001+30$33.00 © 2023 The Author(s). Published by IOP Publishing Ltd Printed in the UK 1
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

most popular techniques for quantum optimisation on gate-based quantum

computers, the quantum approximate optimisation algorithm and the quantum
alternating operator ansatz framework, are discussed in detail. However, it is
still unclear whether these techniques will yield quantum advantage, even with
advancements in hardware and noise reduction. The paper concludes with a
discussion of the challenges facing quantum optimisation techniques and the
need for further research and development to identify new, effective methods
for achieving quantum advantage.

Keywords: quantum optimisation, combinatorial optimisation,

quantum approximate optimisation algorithm,
quantum alternating operator ansatz, gate-based quantum computing

(Some figures may appear in colour only in the online journal)

1. Introduction

The principles of quantum mechanics have enabled the development of a new paradigm of
computing, known as quantum computing. This innovative approach leverages the principles
of superposition and entanglement to offer faster and more efficient solutions to specific
industrial challenges that classical computers struggle to address. For example, the ability
to find prime factors of large numbers very quickly on a quantum computer has the poten-
tial to break widely used encryption algorithms that rely on the difficulty of this problem [1].
Additionally, quantum computing has implications for drug discovery [2], materials science
[3], and machine learning algorithms [4], amongst other fields. While the aforementioned
applications of quantum computing are promising, a significant amount of research is being
carried out on the development of quantum-enhanced optimisation techniques for logistics
[5–8], automotive [9], finance [10–12] and the defence sector [13].
In the classical domain, optimisation problems [14] are defined as mathematical problems
that involve finding the best solution from a set of possible solutions. The goal of optimisa-
tion is to find the solution that maximizes or minimizes an objective function—a measure of
success—subject to a set of constraints limiting the allowable solutions. Typically, classical
optimisation algorithms rely on searching through the space of all possible solutions, evaluat-
ing each solution until the best one is found. This process can be time-consuming and ineffi-
cient, especially for problems with many variables or constraints. As such the search typically
proceeds according to some heuristic approach that can speed up the process considerably.
A well-known example is the traveling salesman problem (TSP) [15], in which a salesman
must visit a set of cities, each only once, and return to his starting point while minimizing
the total distance travelled. This is a combinatorial optimisation problem, which means that
the number of possible solutions grows combinatorially with the number of cities. As a result,
finding the optimal solution to the TSP for many cities becomes time-consuming and com-
putationally expensive for classical computers. In fact, the TSP is an example of an NP-hard
problem, which means that there is no known efficient classical algorithm that can solve it in
polynomial time. The best classical algorithms for the TSP rely on heuristics and approxima-
tion methods, which can still take a long time to reach a solution, especially for large instances
of the problem.
Quantum computers are, in principle, inherently well suited to the task of solving
optimisation problems thanks to the key phenomena of superposition and entanglement
[16]. Superposition refers to the ability of qubits—the basic building blocks of quantum

2
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

computers—to exist in multiple states simultaneously. It is commonly claimed that super-

position will enable a quantum computer to explore all possible solutions to a problem simul-
taneously, thereby leading to an advantage over classical computers. Such claims must be
considered carefully because, in reality the probabilistic nature of measurement outcomes
in quantum mechanics means that superposition alone is not enough to yield an advantage.
Superposition must be combined with clever algorithm design in order for quantum computers
to have high probabilities of outputting useful information. Entanglement, on the other hand,
refers to the strong correlations that exist between qubits in a quantum system. A key prop-
erty of this correlation is that it is non-local i.e. two qubits can be entangled even if they are
separated by arbitrarily large distances. Entanglement is a fundamentally non-classical phe-
nomenon and it is believed that this will be a key resource that enables quantum advantage.
Indeed, entanglement is a key ingredient in many quantum optimisation algorithms, such as
quantum annealing and variational quantum algorithms (discussed in this review), which are
specifically designed to exploit the benefits of quantum mechanics for solving optimisation
problems. However, the actual mechanisms by which we can best exploit entanglement in
algorithms are not currently known. This is primarily because entanglement is still not well
understood in these contexts. Finally, another key reason optimisation has been an early applic-
ation area of quantum computing is the existence of mappings of many problem instances onto
Ising problems which in turn can be mapped onto a quantum computer. In general, it is highly
non-trivial to map a problem to a form that can be executed on a quantum computer. As such
the existence of a simple mapping is itself a good reason to examine a class of problems.
One of the most exciting aspects of quantum computing is the concept of quantum
advantage [17]. Although exact definitions differ somewhat, it is generally accepted that
quantum advantage will have been achieved once a quantum computer can solve a practic-
ally useful problem (significantly) faster than any classical computer. The potential advant-
ages that quantum computers offer in solving complex optimisation problems have made the
topic highly popular among both the scientific community and the public. This is evident from
the significant media coverage that it has received. However, there are several misconceptions
surrounding this topic, particularly regarding the speed and universality of the optimisation
problems that can be addressed. While it is correct that quantum computers may have the
capacity to solve some problems faster than classical computers, true quantum advantage has
not yet been established in practice. Additionally, an optimisation problem that can genuinely
benefit from a quantum device has yet to be identified.
Approaching quantum optimisation to solve real-world industrial challenges can be a daunt-
ing task for practitioners exploring quantum computing as a potential solution. Distinguishing
between facts and myths among the plethora of available articles on this subject can be chal-
lenging. Furthermore, there is currently no clear understanding of the potential benefits of
algorithms or any prospective quantum advantage. In this work, we aim to help practitioners
from various disciplines involving optimisation problems to understand the state of the art
of using quantum computers and algorithms to solve optimisation problems. Specifically, we
provide a parallelism between classical and quantum optimisation problems, analysing analo-
gies and differences, and explaining the kind of problems that will, in principle, benefit from
quantum acceleration. We analyse the difference between digital and analog quantum com-
puters and why quantum algorithms will perform differently. All of this information lays the
foundation to understand what a quantum computer can and cannot do to solve optimisation
problems efficiently, with the ultimate benefit of setting clear community expectations. For
interested readers, we provide a detailed mathematical overview of the most common quantum
algorithms for optimisation, such as quantum alternating operator ansatz (QAOA), useful for
understanding why such algorithms may lead to quantum advantage. Each section provides

3
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

an overview of the some of the most relevant work in the field, redirecting the reader towards
compelling practical case studies to inspire new work. Finally, we conclude this work by dis-
cussing the pros and cons of applied quantum computing to optimisation problems, as well as
future directions.

2. Classical optimisation

The field of classical optimisation encompasses a large number of specific problems that often
have useful, practical applications across a wide variety of fields. Depending on whether the
variables are discrete or continuous, optimisation problems can be divided in two classes,
discrete optimisation and continuous optimisation. In this work we focus almost exclusively
on discrete optimisation problems. In this section we want to provide the reader with a gen-
eral overview of the most common optimisation problems encountered by the scientific com-
munity, explaining why these can be computationally intractable using classical computing
methodologies

2.1. Discrete optimisation

Discrete optimisation problems—also known as combinatorial optimisation problems [18,

19]—involve decision variables that take on discrete or categorical values, such as integers,
binary values, or categories. Combinatorial optimisation problems involve searching through
a finite or countable set of possible solutions, making this class of problems computationally
intractable, as the number of possible solutions can be very large. Many combinatorial optim-
isation problems are known to be NP-hard problems, meaning that as the size of the problem
increases, the time required to find an optimal solution grows exponentially as well (assuming
P̸=NP). This makes this class of optimisation problems incredibly challenging to be solved
exactly in a satisfactory amount of time. There has been a great deal of effort from the com-
munity directed towards solving these problems as accurately and as quickly as possible. This
has resulted in a plethora of heuristic algorithms. Rather than exhaustively searching the space
of all possible solutions for the exact solution, heuristic methods are designed to find a good-
quality approximate solution. These methods are typically able to find solutions efficiently (in
polynomial time) but, this comes at the cost of providing a sub-optimal solution [20, 21].
In practice, the choice of algorithm depends on the size and structure of the problem
instance, as well as the desired trade-off between solution quality and computation time. For
small problem instances, exact methods may be preferred, as they can guarantee optimality.
However, for large problem instances, exact methods may become computationally infeasible,
and approximate methods may be necessary to find a good-quality solution within a reasonable
amount of time. Many combinatorial optimisation problems are in fact known to be APX-hard.
The APX complexity class is comprised of NP problems that are able to approximate in poly-
nomial time to within a constant multiplicative factor of the optimal solution. This has allowed
for the development of many very useful approximate classical algorithms for a wide variety
of combinatorial optimisation problems. It is generally believed that quantum computers will
not be able to exactly solve NP-hard optimisation problems efficiently. This is because the
quantum versions of these problems are QMA-hard [22] (the QMA complexity class can be
thought as the quantum analogue of NP). Quantum algorithms are therefore typically approx-
imate in nature and the real question is whether or not quantum computers are able to yield
better and/or faster approximate solutions relative to classical algorithms.

4
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

Figure 1. An illustration of a simple 5 node Max-Cut problem. The coloring of the nodes
(black and white) demonstrates one example of a partitioning into two separate sets that
yields a maximum cut.

In this subsection, we will discuss some of the most important classes of discrete combinat-
orial optimisation problems, including the travelling salesman problem, Max-Cut, Max-Flows
and the Knapsack problem.

2.1.1. The travelling salesman problem. The travelling salesman problem (TSP) is perhaps
the most well-known combinatorial optimisation problem. In order to setup the problem, con-
sider a salesman that must visit a number of cities, each with various distances between them.
The canonical form of the problem is to find the shortest path that visits all cities. The problem
lends itself well to being mapped to a graph problem. Given an undirected, weighted graph
G = (V, E), each city is a vertex V, and cities are connected by edges E with weights repres-
enting distances. Despite the simple setup of the problem it is very challenging to solve, in fact
it is NP-hard to solve exactly. There are a large number of applications of TSP (or variants of
the problem that have extra constraints) including logistics, circuit board manufacturing and
planning to name but a few.

2.2. The Max-Cut problem

The maximum cut problem is a well-known optimisation problem that requires partitioning the
vertices of an undirected graph into two disjoint sets, such that the number of edges between
the two sets is maximized. Given a graph G = (V, E), a cut represents a partition of the graph
into two distinct subgraphs such that no vertex is shared between either graph. The Max-Cut
problem consists of finding a cut in the graph such that the number of edges between the graph
partitions is maximised. More generally in the case of weighted graphs, the problem extends
to finding a cut that maximises the cumulative value of the edges. An example of the Max-Cut
problem is shown in figure 1.
The Max-Cut problem has applications in a variety of fields, including computer vision,
social network analysis, and computational biology. It is known to be NP-hard, which means
that finding the optimal solution is generally believed to be computationally intractable for
large instances of the problem. Therefore, many heuristic (approximate) algorithms have been
proposed to solve the problem efficiently. The Max-Cut problem can be mapped to an Ising
problem which makes it a good candidate for a quantum computer.

