Quantum Optimization: Potential, Challenges, and the Path Forward∗

Amira Abbas,1 Andris Ambainis,2 Brandon Augustino,3 Andreas Bärtschi,4 Harry Buhrman,1 Carleton Coffrin,4
Giorgio Cortiana,5 Vedran Dunjko,6 Daniel J. Egger,7 Bruce G. Elmegreen,8 Nicola Franco,9 Filippo Fratini,10
Bryce Fuller,11 Julien Gacon,7, 12 Constantin Gonciulea,13 Sander Gribling,14 Swati Gupta,3 Stuart Hadfield,15, 16
Raoul Heese,17 Gerhard Kircher,10 Thomas Kleinert,18 Thorsten Koch,19, 20 Georgios Korpas,21, 22 Steve
Lenk,23 Jakub Marecek,22 Vanio Markov,13 Guglielmo Mazzola,24 Stefano Mensa,25 Naeimeh Mohseni,5
Giacomo Nannicini,26 Corey O’Meara,5 Elena Peña Tapia,7 Sebastian Pokutta,19, 20 Manuel Proissl,7 Patrick
Rebentrost,27 Emre Sahin,25 Benjamin C. B. Symons,25 Sabine Tornow,28 Víctor Valls,29 Stefan Woerner,7
Mira L. Wolf-Bauwens,7 Jon Yard,30 Sheir Yarkoni,31 Dirk Zechiel,18 Sergiy Zhuk,29 and Christa Zoufal7
1
QuSoft and University of Amsterdam
2
University of Latvia
3
Massachusetts Institute of Technology
arXiv:2312.02279v2 [quant-ph] 23 Sep 2024

4
Los Alamos National Laboratory
5
E.ON Digital Technology GmbH
6
Leiden University
7
IBM Quantum, IBM Research Europe – Zurich
8
IBM Research, IBM T.J. Watson Research Center
9
Fraunhofer IKS
10
Erste Group Bank
11
IBM Quantum, IBM T.J. Watson Research Center
12
École Polytechnique Fédérale de Lausanne
13
Wells Fargo
14
Tilburg University
15
Quantum Artificial Intelligence Lab, NASA Ames Research Center
16
USRA Research Institute for Advanced Computer Science
17
Fraunhofer ITWM
18
Quantagonia
19
Zuse Institute Berlin
20
Technische Universität Berlin
21
HSBC Lab, Innovation and Ventures, HSBC, London
22
Czech Technical University in Prague
23
Fraunhofer IOSB-AST
24
University of Zurich
25
The Hartree Centre, STFC
26
University of Southern California
27
Centre for Quantum Technologies, National University of Singapore
28
University of the Bundeswehr Munich
29
IBM Quantum, IBM Research Europe – Dublin
30
Institute for Quantum Computing, Perimeter Institute for Theoretical Physics, University of Waterloo
31
Volkswagen AG
(Dated: September 24, 2024)
Recent advances in quantum computers are demonstrating the ability to solve problems at a
scale beyond brute force classical simulation. As such, a widespread interest in quantum algorithms
has developed in many areas, with optimization being one of the most pronounced domains.
Across computer science and physics, there are a number of different approaches for major classes
of optimization problems, such as combinatorial optimization, convex optimization, non-convex
optimization, and stochastic extensions. This work draws on multiple approaches to study quantum
optimization. Provably exact versus heuristic settings are first explained using computational
complexity theory — highlighting where quantum advantage is possible in each context. Then,
the core building blocks for quantum optimization algorithms are outlined to subsequently define
prominent problem classes and identify key open questions that, if answered, will advance the
field. The effects of scaling relevant problems on noisy quantum devices are also outlined in detail,
alongside meaningful benchmarking problems. We underscore the importance of benchmarking by
proposing clear metrics to conduct appropriate comparisons with classical optimization techniques.
Lastly, we highlight two domains – finance and sustainability – as rich sources of optimization
problems that could be used to benchmark, and eventually validate, the potential real-world impact
of quantum optimization.
Contents I. INTRODUCTION

I Introduction 2 Quantum computing could revolutionize numerous do-

mains in business and science, with optimization fre-
II Quantum Advantage & Complexity quently identified as a prime candidate to profit from
Theory 3 such a revolution. Optimization problems arise almost
A Exact Solutions . . . . . . . . . . . . . . . 3 everywhere, so any improvement over state-of-the-art
B Approximate Solutions . . . . . . . . . . . 6 classical algorithms with quantum computers could have
C Heuristics in Quantum Optimization . . . 8 a huge impact. Moreover, these improvements could oc-
D Takeaways from Complexity Theory in cur across multiple dimensions, such as solution quality,
Quantum Optimization . . . . . . . . . . 9 solution diversity, time-to-solution, and cost-to-solution.
But just how tangible are the benefits of quantum opti-
III Paradigms in the Design of Quantum mization? Techniques like Grover’s search [1], quantum
Optimization Algorithms 9 annealing, and adiabatic quantum optimization [2, 3]
A Grover Search . . . . . . . . . . . . . . . . 10 provided a promising start, but it has become clear that
B Quantum Adiabatic Algorithm . . . . . . 10 more tools are needed to realize a quantum advantage in
C Quantum Phase Estimation . . . . . . . . 10 optimization. In this work, we address the potential of
D Gibbs Sampling . . . . . . . . . . . . . . . 11 quantum optimization from various angles, namely: com-
E Approximation Algorithms . . . . . . . . 12 plexity theory, problem classes and algorithmic design,
F Variational Methods . . . . . . . . . . . . 12 execution on noisy hardware at scale, and fair bench-
marking, while outlining illustrative examples derived
IV Problem Classes & Algorithms 13 from real-world applications. Each topic is self-contained
A Discrete Optimization . . . . . . . . . . . 13 and written for a diverse audience, from optimization re-
B Continuous Optimization . . . . . . . . . 19 searchers to industry professionals. Thus, readers may
C Mixed-Integer Programming . . . . . . . . 22 move directly to any section of interest.
D Dynamic Programming . . . . . . . . . . . 24 For additional context, it is often stated that quan-
E Optimal Control . . . . . . . . . . . . . . 25 tum computers can evaluate all possible combinations of
F Robust Optimization . . . . . . . . . . . . 27 an optimization problem simultaneously. This neglects
G Multi-Objective Optimization . . . . . . . 28 that there remains an exponential list of possible solu-
tions and that a quantum computer requires additional
V Execution on Noisy Digital Hardware effort to find good ones. For some combinatorial opti-
at Scale 28 mization problems in a worst-case setting, the required
A Quantum Stack . . . . . . . . . . . . . . . 29 effort of known classical approaches scales exponentially
B Hardware Benchmarks . . . . . . . . . . . 32 in the problem size [4]. Here, quantum computers may
C Execution Example . . . . . . . . . . . . . 32 offer a quadratic speedup, e.g., via Grover’s Search, but
D Summary . . . . . . . . . . . . . . . . . . 33 a quadratic speedup over an exponential run time is still
an exponential run time. The situation can, however,
VI Benchmarks 33 significantly change if we consider a concrete problem
A Related Work . . . . . . . . . . . . . . . . 33 instance instead of the general worst-case scenario. Clas-
B Goals . . . . . . . . . . . . . . . . . . . . 34 sically, many algorithms and heuristics have been devel-
C Design . . . . . . . . . . . . . . . . . . . . 34 oped that – even for some large problems – can obtain an
D Problems . . . . . . . . . . . . . . . . . . 38 (almost) optimal solution in reasonable time. The most
E Demonstrations . . . . . . . . . . . . . . . 40 prominent example is the Traveling Salesperson Prob-
lem (TSP), where the goal is to find the shortest cycle
to visit a given set of places. It is often used to illus-
VII Illustrative Applications 42
trate the difficulty of combinatorial optimization, but in
A Financial Asset Allocation . . . . . . . . . 42
practice, very large problem instances can be solved close
B Sustainable Energy Transition . . . . . . . 47
to optimality classically [5–10]. In contrast, there exist
problems with less than 100 variables that are difficult to
VIII Conclusion & Outlook 51
tackle classically [11, 12]. While quantum optimization
algorithms will not necessarily improve the performance
References 52 for all problems, they provide additional tools with new
properties that can have significant impact and improve
performance for some problem instances and thus, im-
prove our overall capabilities in optimization. It is these
∗ Summary of the Quantum Optimization Working Group. concepts that we probe more formally in Secs. II, III
Authors are listed in alphabetical order. and IV, by laying the foundation for quantum optimiza-
Contact: Stefan Woerner ([email protected]) tion algorithms within a variety of well-studied problem

2
 II QUANTUM ADVANTAGE & COMPLEXITY THEORY

classes. tion of quantum advantage to strict complexity-theoretic

Another important aspect to consider in quantum com- separations in the context of optimization. The main
puting, is noise. The ultimate strategy to deal with noise takeaway for this section can be summarized as follows:
is quantum error correction, which introduces a signifi-
cant overhead in the number of qubits required [13]. Until Understanding complexity theory is extremely useful for
fault-tolerant error-corrected quantum computers can be gauging possible quantum advantage in optimization,
implemented at scale, other approaches to retrieve mean- but rigorous complexity-theoretic separations are not
ingful results from noisy devices have been considered. A necessary, nor sufficient when seeking a practical
particular strategy is quantum error mitigation [14–16], quantum advantage.
where one reduces noise by cleverly combining multiple A noteworthy example to drive home this point is the
noisy results. In Sec. V, we investigate the possibility of task of factoring a number into its prime constituents.
using error-mitigated noisy quantum computing to bridge This problem is interesting from a complexity point of
the gap between classical and fully fault-tolerant quan- view as it is not believed to be efficiently solvable by a
tum computing for optimization problems. classical computer — although there is no formal proof
In addition to noise, in practice, one often relies on that confirms this. A quantum computer, on the other
optimization heuristics, where the performance of an al- hand, can solve this factoring task efficiently, thanks to
gorithm on a given problem instance is not known up- Shor’s algorithm [17], and is expected to meaningfully
front. This makes comparing quantum and classical al- demonstrate this efficiency once fault-tolerant quantum
gorithms rather tricky. To establish systematic, repro- hardware is available [18]. Thus, our notion of quan-
ducible, and comparable benchmarks, Sec. VI introduces tum advantage in practice should cast a wider net than
fair metrics and outlines relevant optimization problems formal proofs, but within a reasonable range guided by
for which different algorithms can be applied to. This complexity-theoretic arguments. We explain this further
is crucial to identify what strategy works best and guide by delving into the basics of complexity theory, discussing
toward development of new quantum optimization algo- pertinent results and extending the discussion to more
rithms. Sec. VII then dives into two particular real-world practical scenarios for optimization. We refer the inter-
problems to illustrate the nuance of optimization in prac- ested reader to Refs. [19, 20] for a more comprehensive
tice. Combined with reasonable benchmarks, these prob- overview of quantum and classical complexity theory.
lems highlight where we stand and what open questions
must be answered before we can impact a certain domain.
Since quantum computing will not necessarily accelerate A. Exact Solutions
all problems, it is crucial to understand exactly where
advantages might be found. This needs to be probed 1. Decision and relational problems
theoretically, and empirically — which is what we show-
case in this work. Quantum optimization algorithms pro-
Much of complexity theory deals with so-called deci-
vide additional tools that complement existing ones, and
sion problems. As the name suggests, a decision problem
synergies with classical algorithms can certainly lead to
involves determining (i.e. deciding) whether there exists
better performance.
a particular solution to the problem or not. Without
loss of generality, it is often convenient to formalize this
idea using binary strings [21]. A computational decision
II. QUANTUM ADVANTAGE & COMPLEXITY problem on binary strings corresponds to what is called
THEORY a language in complexity theory. A language L ⊆ {0, 1}∗
is a set of binary strings of arbitrary length where the
In theoretical computer science, complexity classes corresponding decision problem is to determine whether
group computational problems together in terms of the a given input string x ∈ {0, 1}∗ is an element of L or not.
amount of resources required to solve them. Naturally, For example, the language could consist of the set of all
the harder the computational problem, the more re- prime numbers encoded in binary form and the problem
sources it will require, where the resources of most im- would involve determining whether x is prime (x ∈ L)
portance consist of time and memory. For this reason, or x is composite (x ̸∈ L). For completeness, an alpha-
complexity theory is often the guiding compass for at- bet Σ, can be thought of as the set of characters from
tempting to find classes of problems that are hard for which the language is formed. In this case, Σ = {0, 1}.
a classical computer to solve, but perhaps less resource- It is also fairly common to think of decision problems on
demanding on a quantum computer. This is a very rea- binary strings as binary functions fn : {0, 1}n → {0, 1}
sonable and rigorous way to formulate a quantum ad- where fn (x) = 1 if and only if x ∈ L, which is referred
vantage, but proving that problems require strictly more to as a “yes” instance. The famous Traveling Salesperson
resources classically than they would on a quantum com- Problem (TSP) can be formulated in such a way, where
puter is often a monumental task. In this section, we a list of cities, the distances between them, and a total
discuss notable strides made in complexity theory, but distance d is described by a bitstring x ∈ {0, 1}n . Then
also highlight issues that arise when restricting our no- fn (x) = 1 if and only if there exists a salesperson’s tour

3
 II QUANTUM ADVANTAGE & COMPLEXITY THEORY

through the cities whose length is ≤ d. This problem be- puts a solution y, whenever one exists. In this regard, FP
comes increasingly difficult as the number of cities grows, generalizes P, as it is the analog of P for functions with
since, even though the input size |x| = n grows polyno- an n-bit output, and importantly, the set FP contains
mially, the number of possible tours grows exponentially; PO. While FP focuses on general relational problems
and the best known classical dynamic programming so- that require finding a specific output, PO is specifically
lution still needs exponential time [22, 23]. In fact, the about optimization problems where the best solution un-
notion of resource scaling in terms of the input size is der certain criteria needs to be found.
how hardness of a problem is formulated in complexity
theory. Similar to P, PSPACE forms the class of decision
problems that can be solved with a polynomial amount of
At this juncture, it is worth noting that decision prob-
space, but there is no restriction on the time used by the
lems only care about the existence of a particular solu-
algorithm. Allowing for randomized computation brings
tion, and not the actual solution itself. Clearly, in prac-
us to the class BPP, which stands for bounded-error
tical optimization settings, finding an actual solution to
probabilistic polynomial time, where a classical proba-
a problem is often necessary. Continuing with TSP as
bilistic machine can solve all instances of a decision prob-
an example, merely knowing that there exists a path
lem in polynomial time, with an error probability ≤ 1/3.
to travel to all cities and back is not enough to bene-
Generalizing further to a computational model that uses
fit in practice, and determining the actual path to take is
quantum mechanics, leads to BQP, the class of decision
needed. This leads to the notion of relational problems,
problems that quantum computers can solve in polyno-
which generalize decision problems in this sense. More
mial time with an error probability ≤ 1/3 on every in-
concretely, given a relation R ⊆ {0, 1}∗ × {0, 1}∗ and an
stance. Crucially, BQP contains problems that are not
input x, the goal of a relational problem is to output
believed to be in BPP or P, like the decision version
any y such that (x, y) ∈ R. So one is concerned about
of factoring large integers for example, and it is widely
finding a y that satisfies the relation (sometimes called a
believed that BQP ̸= BPP [26]. Figure 1 illustrates
“witness” to the original decision problem), not just de-
the relationships between these classes. It is not known
termining the existence of such a y. For the discussions
whether all inclusions are strict. However, there is still
that follow, it is useful to keep in mind that the same
no proof that PSPACE ̸= P. In the highly unlikely sit-
problem can be formulated in different ways. The deci-
uation that PSPACE = P, it follows that BQP = P
sion problem already encompasses the time complexity
and every efficient quantum algorithm can be simulated
of the problem. Namely, having a procedure that deter-
efficiently on a classical computer!
mines whether a TSP tour of length ≤ d exists, can be
used to efficiently, modulo the complexity of the decision
procedure, find a tour of minimal length ≤ d by iterat-
ing over the decision problem after removing edges [24].
Are quantum algorithms strictly more
Looking at the decision version of a problem may be rele-
powerful than randomized algorithms?
vant from a complexity point of view, but perhaps not so
The clear expectation is that the answer is
much from a practical one. Rather, one may explicitly in-
affirmative, but in light of the difficulty of
terpret a relational problem as an optimization problem,
proving PSPACE ̸= P, no proof to answer this
making these closer to problems observed in practice.
question conclusively is known. A long line of
research plausibly suggests that P = BPP. The
reason for this is that good pseudo-random
2. Deterministic, randomized and quantum computation number generators are believed to exist and
these in turn can be used to de-randomize BPP
Decision problems that are efficiently solvable by a de- algorithms. There is also black-box evidence for
terministic machine are grouped together into the com- this. Bernstein and Vazirani [26] introduced the
plexity class P, for polynomial time. Efficient implies problem of Recursive Fourier Sampling (RFS)
the existence of an algorithm which solves the decision which is in BQP but not in BPP relative to an
problem with a run time that grows no faster than a poly- oracle. Interestingly, Yamakawa and
nomial in the size of the input. Crucially, the degree and Zhandry [27] show that, relative to a random
constants of the polynomial are important for practical oracle, there exist problems in BQP which are
purposes. For example, an algorithm with a run time not in BPP. Arora et al. [28] provide further
scaling like n3 is far more practical than one that runs in results relative to a random oracle for shallow
n100 time, even though both approaches scale polynomi- circuits and variational quantum algorithms, but
ally and are therefore deemed computationally efficient it is not clear what practical implications, if any,
in complexity theory. The set of optimization problems these separations relative to random oracles
whose decision versions are then in P is referred to as mean.
PO [25]. Switching to relational problems leads to the
class FP (for functional), which consists of all relations R
for which a deterministic polynomial-time algorithm out-

4
 II QUANTUM ADVANTAGE & COMPLEXITY THEORY

3. Hard versus complete problems

Can quantum computers solve NP-hard
problems? It is generally believed that
If useful problems for optimization, outside of P, could NP-hard problems are not in BQP. An
be solved efficiently by quantum computers, such a find- intuition for this belief is given by the lower
ing would lead to huge breakthroughs in computer sci- bound for search in the black-box [30, 31]
ence, akin to that of Shor’s algorithm. This highlights the setting. This proves that Grover’s search
importance of trying to design quantum algorithms for algorithm [1] is optimal in the black-box setting.
classically hard optimization problems. To identify such The intuition is that this also carries over to the
problems, the class NP for nondeterministic polynomial- Boolean satisfiability problem (SAT), which is
time, becomes relevant. A decision problem is in NP if, the famous standard NP-complete problem. It
given an input x, we can efficiently check that x is a is based on the feeling that a general instance of
“yes” instance of the problem. To do this, we are given a SAT does not contain structure that can help
polynomial-size proof (sometimes called a witness) which finding a satisfying assignment just as in the
certifies this fact. Intriguingly, there is also a relational black-box setting. This would imply that
problem extension of NP, namely FNP. As mentioned, Grover’s algorithm is also optimal for general
problems in NP are concerned with the existence of a instances of SAT. Grover’s algorithm can search
particular solution, such as a satisfying assignment. On over all 2n possible assignments to an n-variable
the other hand, FNP versions of these problems are con- formula and obtains at most a quadratic
cerned with finding the actual value of a solution, if one speedup over classical unstructured search
exists, in keeping with practical requirements for opti- procedures. Since we do not know of either a
mization settings. In particular, the set of optimization quantum or a classical algorithm that can solve
problems within NP is termed NPO, and NPO ⊂ FNP. all SAT instances asymptotically faster than
A problem is called NP-complete if all other problems brute-force-search, it does not seem promising
in NP can be efficiently reduced to it, i.e., with a poly- that quantum computers can efficiently solve
nomial overhead in time. Hence, NP-complete problems NP-complete problems [21, 31]. This does,
are the most difficult problems in NP. Notable examples however, advocate for quantum algorithms that
are TSP and Graph Colorability [29]. More broadly, a exploit structure within certain problem
problem is called NP-hard if it is at least as hard as the instances that arise in practice, rather than
hardest problems in NP — but not necessarily in NP attempting to try to develop fast quantum
itself. Thus, NP-complete problems are merely a subset algorithms for all instances [32].
of NP-hard problems that belong to NP.
When thinking about problems in NP, it is tempting
to imagine a quantum analog of this class, with complete
problems of its own. There are several possible ways to do QMA ⊆ PP [36].
this, see for example [33] for a discussion. We hone in on
one in particular — QMA, for quantum Merlin-Arthur.
Intuitively, QMA encapsulates decision problems where 4. Polynomial Time Hierarchy
“yes” instances can be verified efficiently and with high
probability on a quantum computer, using a quantum A useful way to extend the complexity classes like P
state as a proof. This state should occupy no more than and NP is to give those classes access to an oracle/set
a polynomial number of qubits in the input size, however,
A. This way one can define for example PA , the class of
finding such a state may of course, be computationally
languages decided by polynomial time Turing machines
hard. For the “no” instances all quantum proofs will be
that have oracle access to A. This means that the ma-
rejected (with high probability). Problems may then sim-
chines have access to a subroutine that computes in a
ilarly be classified as QMA-hard or QMA-complete (see
single step whether a query string x ∈ A. For a rigor-
[34]).
ous treatment of oracle Turing machines see for exam-
Other classes relevant to quantum computing that are
ple [37]. One can also define oracle access to a whole
more exotic than NP and QMA exist, albeit not nec-
class of problems: PNP and NPNP . The k th level of
essarily in the context of optimization. Due to a re-
the polynomial hierarchy Σpk is NP with k − 1 stacks of
cent Gaussian Boson Sampling experiment [35] the class NP
#P is worth mentioning. #P is a counting class, de- NP oracles. The third level for example is NPSNP .
fined as the set of all functions f such that for an input The polynomial hierarchy PH is then defined as k Σp k.
x ∈ {0, 1}∗ , f (x) counts the number of accepting compu- It is known that BPP ⊂ NP NP
but remains an open
tation paths of a nondeterministic polynomial-time ma- problem whether BQP is in PH, and interestingly, there
chine on input x. The complexity of problems in #P are oracles relative to whom this is not the case [38].
relates to challenges like computing partition functions Another relevant computational model is presented by
in statistical mechanics. The class PP may be thought Chen et al. [39]. The authors define the complexity
of as the decision analogue of #P and it is known that class NISQ as the set of decision problems solvable by

5
 II QUANTUM ADVANTAGE & COMPLEXITY THEORY

a “typical” instance which may be much easier than the

Decision Problems
hardest, worst-case instance.
EXP-TIME This implies that, even when worst-case complexity-
theoretic results rule out quantum advantage for NP-
PSPACE
hard problems, quantum computers may still exhibit ex-
PP
ponential speedups on instances relevant for practical
problems. More concretely, they may give efficient algo-
rithms for instances that take classical computers expo-
NP QMA nential time. We discuss this in more detail in Sec. II C.
On a similar note, while many problems of practical
interest are NP-hard, if we are interested in measuring
BQP
the computational complexity of problem which admits
NP-intermediate
an algorithm A, as a function of the input length n in
conjunction with some adjustable parameter ϵ, then in-
terestingly, parameterization could render otherwise NP-
P BPP hard instances efficiently solvable — although the depen-
dence could be exponential with respect to 1/ϵ. Such a
problem is termed fixed-parameter tractable (FPT) and
the complexity is sometimes referred to as parameterized
complexity [42, 43]. Therefore, FPT amounts to the pa-
rameterized analogue of P. A common example is that of
the Vertex-Cover problem, which is NP-complete when
Figure 1. The inclusions of various complexity classes con- considering the input size in isolation. However, there
cerned with decision problems, as discussed in Sec. II. While exists an O(n) algorithm when the problem is restricted
these inclusions are not known to be strict, this diagram cap- to constant-size covers [44]. Bremner et al. [45] intro-
tures the gist of prominent, canonical classes in complexity duced quantum parameterized complexity. In particular,
theory. the class FPQT is introduced as the class of parameter-
ized decision problems tractable by a quantum computer,
but intractable otherwise. It is well known that if a pa-
a BPP machine with has access to a noisy quantum de- rameterized problem is in FPT, it admits kernelization,
vice with poly(n) qubits. They then prove relative sep- which could be seen as a pre-processing algorithm that
arations with respect to a particular oracle O, such that reduces the size of an instance of the problem, often sub-
BPPO ⊊ NISQ ⊊ BQPO , which gives evidence that stantially [46, Theorem 1.4]. For benchmarking purposes
the power of NISQ lies strictly between BPP and BQP. in optimization, cf. Sec. VI, the notion of kernelization
It is unlikely, that quantum computers provide expo- is of crucial importance. Formally, a kernelization (also
nential speedups for NP-hard (and thus, NPO-hard) known as kernelization algorithm) for a parameterized
problems. See [40] for stronger hypotheses along these problem can be interpreted as follows: assume a lan-
lines. But exponential speedups for an interesting sub- guage L over an alphabet Σ and parameters in N, an
class of NP problems, not complete but also not in P, algorithm given (x, k) ∈ Σ∗ × N outputs a string x′ ∈ Σ∗
called NP-intermediate, cannot be ruled out. For exam- and an integer parameter k ′ < k, such that (x′ , k ′ ) ∈ L
ple, decision versions of factoring are believed to be of — if the algorithm runs in polynomial time and k can
this type. As depicted in Figure 1, these problems are be expressed as a polynomial of k ′ , the tuple (x′ , k ′ ) is
outside of P and within NP, but are not NP-complete. termed a polynomial kernel. Such polynomial kernels can
Related classes can similarly be studied for the optimiza- be obtained for many combinatorial optimization prob-
tion classes PO and NPO, and relational classes FP lems considered in theoretical computer science [47]. For
and FNP. For instance, the relational version of TSP FPT problems, it thus makes sense to benchmark on the
is FPNP -complete, whereas the decision version is NP- kernel.
complete [8]. MAX-SAT is NPO-complete, whereas the
Vertex-Cover and Bin-Packing problems are not (even
though their decision versions are NP-complete) [41]. B. Approximate Solutions

Thankfully, the story does not end at obtaining exact

5. Worst-case complexity and fixed-parameter tractability solutions efficiently, with high probability — if it did, we
would probably have a difficult time motivating classical
It is important to emphasize the following point: computing, let alone quantum, as a useful paradigm in
(combinatorial) optimization. In reality, finding a solu-
Complexity theory is traditionally focused on worst-case tion that is in some sense, close to the optimal is often suf-
problem instances. In practice, one often needs to solve ficient. The shift from finding the optimal solution, to an

6
 II QUANTUM ADVANTAGE & COMPLEXITY THEORY

For the MAX-E3-SAT problem, one of the earliest ap-

Optimization Problems proximation algorithms showed that a random variable
assignment satisfies each clause with probability 7/8, and
NPO
thus, achieves a 7/8-approximation ratio [48]. It was
APX
later shown that there is no classical polynomial-time al-
gorithm that can achieve a ratio of (7/8+ϵ), for any ϵ > 0
PTAS unless P=NP [49]. In such cases, where inapproximabil-
ity bounds are achieved by classical algorithms, improv-
FPTAS ing these ratios with quantum computers is highly un-
likely, as this would imply that quantum computers can
PO
solve NP-hard problems.
Some inapproximability bounds require additional con-
ditions. For example, in their seminal work in 1995,
Figure 2. Complexity classes for optimization problems. No- Goemans and Williamson introduced an approximation
tice that, similar to #P, the functional problem classes are algorithm for the MAXCUT problem and proved that
outside of the decision problems, and hence, not displayed in it achieves an approximation ratio of c ≈ 0.87856 [50].
this hierarchy. Later, it was shown by Khot et al. that there cannot
exist a polynomial time algorithm that achieves (c + ϵ)-
approximation [51] under the assumption of the Unique
approximate one, leads to the study of approximability Games Conjecture [52]. Again, in such cases, improving
— in particular, how hard is it to approximate solutions these ratios with quantum computers would imply that
to various problems? This leads to interesting complex- quantum computers can solve NP-hard problems or that
ity classes for optimization, induced by approximability. the Unique Games Conjecture is false. Similar results ex-
Intuitively, it may seem like allowing for approximate so- ist for other NP-hard problems, such as graph coloring
lutions could reduce the complexity of a problem, similar problem and set covering [53].
to those in FPT. Unfortunately, it is not that straightfor- There is also a quantum PCP conjecture proposed
ward when dealing with approximation and it seems that by Aharonov et al. [54–56]. At its core is the k-local
in many cases, approximating solutions to hard problems Hamiltonian problem: given a Hamiltonian H with k-
is just as hard as finding the optimal ones. We discuss local terms (i.e. operations acting on up to k qubits),
this so-called hardness of approximation next, followed determine the smallest eigenvalue of H. This problem
by classes of problems that can exhibit efficient approxi- is QMA-complete [57], akin to SAT’s NP-completeness.
mation. The local Hamiltonian problem comes with a promise de-
noted by γ. The ground state energy of H is either less
than some value a, or greater than a + b, where Γ := b
1. Hardness of approximation is the absolute promise and the ratio γ := b/m, where m
is the number of k-local terms, is the relative promise.
In the context of approximate solutions for optimiza- The quantum PCP theorem then states that the k-local
tion problems, a longstanding result called the PCP- Hamiltonian problem is QMA-hard even with a constant
theorem informally states that, for many problems, com- relative promise gap γ > 0, if the interaction graph is
puting an approximate solution with arbitrary precision sparse and all k-qubit interactions have bounded norm.
is as hard as computing the exact solution. To discuss There is some supporting evidence for the quantum PCP
this a bit more formally, the approximation ratio, which theorem, perhaps most notably the recently proved No
— as the name suggests — is the ratio between the (ex- Low-Energy Trivial State (NLTS) theorem [58]. Under
pected) approximate solution and the optimal, quanti- NLTS, there exist families of Hamiltonians for which all
fies the quality of a solution provided by an approxima- low energy states require circuits of super-constant depth
tion algorithm. For many NP-hard problems, there are to prepare. If NLTS were false, this would imply that all
known bounds on the achievable approximation ratios, Hamiltonians have low energy states that are efficiently
and these are referred to as “inapproximability bounds”. simulable — effectively rendering the quantum PCP the-
Such bounds can be derived through many techniques, orem invalid since approximating the ground state energy
including information-theoretic arguments (e.g., any al- could be done efficiently using these low energy states.
gorithm that achieves an approximation factor better While the quantum PCP theorem implies NLTS, under
than c must make at least exponential queries to a func- the assumption that QMA is not equal to NP, the other
tion), or reductions to known NP-hard problems (e.g., if direction does not hold, since there could be other types
a problem can be approximated better than its inapprox- of classically simulable circuits that could prepare low en-
imability bound, then it is equivalent to solving another ergy states, other than low depth circuits. This remains
NP-hard problem). The former holds independent of a very open and interesting research area.
whether P is equal to NP, but the latter inapproxima- Indeed, the hardness of approximation seems to im-
bility bounds are said to be conditional on P ̸= NP. ply – theoretically – that quantum computers will not

7
 II QUANTUM ADVANTAGE & COMPLEXITY THEORY

be very influential, even if approximate solutions are the case of so-called metaheuristic algorithms). By this nega-
goal. However, this is not necessarily true in general, tive definition, our understanding of heuristic algorithms
and less so in practice. There is hope for quantum algo- is rather limited. Having said that, there are multiple at-
rithms to outperform classical methods when there is a tempts to provide more formal definitions. For instance
gap between known inapproximability bounds and prov- Bogdanov et al. [65] define distributional problems, which
able approximation factors (e.g., in metric TSP, [59, 60]). are pairs (L, D) of a language L and a family of distribu-
Further, quantum optimization methods might provide tions D = {Dn }, one for each input size n. Then, a dis-
provable and computational speedups, even if they sim- tributional problem (L, D) is in the class HeurBPP (for
ply match the approximation factors of classical algo- heuristic BPP) if there exists a polynomial-time random-
rithms. ized classical
  algorithm A such that for  all n and δ > 0,
Prx∼Dn Pr A(x, 01/δ ) = L(x) ≥ 2/3 ≥ 1−δ where the
inner probability is taken over the internal randomization
2. Polynomial-time approximation schemes of A. In other words, algorithms that solve problems
in HeurBPP are allowed to err on a small fraction of
If a problem is “approximable”, it is said to belong to instances drawn from Dn . Analogously, we say that a
the class APX which contains the set of NPO prob- distributional problem (L, D) is in HeurBQP if there
lems for which there are polynomial-time approximation exists a polynomial-time quantum algorithm A that sat-
algorithms with approximation ratios bounded above by isfies the same property. This “distributional” view of
some constant c [61]. An interesting subclass of NPO heuristics has been rather popular in relation to quan-
problems that admit a so-called polynomial-time approx- tum advantage.
imation scheme is PTAS [25] [62, Chapter 8]. For any Notably, Pirnay et al. [66] prove a separation for a
problem in PTAS, there is a polynomial-time algorithm combinatorial optimization problem under very special
that is guaranteed to find a solution whose value is within distributions in the sense above, namely those which are
a 1 + ϵ factor of the optimum, where ϵ > 0. For exam- Karp-reduction-images of factoring problems as pointed
ple, by restricting to the Euclidean plane, TSP is then out by Szegedy [67]. The former proves that quantum
in PTAS [63]. However, as with similar issues in FPT, computers can have a super-polynomial advantage over
there are cases where the exponent of the polynomial a certain class of NPO-problems and that this also holds
might depend exponentially on 1/ϵ. To account for this, in the approximation sense. It does so by making use of
one may consider the class FPTAS for fully polynomial- cryptographic tools and by means of notions of computa-
time approximation scheme, which demands the run time tional learning theory, in particular the Occam learning
to be polynomial in the input size, as well as in 1/ϵ. framework. The latter points out that similar separations
As such, FPTAS ⊆ PTAS, and the hierarchy of these hold for other NPO approximation problems, by resort-
classes are displayed in Figure 2. When combined with ing to the PCP theorem. The approach is via faithful
quantum computing, this area makes for an exciting and reductions from well-established problems in NP, which
rather unexplored research direction — where quantum have a large quantum advantage, namely factoring, to
algorithms could serve as new approaches or improve over matching instances of NPO problems. Using techniques
existing classical PTAS algorithms. from complexity theory, it is possible to construct in-
stances with a large approximation gap, which ensure
that classical algorithms cannot approximate the correct
C. Heuristics in Quantum Optimization solutions to these instances, unless factoring is classically
easy. These findings, however, are unlikely to have much
Thus far, the focus has been on worst-case instances. practical value as the proposed quantum solutions will
In practice, heuristic approaches can provide useful so- only work on the contrived instances corresponding to
lutions to special, relevant and typical problem in- factoring, and are not general optimization solvers which
stances [62, e.g. Chapter 28]. Heuristics have also proven could solve other classes of instances efficiently as well.
useful in problems exhibiting structure and average-case The results do, at least, contribute to understanding
instances — which are oftentimes far more practical in what kind of quantum advantages one can hope for in
real-world scenarios [62, 64]. The success of classical combinatorial optimization.
heuristic algorithms indeed motivates the study of quan- Independently, one could relate the distributional view
tum heuristic methods, where the aim of the latter is to of heuristics to numerous studies of so-called land-
design algorithms that leverage quantum computation to scapes [68]. The term landscape here aims to gener-
provide reasonable solutions to problems where classical alize an intuitive mental picture where one associates
approaches either cannot, or simply take too long to do the height of terrain to solution quality and traversing
so. Naturally, the word reasonable depends very much on this terrain corresponds to moving through a space of
the context, making heuristic approaches difficult to an- solutions (which need not be two dimensional in gen-
alyze theoretically. More concretely, heuristic algorithms eral). For example, Refs. [69–71] study the landscape
are understood as algorithms without provable perfor- associated with variational quantum algorithms for spe-
mance guarantees, run time guarantees, or both (as in the cific problems, with the aim of explaining when these

8
 III PARADIGMS IN THE DESIGN OF QUANTUM OPTIMIZATION ALGORITHMS

heuristic methods converge. An important feature in the rithms (and local low-depth quantum algorithms), but it
study of landscapes are so-called barren plateaus [72–79], remains unclear whether these problems can be addressed
especially in conjunction with variational quantum al- by higher-depth and/or more sophisticated quantum al-
gorithms, which correspond to loss landscapes that are gorithms [93–96]. If so, this would be very convincing
exponentially flat in the number of system qubits. In evidence for useful superpolynomial quantum advantage,
fact, this phenomenon provides a central bottleneck in since such problems naturally arise in combinatorial opti-
the scaling of variational quantum algorithms to practi- mization, statistical physics and high-dimensional statis-
cally relevant problem sizes. There are multiple reasons tics. Beyond OGP in the average-case setting, there is
for the appearance of barren plateaus. It was conjec- an emerging literature on the “computational-statistical
tured that they are induced by an ansatz that is (close gaps” of many optimization and learning problems where
to) a t-design [72, 75] or exhibits too much entanglement we could also hope for a superpolynomial quantum ad-
paired with partial measurements [76]. These findings, vantage. Moving into the approximation realm also in-
along with other causes of barren plateaus were united troduces interesting open questions for quantum com-
in theorems which either relate the occurrence to a math- puting. For example, can quantum algorithms saturate
ematical object called the dynamical Lie algebra, which inapproximability bounds that have not yet been reached
is induced by the structure of the generators and the form by classical methods? Moreover, if approximation ratios
of the variational circuit [70, 71], or to the average behav- cannot be improved, there may still be room to speedup
ior of the respectively induced light-cone [80]. Notably, existing approaches with quantum computing.
these phenomena may be avoided if a good warm-starting Another intriguing aspect of complexity theory, is that
strategy is known–in the sense of a parameterization that a restriction of some parameters of a problem, or some
brings the system sufficiently close to the optimum. How- structure present, may drastically change the complex-
ever, it is often unclear how such a warm-start should be ity. A perfect example of this is TSP, which – in its
chosen. most general decision version form – is NP-hard and
There have also been a number of works suggesting the optimization version is APX-hard. If we restrict
that quantum sampling could be a useful subroutine to Euclidean TSP, where all points are in Rd and edge
in a larger classical algorithm. For example, there are weights equal their Euclidean distances, then, remark-
quantum circuits for Metropolis-Hastings algorithms [88] ably, there exists an explicit PTAS to solve such TSP
and quantum enhancement of Markov-chain Monte-Carlo instances [63]. Shifting to heuristic algorithms for TSP,
(MCMC) algorithms [89, 90]. These approaches should one can actually solve problems in practice with up to
be possible to analyze, and thus, turn into randomized millions of variables. Lastly, if we were to consider the
approximation schemes. shortest path problem instead of TSP, where we just want
the shortest path between two places, we would have a
problem in P. This illustrates the nuance of complex-
D. Takeaways from Complexity Theory in ity theory that one should keep in mind when designing
Quantum Optimization quantum optimization algorithms. Akin to TSP classi-
cally, if provable quantum speedups seem unattainable,
Although it is widely believed that quantum computers quantum heuristic algorithms may still offer meaningful
will not offer exponential speedups for NP-hard prob- results. But developing quantum heuristic methods that
lems, the story does not end there. As we have seen, exploit structure in a problem, or enhance classical ap-
most complexity-theoretic statements deal with worst- proaches, is not straightforward either. This serves as
case settings for exact solutions. In practice, typical op- the basis for the next section, where we lay out various
timization problem instances could differ substantially building blocks to carefully design quantum optimization
from the worst-case, and thus, exponential quantum algorithms.
speedups in these more realistic settings are still pos-
sible in principle. Importantly, the picture may drasti-
cally change when dealing with average-case hardness.
Average-case settings could exclude particularly bad in- III. PARADIGMS IN THE DESIGN OF
QUANTUM OPTIMIZATION ALGORITHMS
stances for which most complexity-theoretic statements
hold, and thus, an advantage – certainly one faster than
simply the Grover approach – becomes possible and it There are multiple approaches that lead to exact quan-
is an interesting open problem whether quantum algo- tum optimization algorithms, such as Grover (Adaptive)
rithms can solve optimization problems that are average- Search [97–99] and the Quantum Adiabatic Algorithm
case hard for classical algorithms. For example, there is (QAA). There are lesser known approaches too, like
evidence from [91] that quantum algorithms can solve Quantum Imaginary Time Evolution (QITE) [100], but
random k-SAT faster than Grover and the state-of-the- much of the published work focuses on the more well-
art SAT solvers. Additionally, problems with a so-called known approaches. In this section, we introduce the core
Overlap Gap Property (OGP) [92] have been proven to concepts deployed as subroutines in various quantum op-
be average-case hard for certain families of classical algo- timization algorithms discussed in Sec. IV.

9
 III PARADIGMS IN THE DESIGN OF QUANTUM OPTIMIZATION ALGORITHMS

Table I. An overview of the classical problem classes relevant to optimization problems. The models of analogue computation
below the horizontal line are not considered in much detail in this section.
Complexity class Example problem Key features Year Ref.
FPTAS Knapsack, Subset-Sum Efficiently approximable [81]
PTAS Euclidean TSP Approximable in poly-time [63]
APX-Complete MAXCUT, MAX-3-SAT, Metric-TSP Discrete decisions 1991 [25, 81]
NPO-Complete TSP, Binary IP Discrete decisions 1988 [82]
#P Two-stage SP [83, Theorem 3.1] Stochastic objective function 2006 [83, 84]
PSPACE Multi-stage SP [83, Theorem 4.1] Stochastic objective function 1985 [83, 85]
Banach-Mazur computability Optimal Transport [86, Theorem 5] Real-valued decisions 2003 [86]
NP over Reals Optimal Transport [86, Theorem 5] Real-valued decisions 1996 [87]

A. Grover Search with λ(0) = 0 and λ(1) = 1. If the annealing is performed

slowly enough, in other words, if T is large enough, with
The seminal Grover algorithm for searching an un- a suitable annealing schedule, the Adiabatic Theorem
structured database achieves a quadratic advantage over [106–109] guarantees that the state |ψt ⟩ always remains
classical search in terms of the number of queries to the in the ground state of H(t). Thus, |ψT ⟩ represents the
database (i.e. the query complexity). In exact quan- ground state of H, which corresponds to the solution
tum optimization, this advantage translates to an ad- of the encoded problem. It has been proven that the
vantage for function minimization [97] and can achieve annealing time has to scale inversely with the minimal
a quadratic speedup over brute force search in many squared spectral gap ∆ of H(t) over all t ∈ [0, T ], i.e.,
problems in√combinatorial optimization [101], i.e., a run T = O(1/∆2 ). During annealing, the system undergoes
time of O( 2n ) instead of O(2n ) for n-variable prob- quantum phase transitions, and in the glassy phase, at
lems. In the most general case, without any further as- low values of t/T , the energy gap closes exponentially
sumptions on the problem, for example in terms of or- with increasing system size [110–114]. Consequently, this
acle used in the query complexity, this is the best run implies exponentially long run times if one aims to pre-
time and advantage that we can hope for when using a cisely follow the adiabatic path. Hence, we are limited
quantum computer for discrete optimization [102]. For to at most polynomial speedups.
a concrete problem class, however, brute force search is The practical implementation of the quantum anneal-
rarely the best available classical algorithm, and thus, ing protocol often involves repeated finite-time runs [115]
the actual advantage achieved by Grover Search is usu- or, potentially, diabatic annealing [116]. This variant re-
ally sub-quadratic [103]. Nevertheless, leveraging Grover laxes the constraint that the system must always remain
Search as a sub-routine, together with other (quantum or in the instantaneous ground state of H(t). Another vari-
classical) algorithms, can sometimes help to recover an ant is counter-diabatic annealing, where additional terms
overall quadratic speedup over the best available exact are included in the Hamiltonian to suppress the transi-
classical algorithm [104]. tion to excited states during the annealing [117, 118]. For
example, the counter-diabatic terms can be built up from
nested commutators [119], while the evolution can be en-
B. Quantum Adiabatic Algorithm gineered with optimal control [120]. In its practical im-
plementation, QAA often becomes a heuristic algorithm.
On a gate-based quantum computer, the time evolution
The Quantum Adiabatic Algorithm (QAA) [2, 3] is
to generate |ψt ⟩ can be implemented, e.g., through Trot-
another exact optimization algorithm, reminiscent of ho-
terization [121–123]. The typical long annealing times
motopy methods of numerical mathematics [105]. It as-
then translate to long quantum circuits.
sumes a problem Hamiltonian H whose ground state cor-
responds to the solution of a corresponding optimization
problem, as well as a so-called mixing Hamiltonian HX . C. Quantum Phase Estimation
The mixing Hamiltonian should have an easy-to-prepare
ground state |ψ0 ⟩ that has a non-zero overlap with the
Pn (i) Quantum Phase Estimation (QPE) is an algorithm to
ground state of H. A common choice is HX = i=1 σX , estimate the eigenphases of unitary operators [124]. It al-
(i)
where σX denotes the Pauli X matrix on qubit i, with lows the estimation of the eigenvalues of certain matrices
⊗n
the ground state |+⟩ , i.e., the n-qubit equal superposi- and can be used to find the ground state and correspond-
tion state that has non-zero overlap with every computa- ing energy of various Hamiltonians [125–127]. Famously,
tional basis state. Starting in |ψ0 ⟩, QAA slowly drives the QPE is also leveraged by Shor’s algorithm. Factoring is
state according to H(t) = λ(t/T )H + (1 − λ(t/T ))HX , believed to be in NP-intermediate [41] and Shor’s algo-
and d/dt |ψt ⟩ = −iH(t) |ψt ⟩, for some total annealing rithm [128] achieves an exponential speedup over classical
time T > 0 and an annealing schedule λ : [0, 1] → [0, 1] algorithms, as discussed in Sec. II. Like many other prob-

10
 III PARADIGMS IN THE DESIGN OF QUANTUM OPTIMIZATION ALGORITHMS

lems, factoring can be cast into a quadratic unconstrained (Gibbs) free energy of the system, log Z, falls within the
binary optimization problem (QUBO) [129–131]. Thus, complexity class BPPNP [146, 147]. Here, β is the in-
this gives us an exponential speedup over a subset of verse temperature. Gibbs sampling, then, can be done
QUBOs that happen to correspond to instances of fac- with the Metropolis-Hastings algorithm [148, 149] which,
toring. Similar reductions are possible from other classes despite its exponential scaling with particle number, of-
of problems to factoring [67, 132]. In practice, however, ten exhibits convergence.
we do not expect these instances to appear naturally, and In contrast, quantum Gibbs sampling [150] presents
it is an open question how to harness this relation in the greater challenges and is believed to be at least QMA-
context of optimization. hard [151, 152]. The quantum Gibbs state σ β is (up to
In general, QPE requires an initial state having a poly- normalization) defined as
nomial overlap with the ground state, i.e. an overlap
greater than 1/poly(n). For a diagonal Hamiltonian, ev- dimH
X
ery computational basis state is an eigenstate, and so σ β ∝ e−βH = e−βEi |ψi ⟩ ⟨ψi | , (1)
QPE does not increase the chance of sampling the ground i=1
state. Further, simply sampling from such an initial state
provides a polynomial-time algorithm to find the ground where H is the Hamiltonian operator, |ψi ⟩ are the eigen-
state already. Thus, the true problem is finding such states, as before β is the inverse temperature, and Ei are
an initial state with a polynomial overlap. Nonetheless, the energy eigenvalues.
some algorithms encode an optimization problem into a The problem of thermalizing a quantum state has a
non-diagonal Hamiltonian, potentially making QPE use- long tradition and was studied earlier in [153]. Quan-
ful in quantum optimization. The situation is similar for tumly, the difficulty does not only relate to the hardness
(Krylov) subspace expansions [133, 134], where an ini- of sampling but also to the difficulty to efficiently prepare
tial polynomial overlap is assumed to derive convergence Gibbs states. Quantum adaptations of the Metropolis-
properties towards the ground state of a Hamiltonian. Hastings algorithm that involve quantum phase esti-
To make QPE and related techniques relevant for a mation (QPE) are challenging due to the algorithmic
large set of applications, it is important to have a quan- complexity[154, 155]. Furthermore, Gibbs sampling al-
tum toolbox of matrix linear algebra operations [135– gorithms for Gibbs states of local Hamiltonians at all
140], in analogy to the classical basic linear algebra sub- temperatures have been devised in [156]. It has also been
programs (BLAS) [141]. Quantum signal processing [142] seen that high-temperature Gibbs states are unentangled
enables the simulation of sparse matrices, i.e., the imple- and efficiently preparable [157].
mentation of the time evolution operator e−iHt for a ma-
A variational method for quantum simulators to gen-
trix H and a time t, with optimal complexity. Phase esti-
erate finite temperature Gibbs states through the prepa-
mation then allows one to extract eigenvalues and eigen-
ration of thermofield double states was introduced in
vectors of H. Quantum singular value transformation
[158]. A quantum algorithm to prepare the quantum
[143] generalizes this result and allows efficient simula-
Gibbs state, as well as the partition function within ϵ-
tion of matrix algebra such as A + B, AB, and singular
error, was proposed in [150]. A different VQA approach
value transformations P (A), with P being a low-degree
for the preparation of Gibbs states suitable for NISQ de-
polynomial applied to the singular values of A, where A
vices was presented in [159] where the authors presented
and B are matrices. The block-encoding framework uni-
an algorithm in their paper by applying a truncated Tay-
fies input models for matrices given efficient descriptions,
lor series to compute the free energy, and then select-
sparse access, or quantum memory. These tools find ap-
ing this truncated value as the loss function. Another
plications in preparing Gibbs states, matrix problems,
VQA was proposed in [160] where the variational param-
and continuous optimization.
eters are chosen by minimizing the free energy. Warren et
al. [161] introduced an adaptive VQA approach to pre-
pare Gibbs states by introducing an objective function
D. Gibbs Sampling that is easier to measure than the free-energy and us-
ing dynamically generated, problem-tailored ansätze. Fi-
Gibbs sampling corresponds to generating sample se- nally, a variational method for approximate Gibbs prepa-
quences from the joint probability distribution of multi- ration via imaginary time simulation has been presented
ple variables underlying a Gibbs distribution. Geman in [162]. The difficulty in these variational approaches ei-
and Geman [144] introduced a Markov Chain Monte ther lies in the requirement to continuously compute the
Carlo (MCMC) algorithm to facilitate Gibbs sampling. von Neumann entropy or to correctly propagate imagi-
Its significance in optimization, particularly in classical nary time dynamics via a variational ansatz. If sampling
MCMC algorithms, extends to solving problems like con- from the Gibbs state is sufficient, approaches like “Quan-
straint satisfaction [145]. tum Minimally Entangled Typical Thermal States” [100]
Sampling from a classical many-body Hamiltonians’ H can be used, which are based on repeated imaginary time
Gibbs distributions, and the related problem of
approx- evolution of a weakly entangled initial state. Wild et
imating the partition function Z(β) = Tr e−βH or the al. [163] introduced a set of quantum algorithms that

11
 III PARADIGMS IN THE DESIGN OF QUANTUM OPTIMIZATION ALGORITHMS

provide unbiased samples by creating a state that en- tion algorithms via Quantum subroutines. For instance,
codes the quantum Gibbs distribution in its entirety and Semidefinite Programming (SDP) relaxations, as used
demonstrate that this method can outperform classical in the famous Goemans-Williamson algorithm for MAX-
Markov chain algorithms in various scenarios, including CUT [179, 180] or more general variants for QUBO [181],
the Ising model and sampling from weighted independent may profit from Quantum SDP solvers, cf. Sec. IV B 1.
sets on two distinct graphs. Additionally, recent advance- Eventually, for special cases, these may even achieve an
ments in quantum Gibbs sampling, like the Dissipative exponential speedup for solving some SDPs compared
Quantum Gibbs Sampling algorithm [147], offer a simpler to classical algorithms. However, unless the problem is
and potentially more efficient alternative, as they utilize given in a compact functional form and assuming we are
local update rules and are more feasible for near-term not interested in reading out the full solution but only the
quantum hardware. optimal objective value, the end-to-end advantage will be
Quantum Gibbs sampling plays a crucial role in various reduced to polynomial due to input/output bottlenecks.
optimization algorithms, particularly in convex optimiza- It also raises the question whether classical algorithms
tion as discussed in Sec. IV B 1. Quantum speedups are may be able to leverage this special setting to achieve
possible because, under certain conditions, Gibbs states similar improvements, as has been shown, e.g., in the
can be prepared more efficiently quantumly than classi- context of recommendation systems [182] or solving low-
cally. The block-encoding framework [143] demonstrates rank linear systems [183]. Further, there are many prac-
this, where√an n×n Gibbs state is constructed with com- tical problems to achieve such an advantage for solving
plexity O( n) in the dimension parameter, albeit other SDPs, as will be discussed in more depth in Sec. IV B 1,
parameters also appear in the running time. This method and very likely this will require FTQC.
and its applications in algorithms like the Matrix Multi-
plicative Weights Update for solving general semidefinite
optimization problems are further elaborated in [164–
167]. Notably, these algorithms leverage Gibbs states
for computing trace inner products. F. Variational Methods
Furthermore, the complexity analysis of quantum par-
tition functions, as explored in [168], reinforces the im-
portance of efficient Gibbs sampling methods in quantum Within the near-term intermediate-scale quantum
computing. Encoding Gibbs distributions is at the heart (NISQ) devices, much work is focused on so-called Vari-
of simulated annealing approaches for Gibbs partition ational Quantum Algorithms (VQAs) [184–188], i.e., a
functions [169–172], which can be used to approximately family of randomized search algorithms that solve cer-
count certain combinatorial objects such as k-colorings or tain optimization problems classically while evaluating
matchings of a graph, with genuine quantum speedups in involved expectations quantumly. Specifically, a VQA is
theory. Gibbs sampling methods find application in spe- an algorithmic scheme where one provides some prob-
cialized solutions to semidefinite relaxations of problems lem data D, a parameterized ansatz |Ψ(ϑ)⟩ and a corre-
like MAXCUT [173, 174], see Sec. IV B 1, and are integral sponding cost function C(ϑ) as input. Then, the pa-
to quantum algorithms for zero-sum games [175–177]. Fi- rameterized quantum state is measured in some basis
nally, [178] introduced the notion of a Gibbs objective {Ok }k∈N ∈ Herm(Cn ) to evaluate the cost function. The
function which can be useful for quantum optimization next step involves optimizing over the cost function in
problems. The objective can be evaluated by the means order to obtain updated parameters, which are then fed
of Gibbs sampling, although it can be challenging, as dis- back onto the circuit. See Fig. 3 for a schematic.
cussed previously. We further mention this approach in Although finding the optimal parameters in a VQA is
Sec. IV A 1. in general NP-hard [189], it can still result in a powerful
heuristic to find good solutions, even potentially offering
speedups [190]. This situation is similar to the training
E. Approximation Algorithms of classical artificial neural networks, whose training to
the optimum is also NP-hard [191, 192] (even with only
Grover Search, (Trotterized) QAA, QITE, (and QPE) k = 2 layers, a ReLU neural network cannot be trained
are all expected to require Fault-Tolerant Quantum Com- in time bounded by a polynomial in the dimension of the
puting (FTQC) due to the resulting circuit sizes. How- weights [193]). VQAs differ mainly in the choice of the
ever, they also serve as theoretical motivation for approx- ansatz, the cost function, and the optimizer, as discussed,
imation algorithms and heuristics, which may already for example, in Sec. IV A.
lead to quantum advantages in optimization with noisy Although there are analyses of asymptotic convergence
quantum computers, as discussed later in this section. relating variational methods to the QAA [194], bounds
In addition to the exact algorithms discussed in Sec. II, on iteration complexity of variational methods [195, 196],
there exist also approximation algorithms that guaran- and in some cases, bounds the objective function value
tee a certain approximation ratio. Quantum Computing at the limit point [197, 198], VQAs are still mostly seen
could be used to accelerate known classical approxima- as heuristics.

