Quantum computational chemistry

Sam McArdle,1, ∗ Suguru Endo,1 Alán Aspuru-Guzik,2, 3, 4 Simon C. Benjamin,1 and Xiao Yuan1, †
1
Department of Materials,
University of Oxford,
Parks Road, Oxford OX1 3PH,
United Kingdom
2
Department of Chemistry and Department of Computer Science,
University of Toronto,
Toronto, Ontario M5S 3H6,
Canada
3
Vector Institute for Artificial Intelligence,
Toronto, Ontario M5S 1M1,
Canada
4
Canadian Institute for Advanced Research (CIFAR) Senior Fellow,
arXiv:1808.10402v3 [quant-ph] 27 Jan 2020

Toronto, Ontario M5S 1M1,

Canada
(Dated: January 28, 2020)

One of the most promising suggested applications of quantum computing is solving

classically intractable chemistry problems. This may help to answer unresolved ques-
tions about phenomena like: high temperature superconductivity, solid-state physics,
transition metal catalysis, or certain biochemical reactions. In turn, this increased un-
derstanding may help us to refine, and perhaps even one day design, new compounds
of scientific and industrial importance. However, building a sufficiently large quantum
computer will be a difficult scientific challenge. As a result, developments that enable
these problems to be tackled with fewer quantum resources should be considered very
important. Driven by this potential utility, quantum computational chemistry is rapidly
emerging as an interdisciplinary field requiring knowledge of both quantum computing
and computational chemistry. This review provides a comprehensive introduction to
both computational chemistry and quantum computing, bridging the current knowl-
edge gap. We review the major developments in this area, with a particular focus on
near-term quantum computation. Illustrations of key methods are provided, explicitly
demonstrating how to map chemical problems onto a quantum computer, and solve
them. We conclude with an outlook for this nascent field.

CONTENTS 1. STO-nG and split-valence basis sets 17

2. Correlation-consistent basis sets 18
I. Introduction 2 3. Plane wave basis sets 18
E. Reduction of orbitals 19
II. Quantum computing and simulation 4
A. Quantum computing 4 IV. Quantum computational chemistry mappings 19
B. Quantum simulation 6 A. First quantised encoding methods 20
1. Grid based methods 20
III. Classical computational chemistry 8 2. Basis set methods 21
A. The electronic structure problem 9 B. Second quantised basis set encoding methods 22
1. Chemical systems of interest 9 1. Jordan-Wigner encoding 22
B. First and Second quantisation 10 2. Bravyi–Kitaev encoding 23
1. First quantisation 11 3. Locality preserving mappings 24
a. Grid based methods 11 C. Resource reduction 24
b. Basis set methods 11 1. Hamiltonian reduction 24
2. Second quantisation 12 2. Low rank decomposition techniques 25
a. Basis set methods 12
b. Grid based methods 13 V. Quantum computational chemistry algorithms 25
C. Classical computation methods 13 A. Quantum phase estimation 25
1. Hartree–Fock 14 1. Implementation 25
2. Multiconfigurational self-consistent field 15 2. State preparation 27
3. Configuration interaction 15 3. Hamiltonian simulation 27
4. Coupled cluster 16 a. Product formulae 28
D. Chemical basis sets 17 b. Advanced Hamiltonian simulation methods 28
4. Phase estimation for chemistry simulation 28
a. Gaussian orbital basis sets 29
b. Plane wave basis sets 29
∗
c. Fermi-Hubbard model 30
[email protected] d. Grid based methods 30
† [email protected] 5. Outstanding problems 31
 2

B. Variational algorithms 31 Song et al., 2019, 2017), and photonic systems (Chen
1. Ansätze 32 et al., 2017; Wang et al., 2016; Zhong et al., 2018). It is
a. Hardware efficient ansätze 33
b. Chemically inspired ansätze 33
believed that using quantum systems as our simulation
c. Hamiltonian variational ansatz 34 platform will enable us to tackle classically intractable
2. Measurement 34 problems in chemistry, physics, and materials science.
3. Classical optimisation 35 Classical computational methods have become an im-
a. Previous optimisation studies 35
b. Related methods of optimisation 36
portant investigative tool in areas like transition metal
C. Evaluation of excited states 36 catalysis (Vogiatzis et al., 2019) and high temperature
1. Quantum subspace expansion 37 superconductivity (Dagotto, 1994). Classical simula-
2. Overlap-based methods 38 tions enable us to rationalise experimental results, test
3. Contraction VQE methods 38
physical models, and understand system properties.
VI. Error mitigation for chemistry 38 However, their ability to guide design is often precluded
A. Extrapolation 39 by the computational complexity of realistic models.
B. Probabilistic error cancellation 39 As quantum computers are able to efficiently simulate
C. Quantum subspace expansion 40
D. Symmetry based methods 41 quantum systems, it is believed that they will enable a
E. Other methods of error mitigation 41 more accurate understanding of the models in use today,
as well as the ability to simulate more complex (and
VII. Illustrative examples 41
therefore, more realistic) models. This may lead to an
A. Hydrogen 41
1. STO-3G basis 42 increased understanding which we can leverage to make
2. 6-31G basis 43 advances in areas as diverse as chemistry (Aspuru-Guzik
3. cc-PVDZ basis 44 et al., 2018), biology (Reiher et al., 2017), medicine
B. Lithium Hydride STO-3G basis 44
(Cao et al., 2018), and materials science (Babbush et al.,
VIII. Discussion and Conclusions 46 2018c). It has even been speculated that as quantum
A. Classical limits 46 hardware develops, quantum simulation may one day
B. Quantum resources: medium to long-term 48 progress from being accurate enough to confirm the
C. Quantum resources: near to medium-term 49
D. Summary and outlook 51
results of experiments, to being more accurate than
the experiments themselves. Quantum simulations of
Acknowledgments 51 such high precision may in turn enable the design of
References 52
new, useful compounds. However, we stress that to
achieve this ultimate goal we would need considerable
further developments – both in the technology required
I. INTRODUCTION to build such as powerful quantum computer, and the
theory behind an appropriate algorithm and model.
Quantum mechanics underpins all of modern chem- This can be likened to the aerospace industry, where
istry. One might therefore imagine that we could use computational fluid dynamics calculations on classical
this theory to predict the behaviour of any chemical com- computers have replaced wind tunnel testing in many
pound. This is not the case. As Dirac noted; “The ex- stages of wing design (Jameson, 1999). However, for
act application of these laws leads to equations much too the most demanding parts of aerospace design, neither
complicated to be soluble.” (Dirac, 1929). The problem our largest classical computers, nor the physical models
described by Dirac is that the complexity of the wave- considered, are yet powerful enough to completely
function of a quantum system grows exponentially with replace experimental testing (Malik and Bushnell, 2012).
the number of particles. This leaves classical comput- Instead, the two methods work together in synergy, to
ers unable to exactly simulate quantum systems in an enable increased understanding, with greater efficiency.
efficient way. Feynman proposed a solution to this prob-
lem; using quantum hardware as the simulation platform, To date, several efficient quantum algorithms have
remarking that “If you want to make a simulation of na- been proposed to solve problems in chemistry. The
ture, you’d better make it quantum mechanical, and by runtime and physical resources required by these algo-
golly it’s a wonderful problem, because it doesn’t look so rithms are expected to scale polynomially with both the
easy.” (Feynman, 1982). size of the system simulated and the accuracy required.
Although developing small quantum computers has Experimental developments have accompanied these
taken over 30 years, we may soon be in a position to theoretical milestones, with many groups demonstrating
test Feynman’s proposal, following recent developments proof of principle chemistry calculations. However,
in quantum hardware including ion traps (Ballance limited by hardware capabilities, these experiments
et al., 2016; Gaebler et al., 2016; Harty et al., 2014; focus only on small chemical systems that we are already
Monz et al., 2011), superconducting systems (Arute able to simulate classically. Moreover, the gate counts
et al., 2019; Barends et al., 2014; Chow et al., 2012; currently estimated for transformative chemistry simula-
 3