2.2.1. The Max-Flow problem. A natural problem in transport and logistics is to attempt to
model and optimise the flow of goods and services to customers. In the most basic example,
one could consider the transport of goods from a depot to a destination along a network where
certain routes have only limited capacities for transport. A problem of this form can be mapped

5
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

to a graph problem. The goal is to maximise the flow through the graph given the constraints.
A better value for the max flow corresponds to a more efficient delivery process.
Given a graph G(V, E) with edge capacities cij , a flow f is defined as a function f : E → R+
such that f(i, j) ⩽ cij . Given two vertices s and t labelled the source and sink respectively, we
constrain the flow such that it is equal when entering and exiting at all vertices except the
source and sink vertices,
∑ ∑
f (u, v) = f (v, u) ∀v ∈ V\ {s, t} . (1)
u:(u,v)∈E u:(v,u)∈E

A natural consequence of this constraint is that flow is conserved. For a single source-sink
pair with such a constraint, the flow may be given a value Cf describing the net output of the
source vertex s, that is,
∑ ∑
Cf := f (s, u) − f (u, s) . (2)
u:(s,u)∈E u:(u,s)∈E

The maximum flow problem concerns finding a flow f ∗ such that Cf ∗ is maximum, Cf ∗ ⩾
Cf ∀f : E → R+ . The maximum flow problem represents a canonical optimisation problem
over a weighted graph. Real world applications often require additional or different con-
straints on transport. For example, the vehicle routing problem [23] which requires that mul-
tiple vehicles service multiple destinations (or have different source depots). Variations to the
problem include the introduction of capacity constraints on the vehicles (CVRP) or time win-
dows for delivery (VRPTW).

2.2.2. The knapsack problem. The knapsack problem is an NP-complete combinatorial

optimisation problem that has applications in resource allocation. The knapsack problem is
usually stated as follows: given a set of items, each with a given weight and value, determine
the a collection of items that has total weight less than some maximum whilst also having max-
imum possible value. There are several variations on the problem, the simplest is the so-called
‘0-1 knapsack problem’ that restricts each item to a single copy i.e. each item has xi ∈ {0, 1}
copies. This problem is stated mathematically as,
∑
Maximise vi xi , (3)
i
∑
subject to wi xi ⩽ W, (4)
i

where vi and wi are values and weights respectively and W is the maximum total weight. The
problem can be extended such that multiple copies of each item are permitted or also to include
multiple weights e.g. monetary cost and mass of each item.

2.3. Continuous optimisation

Continuous optimisation problems involve variables that take on continuous values, such as
real numbers [24]. The objective function in a continuous optimisation problem maps a set of
decision variables to a scalar value that represents the objective of the problem. The decision
variables can take on any value within a specified range or domain. The goal is to find the values
of the decision variables that minimize or maximize the objective function. The constraints

6
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

can be expressed as equations or inequalities, and they limit the feasible region of the decision
variables. Examples of continuous optimisation problems include finding the shortest path
between two points in a curved space, determining the optimal allocation of resources for a
production process, or optimising the design of a complex system. Despite the fact that the
fields of continuous and discrete optimisation both aim to solve optimisation problems, they
are in fact rather different and separate fields of study. In the remainder of this work we focus
on discrete optimisation as this has been the sub-field of optimisation that has been studied the
most extensively in the context of quantum computing.

3. Optimisation on analog quantum computers

Analog computing is perhaps an unfamiliar concept in the age of digital computing. Most
are familiar with the digital (gate-based) model that utilises a series of discrete operations to
perform a computation. Analog computing on the other hand, involves the continuous time
evolution of an initial state to a final state. The goal is to setup and evolve the analog com-
puter in such a way that the final state is useful e.g. it is the solution to a problem. There is
a conceptual shift that has to be made when transitioning from thinking about digital to ana-
log computing. In the digital model of computation it is possible to understand computation
without considering the underlying hardware on which the computation is implemented. In the
analog model this is no longer possible because the nature of the continuous evolution of the
system is very much tied to the nature of the system itself i.e. the hardware. It is not possible
to formulate a general hardware-independent model of analog computation.
In the case of quantum computing, the boundary between analog and digital is not as sharp
as it is in the classical realm. Despite the fact that, in the gate-based model, computation is
thought of as a series of discrete gates acting on a quantum state, in reality there is still a
continuous time evolution of a quantum system. This means that, in some sense, all quantum
computers are analog. In both analog and digital quantum computers, the quantum state prior to
measurement is some continuous state that exists in the Hilbert space of the quantum computer.
However, after measurement both types of quantum computer output a bit string (a binary
number) i.e. both analog and digital quantum computers produce a digital output. Despite
these similarities, one of the main distinguishing features of digital quantum computers is
universality. We will discuss the concept of universality in more detail in the section on digital
quantum computers but, it suffices to say here that analog quantum computers are not, in
general, universal. This means that the analog quantum computers that exist today are only
able to tackle very specific problems. The kind of problem a given analog quantum computer
is able to solve is closely related to the form of the hardware itself.
At the time of writing, D-Wave has the most mature technology (superconducting qubits)
in the space of analog quantum computing, specifically they develop quantum annealers [25,
26]. Recently, there have been considerable advances in neutral atom technology, a different
approach that is able to blend the more traditional model of analog quantum computation
with the gate-based model. This allows companies such as Pasqal [27] to, at least in principle,
attain more generality whilst still retaining the advantages of the analog computing. While
we will not discuss neutral atoms in detail, we refer the interested reader to some articles on
the subject [28, 29] and would also like to highlight some impressive recent work that used
Rydberg atoms to solve maximum independent set (MIS) problems with up to 289 vertices
[30]. In this section we will focus on the well established adiabatic quantum computation
(AQC) model of computation that is used to describe computation on certain kinds of analog
quantum computers such as quantum annealers. We will first discuss AQC and then move

7
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

onto quantum annealing (QA) which is closely related but subtly different. There is already
a large literature including many reviews on adiabatic quantum computing [31] and quantum
annealing [32–35]. We will therefore restrict ourselves to a brief introduction and discussion
of the key concepts. The next section will then discuss optimisation for digital (gate-based)
quantum computers which is the main focus of this paper.
To begin an adiabatic quantum computation, a quantum system is initialized in the ground
state of a simple Hamiltonian. An external field is then applied to the system and gradu-
ally changed in order to slowly transform the initial state into the ground state of a different
Hamiltonian that encodes a problem of interest e.g. a combinatorial optimisation problem.
The final Hamiltonian should have a ground state that corresponds to the optimal solution of
the problem. The concept of AQC can be demonstrated mathematically by considering the
time-dependent Hamiltonian,

Ĥ (t) = A (t) Ĥ0 + B (t) Ĥ1 , (5)

where Ĥ0 and Ĥ1 in equation (5) are the initial and problem Hamiltonians respectively. Initially
A(t = 0) = 1 and B(t = 0) = 0, over time this changes until the reverse is true. Provided the
change is gradual enough, the adiabatic theorem guarantees that the quantum state remains
in the ground state throughout the computation. Formally, whether or not the time taken to
implement the change (i.e. the runtime of the computation, tf ) is in fact gradual enough depends
on the size of the energy gap between the ground and first excited states of Ĥ1 , denoted ∆. The
smaller the gap, the more gradual the change has to be to guarantee adiabaticity. Typically tf is
of order O(1/∆2 ), although the worst case is O(1/∆3 ). One of the major concerns is that this
gap can become exponentially small as the problem size increases [36].
Quantum annealing is very closely related to adiabatic quantum computation. Quantum
annealing is essentially an adiabatic quantum computation in which the requirement of adia-
baticity is relaxed somewhat. Rather than requiring the system to remain in the ground state
throughout the computation, the system is allowed to enter an excited state. This relaxation is
largely for practical reasons; it is easier to realise quantum annealing in hardware as opposed
to a perfectly adiabatic quantum computation. Furthermore, while AQC is universal [37], cur-
rent implementations of quantum annealing are not universal. In practice, quantum annealing
is actually stoquastic quantum annealing, another compromise made for the sake of practical-
ity. Stoquastic quantum annealing is limited to stoquastic Hamiltonians, which are defined as
Hamiltonians that, in a given basis, have off-diagonal elements that are real and non-positive.
There is concern that stoquastic Hamiltonians are easy to simulate [31] and therefore stoquastic
quantum annealers may not be able to provide a quantum advantage [38].
One of the main difficulties that arises when developing algorithms for quantum devices
(analog or digital) is how to map a problem of interest to a quantum computer. In other words,
how do we translate the problem into a form that is able to be encoded in qubits and then solved
by a continuous time evolution of a quantum state. Quantum annealers have largely focused on
combinatorial optimisation for two reasons. The first is that such a mapping exists for many
CO problems and is relatively simple. The full details are given in the section on quadratic
unconstrained binary optimisation (QUBO). However, it suffices to say here that many CO
problems can be mapped to a problem of the form of finding the ground state of an Ising
Hamiltonian [39]. It is evident that such problems are well-suited for quantum annealers. The
problem Hamiltonian of equation (5) can simply be made an Ising Hamiltonian. The second
reason is that CO problems have a huge variety of compelling real-world applications.
The current state of the field of quantum annealing is difficult to summarise. Despite the fact
that the field is relatively young, the literature is already large and is growing quickly. It is fair

8
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

to say that there is considerable debate over whether or not quantum annealers will ever achieve
quantum advantage. Proponents of the QA approach are optimistic that once machines reach
sufficient sizes, there is hope for an advantage in specific instances of optimisation problems
[40]. However, at the time of writing, there has been no demonstration of quantum advantage
using a quantum annealer. Sceptics claim that problems such as the exponential decrease in
the energy gap ∆ mean that advantage will never be realised on these devices. In fact, there
is a thriving cottage industry of attempting to demonstrate that quantum annealers will not be
able to provide an advantage [41]. There are also many approaches that take inspiration from
quantum computing to improve classical methods. For example, Fujitsu has developed a so-
called quantum-inspired digital annealer. This is a purpose-built silicon chip that is designed
specifically for combinatorial optimisation problems and is purportedly better at solving these
problems than a classical CPU but is also easier to scale up than an actual quantum annealer
[42]. In summary, the field faces considerable challenges and considerable scepticism. In spite
of this, there is still a large group of scientists that are optimistic and this optimism is not
entirely unfounded. Regardless of the eventual outcome, the field has generated a body of
scientific work that is highly interesting and will likely continue to do so for many years to
come.

4. Optimisation on digital quantum computers

Compared to analog quantum computers, digital quantum computers (DQC) are, at least in
some senses, more similar to conventional classical computers. Analogous to classical logic
gates, digital quantum computers implement a set of quantum logic gates (often DQC’s are
referred to as gate-based quantum computers). Each quantum gate corresponds to a unitary
operator acting on a small number of qubits (typically 1-3 qubits) and they can be combined
to build up a quantum circuit. A given quantum circuit is a representation of a computation
that can be implemented on a quantum computer. A simple example of a quantum circuit to
prepare a Bell state is shown in figure 2. The quantum circuit picture is a useful abstraction
largely because it is a simple and clear way of representing the evolution of a quantum state. It
is often easier to understand what is happening during a quantum computation when looking
at a quantum circuit compared to a long sequence of tensor products of operators. While there
are many similarities between the classical and quantum gate models, there are also some key
differences. For example, the fact that quantum gates correspond to unitary operators means
that quantum gates must be reversible. Classical logic gates have no such restrictions.
One of the most important concepts in the gate-based model of computing is that of univer-
sality. Formally, a set of logic gates (quantum or classical) is universal if it is able to reproduce
the action of all other logic gates. For classical logic gates, the NAND gate (or the NOR gate)
alone forms a universal gate set. A complication arises in the case of quantum gates because
there are uncountably many possible unitary operations that could be applied to a quantum
state (consider a rotation gate with an angle that may take any real value). A quantum gate
set is therefore considered universal if it is able to approximate any unitary operation. The
Solovay–Kitaev theorem [43] ensures that this approximation can be to done to within a given
error efficiently (the number of gates grows logarithmically as a function of the error). Broadly
speaking, the benefit of universality is generality. A universal digital quantum computer can,
at least in principle, be applied to solve any problem of interest. Analog quantum computers on
the other hand are considerably more limited. The form of the hardware determines the exact
type of problem that can be solved on the hardware. Analog quantum computers pay the price
of loss of generality but hope to gain a potential advantage in certain very specific situations.