12
 IV PROBLEM CLASSES & ALGORITHMS

A. Discrete Optimization
Input
{D, U (ϑ), |Ψ0 ⟩, C(ϑ)}
Discrete optimization is a branch of mathematical op-
timization that focuses on problems where, as the name
suggests, variables take on discrete values. This often
implies a combinatorial explosion of possible solutions
with respect to size of the input parameters. We focus
on discrete optimization problems with and without con-
|Ψ0 ⟩ U (ϑ) |Ψ(ϑ)⟩ straints. While problems of one class can be converted
w.r.t. {Ok } into problems of the other class, approaches to solve them
Quantum can differ significantly and leveraging available structure
might be beneficial in many cases. Moreover, the equiv-
alence often only holds if a globally optimal solution is
updated ϑ∗ b
E[{Ok }] found. The remainder of this section discusses exact algo-
rithms, approximations, and heuristics for the considered
problem classes within discrete optimization.

arg min C(ϑ)

ϑ 1. Unconstrained Discrete Optimization

Classical In the realm of unconstrained discrete optimization,

Quadratic Unconstrained Binary Optimization (QUBO)
serves as a foundational example and is given as follows:

min xT Qx, (2)

x ∈ {0, 1}n
Output
where Q ∈ Rn×n is the cost matrix and n ∈ N denotes the
number of variables. Other types of unconstrained dis-
Figure 3. Variational methods iterate between the evaluation crete optimization problems involve, for instance, high-
of empirical estimates of the expectations of a set of observ- order polynomial objective functions [199], also referred
ables {Ok } associated with a parameterized quantum circuit, to as Polynomial Unconstrained Binary Optimization
and classical computation of updated parameters for the pa- (PUBO), or black box objective functions [200] as well
rameterized quantum circuit.
as (bounded) integer variables. While one could also de-
fine linear unconstrained discrete optimization, this class
is of little interest as it can be solved by looking at every
variable independently. Strictly speaking, QUBO also al-
lows one to model linear objective functions when Q is a
diagonal matrix, because binary variables satisfy x2 = x.
IV. PROBLEM CLASSES & ALGORITHMS Some problems, such as MAXCUT, are naturally for-
mulated as a QUBO, and most unconstrained discrete op-
timization problems can be mapped to QUBOs [201–204],
for example, by replacing integer variables by a suitable
There are many different classes of optimization prob- encoding [205, 206] or by converting higher-order polyno-
lems that are defined through the types of variables in- mial terms to additional binary variables and quadratic
volved, the objective function(s) used, presence of con- terms [207, 208]. In addition, certain constraints can
straints, stochasticity, and so on. Furthermore, problems be modeled by translating them into penalty terms and
can be modeled in different ways, which can lead to rep- adding them to the objective function [201, 203]. How-
resentations of the same problem in different problem ever, these conversions usually assume that the problems
classes. Although these formulations might be formally are solved exactly, i.e., that the global optimum is found.
equivalent, the chosen formulation can play a crucial role In the case of approximate solutions, this has to be care-
in how a problem is solved, as different formulations put fully analyzed. For instance, equality constraints might
more or less emphasis on certain properties of a problem. only be satisfied approximately when being represented
In this section, we introduce the most important classes via a penalty term, which, depending on the application,
of optimization problems and discuss existing algorithms might be acceptable or not. Furthermore, binary encod-
to solve them. We then discuss key open questions in ings of integers usually do not reflect the intrinsic order
each domain that could help advance the field of opti- between values, i.e., they lack some important structure
mization. that a solver may leverage to find good solutions.

13
 IV PROBLEM CLASSES & ALGORITHMS

Within this section, we focus primarily on QUBO. QU- mize the expected value of this quantity. The variational
BOs are a natural problem class to consider for quan- principle guarantees that the expectation value is lower
tum optimization as they can easily be converted into a bounded by the ground state energy of H, with equality
ground state problem, often explored in quantum com- if and only if the ground state is reached.
puting, usually in the context of quantum chemistry or The Quantum Approximate Optimization Algorithm
quantum physics [184]. The relationship between solving (QAOA) [185] denotes a special case of VQE that can
a QUBO and finding the ground state of a Hamiltonian is be applied to find good solutions to QUBOs. QAOA is
fundamental to many quantum optimization algorithms. motivated by QAA and has been shown to be a com-
The most common translation involves two steps: first, putationally universal algorithm [212]. More precisely,
substituting binary variables x ∈ {0, 1} with spin vari- QAOA uses a particular problem-dependent ansatz given
ables z = 1 − 2x, where z ∈ {−1, +1}, and secondly, re- by pairs of unitaries in each layer. These unitaries have
placing these spin variables with Pauli Z matrices.
n n
The the form
result is a diagonal Hamiltonian H ∈ R2 ×2 that en-
codes the objective values of Eq. (2) on its diagonal. The UX (β) = e−iβHX , (3)
ground state of this Hamiltonian represents the QUBO’s UP (γ) = e −iγH
, (4)
optimal solution [201].
QUBOs with spin variables z ∈ {−1, +1} natu- and they are termed as mixing and problem unitaries, re-
rally arise in physics and are referred to as Ising spin spectively. The unitaries are repeated alternatingly with
glasses [209]. Here, the present quadratic terms are usu- new parameters for each layer and applied to the ini-
⊗n
ally given via a lattice or graph, and their weight cor- tial state |+⟩ , as shown in Fig. 4. Here, H denotes
responds to an interaction strength between neighboring the problem Hamiltonian, and HX denotes the mixing
sites. By allowing a polynomial resource overhead, many Hamiltonian, see Sec. III B for more details. Both uni-
graph covering and coloring problems can be mapped to taries are alternated and repeated p times, with individ-
an Ising spin glass on a regular lattice [201]. However, ual parameters βj , γj for j = 1, . . . , p.
to save qubit resources, many benchmarks use simpler
Ising Hamiltonians on hardware-efficient lattices without
a specific problem embedding [115].
In general, QUBO problems are NP-hard [209]. In
principle, one could consider, for instance, Grover search
or quantum annealing, as suggested in Sec. III, to solve
QUBOs; however, these algorithms with provable perfor-
mance guarantees are expected to require FTQC. QUBO
problems are also APX-hard, i.e. there cannot exist a Figure 4. Structure of QAOA ansatz: every layer consists of a
PTAS for QUBO problems unless P=NP [210]. But, mixing and a problem unitary with corresponding parameters
there exist PTASes for certain problems that can be for- γj and βj for layer j = 1, . . . , p.
mulated as QUBO problems, like the Euclidean Trav-
eling Salesperson Problem [211]. Little is known about QAOA is an approximation algorithm with worst-case
PTASes on quantum computers; there are many possi- performance bounds for certain problems and algorith-
bilities to try using quantum algorithms to accelerate mic settings. For example, there exist worst-case perfor-
known classical PTASes. An example might be to ap- mance guarantees for QAOA with p = 1, 2, 3 layers for
ply Grover Search or Quantum Dynamic Programming, the MAXCUT problem on 3-regular graphs. Specifically,
cf. Sec. IV D, as sub-routines to achieve (sub-) quadratic for p = 1 Farhi et al. [185] derived a lower bound on
speedups compared to purely classical PTASes. Like the approximation ratio given by 0.692, while Wurtz and
Grover Search, Quantum Dynamic Programming is also Love [213] found that a lower bound given by 0.7559 for
expected to require FTQC. Since exact algorithms and p = 2 and (under certain assumptions) 0.7924 for p = 3.
acceleration of PTASes most likely require FTQC, there Furthermore, the performance of QAOA for (large-girth)
is arguably more hope for a near-term quantum advan- d-regular graphs has been analyzed numerically and it
tage in optimization sourced from quantum approxima- has been shown to outperform many known SDP-based
tion algorithms and heuristics. relaxations from p ≥ 11 [214, 215]. However, for gen-
Originally proposed for quantum chemistry applica- eral QUBO, QAOA is a heuristic without performance
tions, the Variational Quantum Eigensolver (VQE) [184] guarantees.
denotes a class of heuristic VQAs to approximate ground A crucial task for VQAs is to determine the optimal
states of Hamiltonians. This is achieved by choosing a parameters, cf. Sec. III F. The problem-dependent ap-
parametrized quantum state |Ψ(ϑ)⟩ = U (ϑ) |Ψ0 ⟩, where proach QAOA to construct an ansatz implies a relatively
U (ϑ) denotes a parametrized quantum circuit and |Ψ0 ⟩ small number of parameters, which is desirable for the
denotes an initial state, and then using a quantum com- corresponding classical optimization loop. Considerable
puter to evaluate ⟨Ψ(ϑ)|H|Ψ(ϑ)⟩ for given parameter research has been published on training strategies for
values ϑ, and employing a classical optimizer to mini- QAOA as well as on properties of the optimal param-

14
 IV PROBLEM CLASSES & ALGORITHMS

eters. There are different strategies to solve this problem provable guarantees on warm-started QAOA for finite
as we will discuss now. circuit depth with these custom mixers remains an open
The conventional approach to parameter optimization question. Other variants suggest to reuse qubits, lever-
in a VQA is to leverage the quantum computer to eval- aging mid-circuit measurements and dynamic circuits, to
uate the loss function and subsequently use a classical increase the number of variables for a limited number of
optimizer to improve the parameters. Good parameter qubits [226]. It seems also natural to restart the search,
initialization is crucial for this task to succeed and there when one arrives in so-called barren plateau [196], where
is increasing evidence that this requires some physics- the (sub)gradient is not informative enough. Mastropi-
inspired initialization, rather than generic random ini- etro et al. [196] have shown that the more barren regions
tialization [216–218]. Proposed strategies range from there are and the higher their proportion in the parame-
annealing-inspired parameters [217], to machine learning ter space, the higher the speedup obtained by restarting.
based approaches [219]. Similarly, one can try to evalu- Another alternative approach to QAOA is Recursive
ate the loss function for the optimization problem classi- QAOA [227], which uses the quantum computer to pro-
cally and only sample solutions from the quantum com- duce a sequence of reduced problems. Here, at each step
puter. The rationale here is that evaluating (or approx- QAOA is used to estimate correlations between variables,
imating) the (two-)local observables involved in QAOA and the problem Hamiltonian is reduced by fixing the
for QUBO using say, tensor networks or Clifford pertur- strongest one. This process is iterated until the reduced
bation theory, is computationally easier than sampling problem becomes small enough to for (provably exact or
from the complete quantum state [220, 221]. Alterna- heuristic) classical solvers, which then yields an approxi-
tively, for some classes of QUBOs, QAOA has a property mate solution to the input problem. Recent works [228–
called concentration of parameters or transfer of parame- 232] have generalized recursive quantum approaches to
ters [220, 222–224]. For these problems, the optimal pa- problems beyond MAXCUT and to circuit ansätze be-
rameters become instance independent. In other words, yond QAOA.
one can train the parameters for one instance, and reuse In the same paper that introduces Recursive
it for other instances. Depending on the setting, the in- QAOA [227], the authors also show that, in the large
stances can be of the same size, but also of larger size. problem limit, constant-depth QAOA cannot outperform
This may even allow for training of a smaller instance the classical Goemans-Williamson approximation for cer-
classically and reusing the parameters for larger prob- tain instances of MAXCUT on d-regular graphs, i.e.,
lem instances where classical simulation is not possible the depth needs to grow with the problem size. How-
anymore. Moreover, this could enable good solutions ever, these results are not applicable to alternative ap-
to a problem very quickly by just using pre-optimized proaches, e.g., other QAOA algorithm variants or Recur-
parameters. Sometimes, the opposite direction is also sive QAOA. There are further results about the limita-
possible, i.e., optimal parameters for finite instances are tions of QAOA and requirements on circuit depth with
derived from studying the limit of infinite size instances respect to problem size and structure. For instance, Farhi
[214, 225]. Parameter optimization for QAOA in general, et al. [233] show that QAOA “needs to see the whole
and transfer/concentration of parameters in particular, graph” for a particular QAOA approach to the Maxi-
are crucial to progress towards a successful application mum Independent Set (MIS) problem. More precisely,
of QAOA for practical settings. the authors show that for p < log(n)/ log(d/ ln(2))/2,
Several variants of QAOA have been proposed over where n denotes the number of nodes in the graph and
time. Many of them propose different mixers and ini- d its degree, for large enough d, the approximation ra-
tial states to achieve certain goals. The Quantum Al- tio of QAOA is upper bounded by 0.854. However, for
ternating Operator Ansatz, which shares the acronym larger p, no such limitation has been found. Since the
QAOA, denotes a family of QAOA variants that encode logarithm grows slowly in n, this is not too strong of
constraints into the mixer such that they are preserved a restriction. Similar obstructions to low-depth QAOA
and the algorithm is restricted to feasible states. This have also been shown to apply to k-local generalizations
will be discussed in more depth in Sec. IV A 2. Other of MAXCUT [234].
variants adjust QAOA in order to warm-start it from so- QAOA also inspired the development of classical al-
lutions obtained by classical algorithms. The idea is to gorithms, such as the the Mean Field Approximate Op-
use a classical algorithm to compute an approximate so- timization Algorithm [235]. Here, the QAOA circuit is
lution, and then use it to warm-start QAOA to further approximated through mean-field approximations, which
improve upon the obtained solution [197, 198]. These ap- is demonstrated to perform well for certain problem in-
proaches inherit some of the performance guarantees of stances. This may also be of interest as a classical alter-
the classical algorithms used for warm-starting and have native to train or initialize the QAOA parameters before
been shown to significantly outperform standard QAOA sampling solutions from a quantum computer.
on certain instances. With a modification of the ansatz Since the ansatz for QAOA is derived from the prob-
of QAOA based on the warm-starts, one can still show lem, the structure of the problem directly affects the re-
the convergence to MAXCUT under the adiabatic limit quired qubit connectivity of the quantum circuit. Given
(i.e., when the circuit depth p → ∞). However, showing the often sparse-connectivity of quantum devices, this re-

15
 IV PROBLEM CLASSES & ALGORITHMS

quires a transpilation step to map the problem/circuit with the ground state of H. Then, it applies the evolu-
to the hardware. The transpilation and execution of tion
quantum circuits on noisy hardware will be discussed in
depth in Sec. V. However, addressing this issue at the e−Hτ
|ψτ ⟩ = |ψ0 ⟩ , (5)
problem and modeling level might lead to significant im- Zτ
provements through hardware-optimized problem formu- p
lations. A more detailed review of QAOA can be found where τ = it and Zτ = ⟨ψ0 | e−2Hτ |ψ0 ⟩ denotes a nor-
in [236]. malization factor related to the partition function. QITE
Another family of near-term compatible approxima- exponentially suppresses amplitudes of eigenstates with
tion algorithms for MAXCUT is given by Quantum Ran- larger eigenvalues and amplifies the amplitudes of the
dom Access Optimization (QRAO, also called Quantum ground state. The time T to exceed a fixed overlap
Relaxation Algorithms) [237, 238]. Here, the variable-to- with the ground state, assuming at least an exponen-
qubit ratio is chosen to be larger than one and results tially small initial overlap with the ground state, scales
in a non-diagonal Hamiltonian. This allows the encod- inversely to the spectral gap ∆ of H and linear in the
ing of larger optimization problems into a given number problem size, i.e., T = O(n/∆) [241]. Unlike QAA, the
of qubits. However, depending on the chosen encoding Hamiltonian is not time-dependent, and thus, the gap
and resulting ratio, there might be restrictions on the does not depend on the time and not necessarily on the
allowed quadratic terms in the QUBO, i.e., edges in a problem size. In addition, a small gap might also imply
graph. Assuming a feasible graph, and that we can pre- that there are multiple very good solutions, which – for
pare a good enough approximation of the ground state shorter times – would make it an interesting heuristic.
of the corresponding non-diagonal Hamiltonian, which, However, QITE is a non-unitary evolution and can be
in general, cannot be guaranteed, it has been shown that challenging to implement on a quantum computer [100].
QRAO achieves approximation ratios for MAXCUT of If it were possible, it would allow QMA-hard problems
at least 0.555, 0.625, and 0.722, for variable-to-qubit ra- such as generic ground state search or Gibbs state prepa-
tios of 3, 2, and 1.5 respectively. Furthermore, [238] in- ration to be solved [55]. Thus, only heuristic approxima-
troduce another QRAO with ratio 2 that does not im- tions of QITE exist to the best of our knowledge. These
pose any constraints on the graph, but might not always include projecting the non-unitary evolution to unitary
achieve a non-trivial performance guarantee (i.e., > 1/2). operations [100] or embedding it in a unitary of aug-
Since providing the sufficiently accurate approximation mented dimension [242]. Another family of algorithms
of the ground state of the Hamiltonian can in general are VQAs that map the time evolution to an ansatz by
not be guaranteed, QRAO is strictly speaking not an means of variational principles [240, 243] or directly pro-
approximation, but a heuristic algorithm. There also ex- jecting a Taylor-expansion of the time-evolution operator
ist other encoding schemes that represent more variables onto the ansatz [244, 245]. Variational approaches based
than qubits and provide some variational heuristics for on variational principles are also closely related to alter-
optimization algorithms [239]. Note, that in principle, native optimizers, such as Quantum Natural Gradients
one could encode up to 2n binary variables into n qubits. [246] as well as stochastic approximations [247]. The ad-
However, that would usually imply that writing down the vantage of these heuristics is that they do not require
n n
problem (with a cost matrix in R2 ×2 ) would be sim- the ansatz to be derived from the problem, which makes
ilarly expensive as simulating the involved n-qubit cir- it easier to implement them on real hardware and al-
cuits classically. Thus, only moderate, i.e., polynomial, lows one to apply them also to unconstrained black box
encodings make sense unless the problem is given in a binary optimization [200]. However, at the same time,
compact function form and access to the full solution is that is a potential disadvantage since it is unclear what
not needed, only the objective value. That is a similar circuit structure might work and there is less theoretical
situation, as for instance, in semidefinite programming, foundation on why they may lead to a potential quantum
cf. Sec. IV B 1. advantage. The design of suitable circuits, possibly in an
adaptive way, is a key open question for these type of
While neither QAOA nor QRAO achieve the approx- algorithms.
imation ratios of classical approximations for MAX-
CUT, i.e., 0.878567, they are nonetheless very interesting In addition to variational approximations to QITE,
heuristics. As discussed in Sec. II, there exist problems there are also variational approximations to real-time
with approximation algorithms where the known lower evolution [243, 248]. Those can be used to approxi-
bounds do not yet achieve the inapproximability bounds mate QAA and form another family of heuristics. How-
i.e., where there is room for improvement for classical and ever, like for QITE, the design of the ansatz is a cru-
also quantum approximations. This represents a very cial open question. In addition, all of these variational
interesting direction of research and potential quantum time evolution-based algorithms introduce relatively high
advantage over known classical approximations. costs in terms of the number of circuits to be evaluated
Another time evolution-based heuristic is quantum which may lead to a significant bottleneck to scale them
imaginary time evolution (QITE) [100, 240]. Like QAA, [249] — the cost is increased even further if bounds on
QITE assumes an initial state with a non-zero overlap the preparation accuracy are computed [250].

16
 IV PROBLEM CLASSES & ALGORITHMS

While VQAs are often defined as minimizing an ex- scale where exact simulation is not possible anymore, and
pectation value, in quantum optimization we might not even if approximate simulation might be possible, it does
care about an expectation value but rather about sam- not help to learn how to deal with the noise, and thus,
pling good solutions, which gives us some flexibility in how algorithms are scaling in practice. Developing this
the choice of cost function. Alternative cost functions can intuition is key to progress towards a practical quantum
help to increase the robustness against noise or relax the advantage in optimization. While discrete unconstrained
requirements on an ansatz. A frequently used example is optimization already represents a very broad domain of
the Conditional Value at Risk (CVaR) [251, 252], where optimization, many practically relevant problems natu-
we do not average over the objective values correspond- rally come with constraints, such as budget, capacity, or
ing to all samples obtained from measuring a quantum structural constraints. Thus, in the next section, we dis-
state. Instead, we sort them and only take the aver- cuss different approaches to incorporate constraints into
age over the best α-fraction of samples, for a pre-defined quantum algorithms for discrete optimization.
α ∈ [0, 1], where α = 1 corresponds to the full expecta-
tion value and α = 0 corresponds to the single best ob-
served sample. It has been shown by Barron et al. [252] 2. Constrained Discrete Optimization
that choosing α based on the present noise when evaluat-
ing the corresponding circuit on a noisy quantum device, Constrained Discrete Optimization is defined by
CVaR can provably bound noise-free expectation values. adding constraints of the type g(x) = 0 or g(x) ≤ 0
Another proposal is the Gibbs objective function [178]. to an Unconstrained Discrete Optimization problem. In
The rationale is the same as for the CVaR, it puts more the following, for simplicity, we consider QUBO and add
emphasize on good samples than on bad samples, i.e., it linear inequality and equality constraints, although some
biases the expectation value. of the algorithms could be applied to more complex types
Within this section, multiple possible quantum opti- of constraints as well. The conversions for higher-order
mization algorithms have been discussed. All of them polynomial objective functions to quadratic functions
are limited by the number of available qubits on exist- and from integer to binary variables still apply as dis-
ing respectively near-term hardware. However, if an em- cussed in Sec. IV A 1. Thus, the considered problems are
pirical quantum advantage had been demonstrated for a given by
smaller problem, it might be possible to extend it, e.g.,
using decomposition or multi-level schemes to leverage
the quantum computer also for larger problems [253]. min xT Qx
x ∈ {0, 1}n
Unconstrained discrete optimization is a rich research (6)
domain that has already led to plenty of results. Nev- s.t. Ax = b,
ertheless, there are still many open questions to be an- Cx ≤ d
swered regarding the potential for quantum advantage
for this problem class. Important questions can be found for matrices A ∈ Rme ×n , C ∈ Rmi ×n , and vectors
at every level, from quantum-specific problem formula- b ∈ Rme , d ∈ Rmi , where me , mi denote the number of
tions, to better encodings of integer variables, and more equality and inequality constraints, respectively. Note,
efficient encodings of problems in general [206]. Further- that unlike for QUBOs, linear objective functions (i.e., a
more, a key step in classical solvers is pre-processing diagonal Q) are interesting due to the presence of con-
which can simplify problems tremendously. However, straints.
little is known about the potential of quantum-specific Under mild assumptions on A and C, this can always
pre-processing. In addition, classical algorithms, such as be converted to QUBO by adding slack variables to con-
branch-and-bound, often come with a posteriori bounds vert all inequality constraints to equality constraints and
on the optimality gap. This is crucial because in many then converting equality constraints to penalty terms by
cases this results in almost optimal solutions and also squaring the right-hand-side and adding it to the objec-
helps to determine how much compute resources to in- tive function with a large weight [203]. Some constraints
vest in finding good solutions to a given problem. In can also be encoded into penalty terms more efficiently
addition to a priori or a posteriori performance bounds, (see, e.g., [203]). While this is always possible, it may
it is crucial to understand how they carry over when the come at a substantial cost in terms of additional binary
algorithm is executed on noisy quantum hardware and variables and a dramatic increase in non-zero terms in the
how problems can be optimally modeled and mapped cost matrix in Eq. (6); i.e., it often maps an otherwise
to the hardware to minimize the impact of noise. Be- sparse matrix Q to a dense one. Further, adding penalty
cause these questions will often be difficult to answer, terms scaled by a large constant can tremendously am-
systematic benchmarking becomes even more important, plify the range of values of an objective function which
cf. Sec. VI. To conclude, with the availability of quantum can imply numerical instabilities or requirements to con-
computers with more than one hundred qubits, it is cru- trol certain parameters in the software or hardware stack
cial to develop and test quantum optimization heuristics up to an infeasible accuracy. These are some of the rea-
on these devices for non-trivial problem sizes. This is at a sons why explicitly keeping the structure can lead to more

17
 IV PROBLEM CLASSES & ALGORITHMS

efficient algorithms. In the following, we discuss strate- Hamming weight constraints by XY -model [259, 260] or
gies to incorporate constraints directly. Grover-based [190, 261] Hamiltonians. However, all of
Like for unconstrained discrete optimization, we can this leads to more complex mixers, and implementing
leverage Grover Search to achieve a quadratic speedup them can become challenging. Hardware-efficient ansatz
over brute force search [101]. This requires not only an variants [229, 262–264] are one alternative approach
oracle for the objective function, but also one for each toward alleviating this difficulty.
constraint such that we can mark all feasible solutions. Another approach is to leverage quantum Zeno dynam-
Since brute force search is rarely the most efficient classi- ics to repeatedly project the state back to the feasible
cal algorithm the achieved quantum advantage is usually subspace [265]. This requires auxiliary qubits as well as
sub-quadratic or sometimes not present at all. However, projective mid-circuit measurements. While this allows
Grover Search might still be useful as a subroutine in one to include multiple constraints of different types, the
other algorithms. resulting circuits become deeper and are likely to require
Classically, constrained discrete optimization problems FTQC.
are often solved using branch-and-bound algorithms. In Instead of handling constraints explicitly in the quan-
[254], a quantum branch and bound algorithm is intro- tum circuit, unconstrained black box binary optimization
duced that accelerates classical branch-and-bound algo- [200] allows hiding constraints in the black box objective
rithms. It is shown that it solves most instances of the function. Instead of requiring quadratic penalty terms
Sherrington-Kirkpatrick model to optimality with a high for every constraint, a general metric for infeasibility can
probability in time O(20.226n ). The algorithm adapts be defined, which allows one to add a more balanced
Grover Search to be applicable to search branch-and- penalty term to the objective without the disadvantages
bound trees. While this has the potential to accelerate mentioned earlier in this section. However, as mentioned
classical schemes, it has the disadvantage that no a poste- in Sec. IV A 1, it is unclear how to construct promising
riori bounds on the optimality gap are generated, which parameterized circuits.
often allow an early termination of such search schemes.
Further, it requires FTQC, which adds additional over- Alternatively, Iterative or Recursive QAOA variants,
head. cf. Sec. IV A 1, may be adapted to enforce constraints as
suggested in [230]. Here, as for the unconstrained case,
There are multiple proposals to incorporate constraints
problems are solved iteratively and usually one variable is
in VQAs. The most straight-forward approach is to de-
removed at a time. Adjusting the selection rules for how
sign an ansatz that – in the noise-free case – is restricted
variables are fixed can allow one to enforce constraints in
to the feasible space. For instance, suppose a constraint
certain cases.
that fixes the Hamming weight of feasible bit strings, i.e.,
the number of ones. Then, we can prepare as initial state Adding constraints as penalty terms to the objective
a uniform superposition of all these feasible bit strings, function, i.e., casting constrained binary optimization to
i.e., a Dicke state [255]. Or, we can use a parametric QUBO, is often the main driver for the density of the
iSWAP or a Givens rotation gate, like in a particle pre- cost matrix, and thus, the challenges to implement the
serving ansatz in quantum chemistry [256], to construct algorithms on real hardware. If at least some constraints,
an ansatz that only generates feasible bit strings. This is e.g., cardinality, packing, or covering constraints, could
straight-forward for some type of constraints, but more be natively incorporated into an algorithm, this could
difficult in general. Further, it takes the constraints into lead to sparser problems, and thus, could lead to signifi-
account, but not yet the objective function. cant simplifications for implementations on real quantum
The Quantum Alternating Operator hardware. This is sometimes called QUBO-Plus [266],
Ansatz (QAOA’) [257, 258] generalizes the Quan- i.e., QUBO plus special types of constraints. For classical
tum Approximate Optimization Algorithm to much solvers, these constraints can even simplify the problem,
wider classes of problems and encodings, in particular since the knowledge about the optimal solution can be
to problems with hard constraints. Here, more general leveraged within the algorithm. This is also the idea be-
initial states are considered, such as ones from the sub- hind most of the presented quantum algorithms within
space of feasible states for constrained problems. The this section. However, often the complexity is just pushed
QAOA operators of Eq. (3) are generalized to arbitrary to a different part of the algorithm, for instance, from the
families of phase of parameterized operators, again QAOA cost operator to its mixer.
applied in p alternating layers. In particular, mixers can The approaches for constrained discrete optimization
be constructed as ordered products of non-commuting mentioned in this section essentially leverage three strate-
local operators such that problem hard constraints are gies to handle constraints: they use Grover Search (and
guaranteed to remain satisfied. This has the advantage variants), they add penalty terms to the objective func-
to keep the structure of the objective function and tion, and they try to restrict or project states to the
often leads to significantly cheaper implementations feasible space. Since constraints are appearing almost
of cost operators than when adding constraints as everywhere in practice, further progress on how to effi-
penalty terms. Alternatively, one can also directly ciently handle them is crucial on the path towards quan-
replace the mixing Hamiltonian HX in Eq. (3), e.g., for tum advantage in optimization.

18
 IV PROBLEM CLASSES & ALGORITHMS

B. Continuous Optimization sion ϵ to which one solves the problem. We can categorize
quantum algorithms for these problems based on the so-
1. Convex Optimization lution methodology.
First-order methods. The first quantum algorithms
for SDPs were first order methods, based on the Multi-
Convex optimization problems can often be solved ef- plicative Weights Update (MWU) framework [272]. In
ficiently, both in theory and in practice, and find nu- this framework one uses P candidate solutionsP X that are
merous applications in business, science and engineering. m m
proportional to ρ = exp( i=1 yi Ai )/ tr(exp( i=1 yi Ai ))
Quantum algorithms for convex optimization have been for some (sparse) vector y ∈ R . That is, the candidate
m
proposed for several relevant classes of problems. solutions are proportional to Gibbs states, which nat-
One of the most general formulations of a convex op- urally correspond to trace-normalized positive semidefi-
timization problem is the task of minimizing a convex nite matrices and can be efficiently prepared on a quan-
function f : Rn → R over all x ∈ Rn that have at most tum computer under some conditions. To implement
a certain size, say ∥x∥2 ≤ R for a given bound R, when the MWU framework, one additionally needs to compute
one is provided black-box access to evaluate the function trace inner products tr(Aρ) between such Gibbs states ρ
and possibly its derivatives. For example, given access to and a matrix A (which is either one of the Ai ’s or C),
an oracle that on input x ∈ Rn outputs (f (x), ∇f (x)), and essentially solve a search problem based on these
our task is to find an approximate minimizer of f over values. All these steps can be performed on a quantum
the convex set {x ∈ Rn : ∥x∥2 ≤ R}. Gradient descent computer, with various levels of efficiency; the initial run
is an example of a well known algorithm that works in time bounds derived in the pioneering papers [164, 165]
this framework. A natural question is whether quan- were subsequently improved [166, 167, 273], and further
tum computers offer a speedup, be it in query complexity to [167].
or run time. The answer to this question is not always The run time of quantum MWU methods depends on
positive; see, e.g., [267, 268], which consider black-box several natural parameters; the dimension n, the num-
access to the function and its (sub)gradients. In this ber of constraints m, the desired (additive) error ϵ in
setting, classical and quantum algorithms have the same the objective function; and on two instance-specific pa-
lower bound (no n-dependence, but a polynomial scal- rameters, typically denoted by R and r, that are re-
ing in the inverse of the desired precision ϵ) for the query lated to the diameter of the feasible region for the primal
complexity in general. The other regime, where we do al- (tr(X) ≤ R) and dual problems (∥y∥1 ≤ r), and bound
low a polynomial n-dependence, but only a polylogarith- the so-called “width” of the√oracle. The initial
mic dependence on the desired precision, is equally well

works had
a run time scaling as O e mns2 poly Rr , where s is
studied. Under mild assumptions, one can solve general ϵ
e
convex optimization problems in polynomial time using the row sparsity of the input matrices [164, 165] and O
e.g., the ellipsoid method [269]. In this very general set- ignores logarithmic factors,

later works improved this to
Oe (√m + √n) s2 poly Rr . Throughout, the polyno-
ting one aims, for example, to solve a problem of the ϵ
form min cT x s.t. x ∈ K for some bounded convex set K mial was of a highdegree; the state of the  art in a certain



to which we have various types of oracle access (mem- input model is O e √m + √n Rr s Rr 4 , see [167]. In
ϵ ϵ
bership, separation, optimization). Here mild quantum all these bounds, the parameters r, R, and ε appear to-
speedups (and no-go results) are known [270, 271]. gether as one “scale-invariant” parameter γ = Rr/ϵ [165].
Often much more can be said when we consider con- Often, this parameter γ scales poorly in terms of n and
vex optimization problems with a certain structure. The m [165], especially when parameters r, R scale linearly
most notable such classes are arguably linear program- or superlinearly with n. (For example, in the SDP re-
ming (LP) and semidefinite programming (SDP). Here laxation of MAXCUT [179], cf. (9), tr(X) scales lin-
the objective is to minimize a linear function subject early with n.) Notable exceptions are applications to
to linear inequalities on non-negative vectors (LP) or on shadow tomography, quantum state discrimination, and
positive semidefinite matrices (SDP). An SDP can be ex- E-optimal design (i.e., optimal design in which one max-
pressed in the form imizes the smallest eigenvalue of the information ma-
trix) [166, 167]. As an aside, we mention a second fa-
max tr(CX) mous appearance of the MWU framework in the quan-
X⪰0 (7) tum (complexity) literature: it is a key tool in the proof
s.t. tr(Ai X) ≤ bi ∀i ∈ [m] of the equality QIP = PSPACE [274].
For comparison, the classical complexity of a general-
where the problem is defined by symmetric matrices
purpose
 SDP-solver based on the MWU framework is
C, A1 , . . . , Am , each of size n × n, and a vector b ∈ Rm .
4
7
e
O mns Rr + ns Rr , see [165], but the input
Note that LPs correspond to a special case of SDP in ϵ ϵ
which each of the input matrices is diagonal. Quantum models of the classical and quantum algorithms are not
algorithms for these problems usually trade off an im- necessarily equivalent so these results are difficult to com-
proved dependence on the dimension n and the number pare. Namely, quantum algorithms often rely on the use
of inequalities m with a worse dependence on the preci- of Quantum Random Access Memory (QRAM), which is

19
 IV PROBLEM CLASSES & ALGORITHMS

a RAM addressed by qubits. QRAM accessed by a uni- works trying to address the shortcomings in the initial
tary, swaps the state of the addressed memory cell with proposal [287–289]. Because the direction computed in
the state of a target qubit. Although concrete propos- the Newton step is necessary to process the subsequent it-
als exist [275, 276], a scalable fault-tolerant implementa- eration, the quantum linear system algorithm is followed
tion is yet to be found [277–280], making this at best a by the application of a quantum state tomography al-
long-term prospect. The QRAM model of computation gorithm. Specifically, two issues have emerged: the use
is common in the quantum optimization literature, and it of state tomography introduces inexactness in the search
is also studied in some of the papers mentioned in other direction and a 1/ϵ dependence in the run time, where ϵ
sections: we mention it here because the frameworks dis- is the precision to which tomography is performed [290];
cussed in this section are particularly reliant on it. Also, and the use of the quantum linear systems algorithm in-
in the classical literature the MWU framework was spe- troduces a run time dependence on the condition number
cialized (and significantly accelerated) for the SDP re- of the Newton linear system, which, unfortunately, goes
laxation of structured combinatorial problems such as a to ∞ as we approach optimality. The first issue can be
form of QUBO, balanced separator, and sparsest cut, see dealt with by developing an IPM framework that allows
[19, 272]. for inexact search directions [287], and iterative refine-
So far, we have discussed the MWU framework in the ment techniques can alleviate (but not necessarily elimi-
context of SDPs. One can equally well apply this machin- nate) the remaining issues [288, 289]. Overall, assuming
ery to the simpler class of linear programs, often obtain- access to QRAM, quantum IPMs achieve favorable de-
ing much better run times, see, e.g., the works [175–177] pendence on the size of the problem (m and n) for linear
that are based on a zero-sum game approach [281, 282]. and semidefinite optimization compared to classical algo-
In particular, [176] successfully implemented a dynamic rithms, but the remaining dependence on some numeri-
data structure in the quantum setting. It is an interest- cal parameters (e.g., a condition number bound for the
ing open question to extend such a result to SDPs. One Newton systems) makes it unclear if an overall speedup
could also study further first-order methods [283, e.g.], is achieved.
beyond MWU, to SDPs. While these have been pro- Very recently, a quantum speedup for interior point
totyped [283], and there may be some speed-up in per- methods was achieved without introducing a dependence
iteration complexity, their iteration complexity remains on a condition number [291]. Under some mild assump-
the same as in the classical case. tions and access to QRAM, this IPM achieves a quan-
Second order methods. Interior Point Methods tum speedup for linear programs in which the number
(IPMs) are popular second order methods for solving of linear inequalities m is much larger than the number
structured convex optimization problems both from a of variables n, i.e. LPs of the form max cT x s.t. Ax ≥ b
theoretical and practical perspective. At their core, IPMs where A ∈ Rm×n , b ∈ Rm and c ∈ Rn with m ≫ n.
start from a point lying in the interior of a convex fea- In particular, the best classical solver achieves a scal-
sible set X ⊂ Rn , and use Newton’s method to solve a e
ing of O(mn
√ + n ) [? ], whilst the quantum analogue
2.5
sequence of barrier problems of the form scales as m · poly(n, log(1/ϵ)). The key new subroutine
here is a faster quantum algorithm to compute spectral
min η · ⟨c, x⟩ + f (x), (8) approximations of matrices of the form B T B, provided
x∈Rn
query access to a tall matrix B ∈ Rm×n . A spectral
where ⟨·, ·⟩ is an inner product on Rn , η ≥ 0 and approximation of such matrices suffices to speedup the
f : int (X ) 7→ R is a barrier function encoding the con- costly Newton step. This subroutine can be seen as a
straints defining X . That is, f (x) → ∞ as x approaches generalization of a quantum graph sparsifier (by letting
the boundary of X , and thus the only “constraint” present B be the edge-vertex adjacency matrix) [292], which in
in Eq. (8) is that f is only defined on the interior int (X ) turn has been used in optimization algorithms in both
of X . In each iteration, the value of η is increased, and the continuous [293] and discrete settings (e.g. cut ap-
the set of minimizers {x(η) : η ≥ 0} of Eq. (8) define proximations) [292].
the so-called central path, an analytic curve extending
Quantum SDP-solvers for approximating
throughout the interior of the feasible region to the set
MAXCUT. Given the challenges in evidencing end-
of optimal solutions that is uniquely determined by the
to-end speedup for general SDPs posed by the scale
starting point. By iteratively applying Newton’s method,
invariant error parameter Rr/ϵ, the MWU framework
IPMs approximately follow the central path to the opti-
has been specialized to the semidefinite approximation
mal solution via a sequence of local linearizations (i.e.,
of MAXCUT:
Newton steps). Computing the Newton step requires
one to solve a linear system of equations, and consti- max tr(CX)
tutes the dominant operation at each iterate of an IPM. X⪰0 (9)
It is therefore natural to try to use quantum linear sys- s.t. Xii = 1, ∀i ∈ [n].
tem solvers [140, 284] in an effort to accelerate the com-
putation of the Newton step [285]. Work on this line Brandao et al. [181] demonstrated that, after normaliz-
of research was initiated by [286], with several follow-up ing the diagonal constraints by n and relaxing slightly

20
 IV PROBLEM CLASSES & ALGORITHMS

the constraints, feasibility reduces the task of solving the be viewed as a quantization of the Bregman-Lagrangian
MAXCUT SDP to a possibly simpler task: preparing a framework for continuous-time accelerated gradient de-
Gibbs state that is (i) approximately indistinguishable scent introduced by Wibisono et al. [298]. By properly
from the maximally mixed state when measured in the designing a family of Hamiltonians that, as time pro-
computational basis, and (ii) whose trace inner product gresses, lead to the optimal solution of the dynamical
with C is close to an estimate of the optimal objective system, quantum Hamiltonian descent can be shown to
value, determined via binary search. This enables the de- converge to the optimal solution even on certain non-
sign of specialized oracles for testing feasibility within the convex problems (although it may not do so efficiently,
MWU scheme, and leads to an algorithm that, provided as these problems are provably hard). With regard to
access to QRAM, outperforms previous algorithms in the output, QHD returns a classical description of the solu-
problem dimension and sparsity of C. However, the algo- tion without the need for state tomography. The final
rithm exhibits a poor O(ϵ−28 ) dependence on precision, state is a probability distribution concentrated near the
prohibiting an overall speedup. global minimizer of the objective function, and thus a
The impractical error scaling was addressed by Au- bitstring corresponding to a solution is obtained upon
gustino et al. [174], where iterative refinement techniques sampling from the QHD final state.
are used to exponentially improve the dependence on the Along this line, Augustino et al. [299] describe a fully
inverse precision, for both the quantum and classical al- quantum algorithm for solving linear optimization prob-
gorithms proposed by Brandao et al. [181]. Moreover, lems by quantum-mechanical simulation of the central
it is shown that when C is stored in QRAM, one can path, which the authors call the Quantum Central Path
remove the dependence on the sparsity parameter by us- Method (QCPM). While IPMs and QIPMs approxi-
ing Gibbs sampling techniques based on Quantum Singu- mately track the central path using successive lineariza-
lar Value Transformation [143]. This yields a quantum tions of the perturbed KKT conditions (i.e., Newton

steps), the QCPM consists of a single simulation that
algorithm that runs in time O e n1.5 , plus the time it
takes to read C, obtaining a polynomial speedup over works directly with the nonlinear complementarity equa-
the best known classical algorithm. Due to the lack of tions. This is achieved by designing a Hamiltonian which
a classical lower bound for the problem, however, it is encodes the central path in its ground state. Like QHD,
possible that better (and faster) classical algorithms for no intermediate measurement is required, and a classi-
this problem could be developed. Note that the speedup cal description of the solution is obtained upon sampling
in Augustino et al. [174] is reliant on QRAM: the au- from the final state. This approach is faster than any
thors analyze their algorithm in the standard gate model classical or quantum IPM in a certain sparsity/condition
and show that the resulting algorithm is outperformed number regime, while also avoiding the use of QRAM,
by classical approaches. block-encodings, quantum linear system algorithms, and
Quantum algorithms for matrix scaling and ma- state tomography.
trix balancing. Aside from the SDP relaxation of
MAXCUT, another convex optimization problem that
2. Non-Convex Optimization
has enjoyed a provable quantum speedup for first- and
second order methods is the matrix scaling problem. One
version of the matrix scaling problem can be stated as In non-convex optimization, one minimizes a function:
follows: given an entrywise non-negative matrix, find
positive weights for the rows and columns such that the min f (x) (10)
x ∈ Rn
reweighted matrix becomes doubly stochastic. This prob-
lem has many applications, both in theory (e.g., approx- in dimension n ∈ N. Many variants are undecidable [300],
imating the permanent [294]) and in practice (used by as long as one allows for periodic functions f , or suffi-
default as a preconditioner in LAPACK [295]). Recent ciently non-smooth functions. (As suggested in Tab. I,
works [293, 296] have shown how to speedup both first- some are inapproximable in even in the Banach-Mazur
and second-order methods for a natural convex formula- model [86].) One can hence hope to study k-th order
tion of the problem, using a variety of techniques (basic critical points of non-convex functions, in general. In
amplitude amplification, but also graph sparsification). smooth non-convex optimization problems, this means
Importantly, these quantum algorithms do not introduce points where necessary conditions of optimality involving
a dependence on a condition number. first k partial derivatives are satisfied. In non-smooth
Quantum approaches with no direct classical non-convex optimization problems, this often involves
equivalent. Some classical continuous optimization al- necessary conditions of optimality involving the Clarke
gorithms can be phrased in terms of the solution of a cer- subdifferential [301, 302].
tain dynamical system, i.e., ODEs. These ODEs can then Echoing the situation in convex optimization [267],
be cast as the Schrödinger equation, and solved by means there are clear limits as to the speedup in non-convex
of a quantum algorithm, thereby solving the optimiza- optimization [303]. In particular, when first k partial
tion problem. The first work in this line of research is derivatives are available via an oracle or when one has ac-
Quantum Hamiltonian Descent (QHD) [297], which can cess to stochastic gradients (SG), and no further assump-