tions likely signal the need for quantum error correction, map chemistry problems onto a quantum computer.
which requires orders of magnitude more qubits, and We discuss techniques that can be used to reduce
lower error rates, than are currently available (Babbush the simulation resources required, and the quantum
et al., 2018b; Kivlichan et al., 2019a; Mueck, 2015). circuits that can be used. This section seeks to illustrate
Despite ongoing experimental efforts, no group has yet the techniques described throughout the rest of the
demonstrated a single fully error corrected qubit. Even if review, providing worked examples for the reader. We
the significant hardware challenges to build an error cor- conclude this review in Sec. VIII with a comparison
rected quantum computer can be overcome (Gambetta between classical and quantum techniques, and resource
et al., 2017; Ladd et al., 2010; Monroe and Kim, 2013), estimations for the different quantum methods. This
new theoretical developments may be needed to solve section aims to help the reader to understand when, and
classically intractable chemistry problems on a quantum how, quantum computational chemistry may surpass its
computer that we could realistically imagine building in classical counterpart.
the next few decades, such as probing biological nitrogen
fixation, or investigating new metal ion battery designs. A handful of related reviews on this topic exist in
These breakthroughs may be achieved by connecting the literature. Summaries of early theoretical and
researchers working in quantum computing with those experimental work in quantum computational chemistry
working in computational chemistry. We seek to aid were carried out by Kassal et al. (2011) and Lu et al.
this connection with this succinct, yet comprehensive, (2012). More focused discussions of quantum algorithms
review of quantum computational chemistry (using introduced for chemistry simulation before 2015, and the
quantum algorithms, run on quantum computers, to computational complexity of problems in chemistry can
solve problems in computational chemistry) and its be found in works by Kais et al. (2014); Veis and Pittner
foundational fields. (2012); and Yung et al. (2012). A comprehensive review
was recently released by Cao et al. (2019). Said review,
Although quantum algorithms can solve a range of and our own, are complementary; the review of Cao
problems in chemistry, we focus predominantly on the et al. (2019) is well suited to experienced practitioners
problem of finding the low lying energy levels of chem- of classical electronic structure theory, and provides
ical systems. This is known as ‘the electronic structure excellent detail on the computational complexity of
problem’. There are three reasons for this restriction quantum simulations, and how they asymptotically
of scope. Primarily, this problem is a fundamental one compare to cutting edge methods in classical chemistry.
in classical computational chemistry. Knowledge of the Our review provides a more practical guide to the
energy eigenstates enables the prediction of reaction field (especially for those new to electronic structure
rates, location of stable structures, and determination of theory), showing explicitly how the workhorse methods
optical properties (Helgaker et al., 2012). Secondly, the of classical chemistry have been translated to work
machinery developed to solve this problem on quantum on quantum computers, and describing techniques to
computers is easily applied to other types of problems, facilitate experimental demonstrations of quantum
such as finding transition states, or understanding the chemistry algorithms, such as resource reduction and
vibrational structure of molecules. Finally, most of the error mitigation. Together, these reviews provide a
prior work in quantum computational chemistry has complete overview of the progress to date in quantum
focused on this problem. As such, it provides an ideal computational chemistry.
context in which to explain the most important details
of quantum computational chemistry. Despite being a relatively young field, quantum
computational chemistry has grown extremely rapidly,
This review is organised as follows. We first provide a and has already evolved beyond the stage that it can
brief overview of quantum computing and simulation in be fully described by a single review. As such, there
Sec. II. We then introduce the key methods and terminol- are approaches to solving chemistry problems with a
ogy used in classical computational chemistry in Sec. III. quantum computer which we are not able to describe
The methods developed to merge these two fields, includ- fully in this review. As stated above, we have chosen to
ing mapping chemistry problems onto a quantum com- prioritise the canonical topics in the field: using either
puter, are described in Sec. IV. We continue our discus- near-term, or further-future digital quantum computers
sion of quantum computational chemistry in Sec. V by to solve the electronic structure problem ab initio.
describing algorithms for finding the ground and excited We have focused on the most promising methods for
states of chemical systems. Sec. VI highlights the tech- solving this problem: variational algorithms with error
niques developed to mitigate the effects of noise in non- mitigation, and the quantum phase estimation algorithm
error corrected quantum computers, which will be crucial with quantum error correction. Extended discussion is
for achieving accurate simulations in the near-future. reserved for methods which are key to understanding
In Sec. VII we provide several examples of how to how quantum computers can be used to solve general
 4

chemistry problems, or articles which have made im- The canonical circuit model of quantum computation
portant observations on ways to make these simulations is so-named because of its resemblance to the logic cir-
more tractable. It is beyond the scope of this review to cuits used in classical computing. In the classical circuit
summarise work in directions complementary to these, model, logic gates (such as AND, OR and NOT) act on
such as: quantum machine learning based approaches to bits of information. In the quantum case, quantum gates
the electronic structure problem (Xia et al., 2018; Xia are acted upon the basic unit of information, the qubit.
and Kais, 2018), using quantum computers as part of a The qubit lives in a two-dimensional Hilbert space. The
problem decomposition approach to simulation (Bauer basis vectors of the space are denoted as {|0i , |1i}, which
et al., 2016; Dallaire-Demers and Wilhelm, 2016a,b; are referred to as the computational basis states,
Keen et al., 2019; Kreula et al., 2016; Rubin, 2016;    
Rungger et al., 2019), hybrid quantum algorithms for 1 0
density functional theory (Hatcher et al., 2019; Whitfield |0i =   , |1i =   . (1)
et al., 2014), relativistic quantum chemistry (Senjean, 0 1
2019; Veis et al., 2012), gate based methods for sim-
ulating molecular vibrations (McArdle et al., 2018b; A general single qubit state is described by
Sawaya and Huh, 2018; Sawaya et al., 2019), analog  
simulators of molecular vibrations (Chin and Huh, 2018; α
|ϕi = α |0i + β |1i =   , (2)
Clements et al., 2017; Hu et al., 2018a; Huh et al., 2015; β
Huh and Yung, 2017; Joshi et al., 2014; Shen et al.,
2018; Sparrow et al., 2018; Wang et al., 2019), fermionic α, β ∈ C,
quantum computation for chemistry simulation (O’Brien |α| + |β|2 = 1.
2
et al., 2018b), quantum methods for electron-phonon
systems (Macridin et al., 2018a,b; Wu et al., 2002), When quantum logic gates act on the qubits, they ma-
protein folding and molecular docking (Babbush et al., nipulate both basis state vectors at the same time, pro-
2012; Babej et al., 2018; Banchi et al., 2019; Fingerhuth viding (measurement limited) parallelism. Although the
et al., 2018; Lu and Li, 2019; Perdomo et al., 2008; qubit is in a quantum superposition during the algorithm,
Perdomo-Ortiz et al., 2012; Robert et al., 2019), solving when it is measured in the computational basis, it will
problems in chemistry using a quantum annealer (Bab- be found in state |0i or state |1i, not in a superposition.
bush et al., 2014; Genin et al., 2019; Teplukhin et al., These measurement outcomes occur with probability |α|2
2018; Xia et al., 2018), and quantum algorithms for and |β|2 , respectively. For now, we will treat the qubit
finding the eigenvalues of non-hermitian Hamiltoni- as an abstract two level system, before later elaborating
ans (Daskin et al., 2014; Wang et al., 2010). on how they can be physically realised.
If there are n qubits in the system, the state is de-
scribed by a vector in the 2n dimensional Hilbert space
formed by taking the tensor product of the Hilbert spaces
II. QUANTUM COMPUTING AND SIMULATION of the individual qubits. States can be classified as either
‘product’ or ‘entangled’. Product states can be decom-
In this section, we introduce the basic elements of posed into tensor products of fewer qubit wavefunctions,
quantum computing and quantum simulation. We re- such as
fer the reader to the works of Georgescu et al. (2014)
1 1
and Nielsen and Chuang (2002) for more detailed intro- √ (|00i + |01i) = |0i ⊗ √ (|0i + |1i). (3)
ductions. 2 2
Entangled states cannot be decomposed into tensor prod-
ucts, such as the state
A. Quantum computing
1
√ (|00i + |11i). (4)
In this review, we focus on the qubit-based circuit 2
model of quantum computation (Nielsen and Chuang,
2002). Other paradigms that vary to a greater or lesser In this review, we refer to the leftmost qubit in a vector
extent include: adiabatic quantum computing (Aharonov as the (n − 1)th qubit, and the rightmost qubit as the
et al., 2008; Albash and Lidar, 2018; Farhi et al., zeroth qubit. This choice enables us to write numbers
2000), one-way or measurement based quantum com- in binary using computational basis states. For example,
puting (Jozsa, 2005; Raussendorf and Briegel, 2001; we can write |7i = |1i |1i |1i = |111i. We can then place
Raussendorf et al., 2003), and continuous-variable quan- a quantum register of n qubits in a superposition of the
tum computing (Braunstein and van Loock, 2005; Lloyd 2n possible numbers that can be represented by n bits.
P2n −1
and Braunstein, 1999). This is typically written as x=0 cx |xi.
 5