9
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

Figure 2. A simple example of a quantum circuit that prepares a Bell state, drawn using
Qiskit. There are two qubits denoted q0 and q1 , H is the Hadamard gate and the second
gate is a CNOT (2-qubit) gate.

Note, an interesting consequence of universality is that a universal DQC is able to perform any
computation that can be performed in a more restricted framework i.e. it is possible to simu-
late a quantum annealer using a universal DQC. A further advantage of the digital approach
to quantum computation is the potential for error correction. Encoding information digitally
permits the possibility of error correction which is believed to be crucial for achieving fault-
tolerant quantum computing. It remains to be seen which of the analog digital approaches (if
either) will reach quantum advantage first.
The universality of digital quantum computers has enabled the development of quantum
algorithms across a large number of application areas including simulation of many-body
quantum systems, machine learning and optimisation problems. In this work, we will focus on
quantum algorithms for optimisation problems, specifically combinatorial optimisation prob-
lems. Combinatorial optimisation problems are often NP-hard to solve exactly meaning the
most common approaches are heuristic methods that solve these problems approximately.
There have been decades of work developing ever-improved approximate classical algorithms
to solve combinatorial optimisation problems. This work has largely been driven by the fact
that these problems are so widely applicable across a huge number of industries. By compar-
ison to classical algorithms, quantum algorithms for combinatorial optimisation are still very
much in their infancy. There are very few quantum algorithms for optimisation in existence
and, at the time of writing, it is still very much an open question whether or not any of them
will prove to be useful.
Despite the fact that no advantage has been established yet, there are at least theoretical
grounds to think that combinatorial optimisation problems may be a suitable candidate for
quantum advantage. The outcome of both classical and quantum computations is a bit string
that may correspond to a solution of a given problem. However, unlike a classical computer,
while a calculation is in progress on a quantum computer, the quantum state exists in a 2N
dimensional Hilbert space (where N is the number of qubits). Importantly, the quantum state
can be in a superposition of many states in this Hilbert space. The hope is that a well-designed
quantum algorithm can extract a low-lying state (a good quality solution) from such a super-
position more efficiently than classical algorithms which do not have the ability to exploit
superposition. The difficulty arises because, even though a given initial superposition of states
may contain the ground state (the optimal solution to the optimisation problem), if you measure
the superposition, you will obtain one of the states at random. Specifically, for a superposition
of n states, each state will be measured with probability 1/n. This means that superposition

10
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

Figure 3. A schematic of a variational quantum algorithm. Variational parameters θ are

optimised using a classical computer.

alone gives no natural advantage over a classical computer. Tackling this problem is the real
challenge of quantum algorithm design. Fundamentally a quantum algorithm seeks to boost
the probability of measuring some desired state that corresponds to a solution to a problem. In
general, this is a very challenging task. Examples of algorithms that achieve this successfully
include amplitude amplification [44] which is able to take an initial superposition of states and
greatly increase the probability of measuring some desired state.
One of the key limitations faced in the field of quantum optimisation (and more broadly
in quantum computing) is the limited availability of hardware and the noisy nature of the
hardware that is available. We are currently in the so-called noisy intermediate-scale quantum
(NISQ) era. The combination of small numbers of qubits and noise means that only small
problem sizes can be attempted on real hardware with algorithms that are low circuit depth.
These small problems are typically efficiently solvable by classical means and therefore do
not stand to truly benefit from quantum computers. While machines with increasing numbers
of qubits are being developed at a fairly consistent pace, the issue of reducing noise is more
challenging. Circuit depth is a key limitation of current hardware that makes anything but
the simplest algorithms (or simplest versions of a given algorithm)impossible to implement.
Without reduction in noise on the scale of several orders of magnitude, the only real hope is
error correction which itself drastically increases the overhead in terms of numbers of qubits.
In the NISQ era, the challenge is to develop algorithms that have some degree of resilience
to noise and to identify problems that, even for small sizes, may still benefit from quantum
hardware. With this in mind, many of the current approaches are hybrid quantum–classical. A
suitable, difficult part of a given problem is identified and ported to a quantum computer while
the part of the problem that can be efficiently solved classically remains on a conventional
computer. Variational quantum algorithms fall into this category, utilising classical optimisers
to find optimal parameters for quantum circuits.
There is a vast literature on variational quantum algorithms (VQA’s) [45–47], in this work
we will comment on just a few specific algorithms. VQA’s are popular as they are a class of
algorithms that are thought to be particularly well suited to NISQ devices. However, there
is a ‘hidden’ cost that is associated with VQA’s that is not always discussed or factored into

11
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

complexity theory discussions. VQA’s rely on optimisation of a set of variational parameters

which are typically a set of angles θi for single qubit rotation gates. The optimisation of these
parameters is performed by a classical optimisation algorithm (see figure 3 for a schematic
of a VQA). The θi cost landscape generally has many local minima and often has flat regions
(referred to as barren plateaus) that prevent further optimisation of parameters. In general,
the problem of finding the global minimum is an NP-hard problem. The cost landscape can
be made increasingly favourable by overparameterising the VQA i.e. adding more θi [48, 49].
However, the number of parameters (and therefore quantum gates) required to do this for some
of the most common ansatzes increases exponentially making this approach impractical [50].
In this work we will discuss the specific merits (or otherwise) of a given quantum algorithm.
However, when the algorithm is a variational one, it is important to bear this classical overhead
in mind.
When using VQAs in practice, as the size of the problem and therefore the number of
variational parameters increases, the classical optimisation problem can become prohibitively
difficult. This issue is particularly pertinent when discussing quantum algorithms for optim-
isation. Often the optimisation problems that we hope to solve using a quantum computer are
NP-hard. A cynic may notice that, in the VQA paradigm, you must first solve one NP-hard
problem classically (the optimisation of the variational parameters) in order to then solve the
NP-hard optimisation problem. In other words, despite using a quantum computer, you still
have to tackle an NP-hard problem classically. This viewpoint has a degree of merit but, there
are reasons to be optimistic nonetheless. Good classical optimisation algorithms exist that can
efficiently provide approximate solutions. The hope is that these solutions are good enough to
still yield good solutions using the VQA.
In the following sections, we will review the major quantum algorithms for quantum optim-
isation problems. We begin each section by giving an introduction to the conceptual and math-
ematical description of each algorithm. For the interested reader, we then detail a number of
case studies. The case studies are far from exhaustive but are selected because they either show
useful examples of the algorithm in action (simulated or on real hardware) or they propose
interesting modifications to the core algorithm.

4.1. Quantum approximate optimisation algorithm

The quantum approximate optimisation algorithm (QAO algorithm) was introduced by Farhi
et al [51] in 2014 and is a hybrid quantum–classical variational algorithm. Note that we refer to
the quantum approximate optimisation algorithm as the QAO algorithm (rather than QAOA)
in order to distinguish it more easily from the quantum alternating operator ansatz which
we discuss later. In the QAO algorithm, a problem-specific cost function C is mapped to a
Hamiltonian ĤC with an associated parameterised unitary ÛC (γ) = exp(−iγ ĤC ). In particular,
the possible solutions to the problem are each associated with an eigenstate of the Hamiltonian
such that it is diagonal in the computational basis,
n
ĤC |x⟩ = C (x) |x⟩, x ∈ {0, 1} . (6)

The unitary UC (γ) corresponds to time evolution under the cost Hamiltonian ĤC where
the time parameter has been replaced by a variational parameter γ. In addition, a mixer
Hamiltonian ĤM is introduced with associated parameterised unitary ÛM (µ) = exp(−iµĤM ).
A single block of the circuit, as shown in figure 4, corresponds to the application of the cost
and mixer unitaries on an initial state |ϕ0 ⟩,

|ϕ (µ, γ)⟩ = Û (µ, γ) |ϕ0 ⟩ := ÛM (µ) ÛC (γ) |ϕ0 ⟩. (7)

12
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

Figure 4. Schematic of the circuit for the QAO algorithm. The circuit show has depth
p = 1, for higher depths, the entire circuit is repeated p times.

Choosing a depth hyperparameter p ⩾ 1 so that the circuit consists of p blocks Û(µp , γp )

associated with 2p parameters, we can now implement the overall circuit as follows,

∏
p
|ϕ (µ, γ)⟩ = Û (µi , γi ) |ϕ0 ⟩. (8)
i =1

The cost is then evaluated as an expectation value,

⟨ϕ (µ, γ) |ĤC |ϕ (µ, γ)⟩, (9)

where a classical optimiser can be used to iteratively update the parameters µi , γi in order
to extremise the cost. The original formulation of the QAO algorithm utilised the so-called
transverse-field mixing Hamiltonian which is a sum of local operators. In particular ĤM =
∑
i Xi where Xi is the Pauli X-operator (bit flip) acting on the ith qubit. Conceptually, the QAO
algorithm can be thought of as time evolution of a quantum state combined with the power
of ‘hopping’ between states. The time evolution is generated by the cost Hamiltonian and
the ‘hopping’ is facilitated by the mixer unitary. Specifically, the mixer unitary is required
to ensure that the dynamics of the system are non-trivial. If the system were evolved using
only the cost Hamiltonian, the energy associated with the cost Hamiltonian (the cost) would
be conserved i.e. no choice of γ would alter the cost and therefore it would not be pos-
sible to perform an optimisation. By choosing a mixing Hamiltonian that does not commute
with the cost Hamiltonian, the cost is no longer conserved and a minimum can be found by
varying γ, µ.
As we have already discussed, the boundary between analog and digital quantum compu-
tation is not necessarily a sharp one. It is therefore not necessarily straightforward to place
algorithms such as QAO entirely in one category or the other. The vast majority of research
regarding the QAO algorithm is done in the context of DQCs but, there are examples where this
is not the case. For example, Ebadi et al utilised Rydberg atoms and parameterised pulse dur-
ations and phases in order to implement a scheme that strongly resembles the QAO algorithm
but in an analog context [30].
Throughout the paper we will refer to a quantity called the approximation ratio that is often
used to evaluate approximate methods such as the QAO algorithm. The approximation ratio r
is defined in terms of the maximum and minimum (optimal) cost, Cmax and Cmin respectively,
Cmax − C∗
r= . (10)
Cmax − Cmin
As the cost found by the QAO algorithm, C ∗ approaches the optimal value Cmin , the approx-
imation ratio tends towards 1.

13
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

4.1.1. Quadratic unconstrained binary optimisation. Quadratic unconstrained binary optim-

isation (QUBO) is an NP-hard combinatorial optimisation problem. It is primarily of interest
because there are many real world problems for which embeddings into QUBO have been
formulated [39]. Problems include Max-Cut, graph coloring and a number of machine learn-
ing problems to name but a few. Note that despite the utility of QUBO, it is limited by the fact
that it cannot be applied to constrained CO problems. The QUBO problem is formulated by
considering a binary vector of length n, |x⟩ ∈ Bn and an upper triangular matrix Q ∈ Rn×n . The
elements Qij are weights for each pair of indices i, j in |x⟩. Consider the function fQ : Bn → R,

∑
n ∑
n
fQ (x) = ⟨x|Q|x⟩ = Qij xi xj , (11)
i =0 j =i

where xi is the ith element of |x⟩. Solving a QUBO problem means finding a binary vector |x∗ ⟩
that minimises fQ ,

|x∗ ⟩ = argmin fQ (x) . (12)

The difficulty of solving such a problem is readily apparent as the number of binary vec-
tors grows exponentially (2n ) with the size n. Hard instances of QUBO are known to require
exponential time to solve classically. As such, classical methods to solve these methods are
typically heuristics that provide approximate solutions. The hope that is that heuristic quantum
algorithms such as the QAO algorithm may be able to find better approximate solutions to
QUBO problems than existing classical methods.
QUBO and Ising problems are closely linked and in fact a QUBO problem can be linearly
transformed into an simple Ising problem (one with no external magnetic field). Ising problems
are good candidates for quantum algorithms because there is a simple mapping that enables
them to be executed on a quantum computer. With the goal of linking QUBO to an Ising
problem and then mapping this to a quantum computer, we will rewrite the objective function
in equation (11) as a QUBO Hamiltonian,

∑
n ∑
n
ĤQUBO = Qij xi xj + Qii xi , (13)
i <j i =0

where the second term in equation (13) has been simplified using the fact that, for binary
variables, x2i = xi . The Ising Hamiltonian with no external magnetic field can be written in a
similar form,

∑
n ∑
n
ĤIsing = Jij si sj + Jii si . (14)
i <j i =0

In the Ising problem si ∈ {−1, 1} are spin variables. The similarities between equations (13)
and (14) are readily apparent. A simple linear transformation of the form s = 2x − 1 relates the
two problems. Without loss of generality we therefore proceed to map the Ising problem to a
quantum computer by promoting to si , sj to Zi , Zj ,

∑
n ∑
n
ĤI = Jij Zi Zj + Jii Zi . (15)
i <j i =0

14
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

Note that this promotion of spin variables to Pauli Z matrices is sensible because the
eigenvalues of Z are ±1. The Hamiltonian ĤI in equation (15) can now be used as the cost
Hamiltonian in the QAO algorithm thereby allowing it to be applied to a large number of
combinatorial optimisation problems.