21
 IV PROBLEM CLASSES & ALGORITHMS

tions are made, classical lower bounds apply to quantum desired gradient step and we may continue with the
algorithms too. One hence wishes to find suitable special next step. In this setting, Rebentrost et al. [285] con-
cases, such as by considering the numbers of local min- sider certain classes of multivariate polynomials, and in-
ima, local behavior of the functions around those, and cludes Newton steps. While the algorithm may admit
behavior of the trajectories connecting nearby local min- a polylog(n) complexity, amplifying the |0̄⟩ state at the
ima (“the barriers”). end of the gradient √procedure can cost 2T , since at every
Consider the problem of finding an ϵ-stationary point step |α | ≤ |α |/ 2. An efficient circuit U∇ may be
t+1 t

x ∈ Rn of a non-convex function f , which means that constructed only for special cases [285, 306], or could be
approximated via pre-training a parameterized quantum
∥∇f (x)∥ ≤ ϵ. (11) circuit. For affine gradients corresponding to quadratic
optimization problems, Kerenidis and Prakash [305] show
For x ∈ Rn , consider the set of k-order deriva- an efficient implementation using phase estimation.
tives ∇(0,··· ,k) f (x) := {f (x), ∇f (x), · · · ∇k f (x)} with The idea of simulated annealing, finding global min-
universal Lipschitz constants, and consider also the ima assisted by thermal fluctuations, applies to continu-
stochastic gradient g(x, ξ), with ξ a random seed ous non-convex optimization as well. Quantum comput-
and satisfying Eξ [g(x, ξ)] = ∇f (x) and Eξ ∥g(x, ξ) − ers obtain polynomial speedups in executing the classi-
∇f (x)∥ ≤ 1. Assuming quantum access Of |x⟩ |y⟩ =
(k)
cal simulated annealing process via quantum walks [307].
|x⟩ |y ⊕ ∇(0,··· ,k) f (x)⟩, which outputs a binary repre- Quantum fluctuations provide an additional resource al-
sentation of ∇(0,··· ,k) f (x), Zhang and Li [303] prove lowing quantum tunneling through barriers between dif-
the lower bound Ω(ϵ−(k+1)/k ). Assuming quantum ac- ferent minima [308]. Recently, Liu et al. [309] investi-
cess Og |x⟩ |ξ⟩ |y⟩ = |x⟩ |ξ⟩ |y ⊕ g(x, ξ)⟩, produces a lower gated quantum tunneling for SGD, where the SGD pro-
bound of Ω(ϵ−4 ). Both query lower bounds are the same cess is approximated by a continuous-time stochastic dif-
as the classical case. On the other hand, Sidford and ferential equation. Mastropietro et al. [196] investigated
Zhang [304] show query upper bounds for finding critical restarting the search, when one arrives in so-called bar-
points with quantum algorithms for stochastic gradient ren plateau, which allows for speed-up when there are
descent (SGD). With access to a probability-weighted su- many or large barren regions. These quantum walks can
perposition over the bounded-variance stochastic gradi- provide a speedup over classical SGD when the barriers
√ between different local minima are high but thin and the
ent, the algorithm requires O( nϵ−3 ) queries, which is
only better than the classical algorithm O(ϵ−4 ) when the minima are flat.
dimension is < ϵ−2 . This can be improved with√access to There are several promising directions and open ques-
mean-squared smooth SGs, from O(ϵ−3 ) to O( nϵ−5/2 ). tions. What are lower bounds for various settings
The run time for all of these algorithms both classically of quantum polynomial optimization? Beyond lower
and quantumly will in most cases be polynomial in n. bounds, more in-depth focus on possible run time ad-
vantages of quantum algorithms is required. One di-
In high-dimensional situations, the availability of the
rection is about the quantum speedup could one obtain
quantum linear systems algorithm and of quantum linear-
by utilizing moment/sum-of-squares approaches to poly-
algebra methods motivates the direct quantum imple-
nomial optimization and quantum algorithms for SDPs,
mentation of gradient descent over O(polylog(n)) qubits
in analogy to mixed-integer programming discussed be-
[285, 305] to provide descriptions of stationary points
low. Extending to larger classes of polynomials and be-
of non-convex functions. With a non-convex differen-
yond, what classes of non-convex functions would al-
tiable function f (x) : Rn → R and an initial point
low for amplitude-encoded embeddings, efficient quan-
x0 ∈ Rn , consider the gradient descent update x →
tum circuits for gradient computations, and converging
x − η∇f (x) with η > 0. For T ∈ N, the task is to
gradient procedures.
output a quantum state proportional to xT , where the
update is applied T times starting from x0 . To mimic
a classical gradient descent step for fixed Pnt ∈ [T ], de-
fine the log(n)-qubit states |xt ⟩ := ∥x1t ∥2 j=1 xtj |j⟩ and C. Mixed-Integer Programming
Pn
j=1 ∇f (x )j |j⟩. Let us be given a cir-
1
|∇t ⟩ := ∥∇f (x t )∥
t
2
Combining discrete decision variables and continuous
cuit to prepare αt |0⟩1 |xt ⟩ + |ϕt ⟩, where |ϕt ⟩ is some non-
decision variables, Mixed-Integer Programs (MIP) are
normalized orthogonal state and |αt | ≤ 1, and a circuit
non-convex and at least as hard as either of the dis-
U∇ such that U∇ |xt ⟩ = |∇t ⟩. Using an auxiliary qubit
crete optimization (cf. Sec. IV A above) and continuous
and Hadamard gates, the state αt+1 |00⟩12 |xt+1 ⟩+|ϕt+1 ⟩
optimization (cf. Sec. IV B above) problems they sub-
can be prepared, where
sume. We differentiate between Mixed Integer Linear
∥x t
−η∇f (xt )∥2 Programs (MILP), usually used synonymously to MIP,
αt+1 := αt √ √ (12) and its generalization Mixed Integer Quadratic Programs
2 ∥xt ∥22 +∥∇f (xt )∥22
(MIQP). In practice, some instances with unbounded
and |ϕt+1 ⟩ is an orthogonal non-normalized state. Hence, variables are notoriously hard [310]. Having said that,
the |00⟩12 part of the state proportionally encodes the in the case of mixed-integer linear programming, espe-

22
 IV PROBLEM CLASSES & ALGORITHMS

cially when restricted to binary variables, classical solvers be shown to be globally convergent. A similar approach
based on branch-and-bound-and-cut have made an aston- has been tested by Chang et al. [318], in the context of
ishing progress over the past three decades, often solving Benders decomposition.
instances of up to one million binary variables to proven Reformulation to an unconstrained optimiza-
optimality, cf. Fig. 6, and sometimes solving structured tion problem: Braine et al. [319] experimented with a
problems with tens of millions of binary variables [311] decomposition minimizing values of slack variables used
to proven optimality. to convert inequality constraints to equality constraints,
Following Eq. (2), let us consider the following illus- and a Lagrangian relaxation of the resulting equality-
trative MIQP: constrained continuous-valued problem. Another, ap-
proach that optimizes over the Lagrangian function clas-
min xT Qx + y T Ry sically over the circuit parameters and dual variables has
x ∈ {0, 1}nb , y ∈ Rnc been introduced by Le and Kekatos [320]. Such La-
s.t. Ax = a, grangian methods are well understood [321, 322], but
(13) global-convergence results are limited by very restrictive
By = b,
conditions, in general, or require some form of backtrack-
Cx ≤ c, ing. One could also solve the problem approximately
Dy ≤ d by utilizing a test done by Aspman et al. [323] for the
active set remaining fixed subsequently, and only then
where nb , nc ∈ N denote the numbers of binary and con- utilize the Lagrangian relaxation. This could allow for
tinuous variables, respectively, Q ∈ Rnb ×nb and R ∈ easier global convergence proofs without backtracking.
Rnc ×nc are cost matrices, and matrices A ∈ Rme ×nb , It should also be noted that there are multiple options
B ∈ Rme ×nc , C ∈ Rmi ×nb , D ∈ Rmi ×nc , and vectors for working with both the slack variable, penalties for its
a ∈ Rme , b ∈ Rme , c ∈ Rmi , d ∈ Rmi , define 2me equal- non-negativity [324], as well as multiple options for aug-
ity constraints and 2mi inequality constraints. menting the Lagrangian with quadratic or higher-order
There are several potential algorithms to approach terms in order to improve numerical performance. These
such a problem with a quantum computer, whose short are to be explored, yet.
overview is provided in Tab. II. In more detail: Reformulation to high-dimensional LPs: using
Branch-and-Bound-and-Cut: the main workhorse techniques that go back to Dantzig and Wolfe [325], one
in classical optimization. One possibility for a hybrid can reformulate mixed-integer programs as in Eq. (13)
implementation is that discrete decisions are handled as a very large linear program (LP). While quantum al-
by branching and generation of cuts implemented clas- gorithms for LPs have been somewhat overshadowed by
sically, while the continuous, possibly convex relaxations quantum algorithms for SDPs so far (cf. Sec. IV B 1), a
could be solved on the quantum computer. Consider- continuous development of quantum algorithms for LPs
ing the progress in quantum algorithms for convex opti- may render this a viable avenue.
mization (cf. Sec. IV B 1), this seems plausible, although Reformulation to high-dimensional SDPs: using
distant. Perhaps even more ambitiously, one could de- moment/sum of squares (SOS) techniques [326, 327], one
termine the branching decisions on the quantum com- can reformulate mixed-integer programs as in Eq. (13) to
puter as well, which would yield an additional quadratic an SDP in a much higher dimension than n + d. Indeed,
speedup [254, 312–314] over the speedup of quantum al- there are instances where this dimension needs to be ex-
gorithms for convex optimization, while increasing the ponential in n + d [328]. If quantum algorithms for SDPs
requirements on the numbers of qubits and fault tol- (cf. Sec. IV B 1) made it possible to solve such instances
erance substantially. We note that the earlier papers independent of the dimension, it may offer an interest-
[254, 312] consider the speedup in the case of locating all ing avenue for quantum algorithms for mixed-integer pro-
optima (measured against the classical cost of finding all gramming.
optima), whereas Refs. [313, 314] consider the speedup of Reformulation to a completely positive prob-
obtaining one optimum (out of possibly multiple ones), lem: a mathematically elegant reformulation of a mixed-
measured against the classical cost of finding one opti- integer programming problem results in a linear program
mum. Recent research by Dalzell et al. [315] suggests over the dual of the cone of copositive matrices [329]. Re-
that super-quadratic speedups may be possible in some call that copositive matrices are real symmetric matrices
cases. Q ∈ Rn×n , corresponding to quadratic forms xT Qx non-
Decomposition/Splitting/ADMM: this is a large negative on the positive orthant, i.e., for x element-wise
class of algorithms that is based on decomposing the non-negative. For a fixed n, these matrices form a convex
problem. Refs. [316, 317] suggested solving the QUBO cone [330], but even the test of membership in this cone
sub-problem using a quantum computer, while solving is NP-hard [331]. These cones have recently attracted
the continuous sub-problem classically. This approach attention in mathematical optimization [332]. A possi-
is based on the hope that a practical quantum speedup ble way to exploit this reformulation in the context of
for some type of QUBO can be proven. Under some re- a hybrid solver is to use a quantum computer to solve
strictive conditions [317], such an iterative approach can the difficult cut generation problem, whereas the classi-

23
 IV PROBLEM CLASSES & ALGORITHMS

cal computer solves a convex problem. This approach has A dynamic programming algorithm consists of repeat-
been tried in [333], using a QUBO solver to generate cuts. edly optimizing the Bellman equation
We note that to prove optimality one eventually needs to
show that no cut exists, and this in turn requires an exact V (x) = min T (x, a) + V (xa ) (14)
a
solver for the QUBO subproblem. Solving the cut gen-
eration problem exactly can be difficult even for classical where a are the actions that are available in the state x
solvers, see [334] for a discussion of such an approach in and xa = F (x, a) is the state to which we get by apply-
the context of power systems. ing a in the state x. The entries V (x) can be arranged
Next, we review multiple proposals for quantum al- into a dynamic programming table. A dynamic program-
gorithms for MIP and propose a list of new directions. ming algorithm then computes all the entries V (x) in this
There are still many open questions to be answered, how- table. The order in which the entries are computed is
ever, in order to achieve a quantum advantage, which we chosen so that, whenever the algorithm has to compute
express in the remainder of this section. V (x), all V (xa ) have already been computed and V (x)
First, can we provide general statements on provable can be computed from Eq. (14). For example, in the case
quantum advantage for MIP? Is there a general meta- of TSP [22], x are subsets of the set of city and the en-
theorem for combined speedup of solving a convex re- try V (x) is the length of the shortest path through x.
laxation and branch-and-bound? What would be the Thus, the DP algorithm for TSP consists of computing
best possible speedup for branch-and-bound? It seems the shortest paths through every subset of cities.
plausible that it could allow for super-quadratic speedup The complexity of a DP algorithm is O(Sm) where S
for NP-hard problems quite generically, a tantalizing is the number of possible states x for which we have to
prospect. Quartic speedup has been shown, so far, only optimize the Bellman equation and m is the maximum
for learning majorities [335] and the quantum counter- number of actions in one state. In the case of TSP, this
feit coin problem [336]. Further, how could the La- e n ). DP can be used to address many NP-complete
is O(2
grangian approaches benefit from an approximate solu- problems, some of which are listed in Tab. III. For exam-
tion and the test from Aspman et al. [323] for the active ple, the best algorithm with provable running time for
set remaining fixed subsequently? How does one make TSP is based on DP [22]. It is also often used as a sub-
design choices within Lagrangian approaches (penalties routine in approximation algorithms (e.g., some PTAS),
for non-negativity of slack variables, augmenting the and it can be used to solve Markov Decision Problems
Lagrangian with quadratic or higher-order terms) opti- (MDPs), as discussed in more detail in Sec. IV E.
mally? And out of the reformulation approaches sketched Quantum speedups are known for a substantial num-
out above (Dantzig-Wolfe, Moment/SOS, Co(mpletely)- ber of classical DP algorithms. The most important ex-
positive), which ones would allow for the best speedup in amples are shown in Tab. III. The quantum algorithms
the continuous optimization (cf. Sec. IV B above)? This are a combination of classical dynamic programming and
entails bounding the ratio of the inscribed to outscribed Grover’s search. For example, the quantum algorithm for
balls of the reformulations within continuous optimiza- TSP by Ambainis et al. [337] first uses DP to compute
tion. the shortest paths through sets of cities that contain up
Given the practical importance of MIP, it is crucial to 24% of all cities and then uses Grover’s search to find
to advance our understanding of how to approach this the shortest combination of those paths that visits all
problem class with quantum computers. Most of the pro- the cities and returns to the starting point. Similar ideas
posals in the literature consider a quantum sub-routine can be applied to other problems in which the classical
inside a known classical algorithm. However, it is un- algorithm uses an exponentially large DP table indexed
clear what advantage the sub-routine needs to provide by subsets of an n-element set. This includes graph col-
to imply an overall advantage of the optimization algo- oring [338], minimum Steiner tree [339, 341], tree-width
rithm for MIP. This could be a combination of speed, [340] and other problems. The problems amenable to
quality, but also diversity of solutions in order to accel- this approach are typically vertex ordering problems (in
erate branching. Further, there might opportunities to which the task is to find the optimal ordering of vertices
have more quantum-native algorithms. In any case, it is in a graph, as in TSP) or set partitioning problems (in
important to further investigate this problem class and which the task is to find the optimal partition of a set,
systematically benchmark the performance of available as in Set Cover or Graph Coloring). Such problems lead
proposals. to DP tables in the form of Boolean hypercube (2n en-
tries indexed by x ∈ {0, 1}n ) which can often be handled
by computing part of the table and then using Grover’s
D. Dynamic Programming search to find the best solution.
One can formulate a generic DP problem with the
Dynamic Programming (DP) is a very generic mathe- DP table in the form of hypercube and show that, in
matical optimization method that breaks down large op- the generic case, the quantum speedup can be at most
timization problems into smaller ones and combines re- quadratic [337]. It is an open problem whether quan-
sults for the overall problem in a recursive fashion. tum speedups can be obtained for algorithms with a DP

24
 IV PROBLEM CLASSES & ALGORITHMS

Table II. A short history of proposed quantum algorithms for mixed-integer optimization. For approaches without available
guarantees, we write N/A, although this should not be construed as ruling out guarantees.
Algorithm Guarantees Year Ref.
ADMM Asymptotic convergence under restrictive conditions. Speedup of QUBO? 2020 [316, 317]
Benders decomposition N/A 2020 [318]
Branch-and-bound Quadratic speedup + speedup of convex opt. 2020 [254, 313]
Lagrangian methods N/A 2021 [319, 320]
Branch-and-bound Quadratic speedup + speedup of convex opt. 2022 [314]
“Dantzig-Wolfe” Finite convergence, speedup of convex opt. 2023 Here
Moment/SOS Asymptotic convergence, speedup by SDPs 2023 Here
Subadditive duals Asymptotic convergence 2023 Here
Completely-positive N/A 2023 Here

Table III. Most important provable quantum speedups for dynamic programming, where the notation O∗ (f (n)) hides a poly-
nomial factor in n [337].
Problem Quantum Algorithm Best Classical Algorithm Year Ref.
Generic Vertex Ordering Problem O(1.817...n ) O∗ (2n ) 2018 [337]
Traveling Salesperson Problem O(1.728...n ) O∗ (2n ) 2018 [337]
Minimum Vertex Cover O(poly(m, n)1.728...n ) O(nm2n ) 2018 [337]
Graph Coloring O(1.7956...n ) O(2n ) 2023 [338]
Minimum Steiner tree O(1.812...k poly(n)) O(2k poly(n)) 2020 [339]
Treewidth O(1.538...n ) O(1.755...n ) 2022 [340]

table of a different structure. For example, the DP algo- find a control function u with values in a given set U
rithm for the Edit Distance problem [342] (in which one solving the following optimization problem
is given two strings and has to find the smallest number Z T
of symbol insertions/deletions/replacements to transform
min ℓ(x(t), u(t))dt + h(x(T ))
the first string into the the second string) uses a DP ta- u(t) ∈ U 0 (15)
ble that is a two-dimensional array. If the size of the n
array is n × n, the classical DP algorithm takes O(n2 ) s.t. ẋ = f (x, u), x(t) ∈ R , t ∈ [0, T )
steps. Quantumly, no quantum speedup is known but an
e 1.5 ) quantum lower bound is known for the generic for some nice functions ℓ, f, h, e.g., see Ref. [344]. A
Ω(n
globally optimal solution of this problem, the so-called
case [40, 343].
feedback optimal control policy u⋆ (x, t) can be found by
The broad applicability of DP renders it a very in- introducing a value function
teresting approach. However, many open questions re-
main in terms of a potential quantum advantage. What Z T
structures of state space (besides hypercube) allow gen- V (z, t) = min ℓ(x(s), u(s))ds + h(x(T )),
u(t)∈U
eral quantum DP algorithms? Can we add stochastic- t

ity to quantum DP algorithms or extend them to MDPs ẋ = f (x, u), x(t) = z . (16)
(cf. Sec. IV E)? What about the explosion of state space?
Can we prove better lower bounds for quantum speedups
of generic DP algorithms? In addition to provable al- Generalizing classical dynamic programming ideas one
gorithms, can we find good quantum approximations or can demonstrate that V solves the following Hamilton-
heuristics, such as reinforcement learning? Answering Jacobi-Bellman (HJB) equation backwards in time:
any of these questions will help to further understand
∂V 
the potential and practical applicability of quantum DP + min ℓ(x, u) + f ⊤ (x, u)∇x V (x, t) = 0 , (17)
algorithms and some of them will also be discussed in the ∂t u∈U

following Sec. IV E.
subject to the terminal condition V (x, T ) = h(x), for all
x ∈ Rn . Given the solution of Eq. (17) one computes the
optimal control policy in feedback form as follows:
E. Optimal Control 
u∗ (t, x) = argmin ℓ(x, u) + f ⊤ (x, u)∇x V (x, t) . (18)
u∈U
Optimization with differential equations as constraints
is a subject of optimal control theory. A standard finite It should be noted that in many practical situations, ∇x V
horizon optimal control problem takes the following form: does not exist in the classical sense. For this reason the

25
 IV PROBLEM CLASSES & ALGORITHMS

solution of HJB Eq. (17) is understood in the viscosity called contextual multi-armed bandit and in this case
sense: a viscous term represented by ε∆V , ε > 0 is intro- finding V amounts to solving the optimization problem
duced in the right-hand-side of Eq. (17) so that the result- minu∈U ℓ(x, u) for each x.
ing solution Vε becomes differentiable, and then taking The key problem associated with HJB equation, con-
the limit of Vε as ε ↓ 0 one gets the viscosity solution of tinuous or discrete, is the curse of dimensionality. Solv-
Eq. (17). This construction relates Vε to the stochastic ing HJB in Eq. (17) by conventional methods of numer-
optimal control [344] as the Laplacian ε∆V can be seen ical analysis quickly becomes intractable even in rather
as
√ adding "white noise" to the dynamics of x of variance modest dimensions, e.g., if x(t) ∈ Rn for n ≥ 10. In other
ε. words, direct approach to solving HJB equation has slim
chances of succeeding. In what follows we discuss po-
In addition to optimal feedback control design HJB
tential quantum optimization approaches to find global
has many applications: for example, it allows one to
solutions to the three aforementioned problems:
solve the nonlinear filtering / state estimation problem
by constructing reachability sets of nonlinear dynamical 1. Contextual multi-armed bandit problems.
systems [345]. The most relevant applications of HJB
equations in the context of this paper are in discrete op- 2. Markov decision processes (MDPs).
timization: HJB equation for discrete-time dynamics can
be seen as the famous Bellman equation of dynamic pro- 3. Optimization problems constrained by differential
gramming: equations.
Finding ways of solving the above problems will in turn
V (t, x) = min{ℓ(x, u) + V (t + 1, x + f (x, u))} (19)
u∈U suggest ways of solving the corresponding HJB equa-
tions, continuous or discrete. In fact, solving the dis-
with final time condition V (T, x) = h(x). Here time t is crete HJB in Eq. (19) sheds some light back on the so-
discrete, and position x(t) is assumed to take only finitely lution of the continuous one and vice-versa, suggesting
many values (a so-called piece-wise constant approxima- that algorithms developed in the discrete setting can help
tion), so x(t) can be seen as an index taking values in construct solutions of continuous optimization problems.
the set of integers X [344] and hence V (t, x) can be in- Potential algorithms include:
terpreted as a stack of dynamic programming tableaus
indexed by (t, x) (cf. Sec. IV D). 1. Value iteration and policy iteration boil down to
optimization problems, wherein the convex opti-
Yet another important application is in optimiza- mization of Sec. IV B 1 could be applicable, perhaps
tion for discrete stochastic processes: as was mentioned combined with the quadratic speedup of approaches
above, continuous time viscosity solution V of Eq. (17) is of Sec. IV D.
connected to stochastic optimal control for Markov dif-
fusion processes via regularization. Hence it is not sur- 2. Reinforcement learning is a broad family of ap-
prising that the discrete-time HJB in Eq. (19) could be proaches, rather than a single algorithm, which
connected to discrete stochastic processes too: namely, approximate value or policy iteration, for instance
if instead of deterministic transitions x(t + 1) = x(t) + using Least-Squares Temporal-Difference (LSTD).
f (x(t), u(t)), which are singular in terms of probability Again, the convex optimization of Sec. IV B 1 could
distributions, one introduces Markov Decision Process be applicable, perhaps together with the tabular
(MDPs) [346], which jumps from the state x(t) = x to method of Sec. IV D providing additional speedup.
state x′ = x(t+1) with known probability P (x′ |x, u) pro-
vided the action u was applied, then Eq. (19) transforms 3. Quantum reinforcement learning (heuristics). In
into the famous backwards induction for episodic MDPs: the unitary oracle setting, from the perspective of
( ) exact methods, that is, algorithms that are guaran-
X teed to identify optimal policies in arbitrary envi-
′ ′
V (t, x) = min ℓ(x, u) + P (x |x, u)V (t + 1, x )) , ronment, tight quantum upper and lower bounds
u∈U
x′ ∈X have been identified. In Refs. [349, 350], lower
(20) bounds on the more special task of Multi-armed
bandit problems were found, along with essentially
with a final time condition V (T, x) = h(x), [346]. MDPs matching quantum upper bounds, for a number of
are central object of the Reinforcement learning (RL), RL-related tasks, including finding the optimal q-
a family of algorithms in which an agent aims to learn function. These works achieve a quadratic speedup
optimal decisions in unknown environments through the over best possible classical methods, assuming ap-
experience of taking actions and observing the rewards propriate oracle access to the task environment.
gained. In some cases, the environment is not influ-
enced by the actions of the RL agent, in which case We can break the above into two main classes of ap-
P is independent of action u then MDP turns into the proaches, differing in whether the learning agent has a
uncontrolled MDP [347, 348], and if the transition ma- standard classical access to the environment, or if the in-
trix P (x′ |x) is of rank 1 such an MDP becomes the so- teraction itself is quantum. In the latter line, numerous

26
 IV PROBLEM CLASSES & ALGORITHMS

types of quantum oracular access to the task environment lems. These problems are concerned with timing an ac-
have been explored, e.g., where coherent access to a uni- tion for maximizing a reward and are often solved with
tary encoding the transition function is possible [351], or dynamic programming. Optimal stopping appears, for
where the environment closely mimics classical RL set- example, in sequential parameter estimation [362], se-
tings, and allows only actions as inputs and has memory quential hypothesis testing [363], and, most famously, in
[352]. financial derivative pricing [364]. The pricing of Ameri-
We note that the algorithms listed above have signifi- can options involves stochastic modeling of asset prices
cant limitations, including the fact that they require fault and performing a combination of Monte-Carlo estima-
tolerance, and that quadratic speedups may not translate tion, least-squares regression, and dynamic programming
to wall-clock speedups in early fault tolerant quantum for the optimal stopping time [364]. Given quantum ac-
computers. The main conceptual limitation is the re- cess to a sampler for the stochastic process, a quantum
quirement to a coherent access to the task environment, algorithm with a polynomial speedup using Quantum
which could only be possible in very restricted or con- Amplitude Estimation (QAE) [169, 365] was shown by
structed cases, and the fact they deal with exact solu- Doriguello et al. [366]. Key questions are whether we can
tions, and in real-world problems, the search spaces are extend such stopping time algorithms to other problem
simply too large, such as the state space of all 1 megapixel classes and to problems with more decisions, and whether
images. The up-side of the above results is that the we can make the algorithms more near-term quantum
speedups concern the number of interactions with the friendly.
environment, and are not contingent on any assumptions In the following, we outline open questions and promis-
in complexity theory. ing directions to further advance our understanding on
In practice, one often foregoes optimal policies for op- how quantum computers may provide advantages for
tima within a parameterized policy family, as is found optimal control and related problems. In the era of
by policy iteration, and policy gradient algorithms. Al- noisy quantum computers, convex optimization algo-
though the method cannot find an optimal policy in gen- rithms may be hard to use directly. Nevertheless, one
eral, one can still be interested in finding optima with may considering sketching the value iteration [367] or
respect to the function approximation family available. (LSTD) [368], wherein both the random projections and
In Jerbi et al. [353] it was shown that quadratic speedups the convex optimization could be implemented quan-
in policy optimization were possible as long as the poli- tumly. What MDPs could have a bound on the radius
cies satisfy certain regularity conditions, which are, e.g., of the ball outscribed to the (primal, dual) feasible set?
satisfied for policies defined by parameterized quantum Such MDPs could benefit from speedup of quantum al-
circuits. The lower bounds for this problem have been gorithms based on MWU, cf. Sec. IV B 1. Approximat-
established for a more general setting by Cornelissen et ing the value function, discrete or continuous, using the
al. [354] and they show that these quadratic improve- concept of inductive bias recently introduced in the con-
ments cannot be improved further. However, if the envi- text of function approximation by means of quantum ker-
ronments are not arbitrary, but special, more significant nels [369]. Basic idea is to embed a prior knowledge of
improvements are possible, including provable exponen- certain properties of the value function into the design of
tial speedups [355]. The contrived nature of these envi- the approximator: for example, a rather efficient deep
ronments makes these results unlikely to have implica- learning approximation of value function implemented
tions in practice. in AlphaGo, or an assumed linear approximation of the
However, the second research line investigates quan- value function implemented in LinUCB [370]. In this di-
tum enhancements in reinforcement learning when the rection one would benefit from maintaining connections
access to the task environment is restricted to be clas- between discrete and continuous HJB equations for the
sical. Here too speedups are provably possible, at least optimization problem at hand in order to extract induc-
in special cases, but now subject to standard complex- tive bias, e.g. piece-vise smoothness structure of continu-
ity assumptions. For example in Refs. [356, 357] it was ous value function might help in designing a kernel-based
proven that there exist task environments with a super- function approximation.
polynomial advantage for quantum reinforcement learn-
ing algorithms (with respect to finding optimal policies
and optimal q-functions), unless the discrete logarithm F. Robust Optimization
problem has an efficient classical solution. Again, in re-
ality, we are most often not interested in finding prov- Robust optimization addresses the issue of data inaccu-
ably optimal policies, but rather more effective heuristic racy in optimization and allows one to model uncertainty
solutions which perform well empirically, as is done in in the constraints. The convex variant was introduced by
the famous examples of AlphaGo [358] and AlphaStar Ben-Tal and Nemirovski [371] and lead to many follow-
[359]. In this domain, there has been substantial effort up works, for example Refs. [372–374]. The main feature
in defining various types of parameterized-circuit based is a set of parameterized constraints and the requirement
QRL algorithms [356, 360, 361]. that these constraints hold for every choice of parame-
Closely related to MDPs are optimal stopping prob- ters chosen from some uncertainty set. The problem is

27
 V EXECUTION ON NOISY DIGITAL HARDWARE AT SCALE

formally defined as or have fixed weights, which allows one to convert MOO
into a (series of) single objective optimization problems.
min f0 (x) If the full Pareto frontier is required, the number of objec-
x∈D (21) tive functions plays a crucial role. For a small number of
s.t. fi (x, ui ) ≤ 0, ∀ui ∈ U, i ∈ [m], objective functions, there exist efficient strategies, assum-
ing the underlying single objective optimization problems
where f0 , f1 , . . . , fm are convex functions in the parame- can be solved reasonably well. This is usually achieved
ter x and f1 , . . . , fm are concave in the noise parameter by constructing a new objective function as a weighted
u. Moreover, D ⊆ Rn and the uncertainty set U ⊆ Rd sum of the individual objective functions and adjusting
are both convex. Despite many advances, robust opti- the weights based on the previous solutions. Alterna-
mization for large-scale problems comes with a signifi- tively, a single objective can be optimized while the val-
cant computational overhead. Ben-Tal et al. [375] devel- ues of the other objectives are controlled via additional
oped meta-algorithms to approximately solve the robust constraints. While the former might sometimes be easier
version of a given optimization problem using an oracle to implement, it only finds the convex hull of the effi-
for the solving of the original optimization problem. A cient frontier. In contrast, the latter can find also Pareto
quantum version was developed by Lim et al. [376], us- efficient solutions that do not lie on the convex hull. Nei-
ing quantum access to the functions, extending the origi- ther of these strategies is efficient for growing numbers of
nal algorithm to include a stochastic gradient, and using objective functions. If the number of objective functions
quantum sampling subroutines with polynomial advan- increases, MOO problems quickly become very hard.
tage.
In contrast to single-objective optimization, there is
no approximation ratio that can be used to conveniently
quantify the performance of a given solution candidate.
G. Multi-Objective Optimization However, there are multiple other metrics that allow one
to compare different approximations to the efficient fron-
The previous classes were all considering a single ob- tier. The most commonly used ones are the Hyper Vol-
jective function, possibly with constraints. However, in ume (HV) and the Generational Distance (GD) [380].
practice, decision makers often have to find trade-offs be- The HV measures the volume spanned by the found so-
tween different objectives. This is addressed by Multi- lutions, which is measured as the total volume covered
Objective Optimization (MOO) [377]. MOO is defined by all boxes defined by a reference point upper bounding
as (or lower bounding in case of maximization) all objective
functions, and the solution candidates. The GD, in con-
min (f1 (x), . . . , fm (x)), (22) trast, measures the distance between a set of candidates
x∈X and the optimal Pareto front or a reference solution.
where X denotes the feasible set and fi : X → R, There is only very little literature considering quantum
i = 1, . . . , m, denote the multiple objective functions, for optimization algorithms for MOO. Most of the proposals
instance. In this setting, we first need to define what we apply classical strategies and use quantum optimization
consider an optimal solution. This is done using the con- sub-routines for the resulting single objective optimiza-
cept of Pareto optimality. A solution x is called Pareto tion problems [381] or sample from quantum states to
optimal for a problem with multiple objective functions, approximate the efficient frontier [382]. Thus, quantum-
if there exists one objective function fi , that cannot be native algorithms for MOO present a very interesting do-
decreased any further without increasing at least one main for future research.
other objective function fj . The set of all Pareto op- Within this section, we introduced the most important
timal solutions is called the Pareto frontier or efficient classes of optimization problems, reviewed existing quan-
frontier. tum algorithms to approach them and key open questions
It is well known that the cardinality of the Pareto set to be answered to further advance our understanding. In
can be substantial [378], and thus, one would like to the following, we discuss how to execute them on noisy
obtain a PTAS. For certain bi-objective problems, such quantum devices.
as the shortest paths problem, Diakonikolas and Yan-
nakakis [379] have shown that computing a set of so-
lutions that approximates the Pareto curve within ϵ is V. EXECUTION ON NOISY DIGITAL
NP-hard, if the set is to be smaller than twice the opti- HARDWARE AT SCALE
mum number of solutions, while it is possible to obtain
the 2-approximation in polynomial time. We now discuss best practices to execute quantum
Depending on the application, there are different goals optimization on noisy quantum hardware. There are
in MOO. Either, the goal is to determine or approximate many different platforms to consider for optimization
the complete Pareto front to be able to analyze possible tasks [383–389], here, we focus on superconducting
trade-offs. Alternatively, a single Pareto-optimal solution qubits [390]. More precisely, we focus on universal
is sought. For instance, objectives might be prioritized gate-based quantum computing platforms accessible via

28
 V EXECUTION ON NOISY DIGITAL HARDWARE AT SCALE

the cloud, where quantum algorithms researchers can

directly test their algorithms. While there are addi-
tional algorithmic challenges to scale to large problems,
cf. Sec. IV, the goal of this section is to discuss how to
achieve best possible results from available quantum de-
vices.
Superconducting qubits come in many forms, such as
frequency-tunable qubits [391] and fixed-frequency qubits
with static [392] or tunable coupling elements [393, 394].
Over the past years, superconducting qubits have greatly
improved. The number of qubits has increased with sys-
tem sizes featuring up to 433 qubits. Coherence times
have breached the 100 µs barrier with recent devices ex-
hibiting median T1 and T2 times approaching 300 µs and
150 µs, respectively. It should be noted that these fig-
ures are based on IBM Quantum Eagle devices with 127
qubits [395]. Importantly, the qubits are arranged in pla-
nar lattices where each qubit is only connected to its im-
mediate neighbors.
We review each stage of the execution pipeline with Figure 5. High-level summary of the quantum stack. (a) A
quantum algorithm, including an encoding of decision vari-
scalability and best practices in mind. In Sec. V A, we
ables in qubits and any hardware-guided problem simplifica-
describe how a quantum algorithm is translated down tions, results in a high-level quantum circuit. (b) The instruc-
the quantum stack and executed on hardware. We dis- tions in this circuit are then transpiled to hardware native
cuss the impact of modeling, quantum algorithm choice instructions. (c) Error mitigation methods, such as PEC, are
and qubit encoding in Sec. V A 1. Next, Sec. V A 2 dis- often encoded as circuit instructions. (d) At the pulse-level,
cusses transpilation and Sec. V A 3 reviews error mitiga- circuit instructions are represented by the physical pulses that
tion methods in the context of optimization. Sec. V A 4 are played on the qubits. (e) Finally, the circuit is compiled
touches on the pulse-level and the opportunities it into machine-executable waveforms, (f) that are run on the
presents. In Sec. V B we discuss hardware benchmarks quantum processor.
and then briefly discuss an execution example in Sec. V C
before summarizing in Sec. V D.
1. Modeling, quantum algorithms, and encoding

A. Quantum Stack At the top of the stack, an algorithm is described by a

high-level circuit which may not be explicit in the num-
Hardware constraints and noise limit the scalability ber of qubits used and only provides an abstract rep-
and performance of quantum optimization algorithms. resentation of the quantum gates to run. For exam-
An analysis of the quantum stack, the stages shown in ple, Fig. 5(a) shows a high-level representation of the
Fig. 5 through which an algorithm evolves from its ab- QAOA. To arrive at such a circuit one must first ap-
stract design to its tangible execution on a quantum pro- propriately model the optimization problem to solve and
cessor, offers insight into where scalability and perfor- correspondingly choose the quantum algorithm. For in-
mance bottlenecks arise. stance, linear equality constraints can be included in a
At the top of the stack is an abstract expression of QUBO by adding them as quadratic penalty terms which
the quantum algorithm. It is usually free from any hard- then requires appropriately sizing the corresponding mul-
ware constraints and expressed as mathematical opera- tiplicative factor, see Sec. IV A 2. This modeling step also
tions and logical constructs that embody the problem- presents opportunities to consider the impact of choices
solving approach. The transpilation process then maps on the lower levels of the quantum stack. For example,
the abstract operations to the machine-executable in- due to hardware noise a problem model that results in
structions of the chosen target hardware resulting in a more qubits with less connectivity could lower the im-
hardware-native circuit. Here, circuit instructions may pact of noise compared to a model with fewer densely-
be inserted or modified as part of an error mitigation connected variables. Alternatively, Dupont et al. [229]
scheme. The pulse-level compiler converts the circuit in- project a fully-connected QUBO onto the sparse connec-
structions to finely-tuned electromagnetic pulses. These tivity of a quantum computer. Looking forward, addi-
pulses manipulate the quantum states to enact the cho- tional research on quantum tailored modeling of opti-
sen algorithm. Therefore, as the algorithm is lowered mization problems is needed to find the right balance
through the layers of the quantum stack its representa- between model approximations and noise susceptibility.
tion changes. Each level offers optimization opportuni- Since the hardware is noisy, algorithms that efficiently
ties to extract the most out of the noisy hardware. use the quantum resources are required. For exam-

29
 V EXECUTION ON NOISY DIGITAL HARDWARE AT SCALE

ple, a warm-start reduces the number of QAOA lay- candidates.

ers [197, 198]. In addition, the QAOA mixer can be Research for quantum optimization algorithms running
changed to include constraints of the optimization prob- on noisy hardware must carefully consider the hardware
lem [258, 396], e.g., for equality constraints [259, 260]. Al- constraints. Crucially, many of these methods are heuris-
tering the mixer improves the approximation ratio com- tic and can thus only be properly researched at scale on
pared to penalty term constraints (soft constraints) for actual quantum hardware.
constrained optimization problems [397]. Furthermore,
counter diabatic terms may also improve the approxi-
mation ratio [398]. While these mixers may keep the 2. Transpilation
structure of the objective function and lead to cheaper
implementations of the cost operators, they may increase
A transpiler transforms a quantum circuit into an-
mixer circuit depth thereby limiting their near-term util-
other quantum circuit, typically built from instructions
ity [397, 399]. Crucially, from a hardware perspective, the
native to the target quantum hardware. Such low-level
best practice is to choose the methods that limit low-level
circuits can be expressed in domain specific languages
circuit depth and preserve solution quality. For exam-
which include OpenQASM [410], Quil [411], or Black-
ple, adaptive bias-fields [400] and warm-starts [197] only
bird for continuous-variable quantum computers [412].
require additional single-qubit rotations in the QAOA
Since each quantum hardware bears unique character-
mixer which are implemented on hardware with a negli-
istics, tailoring the transpiler to individual architec-
gible overhead. Furthermore, it may be useful to research
tures is fundamental to maximize operational efficacy.
methods that modify the original optimization problem
Furthermore, the design and optimization of hardware-
to facilitate its mapping to the hardware. For example,
executable quantum circuits must include all constraints
in QAOA one may investigate techniques that modify the
of the target hardware such as the native gate-set and
problem Hamiltonian H to produce an H ′ approximat-
the qubit connectivity. On noisy hardware the transpiler
ing H which is however easier to implement on hardware,
′ must also account for properties such as gate error-rates,
i.e., e−iγH has a lower circuit depth than e−iγH .
cross-talk, and qubit properties, e.g., T1 and T2 times.
While both scale and quality are improving over time, The transpiler has a large impact on the gate count and,
quality is typically the limiting factor rather than scale thus, the performance of the algorithm when run on the
for optimization [401–403]. Nevertheless, for many prac- hardware. In general, the transpiler is responsible for (i)
tically relevant optimization problems the scale of cur- selecting the qubits, (ii) decomposing the high-level cir-
rent Quantum Processing Units (QPUs) is likely insuf- cuit into hardware native instructions and (iii) routing
ficient to reach a classically hard regime. It may there- the qubits to overcome any limited qubit connectivity.
fore be necessary to consider denser encoding schemes Step (i) depends on the properties of the hardware and
that encode multiple decision variables in a qubit. The the problem being solved. For example, if we do not
Holevo bound states that n qubits are required to faith- need all the qubits on the processor then we may se-
fully transmit n bits of information [404]. However, en- lect the best ones according to a metric [413], such as
coding schemes such as Quantum Random Access Codes gate fidelity [402], cross-talk [414], or layer fidelity [415].
(QRAC) can encode m bits of information in n qubits This metric should be based on the properties of the ex-
where m > n. The cost is that each bit of information pected circuit. E.g., if the expected circuit is predomi-
can only be retrieved using Projective Operator Valued nantly composed of two-qubit gates then selecting qubits
Measures (POVMs) with some probability p > 1/2 [405– based on their gate fidelity is reasonable [402]. However,
408]. There are many QRACs that, rather than encoding if the algorithm requires many mid-circuit measurements
the two states of a single bit into the |0⟩ and |1⟩ states then measurement quality should also impact qubit selec-
of a qubit, encode more information by utilizing more of tion [416]. Furthermore, in approximate optimization a
the possible states on the Bloch sphere. For example, decision variable is often encoded in a qubit. The choice
the (2, 1, 0.85) QRAC encodes n = 2 bits into a single of which decision variable is mapped to which qubit sig-
qubit (m = 1) with the 4 possible states of 2 bits en- nificantly impacts circuit depth [417].
coded in states that form the corners of a square on the In step (ii), the possibly high-level gate descriptions
Bloch sphere. Here, each bit is recovered with proba- are synthesized into hardware native instructions [418–
bility p ≥ 0.85. For all QRACs, there is a trade-off 420]. To decompose arbitrary unitaries, a quantum com-
between the ratio of classical to quantum bits and the puter must support a universal set of basis gates [421],
retrieval probability. The more bits encoded into a given which impacts the optimal synthesis technique. These
number of qubits, the lower the probability p of being compilation methods are either deterministic or approx-
able to retrieve a given bit and the more complicated the imate. For example, any exponential of Pauli operators
code becomes. As discussed in Sec. IV A 1, this allows can be implemented using a local 2-qubit gates on the
the hardware to tackle problems with more decision vari- involved qubits with a Pauli-Z rotation and single-qubit
ables than it has qubits [180, 237, 238, 409]. However, Cliffords [422], as exemplified in Fig. 5(b). However, due
denser encodings tend to have a lower approximation ra- to hardware noise it might be beneficial to reduce the
tio and introduce additional overhead to readout solution circuit depth by allowing for an approximate compila-

30
 V EXECUTION ON NOISY DIGITAL HARDWARE AT SCALE

tion [423, 424]. processing of quantum circuits and measured results to

Finally, the routing step (iii) must overcome any lim- reduce the impact of hardware noise. While fully elimi-
ited qubit connectivity of the target QPU. Here, general- nating the error comes at an exponential classical over-
purpose heuristic algorithms exist to route arbitrary cir- head, these methods extend the reach of current devices
cuits [425–427]. However, certain algorithms may have a even at a limited classical cost [14, 438, 439]. This is in
predetermined structure. For example, QAOA results in contrast with error correction, which aims to correct er-
circuits with blocks of commuting two-qubit gates. Ex- rors by encoding logical qubits in vast arrays of physical
ploiting this structure results in shallower quantum cir- qubits [440]. Typically, most error mitigation methods,
cuits [401, 417, 428–430]. More precisely, if the topol- such as Probabilistic Error Cancellation (PEC) [15] and
ogy of the problem graph does not perfectly match the Zero-Noise Extrapolation (ZNE) [16, 438, 441], produce
coupling map of the QPU, SWAP gates are necessary noise-mitigated expectation values.
to enable two qubit gates between unconnected qubits. PEC learns the noise channel Λ of layers of Pauli-
The denser the graph, the more SWAP gates are neces- twirled gates [15]. The inverse of the noise channel Λ−1
sary which results in a deeper circuit [401]. Neverthe- is usually not physical, but can be implemented by a
less, limited qubit connectivity remains an issue and the quasi-probability decomposition. PEC thus produces an
problems that scale best on noisy superconducting de- unbiased estimate of a noiseless observable O. The cost
vices are therefore likely to be those that are not overly is an increase by a factor of γ 2 in the measured variance
dense [385, 402]. A key challenge is thus to identify prob- of O. Here, γ > 1 depends on the strength of the terms
lem instances that are not dense but are classically hard. in the learned noise model and is a measure of the fi-
Modern hardware also exposes new circuit instructions delity with which the quantum circuit can be executed.
such as dynamic circuits [431]. In a dynamic circuit some The noise strength γ is also closely related to the Layer
of the qubits are measured and the outcome of these Fidelity (LF) and the Error Per Layered Gate (EPLG),
measurements changes the following circuit instructions two related measures of noise in large-scale quantum de-
within the coherence time of the qubits. For instance, vices [415]. If γ̄ is the average γ per qubit then a circuit
this offers additional possibilities to address limited qubit with n qubits and d layers will have an error mitigation
connectivity [432]. As an example, the SWAP based im- cost that scales as γ̄ 2nd .
plementation of the QAOA problem graph may be re- In ZNE multiple logically equivalent copies of a quan-
placed on grid-like QPU coupling maps by constant depth tum circuit are executed at different noise levels to com-
circuits with a quadratic auxiliary qubit overhead [433]. pute an expectation value. From the noisy expecta-
Additional research is needed to see how to leverage dy- tion values one extrapolates to the zero-noise limit to
namic circuits to reduce the circuit depth and width. obtain a (usually biased) error mitigated expectation
Crucially, not all types of hardware face the same chal- value. The first proposals implemented ZNE by stretch-
lenges. For example, trapped ion devices typically have ing pulses [438], requiring a large calibration overhead,
all-to-all connectivity which permits denser graphs to be or by folding gates, resulting in large stretch factors.
studied without increasing circuit depth [397]. However, Methods such as partial-gate folding [442] or stretching
at the time of writing trapped ions have not reached the of cross-resonance pulses [443] alleviates some of these
same scale as superconducting devices meaning problems issues. Crucially, ZNE does not require more circuits as
sizes are limited. the size of the problem is increased. However, the cir-
cuit depth limits the range of stretch factors which may
render the extrapolation unusable. PEC is unfeasible in
3. Error suppression and mitigation large scale-experiments since they have a very large γ.
For instance, the γ 2 of 60 layers of CNOTs on 127 qubits
After the transpilation, the quantum circuit may be was 10128 in Kim et al. [16]. This prompted the develop-
modified to include additional instruction for error sup- ment of PEA [16], a probabilistic ZNE, in which the noise
pression and mitigation, as shown in Fig. 5(c). Error is learned like in PEC but then amplified (instead of can-
suppression methods reduce the noise in quantum circuits celed) to then extrapolate to the zero-noise limit. This
without having to gather more shots. Most prominently, may also reduce the bias in the extrapolation. Impor-
Dynamical Decoupling (DD) suppresses non-Markovian tantly, circuits designed with regular networks of SWAP
errors by adding pulses in idle times of a quantum gates alternate a small number (e.g., two) of identical
circuit [434–436]. This can improve circuit execution layers of CNOT gates [401]. This minimizes the noise-
quality without introducing additional overhead. In learning overhead of PEC and PEA.
some cases, such as QAOA transpiled with SWAP net- Error mitigation methods that act on expectation val-
works [402], the resulting quantum circuits may be dense ues help train variational parameters [402, 438]. Cru-
which leaves little space for DD pulse sequences. By con- cially, to find good solutions, quantum (approximate)
trast, DD has a large impact when the circuit has many optimization requires sampling from the optimized cir-
idle times, e.g., in hardware native circuits on a heavy- cuit. Indeed, the task is to find a good sample x that,
hex lattice [437]. e.g., minimizes a cost-function f (x). Therefore, devel-
Error mitigation techniques use pre- and post- oping new error mitigation methods to produce error-