A quantum circuit consists of a number of single and |0 Had •

two qubit gates acting on the qubits. The qubits are
initialised in a well defined state, such as the |0̄i state |0
⊗n
(|0̄i = |0i = |0i ⊗ |0i ⊗ · · · ⊗ |0i). A quantum cir-
cuit generally concludes with measurements to extract FIG. 2 A quantum
√ circuit that generates the entangled state
(|00i + |11i)/ 2 and measures each qubit in the computa-
information. It may also employ additional intermediate
tional basis. Time runs from left to right. Here, Had is the
measurements, for example, to check for errors. From a Hadamard gate, defined in the main text. When measured,
mathematical perspective, the gates are unitary matri- the qubits will either both be in the state 0 (|00i), or both 1
ces. Typical gates include the Pauli gates (|11i). Each of these two outcomes occurs with 50 % proba-
      bility.
0 1 0 −i 1 0
X= , Y =  , Z =   , (5)
1 0 i 0 0 −1 ~ where i, j
denote each gate by a unitary operator U i,j (θ),
are the indices of the qubits the gates act on (i = j for
the single qubit rotation gates
single qubit operations), and θ~ are gate parameters (al-
−iθX
) , R (θ) = e( −iθY ) , R (θ) = e( −iθZ though the gates do not have to be parametrized, such as
Rx (θ) = e( 2
y 2
z 2 )
the Pauli gates). We can then mathematically describe
(6)
a quantum circuit by
the Hadamard and T gates Y i ,j
    |ϕi = Ukk k (θ~k ) |0̄i , (10)
k
1 1 1 1 0
Had = √  , T= , (7)
2 1 −1 0 eiπ/4 where k denotes the k th gate in the circuit. The gates are
ordered right to left. For example, the circuit in Fig. 2
and multi-qubit entangling gates, such as the two qubit would be written as
controlled-NOT (CNOT) gate shown in Fig. 1. The ac-
• 1
√ (|00i + |11i) = CNOT0,1 Had0 |00i . (11)
2

FIG. 1 The controlled-NOT (CNOT) gate. ‘•’ denotes the We extract information from the circuits by performing
control qubit and ‘⊕’ denotes the target qubit. measurements of observables, O, which are represented
by Hermitian matrices. Typically, we seek the average
tion of the CNOT gate can be written mathematically value over many measurements, Ō, given by
as
Ō = hϕ| O |ϕi , (12)
|0i h0|C ⊗ IT + |1i h1|C ⊗ XT , (8)
referred to as the expectation value of the operator O.
where T denotes the target qubit, and C denotes the Measuring the expectation value of qubit i in the com-
control qubit. The matrix form of this operation is given putational basis corresponds to hϕ| Zi |ϕi. In practice,
by this means that we repeatedly prepare the state |ϕi, and
  measure the state of qubit i, labelling the outcomes +1
1 0 0 0 (for |0i) and −1 (for |1i). We then take the mean of these



 measurement outcomes. In order to measure qubits in
 0 1 0 0 the X or Y basis, single qubit rotations are first applied
 . (9)

 0 0 0 1
 to change the basis of the relevant qubits, which are then
  measured in the Z basis. We can obtain the outcomes of
0 0 1 0 measuring multi-qubit operators by taking the product
of the measurement outcomes of single qubit operators
These gates are used to create an example quantum measured in the same circuit iteration. For example, the
circuit in Fig. 2. This circuit generates the entangled expectation value of Zi Zj could be obtained by prepar-
state of 2 qubits given by Eq. (4), and then measures ing state |ϕi, measuring Zi on qubit i and Zj on qubit j,
both of the qubits. multiplying these two measurement outcomes together,
With only single qubit operations and CNOT gates, it and averaging the results over many repetitions of this
is possible to approximate an arbitrary multi-qubit gate process. These outcomes are typically correlated for en-
to any desired accuracy (DiVincenzo, 1995). As a result, tangled states. For example, the measurement outcome
the circuit model of quantum computing typically decom- for Z1 Z0 on the state in Eq. (4) is always +1 (+1 × +1
poses all algorithms into single and two qubit gates. We for |00i and −1 × −1 for |11i).
 6