4.1.2. Case studies. The assessment of whether or not the QAO algorithm is a compelling
candidate for quantum supremacy [52] or even advantage has been the subject of consider-
able research and debate. The quantum approximate optimisation algorithm is a metaheuristic,
which is a procedure for generating a heuristic (approximate) algorithm. This makes it very
difficult to derive any rigorous scaling behaviour outside of specific examples. It is therefore
imperative that the QAO algorithm is tested as extensively as possible using simulations and,
more importantly, actual hardware. In this section we will discuss a number of relevant case
studies of that utilise the QAO algorithm in order to give an overview of the current state of
the field.
Initially the QAO algorithm (depth p = 1) was the best known approximate algorithm for
MAX-3-LIN-2 [53] but a better classical algorithm was found soon after [54]. Several studies
have been rather pessimistic in regards to the outlook for the QAO algorithm applied to prob-
lems such as Max-Cut. For example, Guerreschi and Matsuura [55] explored time to solu-
tion for the QAO algorithm on the Max-Cut problem on 3-regular graphs. Here the QAO
algorithm was simulated under ‘realistic’ conditions including noise and circuit decompos-
ition and applied to problems of up to 20 vertices. Their time to solution also included the
cost associated with executing multiple shots for each circuit. It was demonstrated that, under
these conditions, for a given time to solution the QAO algorithm (p = 4,8) was only able to
solve a graph 20 times smaller than the best classical solver. By extrapolation under a few
assumptions, it was concluded that the QAO algorithm would require hundreds or even thou-
sands qubits to meet the threshold for quantum advantage on realistic near-term hardware.
There have been a number of other studies comparing the QAO algorithm in certain instances
to classical algorithms [56, 57]. For example, a paper by Marwaha [58] demonstrated that the
QAO algorithm (p = 2) was outperformed by local classical algorithms for Max-Cut problems
on high-girth regular graphs. Specifically for all D-regular graphs with D ⩾ 2 and girth above
5, they find 2-local classical Max-Cut algorithms outperform the QAO algorithm (p = 2).
Despite the pessimism of the papers that have been discussed so far, there is still hope that
the QAO algorithm might be able to outperform classical algorithms at higher depths than have
currently been explored. Whilst Guerreschi and Matsuura’s work [55] does not bode well for
the QAO algorithm on current generation devices that, at the time of writing have at most 127
qubits (IBM Eagle), their work does suggest that the QAO algorithm might offer an advantage
over classical methods once a sufficient number of qubits is reached. This claim requires an
extrapolation of their data and the authors are careful to make it clear that such an extrapolation
is highly uncertain due to the possibility of small-size effects in their results. An earlier paper
studied the QAO algorithm for Max-Cut problems and found that the approximation ratio
improved as the QAO algorithm depth, p increased [59]. They compared to classical meth-
ods and found that for values of p > 4, the QAO algorithm (simulated) was found to attain
a better approximation ratio than the classical algorithm for problem sizes of 6-17 vertices.
This appears to offer some hope that the QAO algorithm will be useful at larger values of p.
However, as Guerreschi and Matsuura point out in their paper, this study does not compare
the actual time to solution of the quantum and classical methods and therefore neglects the
potentially very costly process of optimising the variational parameters. Indeed Geurreschi

15
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

and Matsuura take a worst case approach and extrapolate their data using an exponential due
to the increasing cost of finding variational parameters as number problem size increases.
As we have already described, the problem of finding optimal values of the QAO algorithm
variational parameters becomes increasingly difficult as the depth p increases (this is a prob-
lem shared by all variational quantum algorithms). As such, there have been several papers that
have attempted to address this problem by exploring improved strategies for setting parameters
[60–65]. For example, Zhou et al [66] employed a heuristic strategy for the initialisation of the
2p variational parameters applied to the Max-Cut problem to reduce the number of classical
optimisation runs to O(poly(p)). This is in comparison to random initialisation which is iden-
tified as requiring 2O(p) optimisation runs (see the worst case exponential scaling adopted by
Geurreschi and Matsuura described above). Specifically, they develop an optimisation heur-
istic for a p + 1 depth circuit based on known optimisations for depth p. Using simulations
it was shown that the optimisation curves for the values of the p-level variational parameters
changed only slightly for p + 1. Hence, it is possible to use results from optimisation over a
p-depth circuit as an appropriate initialisation for the variational parameters in a p + 1 depth
circuit.
Streif and Leib [67] developed a method of determining optimal parameters for the QAO
algorithm that aimed to reduce or even remove the need for a classical optimisation loop. They
considered applications of the QAO algorithm with constant p to the problem of Max-Cut on 3-
regular graphs and 2D spin glasses. They observed that the optimal parameter values tended to
cluster around particular points, with decreasing variance for larger graph sizes. The optimal
parameters were found to be dependent on the topological features of the problem’s graph
rather than the specific problem instance or system size. They use tensor network theory to
develop ‘tree-QAOA’ in which the circuit is reduced to a tensor network, associated with a
tree subgraph for which contraction can be performed classically. This removes the need for
calls to the quantum processing unit (QPU) when determining the variational parameters. Their
numerical results for this method on Max-Cut for 3-regular graphs show results comparable
to using the Adam classical stochastic gradient descent optimiser.
There have been a number of papers that have proposed modifications to the core QAO
algorithm in order to improve performance and/or in an attempt to make the method more
amenable to execution on near term devices [68–70]. Guerreschi took a divide and conquer
approach to QUBO problems using the QAO algorithm and was able to achieve an average
reduction of 42% in estimates of required qubits for 3-regular graphs of up to 600 vertices
[71]. Zhou et al developed an approach to solving large Max-Cut problems with fewer qubits
than the standard approach [72]. Their strategy is to partition the graph into smaller instances
that are optimised with the QAO algorithm in parallel, followed by a merging procedure into
a global solution also using the QAO algorithm. Their procedure, dubbed ‘QAOA-in-QAOA’
(QAOA2 ), potentially offers a solution to tackling large graph optimisation problems with lim-
ited qubit devices. Perhaps surprisingly, they demonstrate that their method is able to perform
on par with or even outperform the classical Goemans-Williamson algorithm [73] on graphs
with 2000 vertices. Their work appears to suggest that the real power of the QAO algorithm
may come when tackling large, dense graphs. However, their results are from simulations only
and therefore more work is needed to determine the utility of the method on actual hardware.
While many studies of the QAO algorithm omit noise, there is a growing body of work
examining the effects of noise on the QAO algorithm [74–76]. For example, França et al
developed a framework for studying quantum optimisation algorithms in the context of noise
[77]. The presence of noise in a quantum computation drives the quantum state towards a max-
imally mixed state which corresponds to a Gibbs state with β = 0 (infinite temperature). They
utilise the fact that there are polynomial time classical algorithms for sampling Gibbs states

16
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

for certain values of β ⩽ βc . As a quantum computation progresses in the presence of noise

it will eventually cross over from the space of Gibbs states that cannot be efficiently sampled
with a classical algorithm to the space of states that can be efficiently sampled. The circuit
depth at which this happens will depend on a number of factors such as the degree of noise
and number of 2 qubit gates etc. Their framework can therefore be used to estimate circuit
depths for which there can be no quantum advantage on current hardware. They present such
a bound for the QAO algorithm. A recent paper built on this work by examining the QAO
algorithm in the context of various SWAP strategies [78]. Note that SWAP gates are required
due to the limited connectivity of qubits in most modern QPU’s. Two qubit gates can only be
applied to connected qubits, if a circuit requires 2 qubit gates to be applied to non-connected
qubits, then a number of SWAP gates must be employed to facilitate this. This leads to the
gate count of the actual circuit that is executed on hardware being larger than the gate count
of the high level circuit. This work therefore provides more realistic estimates of how noise
will impact the QAO algorithm on real quantum hardware. They demonstrate that, at least for
dense problems, gate fidelities would need to be significantly below fault tolerance thresholds
to achieve any advantage.
Finally, it is important to discuss the prospects of the QAO algorithm on current or near term
hardware [78–80]. One important study that tested the QAO algorithm on real hardware was
performed by Harrigan et al using Google’s Sycamore QPU [81]. They studied both planar
and non-planar graph problems using up to 23 qubits. The planar problems had the advantage
of mapping nicely to the topology of the QPU whereas the non-planar problems required com-
pilation using SWAP gates. They found that both planar and non-planar problem instances per-
formed comparably during noiseless simulations. However, when tests were done on quantum
hardware, the extra circuit depth incurred by compilation for non-planar problems meant that
performance quickly degraded as problem size increased. By contrast, the planar graph prob-
lems were found to have performance that was roughly constant regardless of problem size.
This is a useful result given that most modern QPUs have limited connectivity suggesting that
only relatively simple problems can be trivially mapped to them. This casts some doubt over
whether or not the QAO algorithm will be useful for more complex, interesting problems in
the near-term. Further tests investigated the relationship between the QAO algorithm depth
p and performance. Interestingly, performance was found to consistently increase as depth
increased during the noiseless simulations. However, during the tests on the real device, the
extra circuit depth associated with increasing p meant that noise quickly degraded perform-
ance. Despite this, the result is still a good example of empirical evidence to backup an often
made claim about the QAO algorithm which is that poor performance is due to limited depth
p of the algorithm.
A different study by Streif et al [82] also provides some evidence to support the claim that
increasing p can lead to improved performance. They studied the binary paint shop problem
(BPSP) using the QAO algorithm. BPSP is a particularly interesting test case because there
is no known classical algorithm that is able to provide an approximate solution to within a
constant factor of the optimal solution in polynomial time for all problem instances. They ran
numerical studies up to a problem size of 100 (corresponding to 100 qubits) and experiments
on an IONQ QPU of problem sizes up to 11 qubits. They study the probability of achieving an
approximate solution that is to within factor α of the optimal solution. For the hardest problem
instances and at fixed p, they found that the QAO algorithm requires the number of samples
to increase exponentially with problem size in order to maintain a constant α factor approx-
imation. However, allowing p to increase polynomially with system size enables a constant
factor approximation factor to be maintained efficiently. This is tentative evidence to suggest
the QAO algorithm can achieve better performance for BPSP than the best known classical

17
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

algorithms provided p is allowed to increase with problem size. However, the authors make it
clear that their tests were not performed on all possible problem instances meaning no generic
conclusions can be drawn.