31
 V EXECUTION ON NOISY DIGITAL HARDWARE AT SCALE

mitigated samples is crucial for the success of quantum hardware to achieve more accurate estimations of observ-
optimization. ables within a certain time or to pay the cost of error
mitigation which often requires more shots. Therefore,
Circuit Layer Operations Per Second (CLOPS) comple-
4. Pulse-level and compiler ment quality metrics. The CLOPS metric is inspired
from the floating point operations per second metric used
to benchmark classical computers. In essence, CLOPS
Ultimately quantum instructions are implemented by quantifies the throughput of a device by measuring the
physical pulses, scheduled in time, to manipulate the number of gate layers that can be executed within a cer-
quantum states [444], as exemplified in Fig. 5(d). The tain time. This measure includes the classical overhead of
pulse-level makes the differences between quantum ar- updating parameters in the circuit and is therefore a cru-
chitectures more apparent. For instance, tunable cou- cial indicator for the run time of variational workloads.
plers [394] can natively implement exchange-type opera- QV, EPLG, and CLOPS are by design only proxies for
tions which, on cross-resonance hardware [392], are syn- the utility of a quantum device. For example, EPLG and
thesized from RZX rotations. Exchange-type gates may CLOPS on ibm_sherbrooke are 1.7% and 2700, respec-
help reduce schedule duration for applications like QAOA tively. These metrics are general enough to draw compar-
that need SWAP networks with phases to overcome lim- isons between different hardware architectures and track
ited device connectivity. improvements of a platform. However, these general met-
Pulse control enables many different types of optimiza- rics could fail to predict which quantum device will obtain
tions to reduce schedule durations. This may be as sim- the best performance, be it in solution quality, execution
ple as a pulse-efficient transpilation in which variational speed, or any other criterion, for a specific application.
parameters are encoded in durations of cross-resonance Therefore, the best possible measure of a quantum com-
pulses [445]. The resulting reduction in schedule dura- puter’s usefulness for optimization is how well it performs
tion improves the sampling performance of QAOA [402]. on tasks similar to the problem of interest, according to
Furthermore, the extra energy levels of the transmon can the performance criterion of choice. For example, the
help to engineer gates that would otherwise require many ability to create spin-entangled states, i.e., squeezing, re-
CNOTs [446, 447]. Finally, advanced methods of quan- lates directly to a QAOA solving a fully connected MAX-
tum optimal control can help synthesize gates [448, 449], CUT problem [454]. We foresee that more application
measurements [448, 450], and adiabaticity [120] which tailored benchmarks are needed, especially since many
may be useful in the context of approximate quantum algorithms are heuristic in nature.
optimization. However, such methods are not always
compatible with quantum error mitigation as discussed
by Egger et al. [451].
C. Execution Example
The pulse-level is also responsible for quantum pro-
gram compilation which entails loading the pulse-
schedule into waveform memory. Care must be taken We now exemplify some of the considerations discussed
at this step and inefficiencies can be costly to the total in the previous sections by examining the work by Sack
run time. and Egger [402]. Here, the authors executed depth-two
QAOA on hardware with Random-Three-Regular (R3R)
graphs with up to forty nodes on superconducting qubit
processors with 127 qubits. First, since only a third of
B. Hardware Benchmarks the physical qubits are needed, the QAOA is run on the
best qubits as measured by the two-qubit gate fidelity;
In general, the performance of quantum hardware is the most used and error-prone circuit instruction. Sec-
measured by scale, quality, and speed [452]. Scale is sim- ond, the R3R connectivity is engineered with a network
ply the number of qubits. Quality is tracked by low-level of predetermined SWAP gates [385, 401]. This results
performance metrics, e.g., qubit coherence and gate fi- in dense and shallow quantum circuits that transpile
delity, and holistic benchmarks such as Layer Fidelity quickly [401]. Third, Sack and Egger [402] also choose
and Error per Layered Gate [415], Quantum Volume a decision-variable-to-physical-qubit mapping that mini-
(QV) [453] and the γ and γ̄ of PEC [15]. A QPU with a mizes the number of SWAP gates following the approach
2n QV can reliably execute a n qubit circuit with n layers of Matsuo et al. [417]. These best practices for QAOA
of random SU(4) gates [453]. The QV accounts for gate are found open source on GitHub [455].
error rates, error suppression and mitigation, and qubit Finally, a machine-learning-based error mitigation of
connectivity. expectation values allowed the optimization of the QAOA
Neither QV, Layer Fidelity nor EPLG, however, ac- variational parameters resulting in γ ⋆ and β ⋆ . A noise-
count for the speed of the computation. Speed is crucial less simulation of the expected mean of the samples
for optimization tasks if the quantum hardware is sup- drawn with γ ⋆ and β ⋆ showed a large gap with respect
posed to deliver a solution with a certain quality faster to samples drawn from the hardware. This highlights
than classical hardware. Furthermore, speed enables the the need for an error mitigation of samples for quantum

32
 VI BENCHMARKS

approximate optimization. The aim of this section is to describe best practices

for benchmarking and provide a set of interesting bench-
marking problems that provide a thorough, robust, fair,
D. Summary and reproducible setting for comprehensive comparisons
of different classical and quantum approaches. The re-
Sec. V provides an overview of the crucial parts of the mainder of this section is structured as follows. First,
software stack affecting the execution of quantum algo- an overview of related work is given in Sec. VI A. Next,
rithms at scale on noisy hardware. When considering Sec. VI B discusses various reasons for running a bench-
this stack, it becomes evident that users possess varying mark. Notably, the goal behind running a benchmark
degrees of control over the scalability of quantum algo- is crucial for the definition of the setting. Benchmark
rithms on noisy hardware. At the higher levels, such as design choices such as the metrics to be evaluated are
problem recasting, encoding, and high-level circuit de- presented in Sec. VI C. Then, we elaborate on a variety
sign, users enjoy a favorable degree of control (also de- of optimization problems which are difficult for state-of-
pending on the problem being considered), allowing them the-art solvers and may provide interesting benchmark-
to shape the algorithm’s scalability to an extent. Best ing problems for quantum methods in Sec. VI D. Fi-
practices for noisy hardware are to choose algorithms and nally, Sec. VI E discusses existing quantum optimization
their variants that lower the circuit depth without com- demonstrations that may build up to a set of benchmarks.
promising on algorithmic speed and quality.
Moving lower into the quantum stack, transpilation
and pulse-level control also offer the opportunity for A. Related Work
optimization. SWAP networks, SAT mappings, pulse-
efficient transpilation, and problem-tailored gate-sets all Important collections of existing (classical) optimiza-
help reduce circuit depth and duration. These lower- tion benchmarks are publicly hosted benchmarking chal-
level aspects demand a profound understanding of quan- lenges. They have proven as valuable tools for fair and
tum mechanics and hardware intricacies. In some cases, insightful comparisons between different classical solvers
these layers may even conceal these complexities behind and/or platforms. These challenges aim at understand-
a primitive such as a sampler or an estimator. It is nev- ing and improving the practical performance of algo-
ertheless crucial to have transparency on the methods rithms for various optimization problems, particularly
employed. This also enables the design of hardware- those that are theoretically hard. They aid in evaluat-
optimized algorithms and problem formulations, which ing realistic algorithm performance conditioned on the
may further improve the performance of executing quan- resources available at the time of the competition. Some
tum optimization algorithms on noisy hardware. of the most relevant, regularly held optimization chal-
lenges tackling problems with classical digital computers
are the Discrete Mathematics and Theoretical Computer
VI. BENCHMARKS Science (DIMACS) implementation challenges [462] and
the Satisfiability (SAT) competition [463]. While the
Benchmarking can provide insights into realistic algo- problems targeted in the former vary from competition to
rithmic performance when purely analytical methods fail, competition, the latter always focuses on SAT problems
and it can provide new perspectives that motivate deeper which represent a well-defined class of problems that, be-
analytical studies. In particular, theoretical complexity yond their theoretical intrigue, find practical applications
analysis typically pertains to general problem classes and across various fields [464, 465]. Another important re-
considers mostly asymptotic scaling behavior or worst source for optimization benchmarks is the Mixed Integer
case scenarios. In fact, theoretical worst case bounds, as Programming Library (MIPLIB) [466]. This library pro-
they are discussed in Sec. II, often fail to describe the dif- vides regularly updated pure and mixed integer programs
ficulties of solving real-world applications that focus on that are chosen under the consideration of solvability and
a specific problem instance of practical use, nor do they numerical stability. Currently, it holds 240 benchmarking
faithfully describe how well practically relevant problems instances that are solvable with existing methods and an
can be solved for fixed size and data. In this case, we even larger collection of unsolved, and numerically diffi-
need to rely on meaningful and systematic benchmarking cult instances. The performance of many solvers regard-
of optimization algorithms with well-defined assumptions ing several classes of optimization problems is regularly
considering implementation, data, and metrics — a chal- compared in [467].
lenging task. Clear benchmarks that study the utility Optimization competitions and libraries of this form
and robustness of optimization algorithms can also serve that are compatible with (near-term) quantum comput-
to provide trustworthy and verifiable performance indi- ers do not exist as of now. It remains an open task to set
cators for a broader audiences, including policy makers them up and, thereby, enable streamlined benchmarking
and industry leaders. For a general introduction to com- of quantum optimization methods. In fact, most exist-
putational benchmarking, we refer the interested reader ing related work focuses on application-centric quantum
to Refs. [456–461]. hardware benchmarking frameworks that employ opti-

33
 VI BENCHMARKS

mization problems as exemplary application. A bench- ison of run times, solution qualities, etc., of classical and
marking framework presented by Lubinski et al. [468] quantum solvers. Suggestions for suitable benchmarking
suggests the comparison of the output quality to circuit problems are given in Sec. VI D.
width and depth for implementations of Shor’s factoring Benchmarking may also be executed to track the
algorithm, Grover search, Hamiltonian simulation, etc., progress in hardware developments (cf. Sec. VI A) as well
executed in numerical simulations or gate-based quantum as algorithmic improvements. This is particularly impor-
hardware. Another framework introduced by Tomesh et tant for monitoring the impact of the fast development in
al. [469] measures the qubit connectivity, circuit depth, quantum hardware, supporting software, and algorithmic
number of 2-qubit gates, ratio of gates which can be exe- research. For an example of tracking progress of classical
cuted in parallel, number of times qubits are acted upon, hardware and algorithms, we refer the interested reader
and number of mid-circuit measurements for algorithms to Koch et al. [458], which discusses the advancements in
including QAOA, VQE, and Hamiltonian simulation al- mathematical programming solvers over 20 years.
gorithms executed on quantum hardware platforms from Furthermore, benchmarking can give us insights into
different providers. Furthermore, Martiel et al. [470] an algorithm’s scaling behavior: Is there a problem scale
present a metric that measures the maximum problem at which algorithms fail to provide good solutions — pos-
size for which a problem instance can be solved given a sibly due to hardware limitations? How do the resources
pre-defined accuracy level. Notably, the suggested metric needed to find solutions to a problem of a certain quality
does not incorporate time as an explicit factor. Instead, scale with the system size? An example of MIPLIB prob-
the metric is conditioned on algorithms to be run in poly- lem instances is illustrated in Fig. 6. In this case, many
nomial time. This directly excludes the use of error miti- – but not all – larger instances require longer solver run
gation and can easily fail to provide a fair and meaningful times.
comparison between different platforms. Further bench-
marks that investigate different properties of quantum Scaled by number of nonzeros
annealing methods for optimization problems – some in- 104
cluding comparisons to classical and quantum algorithms
– can be found in Refs. [471–478]. A comparison of Ising
machines that checks success probabilities for finding the 103
ground state and the corresponding time-to-solution is
presented by Mohseni et al. [479].
One particularly interesting work is introduced by
102
Finžgar et al. [480]. They compare various QUBO solvers
with respect to the solution quality and time-to-solution
and as such represents a good baseline for an optimiza-
tion benchmark. Although the presented publications 101
do not offer holistic optimization benchmarking frame-
works, they provide interesting insights that can help to
define reproducible, representative, and fair benchmark
libraries, which facilitate identifying advantages and dis- 101 102 103 104 105
advantages of various (quantum) solvers.

Figure 6. The figure illustrates the time it takes the Gurobi

B. Goals solver [481] to find the optimal solution for MIPLIB 2017 [466]
instances of the form minx cT x subject to Ax ≤ b, x being
partly integer. The size of the circles depicts the number of
Reasons for running a benchmark can be diverse. In non-zero coefficients in the constraint matrix A and the in-
the following section, we focus on benchmarking goals stances are ordered (on the x-axis) according to the square
related to understanding the potential of quantum com- root of the number of variables times the number of con-
puters for solving optimization problems in terms of a straints.
fair comparison.
One of the main goals of benchmarking in the present
context is the comparison of classical and quantum algo-
rithms — aiming at demonstrating quantum advantage. C. Design
If classical platforms are compared amongst themselves,
one often relies on benchmarking problems such as SPEC To ensure reproducibility as well as replicability of a
[460] and LINPACK [461], which are established in the benchmark, it is important to clearly define the specifics
community. However, these problems are not well suited in terms of model selection, pre-processing, computa-
to compare classical and quantum platforms. Instead, tional platform, target metrics, and selected algorithm
the community will have to agree on a set of problem- (provably exact, heuristic, etc.) including suitable hy-
centric benchmarks that could enable the direct compar- perparameters. A discussion of relevant algorithmic

34
 VI BENCHMARKS

paradigms and quantum algorithms for optimization we refer the interested reader to Berthold and Csizma-
problem classes can be found in Sec. III and Sec. IV, dia [482]. Some model-dependent benchmarking test in-
respectively. The remaining levels of design choices are stances that are openly available, established, and un-
discussed in the following section. Starting from the de- solved are given in MIPLIB [466], the Quadratic Pro-
cision whether the benchmark is defined with respect to a gramming Library (QPLIB) [483], and the Satisfiability
particular model, and, if so, which one. Secondly, it usu- (SAT) Problem Library [484]. As an example, we can
ally is beneficial to apply pre-processing to reduce the consider (random) SAT instances which exhibit phase
problem to its challenging core. A possibly applied pre- transitions. It may now be of particular interest to in-
processing procedure may already consider the compu- vestigate whether a quantum algorithm is better suited
tational platform that is employed afterwards to execute to handle the difficulty of a phase transition. In fact,
the optimization algorithm. Finally, the algorithm exe- it has been studied by Yu et al. [485] whether an adap-
cution and resulting outcomes can be captured in a com- tive bias version of QAOA might help to avoid transi-
parable fashion by choosing appropriate benchmarking tion phenomena. While the presented study focuses on
metrics. small problem instances, where classical algorithms can
provide solutions, the findings hint at the potential for
advantages in the realm of large problem densities and,
1. Model-(In)Dependence hence, motivate further investigation.

At the core of a practically motivated optimization

task is the solution of a concrete decision problem such 2. Pre-processing
as the selection of an optimal (or at least a good) tour
for TSP. In order to treat this problem in a rigorous The term pre-processing or pre-solving refers to a set
way a mathematical model must be developed. How- of methods that are applied to a problem instance or
ever, the model choice is usually not unique and may data before the actual optimization algorithm is exe-
lead to differently behaving algorithms. In particular, cuted. Standard pre-processing approaches require clas-
certain models may be significantly better or worse suited sical computational effort. Therefore, the pre-processing
for a particular optimization algorithm or may even con- should also be taken into account in a benchmark’s per-
tain characteristics that cannot be handled by a particu- formance evaluation.
lar algorithm, such as non-linearities or constraints. For The goal of pre-processing is to modify the problem in
benchmarks, it is therefore crucial to decide on the level such a way that it allows the optimization algorithm to
a problem is defined: either model-independent, i.e., on work more efficiently. How to implement pre-processing
the level of the original optimization problem, which al- in practice depends on the mathematical formulation
lows different mathematical modeling methods, or model- of the optimization problem, on the data, as well as
dependent, i.e., on the level of a concrete mathematical on the considered optimization algorithm and computa-
model. tional platform. In fact, it is not uncommon for many
Model-independent benchmarks that place few limi- problems such as Euclidean TSP or Steiner tree problems
tations upon admissible solution strategies offer an in- in graphs that pre-processing is able to remove 95% of the
teresting method to identify quantum advantage for op- variables and, hence, reduce the problem size as well as
timization. More specifically, quantum advantage may resource requirements significantly [486, 487]. How this
stem from novel formulations of problems that could can be achieved often depends on the context and char-
not be captured on a model-dependent level. Further- acteristics of the problem and data. A central aspect of
more, quantum hardware is quickly evolving. To enable pre-processing is the reduction of the search space, for ex-
comparability and compatibility for different generations ample by decreasing the number of decision variables, or
of hardware, model-independent benchmarks are better reducing the number of alternative optimal solutions, or
suited, since they may allow one to leverage new features by decomposing the problem into smaller subproblems.
that might not have been available in earlier generations, Considering the example of QUBO, various reduction
such as dynamic quantum circuits [432]. We expect these techniques to eliminate optimization variables have been
benchmarks to eventually show how quantum optimiza- suggested in the literature. A recent review has been
tion algorithms are approaching, meeting, and exceeding provided by Rehfeldt et al. [487]. For example, a set of
the capabilities of classical techniques for solving partic- analytic reduction rules is presented by Glover et al. [488]
ular problem instances. that allows one to conditionally fix certain optimization
Nevertheless, model-dependent benchmarks are also a variables, while guaranteeing that at least one optimal
useful tool. In particular, they can help to study the rel- solution will remain in the search space. In addition
ative performance of different algorithms and platforms to the number of optimization variables, their degree of
for fixed model formulations. At this point, it should be correlation, which is determined by the sparsity of the
noted that model-specific solvers are often significantly QUBO matrix, is also of interest. In the case that certain
faster for the target problem than general solvers. For groups of decision variables are not correlated, QUBOs
specific metrics to compare different types of solvers, can be decomposed trivially into smaller sub-QUBOs to

35
 VI BENCHMARKS

hybrid the binary values 0 and 1. The information in classical

computing digital computations is processed by compositions of log-
ical operations such as AND and NOT gates which can
classical universal be expressed in a circuit model. Digital classical com-
putational models often rely on a combination of Cen-
computing
HPC QC
tral Processing Units (CPUs), which excel at sequential
digital

classical quantum
digital digital and general data processing tasks, and Graphics Process-
specialized ing Units (GPUs), which specialize in parallel processing
hardware digital tasks. There are several other types of processors and
accelerators simulation specialized hardware accelerators designed for specific
analog
tasks and applications, for example Tensor Processing
Units (TPUs) for deep learning or Field-Programmable
computing

simulation
Gate Arrays (FPGAs) for signal processing and embed-
analog

classical quantum ded systems, just to name two. The combination of CPU,
analog analog GPU and special-purpose processors forms the founda-
quantum tion for classical high-performance computation in a wide
annealing range of applications. Due to their widespread use, clas-
sical digital hardware devices span a wide spectrum of
classical quantum computing platforms, including desktop computers, com-
computing computing puting clusters, cloud services, mobile phones, and many
more. All of these platforms can be used to solve opti-
mization problems.
Figure 7. Sketch of the optimization platform ecosystem. We In additional, special purpose systems for optimiza-
focus on digital classical computing as well as digital and ana-
tion are developed. More specifically, these systems sim-
log quantum computing.
ulate quantum phenomena such as quantum annealing,
see Sec. III B, hence, aiming to gradually decrease an en-
ergy function. Well-known examples for this type of tech-
be solved independently. nology are the Fujitsu Digital Annealer [489, 490] that is
based on a hardware-accelerated Markov Chain Monte
Carlo approach, the Hitachi CMOS Annealing Machine
3. Platforms [491] which corresponds to an in-memory computing ar-
chitecture that represents an Ising model with local inter-
Different platforms offer different strengths and suffer actions, and the Toshiba Simulated Bifurcation Machine
from different bottlenecks. Hence, some devices could [492, 493]. It should be noted that the latter does not ap-
be better suited for certain optimization problems than proximate an annealing process but instead implements
others. It may not be directly evident which hardware a highly parallelizable approximation of Ising dynamics
is best suited for a given optimization problem. Fur- with classical hardware (GPU, FPGA). This allows one
thermore, depending on the platform, the optimization to compute heuristic solutions to problems with up to
models and applicable algorithms do vary, resulting in millions of variables in a time frame of seconds.
better or worse performance for a given problem. In digital quantum computing [390] the basic units of
Tailoring the choice of platform to the characteristics of information are qubits instead of bits. The information
the problem at hand – if possible – can lead to better out- processing consists of consecutive quantum gate opera-
comes. This approach aligns with the idea of platform- tions such as CNOT gates and single-qubit rotation gates
aware optimization, ensuring that the selected platform which, together, form a quantum circuit, similar to a clas-
is optimally matched to the unique requirements of the sical logical circuit. Important bottlenecks of existing
problem being addressed. To understand the advantages digital quantum computers are limited coherence times,
and disadvantages of different platforms for different op- which result in limited circuit depths, limited qubit num-
timization problems, performing cross-platform bench- bers, and limited qubit connectivity. There are a variety
marking is crucial. At the same time, such benchmarks of ways to quantify the performance of a digital quan-
pose a particular challenge, as they must allow a fair tum computer. For near-term hardware where noise is
comparison between systems that can function very dif- a dominant factor, one might be tempted to consider
ferently. In the following, we are going to elaborate on metrics such as the coherence times of qubits or the fi-
various hardware types that are relevant for optimization delity with which various quantum operations can be per-
problems, i.e, classical vs. quantum, and analog vs. dig- formed. Such metrics are valuable, however, they can fail
ital. A sketch of the platform ecosystem is shown in to give an accurate description of a device’s potential for
Fig. 7. useful computation. This has motivated proposals for
Classical digital systems are the workhorse of today’s more holistic error metrics, as discussed in Sec. V B.
computer technology. The basic units of information are Quantum analog hardware is realized by quantum an-

36
 VI BENCHMARKS

nealing [2, 3, 471, 494, 495] and quantum analog simula- ber of cores. Quantum computational resources can be
tion [496–504]. In quantum annealing, the energy func- measured in terms of the number of circuits that are run,
tion or Hamiltonian of a system is gradually (but not nec- the number of executed gates per circuit, the number of
essarily perfectly adiabatically) deformed from an initial qubits and shots per circuit execution, and the quality of
ground state to a final state that can be measured to find the qubits. Of course, the resource count should also con-
the solution to an optimization problem. These systems sider resources that are required for problem-dependent
are designed for QUBOs. Important bottlenecks of these calibrations and error mitigation or – if applicable – cor-
platforms include the coherence time, restrictions on the rection. The number of qubits in a circuit strongly de-
driver Hamiltonian, and challenges in embedding practi- pends on the mapping of problem variables to qubits.
cal problems into the device connectivity. The coherence Hence, algorithmic improvements in mapping schemes
time of current publicly available annealing machines is can lead to a reduction in qubit requirements.
shorter than the annealing time [505]. This means that The run time provides an objectively fair metric to
these platforms do not achieve closed system dynamics compare different computational platforms. It is, hence,
but exhibit strong coupling with the environment. Fur- of utmost importance to clearly define what determines a
ther hardware improvements are therefore required [506]. faster algorithm. The effective run time of an algorithm
The primary computational resource is believed to be in- consists of time used for pre-processing, transpilation,
coherent quantum tunneling events [507, 508]. Therefore, embedding, compilation, execution, and post-processing.
recent efforts are directed towards engineering quantum Now, it is not a priori clear, which of these time fac-
drivers that cannot be simulated classically [506, 509]. tors should go into a time benchmarking metric: (i) only
Quantum analog simulation describes a setup consisting the time from start to end of the solver (ii) including
of physical systems with sufficiently controllable parame- the generation of the model for the specific solver, (iii)
ters such that one can manipulate and probe the dynam- including pre-processing from the raw data. Arguably,
ics underlying a family of Hamiltonians. More specifi- a fair comparison considers the total time. Neverthe-
cally, analog quantum simulators enable the emulation of less, it can also be interesting to analyze the composition
continuous quantum dynamics of a target system with a of the total time as this can help to identify the most
simulator that is easily accessible in the lab. This setting costly steps and potential bottlenecks. Given the cur-
has proven itself valuable to simulate physical dynamics rent status of software for quantum computing transpila-
including quench problems, and ground state problems tion, embedding, and compilation times will in fact play
[496] but has not been studied in much detail as a plat- a non-negligible factor. Hence, understanding the rela-
form for optimization applications, yet. tion of time-to-solution and QPU time gives us insights
into the (progress in) efficiency of these steps. However,
these are also areas of active research and development
4. Metrics and we expect them to be significantly reduced in the
near future. Another factor to take into account when
Due to the technological differences between classical measuring the time-to-solution is the potential use of par-
and quantum systems, the selection of a set of metrics allelization, which effectively corresponds to a trade-off
that enable a fair comparison is crucial. The four main between computational resources, i.e., cost, and run time
metric categories that suffice said criteria are resource [510].
cost, run time, quality, and problem complexity. It should also be noted that the nature of a run
The resource cost includes any quantum or classical time benchmark strongly depends on the benchmarking
computational resources employed to solve a given op- goal, e.g., aiming to reach a feasible, proven optimal, or
timization problem including any potential pre- and/or ϵ−close to optimal solution. If the goal is finding a proven
post-processing steps and resources that are being used optimal solution, then it has to be taken into account
in parallel. The respective costs may be quantified in that certifying optimality typically requires a significant
various ways. Firstly, a fair metric would be the amount amount of time compared to merely finding the optimal
of energy that is used to solve a problem. However, in solution. In the case of feasibility problems, we must con-
most cases, the user will not have access to this infor- sider what to do if one method finds a solution and one
mation, particularly, when specialized computational re- does not. A good approach is to have feasibility prob-
sources such as CPUs, GPUs, or quantum computers are lems as a separate class and count how many problems
accessed via the cloud. Similarly, the monetary cost of can be solved within a certain time limit per instance —
running an algorithm – including the pricing of special- even if sometimes none is found. Setting the right time
ized computational resources – could provide an interest- limit is crucial. Short times < 1s are often difficult to
ing metric but will be tough to evaluate on a fair basis measure together with the setup time. Common choices
as these depend on hardware providers and service con- for classical heuristics are therefore < 1s, 10s, 60s, 600s,
tracts. Classical computing resources are typically mea- 3,600s, and 10,000s.
sured with respect to factors such as the used memory, Different forms of optimization methods, for exam-
required memory bandwidth, and central processing unit ple deterministic vs. heuristic, also impact the specifics
(CPU) specifications, including clock speed and the num- of a run time benchmark. Considering, e.g., non-

37
 VI BENCHMARKS

deterministic methods that produce various solutions, are unable to provide provably optimal or feasible solu-
one has to fix whether the run time is calculated as the tions. Furthermore, it is crucial that the problem can
total time, the average time to find a solution, the mini- be formulated as a quantum-compatible model, i.e., with
mum time to find a solution, etc. One proposal to address sufficiently few variables/qubits, limited connection den-
this, is to explore the whole boundary of time-to-target sity, relatively small coefficients, etc. In other words, we
and optimality-gap-at-time values [456, 511]. Another are looking for a sweet spot for current quantum sys-
metric that aims at giving a holistic point of view by tems that enables the investigation of quantum advan-
considering a combined measure for solution quality and tage compared to adequately chosen digital algorithms.
total heuristic run time is suggested by Berthold [512].
It can also be valuable for tracking the progress in algo-
Quality metrics strongly rely on the nature of a prob-
rithmic and hardware development to identify a family
lem. Firstly, one requires a solution to be compatible
of problems that are similar in their nature but differ
with the problem constraints, i.e., to be feasible, which
in their difficulty. Notably, complexity theory typically
can be easily checked. Ultimately, one wants to find the
makes statements about the hardest instances in a prob-
optimal solution or at least a solution that is ϵ−close
lem class and, therefore, seldom provides descriptive ar-
to the optimal solution. It should be noted that the
guments about the hardness of individual instances —
evaluation of optimality may be resource-intensive or in-
which very much depends on the size, structure, and
tractable. Given a feasible solution, the evaluation of the
precise coefficients of the instances. Furthermore, the
optimality gap may be identified via bounds given from
particularly chosen instances may correspond to crafted,
continuous relaxations [513] or dual representations [514].
random, or real-world problems:
Another quality metric for optimization algorithms is so-
lution diversity. Given multiple solutions with equal cost Crafted instances are most likely the best way to
values, one would expect a fair sampling from these solu- find problem instances that are hard to solve for classical
tions. Interestingly, this property is not necessarily given methods. However, there is a risk, that there is (or can be
in quantum annealing algorithms [515–517], and also not developed) also a classical method that takes advantage
easy to realize on current gate-based devices [518, 519]. of the special problem structure, or that the instances
Similarly, one can measure the success rate of a heuris- are of no practical value.
tic algorithm to find solutions that are feasible or suffice
a certain optimality distance in a fixed number of algo- Random instances are different since they lack struc-
rithm execution runs. ture. This can make them, depending on the problem,
The practical difficulty of a problem is often related to particularly easy or very difficult to solve. It certainly
the problem size, density or number of variable connec- sets them apart, but how meaningful these instances are
tions, and the nature of the constraints. The difficulty can be debated.
may be changed by modifying one or more parameters Real-world instances correspond in many practi-
that characterize the system as a whole. These changes cally relevant cases to MIPs that can be solved up to
can even lead to computational phase transitions, which sufficient accuracy with existing solvers. Instances that
are marked by abrupt and significant shifts in the time are relevant and difficult to solve with existing methods
to find a solution of certain quality, or even the feasibility are hard to find in a setting that fits the dimensionality
of finding a solution. Understanding the range of prob- and density limitations of current quantum hardware. In
lems and characteristics where phase transition behavior fact, these problems are generally hard to find because
is observed presents a challenging task. open problems of industrial relevance are rarely published
In order to provide a holistic quantum optimization or may be ill-defined. This highlights a common selection
benchmark, metrics that describe algorithm and solu- bias: We rarely model practically relevant problems in a
tion properties according to these four classes should be way we know we cannot solve.
measured. Depending on the problem, model, algorithm,
and execution platform the respective metrics should be In the following, we list a set of binary problems
wisely chosen and faithfully reported on. that are promising candidates for quantum optimization
benchmarking problems. These problems are random or
crafted instances and relate only to few practical appli-
D. Problems cations. Nevertheless, they offer a good test bed for al-
gorithm development and progress tracking. Classical
methods that can be used to tackle these problems in-
Next, we present a set of model-independent problems clude Branch-and-Cut based Integer (linear, non-linear,
that show promise for establishing standardized quantum semi-definite) programming (ILP) [520], Pseudo-Boolean
optimization benchmarks. The problems should be cho- (PBO) optimization [521], SAT-Solving (SAT) [463, 522],
sen sufficiently hard for state-of-the-art classical solvers Constraint Programming (CP) [523], and several other
to leave a margin for quantum methods but within an general techniques [524, 525] as well as a multitude of
intermediate size range where current and/or near-term specific heuristic approaches.
quantum technologies can be effectively deployed. Ideal
problem instances are those for which existing methods Maximum Independent Set (MIS). Given a

38
 VI BENCHMARKS

weighted graph G = (V, E, c), we are looking to find such that

X X
min cv x v min |sj |
x ∈ {0, 1}|V | v∈V (23) x ∈ {0, 1}n , s ∈ RJ j∈J
X (26)
s.t. xu + xv ≤ 1, ∀(u, v) ∈ E. s.t. wij xi + sj = Cj , ∀j ∈ J.
i∈I
MIS can be formulated as a QUBO for a suitably large
penalty factor P : If designed with carefully selected weights, we get so-
X X called market share problems [533] that are explic-
min cv x2v − P xu xv . itly designed to be difficult to be tackled by classical
(24) branch-and-bound based methods. These problem in-
x ∈ {0, 1}|V | v∈V (u,v)∈E
stances are dense multi-dimensional knapsack problems
This is a classic NPO-hard problem even in case cv = 1 [62, e.g. Chapter 31]. Still, there exist instances with
and there are known instances that are hard to solve less than n = 100 variables and m = 8 constraints which
to proven optimality with existing methods starting at cannot be solved to proven optimality. However, for this
problem sizes of several hundred variables [526]. The problem class the difficulty depends strongly on the par-
allure of this problem class for quantum optimization ticular coefficients, the ratio ci /wi , and C. Problems
benchmarking is that it is well suited for translation into with comparatively small C can be solved quickly by
relatively sparse QUBOs. DPs and if the wi are small compared to C, heuristics
will give good results. For carefully chosen coefficients,
A subclass are MIS with unit disc graphs G (UD-MIS) however, these problems can be difficult to solve with
which are defined by vertices V on a two-dimensional classical methods to proven optimality. We argue that
plane with edges E connecting all pairs of vertices within this problem class provides a good quantum optimization
a unit distance of each other [527, 528]. These problems benchmarking candidate because there is clear evidence
can be naturally mapped onto the structure of Rydberg that it is challenging for classical digital systems to find
quantum computers [529, 530] via the Rydberg block- (approximate) solutions already — even for small system
ade [386]. The optimization itself can then be carried sizes in the multidimensional case.
out either in an analogue or digital fashion. It is also
worth mentioning here that any QUBO can be mapped
constructively to a UD-MIS with weighted vertices (UD- Low Autocorrelation Binary Sequences
MWIS) [531]. (LABS). This is a difficult binary non-linear optimiza-
tion problem [12]. Given a sequence S = (s1 , . . . , sk ) of
Optimization problems that may be modeled as a MIS length k with binary si ∈ { −1, +1 }, the autocorrelations
are, for example, Sudoku and the multidimensional n- of the sequence correspond to
Queens problem. Interestingly, 3D n-Queens problem in- k−j
stances on a 143 board with 2,744 binary variables, and X
Aj (S) = si si+j (27)
most larger board sizes, have no known proven optimal
i=1
solution. As there is by definition only one problem in-
stance per board size, we can nicely track how far we get. for j ∈ {0, . . . , k − 1}. Furthermore, the sequence energy
Furthermore, it should be noted that the maximum clique is defined as
problem can be solved as a MIS on the complementary
graph. k−1
X
E (S) = A2j (S) . (28)
(Multi-dimensional) Knapsack / Market Share. j=1
Given a set I = {1, . . . , n} of items, weights wi , i ∈ I, and
The goal of the optimization problem is to find a sequence
a capacity C, the basic knapsack P constraint regarding a S that minimizes the energy function given above. There
set of binary variables xi , is i∈I wi xi ≤ C. Knapsack
is only one particular instance per size N . Hence, the
problems typically have an objective function, where a
frontier of what is possible is always well defined. So
profit pi is assigned to each item, i.e.,
far problems up to size N = 66 and N = 127 (skew-
X symmetric) can be solved to optimality. Given the spin-
max p i xi
x ∈ {0, 1}n like binary variables, one might be able to find a model
i∈I
X (25) formulation that could provide a natural fit for a quan-
s.t. wi xi ≤ C. tum platform. However, the quartic objective function
i∈I also imposes some challenges to do so [104].
P
A quadratic objective i,j∈I 2 pij xi xj corresponds to the
binary quadratic knapsack problem [532]. In a multi- Quadratic Assignment Problem (QAP). The
dimensional knapsack problem, we have J dimensions QAP [534] has remained one of the great challenges in

39
 VI BENCHMARKS

combinatorial optimization. For I = {1, . . . , n} them provably average-case hard for various families of
X classical algorithms [95]. It remains open as to whether
min aij bkl xik xjl quantum algorithms can do better in these average-case
n×n
x ∈ {0, 1} i,j,k,l∈I settings, and thus, superpolynomial speedups for spin
X glass problems are still possible. Considering all of the
s.t. xij = 1, ∀i ∈ I, (29) above, we argue that they offer an interesting candidate
j∈I
X for quantum benchmarks, especially in a cross-platform
xij = 1, ∀j ∈ I, setting.
i∈I

where aij and bkl describe the problem instance. It is still E. Demonstrations
considered a computationally nontrivial task to solve to
optimality even for modestly sized problems, i.e., n ≈ 30.
If modeled as a binary problem, the number of variables Experimental realizations of large scale quantum op-
is squared. Modeling this problem as a QUBO results timization algorithms have to address several challenges.
in a dense formulation. Nevertheless, the long and un- Not only is it difficult to cope with hardware induced lim-
successful search for improvements in classical methods itations – see Sec. V for further details – but it is already
indicates that there might be room for quantum advan- a non-trivial task to choose an appropriate benchmarking
tage. problem – see Sec. VI D. In fact, most existing hardware
demonstrations of quantum algorithms for optimization
Sports Timetabling Problems. There are regu- are considering QUBO-type problems with a grid struc-
lar competitions held to find the best methodology for ture. Several papers argue in the following way: First, it
tackling timetable scheduling [535]. A mathematical for- is stated that a problem is difficult according to complex-
mulation that describes problems of this form has been ity theory, e.g., it is NPO-hard. Then, a feasible – but
presented, for example by de Werra [536]. These prob- not provably optimal – solution is found by applying a
lems describe scheduling problems of sports leagues or quantum algorithm. However, finding a feasible but not
tournaments and are generated in a way that ensures necessarily optimal solution is not necessarily difficult for
at least one feasible solution exists per instance. These state-of-the-art classical methods [472, 487, 542]. Thus,
problems have a strong combinatorial structure meaning there is an inconsistency in the argument, since classical
that in different feasible solutions many variables need to algorithms are ruled out with a reference to complexity
have different values. theory while allowing to use quantum heuristics. These
Interestingly, for several medium sized instances no results are valuable capability demonstrations, i.e., they
known (and investigated) method was able to find even can be seen as hardware benchmarks rather than per-
a single feasible solution [535, 537]. There are different formance benchmarks. Nevertheless, there is a potential
kinds of requirements, i.e., problems can be generated for better performing quantum heuristics, and to show
with increasing difficulty. From pure binary problems, to this, we require careful and fair systematic benchmark-
problems that rely on the counting of integer variables ing. In this section, we present an overview of selected
or binary representations thereof. What is particularly state-of-the-art experimental quantum optimization re-
interesting about these problems is that one can generate sults run on gate-based quantum computers with more
instances at variable size and difficulty to track progress than 16 qubits that represent excellent starting points for
while it is known that sufficiently large problems are dif- comprehensive quantum optimization benchmarks.
ficult for state-of-the-art methods. Tab. IV summarizes existing quantum optimization
Spin glasses. The study of optimization problems hardware implementations run with (variants of) QAOA
with respect to physics Hamiltonians, such as spin glass or VQE and problems corresponding either to (weighted)
models [471, 538], is of great importance and has inspired MAXCUT, Sherrington-Kirkpatrick, or other instances
the development of classical solvers, including simulated of QUBO/PUBO that use more than 15 qubits. The
annealing [539] and simulated quantum annealing [540]. most commonly employed graph structures are random
Spin glasses defined over different graphs have already 3-regular (R3R) graphs or graphs that are particularly
been used to compare the performance of classical op- fitted to hardware layouts such as heavy-hexagon or near-
timization with quantum annealing methods [471, 538]. est neighbor grid layout [385, 395]. Besides the target
This is enabled by their natural representation as quan- problem and employed algorithm, the table also lists
tum Hamiltonian. More specifically, the mapping of a the problem size in terms of number variables (which
spin glass problem onto a QUBO is trivial and leads equals the number of qubits in the considered cases),
to relatively sparse models. Although classical solvers the density (which quantifies the number connections be-
perform well on spin glasses defined on regular lattices tween variables), the best and mean approximation ratio,
compared to quantum annealing, the possibility of dig- the depth of the ansatz in terms of the number QAOA
ital quantum optimizers having better performance still operator repetitions (except for results from Amaro et
exists — possibly on non-regular lattice structures [541]. al. [544], which corresponds to the repetitions of a “Re-
Moreover, spin glass problems have properties that make alAmplitudes” ansatz with pairwise entanglement [547]),

40
 VI BENCHMARKS

Table IV. An overview of state-of-the-art experimental realizations of optimization algorithms on gate-based quantum computers
with more than 15 variables. In cases where data was not made available in the corresponding publication or the accompanying
data repository, we denote this in the respective field with N/A. AR denotes the approximation ratio, given based on the mean
and the best sample value of the experiment, n.n. grid stands for nearest neighbor grid. Furthermore, JSP, FVQE, QAMPA,
GQAOA, and NDAR abbreviate job shop scheduling problem, filtering variational quantum eigensolver, quantum alternate
mixer-phaser ansatz, greedy QAOA, and noise-directed adaptive remapping, respectively.
AR
Problem Algorithm Qubits Density mean best Depth Year Ref.
Sherrington–Kirkpatrick QAOA 17 100% 0.61 N/A 1≤p≤3 2021 [385]
MAXCUT (R3R) QAOA 20 16% 0.64 1 p=2 2023 [402]
MAXCUT (R3R) QAOA 20 16% 0.94 1 p ≤ 10 2023 [543]
MAXCUT (R3R) QAOA 22 14% 0.67 N/A 1≤p≤3 2021 [385]
MAXCUT (n.n. grid) QAOA 23 13% 0.72 N/A 1≤p≤5 2021 [385]
QUBO (JSP) FVQE 23 N/A 0.88 N/A 1≤p≤2 2022 [544]
MAXCUT (heavy-hex.) QAOA 27 8% N/A 1 p=2 2022 [401]
MAXCUT (R3R) QAOA 30 10% 0.59 0.83 p=2 2023 [402]
MAXCUT (R3R) QAOA 32 10% 0.88 1 p ≤ 10 2023 [543]
MAXCUT (R3R) QAOA 32 10% N/A 1 p=2 2023 [226]
MAXCUT (R3R) QAOA 40 8% 0.58 0.78 p=2 2023 [402]
Sherrington-Kirkpatrick QAMPA 50 100% 0.55 0.83 p=2 2023 [264]
Sherrington-Kirkpatrick QAOA 50 100% 0.54 0.84 p=2 2023 [264]
Sherrington-Kirkpatrick GQAOA 72 100% N/A 0.92 p=1 2023 [229]
Sherrington-Kirkpatrick NDAR-QAOA 82 100% 0.87 0.97 p=1 2024 [545]
QUBO (heavy-hex.) QAOA 127 2% 0.67 0.85 1≤p≤2 2023 [437]
PUBO (heavy-hex.) QAOA 127 2% 0.65 0.84 1≤p≤2 2023 [437]
PUBO (heavy-hex.) QAOA 127 2% 0.73 0.89 1≤p≤5 2023 [546]
QUBO (heavy-hex.) QAOA 414 0.6% 0.57 0.69 p=1 2023 [546]
PUBO (heavy-hex.) QAOA 414 0.6% 0.56 0.68 p=1 2023 [546]

and the year of publication. All presented instances with cost Hamiltonian, with its effectiveness demonstrated for
fewer than 50 qubits employ a standard QAOA or VQE QAOA with 82 qubits [545].
schemes, i.e., they are optimizing the variational param- The table illustrates that most experiments with
eters using the quantum computer. Since the references higher qubit numbers have lower problem densities. In-
report on shot numbers respectively number of parame- creasing the density for a larger number of qubits would
ter update iterations in various ways (or not at all), the require additional SWAP gates, cf. Sec. V, hence, re-
respective numbers are difficult to compare and have, sulting in increased hardware noise. This phenomenon
therefore, not been added to the table. However, for manifests itself also in the resulting approximation ratios.
future benchmarks, providing such numbers in a com- Considering the approximation ratios, it should be noted
parable format would be important. The results of Ma- that for common graph problems such as MAXCUT the
ciejewski et al. [264] with 50 qubits consider both QAOA, objective values usually lie between 0 and a maximum
and QAMPA [263], a hardware-efficient ansatz derived value Cmax , while for general problems such as Ising mod-
from QAOA and using the same number of entangling els usually the objective values are inside an arbitrary
gates. The 72 qubit experiment by Dupont et al. [229] interval [Cmin , Cmax ]. Therefore, a fair comparison of ap-
on the other hand, employs a greedy procedure. In each proximation ratios necessitates that the data of each ref-
iteration cycle, several variables are set to values that erence is expressed consistently. Thus, we first convert a
are evaluated with a classical procedure, hence, produc- problem into a minimization problem, and then, we nor-
ing a smaller optimization problem to be solved on the malize by the range |Cmax − Cmin |, i.e., we define the ap-
quantum computer. Further, the demonstrations with proximation ratio as (Cmax − ⟨C⟩)/(Cmax − Cmin ), where
more than 100 qubits/variables [437, 546] employ either ⟨C⟩ denotes the achieved objective value. In fact, some
a gridsearch over ∼7000 parameter configurations or con- published values do not correspond to range-normalized
centration of parameters [222–224], also cf. Sec. IV A 1, data, such as the ones from Harrigan et al. [385]. There-
where QAOA on large instances can be run on quan- fore, the values listed in the table are computed from the
tum hardware using parameters found for classically corresponding data repositories. Moreover, several refer-
trained smaller (e.g., 16-qubit) instances [546]. Finally, ences give the approximation ratio based on the best sam-
the recently proposed Noise-Directed Adaptive Remap- ple (which may improve with a higher number of shots)
ping meta-heuristic improves algorithm performance on and others on the mean sample value ⟨C⟩. Hence, these
noisy hardware by adaptively gauge transforming the values are listed separately in Tab. IV.