The Pauli operators and identity matrix, multiplied manner (Campbell et al., 2017; Fowler et al., 2012a;
by real coefficients, form a complete basis for any O’Gorman and Campbell, 2017). For example, if we
Hermitian single or multi-qubit operator. Therefore any consider using a quantum computer to factor numbers
multi-qubit observable can be expanded into strings of in polynomial time (Shor, 1994), current estimates (at
Pauli operators, the expectation values of which we can time of writing) suggest that we would require around
measure efficiently with a quantum computer. 20 million physical qubits to factor a number that is too
large to tackle using known classical algorithms (Gidney
It is important to distinguish between the number and Eker, 2019). Building a machine of this size will
of physical and logical qubits in a quantum computer. be extremely difficult, in terms of isolating the qubits
Physical qubits are approximate two-level systems, which from the environment, developing scalable control
can be created in a range of different systems, including, systems, and minimising qubit crosstalk, as discussed
but not limited to: energy levels in trapped ions (Cirac by Gambetta et al. (2017) and Ladd et al. (2010). We
and Zoller, 1995; Leibfried et al., 2003), polarization will discuss the comparably lower resources required for
states of photons (Knill et al., 2001), spins in quantum error corrected chemistry calculations in Sec. VIII.B.
dots (Hanson et al., 2007; Loss and DiVincenzo, 1998)
or silicon (Kane, 1998), and energy levels of supercon- In contrast to the large error corrected machines de-
ducting circuit resonators (Nakamura et al., 1999; Shnir- scribed above, current quantum computers possess only
man et al., 1997; Wendin, 2017). In order to protect our tens of error-prone physical qubits. Nevertheless, quan-
qubits from decoherence caused by coupling to the envi- tum computers with more than around 50 qubits are
ronment (Landauer, 1995; Unruh, 1995), we can encode considered too large to exactly simulate classically, and
m logical qubits into n > m physical qubits. These log- may thus be capable of solving problems which are
ical qubits can simply be thought of as the abstract two intractable on even the largest classical supercomput-
level system described by Eq. (2). The codes used to ers (Arute et al., 2019). However, these problems are
construct logical qubits are analogous to classical error typically artificially constructed examples, rather than
correcting codes, but are in general more complex, due real-world problems (Boixo et al., 2018; Harrow and Mon-
to the delicate nature of quantum information and the tanaro, 2017). Preskill (2018) dubbed these machines
‘no-cloning theorem’ of quantum mechanics. Depending ‘noisy intermediate-scale quantum’ (NISQ) devices, and
on the code used, we can either detect, or detect and observed that it is currently unclear whether they will be
correct the errors which occur. The number of errors we able to outperform classical computers on useful tasks.
can detect and/or correct depends on the code used (it The dichotomy between the resources needed for tackling
is related to the distance of the code). We must also ac- problems like factoring, and the ‘supremacy’ of a machine
count for the fact that the error checking measurements with more than around 50 qubits poses the question;
and correction procedure can cause additional errors to ‘What, if anything, will near-term quantum computers be
occur (Knill et al., 1996). We seek to build circuits in useful for?’. The answer may lie with Feynman’s original
a ‘fault tolerant’ (Aharonov and Ben-Or, 1997; Gottes- proposal; using quantum systems to simulate quantum
man, 1998; Shor, 1996) manner, which limits the spread systems.
of errors during logical blocks of the computation. If this
is achieved, then it is possible to scale up computations
arbitrarily. If the physical error rate in the gates is be- B. Quantum simulation
low a certain (code-dependent) threshold value, the error
rate in the logical operations can be made arbitrarily low, In this review, we focus on the digital quantum
either by concatenation, or for certain codes, by growing simulation of many-body quantum systems. Digital
the code. A more detailed discussion of error correction quantum simulation maps the target problem onto a
is given by Devitt et al. (2013); Lidar and Brun (2013); set of gates which can be implemented by a quantum
Raussendorf (2012); and Terhal (2015). computer. A universal quantum computer can be
One of the most widely studied error correction codes programmed to perform many different simulations.
is the surface code (Kitaev, 1997), which is particularly This can be contrasted with analog quantum simulation,
suitable for 2D grids of qubits with nearest-neighbour where the simulator emulates a specific real system
connectivity. Physical error rates below the surface code of interest. However, analog simulators are generally
threshold of around 1 % (Fowler et al., 2012b; Stephens, considered more robust to noise, and therefore easier to
2014; Wang et al., 2011) have been achieved for trapped construct (Georgescu et al., 2014). To date, there have
ion (Ballance et al., 2016; Gaebler et al., 2016) and been several proposals for the simulation of chemistry
superconducting (Arute et al., 2019; Barends et al., using analog simulators (Argüello-Luengo et al., 2018;
2014) qubits. However, with these error rates, we would Chin and Huh, 2018; Huh et al., 2015; Huh and Yung,
require around 103 − 104 physical qubits per logical 2017; Torrontegui et al., 2011), some of which have been
qubit to perform interesting tasks in a fault-tolerant experimentally realised (Clements et al., 2017; Hu et al.,
 7

2018b; Shen et al., 2018; Smirnov et al., 2007; Sparrow BQP (bounded-error quantum polynomial-time) (Bern-
et al., 2018). Nevertheless, to perform accurate simu- stein and Vazirani, 1997; Lloyd, 1996). It is important
lations of large chemical systems, we will likely require to note that, while it is widely believed to be true, it has
digital quantum simulation, as it is not yet clear how not yet been mathematically proven that BPP 6= BQP
to protect large analog simulations from errors. Digital (where BPP is the complexity class containing problems
quantum simulation is more vulnerable to noise and that are solvable in polynomial time on a probabilistic
device imperfections than analog simulation. While such classical computer). Further discussion of general Hamil-
imperfections can be addressed via error correction, this tonian simulation methods is given in Sec. V.A.3.
requires additional qubits and places stringent require- It has also been shown that time evolution of open and
ments on gate fidelities. In this review we focus solely closed quantum systems can be simulated using varia-
on digital quantum simulation of chemistry problems. tional approaches (Chen et al., 2019; Endo et al., 2018b;
We refer the reader to the works of Aspuru-Guzik and Heya et al., 2019; Li and Benjamin, 2017). A variational
Walther (2012); Blatt and Roos (2012); Georgescu et al. circuit is one with parametrized quantum gates, whose
(2014); Houck et al. (2012); and Schneider et al. (2012) parameters are updated according to an algorithm spe-
for information about digital quantum simulation of cific update rule. They are discussed in more detail in
other physical systems, and analog quantum simulation. Sec. V.B. These techniques may enable simulation of time
evolution using circuits with fewer gates than Trotteriza-
The numerous problems in chemistry that can be tion. However, a variational circuit with set parameters
simulated on a quantum computer can be divided into is tailored to the time evolution of one or more specific
static and dynamics problems. Here, we use ‘dynamics’ initial states, in contrast to a Trotter circuit, which can
to mean evolving wavefunctions in time and seeing be used to time evolve any valid initial state.
how certain observables vary, as opposed to chemical Once the system has been time evolved for the desired
reaction dynamics, which are discussed separately below. duration, we can extract useful dynamical quantities
from these simulations. Examples of such quantities
Methods for solving dynamics problems were for- include the electronic charge density distribution, or
malised by Lloyd (1996) and further developed by particle correlation functions (Abrams and Lloyd,
Abrams and Lloyd (1997). As illustrated in Fig. 3, we 1997). Further information on quantum dynamics sim-
can map the system Hamiltonian, Hs , to a qubit Hamil- ulation can be found in the review by Brown et al. (2010).
tonian, Hq . We similarly map the initial system wave-
function |ψsi i to a qubit representation |ψqi i. We can In chemistry, one is often concerned with determining
then evolve the qubit wavefunction in time by mapping whether two or more sub-systems will react with each
the system time evolution operator, e−itHs , to a series of other, when brought together with a certain energy.
gates. This can be achieved using a Lie-Trotter-Suzuki One might assume that this could be studied by simply
decomposition (Trotter, 1959), commonly referred to as initialising the reactants on the quantum computer,
Trotterization or a Trotter circuit. This means that if and time evolving under the system Hamiltonian, using
the Hamiltonian of the system, Hs can be written as the methods described above. However, it depends
X on the method used to map the chemical problem
Hs = hj , (13) onto the quantum computer. We must take care to
j ensure that our model is able to accurately describe
where hi are local terms which act on a small subset of the system during all parts of the reaction. This is
the particles in the system, then a small time evolution naturally taken care of using grid based methods,
under the Hamiltonian can be decomposed as where the electrons and nuclei are treated on an equal
footing, as discussed in Sec. III.B.1 and Sec. IV.A.1.
In contrast, if the problem is projected onto a finite
P Y
e−iHs δt = e−i j hj δt ≈ e−ihj δt + O(δt2 ). (14)
j
basis set of electron spin-orbitals, we must be careful
to ensure that: 1) The nuclear dynamics are accurately
The number of terms in the Hamiltonian scales polyno- described, either through a precise classical treatment
mially with the number of particles for systems of in- of the nuclear dynamics, or by using additional orbitals
terest in chemistry, due to the two-body nature of the to describe the nuclear motion, and 2) The electron
Coulomb interactions between particles (Helgaker et al., spin-orbitals used are able to accurately describe the
2014). Each of the exponential terms in Eq. (14) can be positions of the electrons at all points in the reaction.
realised efficiently using a quantum computer. As the This may require the orbitals to change in time. To the
dynamics of local Hamiltonians can be efficiently simu- best of our knowledge, only the work by Berry et al.
lated on a quantum computer, but are generally thought (2019a) considers a chemical reaction dynamics cal-
to be inefficient to simulate on a classical computer, this culation on a quantum computer using basis set methods.
problem belongs to the computational complexity class
 8