4.2. Constrained optimisation and the quantum alternating operator ansatz

One of the limitations of the QAO algorithm is that, in its original form, it can only be applied to
unconstrained optimisation problems. However, many interesting combinatorial optimisation
problems have constraints that must be satisfied. Hence, there is considerable motivation to
move beyond unconstrained problems. One approach to account for constraints is to add pen-
alty terms to the cost Hamiltonian that introduce large cost penalties for solutions that violate
constraints [83, 84]. This is often called a soft constraint approach and it allows a constrained
problem to be treated as unconstrained but significantly lowers the probability of getting infeas-
ible (constraint violating) solutions. Solutions that violate constraints must then be pruned post
facto. For unconstrained problems, the total search space of the problem is equal to the space
of feasible (constraint satisfying) solutions. However, constraints reduce the size of the space
of feasible solutions relative to the total search space. Soft constraint approaches may therefore
become inefficient as they search parts of solution space that do not satisfy the constraints and
are therefore infeasible. This is particularly problematic in cases where the infeasible solution
space is much larger than the feasible solution space.
An alternative approach to introducing penalty terms is to construct the mixer unitary in
such a way that only feasible states are explored. The quantum alternating operator ansatz
(QAOA) was proposed by Hadfield et al [85, 86] as a generalisation of the QAO algorithm.
The goal of this generalisation was, at least in part, to address the issue of optimisation prob-
lems with hard constraints. Rather than the cost and mixer unitaries taking the form of time
evolution, they are able to take any form. The authors suggest certain principles that should be
adhered to when designing mixer unitaries:

• The feasibility of the state is preserved, that is, ÛM (µ) maps feasible states to feasible states.
• For any two feasible basis states |x⟩, |y⟩ ∃ µ : ⟨x|ÛM (µ)|y⟩ ̸= 0. That is, for any two feasible
solutions there is always some parameterisation that allows us to evolve one state such that
we measure the other with non-zero probability. In particular, it should be possible to evolve
toward a more optimal feasible solution with the correct parameterisation.

The original quantum alternating operator ansatz paper presented formulations of mixer unit-
aries for a large number of different problems. Since then there has been considerable further
exploration of this topic [87–89]. The hope is that, by designing problem specific mixer unit-
aries, it might be possible to reduce resource requirements and improve the quality of results.
For example, Fuchs et al [90] formulated a generalisation of the XY-mixer. The family of XY-
mixers [85, 91] preserve Hamming weight and are therefore useful in for problems with hard
constraints. Examples include the ring mixer which utilises a Hamiltonian HR ,
∑
N
ĤR = Xi Xi+1 + Yi Yi+1 . (16)
i =0

The ring mixer applies to cases where there are N qubits with periodic (ring) boundary
conditions. An alternative is the complete graph mixer Hamiltonian,
∑
ĤK = Xi Xj + Yi Yj (17)
(i,j)∈E(Kn )

18
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

where the sum is over all pairs of qubits on a complete graph Kn . The generalisation of Fuchs
et al enables the construction of mixers that preserve a feasible subspace specified by an arbit-
rary set of feasible basis states. The issue with their approach is that the exact way of construct-
ing mixers scales poorly as the number of basis states increases. This is especially problematic
as this number can often increase rapidly with problem size. The authors suggest the need for
heuristic algorithms to approximately construct mixers which may solve this issue but, this is
still an open problem.
Saleem et al [92] developed a modified version of the quantum alternating operator ansatz
algorithm called dynamic quantum variational ansatz (DQVA). This approach utilises a ‘warm
start’ in which a polynomial time classical algorithm is used to find local minima that are used
as inputs for the quantum algorithm. DQVA also employs dynamic switching on or off of parts
of the ansatz in order to reduce resource requirements for near term devices. Their results are
promising on simulators but are limited to relatively small graphs of up to 14 nodes (with edge
probability between 20% and 80%) due to lack of computational resources. Note that there
is a large literature on warm starting quantum optimisation algorithms that is not covered in
detail here [93–95].

4.2.1. Physics inspired approaches. One of the more interesting strategies for designing
mixers in the quantum alternating operator ansatz framework is the method of appealing to
physics for inspiration. In this section we will discuss several papers that take this approach
to design mixers for both physics and non-physics applications. For example, Kremenetski
et al [96] applied QAOA to quantum chemistry problems. The authors utilise the full second-
quantized electronic Hamiltonian as their cost function. Their mixer unitary is time evolution
generated by the Hartree–Fock (HF) Hamiltonian. This mixer naturally preserves feasibility
as states can only evolve into other chemically allowed states. The method is tested on P2 ,
CO2 , Cl2 and CH2 using classical simulations. The paper represents a nice proof of concept
of a physics inspired mixer for a physics application of QAOA.
Another recent paper applied the quantum alternating operator ansatz to the problem of
protein folding on a lattice. Building on previous quantum annealing work, [97] Fingerhuth
et al developed a novel encoding for a lattice protein folding problem on gate-based quantum
computers [98]. The method utilises ‘one-hot’ encoding to encode turns on a lattice in order
to simulate protein folding. They also develop several problem specific mixer Hamiltonians
that encode constraints that, for example, ensure the protein does not fold back on itself. The
authors find that their results improve significantly when they initialise the system in a super-
position over feasible states rather than a superposition over all states. The authors demon-
strated a proof of concept of their method on a Rigetti 19Q-Acorn QPU. However, hardware
constraints meant that they had to split the problem into two parts, each only 4 qubits and
had to initialise the state as a uniform superposition. They achieved a 6.05% probability of
reaching the ground state in their test on the real quantum device.
There has been some work exploring symmetries in the context of QAOA [99]. Recent work
by Zhang et al exploits the fact that symmetries and conserved quantities are intimately linked
[100]. In this case, the authors exploit gauge invariance in quantum electrodynamics (QED)
to formulate a mixer Hamiltonian that naturally conserves flow in the context of network flow
problems. Network flow problems satisfy the constraint that flow must originate and terminate
only at sources and sinks respectively. A consequence of this constraint is conservation of
flow i.e. the amount flowing into a given node must equal the amount flowing out of a given
node (unless it is a source or sink). The authors draw an analogy between these properties of

19
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

network flow problems and lattice QED. The analogy is justified by noticing that, when Gauss’
law ∇ · ⃗E = ρ is written in discrete form, it is of the same form as equation (1). The authors
therefore equate positive and negative charges to sources and sinks respectively. Electric charge
is equivalent to the amount of goods and the electric field ⃗E represents a flow.
In order to derive their mixer Hamiltonian, Zhang et al write down a minimal gauge invari-
ant lattice QED Hamiltonian. Making use of the analogy outlined above, the authors propose
this as a QED-mixer Hamiltonian for network flow problems. The gauge symmetry of the
Hamiltonian means that flow is naturally conserved without the need for extra penalty terms.
Despite satisfying the constraint of flow conservation, the QED-mixer Hamiltonian can gen-
erate solutions that contain isolated closed loops of flow that are unconnected to any source
or sink and are therefore not feasible solutions. The authors therefore propose a modified,
restricted QED-mixer that solves this problem at the cost of increased circuit complexity.
Zhang et al show that their mixer performs well compared to the standard X-mixer for
a number of relatively simple test cases. Specifically they study single source shortest path
(SSSP) problems on triangle graphs with 2-4 triangles and 2-pair edge-disjoint path (EDP)
problems on undirected graphs with 3 × 3, 3 × 4 and 4 × 4 grids. All test cases were per-
formed using numerical simulations rather than actual quantum devices. Despite the fact that
the test cases are relatively limited in scope, the paper clearly demonstrates that potential of
developing physics-inspired mixers. Considerable research effort relating to QAOA involves
the development of improved mixers and physics inspired approaches like the ones discussed
in this section offer a clear, rigorous path to improvements.

4.2.2. Case studies of constrained optimisation. Case studies of QAOA with hard con-
straints on actual quantum devices have been somewhat rare due to hardware limitations both
in terms of number and fidelity of qubits. However, as hardware continues to improve, QAOA
is increasingly able to be tested. For example, Niroula et al [101] recently performed an insight-
ful study comparing unconstrained (penalty term) QAOA, XY-QAOA and the layer variational
quantum eigensolver (L-VQE) algorithm [102]. They studied the problem of extractive sum-
marisation (ES) which is an example of a constrained optimisation problem that has real world
application. Tests were performed using noiseless and noisy simulators as well as Honeywell’s
H1-1 20 qubit trapped ion quantum device. In their tests, both L-VQE and QAOA account for
constraints by including a penalty term whereas in XY-QAOA the constraints are encoded dir-
ectly into the mixer. The XY-QAOA method also requires preparation of an initial state that is
in a superposition over all feasible states. 14 and 20 qubit experiments were performed with 10
repetitions for all methods apart from XY-QAOA which was only repeated three times due to
high circuit depth. Their results contain several key insights into how best to treat constraints
and the impact of noise.
Across all platforms (noiseless/noisy simulator and QPU) the penalty term QAOA had very
poor probability of generating feasible solutions (worse than random in the 14 qubit tests).
The authors show that this poor performance arises because introducing a penalty term leads
to a trade-off between terms in the Hamiltonian. L-VQE does not suffer from the same issue
in these test cases because it is expressive enough to be able to solve the problems exactly.
However, this will not be the case in general for L-VQE when applied to larger problem
instances that become more difficult to optimise as the number of parameters increases. These
tests also demonstrate that XY-QAOA is highly sensitive to noise. On the noiseless simulator,
XY-QAOA delivers feasible solutions with 100% probability but on the noisy simulator and
quantum device, the probability drops to less than 40%. The authors suggest that this is likely

20
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

due to the increased circuit depth and number of 2 qubit gates required for XY-QAOA relative
to both the other methods. In particular, the state preparation step for XY-QAOA is non-trivial
and already introduces noise.
These results appear to support the idea that, at least for QAOA, it is crucial to setup the
mixer unitary in such a way that it naturally encode constraints. However, the increased cost
in terms of circuit depth and number of 2 qubit gates associated with this approach means that
such implementations are more susceptible to noise and are therefore less likely to be useful on
current generation devices. The comparison between L-VQE and XY-QAOA in this instance
is inconclusive as, for the small problems studied, the two methods perform comparably.
Baker and Radha [103] were the first to test QAOA with hard constraints combined with
initial preparation of a superposition of feasible states (Dicke state) [104] on a QPU. Note that
the aforementioned protein folding study [98] came first and utilised hard constraints but the
initial state preparation was not performed on the QPU. Baker et al solved mean-variance port-
folio optimisation (MVPO) problems using QPU’s from multiple hardware providers: Rigetti
(Aspen 10), IonQ (11 qubit device) and IBM (multiple devices). Their tests provide a direct
comparison between two kinds of hardware: superconducting and trapped ion, each of which
have different qubit connectivities. The test cases varied both QAOA depth 1 ⩽ p ⩽ 5 and the
problem size from 2 to 5 assets. The methods tested were soft constraint QAOA (penalty term)
as well as two hard constraint mixers: ring and complete graph, with random and Dicke state
initialisation respectively.
The tests performed were comprehensive and worth reading in full detail. Here we will
summarise some of the key results. In a large number of cases QAOA outperforms random
although, as p and problem size increase, this is not necessarily always the case. QAOA with
soft constraints performs best across the QPU tests. However, the authors state that the total
number of gates required for the hard constraint methods was roughly an order of magnitude
larger than for the soft constraint circuits. They therefore attribute the poor performance of
hard constraints to noise. This is backed up by the fact that the hard constraint approach with
random state initialisation, which requires fewer gates, out performs the Dicke state approach.
This further confirms the findings of Niroula et al that, while hard constraint QAOA may be
preferable, the increased circuit depth makes the approach less suitable for current devices.
The authors also find a concerning variability in result quality across repetitions on all QPU’s
(regardless of hardware type) that far exceeds shot-based stochastic variability.
Finally, a paper by Wang et al deserves a brief discussion [91]. They performed numerical
simulations of penalty term and XY-mixer QAOA for graph colouring problems on small (5-7
vertices), hard to colour graphs. Consistent with the findings already discussed, they discover
that penalty term QAOA has a poor approximation ratio compared to the XY-mixer. In their
tests, they also find that the complete graph mixer outperforms the ring mixer. Interestingly
they demonstrate that increasing QAOA circuit depth consistently improves approximation
ratio. However, as we have already seen, the effects of noise cannot be ignored for QAOA.
These tests therefore need to be considered in the full context of the studies that utilised real
hardware.