41
 VII ILLUSTRATIVE APPLICATIONS

In addition, we would like to highlight a work by Fuller course for research and link to problem classes discussed
et al. [237] and Moses et al. [226], where strategies have in Sec. IV. Further, the discussion highlights that being
been investigated to address problems with more vari- practically relevant requires taking into account many
ables than qubits. In the former, MAXCUT problems on details and complications that most published results in
a planar graph with up to 40 variables are solved with the related quantum optimization literature are lacking.
QRAO using up to 15 qubits, cf. also Sec. IV A 1. In fact, Thus, defining a family of problems with increasing com-
the approach achieves an approximation ratio of 0.905 for plexity helps to determine progress on the quest towards
the 40 variable problem executed on quantum hardware. quantum advantage in optimization.
In the latter, a variant of QAOA with mid-circuit mea-
surements and qubit re-use has been tested which could
be scaled to 130 variables on 32 qubits, achieving an ap- A. Financial Asset Allocation
proximation ratio around 0.8. In general, the intrinsic
differences between the method employed in this work Optimization problems in finance span a wide range of
and standard QAOA / VQE approaches make it diffi- applications, from the quantification and management of
cult to compare these results with the ones presented in risks, to asset allocation, option pricing, macroeconomic
the rest of this section. Nevertheless, we believe that in- modeling, algorithmic trading, lending and more [554].
creasing the variable-to-qubit-ratio beyond one is a very The goal of this section is to present a variety of chal-
interesting direction to scale towards relevant problems. lenging optimization problems from the financial realm
To conclude, these demonstrations build a foundation that may provide interesting use cases for quantum algo-
for future quantum optimization benchmarks. They also rithmic development, as the technology advances.
highlight the difficulty as well as importance to agree on Firstly, we give a short overview of the field and a dis-
common metrics to achieve comparable and fair bench- cussion of advancements in quantitative financial theory
marks. This section is a first attempt for a coordinated pertaining to optimization. We provide a few example
effort to achieve this goal. problems and a discussion of their practical limitations in
Sec. VII A 1. Then, we present the current state of quan-
tum optimization and early investigations of use cases
VII. ILLUSTRATIVE APPLICATIONS in finance in Sec. VII A 2. Next, we provide a deep-dive
into a particularly hard problem with many unresolved
Identifying optimization problems that are promising questions and evident practical relevance: the asset al-
candidates for a practically relevant quantum advantage location problem and different formulations thereof, with
is challenging. This is due to the many open questions increasing levels of complexity in Sec. VII A 3. Finally,
about future performance of quantum optimization algo- in Sec. VII A 4, we conclude and outline further research
rithms, as well as the heuristic nature of many of them. directions that could act as an exploration ground for
Further, benchmarking problem instances known to be quantum optimization applications in finance.
hard for today’s solvers are often (hand-)crafted. Thus,
even if a quantum advantage could be shown for one,
it would not necessarily generalize to other problem in- 1. Overview
stances. While these benchmarks are still crucial to iden-
tify promising problem structures and track algorithmic The global financial system is the backbone of eco-
progress, we also need to look into real-world problem nomics, embedding into every part of society through
instances. Further, there might be a selection bias, as either direct digital channels, such as monetary trans-
already discussed in Sec. VI D. We are often focusing on actions and contracts, or in the form of indirect physical
models where we have an idea about how to solve them. carriers, such as transferable goods and services. It repre-
New formulations of a problem, for which no solver may sents the world’s largest human-made regulated stochas-
exist yet, are often unconsciously discarded. tic network, hosting trillions of interactions between mar-
Thus, within this section, we discuss two exemplary in- ket makers, traders, consumers, regulatory control and
dustries, i.e., finance and sustainable energy — two fields monitoring entities. The system is powered by a variety
that have many outstanding (optimization) challenges. of distributed computing technologies and operates on
For general discussions on the potential applicability of the shoulders of over 120 years of financial theory, heuris-
quantum algorithms for finance and sustainability we re- tic analysis, and a broad range of mathematical tools and
fer the interested reader to the following resources [548– probabilistic models informing decision making — within
553]. Here, we discuss families of optimization problems regulatory and business constraints. This enables one to
arising in the two industries with increasing model com- manage risk, counter financial crime, derive investment
plexity. Importantly, the goal of the following discus- and business decisions, and service consumer needs with
sion is not to claim that the selected use cases are the personalized experiences [555].
most likely candidates for near-term quantum advantage. Despite a century of progress, many models in regu-
Instead, they should serve as a playground of (possibly lated financial institutions suffer from different degrees
simplified) real-world problems that can help to set the of approximations and model trade-offs due to irregular

42
 VII ILLUSTRATIVE APPLICATIONS

and/or scarce information and the need to balance model simplex methods for linear and quadratic programs, or
complexity, computational feasibility, and regulatory, as branch-and-bound and combinations with cutting-plane
well as business constraints. This is generally driven by methods for integer programs. Nevertheless, these prob-
challenges originating from fundamental unknowns un- lems form an excellent ground to benchmark and develop
derlying the interaction dynamics between market par- quantum optimization algorithms and potentially detect
ticipants [556, 557], data quality anomalies in observables value in specific financial applications.
and the overall fast-paced change of the world. The latter Financial optimization problems of substantial com-
may gradually become a notable challenge for many mod- plexity that face limitations when being approached with
eling trade-offs relied on today, and thus, could influence state-of-the-art solvers may, hence, be interesting sub-
the financial system’s systemic resilience [558–560]. For jects for the exploration of quantum algorithms and are
example, more complex probabilistic drivers are emerging summarized here with a few examples:
from new consumer-banking interactions [561], climate
change and energy transmission risk guidelines [562], to • Dynamic programming problems (see Sec. IV D) for
the economic impact of changing geopolitical dynam- pricing and hedging of derivatives on binomial lat-
ics [563]. This demands the development and analysis tices, or structuring of collateralized mortgage obli-
of new modeling techniques and optimization algorithms gations with maximized profits from issuance;
to better inform decision making under uncertainty. In • Stochastic programming problems (see Sec. IV E)
many finance problems, even small improvements can for minimizing bond portfolio credit risk using
have a significant impact. Although, it is still unclear a Conditional Value-at-Risk measure, or asset-
if and how a potential quantum advantage could be real- liability management maximizing wealth, or retir-
ized for financial optimization problems – even as the un- ing outstanding debt at minimal cost, or creating
derlying technology advances – a thorough understand- a synthetic option strategy to reach desired payoff
ing of the broad spectrum of classical models and their at the end of a planning horizon;
limitations in the financial industry could help guide the
development of quantum optimization approaches for use • Robust programming problems (see Sec. IV F) for
cases that deal with uncertainty in objective functions, optimizing portfolios while taking estimation risk
decision variables or constraints. of input parameters into account, or determining
lower and upper bounds on the price of a security;
While not aiming for rigor or completeness, the fol-
lowing provides a bigger picture and short recap of ad- • Multi-objective programming problems (see
vances and challenges in financial optimization problem Sec. IV G) for dynamic margin-volume bal-
domains for their respective input dependencies. This ancing to price mortgages combining optimal
links to the roots of quantitative financial theory and the revenue, balance sheet and business objectives,
understanding that we cannot explain with confidence or optimizing an investment portfolio of stocks
why the prices of financial assets move, but only attempt combining return, risk diversification, incentivizing
to model how they move. As a result, the modeling decarbonization and Environmental, Social, and
of price dynamics and volatility has become a core fo- Corporate Governance (ESG) [574] objectives.
cal point in finance with many models aiming to cap-
ture its essence: from Brownian motion [556] to per-
fect delta-hedging — assuming no price jumps and fi- 2. Related Work
nancial crashes [564]; capturing volatility clustering phe-
nomena [565] with GARCH [566] and Heston [567] mod- The majority of existing quantum optimization re-
els — assuming volatility fluctuations decay over one search in finance focuses on investigating quantum vari-
timescale; Multifractal Random Walks [568] and Rough ants of linear and quadratic optimization problems and
(Heston) Volatility models [569–571] capturing most styl- simplified (mixed) integer programs. Most publications
ized facts [572] from empirical observations of financial employ a QUBO formulation, combined with a ground
time series, such as heavy-tailed probability distribu- state solver such as QAOA or VQE (see Sec. IV A) that is
tions [573]. This list summarizes just a few milestones executed on numerical simulators, gate-based hardware,
over a century of financial modeling, effectively demon- or annealers, typically for systems with less than 50 vari-
strating that capturing market dynamics with such com- ables. In the following, we present a selection of the
plex input patterns is difficult. corresponding papers.
There are also many optimization problems in finance Brandhofer et al. [575] study portfolio optimization us-
that do not need to account for input uncertainties to ing QAOA for quadratic binary optimization constrained
prove valuable for business, such as linear programming by the number of assets. That is, the presence or absence
problems (see Sec. IV B), quadratic programming prob- of a particular stock in a portfolio has a value of 1 or
lems (see Sec. IV A), and (mixed) integer programming 0 respectively, and the sum of the assets is the value of
problems (see Sec. IV C). However, for most instances the investment, which is a (simplified) budget constraint.
corresponding to these classes, there are established, well- Risk is minimized by having the smallest combined co-
performing classical solvers, such as interior-point and variance between all pairs of stock values as represented

43
 VII ILLUSTRATIVE APPLICATIONS

by a covariance matrix; this minimum ensures portfolio level 3

diversification.
In a numerical study, Hodson et al. [576] compare a level 2

probability space
Quantum Alternating Operator Ansatz to QAOA for level 1
models that incorporate trading and investment con-
straints. Baker and Radha [399] evaluated the solution
quality as functions of qubit number and circuit depth. model
Mugel et al. [577] compare methods for dynamic port- definition
folio optimization on quantum processors using 8 years
of data for 52 assets with quantum annealing, VQE, and
a quantum-inspired optimizer based on Tensor Networks time
using pre-processing to reduce the problem dimension.
Venturelli et al. [578] compare the performance of reverse
Figure 8. An illustration of the optimal asset allocation at
quantum annealing for portfolio optimization problems time t0 (model definition), considering perfect predictors of
formulated as QUBOs with classically executed genetic asset valuation at the end of investment horizon t1 given
algorithms. Constrained QAOA solutions to portfolio op- prior statistics ω1 , ω2 , . . . (level 1), then taking uncertainties
timization are also presented in Herman et al. [265], who of these predictors into account but still with a limited view
use Zeno dynamics with rapid sampling to restrict the until t1 (level 2), before anticipating a multi-period allocation
solutions to a constrained subspace. Giron et al. [579] strategy and respective uncertainties beyond t1 (level 3).
use QAOA to study the problem of collateral optimiza-
tion — the optimal allocation of financial assets to sat-
isfy obligations at minimum cost. In a broader context, a However, we stress that it remains unclear how to sys-
quantum optimized portfolio allocation could feed into a tematically advance towards the complexity and value
risk model evaluated on a quantum computer [580–582]. frontier at which a real-world practical impact can be
Other finance optimization applications that have been realized.
proposed for quantum algorithms include transaction In this section, we introduce the concept of value-
settlements, where the goal is to find the maximum guided levels with examples representing incremental in-
number of cash transactions between multiple parties crease in use case complexity by respective problem for-
given their liquidity [319], and Anti Money Laundering, mulations — for reasons of simplification in the mean-
which matches network motifs on snapshots of transac- variance framework, as illustrated in Fig. 8. The aim is
tion graphs [583]. to systematically explore different challenges of this use
case and possibly inspire the exploration and research of
quantum optimization algorithms in this domain.
3. Deep Dive: Optimal Asset Allocation Model Definition. From the ground up, it needs to
be well understood within which environment we are at-
Among the hardest problems in the financial industry tempting to construct an asset allocation problem. For
is the market adaptive search for an optimal portfolio instance, we obviously cannot comprehend all the mar-
composition with an investment allocation in financial kets with interactions between millions of traders with
assets that maximizes the return of investment within a each a different investment style, behavior, risk limits or
given risk appetite and time frame, while complying to trading frequency. But rather than accepting the entire
possibly other objectives and constraints. Reasons for universe of possibly feasible portfolios S, a preliminary
its complexity relate to the stochastic nature of the fi- steering of the downstream complexity for all other lev-
nancial system and various statistical and computational els can help build up and realistically frame the use case.
challenges as discussed in Sec. VII A 1. It is also a prob- The field of investing is found with very different business
lem where marginal improvements in practical solutions domains, expectations and requirements that yield var-
can have substantial impact. In this context, the term ious degrees of complexity to start with. For example,
solution should be discussed. It is difficult to find a prov- a wealth manager for retail, affluent or high-net-worth
able or universal solution, instead one accepts an outcome individuals may have different requirements for compos-
with different forms of trade-offs to encompass particular ing or rebalancing investment portfolios than say, a large
use case requirements, market environments, asset class asset manager. While this seems rather obvious, it is key
restrictions or investment scenarios. While over the past to formulate the model purpose, and thus, its expected
70 years, both academic researchers and industry practi- performance. In particular, an informed pre-selection of
tioners have created a large repertoire of targeted heuris- diversified asset classes, industry exposures, risk profiles
tic approaches and rigorous mathematical tools, it re- or expected dynamics for a given investment time hori-
mains one of the most hunted challenges in finance [584]. zon helps build a more robust starting point, t0 , for the
This partially explains why the quantum computing com- optimization problem.
munity is exploring this use case in various quantum op- Level 1. At this stage we begin with the simplest
timization for finance publications [575, 576, 585, 586]. problem formulation based on Harry Markowitz’s intro-

44
 VII ILLUSTRATIVE APPLICATIONS

duction of modern portfolio theory in 1952 [587], by for- Level 2. A severe limitation of Level 1 is the ig-
malizing portfolio diversification and optimal return-risk norance of uncertainty in the inputs and instability of
trade-offs with a mean-variance framework. It assumes the optimization [593]. Thus, it is of limited practical
a risk-averse myopic investor in a frictionless, one-period value to derive actual investment decisions. In fact, the
financial market with exact knowledge of parameters cap- optimization is highly sensitive to small changes in the
turing asset price dynamics. The aim is to find an op- estimated expected returns µ and covariance matrix Σ,
timal portfolio x = (x1 , ..., xN ) ∈ RN from permissi- leading to large differences in the resulting optimal port-
ble portfolios S with fractions xi of a fixed budget to folio x. At this stage, the problem formulation is ex-
be fully invested into N pre-selected, non-redundant as- tended to take the stochastic nature of underlying ob-
sets at time t0 , based on estimates of expected returns servables into account. The root challenge comes from
µ ∈ RN and their positive semi-definite covariance ma- the portfolio-specific operating regime of the observation
trix Σ ∈ RN ×N at the end of one investment period t1 . ratio q = N/T given the prior number of empirical (sam-
This can be formulated in three equivalent variants as ple) realizations T for the N considered assets at t0 . In
quadratic programming problems with linear constraints the large limit with q → 0, the sample estimators µ and
and without uncertainty in the parameters: Σ would converge to the true realized (population) µtrue
and Σtrue , respectively. But in reality most asset alloca-
1 ⊤ tion problems suffer from a lack of data (except maybe in
min x Σx
x ∈ RN 2 high-frequency trading) with q > 0. For example, a real-
istic scenario for stocks could be N = 600 with T = 2500
s.t. µ⊤ x ≥ rmin , (30)
of about 10 years of daily returns and, thus, q = 0.24. In
1 x = 1,
⊤
case of a few thousand stocks, the issue of stock lifetime
x ∈ S, and structural evolution of markets over time becomes
relevant, shifting the problem into the large dimensional
where the convex quadratic objective minimizes the port- limit regime with q ∼ 1 and fast diverging estimators
folio risk xT Σx with an expected minimal portfolio re- with q > 1. If one would aim to increase the T statis-
turn rmin . Instead, one can also maximize the expected tics with intra-day observations, one may not capture the
return µT x while bounding the risk xT Σx ≤ γ 2 as a dynamics for lower frequency trading strategies anymore
quadratic constraint, which can be transformed into conic and possibly introduce biases in the resulting optimal x.
form [588]. Another alternative is to maximize a con- This makes the problem very complex, even for a small
cave quadratic utility function µT x − λ/2 xT Σx with the number of assets, and usually forces the industry to sim-
risk-aversion factor λ for steering the return-risk trade- plify S at the level of model definition. In the following,
off and variable focus on tail risk. However, all of these we are going to present several approaches to address
formulations are relatively trivial to solve, especially in this challenge, ordered by increasing levels of complex-
the continuous setting, where a closed-form solution may ity: first we de-bias and reduce the estimation error of
be derived using Lagrangian multipliers [589] considering µ and Σ prior to the optimization, then adding regular-
equality constraints [590] and with Karush-Kuhn-Tucker izing techniques to the optimization itself, before comb-
conditions considering also inequality constraints [591]. ing both. Among various methods, the Black-Litterman
Notably, realistic market conditions are not captured by model [594] could be leveraged to create an expected re-
this model formulation. Nevertheless, this simple model turn prediction by combining actual market returns with
offers excellent benchmarking opportunities to evaluate independent expert beliefs about their future, taking into
advances in quantum technology. In fact, the model com- account, for instance, recent news, economic forecasts,
plexity can also be increased in a controlled manner by analyst ratings, or financial statements [595, 596]. Alter-
adding additional constraints that might match a pre- natively, a robust optimization problem can be formu-
defined investment strategy and, as the term x ∈ S is lated to achieve a stabilizing effect on µ during the opti-
meant to reflect, to further constrain the universe of port- mization itself. For this purpose, we define the ellipsoidal
folio configurations and test with an increasing number uncertainty set Uµ = {µ | (µ − µ0 )⊤ Q−1 2
µ (µ − µ0 ) ≤ κ },
of additional linear constraints that match a pre-defined where Qµ is the covariance matrix of estimation errors in
investment strategy. In practice, for a sufficiently large µ, and κ is the uncertainty aversion defining the width of
number of constraints one often has to rely on numerical the uncertainty. This induces a new optimization prob-
approaches. As discussed in Sec. VII A 2, so far, most lem by replacing µ with:
quantum approaches to this use case employ QUBO for- p
mulations with binary or integer encodings, which result µ̃(x) := argmin µ⊤ ⊤
0 x − κ x Qx. (31)
in an NPO-hard combinatorial problem and constraints µ∈Uµ

that are converted into penalty terms [592] of the ob-

jective function. Since these penalty terms only facili- Even if the resulting improved estimate of µ, denoted
tate approximate encoding of the constraints, additional by µ̃, still substantially deviates from µtrue , finding the
substantial post-processing would be required to achieve minimal out-of-sample risk portfolio for it may still be
reliable results. valuable in practice. The sample covariance matrix Σ for
q ∈ (0, 1] can suffer from small eigenvalues near zero,

45
 VII ILLUSTRATIVE APPLICATIONS

leading to a potential underestimation of the realized return forecasts, cost-favorable positions for trading in
(out-of-sample) risk exposure at t1 . To address this issue, following periods, or foreseeable events impacting risk,
Random Matrix Theory [597–599] provides efficient tools trading volume or liquidity. Although there is a lot of
to de-bias the eigenvalues of Σ such that the resulting Σ e research interest, given the computational hardness of
is closer to Σtrue [600–604]. A factor modeling method to the multi-period domain, most practitioners, as well as
represent the portfolio returns with a small number M of classical enterprise solver suites, cannot effectively tackle
common factors [605] can additionally help to circumvent these use cases [584]. Essentially, the h-period optimiza-
the curse of dimensionality if M ≪ N, T . Furthermore, tion problem can be generally formulated as follows:
one may want to further regularize Σ by adding L1 - or
L2 -norm constraints to the optimization. Arguably one max
xt+1 , . . . , xt+h
E [X (xt+1 , . . . , xt+h ) | Ft ]
of the most important ones is the consideration of a trans- (33)
s.t. x ∈ Θ,
action cost function C(x | x0 ) that estimates the cost for
rebalancing the portfolio x0 at t0 towards the optimal where X is the inter-temporal utility function given the
portfolio x at t1 , taking for instance broker fees and im- filtration F associated to the probability space at the
plementation shortfall into account. This results in the tth period with a set Θ of constraints. If we can as-
following problem formulation: sume that X is separable in time, then the objective
λ ⊤e function can be written as minx {a(x)P+ b(x)} with a
max µ̃⊤ x − C(x | x0 ) − x Σx static (forward-looking) part a(x)P = τ aτ (xτ ) and a
x ∈ RN 2
(32) dynamic (coupling) part b(x) = τ b τ (x τ −1 , xτ ) with
s.t. 1⊤ x + C(x | x0 ) = 1, τ = t + 1, . . . , t + h, and their respective separation of
x ∈ S, constraints becomes x ∈ Θ(a) ∩ Θ(b) . Here, aτ rep-
resents a one-period optimization program such as the
where C determines the complexity of the problem and mean-variance terms discussed in Levels 1 and 2, while
depends on how realistic transaction costs should be rep- bτ serves as an intra-period regularization penalty such
resented. For example, a piecewise linear non-convex as the C(xτ |xτ −1 ) transaction cost functions discussed in
function capturing transaction volume and fixed costs for Level 2. Problem formulations of this form and other
buying and selling an asset can be formulated, or a more multi-objective settings are gaining substantial interest
realistic variant with stronger penalization on turnover in the financial industry, and will play a key role in the
considers costs proportional to the L1 -norm of rebalanc- coming years. Thus, it is interesting to investigate new
ing with C(x | x0 ) = β∥x − x0 ∥1 , where β > 0 denotes a approaches for this use case with quantum technology
cost parameter. While formulations of this problem that and inspire the development of new quantum algorithms.
are based on C employ strong regularization on Σ, de- This concludes our short deep-dive illustration with
pending on the respective q regime set out in the model increasing levels of complexity and selected examples
definition, the resulting x may still be biased and sensi- within the mean-variance framework. But notably, this
tive to estimation errors. This can be further addressed can be extended as our exploration of quantum solutions
with two methods. The first one uses a non-parametric evolve. In particular, the development of new quan-
bootstrap resampling approach [606] and the second one tum optimization solutions should transition from cur-
uses a nested clustering approach [607] that partitions rent Level 1 investigations to Level 2 and 3 challenges,
e by forming clusters of highly-correlated assets. The
Σ and with that drive research towards handling optimiza-
latter approach can help to prevent error propagation of tion problems under uncertainty in order to aim for valu-
intra-cluster noise across clusters. Due to a reduction of able contributions to the financial industry.
dimensionality and better control of noise isolation with
smaller sample sizes, the exploration of near-term varia-
tional quantum algorithms with realistic stock portfolio 4. Outlook
sizes may become possible in this setting.
Level 3. Until this point we have restricted our- The financial industry offers a broad spectrum of hard
selves to a static one-period market environment. At optimization problems and demands the exploration of
this level, we introduce the next complexity layer by new solution approaches to tackle many of today’s trade-
considering market dynamics and temporal dependency. offs. This short condensed review provides a satellite
For example, similar to volatility, correlations fluctuate, view of open optimization challenges relevant to the fi-
evolve with time and jump or even flip signs [608, 609]. nancial industry. A selection of use cases are discussed
However, our understanding of correlation dynamics is along with their classical challenges and how optimiza-
nowhere near our understanding of volatility, which cre- tion problems can be constructed with increasing com-
ates an interesting domain to explore with quantum com- plexity levels. The main challenge remains to effec-
puting. Similarly, this is the case for multi-period op- tively combine both the modeling and optimization of
timization, which could help capture inter-temporal ef- problems to achieve practically relevant results for in-
fects, such as accounting for time-varying constraints, dustry use cases. However, there are many more di-

46
 VII ILLUSTRATIVE APPLICATIONS

rections to investigate, such as handling skewed, non- cost and scalability. Additionally, power grid digitiza-
Gaussian heavy tailed probability distributions, Bayesian tion introduces multiple challenges including cybersecu-
approaches and industry sector or subgroup specific pri- rity and data privacy. Adapting grid systems to this new
ors, or non-Markovian decision processes. All of these paradigm will necessitate substantial investment and in-
aim to inspire and guide the exploration of research, novative solutions, potentially surpassing the capabilities
development and benchmarking of potential industry- of classical computer systems.
specific quantum optimization applications.
In this section, we discuss three types of optimization
problems in power grids. We start with a brief overview
B. Sustainable Energy Transition of the area in Sec. VII B 1 and related quantum opti-
mization work, in Sec. VII B 2. In Sec. VII B 3, we in-
troduce electric mobility (e-Mobility) and present three
Sustainability describes the goal of a responsible use of families of e-Mobility use cases and how they connect to
resources to guarantee a long-term existence of humanity the problem classes and algorithms in Sec. IV. Finally,
on earth in a healthy environment. While sustainability Sec. VII B 4 concludes and outlines potential research di-
has multiple dimensions, a common focus is put on envi- rections.
ronmental problems, such as countering climate change.
For this paper, we focus on a subsection of sustainability,
that of sustainable transition of the energy sector. It is
a particularly relevant area as it is rich with optimiza-
tion problems with societal benefit. We discuss multiple
problems in this domain that can serve as illustrative test
cases with different challenges to benchmark optimiza- 2. Related Work
tion algorithms. The use cases described here serve as
an interesting playground to generate examples derived
from real-world data to study corresponding quantum Research in quantum optimization for power grids is
optimization algorithms. While they might not necessar- currently in its early stages, with a predominant focus
ily be candidates for a practical quantum advantage in on the unit commitment problem (UCP) [311, 613]. Ko-
near future, they can help to improve our understanding retsky et al. [614] consider the UCP for power networks
about problem characteristics where quantum optimiza- by combining quantum and classical methods: QAOA
tion may have some potential over classical algorithms handles the binary on-off variables for each unit and a
and where not. classical optimizer handles the continuous variables for
how much power each unit should provide. Takahashi
et al. [615] writes the UCP as a QUBO problem for net-
1. Power Grid Overview work switches equal to 0 or 1 if open or closed, subject to
connection, voltage, and maximum current constraints.
They model the constraints as penalties in the objective
For decades, society has relied on carbon-intense and function and solve the problem with annealing. Simi-
centralized energy sources [610]. Few powerful genera- larly, Halffmann et al. [616] and Braun et al. [617] for-
tion sites were responsible for meeting energy demands, mulate the UCP as a QUBO with penalty terms. Mahroo
and adjusting energy generation was relatively simple and Kargarian [618] decompose the UCP into a sequence
with fossil fuels. However, the energy sector is changing of quadratic continuous and binary (unconstrained) sub-
rapidly. Technological advancements and climate-change problems, similarly to ADMM [316, 317], cf. Sec. IV C.
awareness have led to an increase in renewable energy Other non-UCP quantum optimization works in power
sources. Power grid digitization [611, 612] has resulted in grid include [619, 620], which formulate a power flow
greater connectivity and interactions between customers, problem as a QUBO with constraints encoded as penal-
producers, utilities, and grid operators. This has revolu- ties in the objective, and [621], which addresses an energy
tionized how we manage and optimize energy resources supply scheduling problem with storage, where the opti-
by enabling real-time monitoring, data-driven decision- mization is formulated as a constrained quadratic pro-
making, and enhanced control over energy consumption gram and solved with annealing.
and distribution.
However, the transition to renewable energy sources All of the works mentioned above have in common
and the digitization of the power grid are not without that they focus on small, illustrative problems and that
challenges. Renewable energy sources such as solar and they explore simple mathematical models (mostly QU-
wind are intermittent and probabilistic in nature, and, BOs) that do not capture the complexity of real-world
they require the coordination of a large number of as- use cases. Although these works may not have a direct
sets; for example, to match energy supply with demand. practical impact yet, they could help to pave the way
Energy storage solutions are essential for managing the for advancing the corresponding theory, problem formu-
variability associated with renewable energy production, lations, algorithms, and to understand practical require-
but they present their own set of challenges, including ments on optimization solutions to have practical impact.

47
 VII ILLUSTRATIVE APPLICATIONS

3. Leveraging the Flexibility of e-Mobility

As outlined in Sec. VII B 1, the world is progressively

adopting carbon-neutral energy generation methods [622, 10 kWh

EVs
5 kWh
623]. While renewable sources like wind and solar power 1 kWh
play a vital role, they often produce energy when demand 0 kWh

is low. To address this disparity, it is crucial to invest

in energy storage solutions, such as batteries, and the 8 am time 10 pm
optimized coordination thereof.
Electric vehicles (EVs) are an energy storage solution
of increasing importance. In 2022, the number of EVs Figure 9. Illustrating a possible instance for the problem of
scheduling the charging of EVs in a parking site (use case 1).
on the road exceeded 26 million [624, pp. 14], and the
global fleet of EVs is expected to grow to about 240 mil-
lion in 2030 — 10% of the road vehicle fleet [624, pp.
109]. To put this opportunity in perspective, EV batter- the number of EVs and K the total number of intervals
ies can typically store 20 to 100 kWh and can have an (e.g., 24 intervals, where each interval corresponds to an
output power of 10 kWh when connected to a Type 2 hour). Next, let un ∈ {0, 1}K be a vector that indicates
charger. Thus, 400 EVs, in principle, can deliver 4 MWh the (known) presence of EV n ∈ {1, . . . , N } at time k ∈
of energy to the grid, which is approximately the same {0, . . . , K−1}, and vn ∈ RK be the value that the parking
output power of an energy storage substation [625]. site will obtain by admitting EV n (e.g., dollars). Thus,
the inner product vnT un denotes the total value for the
To leverage e-Mobility’s potential to stabilize the en-
parking site if admitting EV n. Also, let dn ∈ RK , where
ergy network and balance supply and demand, it is cru-
the k-th entry indicates the energy requested by EV n at
cial to address three challenges: The first challenge is
time k. The parking site has a maximum input power of
to increase the EVs’ penetration, for example, by facil-
E kWh and has a maximum capacity of M EVs. Then,
itating access to charging infrastructure [626]. This is
the optimization problem is the following:
particularly important in densely populated areas where
overnight charging may only be available for some EVs. N
X
The second challenge is to effectively control how EVs max xn (vnT un )
store energy [627]. Unlike conventional energy storage x ∈ {0, 1}N n=1
systems, EVs consume energy from the grid, and their N
batteries must be sufficiently charged by strict deadlines. X
s.t. xn un,k ≤ M, ∀k, (34)
The final challenge is to integrate EVs into the power grid
n=1
infrastructure [627], enabling energy retailers to store en- N
ergy in EVs that they can then inject back into the grid, X
xn dn,k un,k ≤ E, ∀k,
i.e., an EV acts as a “virtual power plant.”
n=1
Next, we discuss three optimization-related e-Mobility
use cases, each addressing one of the challenges men- where xn denotes the decision variable of whether to ac-
tioned earlier. We start with a “core” mathematical prob- cept EV n to a parking space where it can be plugged in
lem and then extend this to a family of problems with or not. That is, Eq. (34) is a binary optimization with
increasing practical relevance, linking it to the problem 2K knapsack constraints; K knapsacks for the parking
classes discussed in Sec. IV. spaces, and K knapsacks for the energy available in the
Use case 1: Scheduling EVs in a parking site. parking site.
This use case consists of allocating EVs to parking spaces Some instances of the multi-dimensional knapsack
in a parking site [628, 629]. In short, consider a parking problems can be challenging to solve, even when the prob-
site that has two types of parking spaces: regular park- lem instances are relatively small, cf. Sec. VI D. Thus,
ing spaces and parking spaces where EVs can be plugged studying problems of the form in Eq. (34) can be of inter-
in. EVs arrive at the parking site indicating the kWh est for near-term quantum optimization algorithms with
they want to replenish at each time, e.g., in each hour. a potential practical application. Of course, we need
With that information, the parking site administrator de- to benchmark specific practically relevant problem in-
cides where to allocate the EVs to maximize the charging stances to assess whether they are difficult classically. If
volumes and thereby profit. To simplify the problem for- the instances prove difficult, we can proceed to explore
mulation even further, let us assume that the EVs arrival whether there is an opportunity for a practically rele-
and departure times are known in advance. Figure 9 vant quantum advantage. Since reality and real world
shows the problem schematically, where the y-axis in the scenarios are often more delicate, there will likely be
figure represents the EVs. The goal is to select EVs such additional constraints and cost function adjustments to
that the number of EVs at a given time does not exceed Eq. (34) thereby increasing its general difficulty for clas-
the number of chargers and the input power of the site. sical solvers. Some interesting problem extensions for
We can formalize the problem as follows. Let N be Eq. (34) towards practically relevant applications are the

48
 VII ILLUSTRATIVE APPLICATIONS

following: at time k. In each time slot k, the controller selects an

action from the action set A(xk ) ⊂ RN containing the
1. The power available in a parking site may vary over possible input/output powers of each EV, e.g., ±10 kWh.
time; for example, when the parking site shares its The action set at time k depends on the state xk since,
connection to the grid with another infrastructure for example, it is not possible to discharge an EV if its
or a local renewable energy source is considered, battery is empty. The choice of action ak ∈ A(xk ) affects
such as a solar power plant. the EVs’ state of charge in the next time slot k + 1. In
2. The decision of admitting/rejecting EVs is carried particular, xk+1 = F (xk , ak ) where F is a function that
out via a pricing policy that affects the willingness deterministically maps a state xk given an action ak to
of the drivers to use the parking site. the next state xk+1 . Next, let Tk (xk , ak ) be the cost
obtained by selecting action ak in state xk at time k.
3. Increase the EV charging flexibility. EVs can have The resulting optimization problem is given as following:
a desired state of charge at one or multiple points
K−1
X
in time (e.g., at least 60 kWh by 6 am).
min Tk (xk , ak )
4. The EV arrivals and departure times may not be ak ∈ A(xk ) k=0 (35)
known in advance, as described at the beginning of s.t. x K = x′ ,
the section. Additionally, the forecast of renewable xk+1 = F (xk , ak ),
energies includes uncertainties if considered in (1).
The original problem in Eq. (34) can be modeled as where x0 ∈ RN ≥0 denotes the EVs’ initial state of charge
a QUBO. The original objective function is linear, i.e., and x′ the state that all EVs are fully charged. Thus, the
the corresponding cost matrix will be a diagonal matrix, goal is to select actions ak ∈ A(xk ) to minimize the sum
cf. Sec. IV A 1, and the constraints can, for instance, be of the costs given that xk+1 = F (xk , ak ) and the final
encoded as penalties into the objective, cf. Sec. IV A 1, state, xK , is equal to x′ . We have modeled the problem
which, however, will immediately imply a dense cost as a discrete optimal control problem, cf. Sec. IV E, but
matrix. The extensions (1)-(4) affect the difficulty of other formulations are possible as well. For instance, we
Eq. (34) in different ways: (1) imposes additional con- can formulate it as a MIP with K + 1 constraints for the
straints that may result in an increasing problems size, dynamics, plus the constraints for enforcing feasibility of
which may result in more qubits or having to encode actions. However, unlike the previous use case, Eq. (35)
more binary variables per qubit. (2) will imply adding is a less promising problem candidate for demonstrating
continuous variables and results in a MIP. (3) adds flex- industrial quantum advantage in the short term. The
ibility to how EVs charge and discharge, which may fur- problem formulation can be mapped to a UCP (via DP)
ther increase the number of decision variables. Finally, [613], which commercial solvers can solve efficiently with
(4) introduces uncertainty to the optimization problem, hundreds of units [630, Sec. 6]. Nonetheless, we can add
potentially necessitating a shift in the formulation from features to Eq. (35) to make it more difficult to solve
offline optimization to online or stochastic optimization classically, and at the same time, closer to real-world sce-
such as MDP, cf. Sec. IV E. narios:
Use case 2: Charging and discharging EVs. Con- 1. Energy constraints that affect how an EV or group
sider the scenario where a fleet of EVs are plugged in of EVs can charge or discharge at a given time.
from 6pm to 6am the next day, e.g., delivery vehicles.
The fleet owner allows the energy supplier to use the 2. EVs arriving and departing at different (stochastic)
EVs as energy storage subject to the constraint that the times instead of being parked overnight.
EVs are fully charged at departure time (i.e., by 6am the
3. Stochastic cost function that captures, for example,
next day). The goal of the energy supplier is to pro-
the uncertainty in the electricity demand.
vide electricity to the end customers at minimum cost.
In particular, the energy supplier would like to store en- We can approach the problem in Eq. (35), for instance,
ergy in the EVs when this is cheap, and use the energy with DP. As mentioned in Sec. IV D, DP requires search-
in the EVs to power the grid when energy production is ing over all possible states and actions, which is often
low or expensive to buy from the energy markets. The exponential in size. One approach to speedup DP is us-
problem is challenging because the energy demand and ing Grover’s search algorithm as a subroutine. Similar to
energy prices may vary over time. [337] for the TSP, the quantum algorithm could first use
We can formulate the problem above as a deterministic DP to compute multiple paths from state x0 to xK = s′
discrete-time control problem, cf. Sec. IV E. In short, we and then use Grover’s search to find a combination of
divide time in slots k ∈ {0, . . . , K − 1} of equal duration, those paths that yields lower cost. Nonetheless, such an
where K defines the total considered time horizon. The approach is unlikely to yield practical speedups in the
duration of a slot can be, for example, 1 hour, and the short term since Grover’s search requires FTQC and a
horizon 12 hours. Next, let N be the number of EVs quadratic speedup might not be sufficient for the con-
and xk ∈ RN ≥0 represent the state of charge of the EVs sidered problems and scales. An alternative is to model

49
 VII ILLUSTRATIVE APPLICATIONS

the problem as a MIP, and try algorithms discussed in energy, and the expected cost of buying expensive energy.
Sec. IV C. Note that [Dk − ek − Bk ]+ captures the energy demand
The extensions (1)-(3) add different complications to that cannot be met with the energy bought (i.e., ek ) and
Eq. (35): (1-2) add constraints to enforce certain target the energy in storage (i.e., Bk ).
states at given times or feasible actions, e.g., minimum The optimization in Eq. (36) is convex since the ob-
constant charge or discharge times, which affects both, jective function is piece-wise linear and convex. Thus, if
DP and MIP formulations as shown for uni-directional Dk , Bk , and Pkhigh are random variables independent of
charging by Federer et al. [631]. In case of stochastic ar- the decision variables, we can efficiently solve the prob-
rival/departure times, (2) will require more complex ap- lem with conventional convex optimization tools, in fact,
proaches, e.g. MDP, to handle the uncertainty. Similar by solving K separate problems — one for every term in
to the stochastic case in (2), (3) will require approaches the sum in Eq. (36). Hence, Eq. (36), in its current form,
like MDP to compute a policy that minimizes the costs is not a promising problem candidate for demonstrating
under uncertainty. MDPs quickly become very difficult quantum advantage. Yet, we can make it more challeng-
to solve due to the curse of dimensionality, cf. Sec. IV E. ing and closer to real-world problems with the following
Thus, this might be an interesting domain for first small three extensions:
test cases to analyze the potential of corresponding quan-
tum approaches. Since the result of MDP is a policy, the 1. The energy available in storage at time k depends
optimization and inference are split, which may allow one on the previous decisions taken; for example, on
to relax requirements on the time-to-solution. how EVs charge and discharge, as in the previous
use case 2.
Use case 3: Energy retailing with storage. Con-
sider an energy retailer that delivers energy to end cus- 2. Convex or concave cost functions, i.e., the cost to
tomers. The energy supplier buys energy from energy buy energy might depend on the quantity ordered.
producers in advance, e.g., on the day-ahead energy trad-
ing market, in one-hour time intervals, where the elec- 3. The energy supply has to meet the energy demand.
tricity price depends on several factors such as energy
demand, renewable energy generation, etc. Additionally, The extensions add different layers of complexity. In
the energy supplier has access to an energy storage sys- particular, (1) couples the random variables with the de-
tem, e.g., a fleet of EVs that act as a virtual power plant, cision variables and “links” the K time steps, i.e., terms in
that can accumulate energy when this is cheap, to later the objective function Eq. (36). This is similar to the sys-
use it when energy demand is high. The energy sup- tem dynamics function in optimal control, cf. Sec. IV E.
plier’s goal is to deliver energy to its customers at the This extension is important since assuming that the en-
lowest possible cost. Thus, it is incentivized to lever- ergy in storage (i.e., Bk ) is independent of the decision
age the flexibility provided by its energy storage assets. variables does not allow us to capture, for example, that
The problem is challenging because the electricity de- the energy retailer buys energy in advance to keep in
mand and the energy available in storage are not exactly storage for later use, i.e., the decisions in the past affect
known in advance. Furthermore, if the energy demand the events in the future. We can recast Eq. (36) as an
is not met, the energy supplier has to buy energy from optimal control problem where the system dynamics are
other retailers that often sell energy at a higher price. given by the algorithm in the second use case, i.e., the en-
We can formulate the problem mathematically as fol- ergy available by how EVs charge and discharge. That is,
lows. Let ek ∈ R≥0 be the decision variable indicat- we are coupling this use case with the previous one, and,
ing the MWh the energy supplier buys in each hour indirectly, inheriting the quantum challenges already dis-
k ∈ {0, 1, . . . , K − 1} in advance, usually with K = 24, cussed. In other words, we are adding the complications
and plow ∈ R the price of the cheap energy (which some- of use case 2 on top of use case 3. (2) adds a constraint
k
times might even be negative!). Similarly, let Dk , Bk , of the form g(Dk − ek − Bk ) ≤ 0, where g is a func-
tion such as CVaR that measures risk. The problem falls
Pkhigh be random variables that capture, respectively, the
within the realm of robust optimization, see Sec. IV F,
electricity demand at time k in MWh, the energy avail-
or chance constraints [632]. In case of (3), depending on
able in storage at time k in MWh, and the price of the
the exact properties of cost functions, the problem may
expensive energy at time k. The optimization problem is
remain convex or not. In case of a non-convex problem,
the following:
this can quickly become very challenging to be solved.
X
K−1 h
high
i
min plow
k e k + E Pk [Dk − ek − Bk ]+
ek ≥ 0 k=0 4. Outlook
(36)
where [·]+ := max{0, ·} and the expectation is with re- Quantum computing has the potential to significantly
spect to Dk , Bk , and Pkhigh . That is, we have a stochastic advance optimization problems related to the sustainable
unconstrained optimization where the objective function energy transition. However, identifying the most promis-
at time k consists of two terms: the cost of buying cheap ing applications is an ongoing challenge. Here, we have

50
 VIII CONCLUSION & OUTLOOK

introduced illustrative examples of optimization prob- encourage advancement in quantum optimization and ad-
lems that may serve as test cases derived from real-world vocate responsible use of the insights presented in this
data. Under certain assumptions on the problem data, work. While we showcase where benefits for optimiza-
some instances may prove to be classically challenging tion algorithms may lie, it is important not to overstate
even with few variables. We have shown how the prob- or misinterpret the results of this paper and its potential
lem classes introduced in Sec. IV relate to these families applications. Such considerations motivate the need to
of use cases with increasing complexity and of increas- establish clear benchmarks, cf. Sec. VI, which enable a re-
ing relevance to real-world applications. The discussion liable interpretation of the scientific insights by a broader
underlines the need for additional research on quantum audience. Moreover, where use case applications of quan-
optimization algorithms to approach the complications tum optimization algorithms are chosen and funded, we
required for practical impact. It further shows that it is encourage the prioritization of those with positive social
crucial to benchmark the problem instances to identify impact, for example, sustainable energy transition as dis-
which ones are truly difficult classically and where there cussed in Sec. VII B.
is room for a potential quantum advantage. In conclusion, quantum optimization holds vast poten-
tial for various applications, but significant challenges re-
main in demonstrating a practical quantum advantage.
VIII. CONCLUSION & OUTLOOK Once a tangible quantum advantage is demonstrated, we
expect that quantum optimization would rapidly influ-
In this paper, we provided a comprehensive overview ence many domains given the widespread relevance of
of the potential, challenges, and emerging research areas optimization. This paper has provided a blueprint to
in quantum optimization. We observed that, while com- define and measure progress in quantum optimization,
plexity theory is useful to guide towards provable perfor- which unquestionably has exciting prospects.
mance guarantees, it may not always be useful for finding
practical quantum advantage. This underscores the need Acknowledgments. The authors thank Jens Eisert and
to develop and analyze quantum optimization heuristics Mark Wilde for their valuable feedback and suggestions
to better understand their effectiveness. We highlighted to further improve this paper.
key problem classes in optimization, reviewed existing Andris Ambainis acknowledges the support of
algorithms and suggested new research directions. Fur- the Latvian Quantum Initiative under European
ther, we discussed the challenges of executing and scal- Union Recovery and Resilience Facility Project
ing these algorithms on noisy hardware and the impor- No. 2.3.1.1.i.0/1/22/I/CFLA/001 and the QuantERA II
tance of benchmarking for identifying quantum advan- ERA-NET Cofund projects QOPT and HQCC. Brandon
tages. Lastly, we explored two illustrative application Augustino and Swati Gupta acknowledge support from
domains, emphasizing the need to expand our focus be- the US Defense Advanced Research Projects Agency
yond problem classes like QUBO to truly impact these (DARPA) Contract No. HR001120C0046. Vedran
fields. Dunjko acknowledges the support of the Dutch Research
As quantum hardware evolves with more qubits, re- Council (NWO/OCW), as part of the Quantum Software
duced errors, and faster circuit execution, a new era is Consortium programme (Project No. 024.003.037), and
emerging where progress in quantum algorithm research of the Dutch National Growth Fund (NGF), as part of
comes from both theoretical analysis and empirical meth- the Quantum Delta NL programme. Stuart Hadfield
ods. While there remains a strong need to discover was supported by NASA Academic Mission Services
new algorithms with provable performance guarantees, Contract No. NNA16BD14C, and by DARPA under
we must also prioritize the development of quantum al- interagency agreement IAA 8839, Annex 114. Thorsten
gorithms with no such guarantees, and benchmark their Koch acknowledges support by the BMBF Research
performance on real quantum computers. The improving Campus MODAL (05M14ZAM, 05M20ZBM). Steve
abilities to test ideas in practice unlocks unprecedented Lenk was funded by Free State of Thuringia (Thüringer
opportunities for the advancement of quantum optimiza- Aufbaubank) through project Quantum Hub Thüringen
tion. Even if it does not immediately imply a practical (2021 FGI 0047), by Bundesministerium für Wirtschaft
quantum advantage, the intuition gained from running und Energie, Germany through the project “EnerQuant”
experiments on real hardware empowers one to quickly (Project No. 03EI1025C), and by the European Union
validate proposals. under Horizon Europe Programme. Views and opinions
In addition to technical advances in quantum optimiza- expressed are however those of the author(s) only and
tion, there is also the question of responsible research do not necessarily reflect those of the European Union
and use of this new technology. There are a variety of or European Climate, Infrastructure and Environment
causes, arising from different applications. Thus, further Executive Agency (CINEA). Neither the European
research should be done with the awareness that opti- Union nor the granting authority can be held responsible
mization use cases occur in a multitude of contexts — for them. Grant Agreement 101080086 - NeQST. Jakub
from those with positive societal impact, to those with Marecek acknowledges the support of the Czech Science
nefarious intent. We hope that open scientific discussions Foundation (23-07947S). Stefano Mensa, Emre Sahin

51
 VIII CONCLUSION & OUTLOOK

and Benjamin Symons work was supported by the support by the National Research Foundation, Singa-
Hartree National Centre for Digital Innovation, a UK pore. The work of Jon Yard was supported in part
Government-funded collaboration between STFC and by the NSERC Discovery under Grant No. RGPIN-
IBM. IBM, the IBM logo, and ibm.com are trademarks 2018-04742, the NSERC project FoQaCiA under Grant
of International Business Machines Corp., registered No. ALLRP-569582-21 and the Perimeter Institute for
in many jurisdictions worldwide. Other product and Theoretical Physics. Research at Perimeter Institute
service names might be trademarks of IBM or other is supported by the Government of Canada through
companies. The current list of IBM trademarks is Innovation, Science and Economic Development Canada
available at https://www.ibm.com/legal/copytrade. and by the Province of Ontario through the Ministry of
Giacomo Nannicini acknowledges support by ONR Research, Innovation and Science. Los Alamos unlimited
(N000142312585). Patrick Rebentrost acknowledges release LA-UR-23-33327.

[1] Lov K. Grover, “A fast quantum mechanical algorithm cuits,” Physical Review Letters 119 (2017).
for database search,” in Proceedings of the Twenty- [15] Ewout van den Berg, Zlatko K. Minev, Abhinav Kan-
Eighth Annual ACM Symposium on Theory of Comput- dala, and Kristan Temme, “Probabilistic error cancel-
ing, STOC ’96 (Association for Computing Machinery, lation with sparse pauli–lindblad models on noisy quan-
New York, NY, USA, 1996) p. 212–219. tum processors,” Nature Physics 19, 1116–1121 (2023).
[2] Tadashi Kadowaki and Hidetoshi Nishimori, “Quantum [16] Youngseok Kim, Andrew Eddins, Sajant Anand,
annealing in the transverse Ising model,” Physical Re- Ken Xuan Wei, Ewout van den Berg, Sami Rosenblatt,
view E 58, 5355 (1998). Hasan Nayfeh, Yantao Wu, Michael Zaletel, Kristan
[3] Edward Farhi, Jeffrey Goldstone, Sam Gutmann, and Temme, and Abhinav Kandala, “Evidence for the util-
Michael Sipser, “Quantum computation by adiabatic ity of quantum computing before fault tolerance,” Na-
evolution,” (2000), arXiv:quant-ph/0001106. ture 618, 500–505 (2023).
[4] Maciej Liśkiewicz and Martin R Schuster, “A new up- [17] P.W. Shor, “Algorithms for quantum computation: dis-
per bound for the traveling salesman problem in cu- crete logarithms and factoring,” in Proceedings 35th An-
bic graphs,” Journal of Discrete Algorithms 27, 1–20 nual Symposium on Foundations of Computer Science
(2014). (1994) pp. 124–134.
[5] David L. Applegate, Robert E. Bixby, Vašek Chvátal, [18] Craig Gidney and Martin Ekerå, “How to factor 2048 bit
and William J. Cook, “On the solution of traveling RSA integers in 8 hours using 20 million noisy qubits,”
salesman problems.” Documenta Mathematica , 645– Quantum 5, 433 (2021).
656 (1998). [19] Sanjeev Arora and Boaz Barak, Computational Com-
[6] David L. Applegate, Robert E. Bixby, Vasek Chvátal, plexity: A Modern Approach (Cambridge University
and William J. Cook, “Finding Tours in the TSP,” Press, 2009).
(1999). [20] John Watrous, “Quantum Computational Complexity,”
[7] William J Cook, William H Cunningham, William R in Encyclopedia of Complexity and Systems Science,
Pulleyblank, and Alexander Schrijver, “Wiley- edited by Robert A. Meyers (Springer New York, New
interscience series in discrete mathematics and opti- York, NY, 2009) pp. 7174–7201.
mization,” in Combinatorial Optimization (John Wiley [21] Ronald de Wolf, “Quantum computing: Lecture notes,”
& Sons, Ltd, 1997) pp. 356–356. (2023), arXiv:1907.09415 [quant-ph].
[8] William J Cook, David L Applegate, Robert E Bixby, [22] Richard Bellman, “Dynamic programming treatment of
and Vasek Chvatal, The traveling salesman problem: a the travelling salesman problem,” J. ACM 9, 61–63
computational study (Princeton university press, 2011). (1962).
[9] William J Cook, In Pursuit of the Traveling Salesman: [23] Michael Held and Richard M. Karp, “A Dynamic Pro-
Mathematics at the Limits of Computation (Princeton gramming Approach to Sequencing Problems,” Journal
university press, 2012). of the Society for Industrial and Applied Mathematics
[10] “3D Star TSPs,” (2023), accessed Dec 3, 2023. 10, 196–210 (1962).
[11] Jakob Puchinger, Günther R. Raidl, and Ulrich Pfer- [24] Andreas Björklund, Petteri Kaski, and Lukasz Kowalik,
schy, “The Multidimensional Knapsack Problem: Struc- “Fast Witness Extraction using a Decision Oracle,” in
ture and Algorithms,” INFORMS J. on Computing 22, Algorithms - ESA 2014/Lecture notes in computer sci-
250–265 (2010). ence, Vol. 8737, edited by Andreas Schulz and Dorothea
[12] Tom Packebusch and Stephan Mertens, “Low autocor- Wagner (Springer, 2014) pp. 149–160.
relation binary sequences,” Journal of Physics A: Math- [25] Giorgio Ausiello, Pierluigi Crescenzi, Giorgio Gam-
ematical and Theoretical 49, 165001 (2016). bosi, Viggo Kann, Alberto Marchetti-Spaccamela, and
[13] Sergey Bravyi, Andrew W. Cross, Jay M. Gambetta, Marco Protasi, Complexity and approximation: Com-
Dmitri Maslov, Patrick Rall, and Theodore J. Yoder, binatorial optimization problems and their approxima-
“High-threshold and low-overhead fault-tolerant quan- bility properties (Springer Science & Business Media,
tum memory,” (2023), arXiv:2308.07915 [quant-ph]. 2012).
[14] Kristan Temme, Sergey Bravyi, and Jay M. Gam- [26] Ethan Bernstein and Umesh Vazirani, “Quantum com-
betta, “Error mitigation for short-depth quantum cir- plexity theory,” SIAM Journal on Computing 26, 1411–

52
 VIII CONCLUSION & OUTLOOK

1473 (1997). Completeness (W. H. Freeman & Co., USA, 1990).