𝑈 𝑡 = 𝑒 %&'( )

|0⟩ Had Um
U
|0⟩ Had Um
U
|0⟩ Had Um
…
…

𝑒 %&'- .) 𝑒 %&'- .) 𝑒 %&'- .) 𝑒 %&'- .)

|0⟩ Had Um

State preparation Evolution Measurement

FIG. 3 Digital quantum simulation of time evolution of a spin chain, using a canonical Trotter-type method. We first map
the system Hamiltonian, Hs , to a qubit Hamiltonian, Hq . Then the initial system wavefunction |ψsi i is mapped to a qubit
wavefunction |ψqi i. The time evolution of the system can be mapped to a Trotterized circuit that acts on the initial qubit
wavefunction. Finally, well chosen measurements are applied P to extract the desired
P information, such as particle correlation
functions. For a spin chain with an Ising Hamiltonian, H = hi,ji Jij Zi Zj + i Bi Xi , where the first sum is over nearest-
neighbour spins i and j, the unitaries Uij are given by Uij = CNOTi,j Rzj (2Jij ) CNOTi,j Rxi (2Bi ).

We can obtain static properties by mapping the tar- on finding the low lying energy levels of chemical sys-
get wavefunction (such as the ground state wavefunction tems (solving the electronic structure problem). It is
of the system) onto a qubit wavefunction. We can then widely believed that this problem should be efficiently
use the quantum computer to calculate the expectation solvable with a quantum computer, for physically rel-
value of the desired observable, hψq | Oq |ψq i. In partic- evant systems (Whitfield et al., 2013). The electronic
ular, Abrams and Lloyd (1999) showed that the phase structure problem has received significant attention since
estimation algorithm (Kitaev, 1995) can be used to find it was first introduced in the context of quantum com-
the energy of a quantum system, and collapse it into the putational chemistry by Aspuru-Guzik et al. (2005), and
desired energy eigenstate. We will discuss this approach is widely considered to be one of the first applications
in Sec. V.A. The ground state problem can also be tack- of quantum computing. Solving the electronic structure
led using variational algorithms (Peruzzo et al., 2014), problem is often a starting point for more complex cal-
which we will discuss in detail in Sec. V.B. culations in chemistry, including the calculation of re-
Finding the low lying energy levels of a quantum action rates, the determination of molecular geometries
Hamiltonian is in general an exponentially difficult prob- and thermodynamic phases, and calculations of optical
lem for classical computers. Moreover, it is important properties (Helgaker et al., 2012).
to note that solving the ground state problem for a Before discussing how the electronic structure problem
completely general local Hamiltonian is QMA-complete can be solved using a quantum computer in Sec. IV and
(quantum Merlin-Arthur complete, the quantum ana- Sec. V, we first summarise the classical methods used
logue of NP-complete) (Cubitt and Montanaro, 2016; to solve this problem in Sec. III. Many of these meth-
Kempe et al., 2006). Problems in this complexity class ods have formed the basis of the work done thus far in
are not believed to be efficiently solvable with either a quantum computational chemistry.
classical or quantum computer. For such systems, it
would appear that Nature itself cannot efficiently find III. CLASSICAL COMPUTATIONAL CHEMISTRY
the ground state.
Despite this, the situation is not as bleak as it may In this section, we introduce a selection of the most
initially seem. As stated in the introduction, we focus widely used techniques in ab initio classical computa-
 9

tional chemistry. As discussed in the introduction, we Our aim is to find energy eigenstates |Ei i and the
focus on tools developed to solve the electronic structure corresponding energy eigenvalues Ei of the Hamiltonian
problem. The problem is formulated in Sec. III.A, and He . In the rest of this review, we drop the subscript
translated into the language of first and second quantisa- e. In particular, we are interested in the ground state
tion in Sec. III.B. In Sec. III.C we describe the different energy and the lowest excited state energies. We can
approximations that can be used to make this problem solve this equation for a range of nuclear configurations
tractable for classical computers. In Sec. III.D we review to map out the potential energy surfaces of the molecule.
some of the common spin-orbital basis functions used in Mapping out these potential energy curves explicitly
basis set approaches to the molecular electronic structure is exponentially costly in the degrees of freedom of
problem. We discuss orbital basis changes, and their the molecule, and that there are a variety of methods
use in reducing the simulation resources in Sec. III.E. being developed to solve this difficult problem more
We have sought to produce a self-contained summary efficiently (Christiansen, 2012).
of the essential knowledge required for quantum compu-
tational chemistry, and we refer the reader to Helgaker We wish to measure the energy to an accuracy of at
et al. (2014) and Szabo and Ostlund (2012) for further least 1.6 × 10−3 Hartree, known as ‘chemical accuracy’.
information. If the energy is known to chemical accuracy, then the
chemical reaction rate at room temperature can be pre-
dicted to within an order of magnitude using the Eyring
A. The electronic structure problem equation (Evans and Polanyi, 1935; Eyring, 1935)
The Hamiltonian of a molecule consisting of K nuclei Rate ∝ e−∆E/kB T , (18)
and N electrons is
X ~2 X ~2 X e2 ZI where T is the temperature of the system, and ∆E is
H =− ∇2i − ∇2I −
2m e 2M I 4π |r
0 i − RI | the energy difference between the reactant and product
i I i,I
states. In computational chemistry, we are typically more
1 X e2 1 1 X e2 ZI ZJ interested in the relative energies of two points on the
+ + ,
2 4π0 |ri − rj | 2 4π0 |RI − RJ | potential energy surface than the absolute energy of a
i6=j I6=J
(15) single point. Even if the individual energy values can-
where MI , RI , and ZI denote the mass, position, and not be measured to within chemical accuracy, there is
atomic number of the I th nucleus, and ri is the posi- often a fortuitous cancellation of errors, which leads to
tion of the ith electron. The first two sums in H are the energy difference being found to chemical accuracy.
the kinetic terms of the electrons and nuclei, respectively. However, in this review we consider chemical accuracy to
The final three sums represent the Coulomb repulsion be- mean an error of less than 1.6 × 10−3 Hartree in the en-
tween: the electrons and nuclei, the electrons themselves, ergy value at every point on the potential energy surface.
and the nuclei themselves, respectively. For conciseness,
we work in atomic units, where the unit of length is
a0 = 1 Bohr (0.529 × 10−10 m), the unit of mass is
the electron mass me , and the unit of energy is 1 Hartree 1. Chemical systems of interest
(1 Hartree = e2 /4π0 a0 = 27.211 eV). Denoting
MI0 = MI /me , the Hamiltonian in atomic units becomes While classical computational chemistry has made
X ∇2 X ∇2 tremendous progress in describing and predicting the
i I
X ZI properties of a multitude of systems, there are some
H =− − 0 −
i
2 2MI |ri − RI | systems that appear classically intractable to simulate
I i,I
(16) with currently known techniques. Consequently, there
1X 1 1 X ZI ZJ
+ + . has been significant interest in the possibility of using
2 |ri − rj | 2 |RI − RJ |
i6=j I6=J quantum computers to efficiently solve these problems.
We are predominantly interested in the electronic In particular, we are interested in solving so-called
structure of the molecule. As a nucleon is over one thou- ‘strongly correlated systems’ (we will explain this term
sand times heavier than an electron, we apply the Born- more carefully in Sec. III.C.1. Here, it suffices to say
Oppenheimer approximation, treating the nuclei as clas- that these are systems which possess wavefunctions
sical point charges. As a result, for a given nuclear con- with a high degree of entanglement. Many systems of
figuration one only needs to solve the electronic Hamil- commercial and scientific interest, such as catalysts and
tonian high temperature superconductors, are believed to be
X ∇2 X strongly correlated. In this section, we discuss two such
i ZI 1X 1
He = − − + . (17) strongly correlated systems that have been identified
i
2 |ri − R I | 2 |r i − rj | as interesting potential targets for a future quantum
i,I i6=j
 10