4.3. Quantum semidefinite programming

Semidefinite programming (SDP) is a field of convex optimisation that is interesting because

a number of practical problems (e.g. combinatorial optimisation) are able to be modelled or
approximated using SDP. In this section we will outline the basic formulation of an SDP prob-
lem and then discuss recent work that has been done to develop quantum algorithms for semi-

21
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

definite programming. Given matrices (A1 , . . ., Am , C) ∈ Sn where Sn is the space of symmetric

n × n matrices and real numbers (b1 , . . ., bm ) a general SDP problem takes the form,

min ⟨C, X⟩Sn , (18)

X∈Sn

subject to the constraints,

⟨Ai , X⟩Sn ⩽ bi . (19)

Where X is a positive semidefinite matrix, denoted X ⪰ 0. Positive semidefinite matrices

are self-adjoint matrices that have no negative eigenvalues. Stated in words, the SDP prob-
lem requires finding a matrix X that minimises the cost which is given by the inner product
in equation (18), while satisfying the constraints (equation (19)). There are many classical
algorithms to solve SDP problems [105–107]. We have already mentioned one example when
discussing QAOA: the GW algorithm for Max-Cut [73]. It is important to note that SDP applies
only to convex optimisation problems. Convex problems have a single minimum which, in
most cases, makes them easier to solve than non-convex problems that typically have many
local minima. While many problems that are of real world interest are non-convex (e.g. optim-
ising variational parameters for VQA’s), SDP is useful nonetheless. Often SDP can be used
to approximately solve problems that are not necessarily convex by performing a so-called
‘convex relaxation’. This is a relaxation of the original problem in order to make it convex.
GW is an example of an algorithm that makes use of this procedure.
To date there have been only a relatively small number of papers that propose quantum
algorithms for semidefinite programming [108–112]. Semidefinite programming as a field is
relatively young and quantum semidefinite programming is even more so. The considerable
success of classical SDP algorithms has motivated the exploration of quantum SDP algorithms.
In the remainder of this section we will briefly discuss a small number of these works. Brandão
and Svore proposed a quantum SDP algorithm in 2017 [113] that utilises a combination of
Gibbs state preparation and amplitude amplification. They utilise the fact that it is possible to
solve an SDP problem by preparing Gibbs states of Hamiltonians given by linear combinations
of the input matrices Ai , and C. They then utilise amplitude amplification to reduce the cost
of preparing the Gibbs states. Note that their paper presents just the algorithm without any
simulations. Therefore, despite the authors giving bounds on properties such as run time, we
do not know whether or not the algorithm is practically useful.
Bharti et al developed a quantum SDP solver specifically targeting NISQ hardware [111].
They achieve this by developing a variational quantum algorithm. One of the most interesting
aspects of the paper is their approach to optimising the variational parameters. The cost land-
scape is typically highly non-convex and finding the global minimum is an NP-hard problem.
Classical optimisation of variational parameters is therefore often plagued by issues such as
local minima and barren plateaus. Bharti et al use a classical SDP solver to optimise their
variational parameters. This setup limits the dimension of the classical SDP problem to that
of the variational ansatz rather than the actual quantum SDP problem that is being solved.
However, SDP can only be used for convex optimisation problems and optimising VQAs is
known to be non-convex in general. As such, Bharti et al formulate∑ the problem in terms of
density matrices. They consider a hybrid density matrix, ρ = i βi |ψi ⟩⟨ψi | where β i are the
parameters optimised by the classical SDP. It is shown in the paper that this is a convex optim-
isation problem and as such a classical SDP solver can be used in order to solve the problem
efficiently.

22
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

5. Discussion & analysis

In this paper we have explored some of the most popular quantum optimisation techniques
currently being considered, particularly with a view to deployment on near term (noisy inter-
mediate scale) quantum devices. In both classical and quantum optimisation, the techniques
used are most commonly heuristic methods that provide approximate solutions. Applications
of quantum optimisation are still limited to what could be considered toy problems, and it
remains to be seen how these approaches may be translated into practical industrial applica-
tion of quantum advantage. Indeed, as much of the literature shows, there are still strides to
be made when it comes to realising quantum advantage in a narrower context. Nonetheless it
is possible to identify some of the potential advantages and shortfalls inherent in these tech-
niques, which we hope will guide practitioners in the field toward the development of new
effective methods for yielding quantum advantage.
In this work we have discussed two of the main quantum hardware paradigms that cur-
rently exist. Quantum annealers represent an analog approach, deploying continuous time
evolution to solve optimisation problems that can be appropriately translated into a well-
behaved quantum mechanical problem. However, the use cases remain constrained, and exist-
ing quantum annealers are not able to implement universal quantum computation. Nonetheless,
the application of quantum annealers to discrete optimisation problems has captured a great
deal of interest in part owing to the availability of D-Wave Systems’ commercial platform.
Typically quantum annealing is considered in the context of adiabatic quantum computation
(AQC), where the system is evolved slowly between the ground states of two Hamiltonians,
with the final state representing the solution of the optimisation problem. The adiabatic the-
orem guarantees that the system remains in the ground state as long as the schedule is suf-
ficiently slow. However, in many practical cases this restriction is unrealistic. Small gaps
between the ground state and first excited state energy can lead to long annealing times or
error-prone computation which may negate quantum advantage. However, proposals have been
made to allow annealing to take advantage of small jumps in energy, especially early in the
annealing process [114]. This ‘diabatic’ annealing pathway shows some promise due to its
more general regime and resistance to classical simulation.
The other hardware paradigm, and the main focus of this paper, is the ‘digital’ or gate-
based quantum computer. In contrast to annealers, digital quantum computers offer universal
quantum computation. While digital quantum computers are distinct from their analog coun-
terparts in many ways, there are also similarities. For example, the quantum state of a digital
quantum computer still undergoes a continuous evolution through time from an initial to a final
state making it analog in a sense. Despite this, certain guarantees on the approximation of gates
and the separability of their operations make the analogy to digital circuits a useful abstraction
for modelling quantum computation. To date there have been many quantum algorithms pro-
posed that utilise the quantum circuit model. Many promise to enhance classical methods or
to provide quantum advantage but in practice they are hampered by the limitations of current
hardware.
The quantum approximate optimisation algorithm and the quantum alternating operator
ansatz framework represent the vast majority of the field of quantum optimisation on digital
quantum computers. Since their inception, these methods have dominated the field of quantum
optimisation. We have discussed a wide variety of papers that implement variations of the
quantum approximate optimisation algorithm as well many papers operating in the quantum
alternating operator ansatz framework. Despite this, we have barely scratched the surface of
this literature. Across the papers that we have reviewed in this work, there are some common
themes. In many cases hardware limitations are blamed for poor performance of QAOA or for

23
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

Table 1. List of experiments performed on quantum hardware.

Reference Method Problem Hardware Qubits used Year

[98] QAO Protein folding Rigetti 19Q 4 2018

(superconducting)
[81] QAO Max-Cut 3-regular Google Sycamore 14 2021
(superconducting)
[78] QAO Weighted Max-Cut IBM Nairobi 7 2022
(superconducting)
IBM Mumbai 27
(superconducting)
[103] QAO, MVPO Rigetti Aspen-10 2-5 2022
QAOA (superconducting)
IonQ (trapped ion) 2-5
IBM (5 devices, 2-5
superconducting)
[30] QAO MIS QuEra Aquila 39-289 2022
(analog) (Rydberg atoms)
[101] QAO, Extractive H1-1 (trapped ion) 14, 20 2022
QAOA summarisation

the lack of testing of QAOA on hardware. The noise of qubits limits circuit depth and therefore
the QAOA depth p. Furthermore, the small number of qubits available limits the size of prob-
lems that can be studied to regimes that, for all optimisation problems of real world interest, are
efficiently solvable classically. These are valid concerns and there is evidence to suggest that
improving fidelity and number of qubits will improve the utility of QAOA. However, there
are several key studies we have mentioned that suggest improvements in fidelity and qubit
counts will have to be substantial in order for QAOA to yield any advantage [55, 78]. If these
studies are correct in their assessments, we will require, at best, several further generations of
quantum computers to be developed or at worst fault tolerant devices to be developed before
we can expect advantage. Furthermore, even if qubit counts increase and fidelities improve, the
corresponding increase in number of variational parameters will start to cause serious prob-
lems for the classical optimisation part of the algorithm. Efforts have been made to reduce the
overhead in the classical optimisation loop including heuristics to ‘pre-optimise’ the paramet-
ers for circuit based on values found in a shallower circuit [66], or by considering the topology
of the optimisation circuit as a tensor network [67]. Despite this, there is no guarantee that
we will be able to find good variational parameters efficiently as problem sizes and depths
grow. It is therefore far from certain that simply waiting for better hardware to be available
will naturally lead to advantage at some point in the future.
Another common theme in the papers we have discussed is the use of noiseless simulators
when testing implementations of QAOA. While many of these studies are interesting proofs
of concept, the lack of noise makes them highly unrealistic if the goal is to assess the prospect
of quantum advantage. We have discussed a small number of papers that studied noise in a
realistic manner or even ran on real hardware (see table 1 for a list of experiments on quantum
hardware mentioned in this paper). In general, the conclusions are that going beyond low
depth implementations of the basic quantum approximate optimisation algorithm is essentially
infeasible on current hardware due to the impact of noise [81, 101, 103]. This is problematic
in the context of constrained problems because the evidence we have reviewed appears to
suggest that hard constraints are in fact necessary to get good quality results. However, the

24
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

hard constraint approach increases circuit depth to the point that executing on noisy hardware
is no longer possible. For the foreseeable future quantum computers will be noisy and it is
therefore crucial for any algorithm that hopes to obtain near-term advantage to be resilient in
the presence of noise. The research that has been done to date highlights how crucial it is to
test on real hardware or, at minimum, with realistic noise models.
Although our assessment of QAOA so far has focused on the challenges and reasons to be
pessimistic, we believe there are also reasons to be optimistic. There is some evidence, both
empirical and numerical, that suggests the performance of the QAO algorithm improves as
p increases [81, 82]. We have also reviewed a considerable number of papers that operate in
the quantum alternating operator ansatz framework, designing problem-specific mixers. It is
possible that there exist certain highly specific optimisation problems for which an effective
and efficient problem-specific mixer can be designed. This may be an avenue to achieving
advantage sooner, although by nature any such advantage would be limited in scope. Another
practical avenue for near-term advantage is the utilisation of multiple QPU’s in parallel. We
have discussed some papers that employ strategies of breaking up problems into smaller parts
[71, 72]. If approaches like this are able to be parallelised across multiple QPU’s then it is pos-
sible that some of the hardware limitations of current generation devices can be circumvented
sooner than expected.
Although QAOA is by far the most popular method for quantum optimisation, there are a
number of different methods that exist but are less mature. We have briefly discuss quantum
semidefinite programming which is one such example. Quantum SDP can be applied to convex
optimisation problems which excludes a number of problems of interest that are non convex.
However, quantum SDP is still a promising and growing area of research. The hope is that the
success of classical SDP in recent decades can be replicated in the quantum realm. Beyond the
scope of what we have discussed here, other optimisation techniques have been explored in the
literature including variations of VQE [102, 115] and the use of branch & bound algorithms
[116]. Other examples include a recently proposed technique of logarithmic encoding that
considerably reduces the resource overhead in solving many NP-hard graph problems [117,
118] and a quantum enhanced greedy solver [119]. There is also a small but growing liter-
ature on Quantum Metropolis sampling, first proposed by Temme et al [120] The algorithm
is inspired by the classical metropolis algorithm and it can be applied to optimisation prob-
lems among other things. Since the original method was proposed, there have been a number
of developments [121, 122]. Chen et al provide a thorough overview of the developments to
date in their recent paper [123]. Finally, while we have focused the discussion here on dis-
crete optimisation, there has been some work on solving continuous optimisation problems.
For example, Matwiejew et al developed a continuous-time quantum walk mixer within the
QAOA framework to solve continuous multivariable optimisation problems [124].
Overall optimisation is a promising application area for quantum computing, not least
because the potential real world gain is huge if quantum advantage were to be achieved. There
are reasons to be optimistic that quantum computers will be able to efficiently solve optim-
isation problems, at least principle. However, in practice there has been no realisation of this
potential advantage and there are many reasons to be sceptical that such advantage is even close
to being achieved. In order to advance the field we require improved, noise resilient algorithms
as well as techniques that utilise hardware in more creative ways. In order to achieve this, there
is a desperate need for thorough benchmarking of existing methods using noisy simulations
and real hardware in order to better understand the impact of noise and performance on NISQ
hardware. Such benchmarks would be invaluable in guiding future research efforts in the field.