[27] Takashi Yamakawa and Mark Zhandry, “Verifiable [42] R. G. Downey and M. R. Fellows, Parameterized Com-
quantum advantage without structure,” in 2022 IEEE plexity (Springer New York, 1999).
63rd Annual Symposium on Foundations of Computer [43] Liming Cai and Jianer Chen, “On fixed-parameter
Science (FOCS) (IEEE, 2022) pp. 69–74. tractability and approximability of np optimization
[28] Atul Singh Arora, Andrea Coladangelo, Matthew problems,” Journal of Computer and System Sciences
Coudron, Alexandru Gheorghiu, Uttam Singh, and 54, 465–474 (1997).
Hendrik Waldner, “Quantum depth in the random or- [44] Jianer Chen, Iyad A. Kanj, and Ge Xia, “Improved
acle model,” in Proceedings of the 55th Annual ACM upper bounds for vertex cover,” Theoretical Computer
Symposium on Theory of Computing (2023) pp. 1111– Science 411, 3736–3756 (2010).
1124. [45] Michael J. Bremner, Zhengfeng Ji, Ryan L. Mann, Luke
[29] Richard M. Karp, “Reducibility among combinatorial Mathieson, Mauro E. S. Morales, and Alexis T. E.
problems,” in Complexity of Computer Computations Shaw, “Quantum parameterized complexity,” (2022),
(Springer US, 1972) pp. 85–103. arXiv:2203.08002 [quant-ph].
[30] Robert Beals, Harry Buhrman, Richard Cleve, Michele [46] Fedor V Fomin, Daniel Lokshtanov, Saket Saurabh, and
Mosca, and Ronald de Wolf, “Quantum lower bounds Meirav Zehavi, Kernelization: theory of parameterized
by polynomials,” Journal of the ACM (JACM) 48, 778– preprocessing (Cambridge University Press, 2019).
797 (2001). [47] Hans L. Bodlaender, Fedor V. Fomin, Daniel Loksh-
[31] Charles H Bennett, Ethan Bernstein, Gilles Brassard, tanov, Eelko Penninkx, Saket Saurabh, and Dim-
and Umesh Vazirani, “Strengths and weaknesses of itrios M. Thilikos, “(meta) kernelization,” J. ACM 63
quantum computing,” SIAM journal on Computing 26, (2016), 10.1145/2973749.
1510–1523 (1997). [48] David S. Johnson, “Approximation algorithms for com-
[32] Scott Aaronson, “How much structure is needed for binatorial problems,” Journal of Computer and System
huge quantum speedups?” (2022), talk delivered at Sciences 9, 256–278 (1974).
the 28th Solvay Physics Conference, arXiv:2209.06930 [49] Johan Håstad, “Some optimal inapproximability re-
[quant-ph]. sults,” Journal of the ACM 48, 798–859 (2001).
[33] Sevag Gharibian, “The 7 faces of quantum NP,” (2023), [50] Michel X. Goemans and David P. Williamson, “Im-
arXiv:2310.18010 [quant-ph]. proved approximation algorithms for maximum cut
[34] There is a slight technical complication here, because and satisfiability problems using semidefinite program-
we deal with a class that has a promise: either there ming,” J. ACM 42, 1115–1145 (1995).
exists a quantum proof that makes the quantum verifier [51] Subhash Khot, Guy Kindler, Elchanan Mossel, and
accept with high probability or for every proof the veri- Ryan O’Donnell, “Optimal Inapproximability Results
fier rejects with high probability. In the strict sense such for MAX-CUT and Other 2-Variable CSPs?” SIAM
a promise class is not known to contain complete prob- Journal on Computing 37, 319–357 (2007).
lems. In order to still use the notion of completeness a [52] Subhash Khot, “On the power of unique 2-prover 1-
framework of promise problems is introduced [633] for round games,” in 34th Annual ACM Symposium on The-
additional technical details. ory of Computing STOC’02 (2002) pp. 767–775.
[35] Lars S. Madsen, Fabian Laudenbach, Mohsen Fala- [53] Carsten Lund and Mihalis Yannakakis, “On the hard-
marzi. Askarani, Fabien Rortais, Trevor Vincent, Jacob ness of approximating minimization problems,” Journal
F. F. Bulmer, Filippo M. Miatto, Leonhard Neuhaus, of the ACM 41, 960–981 (1994).
Lukas G. Helt, Matthew J. Collins, Adriana E. Lita, [54] Dorit Aharonov, Itai Arad, Zeph Landau, and Umesh
Thomas Gerrits, Sae Woo Nam, Varun D. Vaidya, Mat- Vazirani, “The detectability lemma and quantum gap
teo Menotti, Ish Dhand, Zachary Vernon, Nicolás Que- amplification,” (2008), arXiv:0811.3412 [quant-ph].
sada, and Jonathan Lavoie, “Quantum computational [55] Dorit Aharonov and Tomer Naveh, “Quantum np - a
advantage with a programmable photonic processor,” survey,” (2002), arXiv:quant-ph/0210077 [quant-ph].
Nature 606, 75–81 (2022). [56] Dorit Aharonov, Itai Arad, and Thomas Vidick, “Guest
[36] Chris Marriott and John Watrous, “Quantum column: The quantum pcp conjecture,” SIGACT News
arthur–merlin games,” computational complexity 44, 47–79 (2013).
14, 122–152 (2005). [57] Alexei Yu Kitaev, Alexander Shen, and Mikhail N Vya-
[37] Sanjeev Arora and Boaz Barak, Computational Com- lyi, Classical and quantum computation, Vol. 47 (Amer-
plexity (Cambridge University Press, 2009). ican Mathematical Soc., 2002).
[38] Ran Raz and Avishay Tal, “Oracle separation of bqp [58] Anurag Anshu, Nikolas P. Breuckmann, and Chinmay
and ph,” ACM Journal of the ACM (JACM) 69, 1–21 Nirkhe, “Nlts hamiltonians from good quantum codes,”
(2022). in Proceedings of the 55th Annual ACM Symposium
[39] Sitan Chen, Jordan Cotler, Hsin-Yuan Huang, and on Theory of Computing, STOC 2023 (Association for
Jerry Li, “The complexity of NISQ,” Nature Commu- Computing Machinery, New York, NY, USA, 2023) p.
nications 14, 6001 (2023). 1090–1096.
[40] Harry Buhrman, Subhasree Patro, and Florian Speel- [59] Marek Karpinski, Michael Lampis, and Richard
man, “A framework of quantum strong exponential-time Schmied, “New inapproximability bounds for tsp,” Jour-
hypotheses,” in 38th International symposium on theo- nal of Computer and System Sciences 81, 1665–1677
retical aspects of computer science (stacs 2021) (Schloss (2015).
Dagstuhl-Leibniz-Zentrum für Informatik, 2021). [60] Anna R. Karlin, Nathan Klein, and Shayan Oveis Gha-
[41] Michael R. Garey and David S. Johnson, Comput- ran, “A (slightly) improved approximation algorithm for
ers and Intractability; A Guide to the Theory of NP- metric tsp,” in 53rd Annual ACM SIGACT Symposium

53
 VIII CONCLUSION & OUTLOOK

on Theory of Computing STOC’21 (2021) pp. 32–45. Wiebe, “Entanglement-induced barren plateaus,” PRX
[61] Arthur Lee and Bruce Xu, “Classifying approximation Quantum 2, 040316 (2021).
algorithms: Understanding the apx complexity class,” [77] Samson Wang, Enrico Fontana, M. Cerezo, Kunal
(2021), arXiv:2111.01551 [cs.CC]. Sharma, Akira Sone, Lukasz Cincio, and Patrick J.
[62] Teofilo F. Gonzalez, ed., Handbook of Approximation Coles, “Noise-induced barren plateaus in variational
Algorithms and Metaheuristics: Methologies and Tra- quantum algorithms,” Nature Communications 12
ditional Applications, Vol. 1 (CRC Press, 2018). (2021).
[63] Sanjeev Arora, “Polynomial time approximation [78] John Napp, “Quantifying the barren plateau phe-
schemes for euclidean traveling salesman and other nomenon for a model of unstructured variational an-
geometric problems,” Journal of the ACM (JACM) 45, sätze,” (2022), arXiv:2203.06174 [quant-ph].
753–782 (1998). [79] A V Uvarov and J D Biamonte, “On barren plateaus
[64] Daniel A Spielman and Shang-Hua Teng, “Smoothed and cost function locality in variational quantum algo-
analysis: an attempt to explain the behavior of algo- rithms,” Journal of Physics A: Mathematical and The-
rithms in practice,” Communications of the ACM 52, oretical 54, 245301 (2021).
76–84 (2009). [80] Alistair Letcher, Stefan Woerner, and Christa Zoufal,
[65] Andrej Bogdanov, Luca Trevisan, et al., “Average-case “From tight gradient bounds for parameterized quan-
complexity,” Foundations and Trends® in Theoretical tum circuits to the absence of barren plateaus in qgans,”
Computer Science 2, 1–106 (2006). (2023), arXiv:2309.12681 [quant-ph].
[66] Niklas Pirnay, Vincent Ulitzsch, Frederik Wilde, Jens [81] Christos H. Papadimitriou and Mihalis Yannakakis,
Eisert, and Jean-Pierre Seifert, “An in-principle “Optimization, approximation, and complexity classes,”
super-polynomial quantum advantage for approximat- Journal of Computer and System Sciences 43, 425–440
ing combinatorial optimization problems,” (2022), (1991).
arXiv:2212.08678. [82] Mark W. Krentel, “The complexity of optimization
[67] Mario Szegedy, “Quantum advantage for combina- problems,” Journal of Computer and System Sciences
torial optimization problems, Simplified,” (2022), 36, 490–509 (1988).
arXiv:2212.12572 [quant-ph]. [83] Martin Dyer and Leen Stougie, “Computational com-
[68] Marshall Ball, Alon Rosen, Manuel Sabin, and plexity of stochastic programming problems,” Mathe-
Prashant Nalini Vasudevan, “Average-case fine-grained matical Programming 106, 423–432 (2006).
hardness,” in Proceedings of the 49th Annual ACM [84] Leslie G Valiant, “The complexity of computing the
SIGACT Symposium on Theory of Computing (2017) permanent,” Theoretical computer science 8, 189–201
pp. 483–496. (1979).
[69] Kaining Zhang, Liu Liu, Min-Hsiu Hsieh, and Dacheng [85] Christos H. Papadimitriou, “Games against nature,”
Tao, “Escaping from the barren plateau via gaussian ini- Journal of Computer and System Sciences 31, 288–301
tializations in deep variational quantum circuits,” Ad- (1985).
vances in Neural Information Processing Systems 35, [86] Yunseok Lee, Holger Boche, and Gitta Ku-
18612–18627 (2022). tyniok, “Computability of optimizers,” (2023),
[70] Enrico Fontana, Dylan Herman, Shouvanik arXiv:2301.06148 [math.OC].
Chakrabarti, Niraj Kumar, Romina Yalovetzky, [87] Lenore Blum, Felipe Cucker, Mike Shub, and Steve
Jamie Heredge, Shree Hari Sureshbabu, and Marco Smale, “Complexity and real computation: A mani-
Pistoia, “The adjoint is all you need: Characteriz- festo,” International Journal of Bifurcation and Chaos
ing barren plateaus in quantum ansätze,” (2023), 6, 3–26 (1996).
arXiv:2309.07902 [quant-ph]. [88] Jessica Lemieux, Bettina Heim, David Poulin, Krysta
[71] Michael Ragone, Bojko N. Bakalov, Frédéric Sauvage, Svore, and Matthias Troyer, “Efficient Quantum Walk
Alexander F. Kemper, Carlos Ortiz Marrero, Martin Circuits for Metropolis-Hastings Algorithm,” Quantum
Larocca, and M. Cerezo, “A unified theory of barren 4, 287 (2020).
plateaus for deep parametrized quantum circuits,” arXiv [89] David Layden, Guglielmo Mazzola, Ryan V Mish-
preprint - arXiv:2309.09342 (2023). mash, Mario Motta, Pawel Wocjan, Jin-Sung Kim,
[72] Jarrod McClean, Sergio Boixo, Vadim N. Smelyanskiy, and Sarah Sheldon, “Quantum-enhanced Markov chain
Ryan Babbush, and Hartmut Neven, “Barren plateaus Monte Carlo,” Nature 619, 282–287 (2023).
in quantum neural network training landscapes,” Nature [90] Philip Intallura, Georgios Korpas, Sudeepto
Communications 9 (2018). Chakraborty, Vyacheslav Kungurtsev, and Jakub
[73] M. Cerezo, Akira Sone, Tyler Volkoff, Lukasz Cincio, Marecek, “A survey of quantum alternatives to ran-
and Patrick J. Coles, “Cost function dependent barren domized algorithms: Monte carlo integration and
plateaus in shallow parametrized quantum circuits,” Na- beyond,” (2023), arXiv:2303.04945 [quant-ph].
ture Communications 12 (2021), 10.1038/s41467-021- [91] Sami Boulebnane and Ashley Montanaro, “Solving
21728-w. boolean satisfiability problems with the quantum ap-
[74] M Cerezo and Patrick J Coles, “Higher order deriva- proximate optimization algorithm,” arXiv preprint
tives of quantum neural networks with barren plateaus,” arXiv:2208.06909 (2022).
Quantum Science and Technology 6 (2021). [92] David Gamarnik, “The overlap gap property: A topo-
[75] Zoë Holmes, Kunal Sharma, M. Cerezo, and Patrick J. logical barrier to optimizing over random structures,”
Coles, “Connecting ansatz expressibility to gradient Proceedings of the National Academy of Sciences 118,
magnitudes and barren plateaus,” PRX Quantum 3 e2108492118 (2021).
(2022), 10.1103/PRXQuantum.3.010313. [93] David Gamarnik, Aukosh Jagannath, and Alexander S
[76] Carlos Ortiz Marrero, Mária Kieferová, and Nathan Wein, “Hardness of random optimization problems for

54
 VIII CONCLUSION & OUTLOOK

boolean circuits, low-degree polynomials, and langevin ical Physics 48 (2007).

dynamics,” arXiv preprint arXiv:2004.12063 (2020). [108] Daniel A Lidar, Ali T Rezakhani, and Alioscia Hamma,
[94] Brice Huang and Mark Sellke, “Tight lipschitz hardness “Adiabatic approximation with exponential accuracy for
for optimizing mean field spin glasses,” in 2022 IEEE many-body systems and quantum computation,” Jour-
63rd Annual Symposium on Foundations of Computer nal of Mathematical Physics 50 (2009).
Science (FOCS) (IEEE, 2022) pp. 312–322. [109] Donny Cheung, Peter Høyer, and Nathan Wiebe, “Im-
[95] Joao Basso, David Gamarnik, Song Mei, and Leo Zhou, proved error bounds for the adiabatic approximation,”
“Performance and limitations of the qaoa at constant Journal of Physics A: Mathematical and Theoretical 44,
levels on large sparse hypergraphs and spin glass mod- 415302 (2011).
els,” in 2022 IEEE 63rd Annual Symposium on Foun- [110] A.B. Finnila, M.A. Gomez, C. Sebenik, C. Stenson, and
dations of Computer Science (FOCS) (IEEE, 2022) pp. J.D. Doll, “Quantum annealing: A new method for min-
335–343. imizing multidimensional functions,” Chemical Physics
[96] Antares Chen, Neng Huang, and Kunal Marwaha, Letters 219, 343–348 (1994).
“Local algorithms and the failure of log-depth quan- [111] Victor Bapst, Laura Foini, Florent Krzakala, Guilhem
tum advantage on sparse random csps,” arXiv preprint Semerjian, and Francesco Zamponi, “The quantum adi-
arXiv:2310.01563 (2023). abatic algorithm applied to random optimization prob-
[97] Christoph Durr and Peter Hoyer, “A quantum algo- lems: The quantum spin glass perspective,” Physics Re-
rithm for finding the minimum,” (1999), arXiv:quant- ports 523, 127–205 (2013).
ph/9607014 [quant-ph]. [112] Sergey Knysh, “Zero-temperature quantum annealing
[98] D. Bulger, W. P. Baritompa, and G. R. Wood, “Im- bottlenecks in the spin-glass phase,” Nature communi-
plementing pure adaptive search with grover's quantum cations 7, 12370 (2016).
algorithm,” Journal of Optimization Theory and Appli- [113] Sergei V Isakov, Guglielmo Mazzola, Vadim N Smelyan-
cations 116, 517–529 (2003). skiy, Zhang Jiang, Sergio Boixo, Hartmut Neven, and
[99] W. P. Baritompa, D. W. Bulger, and G. R. Wood, Matthias Troyer, “Understanding quantum tunneling
“Grover's quantum algorithm applied to global opti- through quantum monte carlo simulations,” Physical re-
mization,” SIAM Journal on Optimization 15, 1170– view letters 117, 180402 (2016).
1184 (2005). [114] Naeimeh Mohseni, Carlos Navarrete-Benlloch, Tim
[100] Mario Motta, Chong Sun, Adrian TK Tan, Matthew J Byrnes, and Florian Marquardt, “Deep recurrent net-
O’Rourke, Erika Ye, Austin J Minnich, Fernando GSL works predicting the gap evolution in adiabatic quan-
Brandao, and Garnet Kin-Lic Chan, “Determining tum computing,” Quantum 7, 1039 (2023).
eigenstates and thermal states on a quantum com- [115] Troels F. Rønnow, Zhihui Wang, Joshua Job, Sergio
puter using quantum imaginary time evolution,” Nature Boixo, Sergei V. Isakov, David Wecker, John M. Mar-
Physics 16, 205–210 (2020). tinis, Daniel A. Lidar, and Matthias Troyer, “Defining
[101] Austin Gilliam, Stefan Woerner, and Constantin Gonci- and detecting quantum speedup,” Science 345, 420–424
ulea, “Grover adaptive search for constrained polyno- (2014).
mial binary optimization,” Quantum 5, 428 (2021). [116] EJ Crosson and DA Lidar, “Prospects for quantum en-
[102] Charles H. Bennett, Ethan Bernstein, Gilles Brassard, hancement with diabatic quantum annealing,” Nature
and Umesh Vazirani, “Strengths and weaknesses of Reviews Physics 3, 466–489 (2021).
quantum computing,” SIAM Journal on Computing 26, [117] Mustafa Demirplak and Stuart A. Rice, “Adiabatic pop-
1510–1523 (1997). ulation transfer with control fields,” The Journal of
[103] Chris Cade, Marten Folkertsma, Ido Niesen, and Physical Chemistry A 107, 9937–9945 (2003).
Jordi Weggemans, “Quantifying grover speed-ups be- [118] Adolfo del Campo, “Shortcuts to adiabaticity by coun-
yond asymptotic analysis,” Quantum 7, 1133 (2023). terdiabatic driving,” Phys. Rev. Lett. 111, 100502
[104] Ruslan Shaydulin, Changhao Li, Shouvanik (2013).
Chakrabarti, Matthew DeCross, Dylan Herman, [119] Pieter W. Claeys, Mohit Pandey, Dries Sels, and Ana-
Niraj Kumar, Jeffrey Larson, Danylo Lykov, Pierre toli Polkovnikov, “Floquet-engineering counterdiabatic
Minssen, Yue Sun, Yuri Alexeev, Joan M. Dreiling, protocols in quantum many-body systems,” Phys. Rev.
John P. Gaebler, Thomas M. Gatterman, Justin A. Lett. 123, 090602 (2019).
Gerber, Kevin Gilmore, Dan Gresh, Nathan Hewitt, [120] T. Chasseur, L. S. Theis, Y. R. Sanders, D. J. Eg-
Chandler V. Horst, Shaohan Hu, Jacob Johansen, ger, and F. K. Wilhelm, “Engineering adiabaticity at
Mitchell Matheny, Tanner Mengle, Michael Mills, an avoided crossing with optimal control,” Phys. Rev.
Steven A. Moses, Brian Neyenhuis, Peter Siegfried, A 91, 043421 (2015).
Romina Yalovetzky, and Marco Pistoia, “Evidence [121] Hale F Trotter, “On the product of semi-groups of oper-
of scaling advantage for the quantum approximate ators,” Proceedings of the American Mathematical So-
optimization algorithm on a classically intractable ciety 10, 545–551 (1959).
problem,” (2023), arXiv:2308.02342 [quant-ph]. [122] Masuo Suzuki, “Fractal decomposition of exponential
[105] Andrew J Sommese, Charles W Wampler, et al., The operators with applications to many-body theories and
Numerical solution of systems of polynomials arising in Monte Carlo simulations,” Physics Letters A 146, 319–
engineering and science (World Scientific, 2005). 323 (1990).
[106] M. Born and V. Fock, “Beweis des Adiabatensatzes,” [123] Masuo Suzuki, “General theory of fractal path integrals
Zeitschrift fur Physik 51, 165–180 (1928). with applications to many-body theories and statistical
[107] Sabine Jansen, Mary-Beth Ruskai, and Ruedi Seiler, physics,” Journal of Mathematical Physics 32, 400–407
“Bounds for the adiabatic approximation with applica- (1991).
tions to quantum computation,” Journal of Mathemat- [124] Michael A. Nielsen and Isaac L. Chuang, Quantum

55
 VIII CONCLUSION & OUTLOOK

Computation and Quantum Information (Cambridge Robin Kothari, and Rolando D. Somma, “Simulating
University Press, 2000). hamiltonian dynamics with a truncated taylor series,”
[125] Daniel S. Abrams and Seth Lloyd, “Quantum algorithm Phys. Rev. Lett. 114, 090502 (2015).
providing exponential speed increase for finding eigen- [140] Andrew M Childs, Robin Kothari, and Rolando D
values and eigenvectors,” Phys. Rev. Lett. 83, 5162– Somma, “Quantum algorithm for systems of linear equa-
5165 (1999). tions with exponentially improved dependence on pre-
[126] Alán Aspuru-Guzik, Anthony D. Dutoi, Pe- cision,” SIAM Journal on Computing 46, 1920–1950
ter J. Love, and Martin Head-Gordon, (2017).
“Simulated quantum computation of molecu- [141] C. L. Lawson, R. J. Hanson, D. R. Kincaid, and F. T.
lar energies,” Science 309, 1704–1707 (2005), Krogh, “Basic linear algebra subprograms for fortran
https://www.science.org/doi/pdf/10.1126/science.1113479. usage,” ACM Trans. Math. Softw. 5, 308–323 (1979).
[127] Seunghoon Lee, Joonho Lee, Huanchen Zhai, Yu Tong, [142] Guang Hao Low and Isaac L. Chuang, “Optimal hamil-
Alexander M. Dalzell, Ashutosh Kumar, Phillip Helms, tonian simulation by quantum signal processing,” Phys.
Johnnie Gray, Zhi-Hao Cui, Wenyuan Liu, Michael Kas- Rev. Lett. 118, 010501 (2017).
toryano, Ryan Babbush, John Preskill, David R. Reich- [143] András Gilyén, Yuan Su, Guang Hao Low, and Nathan
man, Earl T. Campbell, Edward F. Valeev, Lin Lin, and Wiebe, “Quantum singular value transformation and be-
Garnet Kin-Lic Chan, “Evaluating the evidence for ex- yond: exponential improvements for quantum matrix
ponential quantum advantage in ground-state quantum arithmetics,” in Proceedings of the 51st Annual ACM
chemistry,” Nature Communications 14, 1952 (2023). SIGACT Symposium on Theory of Computing (2019)
[128] P.W. Shor, “Algorithms for quantum computation: dis- pp. 193–204.
crete logarithms and factoring,” in Proceedings 35th An- [144] Stuart Geman and Donald Geman, “Stochastic relax-
nual Symposium on Foundations of Computer Science ation, gibbs distributions, and the bayesian restoration
(1994) pp. 124–134. of images,” IEEE Trans. Pattern Anal. Mach. Intell.
[129] Shuxian Jiang, Keith A. Britt, Alexander J. McCaskey, PAMI-6, 721–741 (1984).
Travis S. Humble, and Sabre Kais, “Quantum anneal- [145] Florent Krzakała, Andrea Montanari, Federico Ricci-
ing for prime factorization,” Scientific Reports 8 (2018), Tersenghi, Guilhem Semerjian, and Lenka Zdeborová,
10.1038/s41598-018-36058-z. “Gibbs states and the set of solutions of random con-
[130] Vivian Phan, Arttu Pönni, Matti Raasakka, and straint satisfaction problems,” Proceedings of the Na-
Ilkka Tittonen, “On quantum factoring using noisy tional Academy of Sciences 104, 10318–10323 (2007).
intermediate scale quantum computers,” (2022), [146] Larry Stockmeyer, “The complexity of approximate
arXiv:2208.07085 [quant-ph]. counting,” in Proceedings of the fifteenth annual ACM
[131] Kyungtaek Jun and Hyunju Lee, “HUBO and QUBO symposium on Theory of computing - STOC ’83 , STOC
models for prime factorization,” Scientific Reports 13 ’83 (ACM Press, 1983).
(2023), 10.1038/s41598-023-36813-x. [147] Daniel Zhang, Jan Lukas Bosse, and Toby Cu-
[132] Niklas Pirnay, Vincent Ulitzsch, Frederik Wilde, Jens bitt, “Dissipative quantum gibbs sampling,” (2023),
Eisert, and Jean-Pierre Seifert, “An in-principle arXiv:2304.04526 [quant-ph].
super-polynomial quantum advantage for approximat- [148] Nicholas Metropolis, Arianna W. Rosenbluth, Mar-
ing combinatorial optimization problems,” (2023), shall N. Rosenbluth, Augusta H. Teller, and Edward
arXiv:2212.08678 [quant-ph]. Teller, “Equation of state calculations by fast comput-
[133] Philippe Suchsland, Francesco Tacchino, Mark H. Fis- ing machines,” The Journal of Chemical Physics 21,
cher, Titus Neupert, Panagiotis Kl. Barkoutsos, and 1087–1092 (1953).
Ivano Tavernelli, “Algorithmic error mitigation scheme [149] W. K. Hastings, “Monte Carlo sampling methods using
for current quantum processors,” Quantum 5, 492 Markov chains and their applications,” Biometrika 57,
(2021). 97–109 (1970).
[134] William Kirby, Mario Motta, and Antonio Mezzacapo, [150] David Poulin and Pawel Wocjan, “Sampling from the
“Exact and efficient lanczos method on a quantum com- thermal quantum gibbs state and evaluating partition
puter,” Quantum 7, 1018 (2023). functions with a quantum computer,” Physical review
[135] Seth Lloyd, “Universal quantum simulators,” Science letters 103, 220502 (2009).
273, 1073–1078 (1996). [151] Dorit Aharonov, Itai Arad, and Thomas Vidick, “The
[136] Andrew M. Childs, Richard Cleve, Enrico Deotto, Ed- quantum pcp conjecture,” (2013), arXiv:1309.7495
ward Farhi, Sam Gutmann, and Daniel A. Spielman, [quant-ph].
“Exponential algorithmic speedup by a quantum walk,” [152] Sergey Bravyi, Anirban Chowdhury, David Gosset, and
in Proceedings of the Thirty-Fifth Annual ACM Sympo- Pawel Wocjan, “Quantum Hamiltonian complexity in
sium on Theory of Computing, STOC ’03 (Association thermal equilibrium,” Nature Physics 18, 1367–1370
for Computing Machinery, New York, NY, USA, 2003) (2022).
p. 59–68. [153] Barbara M Terhal and David P DiVincenzo, “Problem
[137] Dominic W. Berry, Graeme Ahokas, Richard Cleve, of equilibration and the computation of correlation func-
and Barry C. Sanders, “Efficient quantum algorithms tions on a quantum computer,” Physical Review A 61,
for simulating sparse hamiltonians,” Communications in 022301 (2000).
Mathematical Physics 270, 359–371 (2007). [154] K. Temme, T. J. Osborne, K. G. Vollbrecht, D. Poulin,
[138] Aram W. Harrow, Avinatan Hassidim, and Seth Lloyd, and F. Verstraete, “Quantum metropolis sampling,” Na-
“Quantum algorithm for linear systems of equations,” ture 471, 87–90 (2011).
Phys. Rev. Lett. 103, 150502 (2009). [155] Man-Hong Yung and Alán Aspuru-Guzik, “A quan-
[139] Dominic W. Berry, Andrew M. Childs, Richard Cleve, tum–quantum metropolis algorithm,” Proceedings of

56
 VIII CONCLUSION & OUTLOOK

the National Academy of Sciences 109, 754–759 (2012). [171] Srinivasan Arunachalam, Vojtech Havlicek, Giacomo
[156] Cambyse Rouzé, Daniel Stilck França, and Álvaro Nannicini, Kristan Temme, and Pawel Wocjan, “Sim-
M. Alhambra, “Efficient thermalization and universal pler (classical) and faster (quantum) algorithms for
quantum computing with quantum gibbs samplers,” gibbs partition functions,” Quantum 6, 789 (2022).
(2024), arXiv:2403.12691 [quant-ph]. [172] Arjan Cornelissen and Yassine Hamoudi, “A sublinear-
[157] Ainesh Bakshi, Allen Liu, Ankur Moitra, and Ewin time quantum algorithm for approximating partition
Tang, “High-temperature gibbs states are unentangled functions,” in Proceedings of the 2023 Annual ACM-
and efficiently preparable,” (2024), arXiv:2403.16850 SIAM Symposium on Discrete Algorithms (SODA)
[quant-ph]. (SIAM, 2023) pp. 1245–1264.
[158] Jingxiang Wu and Timothy H. Hsieh, “Variational ther- [173] Fernando GS L Brandao, Richard Kueng, and
mal quantum simulation via thermofield double states,” Daniel Stilck França, “Faster quantum and classical
Phys. Rev. Lett. 123, 220502 (2019). sdp approximations for quadratic binary optimization,”
[159] Youle Wang, Guangxi Li, and Xin Wang, “Variational Quantum 6, 625 (2022).
quantum gibbs state preparation with a truncated tay- [174] Brandon Augustino, Giacomo Nannicini, Tamás Ter-
lor series,” Phys. Rev. Appl. 16, 054035 (2021). laky, and Luis Zuluaga, “Solving the semidefinite relax-
[160] Mirko Consiglio, Jacopo Settino, Andrea Giordano, ation of qubos in matrix multiplication time, and faster
Carlo Mastroianni, Francesco Plastina, Salvatore with a quantum computer,” (2023), arXiv:2301.04237
Lorenzo, Sabrina Maniscalco, John Goold, and Tony [quant-ph].
J. G. Apollaro, “Variational gibbs state preparation on [175] Joran van Apeldoorn and András Gilyén, “Quan-
nisq devices,” (2023), arXiv:2303.11276 [quant-ph]. tum algorithms for zero-sum games,” (2019),
[161] Ada Warren, Linghua Zhu, Nicholas J. Mayhall, Ed- arXiv:1904.03180 [quant-ph].
win Barnes, and Sophia E. Economou, “Adaptive varia- [176] Adam Bouland, Yosheb M Getachew, Yujia Jin, Aaron
tional algorithms for quantum gibbs state preparation,” Sidford, and Kevin Tian, “Quantum speedups for zero-
(2022), arXiv:2203.12757 [quant-ph]. sum games via improved dynamic gibbs sampling,” in
[162] Christa Zoufal, Aurélien Lucchi, and Stefan Woerner, International Conference on Machine Learning (PMLR,
“Variational quantum boltzmann machines,” Quantum 2023) pp. 2932–2952.
Machine Intelligence 3, 7 (2021). [177] Minbo Gao, Zhengfeng Ji, Tongyang Li, and Qisheng
[163] Dominik S. Wild, Dries Sels, Hannes Pichler, Cristian Wang, “Logarithmic-regret quantum learning algo-
Zanoci, and Mikhail D. Lukin, “Quantum sampling al- rithms for zero-sum games,” (2023), arXiv:2304.14197
gorithms for near-term devices,” Phys. Rev. Lett. 127, [quant-ph].
100504 (2021). [178] Li Li, Minjie Fan, Marc Coram, Patrick Riley, and Ste-
[164] F. G. S. L. Brandao and K. M. Svore, “Quantum speed- fan Leichenauer, “Quantum optimization with a novel
ups for solving semidefinite programs,” in 2017 IEEE gibbs objective function and ansatz architecture search,”
58th Annual Symposium on Foundations of Computer Phys. Rev. Res. 2, 023074 (2020).
Science (FOCS) (2017) pp. 415–426. [179] Michel X. Goemans and David P. Williamson, “Im-
[165] Joran Van Apeldoorn, András Gilyén, Sander Gribling, proved approximation algorithms for maximum cut
and Ronald de Wolf, “Quantum SDP-solvers: Better and satisfiability problems using semidefinite program-
upper and lower bounds,” Quantum 4, 230 (2020). ming,” J. ACM 42, 1115–1145 (1995).
[166] F. G. S. L. Brandao, Amir Kalev, Tongyang Li, Cedric [180] Taylor L. Patti, Jean Kossaifi, Anima Anandkumar,
Yen-Yu Lin, K. M. Svore, and Xiaodi Wu, “Quan- and Susanne F. Yelin, “Quantum goemans-williamson
tum sdp solvers: Large speed-ups, optimality, and ap- algorithm with the hadamard test and approximate am-
plications to quantum learning,” in 46th International plitude constraints,” Quantum 7, 1057 (2023).
Colloquium on Automata, Languages, and Programming [181] Fernando G.S L. Brandã o, Richard Kueng, and
(ICALP 2019), Vol. 132, edited by Christel Baier, Paola Daniel Stilck França, “Faster quantum and classical
Flocchini, and Stefano Leonardi (Schloss Dagstuhl– SDP approximations for quadratic binary optimiza-
Leibniz-Zentrum fuer Informatik, 2019) p. 27. tion,” Quantum 6, 625 (2022).
[167] Joran van Apeldoorn and András Gilyén, “Improve- [182] Ewin Tang, “A quantum-inspired classical algorithm for
ments in quantum sdp-solving with applications,” in recommendation systems,” in Proceedings of the 51st
Proceedings of 46th International Colloquium on Au- Annual ACM SIGACT Symposium on Theory of Com-
tomata, Languages, and Programming (ICALP 2019 puting (ACM, 2019).
(2019). [183] Nai-Hui Chia, András Gilyén, Han-Hsuan Lin, Seth
[168] Sergey Bravyi, Anirban Chowdhury, David Gosset, Vo- Lloyd, Ewin Tang, and Chunhao Wang, “Quantum-
jtech Havlicek, and Guanyu Zhu, “Quantum complexity Inspired Algorithms for Solving Low-Rank Linear Equa-
of the kronecker coefficients,” (2023), arXiv:2302.11454 tion Systems with Logarithmic Dependence on the Di-
[quant-ph]. mension,” in 31st International Symposium on Algo-
[169] Ashley Montanaro, “Quantum speedup of monte carlo rithms and Computation (ISAAC 2020), Leibniz Inter-
methods,” Proceedings of the Royal Society A: Math- national Proceedings in Informatics (LIPIcs), Vol. 181,
ematical, Physical and Engineering Sciences 471, edited by Yixin Cao, Siu-Wing Cheng, and Minming
20150301 (2015). Li (Schloss Dagstuhl–Leibniz-Zentrum für Informatik,
[170] Aram W Harrow and Annie Y Wei, “Adaptive quantum Dagstuhl, Germany, 2020) pp. 47:1–47:17.
simulated annealing for bayesian inference and estimat- [184] Alberto Peruzzo, Jarrod McClean, Peter Shadbolt,
ing partition functions,” in Proceedings of the Fourteenth Man-Hong Yung, Xiao-Qi Zhou, Peter J. Love, Alán
Annual ACM-SIAM Symposium on Discrete Algorithms Aspuru-Guzik, and Jeremy L. O’Brien, “A variational
(SIAM, 2020) pp. 193–212. eigenvalue solver on a photonic quantum processor,” Na-

57
 VIII CONCLUSION & OUTLOOK

ture Communications 5 (2014), 10.1038/ncomms5213. tional quantum algorithm for unconstrained black box
[185] Edward Farhi, Jeffrey Goldstone, and Sam Gutmann, binary optimization: Application to feature selection,”
“A quantum approximate optimization algorithm,” Quantum 7, 909 (2023).
(2014), arXiv:1411.4028. [201] Andrew Lucas, “Ising formulations of many NP prob-
[186] Jarrod R McClean, Jonathan Romero, Ryan Babbush, lems,” Frontiers in Physics 2 (2014).
and Alán Aspuru-Guzik, “The theory of variational [202] Stuart Hadfield, “On the representation of Boolean and
hybrid quantum-classical algorithms,” New Journal of real functions as Hamiltonians for quantum comput-
Physics 18, 023023 (2016). ing,” ACM Transactions on Quantum Computing 2, 1–
[187] Marco Cerezo, Andrew Arrasmith, Ryan Babbush, Si- 21 (2021).
mon C Benjamin, Suguru Endo, Keisuke Fujii, Jarrod R [203] F. Glover, G. Kochenberger, R. Hennig, and Y. Du,
McClean, Kosuke Mitarai, Xiao Yuan, Lukasz Cincio, “Quantum bridge analytics I: a tutorial on formulating
et al., “Variational quantum algorithms,” Nature Re- and using QUBO models.” Ann Oper Res 314, 141–183
views Physics 3, 625–644 (2021). (2022).
[188] Kishor Bharti, Alba Cervera-Lierta, Thi Ha Kyaw, [204] Pedro Maciel Xavier, Pedro Ripper, Tiago Andrade,
Tobias Haug, Sumner Alperin-Lea, Abhinav Anand, Joaquim Dias Garcia, Nelson Maculan, and David
Matthias Degroote, Hermanni Heimonen, Jakob S E Bernal Neira, “Qubo. jl: A julia ecosystem for
Kottmann, Tim Menke, et al., “Noisy intermediate- quadratic unconstrained binary optimization,” arXiv
scale quantum (nisq) algorithms (2021),” arXiv preprint preprint arXiv:2307.02577 (2023).
arXiv:2101.08448 (2021). [205] Sahar Karimi and Pooya Ronagh, “Practical integer-
[189] Lennart Bittel and Martin Kliesch, “Training variational to-binary mapping for quantum annealers,” Quantum
quantum algorithms is NP-hard,” Physical Review Let- Information Processing 18 (2019), 10.1007/s11128-019-
ters 127 (2021), 10.1103/physrevlett.127.120502. 2213-x.
[190] John Golden, Andreas Bärtschi, Stephan Eidenbenz, [206] Nicolas PD Sawaya, Albert T Schmitz, and Stuart
and Daniel O’Malley, “Numerical Evidence for Expo- Hadfield, “Encoding trade-offs and design toolkits in
nential Speed-up of QAOA over Unstructured Search quantum algorithms for discrete optimization: coloring,
for Approximate Constrained Optimization,” in IEEE routing, scheduling, and other problems,” Quantum 7,
International Conference on Quantum Computing and 1111 (2023).
Engineering QCE’23 (2023) pp. 496–505. [207] JD Biamonte, “Nonperturbative k-body to two-body
[191] Avrim L. Blum and Ronald L. Rivest, “Training a 3- commuting conversion hamiltonians and embedding
node neural network is NP-complete,” Neural Networks problem instances into ising spins,” Physical Review A
5, 117–127 (1992). 77, 052331 (2008).
[192] Sammy Khalife, Hongyu Cheng, and Amitabh Basu, [208] Ryan Babbush, Bryan O’Gorman, and Alán Aspuru-
“Neural networks with linear threshold activations: Guzik, “Resource efficient gadgets for compiling adi-
structure and algorithms,” Mathematical Programming abatic quantum optimization problems,” Annalen der
(2023), 10.1007/s10107-023-02016-5. Physik 525, 877–888 (2013).
[193] Vincent Froese and Christoph Hertrich, “Training neu- [209] Francisco Barahona, “On the computational complex-
ral networks is np-hard in fixed dimension,” (2023), ity of ising spin glass models,” Journal of Physics A:
arXiv:2303.17045 [cs.CC]. Mathematical and General 15, 3241 (1982).
[194] Lennart Binkowski, Gereon Koßmann, Timo Ziegler, [210] A.P. Punnen, The Quadratic Unconstrained Binary Op-
and René Schwonnek, “Elementary proof of qaoa con- timization Problem: Theory, Algorithms, and Applica-
vergence,” (2023), arXiv:2302.04968 [quant-ph]. tions (Springer International Publishing, 2022).
[195] Vyacheslav Kungurtsev, Georgios Korpas, Jakub Mare- [211] Sanjeev Arora, “Polynomial time approximation
cek, and Elton Yechao Zhu, “Iteration complex- schemes for euclidean traveling salesman and other
ity of variational quantum algorithms,” (2023), geometric problems,” Journal of the ACM 45, 753–782
arXiv:2209.10615 [quant-ph]. (1998).
[196] Daniel Mastropietro, Georgios Korpas, Vyacheslav Kun- [212] Seth Lloyd, “Quantum approximate optimization is
gurtsev, and Jakub Marecek, “Fleming-Viot helps computationally universal,” (2018), arXiv:1812.11075
speed up variational quantum algorithms in the pres- [quant-ph].
ence of barren plateaus,” (2023), arXiv:2311.18090 [213] Jonathan Wurtz and Peter Love, “Maxcut quantum ap-
[quant-ph]. proximate optimization algorithm performance guaran-
[197] Daniel J Egger, Jakub Mareček, and Stefan Woerner, tees for,” Phys. Rev. A 103, 042612 (2021).
“Warm-starting quantum optimization,” Quantum 5, [214] Edward Farhi, Jeffrey Goldstone, Sam Gutmann, and
479 (2021). Leo Zhou, “The quantum approximate optimization al-
[198] R. Tate, M. Farhadi, C. Herold, G. Mohler, and gorithm and the Sherrington-Kirkpatrick model at infi-
S. Gupta, “Bridging classical and quantum with SDP nite size,” Quantum 6, 759 (2022).
initialized warm-starts for QAOA,” ACM Transactions [215] Joao Basso, Edward Farhi, Kunal Marwaha, Benjamin
on Quantum Computing 4, 1–39 (2023). Villalonga, and Leo Zhou, “The Quantum Approximate
[199] Yves Crama, Sourour Elloumi, Amélie Lambert, and Optimization Algorithm at High Depth for MaxCut
Elisabeth Rodriguez-Heck, “Quadratization and convex- on Large-Girth Regular Graphs and the Sherrington-
ification in polynomial binary optimization,” (2022), Kirkpatrick Model,” in 17th Conference on the Theory
working paper or preprint. of Quantum Computation, Communication and Cryp-
[200] Christa Zoufal, Ryan V. Mishmash, Nitin Sharma, Ni- tography (TQC 2022), Leibniz International Proceed-
raj Kumar, Aashish Sheshadri, Amol Deshmukh, Noelle ings in Informatics (LIPIcs), Vol. 232, edited by François
Ibrahim, Julien Gacon, and Stefan Woerner, “Varia- Le Gall and Tomoyuki Morimae (Schloss Dagstuhl –

58
 VIII CONCLUSION & OUTLOOK

Leibniz-Zentrum für Informatik, Dagstuhl, Germany, R. T. Sprenkle, R. P. Stutz, M. Swallows, R. I. Tobey,

2022) pp. 7:1–7:21. A. Tran, T. Tran, E. Vogt, C. Volin, J. Walker, A. M.
[216] Leo Zhou, Sheng-Tao Wang, Soonwon Choi, Hannes Zolot, and J. M. Pino, “A race track trapped-ion quan-
Pichler, and Mikhail D. Lukin, “Quantum approximate tum processor,” (2023), arXiv:2305.03828 [quant-ph].
optimization algorithm: Performance, mechanism, and [227] Sergey Bravyi, Alexander Kliesch, Robert Koenig, and
implementation on near-term devices,” Phys. Rev. X 10, Eugene Tang, “Obstacles to variational quantum opti-
021067 (2020). mization from symmetry protection,” Phys. Rev. Lett.
[217] Stefan H Sack and Maksym Serbyn, “Quantum anneal- 125, 260505 (2020).
ing initialization of the quantum approximate optimiza- [228] Sergey Bravyi, Alexander Kliesch, Robert Koenig, and
tion algorithm,” Quantum 5, 491 (2021). Eugene Tang, “Hybrid quantum-classical algorithms for
[218] Giuseppe Scriva, Nikita Astrakhantsev, Sebastiano Pi- approximate graph coloring,” Quantum 6, 678 (2022).
lati, and Guglielmo Mazzola, “Challenges of variational [229] Maxime Dupont, Bram Evert, Mark J. Hodson, Bhu-
quantum optimization with measurement shot noise,” vanesh Sundar, Stephen Jeffrey, Yuki Yamaguchi, Den-
(2023), arXiv:2308.00044 [quant-ph]. nis Feng, Filip B. Maciejewski, Stuart Hadfield, M. So-
[219] Mahabubul Alam, Abdullah Ash-Saki, and Swa- haib Alam, Zhihui Wang, Shon Grabbe, P. Aaron
roop Ghosh, “Accelerating quantum approximate opti- Lott, Eleanor G. Rieffel, Davide Venturelli, and
mization algorithm using machine learning,” (2020), Matthew J. Reagor, “Quantum-enhanced greedy com-
arXiv:2002.01089 [cs.ET]. binatorial optimization solver,” Science Advances 9
[220] Michael Streif and Martin Leib, “Training the quantum (2023), 10.1126/sciadv.adi0487.
approximate optimization algorithm without access to [230] Lucas T. Brady and Stuart Hadfield, “Iterative quantum
a quantum processing unit,” (2019), arXiv:1908.08862 algorithms for maximum independent set: A tale of low-
[quant-ph]. depth quantum algorithms,” (2023), arXiv:2309.13110
[221] Tomislav Begušić, Kasra Hejazi, and Garnet Kin- [quant-ph].
Lic Chan, “Simulating quantum circuit expectation [231] Jernej Rudi Finžgar, Aron Kerschbaumer, Martin J. A.
values by clifford perturbation theory,” (2023), Schuetz, Christian B. Mendl, and Helmut G. Katz-
arXiv:2306.04797 [quant-ph]. graber, “Quantum-informed recursive optimization al-
[222] Fernando G. S. L. Brandao, Michael Broughton, Ed- gorithms,” (2023), arXiv:2308.13607 [quant-ph].
ward Farhi, Sam Gutmann, and Hartmut Neven, “For [232] Yash J Patel, Sofiene Jerbi, Thomas Bäck, and Ve-
fixed control parameters the quantum approximate op- dran Dunjko, “Reinforcement learning assisted recursive
timization algorithm’s objective function value concen- qaoa,” EPJ Quantum Technology 11, 6 (2024).
trates for typical instances,” (2018), arXiv:1812.04170 [233] Edward Farhi, David Gamarnik, and Sam Gutmann,
[quant-ph]. “The quantum approximate optimization algorithm
[223] V. Akshay, D. Rabinovich, E. Campos, and J. Bia- needs to see the whole graph: A typical case,” (2020),
monte, “Parameter concentrations in quantum approx- arXiv:2004.09002 [quant-ph].
imate optimization,” Physical Review A 104 (2021), [234] Kunal Marwaha and Stuart Hadfield, “Bounds on ap-
10.1103/physreva.104.l010401. proximating max k xor with quantum and classical local
[224] Alexey Galda, Xiaoyuan Liu, Danylo Lykov, Yuri algorithms,” Quantum 6, 757 (2022).
Alexeev, and Ilya Safro, “Transferability of optimal [235] Aditi Misra-Spieldenner, Tim Bode, Peter K. Schuh-
qaoa parameters between random graphs,” (2021), macher, Tobias Stollenwerk, Dmitry Bagrets, and
arXiv:2106.07531 [quant-ph]. Frank K. Wilhelm, “Mean-field approximate op-
[225] Sami Boulebnane and Ashley Montanaro, “Predicting timization algorithm,” PRX Quantum 4 (2023),
parameters for the quantum approximate optimiza- 10.1103/prxquantum.4.030335.
tion algorithm for max-cut from the infinite-size limit,” [236] Kostas Blekos, Dean Brand, Andrea Ceschini, Chiao-
(2021), arXiv:2110.10685 [quant-ph]. Hui Chou, Rui-Hao Li, Komal Pandya, and Alessan-
[226] S. A. Moses, C. H. Baldwin, M. S. Allman, R. An- dro Summer, “A review on quantum approximate
cona, L. Ascarrunz, C. Barnes, J. Bartolotta, B. Bjork, optimization algorithm and its variants,” (2023),
P. Blanchard, M. Bohn, J. G. Bohnet, N. C. Brown, arXiv:2306.09198 [quant-ph].
N. Q. Burdick, W. C. Burton, S. L. Campbell, J. P. [237] Bryce Fuller, Charles Hadfield, Jennifer R. Glick,
Campora III au2, C. Carron, J. Chambers, J. W. Chan, Takashi Imamichi, Toshinari Itoko, Richard J. Thomp-
Y. H. Chen, A. Chernoguzov, E. Chertkov, J. Colina, son, Yang Jiao, Marna M. Kagele, Adriana W. Blom-
J. P. Curtis, R. Daniel, M. DeCross, D. Deen, C. De- Schieber, Rudy Raymond, and Antonio Mezzacapo,
laney, J. M. Dreiling, C. T. Ertsgaard, J. Esposito, “Approximate Solutions of Combinatorial Problems
B. Estey, M. Fabrikant, C. Figgatt, C. Foltz, M. Foss- via Quantum Relaxations,” (2021), arXiv:2111.03167
Feig, D. Francois, J. P. Gaebler, T. M. Gatterman, [quant-ph].
C. N. Gilbreth, J. Giles, E. Glynn, A. Hall, A. M. Han- [238] Kosei Teramoto, Rudy Raymond, Eyuri Wakakuwa,
kin, A. Hansen, D. Hayes, B. Higashi, I. M. Hoffman, and Hiroshi Imai, “Quantum-relaxation based opti-
B. Horning, J. J. Hout, R. Jacobs, J. Johansen, L. Jones, mization algorithms: Theoretical extensions,” (2023),
J. Karcz, T. Klein, P. Lauria, P. Lee, D. Liefer, C. Lytle, arXiv:2302.09481 [quant-ph].
S. T. Lu, D. Lucchetti, A. Malm, M. Matheny, B. Math- [239] Taylor L. Patti, Jean Kossaifi, Anima Anandku-
ewson, K. Mayer, D. B. Miller, M. Mills, B. Neyen- mar, and Susanne F. Yelin, “Variational quantum
huis, L. Nugent, S. Olson, J. Parks, G. N. Price, optimization with multi-basis encodings,” (2022),
Z. Price, M. Pugh, A. Ransford, A. P. Reed, C. Ro- arXiv:2106.13304 [quant-ph].
man, M. Rowe, C. Ryan-Anderson, S. Sanders, J. Sed- [240] Sam McArdle, Tyson Jones, Suguru Endo, Ying Li, Si-
lacek, P. Shevchuk, P. Siegfried, T. Skripka, B. Spaun, mon C Benjamin, and Xiao Yuan, “Variational ansatz-