computer. et al., 2006). While more complex models are likely re-
quired to fully understand the mechanism of cuprate su-
Many transition metals have found use as cata- perconductivity, a complete understanding of even the
lysts (Vogiatzis et al., 2019). However, many of these Fermi-Hubbard model is still elusive. Close to half fill-
systems show strong correlation, often due to their open- ing, at intermediate interaction strengths, the system ap-
shell nature, and spatially degenerate states (Lyakh pears to show several competing orders in its phase di-
et al., 2012). This strong correlation often precludes agram, which makes it difficult to reliably extract the
their study in silico, forcing us to carry out expen- ground state properties from classical numerical calcu-
sive, trial-and-error based discovery – especially for lations (Fradkin et al., 2015). Interestingly, this is the
transition metal based biological catalysts (Podewitz same regime that is believed to be most relevant to under-
et al., 2011). One such system is the biological en- standing cuprate superconductors (LeBlanc et al., 2015).
zyme nitrogenase, which enables microorganisms which Properties are typically obtained by performing ground
contain it to convert atmospheric dinitrogen (N2 ) to state calculations on the Fermi-Hubbard model for a
ammonia (NH3 ) under ambient conditions. This process range of different system sizes, and then extrapolating
is known as nitrogen fixation, and is one of the two to the thermodynamic limit (LeBlanc et al., 2015). We
main ways of producing ammonia for fertiliser – the discuss the system sizes that can be tackled using modern
other is the energy intensive Haber-Bosch process. The classical techniques in Sec. VIII.A. We then examine the
Haber-Bosch process requires high temperatures and quantum resources required to surpass these calculations
pressures, and is believed to consume up to 2 % of in Sec. VIII.B.
the world’s energy output (Reiher et al., 2017). While
there has been significant experimental progress in
understanding the structure of the nitrogenase enzyme B. First and Second quantisation
over the last 100 years, the reaction mechanism is
not yet fully understood (Burgess and Lowe, 1996). As a consequence of the Pauli exclusion principle, the
The crux of understanding the enzyme appears to be electronic wavefunction must be antisymmetric under the
several transition metal compounds found within it: exchange of any two electrons. This antisymmetrisation
the P cluster (containing iron and sulphur) and the can be accounted for in two ways, known as first and
iron molybdenum cofactor (FeMo-co, containing iron, second quantisation. These names are largely histori-
molybdenum, carbon, hydrogen and oxygen) (Hoffman cal; first quantisation was the approach initially taken
et al., 2014). Computational models of FeMo-co have by the pioneers of quantum mechanics, whereby variables
been proposed which are beyond the reach of current like position and momentum are promoted to being op-
classical methods for solving strongly correlated systems erators (they are ‘quantised’). Second quantisation was
(see Sec. VIII.A) but would be accessible with a small developed afterwards, and quantises fields, rather than
error corrected quantum computer (Berry et al., 2019b; variables. There are key differences between these repre-
Reiher et al., 2017). We discuss the resources required sentations, which affect how simulations of physical sys-
for these simulations in Sec. VIII.B. tems are carried out using a quantum computer.
As will be discussed in Sec. III.B.1, first quantised
High temperature superconductors are also believed methods explicitly retain the antisymmetry in the wave-
to be strongly correlated systems. Since their discov- function. In contrast, second quantisation maintains the
ery in the 1980’s, there has been significant experimen- correct exchange statistics through the properties of the
tal and theoretical work to understand these compounds, operators which are applied to the wavefunction, as we
which are not well described by the Bardeen-Cooper- will show in Sec. III.B.2. These differences will become
Schrieffer theory of superconductivity. In the case of the more apparent in the context of quantum computational
so-called cuprate superconductors, it is widely believed chemistry mappings, which we will discuss in Sec. IV.
that the mechanism of high temperature superconductiv- It is important to note that whether a simulation is
ity is closely linked to the physics of the copper-oxygen carried out in first or second quantisation is distinct from
planes that comprise them. A complete analytic or com- whether a ‘basis set’ or ‘grid based’ method is used. This
putational understanding of these layers is beyond the ca- will be elaborated on in more detail in the following sec-
pabilities of current classical techniques. Several simpli- tions. However, one key difference is that using a basis
fied models have been proposed, which have been found set is known as a ‘Galerkin discretisation’, which ensures
to capture some of the important behaviour of high tem- that the energy converges to the correct value from above,
perature superconductors (Dagotto, 1994). In particu- as the number of basis functions tends to infinity. This
lar, models of fermions hopping on a 2D square lattice, property does not hold for grid based methods. A more
such as the one-band Fermi-Hubbard model (Hubbard, detailed discussion on the differences between grid based
1963), and the related t-J model, appear to reproduce and basis set methods can be found in the main text and
many experimental observations (Anderson, 2002; Lee Appendix A of Babbush et al. (2018c).
 11

1. First quantisation We project the Hamiltonian onto M basis wavefunc-

tions, {φp (xi )}, where xi is the spatial and spin coordi-
Here, we focus on classical first quantised simulation nate of the ith electron, xi = (ri , σi ). These basis func-
methods. Discussion of first quantised chemistry simula- tions approximate electron spin-orbitals. The grid based
tion on quantum computers is postponed until Sec. IV.A. method described above directly stores the wavefunction
without exploiting any knowledge we may have about the
general spatial form of the orbitals. In contrast, basis set
methods exploit this knowledge to reduce the resources
a. Grid based methods needed to simulate chemical systems. We write the many
We consider the wavefunction in the position electron wavefunction as a Slater determinant, which is
representation, {|ri}, which must be explicitly anti- an antisymmetrised product of the single electron basis
symmetrised to enforce exchange symmetry. Mathemat- functions. The wavefunction is given by
ically, we describe the N electron wavefunction as
Z ψ(x0 . . . xN−1 ) =
|Ψi = ψ(x1 , . . . , xN )A (|x1 , . . . , xN i) dx1 , . . . , dxN
 
 
φ (x ) φ (x ) ... φ (x
M −1 0 )
x1 ,...,xN  0 0 1 0
 

(19) 
 φ0 (x1 ) φ1 (x1 ) ... φM −1 (x1 ) 

where A denotes antisymmetrisation, xi = (ri , σi ) =  
th  (21)
 
(xi , yi , zi , σi ) gives the position and spin of the i elec- 1   . . . .
√ .
tron and ψ(x1 , x2 , . . . , xN ) = hx1 , x2 , . . . , xN |Ψi. We N!  
. . . .


can simulate this system on a classical computer by eval- 



uation of the wavefunction on a discretised spatial grid.  . . . . 
 