25
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

Data availability statement

No new data were created or analysed in this study.

Acknowledgments

This work was supported by the Hartree National Centre for Digital Innovation, a UK
Government-funded collaboration between STFC and IBM.

ORCID iDs

Benjamin C B Symons  https://orcid.org/0000-0001-5742-1082

David Galvin  https://orcid.org/0009-0006-6047-2730
Emre Sahin  https://orcid.org/0000-0002-5996-0407
Vassil Alexandrov  https://orcid.org/0000-0002-4889-1207
Stefano Mensa  https://orcid.org/0000-0002-0938-144X

References

[1] Shor P W 1994 Algorithms for quantum computation: discrete logarithms and factoring Proc.
35th Annual Symp. on Foundations of Computer Science pp 124–34
[2] Mensa S, Sahin E, Tacchino F, Barkoutsos P K and Tavernelli I 2023 Quantum machine learning
framework for virtual screening in drug discovery: a prospective quantum advantage Mach.
Learn.: Sci. Technol. 4 015023
[3] Bauer B, Bravyi S, Motta M and Chan G K-L 2020 Quantum algorithms for quantum chemistry
and quantum materials science Chem. Rev. 120 12685
[4] Biamonte J, Wittek P, Pancotti N, Rebentrost P, Wiebe N and Lloyd S 2017 Quantum machine
learning Nature 549 195
[5] Vogiatzis C and Pardalos P M 2013 Combinatorial optimization in transportation and logistics
networks Handbook of Combinatorial Optimization (Springer) pp 673–722
[6] Bentley C D B, Marsh S, Carvalho A R R, Kilby P and Biercuk M J Quantum computing for
transport optimization (https://doi.org/10.48550/arXiv.2206.07313)
[7] Clark J, West T, Zammit J, Guo X, Mason L and Russell D 2019 Towards real time multi-
robot routing using quantum computing technologies Proc. Int. Conf. on High Performance
Computing in Asia-Pacific Region pp 111–9
[8] Osaba E, Villar-Rodriguez E and Oregi I 2022 A systematic literature review of quantum com-
puting for routing problems IEEE Access 10 55805
[9] Yarkoni S, Neukart F, Tagle E M G, Magiera N, Mehta B, Hire K, Narkhede S and Hofmann M
2020 Quantum shuttle: traffic navigation with quantum computing Proc. 1st ACM SIGSOFT
Int. Workshop on Architectures and Paradigms for Engineering Quantum Software pp 22–30
[10] Slate N, Matwiejew E, Marsh S and Wang J B 2021 Quantum walk-based portfolio optimisation
Quantum 5 513
[11] Phillipson F and Bhatia H S 2021 Portfolio optimisation using the d-wave quantum annealer
Computational Science–ICCS 2021: 21st Int. Conf. (Krakow, Poland, Proc., Part VI) (Springer)
pp 45–59
[12] Herman D, Googin C, Liu X, Galda A, Safro I, Sun Y, Pistoia M and Alexeev Y 2022 A survey
of quantum computing for finance (arXiv:2201.02773)
[13] Krelina M 2021 Quantum technology for military applications EPJ Quantum Technol. 8 24
[14] Boyd S P and Vandenberghe L 2004 Convex Optimization (Cambridge University Press)
[15] Applegate D L, Bixby R E, Chvátal V and Cook W J 2011 The traveling salesman problem The
Traveling Salesman Problem (Princeton University Press)
[16] Nielsen M A and Chuang I 2002 Quantum Computation and Quantum Information (American
Association of Physics Teachers)

26
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

[17] Daley A J, Bloch I, Kokail C, Flannigan S, Pearson N, Troyer M and Zoller P 2022 Practical
quantum advantage in quantum simulation Nature 607 667
[18] Korte B H, Vygen J, Korte B and Vygen J 2011 Combinatorial Optimization vol 1 (Springer)
[19] Korte B and Vygen J 2018 Combinatorial Optimization (Algorithms and Combinatorics) 6th edn
(Springer)
[20] Papadimitriou C H and Steiglitz K 1998 Combinatorial Optimization: Algorithms and Complexity
(Courier Corporation)
[21] Blum C and Roli A 2003 Metaheuristics in combinatorial optimization: overview and conceptual
comparison ACM Comput. Surv. 35 268
[22] Bookatz A D 2014 QMA-complete problems Quantum Inf. Comput. 14 361
[23] Toth P and Vigo D 2002 The Vehicle Routing Problem (SIAM)
[24] Strang G and Freund L 1986 Introduction to applied mathematics J. Appl. Mech. 53 480
[25] King A D et al 2022 Coherent quantum annealing in a programmable 2,000 qubit Ising chain Nat.
Phys. 18 1324–8
[26] King A D et al 2023 Quantum critical dynamics in a 5,000-qubit programmable spin glass Nature
617 61
[27] Scholl P et al 2021 Quantum simulation of 2D antiferromagnets with hundreds of Rydberg atoms
Nature 595 233
[28] Saffman M 2016 Quantum computing with atomic qubits and Rydberg interactions: progress and
challenges J. Phys. B: At. Mol. Opt. Phys. 49 202001
[29] Henriet L, Beguin L, Signoles A, Lahaye T, Browaeys A, Reymond G-O and Jurczak C 2020
Quantum computing with neutral atoms Quantum 4 327
[30] Ebadi S et al 2022 Quantum optimization of maximum independent set using Rydberg atom arrays
Science 376 1209–15
[31] Albash T and Lidar D A 2018 Adiabatic quantum computation Rev. Mod. Phys. 90 015002
[32] Hauke P, Katzgraber H G, Lechner W, Nishimori H and Oliver W D 2020 Perspectives of quantum
annealing: methods and implementations Rep. Prog. Phys. 83 054401
[33] Yarkoni S, Raponi E, Bäck T and Schmitt S 2022 Quantum annealing for industry applications:
introduction and review Rep. Prog. Phys. 85 104001
[34] Das A and Chakrabarti B K 2008 Colloquium: quantum annealing and analog quantum computa-
tion Rev. Mod. Phys. 80 1061
[35] Santoro G E and Tosatti E 2006 Optimization using quantum mechanics: quantum annealing
through adiabatic evolution J. Phys. A: Math. Gen. 39 R393
[36] Jörg T, Krzakala F, Kurchan J and Maggs A C 2008 Simple glass models and their quantum
annealing Phys. Rev. Lett. 101 147204
[37] Aharonov D, van Dam W, Kempe J, Landau Z, Lloyd S and Regev O 2008 Adiabatic quantum
computation is equivalent to standard quantum computation SIAM Rev. 50 755
[38] Bravyi S and Terhal B 2010 Complexity of stoquastic frustration-free hamiltonians SIAM J.
Comput. 39 1462
[39] Lucas A 2014 Ising formulations of many NP problems Front. Phys. 2 5
[40] King A D et al 2021 Scaling advantage over path-integral Monte Carlo in quantum simulation of
geometrically frustrated magnets Nat. Commun. 12 1113
[41] Mandrá S, Katzgraber H G and Thomas C 2017 The pitfalls of planar spin-glass benchmarks:
raising the bar for quantum annealers (again) Quantum Sci. Technol. 2 038501
[42] Sao M, Watanabe H, Musha Y and Utsunomiya A 2019 Application of digital annealer for faster
combinatorial optimization Fujitsu Sci. Tech. J. 55 45–51
[43] Kitaev A Y 1997 Quantum computations: algorithms and error correction Russ. Math. Surv.
52 1191
[44] Grover L K 1996 A fast quantum mechanical algorithm for database search (arXiv:quant-ph/
9605043)
[45] McClean J R, Romero J, Babbush R and Aspuru-Guzik A 2016 The theory of variational hybrid
quantum-classical algorithms New J. Phys. 18 023023
[46] Cerezo M et al 2021 Variational quantum algorithms Nat. Rev. Phys. 3 625
[47] Tilly J et al 2022 The variational quantum eigensolver: a review of methods and best practices
Phys. Rep. 986 1
[48] Lee J, Magann A B, Rabitz H A and Arenz C 2021 Progress toward favorable landscapes in
quantum combinatorial optimization Phys. Rev. A 104 032401