59
 VIII CONCLUSION & OUTLOOK

based quantum simulation of imaginary time evolution,” pp. 15–21.

npj Quantum Information 5, 75 (2019). [258] Stuart Hadfield, Zhihui Wang, Bryan O’Gorman,
[241] Christa Zoufal, “Generative quantum machine learn- Eleanor G Rieffel, Davide Venturelli, and Rupak
ing,” (2021), arXiv:2111.12738 [quant-ph]. Biswas, “From the quantum approximate optimization
[242] Sheng-Hsuan Lin, Rohit Dilip, Andrew G. Green, Adam algorithm to a quantum alternating operator ansatz,”
Smith, and Frank Pollmann, “Real- and Imaginary- Algorithms 12, 34 (2019).
Time Evolution with Compressed Quantum Circuits,” [259] Zhihui Wang, Nicholas C. Rubin, Jason M. Dominy,
PRX Quantum 2, 010342 (2021). and Eleanor G. Rieffel, “xy mixers: Analytical and nu-
[243] Xiao Yuan, Suguru Endo, Qi Zhao, Ying Li, and Si- merical results for the quantum alternating operator
mon C. Benjamin, “Theory of variational quantum sim- ansatz,” Physical Review A 101, 012320 (2020).
ulation,” Quantum 3, 191 (2019). [260] Jeremy Cook, Stephan Eidenbenz, and Andreas
[244] Marcello Benedetti, Mattia Fiorentini, and Michael Bärtschi, “The Quantum Alternating Operator Ansatz
Lubasch, “Hardware-efficient variational quantum algo- on Maximum k-Vertex Cover,” in IEEE International
rithms for time evolution,” Physical Review Research 3, Conference on Quantum Computing and Engineering
033083 (2021). QCE’20 (2020) pp. 83–92.
[245] Lucas Slattery, Benjamin Villalonga, and Bryan K. [261] Andreas Bärtschi and Stephan Eidenbenz, “Grover Mix-
Clark, “Unitary block optimization for variational quan- ers for QAOA: Shifting Complexity from Mixer Design
tum algorithms,” Physical Review Research 4, 023072 to State Preparation,” in IEEE International Confer-
(2022). ence on Quantum Computing and Engineering QCE’20
[246] James Stokes, Josh Izaac, Nathan Killoran, and (2020) pp. 72–82.
Giuseppe Carleo, “Quantum Natural Gradient,” Quan- [262] Nikolaj Moll, Panagiotis Barkoutsos, Lev S Bishop,
tum 4, 269 (2020). Jerry M Chow, Andrew Cross, Daniel J Egger, Ste-
[247] Julien Gacon, Christa Zoufal, Giuseppe Carleo, and fan Filipp, Andreas Fuhrer, Jay M Gambetta, Marc
Stefan Woerner, “Stochastic approximation of varia- Ganzhorn, et al., “Quantum optimization using vari-
tional quantum imaginary time evolution,” (2023), ational algorithms on near-term quantum devices,”
arXiv:2305.07059 [quant-ph]. Quantum Science and Technology 3, 030503 (2018).
[248] Stefano Barison, Filippo Vicentini, and Giuseppe Car- [263] Ryan LaRose, Eleanor Rieffel, and Davide Venturelli,
leo, “An efficient quantum algorithm for the time evolu- “Mixer-phaser ansätze for quantum optimization with
tion of parameterized circuits,” Quantum 5, 512 (2021). hard constraints,” Quantum Machine Intelligence 4, 17
[249] Pauline J. Ollitrault, Alexander Miessen, and Ivano (2022).
Tavernelli, “Molecular quantum dynamics: A quantum [264] Filip B Maciejewski, Stuart Hadfield, Benjamin Hall,
computing perspective,” Accounts of Chemical Research Mark Hodson, Maxime Dupont, Bram Evert, James
54, 4229–4238 (2021). Sud, M Sohaib Alam, Zhihui Wang, Stephen Jeffrey,
[250] Christa Zoufal, David Sutter, and Stefan Woerner, et al., “Design and execution of quantum circuits using
“Error bounds for variational quantum time evolution,” tens of superconducting qubits and thousands of gates
Phys. Rev. Appl. 20, 044059 (2023). for dense ising optimization problems,” arXiv preprint
[251] Panagiotis Kl. Barkoutsos, Giacomo Nannicini, Anton arXiv:2308.12423 (2023), arXiv:2308.12423 [quant-ph].
Robert, Ivano Tavernelli, and Stefan Woerner, “Improv- [265] Dylan Herman, Ruslan Shaydulin, Yue Sun, Shouvanik
ing Variational Quantum Optimization using CVaR,” Chakrabarti, Shaohan Hu, Pierre Minssen, Arthur Rat-
Quantum 4, 256 (2020). tew, Romina Yalovetzky, and Marco Pistoia, “Con-
[252] Samantha V. Barron, Daniel J. Egger, Elijah strained optimization via quantum zeno dynamics,”
Pelofske, Andreas Bärtschi, Stephan Eidenbenz, Communications Physics 6, 219 (2023).
Matthis Lehmkuehler, and Stefan Woerner, “Provable [266] Fred Glover, Gary Kochenberger, Moses Ma, and
bounds for noise-free expectation values computed from Yu Du, “Quantum bridge analytics ii: Qubo-plus, net-
noisy samples,” (2023), arXiv:2312.00733 [quant-ph]. work optimization and combinatorial chaining for asset
[253] Anthony Angone, Xioayuan Liu, Ruslan Shaydulin, exchange,” Annals of Operations Research 314, 185–212
and Ilya Safro, “Hybrid quantum-classical multilevel (2022).
approach for maximum cuts on graphs,” (2023), [267] Ankit Garg, Robin Kothari, Praneeth Netrapalli, and
arXiv:2309.08815 [quant-ph]. Suhail Sherif, “No quantum speedup over gradient de-
[254] Ashley Montanaro, “Quantum speedup of branch-and- scent for non-smooth convex optimization,” in 12th In-
bound algorithms,” Physical Review Research 2, 013056 novations in Theoretical Computer Science Conference
(2020). (ITCS 2021) (Schloss Dagstuhl-Leibniz-Zentrum für In-
[255] Andreas Bärtschi and Stephan Eidenbenz, “Short- formatik, 2021).
Depth Circuits for Dicke State Preparation,” in IEEE [268] Ankit Garg, Robin Kothari, Praneeth Netrapalli, and
International Conference on Quantum Computing and Suhail Sherif, “Near-optimal lower bounds for convex
Engineering QCE’22 (2022) pp. 87–96. optimization for all orders of smoothness,” in Advances
[256] Google AI Quantum, Collaborators, et al., “Hartree- in Neural Information Processing Systems, Vol. 34,
fock on a superconducting qubit quantum computer,” edited by M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S.
Science 369, 1084–1089 (2020). Liang, and J. Wortman Vaughan (Curran Associates,
[257] Stuart Hadfield, Zhihui Wang, Eleanor G Rieffel, Bryan Inc., 2021) pp. 29874–29884.
O’Gorman, Davide Venturelli, and Rupak Biswas, [269] Martin Grötschel, László Lovász, and Alexander Schri-
“Quantum approximate optimization with hard and soft jver, Geometric algorithms and combinatorial optimiza-
constraints,” in Proceedings of the Second International tion, Vol. 2 (Springer Science & Business Media, 2012).
Workshop on Post Moores Era Supercomputing (2017) [270] Joran van Apeldoorn, András Gilyén, Sander Gribling,

60
 VIII CONCLUSION & OUTLOOK

and Ronald de Wolf, “Convex optimization using quan- [287] Brandon Augustino, Giacomo Nannicini, Tamás Ter-
tum oracles,” Quantum 4 (2020), 10.22331/q-2020-01- laky, and Luis F Zuluaga, “Quantum interior point
13-220. methods for semidefinite optimization,” Quantum 7,
[271] Shouvanik Chakrabarti, Andrew M Childs, Tongyang 1110 (2023).
Li, and Xiaodi Wu, “Quantum algorithms and lower [288] Mohammadhossein Mohammadisiahroudi, Ramin
bounds for convex optimization,” Quantum 4, 221 Fakhimi, and Tamás Terlaky, “Efficient use of quan-
(2020). tum linear system algorithms in interior point methods
[272] Sanjeev Arora, Elad Hazan, and Satyen Kale, “Fast al- for linear optimization,” (2023), arXiv:2205.01220
gorithms for approximate semidefinite programming us- [quant-ph].
ing the multiplicative weights update method,” in 46th [289] Zeguan Wu, Mohammadhossein Mohammadisiahroudi,
Annual IEEE Symposium on Foundations of Computer Brandon Augustino, Xiu Yang, and Tamás Terlaky,
Science (FOCS’05) (IEEE, 2005) pp. 339–348. “An inexact feasible quantum interior point method for
[273] Joran van Apeldoorn, A quantum view on convex op- linearly constrained quadratic optimization,” Entropy
timization, Ph.D. thesis, University of Amsterdam In- 25, 330 (2023).
stitute for Logic, Language and Computation (ILLC) [290] Joran van Apeldoorn, Arjan Cornelissen, András Gi-
(2020). lyén, and Giacomo Nannicini, “Quantum tomography
[274] Rahul Jain, Zhengfeng Ji, Sarvagya Upadhyay, and using state-preparation unitaries,” in Proceedings of the
John Watrous, “Qip = pspace,” J. ACM 58 (2011), 2023 Annual ACM-SIAM Symposium on Discrete Algo-
10.1145/2049697.2049704. rithms (SODA) (SIAM, 2023) pp. 1265–1318.
[275] Vittorio Giovannetti, Seth Lloyd, and Lorenzo Mac- [291] Simon Apers and Sander Gribling, “Quantum speedups
cone, “Quantum random access memory,” Phys. Rev. for linear programming via interior point methods,”
Lett. 100, 160501 (2008), arxiv:0708.1879 [quant-ph]. (2023), arXiv:2311.03215 [quant-ph].
[276] Vittorio Giovannetti, Seth Lloyd, and Lorenzo Mac- [292] Simon Apers and Ronald de Wolf, “Quantum speedup
cone, “Architectures for a quantum random access mem- for graph sparsification, cut approximation, and Lapla-
ory,” Phys. Rev. A 78, 052310 (2008). cian solving,” SIAM Journal on Computing 51, 1703–
[277] Scott Aaronson, “Read the fine print,” Nature Phys 11, 1742 (2022).
291–293 (2015). [293] Sander Gribling and Harold Nieuwboer, “Improved
[278] Olivia Di Matteo, Vlad Gheorghiu, and Michele Mosca, quantum lower and upper bounds for matrix scaling,” in
“Fault tolerant resource estimation of quantum random- Proceedings of 39th International Symposium on Theo-
access memories,” IEEE Trans. Quantum Eng. 1, 1–13 retical Aspects of Computer Science (STACS’22), Vol.
(2020), arxiv:1902.01329 [quant-ph]. 219 (2022) pp. 35:1–35:23, arXiv:2109.15282v1.
[279] B. David Clader, Alexander M. Dalzell, Nikitas Stam- [294] Nathan Linial, Alex Samorodnitsky, and Avi Wigder-
atopoulos, Grant Salton, Mario Berta, and William J. son, “A deterministic strongly polynomial algorithm for
Zeng, “Quantum Resources Required to Block-Encode matrix scaling and approximate permanents,” Combi-
a Matrix of Classical Data,” (2022), arxiv:2206.03505 natorica 20, 545–568 (2000).
[quant-ph]. [295] Edward Anderson, Zhaojun Bai, Christian Bischof, Su-
[280] Samuel Jaques and Arthur G. Rattew, “QRAM: A Sur- san Blackford, James Demmel, Jack Dongarra, Jeremy
vey and Critique,” (2023), arxiv:2305.10310 [quant-ph]. Du Croz, Anne Greenbaum, Sven Hammarling, Alan
[281] Michael D. Grigoriadis and Leonid G. Khachiyan, “Ap- McKenney, et al., LAPACK Users’ guide (SIAM, 1999).
proximate solution of matrix games in parallel,” in Ad- [296] Joran van Apeldoorn, Sander Gribling, Yinan Li,
vances in Optimization and Parallel Computing, edited Harold Nieuwboer, Michael Walter, and Ronald
by Panos M. Pardalos (Elsevier, 1991). de Wolf, “Quantum algorithms for matrix scaling and
[282] Michael D. Grigoriadis and Leonid G. Khachiyan, “A matrix balancing,” in Proceedings of 48th International
sublinear-time randomized approximation algorithm for Colloquium on Automata, Languages, and Program-
matrix games,” Operations Research Letters 18, 53–58 ming (ICALP’21), Vol. 198 (2021) pp. 110:1–110:17,
(1995). arXiv:2011.12823v1.
[283] Jakub Marecek and Albert Akhriev, “A cutting-plane [297] Jiaqi Leng, Ethan Hickman, Joseph Li, and Xiaodi
method for semidefinite programming with poten- Wu, “Quantum hamiltonian descent,” arXiv preprint
tial applications on noisy quantum devices,” (2021), arXiv:2303.01471 (2023).
arXiv:2110.03400 [quant-ph]. [298] Andre Wibisono, Ashia C Wilson, and Michael I Jor-
[284] Shantanav Chakraborty, András Gilyén, and Stacey dan, “A variational perspective on accelerated methods
Jeffery, “The power of block-encoded matrix powers: in optimization,” proceedings of the National Academy
Improved regression techniques via faster hamiltonian of Sciences 113, E7351–E7358 (2016).
simulation,” Leibniz international proceedings in infor- [299] Brandon Augustino, Jiaqi Leng, Giacomo Nannicini,
matics 132 (2019). Tamás Terlaky, and Xiaodi Wu, “A quantum central
[285] Patrick Rebentrost, Maria Schuld, Leonard Wossnig, path algorithm for linear optimization,” arXiv preprint
Francesco Petruccione, and Seth Lloyd, “Quantum arXiv:2311.03977 (2023).
gradient descent and newton’s method for constrained [300] Leo Liberti, “Undecidability and hardness in mixed-
polynomial optimization,” New Journal of Physics 21, integer nonlinear programming,” RAIRO-Operations
073023 (2019). Research 53, 81–109 (2019).
[286] Iordanis Kerenidis and Anupam Prakash, “A quan- [301] Frank H. Clarke, Optimization and Nonsmooth Analysis
tum interior point method for lps and sdps,” ACM (Society for Industrial and Applied Mathematics, 1990)
Transactions on Quantum Computing 1 (2020), https://epubs.siam.org/doi/pdf/10.1137/1.9781611971309.
10.1145/3406306. [302] Jiajin Li, Anthony Man-Cho So, and Wing-Kin Ma,

61
 VIII CONCLUSION & OUTLOOK

“Understanding notions of stationarity in nonsmooth Quantum Engineering 1, 1–22 (2020).

optimization: A guided tour of various constructions of [318] Chin-Yao Chang, Eric Jones, Yiyun Yao, Peter Graf,
subdifferential for nonsmooth functions,” IEEE Signal and Rishabh Jain, “On hybrid quantum and classical
Processing Magazine 37, 18–31 (2020). computing algorithms for mixed-integer programming,”
[303] Chenyi Zhang and Tongyang Li, “Quantum lower arXiv preprint arXiv:2010.07852 (2020).
bounds for finding stationary points of nonconvex func- [319] Lee Braine, Daniel J. Egger, Jennifer Glick, and Ste-
tions,” in International Conference on Machine Learn- fan Woerner, “Quantum algorithms for mixed binary
ing (PMLR, 2023) pp. 41268–41299. optimization applied to transaction settlement,” IEEE
[304] Aaron Sidford and Chenyi Zhang, “Quantum speedups Transactions on Quantum Engineering 2, 1–8 (2021).
for stochastic optimization,” (2023), arXiv:2308.01582 [320] Thinh Viet Le and Vassilis Kekatos, “Variational Quan-
[quant-ph]. tum Eigensolver with Constraints (VQEC): Solving
[305] Iordanis Kerenidis and Anupam Prakash, “Quantum Constrained Optimization Problems via VQE,” (2023),
gradient descent for linear systems and least squares,” arXiv:2311.08502 [quant-ph].
Phys. Rev. A 101, 022316 (2020). [321] Dimitri P Bertsekas, Constrained optimization and La-
[306] Keren Li, Shijie Wei, Pan Gao, Feihao Zhang, Zengrong grange multiplier methods (Academic press, 2014).
Zhou, Tao Xin, Xiaoting Wang, Patrick Rebentrost, [322] Ernesto G Birgin and José Mario Martínez, Practi-
and Guilu Long, “Optimizing a polynomial function on cal augmented Lagrangian methods for constrained op-
a quantum processor,” npj Quantum Information 7, 16 timization (SIAM, 2014).
(2021). [323] Johannes Aspman, Gilles Bareilles, Vyacheslav Kun-
[307] Rolando D Somma, Sergio Boixo, Howard Barnum, and gurtsev, Jakub Marecek, and Martin Takáč, “Hy-
Emanuel Knill, “Quantum simulations of classical an- brid methods in polynomial optimisation,” (2023),
nealing processes,” Physical review letters 101, 130504 arXiv:2305.16122 [math.OC].
(2008). [324] Michal Kočvara and Michael Stingl, “Pennon: A code
[308] Satoshi Morita and Hidetoshi Nishimori, “Mathematical for convex nonlinear and semidefinite programming,”
foundation of quantum annealing,” Journal of Mathe- Optimization methods and software 18, 317–333 (2003).
matical Physics 49 (2008). [325] George B Dantzig and Philip Wolfe, “Decomposition
[309] Yizhou Liu, Weijie J Su, and Tongyang Li, “On quan- principle for linear programs,” Operations research 8,
tum speedups for nonconvex optimization via quantum 101–111 (1960).
tunneling walks,” Quantum 7, 1030 (2023). [326] Pablo A Parrilo, Structured semidefinite programs and
[310] Immanuel M. Bomze, Jianqiang Cheng, Peter J. C. semialgebraic geometry methods in robustness and opti-
Dickinson, Abdel Lisser, and Jia Liu, “Notoriously hard mization (California Institute of Technology, 2000).
(mixed-)binary qps: empirical evidence on new com- [327] Jean B. Lasserre, “Global optimization with polynomi-
pletely positive approaches,” Comput. Manag. Sci. 16, als and the problem of moments,” SIAM Journal on op-
593–619 (2019). timization 11, 796–817 (2001).
[311] Wim Van Ackooij, Irene Danti Lopez, Antonio Fran- [328] Grant Schoenebeck, “Linear level lasserre lower bounds
gioni, Fabrizio Lacalandra, and Milad Tahanan, “Large- for certain k-csps,” in 2008 49th Annual IEEE Sympo-
scale unit commitment under uncertainty: an updated sium on Foundations of Computer Science (IEEE, 2008)
literature survey,” Annals of Operations Research 271, pp. 593–602.
11–85 (2018). [329] Samuel Burer, “On the copositive representation of bi-
[312] Ashley Montanaro, “Quantum walk speedup of back- nary and continuous nonconvex quadratic programs,”
tracking algorithms,” arXiv preprint arXiv:1509.02374 Mathematical Programming 120, 479–495 (2009).
(2015). [330] Peter JC Dickinson, “Geometry of the copositive and
[313] Andris Ambainis and Martins Kokainis, “Quantum al- completely positive cones,” Journal of Mathematical
gorithm for tree size estimation, with applications to Analysis and Applications 380, 377–395 (2011).
backtracking and 2-player games,” in Proceedings of the [331] Peter JC Dickinson and Luuk Gijben, “On the com-
49th Annual ACM SIGACT Symposium on Theory of putational complexity of membership problems for the
Computing (2017) pp. 989–1002. completely positive cone and its dual,” Computational
[314] Shouvanik Chakrabarti, Pierre Minssen, Romina optimization and applications 57, 403–415 (2014).
Yalovetzky, and Marco Pistoia, “Universal quan- [332] Immanuel M Bomze, “Copositive optimization–recent
tum speedup for branch-and-bound, branch-and- developments and applications,” European Journal of
cut, and tree-search algorithms,” arXiv preprint Operational Research 216, 509–520 (2012).
arXiv:2210.03210 (2022). [333] Robin Brown, David E. Bernal Neira, Davide Ven-
[315] Alexander M Dalzell, Nicola Pancotti, Earl T Campbell, turelli, and Marco Pavone, “A copositive framework
and Fernando GSL Brandão, “Mind the gap: Achiev- for analysis of hybrid ising-classical algorithms,” (2023),
ing a super-grover quantum speedup by jumping to the arXiv:2207.13630 [math.OC].
end,” in Proceedings of the 55th Annual ACM Sympo- [334] Cheng Guo, Merve Bodur, and Joshua A. Taylor,
sium on Theory of Computing (2023) pp. 1131–1144. “Copositive duality for discrete markets and games,”
[316] Claudio Gambella, Andrea Simonetto, Martin Mevis- (2021), arXiv:2101.05379 [math.OC].
sen, and Jakub Marecek, “Mixed-binary constrained [335] Aleksandrs Belovs, “Quantum algorithms for learning
optimization on quantum computers,” (2020), uS symmetric juntas via the adversary bound,” computa-
Patent App. 16/395,560. tional complexity 24, 255–293 (2015).
[317] Claudio Gambella and Andrea Simonetto, “Multiblock [336] Kazuo Iwama, Harumichi Nishimura, Rudy Raymond,
admm heuristics for mixed-binary optimization on clas- and Junichi Teruyama, “Quantum counterfeit coin prob-
sical and quantum computers,” IEEE Transactions on lems,” Theoretical Computer Science 456, 51–64 (2012).

62
 VIII CONCLUSION & OUTLOOK

[337] Andris Ambainis, Kaspars Balodis, Janis Iraids, Mar- model,” in Proceedings of the 38th International Con-
tins Kokainis, Krisjanis Prusis, and Jevgenijs Vihrovs, ference on Machine Learning, Proceedings of Machine
“Quantum speedups for exponential-time dynamic pro- Learning Research, Vol. 139, edited by Marina Meila
gramming algorithms,” in Proceedings of the Thirti- and Tong Zhang (PMLR, 2021) pp. 10916–10926.
eth Annual ACM-SIAM Symposium on Discrete Algo- [351] Arjan Cornelissen, “Quantum gradient estimation and
rithms, SODA 2019, San Diego, California, USA, Jan- its application to quantum reinforcement learning,” TU
uary 6-9, 2019 , edited by Timothy M. Chan (SIAM, Delft Library (2018).
2019) pp. 1783–1793. [352] Vedran Dunjko, Jacob M. Taylor, and Hans J.
[338] Serge Gaspers and Jerry Zirui Li, “Quantum algorithms Briegel, “Quantum-enhanced machine learning,” Phys.
for graph coloring and other partitioning, covering, and Rev. Lett. 117, 130501 (2016).
packing problems,” (2023), arXiv:2311.08042 [cs.DS]. [353] Sofiene Jerbi, Arjan Cornelissen, Maris Ozols, and
[339] Masayuki Miyamoto, Masakazu Iwamura, Koichi Kise, Vedran Dunjko, “Quantum Policy Gradient Algo-
and François Le Gall, “Quantum speedup for the min- rithms,” in 18th Conference on the Theory of Quantum
imum steiner tree problem,” in International Comput- Computation, Communication and Cryptography (TQC
ing and Combinatorics Conference (Springer, 2020) pp. 2023), Leibniz International Proceedings in Informatics
234–245. (LIPIcs), Vol. 266, edited by Omar Fawzi and Michael
[340] Vladislavs Klevickis, Krisjanis Prusis, and Jevgenijs Walter (Schloss Dagstuhl – Leibniz-Zentrum für Infor-
Vihrovs, “Quantum speedups for treewidth,” in 17th matik, Dagstuhl, Germany, 2023) pp. 13:1–13:24.
Conference on the Theory of Quantum Computation, [354] Arjan Cornelissen, Yassine Hamoudi, and Sofiene
Communication and Cryptography, TQC 2022, July 11- Jerbi, “Near-optimal quantum algorithms for multivari-
15, 2022, Urbana Champaign, Illinois, USA, LIPIcs, ate mean estimation,” in Proceedings of the 54th An-
Vol. 232, edited by François Le Gall and Tomoyuki nual ACM SIGACT Symposium on Theory of Comput-
Morimae (Schloss Dagstuhl - Leibniz-Zentrum für In- ing, STOC 2022 (Association for Computing Machinery,
formatik, 2022) pp. 11:1–11:18. New York, NY, USA, 2022) pp. 33–43.
[341] Kazuya Shimizu and Ryuhei Mori, “Exponential-time [355] Vedran Dunjko, Yi-Kai Liu, Xingyao Wu, and
quantum algorithms for graph coloring problems,” in Jacob M. Taylor, “Exponential improvements for
LATIN 2020: Theoretical Informatics - 14th Latin quantum-accessible reinforcement learning,” (2017),
American Symposium, São Paulo, Brazil, January 5- arXiv:1710.11160.
8, 2021, Proceedings, Lecture Notes in Computer Sci- [356] Sofiene Jerbi, Casper Gyurik, Simon Marshall, Hans
ence, Vol. 12118, edited by Yoshiharu Kohayakawa and Briegel, and Vedran Dunjko, “Parametrized quantum
Flávio Keidi Miyazawa (Springer, 2020) pp. 387–398. policies for reinforcement learning,” in Advances in Neu-
[342] Robert A Wagner and Michael J Fischer, “The string- ral Information Processing Systems, Vol. 34, edited by
to-string correction problem,” Journal of the ACM M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang,
(JACM) 21, 168–173 (1974). and J. Wortman Vaughan (Curran Associates, Inc.,
[343] Andris Ambainis, Kaspars Balodis, Janis Iraids, Kamil 2021) pp. 28362–28375.
Khadiev, Vladislavs Klevickis, Krisjanis Prusis, Yixin [357] Andrea Skolik, Sofiene Jerbi, and Vedran Dunjko,
Shen, Juris Smotrovs, and Jevgenijs Vihrovs, “Quan- “Quantum agents in the Gym: a variational quantum
tum bounds for 2d-grid and dyck language,” Quantum algorithm for deep Q-learning,” Quantum 6, 720 (2022).
Inf. Process. 22, 194 (2023). [358] David Silver, Aja Huang, Chris J. Maddison, Arthur
[344] Martino Bardi, Italo Capuzzo Dolcetta, et al., Opti- Guez, Laurent Sifre, George van den Driessche, Julian
mal control and viscosity solutions of Hamilton-Jacobi- Schrittwieser, Ioannis Antonoglou, Veda Panneershel-
Bellman equations, Vol. 12 (Springer, 1997). vam, Marc Lanctot, Sander Dieleman, Dominik Grewe,
[345] John S Baras and A Kurzhanski, “Nonlinear filter- John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy
ing: The set-membership (bounding) and the h∞ tech- Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore
niques,” IFAC Proceedings Volumes 28, 409–415 (1995). Graepel, and Demis Hassabis, “Mastering the game of
[346] Martin L Puterman, Markov decision processes: discrete Go with deep neural networks and tree search,” Nature
stochastic dynamic programming (John Wiley & Sons, 529, 484–489 (2016).
2014). [359] Kai Arulkumaran, Antoine Cully, and Julian Togelius,
[347] Giuseppe C. Calafiore and Fabrizio Dabbene, “A prob- “Alphastar: an evolutionary computation perspective,”
abilistic analytic center cutting plane method for fea- in Proceedings of the Genetic and Evolutionary Com-
sibility of uncertain lmis,” Automatica 43, 2022–2033 putation Conference Companion, GECCO ’19 (ACM,
(2007). 2019).
[348] Jonathan P Epperlein, Roman Overko, Sergiy Zhuk, [360] Owen Lockwood and Mei Si, “Reinforcement learn-
Christopher King, Djallel Bouneffouf, Andrew Cullen, ing with quantum variational circuits,” (2020),
and Robert Shorten, “Reinforcement learning with al- arXiv:2008.07524.
gorithms from probabilistic structure estimation,” Au- [361] Samuel Yen-Chi Chen, Chao-Han Huck Yang, Jun Qi,
tomatica 144, 110483 (2022). Pin-Yu Chen, Xiaoli Ma, and Hsi-Sheng Goan, “Vari-
[349] Daochen Wang, Xuchen You, Tongyang Li, and An- ational quantum circuits for deep reinforcement learn-
drew M. Childs, “Quantum exploration algorithms for ing,” IEEE Access 8, 141007–141024 (2020).
multi-armed bandits,” Proceedings of the AAAI Confer- [362] Alexey Muravlev and Mikhail Zhitlukhin, “A Bayesian
ence on Artificial Intelligence 35, 10102–10110 (2021). sequential test for the drift of a fractional Brownian mo-
[350] Daochen Wang, Aarthi Sundaram, Robin Kothari, tion,” Advances in Applied Probability 52, 1308–1324
Ashish Kapoor, and Martin Roetteler, “Quantum al- (2020).
gorithms for reinforcement learning with a generative [363] Constantinos Daskalakis and Yasushi Kawase, “Opti-

63
 VIII CONCLUSION & OUTLOOK

mal stopping rules for sequential hypothesis testing,” in [379] Ilias Diakonikolas and Mihalis Yannakakis, “Small ap-
25th Annual European Symposium on Algorithms (ESA) proximate pareto sets for biobjective shortest paths
(2017). and other problems,” SIAM Journal on Computing 39,
[364] Francis A. Longstaff and Eduardo S. Schwartz, “Valuing 1340–1371 (2010).
american options by simulation: a simple least-squares [380] Nery Riquelme, Christian Von Lücken, and Benjamin
approach,” The review of financial studies 14, 113–147 Baran, “Performance metrics in multi-objective opti-
(2001). mization,” in 2015 Latin American Computing Confer-
[365] Gilles Brassard, Peter Hoyer, Michele Mosca, and Alain ence (CLEI) (2015) pp. 1–11.
Tapp, “Quantum amplitude amplification and estima- [381] Shao-Hen Chiew, Kilian Poirier, Rajesh Mishra, Ul-
tion,” Contemporary Mathematics 305, 53–74 (2002). rike Bornheimer, Ewan Munro, Si Han Foon, Christo-
[366] J. F. Doriguello, Alessandro Luongo, Jinge Bao, Patrick pher Wanru Chen, Wei Sheng Lim, and Chee Wei
Rebentrost, and Miklos Santha, “Quantum Algorithm Nga, “Multi-objective optimization and network rout-
for Stochastic Optimal Stopping Problems with Appli- ing with near-term quantum computers,” (2023),
cations in Finance,” in 17th Conference on the Theory of arXiv:2308.08245 [quant-ph].
Quantum Computation, Communication and Cryptog- [382] Linus Ekstrom, Hao Wang, and Sebastian Schmitt,
raphy (TQC 2022), Leibniz International Proceedings “Variational quantum multi-objective optimization,”
in Informatics (LIPIcs), Vol. 232, edited by François (2024), arXiv:2312.14151 [quant-ph].
Le Gall and Tomoyuki Morimae (Schloss Dagstuhl – [383] R. Barends, A. Shabani, L. Lamata, J. Kelly, A. Mez-
Leibniz-Zentrum für Informatik, Dagstuhl, Germany, zacapo, U. Las Heras, R. Babbush, A. G. Fowler,
2022) pp. 2:1–2:24. B. Campbell, Yu Chen, and et al., “Digitized adia-
[367] Yangchen Pan, Erfan Sadeqi Azer, and Martha White, batic quantum computing with a superconducting cir-
“Effective sketching methods for value function approx- cuit,” Nature 534, 222–226 (2016).
imation,” (2017), arXiv:1708.01298 [cs.LG]. [384] Clemens Dlaska, Kilian Ender, Glen Bigan Mbeng, An-
[368] Haifang Li, Yingce Xia, and Wensheng Zhang, “Finite dreas Kruckenhauser, Wolfgang Lechner, and Rick van
sample analysis of lstd with random projections and eli- Bijnen, “Quantum optimization via four-body rydberg
gibility traces,” in Proceedings of the 27th International gates,” Phys. Rev. Lett. 128, 120503 (2022).
Joint Conference on Artificial Intelligence (2018) pp. [385] Matthew P. Harrigan, Kevin J. Sung, Matthew Neeley,
2390–2396. Kevin J. Satzinger, Frank Arute, Kunal Arya, Juan Ata-
[369] Jonas Kübler, Simon Buchholz, and Bernhard laya, Joseph C. Bardin, Rami Barends, Sergio Boixo,
Schölkopf, “The inductive bias of quantum kernels,” Ad- and et al., “Quantum approximate optimization of non-
vances in Neural Information Processing Systems 34, planar graph problems on a planar superconducting pro-
12661–12673 (2021). cessor,” Nat. Phys. 17, 332–336 (2021).
[370] Wei Chu, Lihong Li, Lev Reyzin, and Robert Schapire, [386] S. Ebadi, A. Keesling, M. Cain, T. T. Wang, H. Levine,
“Contextual bandits with linear payoff functions,” in D. Bluvstein, G. Semeghini, A. Omran, J.-G. Liu,
Proceedings of the Fourteenth International Conference R. Samajdar, X.-Z. Luo, B. Nash, X. Gao, B. Barak,
on Artificial Intelligence and Statistics (JMLR Work- E. Farhi, S. Sachdev, N. Gemelke, L. Zhou, S. Choi,
shop and Conference Proceedings, 2011) pp. 208–214. H. Pichler, S.-T. Wang, M. Greiner, V. Vuletić, and
[371] Aharon Ben-Tal and Arkadi Nemirovski, “Robust con- M. D. Lukin, “Quantum optimization of maximum in-
vex optimization,” Mathematics of operations research dependent set using Rydberg atom arrays,” Science 376,
23, 769–805 (1998). 1209–1215 (2022).
[372] Aharon Ben-Tal and Arkadi Nemirovski, “Robust [387] Guido Pagano, Aniruddha Bapat, Patrick Becker,
optimization–methodology and applications,” Mathe- Katherine S. Collins, Arinjoy De, Paul W. Hess, Har-
matical programming 92, 453–480 (2002). vey B. Kaplan, Antonis Kyprianidis, Wen Lin Tan,
[373] Aharon Ben-Tal and Arkadi Nemirovski, “Selected top- Christopher Baldwin, Lucas T. Brady, Abhinav Desh-
ics in robust convex optimization,” Mathematical Pro- pande, Fangli Liu, Stephen Jordan, Alexey V. Gor-
gramming 112, 125–158 (2008). shkov, and Christopher Monroe, “Quantum approxi-
[374] Dimitris Bertsimas, David B. Brown, and Constantine mate optimization of the long-range ising model with
Caramanis, “Theory and applications of robust opti- a trapped-ion quantum simulator,” Proceedings of the
mization,” SIAM review 53, 464–501 (2011). National Academy of Sciences 117, 25396–25401 (2020).
[375] Aharon Ben-Tal, Elad Hazan, Tomer Koren, and Shie [388] Phillip C. Lotshaw, Kevin D. Battles, Bryan Gard,
Mannor, “Oracle-based robust optimization via online Gilles Buchs, Travis S. Humble, and Creston D. Herold,
learning,” Operations Research 63, 628–638 (2015). “Modeling noise in global mølmer-sørensen interactions
[376] Debbie Lim, J. F. Doriguello, and Patrick Reben- applied to quantum approximate optimization,” Physi-
trost, “Quantum algorithm for robust optimization cal Review A 107 (2023), 10.1103/physreva.107.062406.
via stochastic-gradient online learning,” (2023), [389] Elijah Pelofske, Andreas Bärtschi, John Golden, and
arXiv:2304.02262 [quant-ph]. Stephan Eidenbenz, “High-Round QAOA for MAX k-
[377] Nyoman Gunantara, “A review of multi-objective opti- SAT on Trapped Ion NISQ Devices,” in IEEE Inter-
mization: Methods and its applications,” Cogent Engi- national Conference on Quantum Computing and Engi-
neering 5, 1502242 (2018). neering QCE’23 (2023) pp. 506–517.
[378] Christos H Papadimitriou and Mihalis Yannakakis, “On [390] Philip Krantz, Morten Kjaergaard, Fei Yan, Terry P.
the approximability of trade-offs and optimal access of Orlando, Simon Gustavsson, and William D. Oliver, “A
web sources,” in Proceedings 41st annual symposium on quantum engineer’s guide to superconducting qubits,”
foundations of computer science (IEEE, 2000) pp. 86– Appl. Phys. Rev. 6, 021318 (2019).
92. [391] R. Barends, J. Kelly, A. Megrant, A. Veitia,

64
 VIII CONCLUSION & OUTLOOK

D. Sank, E. Jeffrey, T. C. White, J. Mutus, A. G. 383.

Fowler, B. Campbell, Y. Chen, Z. Chen, B. Chiaro, [406] A. Ambainis, A. Nayak, A. Ta-shma, and U. Vazirani,
A. Dunsworth, C. Neill, P. O’Malley, P. Roushan, “Dense quantum coding and quantum finite automata,”
A. Vainsencher, J. Wenner, A. N. Korotkov, A. N. Cle- Journal of the ACM 49, 496–511 (2002).
land, and John M. Martinis, “Superconducting quan- [407] T. Imamichi and R. Raymond, “Constructions of quan-
tum circuits at the surface code threshold for fault tol- tum random access codes,” in Asian Quantum Informa-
erance,” Nature 508, 500–503 (2014). tion Symposium (AQIS), Vol. 66 (2018).
[392] Chad Rigetti and Michel Devoret, “Fully microwave- [408] Laurin E. Fischer, Daniel Miller, Francesco Tacchino,
tunable universal gates in superconducting qubits with Panagiotis Kl. Barkoutsos, Daniel J. Egger, and Ivano
linear couplings and fixed transition frequencies,” Phys. Tavernelli, “Ancilla-free implementation of generalized
Rev. B 81, 134507 (2010). measurements for qubits embedded in a qudit space,”
[393] David C. McKay, Stefan Filipp, Antonio Mezzacapo, Phys. Rev. Res. 4, 033027 (2022).
Easwar Magesan, Jerry M. Chow, and Jay M. Gam- [409] Maxime Dupont and Bhuvanesh Sundar, “Extend-
betta, “Universal gate for fixed-frequency qubits via a ing relax-and-round combinatorial optimization solvers
tunable bus,” Phys. Rev. Appl. 6, 064007 (2016). with quantum correlations,” Physical Review A 109,
[394] M. Ganzhorn, G. Salis, D. J. Egger, A. Fuhrer, M. Mer- 012429 (2024).
genthaler, C. Müller, P. Müller, S. Paredes, M. Pechal, [410] Andrew Cross, Ali Javadi-Abhari, Thomas Alexander,
M. Werninghaus, and S. Filipp, “Benchmarking the Niel De Beaudrap, Lev S Bishop, Steven Heidel, Colm A
noise sensitivity of different parametric two-qubit gates Ryan, Prasahnt Sivarajah, John Smolin, Jay M Gam-
in a single superconducting quantum computing plat- betta, et al., “Openqasm 3: A broader and deeper quan-
form,” Phys. Rev. Res. 2, 033447 (2020). tum assembly language,” ACM Transactions on Quan-
[395] IBM, “IBM Quantum Platform - Compute resources,” tum Computing 3, 1–50 (2022).
(2023). [411] Robert S. Smith, Michael J. Curtis, and William J.
[396] S. Hadfield, T. Hogg, and E. G. Rieffel, “Analytical Zeng, “A Practical Quantum Instruction Set Architec-
framework for quantum alternating operator ansätze,” ture,” (2017), arXiv:1608.03355 [quant-ph].
Quantum Science and Technology 8, 015017 (2022). [412] Nathan Killoran, Josh Izaac, Nicolás Quesada, Ville
[397] P. Niroula, R. Shaydulin, R. Yalovetsky, P. Minssen, Bergholm, Matthew Amy, and Christian Weedbrook,
D. Herman, S. Hu, and M. Pistoia, “Constrained quan- “Strawberry fields: A software platform for photonic
tum optimization for extractive summarization on a quantum computing,” Quantum 3, 129 (2019).
trapped-ion quantum computer,” Scientific Reports 12 [413] Paul D. Nation and Matthew Treinish, “Suppressing
(2022), 10.1038/s41598-022-20853-w. quantum circuit errors due to system variability,” PRX
[398] P. Chandarana, N. N. Hegade, K. Paul, F. Albarrán- Quantum 4, 010327 (2023).
Arriagada, E. Solano, A. del Campo, and Xi Chen, [414] Fei Hua, Yuwei Jin, Ang Li, Yanhao Chen, Chi Zhang,
“Digitized-counterdiabatic quantum approximate opti- Ari Hayes, Hang Gao, and Eddy Z. Zhang, “A syn-
mization algorithm,” Phys. Rev. Res. 4, 013141 (2022). ergistic compilation workflow for tackling crosstalk in
[399] J. S. Baker and S. K. Radha, “Wasserstein solu- quantum machines,” (2022), arXiv:2207.05751 [quant-
tion quality and the quantum approximate optimiza- ph].
tion algorithm: A portfolio optimization case study,” [415] David C. McKay, Ian Hincks, Emily J. Pritchett, Mal-
arXiv:2202.06782 (2022). colm Carroll, Luke C. G. Govia, and Seth T. Merkel,
[400] Yunlong Yu, Chenfeng Cao, Carter Dewey, Xiang-Bin “Benchmarking Quantum Processor Performance at
Wang, Nic Shannon, and Robert Joynt, “Quantum ap- Scale,” (2023), arXiv:2311.05933 [quant-ph].
proximate optimization algorithm with adaptive bias [416] Fei Hua, Yuwei Jin, Yanhao Chen, Suhas Vittal, Kevin
fields,” Phys. Rev. Res. 4, 023249 (2022). Krsulich, Lev S. Bishop, John Lapeyre, Ali Javadi-
[401] Johannes Weidenfeller, Lucia C Valor, Julien Gacon, Abhari, and Eddy Z. Zhang, “Exploiting qubit reuse
Caroline Tornow, Luciano Bello, Stefan Woerner, and through mid-circuit measurement and reset,” (2023),
Daniel J Egger, “Scaling of the quantum approximate arXiv:2211.01925 [quant-ph].
optimization algorithm on superconducting qubit based [417] A. Matsuo, S. Yamashita, and D. J. Egger, “A SAT ap-
hardware,” Quantum 6, 870 (2022). proach to the initial mapping problem in SWAP gate
[402] Stefan H. Sack and Daniel J. Egger, “Large-scale quan- insertion for commuting gates,” IEICE Transactions
tum approximate optimization on non-planar graphs on Fundamentals of Electronics, Communications and
with machine learning noise mitigation,” (2023), Computer Sciences , 2022EAP1159 (2023).
arXiv:2307.14427 [quant-ph]. [418] Jun Zhang, Jiri Vala, Shankar Sastry, and K. Birgitta
[403] Aidan Pellow-Jarman, Shane McFarthing, Ilya Whaley, “Minimum construction of two-qubit quantum
Sinayskiy, Anban Pillay, and Francesco Petruc- operations,” Phys. Rev. Lett. 93, 020502 (2004).
cione, “QAOA Performance in Noisy Devices: The [419] Byron Drury and Peter Love, “Constructive quantum
Effect of Classical Optimizers and Ansatz Depth,” Shannon decomposition from Cartan involutions,” Jour-
(2023), arXiv:2307.10149 [quant-ph]. nal of Physics A: Mathematical and Theoretical 41,
[404] A. S. Holevo, “On capacity of a quantum communica- 395305 (2008).
tions channel,” Problems of Information Transmission [420] Dmitri Maslov, “Basic circuit compilation techniques for
15, 247–253 (1979). an ion-trap quantum machine,” New Journal of Physics
[405] A. Ambainis, A. Nayak, A. Tashma, and U. Vazirani, 19, 023035 (2017).
“Dense quantum coding and a lower bound for 1-way [421] David P. DiVincenzo, “The physical implementation
quantum automata,” in Proceedings of 31st ACM Sym- of quantum computation,” Fortschritte der Physik 48,
posium on Theory of Computing (ACM, 1999) pp. 376– 771–783 (2000).