However, the cost of storing the wavefunction scales ex- 
φ0 (xN−1 ) φ1 (xN−1 ) ... φM −1 (xN−1 )

ponentially with the number of electrons, N . Suppose
each axis of space is discretised into P equidistant points. Swapping the positions of any two electrons is equiva-
The discretised wavefunction is given by lent to interchanging two rows of the Slater determinant,
X which changes the sign of the wavefunction. This pro-
|Ψi = ψ(x1 , . . . , xN )A (|x1 , . . . , xN i) , (20) vides the correct exchange symmetry for the fermionic
x1 ,...,xN
wavefunction. While the number of spin-orbitals consid-
where |xi i = |ri i |σi i is a spatial and spin-coordinate, ered, M , is typically larger than the number of electrons
|ri i = |xi i |yi i |zi i , ∀i ∈ {1, 2, . . . , N }, xi , yi , zi ∈ in the system, N , the electrons can only occupy N of
{0, 1, . . . , P − 1}, and σi ∈ {0, 1}. In total, there are the spin-orbitals in a given Slater determinant. As a re-
P 3N N
× 2 complex amplitudes, showing that the mem- sult, the Slater determinant only contains the N occupied
ory required scales exponentially with the size of the sim- spin-orbitals. For example, imagine a fictitious system
ulated system. This makes it classically intractable to with two electrons, distributed among basis functions
simulate more than a few particles on a grid using a clas- A(x) and B(x). Each of these basis functions could be oc-
sical computer. Consequently, we do not discuss classical cupied by an electron of either spin, so we effectively work
methods that are specific to the grid based mapping in with four basis functions {A↑ (x), A↓ (x), B↑ (x), B↓ (x)}.
this review. We consider the case where both electrons are in the A
Grid based methods are useful when considering orbitals (and therefore have opposite sz values, where sz
chemical reaction dynamics, or when simulating systems is the z component of the spin of the electron). As dis-
for which the Born-Oppenheimer approximation is not cussed above, the Slater determinant only contains the
appropriate. In these scenarios, we must include the N occupied spin-orbitals. We use Eq. (21) to obtain the
motion of the nuclei. If we consider the nuclear motion wavefunction
separately, we need to obtain the potential energy ψ(x0 , x1 ) =
surfaces from electronic structure calculations. As  
 
mentioned in the previous section, mapping out these 1  A (x
↑ 0 ) A (x
↓ 0 )
potential energy surfaces is exponentially costly. As √  =
2 A↑ (x1 ) A↓ (x1 ) (22)
such, it may be better to treat the nuclei and electrons  
on an equal footing, which is best achieved with grid 1
√ A↑ (x0 )A↓ (x1 ) − A↓ (x0 )A↑ (x1 ) .
based methods. This is discussed further by Kassal et al. 2
(2008).
This wavefunction is antisymmetric under the exchange
of the two electrons, as required.
As we will see in Sec. III.B.2, the information encoded
b. Basis set methods by a Slater determinant can be compressed by moving
 12

to the second quantised formalism. As a consequence, fermionic operators on the determinants are given by
first quantised basis set methods are rarely, if ever, used
ap |fM −1 , fM −2 , . . . , f0 i
in classical computational chemistry calculations. Nev- Pp−1
ertheless, it is important to be aware of what the wave- fi
=δfp ,1 (−1) i=0 |fM −1 , fM −2 , . . . , fp ⊕ 1, . . . , f0 i ,
function looks like in the first quantised basis set rep-
resentation for two reasons. Firstly, as discussed above,
the second quantised basis set method follows directly a†p |fM −1 , fM −2 , . . . , f0 i
Pp−1
from the first quantised basis set approach. Secondly, =δfp ,0 (−1) fi
|fM −1 , fM −2 , . . . , fp ⊕ 1, . . . , f0 i ,
i=0

first quantised basis set approaches have found use in (25)

quantum computational chemistry, as we will discuss in where ⊕ denotes addition modulo 2,P such that 0 ⊕ 1 = 1,
Sec. IV.A.2. p−1
1 ⊕ 1 = 0. The phase term (−1) i=0 fi enforces the
exchange anti-symmetry of fermions. The spin-orbital
occupation operator is given by
2. Second quantisation n̂i = a†i ai ,
n̂i |fM −1 , . . . , fi , . . . , f0 i = fi |fM −1 , . . . , fi , . . . , f0 i ,
a. Basis set methods (26)
The second quantised basis set approach follows nat- and counts the number of electrons in a given spin-
urally from the first quantised basis set method discussed orbital.
in Sec. III.B.1. We again project the Hamiltonian onto
M basis wavefunctions, {φp (xi )}, and consider many- Observables must be independent of the representa-
electron wavefunctions that must be antisymmetric un- tion used. Therefore, the expectation values of second
der the exchange of any two electrons. As we saw in quantised operators must be equivalent to the expecta-
Sec. III.B.1, to write down a Slater determinant we only tion values of the corresponding first quantised opera-
need to indicate which spin-orbitals are occupied by elec- tors. As first quantised operators conserve the number
trons. This enables the introduction of a convenient of electrons, the second quantised operators must contain
shorthand for Slater determinants (Szabo and Ostlund, an equal number of creation and annihilation operators.
2012) We can use these requirements to obtain the second quan-
tised form of the electronic Hamiltonian (Helgaker et al.,
ψ(x0 . . . xN−1 ) = |fM −1 , . . . , fp , . . . , f0 i = |f i , (23) 2014; Szabo and Ostlund, 2012).
X 1 X
where fp = 1 when φp is occupied (and therefore present H= hpq a†p aq + hpqrs a†p a†q ar as , (27)
p,q
2 p,q,r,s
in the Slater determinant), and fp = 0 when φp is empty
(and therefore not present in the determinant). The with
vector |f i is known as an occupation number vector, !
∇2 X
Z
and the space of all such vectors is known as ‘Fock ZI
hpq = dxφ∗p (x) − − φq (x),
space’. The second quantised formalism is concerned 2 |r − RI |
I
with manipulating these occupation number vectors. Z
φ∗p (x1 )φ∗q (x2 )φr (x2 )φs (x1 )
As these occupation number vectors are a convenient hpqrs = dx1 dx2 .
|r1 − r2 |
short-hand for Slater determinants, we will refer to them
(28)
throughout this review as Slater determinants. This is
The first integral represents the kinetic energy terms
common practice in computational chemistry (Szabo
of the electrons, and their Coulomb interaction with
and Ostlund, 2012).
the nuclei. The second integral is due to the electron-
electron Coulomb repulsion. The Hamiltonian only con-
Electrons are excited into the single electron spin- tains terms with up to four creation and annihilation
orbitals by fermionic creation operators, a†p . They are operators (two creation, two annihilation) because the
de-excited by annihilation operators, ap . These opera- Coulomb interaction between the electrons is a two-body
tors obey fermionic anti-commutation relations interaction. As a result, the Hamiltonian contains up
to M 4 terms, depending on the basis functions used.
{ap , a†q } = ap a†q + a†q ap = δpq , We examine the form of these basis functions, and how
(24) to select them in Sec. III.D. A special case of the elec-
{ap , aq } = {a†p , a†q } = 0.
tronic structure Hamiltonian is obtained for the Fermi-
Hubbard model, introduced in Sec. III.A.1. The Fermi-
The determinants |f i form an orthonormal basis for Hubbard Hamiltonian considers fermions hopping be-
the Fock space of the system. The actions of the tween nearest-neighbour lattice sites with strength t.
 13