27
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

[49] Kim J, Kim J and Rosa D 2021 Universal effectiveness of high-depth circuits in variational eigen-
problems Phys. Rev. Res. 3 023203
[50] Larocca M, Ju N, García-Martín D, Coles P J and Cerezo M 2021 Theory of overparametrization
in quantum neural networks (arXiv:2109.11676)
[51] Farhi E, Goldstone J and Gutmann S 2014 A quantum approximate optimization algorithm
(arXiv:1411.4028)
[52] Farhi E and Harrow A W 2019 Quantum supremacy through the quantum approximate optimiz-
ation algorithm (arXiv:1602.07674)
[53] Farhi E, Goldstone J and Gutmann S 2015 A quantum approximate optimization algorithm applied
to a bounded occurrence constraint problem (arXiv:1412.6062)
[54] Barak B, Moitra A, O’Donnell R, Raghavendra P, Regev O, Steurer D, Trevisan L,
Vijayaraghavan A, Witmer D and Wright J 2015 Beating the random assignment on constraint
satisfaction problems of bounded degree (arXiv:1505.03424)
[55] Guerreschi G G and Matsuura A Y 2019 QAOA for Max-Cut requires hundreds of qubits for
quantum speed-up Sci. Rep. 9 6903
[56] Wurtz J and Love P 2021 MaxCut quantum approximate optimization algorithm performance
guarantees for p > 1 Phys. Rev. A 103 042612
[57] Marwaha K and Hadfield S 2022 Bounds on approximating max κXOR with quantum and clas-
sical local algorithms Quantum 6 757
[58] Marwaha K 2021 Local classical MAX-CUT algorithm outperforms p = 2 QAOA on high-girth
regular graphs Quantum 5 437
[59] Crooks G E 2018 Performance of the quantum approximate optimization algorithm on the max-
imum cut problem (arXiv:1811.08419)
[60] Wang Z, Hadfield S, Jiang Z and Rieffel E G 2018 Quantum approximate optimization algorithm
for MaxCut: a fermionic view Phys. Rev. A 97 022304
[61] Brandão F G S L, Broughton M, Farhi E, Gutmann S and Neven H 2018 For fixed control paramet-
ers the quantum approximate optimization algorithm’s objective function value concentrates for
typical instances (arXiv:1812.04170)
[62] Sack S H and Serbyn M 2021 Quantum annealing initialization of the quantum approximate
optimization algorithm Quantum 5 491
[63] Galda A, Liu X, Lykov D, Alexeev Y and Safro I 2021 Transferability of optimal QAOA para-
meters between random graphs (arXiv:2106.07531)
[64] Akshay V, Rabinovich D, Campos E and Biamonte J 2021 Parameter concentrations in quantum
approximate optimization Phys. Rev. A 104 L010401
[65] Boulebnane S and Montanaro A 2021 Predicting parameters for the quantum approximate optim-
ization algorithm for MAX-CUT from the infinite-size limit (arXiv:2110.10685)
[66] Zhou L, Wang S-T, Choi S, Pichler H and Lukin M D 2020 Quantum approximate optimization
algorithm: performance, mechanism and implementation on near-term devices Phys. Rev. X
10 021067
[67] Streif M and Leib M 2020 Training the quantum approximate optimization algorithm without
access to a quantum processing unit Quantum Sci. Technol. 5 034008
[68] Zhu L, Tang H L, Barron G S, Calderon-Vargas F A, Mayhall N J, Barnes E and Economou S E
2022 Adaptive quantum approximate optimization algorithm for solving combinatorial prob-
lems on a quantum computer Phys. Rev. Res. 4 033029
[69] Majumdar R, Madan D, Bhoumik D, Vinayagamurthy D, Raghunathan S and Sur-Kolay S 2021
Optimizing ansatz design in QAOA for Max-cut (arXiv:2106.02812)
[70] Ayanzadeh R, Alavisamani N, Das P and Qureshi M 2022 Frozenqubits: boosting fidelity of
QAOA by skipping hotspot nodes (arXiv:2210.17037)
[71] Guerreschi G G 2021 Solving quadratic unconstrained binary optimization with divide-and-
conquer and quantum algorithms (arXiv:2101.07813)
[72] Zhou Z, Du Y, Tian X and Tao D 2023 QAOA-in-QAOA: solving large-scale MaxCut problems
on small quantum machines Phys. Rev. Appl. 19 024027
[73] Goemans M X and Williamson D P 1995 Improved approximation algorithms for maximum cut
and satisfiability problems using semidefinite programming J. ACM 42 1115
[74] Alam S G M, Ash-Saki A 2019 Analysis of quantum approximate optimization algorithm under
realistic noise in superconducting qubits (arXiv:1907.09631)
[75] Alam S G M and Ash-Saki A 2020 Design-space exploration of quantum approximate optimiza-
tion algorithm under noise IEEE Custom Integrated Circuits Conf. p 1

28
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

[76] Lotshaw P C, Nguyen T, Santana A, McCaskey A, Herrman R, Ostrowski J, Siopsis G and

Humble T S 2022 Scaling quantum approximate optimization on near-term hardware Sci. Rep.
12 12388
[77] Stilck França D and Garcia-Patron R 2021 Limitations of optimization algorithms on noisy
quantum devices Nat. Phys. 17 1221
[78] Weidenfeller J, Valor L C, Gacon J, Tornow C, Bello L, Woerner S and Egger D J 2022 Scaling
of the quantum approximate optimization algorithm on superconducting qubit based hardware
Quantum 6 870
[79] Willsch M, Willsch D, Jin F, De Raedt H and Michielsen K 2020 Benchmarking the quantum
approximate optimization algorithm Quantum Inf. Process. 19 1
[80] Lubinski T, Coffrin C, McGeoch C, Sathe P, Apanavicius J and Neira D E B 2023 Optimization
applications as quantum performance benchmarks (arXiv:2302.02278)
[81] Harrigan M P et al 2021 Quantum approximate optimization of non-planar graph problems on a
planar superconducting processor Nat. Phys. 17 332
[82] Streif M, Yarkoni S, Skolik A, Neukart F and Leib M 2021 Beating classical heuristics for the
binary paint shop problem with the quantum approximate optimization algorithm Phys. Rev. A
104 012403
[83] Farhi E, Gamarnik D and Gutmann S 2020 The quantum approximate optimization algorithm
needs to see the whole graph: worst case examples (arXiv:2005.08747)
[84] Farhi E, Gamarnik D and Gutmann S 2020 The quantum approximate optimization algorithm
needs to see the whole graph: a typical case (arXiv:2004.09002)
[85] Hadfield S, Wang Z, O’Gorman B, Rieffel E G, Venturelli D and Biswas R 2019 From the quantum
approximate optimization algorithm to a quantum alternating operator ansatz Algorithms 12 34
[86] Hadfield S, Hogg T and Rieffel E G 2022 Analytical framework for quantum alternating operator
ansätze Quantum Sci. Technol. 8 015017
[87] Ruan Y, Marsh S, Xue X, Li X, Liu Z and Wang J 2020 Quantum approximate algorithm for NP
optimization problems with constraints (arXiv:2002.00943)
[88] Fuchs F G, Kolden H, Aase N H and Sartor G 2021 Efficient encoding of the weighted MAX k
-CUT on a quantum computer using QAOA SN Comput. Sci. 2 89
[89] Marsh S and Wang J B 2020 Combinatorial optimization via highly efficient quantum walks Phys.
Rev. Res. 2 023302
[90] Fuchs F G, Lye K O, Nilsen H M, Stasik A J and Sartor G 2022 Constraint preserving mixers for
QAOA Algorithms 15 202
[91] Wang Z, Rubin N C, Dominy J M and Rieffel E G 2020 XY mixers: analytical and numerical
results for the quantum alternating operator ansatz Phys. Rev. A 101 012320
[92] Saleem Z H, Tomesh T, Tariq B and Suchara M 2023 Approaches to constrained quantum approx-
imate optimization SN Comput. Sci. 4 183
[93] Egger D J, mareček J and Woerner S 2021 Warm-starting quantum optimization Quantum 5 479
[94] Tate R, Farhadi M, Herold C, Mohler G and Gupta S 2023 Bridging classical and quantum with
SDP initialized warm-starts for QAOA ACM Trans. Quantum Comput. 4 1
[95] Okada K N, Nishi H, Kosugi T and Matsushita Y 2022 Systematic study on the dependence
of the warm-start quantum approximate optimization algorithm on approximate solutions
(arXiv:2209.02942)
[96] Kremenetski V, Hogg T, Hadfield S, Cotton S J and Tubman N M 2021 Quantum altern-
ating operator ansatz (QAOA) phase diagrams and applications for quantum chemistry
(arXiv:2108.13056)
[97] Babej T, Ing C and Fingerhuth M 2018 Coarse-grained lattice protein folding on a quantum
annealer (arXiv:1811.00713)
[98] Fingerhuth M, Babej T and Ing C 2018 A quantum alternating operator ansatz with hard and soft
constraints for lattice protein folding (arXiv:1810.13411)
[99] Shaydulin R, Hadfield S, Hogg T and Safra I 2021 Classical symmetries and the quantum approx-
imate optimization algorithm Quantum Inf. Process. 20 359
[100] Zhang Y, Zhang R and Potter A C 2021 QED driven QAOA for network-flow optimization
Quantum 5 510
[101] Niroula P, Shaydulin R, Yalovetsky R, Minssen P, Herman D, Hu S and Pistoia M 2022
Constrained quantum optimization for extractive summarization on a trappedion quantum com-
puter Sci. Rep. 12 17171

29
J. Phys. A: Math. Theor. 56 (2023) 453001 Topical Review

[102] Liu X, Angone A, Shaydulin R, Safro I, Alexeev Y and Cincio L 2022 Layer VQE: a variational
approach for combinatorial optimization on noisy quantum computers IEEE Trans. Quantum
Eng. 3 1
[103] Baker J S and Radha S K 2022 Wasserstein solution quality and the quantum approximate optim-
ization algorithm: a portfolio optimization case study (arXiv:2202.06782)
[104] Aktar S, Bärtschi A, Badawy A-H A and Eidenbenz S 2022 A divide-and-conquer approach to
Dicke state preparation IEEE Trans. Quantum Eng. 3 1
[105] Lee Y T, Sidford A, Wong S C and Chiu-Wai S 2015 A faster cutting plane method and its implica-
tions for combinatorial and convex optimization 2015 IEEE 56th Annual Symp. on Foundations
of Computer Science pp 1049–65
[106] Arora S, Hazan E and Kale S 2005 Fast algorithms for approximate semidefinite programming
using the multiplicative weights update method 46th Annual IEEE Symp. on Foundations of
Computer Science (FOCS’05) pp 339–48
[107] Arora S, Hazan E and Kale S 2012 The multiplicative weights update method: a meta-algorithm
and applications Theory Comput. 8 121
[108] Chakrabarti S, Childs A M, Li T and Wu X 2020 Quantum algorithms and lower bounds for
convex optimization Quantum 4 221
[109] van Apeldoorn J, Gilyén A, Gribling S and de Wolf R 2020 Quantum SDP-solvers: better upper
and lower bounds Quantum 4 230
[110] Kerenidis I and Prakash A 2020 A quantum interior point method for LPs and SDPs ACM Trans.
Quantum Comput. 1 1
[111] Bharti K, Haug T, Vedral V and Kwek L-C 2022 Noisy intermediate-scale quantum algorithm for
semidefinite programming Phys. Rev. A 105 052445
[112] Huang B, Jiang S, Song Z, Tao R and Zhang R 2023 A faster quantum algorithm for semidefinite
programming via robust IPM framework (arXiv:2207.11154)
[113] Brandão F G and Svore K M 2017 Quantum speed-ups for solving semidefinite programs 2017
IEEE 58th Annual Symp. on Foundations of Computer Science (FOCS) pp 415–26
[114] Crosson E and Lidar D 2021 Prospects for quantum enhancement with diabatic quantum annealing
Nat. Rev. Phys. 3 466
[115] Amaro D, Modica C, Rosenkranz M, Fiorentini M, Benedetti M and Lubasch M 2022
Filtering variational quantum algorithms for combinatorial optimization Quantum Sci. Technol.
7 015021
[116] Montanaro A 2020 Quantum speedup of branch-and-bound algorithms Phys. Rev. Res. 2 013056
[117] Rančić M J 2021 Noisy intermediate-scale quantum computing algorithm for solving an n-vertex
MaxCut problem with log(n) qubits Phys. Rev. Res. 5 12021
[118] Chatterjee Y, Bourreau E and Rančić M J 2023 Solving various np-hard problems using exponen-
tially fewer qubits on a quantum computer (arXiv:2301.06978)
[119] Dupont M et al 2023 Quantum enhanced greedy solver for optimization problems
(arXiv:2303.05509)
[120] Temme K, Osborne T J, Vollbrecht K G, Poulin D and Verstraete F 2011 Quantum metropolis
sampling Nature 471 87
[121] Yung M-H and Aspuru-Guzik A 2012 A quantum-quantum metropolis algorithm Proc. Natl Acad.
Sci. 109 754
[122] Shtanko O and Movassagh R 2021 Preparing thermal states on noiseless and noisy programmable
quantum processors (arXiv:2112.14688)
[123] Chen C-F, Kastoryano M J, Brandão F G S L and Gilyén A 2023 Quantum thermal state prepar-
ation (arXiv:2303.18224)
[124] Matwiejew E, Pye J and Wang J B 2023 Quantum optimisation for continuous multivariable func-
tions by a structured search Quantum Sci. Technol. 8 045013

30