65
 VIII CONCLUSION & OUTLOOK

[422] Gushu Li, Anbang Wu, Yunong Shi, Ali Javadi-Abhari, Quantum Computing and Engineering QCE’23 (2023)
Yufei Ding, and Yuan Xie, “Paulihedral: A generalized pp. 120–130.
block-wise compiler optimization framework for quan- [434] L. Viola and S. Lloyd, “Dynamical suppression of de-
tum simulation kernels,” in Proceedings of the 27th ACM coherence in two-state quantum systems,” Phy. Rev. A
International Conference on Architectural Support for 58, 2733 (1998).
Programming Languages and Operating Systems, ASP- [435] D. Vitali and P. Tombesi, “Using parity kicks for deco-
LOS ’22 (Association for Computing Machinery, New herence control,” Phy. Rev. A 59, 4178 (1999).
York, NY, USA, 2022) pp. 554–569. [436] B. Pokharel, N. Anand, B. Fortman, and D. A. Lidar,
[423] Liam Madden and Andrea Simonetto, “Best Ap- “Demonstration of fidelity improvement using dynami-
proximate Quantum Compiling Problems,” ACM cal decoupling with superconducting qubits,” Phys. Rev.
Transactions on Quantum Computing 3 (2022), Lett. 121, 220502 (2018).
10.1145/3505181, place: New York, NY, USA Publisher: [437] Elijah Pelofske, Andreas Bärtschi, and Stephan Eiden-
Association for Computing Machinery. benz, “Short-depth QAOA circuits and quantum an-
[424] Ethan Smith, Marc Grau Davis, Jeffrey Larson, Ed You- nealing on higher-order ising models,” npj Quantum In-
nis, Lindsay Bassman Oftelie, Wim Lavrijsen, and formation 10, 30 (2024).
Costin Iancu, “LEAP: Scaling Numerical Optimiza- [438] Abhinav Kandala, Kristan Temme, Antonio D. Cor-
tion Based Synthesis Using an Incremental Approach,” coles, Antonio Mezzacapo, Jerry M. Chow, and Jay M.
ACM Transactions on Quantum Computing 4 (2023), Gambetta, “Error mitigation extends the computational
10.1145/3548693, place: New York, NY, USA Publisher: reach of a noisy quantum processor,” Nature 567, 491–
Association for Computing Machinery. 495 (2018).
[425] Gushu Li, Yufei Ding, and Yuan Xie, “Tackling the [439] Christophe Piveteau, David Sutter, and Stefan
Qubit Mapping Problem for NISQ-Era Quantum De- Woerner, “Quasiprobability decompositions with re-
vices,” in Proceedings of the Twenty-Fourth Interna- duced sampling overhead,” npj Quantum Information
tional Conference on Architectural Support for Program- 8 (2022), 10.1038/s41534-022-00517-3.
ming Languages and Operating Systems, ASPLOS ’19 [440] Daniel Gottesman, “An introduction to quantum error
(Association for Computing Machinery, New York, NY, correction and fault-tolerant quantum computation,”
USA, 2019) pp. 1001–1014. (2009), arXiv:0904.2557 [quant-ph].
[426] A. Zulehner, A. Paler, and R. Wille, “An efficient [441] Ying Li and Simon C. Benjamin, “Efficient variational
methodology for mapping quantum circuits to the ibm quantum simulator incorporating active error minimiza-
qx architectures,” IEEE Transactions on Computer- tion,” Physical Review X 7 (2017).
Aided Design of Integrated Circuits and Systems 38, [442] Ryan LaRose, Andrea Mari, Sarah Kaiser, Peter J.
1226–1236 (2019). Karalekas, Andre A. Alves, Piotr Czarnik, Mo-
[427] Alexander Cowtan, Silas Dilkes, Ross Duncan, Alexan- hamed El Mandouh, Max H. Gordon, Yousef Hindy,
dre Krajenbrink, Will Simmons, and Seyon Sivarajah, Aaron Robertson, Purva Thakre, Misty Wahl, Danny
“On the Qubit Routing Problem,” (2019). Samuel, Rahul Mistri, Maxime Tremblay, Nick Gard-
[428] Lingling Lao and Dan E. Browne, “2qan: A quantum ner, Nathaniel T. Stemen, Nathan Shammah, and
compiler for 2-local qubit hamiltonian simulation algo- William J. Zeng, “Mitiq: A software package for error
rithms,” in Proceedings of the 49th Annual International mitigation on noisy quantum computers,” Quantum 6,
Symposium on Computer Architecture, ISCA ’22 (Asso- 774 (2022).
ciation for Computing Machinery, New York, NY, USA, [443] Almudena Carrera Vazquez, Daniel J. Egger, David
2022) pp. 351–365. Ochsner, and Stefan Woerner, “Well-conditioned multi-
[429] Yuwei Jin, Jason Luo, Lucent Fong, Yanhao Chen, product formulas for hardware-friendly hamiltonian
Ari B. Hayes, Chi Zhang, Fei Hua, and Eddy Z. simulation,” Quantum 7, 1067 (2023).
Zhang, “A Structured Method for Compilation of [444] Thomas Alexander, Naoki Kanazawa, Daniel J Egger,
QAOA Circuits in Quantum Computing,” (2022), Lauren Capelluto, Christopher J Wood, Ali Javadi-
arXiv:2112.06143 [quant-ph]. Abhari, and David C McKay, “Qiskit pulse: pro-
[430] Mahabubul Alam, Abdullah Ash-Saki, and Swaroop gramming quantum computers through the cloud with
Ghosh, “Circuit compilation methodologies for quantum pulses,” Quantum Science and Technology 5, 044006
approximate optimization algorithm,” in 2020 53rd An- (2020).
nual IEEE/ACM International Symposium on Microar- [445] Nathan Earnest, Caroline Tornow, and Daniel J. Egger,
chitecture (MICRO) (2020) pp. 215–228. “Pulse-efficient circuit transpilation for quantum appli-
[431] A. D. Córcoles, Maika Takita, Ken Inoue, Scott Lekuch, cations on cross-resonance-based hardware,” Phys. Rev.
Zlatko K. Minev, Jerry M. Chow, and Jay M. Gam- Res. 3, 043088 (2021).
betta, “Exploiting dynamic quantum circuits in a quan- [446] Alexey Galda, Michael Cubeddu, Naoki Kanazawa,
tum algorithm with superconducting qubits,” Phys. Prineha Narang, and Nathan Earnest-Noble, “Im-
Rev. Lett. 127, 100501 (2021). plementing a Ternary Decomposition of the Toffoli
[432] Elisa Bäumer, Vinay Tripathi, Derek S. Wang, Gate on Fixed-Frequency Transmon Qutrits,” (2021),
Patrick Rall, Edward H. Chen, Swarnadeep Majumder, arXiv:2109.00558 [quant-ph].
Alireza Seif, and Zlatko K. Minev, “Efficient long- [447] Laurin E. Fischer, Alessandro Chiesa, Francesco
range entanglement using dynamic circuits,” (2023), Tacchino, Daniel J. Egger, Stefano Carretta, and Ivano
arXiv:2308.13065 [quant-ph]. Tavernelli, “Universal qudit gate synthesis for trans-
[433] Anette Messinger, Michael Fellner, and Wolfgang Lech- mons,” PRX Quantum 4, 030327 (2023).
ner, “Constant Depth Code Deformations in the Par- [448] D. J. Egger and F. K. Wilhelm, “Optimal control of
ity Architecture,” in IEEE International Conference on a quantum measurement,” Phys. Rev. A 90, 052331

66
 VIII CONCLUSION & OUTLOOK

(2014). [464] Emily Yu, Armin Biere, and Keijo Heljanko, “Progress
[449] Pranav Gokhale, Yongshan Ding, Thomas Propson, in certifying hardware model checking results,” in Com-
Christopher Winkler, Nelson Leung, Yunong Shi, puter Aided Verification: 33rd International Confer-
David I. Schuster, Henry Hoffmann, and Frederic T. ence, CAV 2021, Virtual Event, July 20–23, 2021, Pro-
Chong, “Partial compilation of variational algorithms ceedings, Part II 33 (Springer, 2021) pp. 363–386.
for noisy intermediate-scale quantum machines,” in Pro- [465] Koen Claessen, Niklas Een, Mary Sheeran, and Niklas
ceedings of the 52nd Annual IEEE/ACM International Sorensson, “Sat-solving in practice,” in 2008 9th Inter-
Symposium on Microarchitecture, MICRO ’52 (Associ- national Workshop on Discrete Event Systems (IEEE,
ation for Computing Machinery, New York, NY, USA, 2008) pp. 61–67.
2019) pp. 266–278. [466] Ambros Gleixner, Gregor Hendel, Gerald Gamrath, To-
[450] D. T. McClure, Hanhee Paik, L. S. Bishop, M. Steffen, bias Achterberg, Michael Bastubbe, Timo Berthold,
Jerry M. Chow, and Jay M. Gambetta, “Rapid driven Philipp M. Christophel, Kati Jarck, Thorsten Koch, Jeff
reset of a qubit readout resonator,” Phys. Rev. Appl. 5, Linderoth, Marco Lübbecke, Hans Mittelmann, Derya
011001 (2016). Ozyurt, Ted Ralphs, Domenico Salvagnin, and Yuji Shi-
[451] Daniel J. Egger, Chiara Capecci, Bibek Pokharel, Pana- nano, “MIPLIB 2017: Data-Driven Compilation of the
giotis Kl. Barkoutsos, Laurin E. Fischer, Leonardo 6th Mixed-Integer Programming Library,” Mathemati-
Guidoni, and Ivano Tavernelli, “Pulse variational quan- cal Programming Computation 13, 443 – 490 (2021).
tum eigensolver on cross-resonance-based hardware,” [467] Hans Mittelmann, “Decision Tree for Optimization Soft-
Phys. Rev. Res. 5, 033159 (2023). ware,” (2023).
[452] Andrew Wack, Hanhee Paik, Ali Javadi-Abhari, Petar [468] Thomas Lubinski, Sonika Johri, Paul Varosy, Jeremiah
Jurcevic, Ismael Faro, Jay M. Gambetta, and Blake R. Coleman, Luning Zhao, Jason Necaise, Charles H. Bald-
Johnson, “Quality, Speed, and Scale: three key at- win, Karl Mayer, and Timothy Proctor, “Application-
tributes to measure the performance of near-term quan- Oriented Performance Benchmarks for Quantum Com-
tum computers,” (2021), arXiv:2110.14108 [quant-ph]. puting,” arXiv preprint (2023), arXiv:2110.03137.
[453] Andrew W. Cross, Lev S. Bishop, Sarah Sheldon, [469] T. Tomesh, P. Gokhale, V. Omole, G. Ravi, K. N.
Paul D. Nation, and Jay M. Gambetta, “Validating Smith, J. Viszlai, X. Wu, N. Hardavellas, M. R.
quantum computers using randomized model circuits,” Martonosi, and F. T. Chong, “SupermarQ: A Scal-
Phys. Rev. A 100, 032328 (2019). able Quantum Benchmark Suite,” in 2022 IEEE In-
[454] Gopal Chandra Santra, Fred Jendrzejewski, Philipp ternational Symposium on High-Performance Computer
Hauke, and Daniel J. Egger, “Squeezing and quantum Architecture (HPCA) (IEEE Computer Society, Los
approximate optimization,” (2023), arXiv:2205.10383 Alamitos, CA, USA, 2022) pp. 587–603.
[quant-ph]. [470] Simon Martiel, Thomas Ayral, and Cyril Al-
[455] “Best practices in quantum optimization,” (2023). louche, “Benchmarking quantum coprocessors in an
[456] Iain Dunning, Swati Gupta, and John Silberholz, application-centric, hardware-agnostic, and scalable
“What works best when? a systematic evaluation of way,” IEEE Transactions on Quantum Engineering 2
heuristics for max-cut and qubo,” INFORMS J. Com- (2021), 10.1109/TQE.2021.3090207.
put. 30, 608–624 (2018). [471] Troels F. Rønnow, Zhihui Wang, Joshua Job, Sergio
[457] Raj Jain, The Art of Computer Systems Performance Boixo, Sergei V. Isakov, David Wecker, John M. Mar-
Analysis: Techniques For Experimental Design, Mea- tinis, Daniel A. Lidar, and Matthias Troyer, “Defining
surement, Simulation, and Modeling (Wiley, 1991). and detecting quantum speedup,” Science 345, 420–424
[458] Thorsten Koch, Timo Berthold, Jaap Pedersen, and (2014).
Charlie Vanaret, “Progress in mathematical program- [472] Michael Juenger, Elisabeth Lobe, Petra Mutzel, Ger-
ming solvers from 2001 to 2020,” EURO Journal on hard Reinelt, Franz Rendl, Giovanni Rinaldi, and
Computational Optimization 10, 100031 (2022). Tobias Stollenwerk, “Performance of a Quantum An-
[459] Thomas Bartz-Beielstein, Carola Doerr, Daan van den nealer for Ising Ground State Computations on Chimera
Berg, Jakob Bossek, Sowmya Chandrasekaran, Tome Graphs,” (2019), arXiv:1904.11965.
Eftimov, Andreas Fischbach, Pascal Kerschke, [473] Michael Jünger, Elisabeth Lobe, Petra Mutzel, Gerhard
William La Cava, Manuel Lopez-Ibanez, Kather- Reinelt, Franz Rendl, Giovanni Rinaldi, and Tobias
ine M. Malan, Jason H. Moore, Boris Naujoks, Patryk Stollenwerk, “Quantum annealing versus digital com-
Orzechowski, Vanessa Volz, Markus Wagner, and puting: An experimental comparison,” ACM J. Exp.
Thomas Weise, “Benchmarking in Optimization: Best Algorithmics 26 (2021), 10.1145/3459606.
Practice and Open Issues,” (2020), arXiv:2007.03488 [474] Alejandro Perdomo-Ortiz, Alexander Feldman, Asier
[cs.NE]. Ozaeta, Sergei V. Isakov, Zheng Zhu, Bryan O’Gorman,
[460] Matthias Müller, Brian Whitney, Robert Henschel, and Helmut G. Katzgraber, Alexander Diedrich, Hartmut
Kalyan Kumaran, “Spec benchmarks,” in Encyclopedia Neven, Johan de Kleer, Brad Lackey, and Rupak
of Parallel Computing (Springer US, Boston, MA, 2011) Biswas, “Readiness of quantum optimization machines
pp. 1886–1893. for industrial applications,” Phys. Rev. Appl. 12, 014004
[461] Jack J Dongarra, Cleve Barry Moler, James R Bunch, (2019).
and Gilbert W Stewart, LINPACK users’ guide (SIAM, [475] Erica Grant, Travis S. Humble, and Benjamin Stump,
1979). “Benchmarking quantum annealing controls with port-
[462] “DIMACS Implementation Challenges,” (2023). folio optimization,” Phys. Rev. Appl. 15, 014012 (2021).
[463] Nils Froleyks, Marijn Heule, Markus Iser, Matti [476] Byron Tasseff, Tameem Albash, Zachary Morrell, Marc
Järvisalo, and Martin Suda, “Sat competition 2020,” Vuffray, Andrey Y. Lokhov, Sidhant Misra, and Car-
Artificial Intelligence 301, 103572 (2021). leton Coffrin, “On the Emerging Potential of Quantum

67
 VIII CONCLUSION & OUTLOOK

Annealing Hardware for Combinatorial Optimization,” mization problems and its applications,” in 2020 25th
(2022), arXiv:2210.04291 [math.OC]. Asia and South Pacific Design Automation Conference
[477] Thomas Lubinski, Carleton Coffrin, Catherine Mc- (ASP-DAC) (2020) pp. 667–672.
Geoch, Pratik Sathe, Joshua Apanavicius, and David [491] Chihiro Yoshimura, Masato Hayashi, Takashi Take-
E. Bernal Neira, “Optimization applications as quantum moto, and Masanao Yamaoka, “Cmos annealing ma-
performance benchmarks,” (2024), arXiv:2302.02278 chine: A domain-specific architecture for combinatorial
[quant-ph]. optimization problem,” in 2020 25th Asia and South Pa-
[478] Madita Willsch, Dennis Willsch, Fengping Jin, Hans cific Design Automation Conference (ASP-DAC) (2020)
De Raedt, and Kristel Michielsen, “Benchmarking the pp. 673–678.
quantum approximate optimization algorithm,” Quan- [492] Hayato Goto, “Bifurcation-based adiabatic quantum
tum Information Processing 19 (2020), 10.1007/s11128- computation with a nonlinear oscillator network,” Sci-
020-02692-8. entific Reports 6, 21686 (2016).
[479] Naeimeh Mohseni, Peter L McMahon, and Tim Byrnes, [493] Juntao Wang, Daniel Ebler, K. Y. Michael Wong, David
“Ising machines as hardware solvers of combinatorial op- Shui Wing Hui, and Jie Sun, “Bifurcation behaviors
timization problems,” Nature Reviews Physics 4, 363– shape how continuous physical dynamics solves discrete
379 (2022). ising optimization,” Nature Communications 14, 2510
[480] Jernej Rudi Finžgar, Philipp Ross, Leonhard Hölscher, (2023).
Johannes Klepsch, and Andre Luckow, “Quark: A [494] Tameem Albash and Daniel A. Lidar, “Adiabatic quan-
framework for quantum computing application bench- tum computation,” Rev. Mod. Phys. 90, 015002 (2018).
marking,” in 2022 IEEE International Conference on [495] Catherine C. McGeoch, “Adiabatic quantum computa-
Quantum Computing and Engineering (QCE) (2022) tion,” in Adiabatic Quantum Computation and Quantum
pp. 226–237. Annealing: Theory and Practice (Springer International
[481] Gurobi Optimization, GUROBI Optimizer reference Publishing, Cham, 2014) pp. 9–27.
manual version 10.0 , Tech. Rep. (Gurobi Optimization, [496] D. Hangleiter, J. Carolan, and K.P.Y. Thébault, Ana-
LLC, 2023). logue Quantum Simulation: A New Instrument for Sci-
[482] Timo Berthold and Zsolt Csizmadia, “The confined pri- entific Understanding (Springer International Publish-
mal integral: a measure to benchmark heuristic minlp ing, 2022).
solvers against global minlp solvers,” Mathematical Pro- [497] Immanuel Bloch, Jean Dalibard, and Sylvain Nascim-
gramming 188, 523–537 (2021). bène, “Quantum simulations with ultracold quantum
[483] Fabio Furini, Emiliano Traversi, Pietro Belotti, Anto- gases,” Nature Physics 8, 267–276 (2012).
nio Frangioni, Ambros Gleixner, Nick Gould, Leo Lib- [498] S. Trotzky, Y-A. Chen, A. Flesch, I. P. McCulloch,
erti, Andrea Lodi, Ruth Misener, Hans Mittelmann, U. Schollwöck, J. Eisert, and I. Bloch, “Probing the
Nikolaos Sahinidis, Stefan Vigerske, , and Angelika relaxation towards equilibrium in an isolated strongly
Wiegele, “QPLIB: A library of quadratic programming correlated one-dimensional bose gas,” Nature Physics
instances,” Mathematical Programming Computation 8, 325–330 (2012).
(2018), 10.1007/s12532-018-0147-4. [499] J.-Y. Choi, Sebastian Hild, Johannes Zeiher, Peter
[484] Holger Hoos and Thomas Stützle, “SATLIB: An online Schauß, Antonio Rubio-Abadal, Tarik Yefsah, Vedika
resource for research on SAT,” (Kluwer Academic Pub- Khemani, David A. Huse, Immanuel Bloch, and
lishers, 2000) pp. 283–292. Christian Gross, “Exploring the many-body localization
[485] Yunlong Yu, Chenfeng Cao, Xiang-Bin Wang, Nic Shan- transition in two dimensions,” Science 352, 1547–1552
non, and Robert Joynt, “Solution of sat problems with (2016).
the adaptive-bias quantum approximate optimization [500] Maxwell F. Parsons, Anton Mazurenko, Christie S.
algorithm,” Physical Review Research 5, 023147 (2023). Chiu, Geoffrey Ji, Daniel Greif, and Markus Greiner,
[486] David L. Applegate, Robert E. Bixby, Vašek Chvatál, “Site-resolved measurement of the spin-correlation func-
and William J. Cook, The Traveling Salesman Problem: tion in the fermi-hubbard model,” Science 353, 1253–
A Computational Study (Princeton University Press, 1256 (2016).
2006). [501] Hannes Bernien, Sylvain Schwartz, Alexander Keesling,
[487] Daniel Rehfeldt, Thorsten Koch, and Yuji Shinano, Harry Levine, Ahmed Omran, Hannes Pichler, Soon-
“Faster exact solution of sparse maxcut and qubo prob- won Choi, Alexander S. Zibrov, Manuel Endres, Markus
lems,” Mathematical Programming Computation 15, Greiner, Vladan Vuletić, and Mikhail D. Lukin, “Prob-
445–470 (2023). ing many-body dynamics on a 51-atom quantum simu-
[488] Fred Glover, Mark Lewis, and Gary Kochenberger, lator,” Nature 551, 579–584 (2017).
“Logical and inequality implications for reducing the [502] R. Blatt and C. F. Roos, “Quantum simulations with
size and difficulty of quadratic unconstrained binary op- trapped ions,” Nature Physics 8, 277–284 (2012).
timization problems,” European Journal of Operational [503] J. Zhang, G. Pagano, P. W. Hess, A. Kyprianidis,
Research 265, 829–842 (2018). P. Becker, H. Kaplan, A. V. Gorshkov, Z. X. Gong,
[489] Sanroku Tsukamoto, M. Takatsu, S. Matsubara, and and C. Monroe, “Observation of a many-body dynam-
Hirotaka Tamura, “An accelerator architecture for com- ical phase transition with a 53-qubit quantum simula-
binatorial optimization problems,” Fujitsu Scientific and tor,” Nature 551, 601–604 (2017).
Technical Journal 53, 8–13 (2017). [504] Iñigo Arrazola, Julen S. Pedernales, Lucas Lamata, and
[490] Satoshi Matsubara, Motomu Takatsu, Toshiyuki Enrique Solano, “Digital-analog quantum simulation of
Miyazawa, Takayuki Shibasaki, Yasuhiro Watanabe, spin models in trapped ions,” Scientific Reports 6, 30534
Kazuya Takemoto, and Hirotaka Tamura, “Digital (2016).
annealer for high-speed solving of combinatorial opti- [505] Andrew D. King, Sei Suzuki, Jack Raymond, Alex

68
 VIII CONCLUSION & OUTLOOK

Zucca, Trevor Lanting, Fabio Altomare, Andrew J. Physical Review Applied 17, 044046 (2022).
Berkley, Sara Ejtemaee, Emile Hoskinson, Shuiyuan [518] John Golden, Andreas Bärtschi, Daniel O’Malley, and
Huang, Eric Ladizinsky, Allison J. R. MacDonald, Stephan Eidenbenz, “Fair Sampling Error Analysis on
Gaelen Marsden, Travis Oh, Gabriel Poulin-Lamarre, NISQ Devices,” ACM Transactions on Quantum Com-
Mauricio Reis, Chris Rich, Yuki Sato, Jed D. Whittaker, puting 3, 8:1–8:23 (2022).
Jason Yao, Richard Harris, Daniel A. Lidar, Hidetoshi [519] Elijah Pelofske, John Golden, Andreas Bärtschi, Daniel
Nishimori, and Mohammad H. Amin, “Coherent quan- O’Malley, and Stephan Eidenbenz, “Sampling on NISQ
tum annealing in a programmable 2,000 qubit Ising Devices: “Who’s the Fairest One of All?”,” in IEEE
chain,” Nature Physics 18, 1324–1328 (2022). International Conference on Quantum Computing and
[506] Andrew D. King, Jack Raymond, Trevor Lanting, Engineering QCE’21 (2021) pp. 207–217.
Richard Harris, Alex Zucca, Fabio Altomare, Andrew J. [520] Laurence Wolsey, Integer Programming: 2nd Edition
Berkley, Kelly Boothby, Sara Ejtemaee, Colin Enderud, (John Wiley & Sons, Ltd, 2020).
Emile Hoskinson, Shuiyuan Huang, Eric Ladizinsky, Al- [521] Jo Devriendt, Stephan Gocht, Emir Demirović, Jakob
lison J. R. MacDonald, Gaelen Marsden, Reza Molavi, Nordström, and Peter J. Stuckey, “Cutting to the
Travis Oh, Gabriel Poulin-Lamarre, Mauricio Reis, core of pseudo-boolean optimization: Combining core-
Chris Rich, Yuki Sato, Nicholas Tsai, Mark Volkmann, guided search with cutting planes reasoning,” Proceed-
Jed D. Whittaker, Jason Yao, Anders W. Sandvik, and ings of the AAAI Conference on Artificial Intelligence
Mohammad H. Amin, “Quantum critical dynamics in 35, 3750–3758 (2021).
a 5,000-qubit programmable spin glass,” Nature 617, [522] Mukul R. Prasad, Armin Biere, and Aarti Gupta, “A
61–66 (2023). survey of recent advances in SAT-based formal verifica-
[507] Sergio Boixo, Troels F Rønnow, Sergei V Isakov, Zhihui tion,” International Journal on Software Tools for Tech-
Wang, David Wecker, Daniel A Lidar, John M Martinis, nology Transfer 7, 156–173 (2005).
and Matthias Troyer, “Evidence for quantum annealing [523] Nicholas Nethercote, Peter J. Stuckey, Ralph Becket,
with more than one hundred qubits,” Nat. Phys. 10, Sebastian Brand, Gregory J. Duck, and Guido Tack,
218–224 (2014). “Minizinc: Towards a standard cp modelling language,”
[508] Vasil S Denchev, Sergio Boixo, Sergei V Isakov, Nan in Principles and Practice of Constraint Programming –
Ding, Ryan Babbush, Vadim Smelyanskiy, John Marti- CP 2007 , edited by Christian Bessière (Springer Berlin
nis, and Hartmut Neven, “What is the computational Heidelberg, 2007) pp. 529–543.
value of finite-range tunneling?” Phys. Rev. X 6, 031015 [524] “Hexaly,” (2023), localSolver.
(2016). [525] Ryo Kuroiwa and J. Christopher Beck, “Solving domain-
[509] Hidetoshi Nishimori and Kabuki Takada, “Exponential independent dynamic programming problems with any-
Enhancement of the Efficiency of Quantum Annealing time heuristic search,” Proceedings of the International
by Non-Stoquastic Hamiltonians,” Frontiers in ICT 4 Conference on Automated Planning and Scheduling 33,
(2017). 245–253 (2023).
[510] Edson L. Padoin, Daniel A.G. de Oliveira, Pedro [526] N. J. A. Sloane, “Challenge problems: Independent sets
Velho, and Philippe O.A. Navaux, “Time-to-solution in graphs,” (2000).
and energy-to-solution: A comparison between arm and [527] Brent N. Clark, Charles J. Colbourn, and David S.
xeon,” in 2012 Third Workshop on Applications for Johnson, “Unit disk graphs,” Discrete Mathematics 86,
Multi-Core Architecture (2012). 165–177 (1990).
[511] Thomas Lubinski, Carleton Coffrin, Catherine Mc- [528] Ruben S. Andrist, Martin J. A. Schuetz, Pierre Minssen,
Geoch, Pratik Sathe, Joshua Apanavicius, and David Romina Yalovetzky, Shouvanik Chakrabarti, Dylan
E. Bernal Neira, “Optimization Applications as Quan- Herman, Niraj Kumar, Grant Salton, Ruslan Shaydulin,
tum Performance Benchmarks,” arXiv preprint (2023), Yue Sun, Marco Pistoia, and Helmut G. Katzgraber,
arXiv:2302.02278. “Hardness of the maximum-independent-set problem on
[512] Timo Berthold, “Measuring the impact of primal heuris- unit-disk graphs and prospects for quantum speedups,”
tics,” Operations Research Letters 41, 611–614 (2013). Physical Review Research 5 (2023), 10.1103/physrevre-
[513] Jorge Nocedal and Stephen J. Wright, Numerical Op- search.5.043277.
timization, 2nd ed. (Springer, New York, NY, USA, [529] Hannes Pichler, Sheng-Tao Wang, Leo Zhou, Soonwon
2006). Choi, and Mikhail D. Lukin, “Quantum optimization
[514] J.F. Bonnans, J.C. Gilbert, C. Lemarechal, and C.A. for maximum independent set using rydberg atom ar-
Sagastizábal, Numerical Optimization: Theoretical and rays,” (2018), arXiv:1808.10816 [quant-ph].
Practical Aspects, Universitext (Springer Berlin Heidel- [530] Xiaoling Wu, Xinhui Liang, Yaoqi Tian, Fan Yang,
berg, 2006). Cheng Chen, Yong-Chun Liu, Meng Khoon Tey, and
[515] Masayuki Yamamoto, Masayuki Ohzeki, and Kazuyuki Li You, “A concise review of Rydberg atom based quan-
Tanaka, “Fair sampling by simulated annealing on quan- tum computation and quantum simulation*,” Chinese
tum annealer,” Journal of the Physical Society of Japan Physics B 30, 020305 (2021).
89, 025002 (2020). [531] Minh-Thi Nguyen, Jin-Guo Liu, Jonathan Wurtz,
[516] Mario S. Könz, Guglielmo Mazzola, Andrew J. Ochoa, Mikhail D. Lukin, Sheng-Tao Wang, and Hannes Pich-
Helmut G. Katzgraber, and Matthias Troyer, “Uncer- ler, “Quantum optimization with arbitrary connectivity
tain fate of fair sampling in quantum annealing,” Phys. using Rydberg atom arrays,” PRX Quantum 4, 010316
Rev. A 100, 030303 (2019). (2023).
[517] Jon Nelson, Marc Vuffray, Andrey Y. Lokhov, Tameem [532] David Pisinger, “The quadratic knapsack problem—a
Albash, and Carleton Coffrin, “High-Quality Thermal survey,” Discrete Applied Mathematics 155, 623–648
Gibbs Sampling with Quantum Annealing Hardware,” (2007).

69
 VIII CONCLUSION & OUTLOOK

[533] G. Cornuéjols and M. Dawande, “A class of hard small prospects,” Rev. Phys. 4, 100028 (2019).
01 programs,” INFORMS Journal on Computing 11, [550] Dylan Herman, Cody Googin, Xiaoyuan Liu, Yue Sun,
205–210 (1999), also published in LNCS 1412 doi: Alexey Galda, Ilya Safro, Marco Pistoia, and Yuri Alex-
10.1007/3-540-69346-7_22. eev, “Quantum computing for finance,” Nature Reviews
[534] Rainer E. Burkard, Stefan E. Karisch, and Franz Rendl, Physics 5, 450–465 (2023).
“Qaplib –a quadratic assignment problem library,” Jour- [551] Casey Berger, Agustin Di Paolo, Tracey Forrest, Stuart
nal of Global Optimization 10, 391–403 (1997). Hadfield, Nicolas Sawaya, Michał Stęchły, and Karl
[535] David Van Bulck and Dries Goossens, “The interna- Thibault, “Quantum technologies for climate change:
tional timetabling competition on sports timetabling Preliminary assessment,” (2021), arXiv:2107.05362
(itc2021),” European Journal of Operational Research [quant-ph].
308, 1249–1267 (2023). [552] Yifan Zhou, Zefan Tang, Nima Nikmehr, Pouya Baba-
[536] D. de Werra, “Scheduling in sports,” in Annals of Dis- hajiani, Fei Feng, Tzu-Chieh Wei, Honghao Zheng, and
crete Mathematics (11), North-Holland Mathematics Peng Zhang, “Quantum computing in power systems,”
Studies, Vol. 59, edited by P. Hansen (North-Holland, iEnergy 1, 170–187 (2022).
1981) pp. 381–395. [553] Hari P. Paudel, Madhava Syamlal, Scott E. Crawford,
[537] David Van Bulck, Dries Goossens, Jan-Patrick Clarner, Yueh-Lin Lee, Roman A. Shugayev, Ping Lu, Paul R.
Angelos Dimitsas, George H. G. Fonseca, Carlos Lamas- Ohodnicki, Darren Mollot, and Yuhua Duan, “Quan-
Fernandez, Martin Mariusz Lester, Jaap Pedersen, tum computing and simulations for energy applications:
Antony E. Phillips, and Roberto Maria Rosati, “Which Review and perspective,” ACS Engineering 2, 151–196
algorithm to select in sports timetabling?” (2023), (2022).
arXiv:2309.03229 [cs.AI]. [554] Stavros Zenios, Practical Financial Optimization. De-
[538] Tameem Albash and Daniel A. Lidar, “Demonstration cision making for financial engineers (Wiley-Blackwell,
of a scaling advantage for a quantum annealer over sim- 2007).
ulated annealing,” Phys. Rev. X 8, 031016 (2018). [555] Shanker Ramamurthy, John J Duigenan, Hans Tes-
[539] Scott Kirkpatrick, C Daniel Gelatt, and Mario P selaar, Hector Arias, and Paolo Sironi, “Embedded
Vecchi, “Optimization by simulated annealing,” Science finance: Creating the everywhere, everyday bank,”
220, 671–680 (1983). (2023).
[540] Giuseppe E Santoro, Roman Martoňák, Erio Tosatti, [556] L. Bachelier, “Theory of speculation,” (1900).
and Roberto Car, “Theory of quantum annealing of an [557] F. A. Hayek, “The use of knowledge in society,” The
Ising spin glass,” Science 295, 2427–2430 (2002). American Economic Review 35, 519–530 (1945).
[541] Francesco Slongo, Philipp Hauke, Pietro Faccioli, and [558] Raphael A Espinoza, Miguel A. Segoviano, Ji Yan, “Sys-
Cristian Micheletti, “Quantum-inspired encoding en- temic risk modeling: How theory can meet statistics,”
hances stochastic sampling of soft matter systems,” Sci- (2020).
ence Advances 9 (2023). [559] European Central Bank, “Financial stability review,”
[542] Abraham P. Punnen, ed., The Quadratic Unconstrained (2023).
Binary Optimization Problem (Springer, 2022). [560] International Monetary Fund, “Global financial stabil-
[543] Ruslan Shaydulin and Marco Pistoia, “QAOA with ity report: Financial and climate policies for a high-
N · p ≥ 200,” in IEEE International Conference on interest-rate era,” (2023).
Quantum Computing and Engineering QCE’23 (2023) [561] European Comission, “Proposal for a regulation of the
pp. 1074–1077. european parliament and of the council laying down har-
[544] David Amaro, Matthias Rosenkranz, Nathan Fitz- monised rules on artificial intelligence (artificial intelli-
patrick, Koji Hirano, and Mattia Fiorentini, “A case gence act) and amending certain union legislative acts,”
study of variational quantum algorithms for a job shop (2023).
scheduling problem,” EPJ Quantum Technology 9, 5 [562] Basel Committee on Banking Supervision, Bank of In-
(2022). ternational Settlements, “Climate-related risk drivers
[545] Filip B Maciejewski, Jacob Biamonte, Stuart Hadfield, and their transmission channels,” (2021).
and Davide Venturelli, “Improving quantum approxi- [563] Tom Donilon, Catherine Kress, and Stephanie Lee
mate optimization by noise-directed adaptive remap- Jackson Spivey, “Geopolitical risk dashboard,” (2023).
ping,” arXiv preprint arXiv:2404.01412 (2024). [564] Fishcer Black and Myron Scholes, “The pricing of op-
[546] Elijah Pelofske, Andreas Bärtschi, Lukasz Cincio, John tions and corporate liabilities,” Journal of political econ-
Golden, and Stephan Eidenbenz, “Scaling Whole-Chip omy 81, 637 (1973).
QAOA for Higher-Order Ising Spin Glass Models on [565] Rama Cont, “Volatility clustering in financial markets:
Heavy-Hex Graphs,” (2023), arXiv:2312.00997 [quant- Empirical facts and agent-based models,” in Long Mem-
ph]. ory in Economics (Springer Berlin Heidelberg, Berlin,
[547] Qiskit contributors, “Qiskit: An Open-source Frame- Heidelberg, 2007) pp. 289–309.
work for Quantum Computing,” (2023). [566] Tim Bollerslev, “Generalized autoregressive conditional
[548] Daniel J. Egger, Claudio. Gambella, Jakub Marecek, heteroskedasticity,” Journal of Econometrics 31, 307–
Scott McFaddin, Martin Mevissen, Rudy Raymond, An- 327 (1986).
drea Simonetto, Stefan Woerner, and Elena Yndurain, [567] Steven L. Heston, “A Closed-Form Solution for Options
“Quantum computing for finance: state of the art and with Stochastic Volatility with Applications to Bond
future prospects,” IEEE Transactions on Quantum En- and Currency Options,” The Review of Financial Stud-
gineering 1, 1–24 (2020). ies 6, 327–343 (1993).
[549] Roman Orus, Samuel Mugel, and Enrique Lizaso, [568] J. Muzy, J. Delour, and E. Bacry, “Modelling fluctu-
“Quantum computing for finance: Overview and ations of financial time series: from cascade process to

70
 VIII CONCLUSION & OUTLOOK

stochastic volatility model,” Eur. Phys. J. B 17 , 537–548 and current trends,” European Journal of Operational
(2000). Research 234, 356–371 (2014), 60 years following Harry
[569] Thibault Jaisson Jim Gatheral and Mathieu Rosen- Markowitz’s contribution to portfolio theory and oper-
baum, “Volatility is rough,” Quantitative Finance 18, ations research.
933–949 (2018). [585] Dylan Herman, Cody Googin, Xiaoyuan Liu, Yue Sun,
[570] J. Donier P. Blanc and J.-P. Bouchaud, “Quadratic Alexey Galda, Ilya Safro, Marco Pistoia, and Yuri Alex-
hawkes processes for financial prices,” Quantitative Fi- eev, “Quantum computing for finance,” Nature Reviews
nance 17, 171–188 (2017). Physics 5, 450–465 (2023).
[571] Aditi Dandapani, Paul Jusselin, and Mathieu Rosen- [586] Alexander M. Dalzell, Sam McArdle, Mario Berta, Prze-
baum, “From quadratic Hawkes processes to super- myslaw Bienias, Chi-Fang Chen, András Gilyén, Con-
Heston rough volatility models with Zumbach effect,” nor T. Hann, Michael J. Kastoryano, Emil T. Khabi-
(2021), arXiv:1907.06151 [q-fin.ST]. boulline, Aleksander Kubica, Grant Salton, Samson
[572] Rama Cont, “Empirical properties of asset returns: styl- Wang, and Fernando G. S. L. Brandão, “Quantum al-
ized facts and statistical issues,” Quantitative Finance gorithms: A survey of applications and end-to-end com-
1, 223 (2001). plexities,” (2023), arXiv:2310.03011 [quant-ph].
[573] Jayakrishnan Nair, Adam Wierman, and Bert Zwart, [587] Harry Markowitz, “Portfolio selection,” The Journal of
The Fundamentals of Heavy Tails: Properties, Emer- Finance 7, 77–91 (1952).
gence, and Estimation, Cambridge Series in Statistical [588] Aharon Ben-Tal and Arkadi Nemirovski, Lectures on
and Probabilistic Mathematics (Cambridge University Modern Convex Optimization (Society for Industrial
Press, 2022). and Applied Mathematics, 2001).
[574] Organisation for Economic Co-operation and Develop- [589] Robert C. Merton, “An analytic derivation of the effi-
ment, “Environmental social and governance (esg) in- cient portfolio frontier,” The Journal of Financial and
vesting,” (2023). Quantitative Analysis 7, 1851–1872 (1972).
[575] Sebastian Brandhofer, Daniel Braun, Vanessa Dehn, [590] Nicholas I. M. Gould, Mary E. Hribar, and Jorge No-
Gerhard Hellstern, Matthias Hüls, Yanjun Ji, Ilia Po- cedal, “On the solution of equality constrained quadratic
lian, Amandeep Singh Bhatia, and Thomas Wellens, programming problems arising in optimization,” SIAM
“Benchmarking the performance of portfolio optimiza- Journal on Scientific Computing 23, 1376–1395 (2001).
tion with QAOA,” Quantum Information Processing 22, [591] William Karush, Minima of Functions of Several Vari-
25 (2022). ables with Inequalities as Side Conditions, Master’s the-
[576] Mark Hodson, Brendan Ruck, Hugh Ong, David sis, Department of Mathematics, University of Chicago,
Garvin, and Stefan Dulman, “Portfolio rebalancing Chicago, IL, USA (1939).
experiments using the Quantum Alternating Operator [592] Fred Glover, Gary Kochenberger, and Yu Du, “Quan-
Ansatz,” (2019), arXiv:1911.05296 [quant-ph]. tum Bridge Analytics I: a tutorial on formulating and
[577] Samuel Mugel, Carlos Kuchkovsky, Escolástico Sánchez, using QUBO models,” 4OR 17, 335–371 (2019).
Samuel Fernández-Lorenzo, Jorge Luis-Hita, Enrique [593] Richard O. Michaud and Tongshu Ma, “Efficient As-
Lizaso, and Román Orús, “Dynamic portfolio opti- set Management: A Practical Guide to Stock Portfolio
mization with real datasets using quantum processors Optimization and Asset Allocation.” The Review of Fi-
and quantum-inspired tensor networks,” Physical Re- nancial Studies 14, 901–904 (2015).
view Research 4, 013006 (2022). [594] Fischer Black and Robert Litterman, “Global portfo-
[578] Davide Venturelli and Alexei Kondratyev, “Reverse lio optimization,” Financial Analysts Journal 48, 28–43
quantum annealing approach to portfolio optimiza- (1992).
tion problems,” Quantum Machine Intelligence 1, 17–30 [595] Doron Avramov and Guofu Zhou, “Bayesian portfolio
(2019). analysis,” Annual Review of Financial Economics 2, 25–
[579] Megan C. Giron, Georgios Korpas, Waqas Parvaiz, 47 (2010).
Prashant Malik, and Johannes Aspman, “Approach- [596] Gérard Cornuéjols, Javier Peña, and Reha Tütüncü,
ing Collateral Optimization for NISQ and Quantum- Optimization Methods in Finance, 2nd ed. (Cambridge
Inspired Computing,” IEEE Transactions on Quantum University Press, 2018).
Engineering 4, 1–18 (2023). [597] M. L. Mehta, “Random matrices,” (2004).
[580] Stefan Woerner and Daniel J. Egger, “Quantum risk [598] Eugene P. Wigner, “Characteristic vectors of bordered
analysis,” npj Quantum Inf. 5, 15 (2019). matrices with infinite dimensions,” Annals of Mathe-
[581] Daniel J. Egger, Ricardo G. Gutierrez, Jordi Cahue matics 62, 548–564 (1955).
Mestre, and Stefan Woerner, “Credit risk analysis us- [599] John von Neumann and H. H. Goldstine, “Numerical in-
ing quantum computers,” IEEE Trans. Comput. 70, 1–1 verting of matrices of high order,” Bulletin of the Amer-
(5555). ican Mathematical Society 53, 1021 – 1099 (1947).
[582] Nikitas Stamatopoulos, Daniel J. Egger, Yue Sun, [600] Olivier Ledoit and Sandrine Péché, “Eigenvectors of
Christa Zoufal, Raban Iten, Ning Shen, and Stefan some large sample covariance matrix ensembles,” Prob-
Woerner, “Option pricing using quantum computers,” ability Theory and Related Fields 151, 233–264 (2011).
Quantum 4, 291 (2020). [601] Joël Bun, Romain Allez, Jean-Philippe Bouchaud, and
[583] Cristian S. Calude, Michael J. Dinneen, and Richard Marc Potters, “Rotational invariant estimator for gen-
Hua, “Qubo formulations for the graph isomorphism eral noisy matrices,” IEEE Transactions on Information
problem and related problems,” Theoretical Computer Theory 62, 7475–7490 (2016).
Science 701, 54–69 (2017). [602] Florent Benaych-Georges, Jean-Philippe Bouchaud,
[584] Petter N. Kolm, Reha Tütüncü, and Frank J. Fabozzi, and Marc Potters, “Optimal cleaning for singular values
“60 years of portfolio optimization: Practical challenges of cross-covariance matrices,” (2021), arXiv:1901.05543

71
 VIII CONCLUSION & OUTLOOK

[math.ST]. and F. Lorenz, “Towards optimization under uncer-

[603] Joël Bun, Jean-Philippe Bouchaud, and Marc Potters, tainty for fundamental models in energy markets using
“Cleaning large correlation matrices: Tools from ran- quantum computers,” (2023), arXiv:2301.01108 [quant-
dom matrix theory,” Physics Reports 666, 1–109 (2017), ph].
cleaning large correlation matrices: tools from random [618] Reza Mahroo and Amin Kargarian, “Hybrid quantum-
matrix theory. classical unit commitment,” in 2022 IEEE Texas Power
[604] Olivier Ledoit and Michael Wolf, “Nonlinear Shrink- and Energy Conference (TPEC) (IEEE, 2022).
age of the Covariance Matrix for Portfolio Selection: [619] Giuseppe Colucci, Stan van der Linde, and Frank
Markowitz Meets Goldilocks,” The Review of Financial Phillipson, “Power network optimization: A quantum
Studies 30, 4349–4388 (2017). approach,” IEEE Access 11, 98926–98938 (2023).
[605] Gary Chamberlain and Michael Rothschild, “Arbitrage, [620] Thomas Morstyn, “Annealing-based quantum comput-
factor structure, and mean-variance analysis on large ing for combinatorial optimal power flow,” IEEE Trans-
asset markets,” Econometrica 51, 1281–1304 (1983). actions on Smart Grid 14, 1093–1102 (2023).
[606] Richard O. Michaud and Robert O. Michaud, Efficient [621] Philipp Schworm, Xiangqian Wu, Marcel Wagner,
Asset Management: A Practical Guide to Stock Portfo- Svenja Ehmsen, Moritz Glatt, and Jan C. Aurich, “En-
lio Optimization and Asset Allocation, OUP Catalogue ergy supply scheduling in manufacturing systems using
No. 9780195331912 (Oxford University Press, 2008). quantum annealing,” Manufacturing Letters 38, 47–51
[607] Marcos Prado, “A robust estimator of the effi- (2023).
cient frontier,” SSRN Electronic Journal (2019), [622] European Comission, “Quarterly report on European
10.2139/ssrn.3469961. electricity markets. Volume 15. Issue 4. ,” (2023).
[608] Pierre-Alain Reigneron, Romain Allez, and Jean- [623] Financial Times, “France sets course for a nuclear re-
Philippe Bouchaud, “Principal regression analysis and naissance,” (2023).
the index leverage effect,” Physica A: Statistical Me- [624] INTERNATIONAL ENERGY AGENCY, “Global EV
chanics and its Applications 390, 3026–3035 (2011). Outlook 2023,” (2023).
[609] Armine Karami, Raphael Benichou, Michael Benza- [625] Renewable Energy World, “Aggreko delivers 2 MW/3.8
quen, and Jean-Philippe Bouchaud, “Conditional cor- MWh energy storage system for National Grid US,”
relations and principal regression analysis for futures,” (2023).
Wilmott 2021, 63–73 (2021). [626] Stergios Statharas, Yannis Moysoglou, Pelopidas Siskos,
[610] Vaclav Smil, “Distributed generation and megacities: Georgios Zazias, and Pantelis Capros, “Factors influ-
Are renewables the answer?” IEEE Power and Energy encing electric vehicle penetration in the EU by 2030:
Magazine 17, 37–41 (2019). A model-based policy assessment,” Energies 12 (2019),
[611] Naser Hossein Motlagh, Mahsa Mohammadrezaei, Ju- 10.3390/en12142739.
lian Hunt, and Behnam Zakeri, “Internet of things [627] Henrik Lund and Willett Kempton, “Integration of re-
(IoT) and the energy sector,” Energies 13 (2020), newable energy into the transport and electricity sectors
10.3390/en13020494. through V2G,” Energy Policy 36, 3578–3587 (2008).
[612] Guneet Bedi, Ganesh Kumar Venayagamoorthy, Rajen- [628] Bo Sun, Ali Zeynali, Tongxin Li, Mohammad Ha-
dra Singh, Richard R. Brooks, and Kuang-Ching Wang, jiesmaili, Adam Wierman, and Danny H.K. Tsang,
“Review of internet of things (iot) in electric power and “Competitive algorithms for the online multiple knap-
energy systems,” IEEE Internet of Things Journal 5, sack problem with application to electric vehicle charg-
847–870 (2018). ing,” Proc. ACM Meas. Anal. Comput. Syst. 4 (2021),
[613] N.P. Padhy, “Unit commitment-a bibliographical sur- 10.1145/3428336.
vey,” IEEE Transactions on Power Systems 19, 1196– [629] Bo Sun, Lin Yang, Mohammad Hajiesmaili, Adam
1205 (2004). Wierman, John C. S. Lui, Don Towsley, and
[614] Samantha Koretsky, Pranav Gokhale, Jonathan M. Danny H.K. Tsang, “The online knapsack problem with
Baker, Joshua Viszlai, Honghao Zheng, Niroj Gurung, departures,” Proc. ACM Meas. Anal. Comput. Syst. 6
Ryan Burg, Esa Aleksi Paaso, Amin Khodaei, Rozhin (2022), 10.1145/3570618.
Eskandarpour, and Frederic T. Chong, “Adapting [630] Bernard Knueven, James Ostrowski, and Jean-Paul
quantum approximation optimization algorithm (qaoa) Watson, “On mixed-integer programming formulations
for unit commitment,” in Proceedings of the 2021 IEEE for the unit commitment problem,” INFORMS Journal
International Conference on Quantum Computing and on Computing 32, 857–876 (2020).
Engineering (QCE) (2021) pp. 181–187. [631] Marika Federer, Daniel Müssig, Stefan Klaiber, Jörg
[615] Masaya Takahashi, Hiroaki Nishioka, Masahiro Hi- Lässig, Peter Bretschneider, and Steve Lenk, “Appli-
rai, and Hidetaka Takano, “A study of the opti- cation benchmark for quantum optimization on electro-
mization problem on the combination of sectionaliz- mobility use case,” in 2022 IEEE Vehicle Power and
ing switches in power grid with quantum annealing,” Propulsion Conference (VPPC) (2022) pp. 1–6.
(2023), arXiv:2305.07899 [quant-ph]. [632] Arkadi Nemirovski and Alexander Shapiro, “Convex ap-
[616] Pascal Halffmann, Patrick Holzer, Kai Plociennik, and proximations of chance constrained programs,” SIAM
Michael Trebing, “A quantum computing approach for Journal on Optimization 17, 969–996 (2007).
the unit commitment problem,” in International Con- [633] Oded Goldreich, “On promise problems (a survey in
ference on Operations Research, edited by O. Grothe, memory of shimon even [1935-2004]),” ECCC, TR05-
S. Nickel, S. Rebennack, and O. Stein (Springer, 2022) 018 127, 128 (2005).
pp. 113–120.
[617] M. C. Braun, T. Decker, N. Hegemann, S. F. Kerstan,

72