These fermions feel a repulsive (or attractive) force U scales exponentially with the number of electrons, mak-
when they occupy the same lattice site, i. The Hamilto- ing large calculations classically intractable.
nian is given by Second quantised basis set methods are the most
X  †  widely used approach in classical computational chem-
ai,σ aj,σ + a†j,σ ai,σ + U
X
H = −t ni,↑ ni,↓ (29) istry, and have formed the basis of most of the work
hi,ji,σ i done so far in quantum computational chemistry. As a
result, we discuss some of the approximate methods used
where hi, ji denotes a sum over nearest-neighbour lattice in second quantised basis set simulations in Sec. III.C
sites, and σ is a spin-coordinate. This Hamiltonian has and Sec. III.D. In Sec. III.C we consider making ground
only O(M ) terms, where M is the number of spin-sites. state calculations classically tractable by approximating
For convenience, throughout the rest of this review we the exact ground state wavefunction with a restricted
will refer to both molecular spin-orbitals and lattice spin number of Slater determinants. In Sec. III.D we consider
sites as spin-orbitals. approximating the exact wavefunction by considering
only the most important molecular orbitals. However,
Let us consider general and approximate solutions of first we briefly discuss the limited work that has been
the electronic structure Hamiltonian. If the electron- done on second quantised grid based methods, for the
electron Coulomb interaction term in Eq. (17) is ne- sake of completeness.
glected, we obtain a new Hamiltonian which describes
the behaviour of N independent electrons. We can de-
fine a suitable basis for this fictitious system as the set of
molecular orbitals which diagonalise the non-interacting b. Grid based methods
Hamiltonian. These molecular orbitals are typically lin- To the best of our knowledge, second quantised
ear combinations of orbitals localised around each of the grid based methods have only been discussed by Bab-
atoms. We note that in practice, the molecular orbitals bush et al. (2018c) in their Appendix A. Nevertheless,
obtained by diagonalising the non-interacting part of the these methods follow naturally from the discussion of
Hamiltonian will likely form a poor basis for the sys- second quantised basis sets above. Our discussion of
tem. Instead, a mean-field approximation (the Hartree- the topic closely follows the derivation of Babbush et al.
Fock procedure, to be described in Sec. III.C.1) can be (2018c). As a first step, we can consider our real space
used to obtain more suitable molecular orbitals. As grid to be described by a set of basis functions that are
these single-particle molecular orbitals are chosen such delta functions, δ(r − ri ), each positioned at grid point
that they diagonalise the non-interacting (or mean-field) ri = (xi , yi , zi ). The creation operators a†i,σ then become
Hamiltonian, energy eigenstates can be formed by tak-
a†xi ,yi ,zi ,σ ; rather than creating an electron in spin-orbital
ing tensor products of each electron in a different molec-
i, they now create an electron with spin σ at the grid
ular spin-orbital. In order to obey the Pauli exclusion
point (xi , yi , zi ) in 3D space. As the basis functions do
principle, these tensor products must be correctly anti-
not overlap in space, the kinetic energy operator must be
symmetrised. This can be achieved by creating Slater
defined using a finite difference formula, rather than the
determinants from the molecular orbitals, as described
integral in Eq. (28). We must also define a suitable in-
by Eq. (21).
ner product between functions defined on the grid. These
As they are eigenstates of a Hermitian operator, these
steps allow us to calculate the coefficients of each term in
Slater determinants form a complete basis of the problem
the Hamiltonian. As discussed above, this second quan-
Hilbert space. Consequently, the eigenstates of the true
tised grid based method has not yet (to the best of our
Hamiltonian can be expressed as linear combinations of
knowledge) been used in any classical or quantum com-
these Slater determinants, written as
putational chemistry algorithms.
X
|Ψi = αf |f i , (30)
f
C. Classical computation methods
where αf are complex coefficients which we refer to herein
as the determinant amplitudes. These solutions are ex- In this section, we review four methods for approximat-
act, provided that the molecular orbitals form a complete ing the ground state wavefunction with a restricted num-
basis for the single particle states, and the N -electron ber of Slater determinants: the Hartree-Fock (HF), mul-
wavefunction contains all of the determinants that these ticonfigurational self-consistent field (MCSCF), configu-
molecular orbitals can generate (Helgaker et al., 2014; ration interaction (CI), and coupled cluster (CC) meth-
Szabo and Ostlund, 2012). If all M


N determinants are ods. These methods create parametrized trial states,
included, the wavefunction is known as the full config- which can then be optimised to approach the ground
uration interaction (FCI) wavefunction. However, this state (to an accuracy determined by the approximations
wavefunction contains a number of determinants which made). In order to isolate the errors arising from the
 14

method used, we assume that we are working in the the old orbitals). This process is repeated until the or-
full molecular orbital basis for our molecule, although bitals converge (Szabo and Ostlund, 2012). The new or-
in practice this would be classically intractable for large bitals obtained are referred to as the canonical orbitals.
system sizes (with the size of the system dependent upon This procedure generates single particle molecular or-
the accuracy of the method used). Restricting the size of bitals from combinations of the single particle atomic
the basis set will be discussed in Sec. III.D. orbitals.
The methods discussed below are considered in the The term hikkj describes the Coulomb interaction of
context of second quantised basis set calculations, as an electron with the charge distribution of the other elec-
these translate most easily to the methods used in quan- trons, while the term hikjk describes exchange effects
tum computational chemistry. These methods are among (also called Fermi correlation) arising from the required
the most straightforward and widely used in classical antisymmetrisation. However, as a mean-field solution,
computational chemistry. the HF method neglects the effects of dynamic and static
correlation in the wavefunction.
Dynamic correlation is a typically small correc-
1. Hartree–Fock tion (Hattig et al., 2012), arising from the Coulomb
repulsion between electrons. Wavefunctions display-
The Hartree–Fock (HF) method is a mean-field tech- ing dynamic correlation are often dominated by the
nique which aims to find the dominant Slater determi- Hartree-Fock determinant, and have small additional
nant in the system wavefunction. This is achieved by op- contributions from a (potentially large) number of ex-
timising the spatial form of the spin-orbitals to minimise cited state determinants. Static correlation occurs when
the energy of the wavefunction. We generally consider a more than one Slater determinant is equally dominant
set of spin-orbitals, M , that is larger than the number of in the wavefunction (Hattig et al., 2012). In this case,
electrons in the molecule, N . As we only consider a single the Hartree-Fock method provides a poor approximation
Slater determinant, we are essentially assuming that N to the ground state wavefunction. The presence of
of the spin-orbitals are occupied, and M − N are left un- strong static correlation can be evidenced by multiple
occupied, or virtual. In the HF method, we first neglect near-degenerate solutions of the Hartree-Fock procedure.
the Coulomb repulsion term in the electronic structure Static correlation can arise because the wavefunction
Hamiltonian (Eq. (17)), reducing the problem to one of may require several determinants to be coupled in a
N independent electrons. We then assume that each elec- proper spin configuration, or during bond breaking in
tron moves in the average charge distribution of all of the order to account for the separation of the electrons
other electrons, which introduces an effective potential. onto the products (the latter case is often referred to
We can solve the N coupled equations iteratively; first as non-dynamic, or left-right correlation) (Lyakh et al.,
calculating the position of each electron, then updating 2012). The relative contribution of these effects depends
the potential, and repeating this process until the orbitals on the orbital basis used. Lyakh et al. (2012) note
converge. In the second quantised formalism, this proce- that, while it is sometimes possible to reduce the level
dure is carried out by using the orbitals to construct the of static correlation by manually enforcing correct spin
‘Fock operator’, and diagonalising the Fock operator to symmetries, and using appropriately localised orbitals,
obtain new orbitals. This process is repeated until the it is in general not possible to avoid strong static
orbitals converge, and so HF is also referred to as the correlation when considering the whole of the potential
self-consistent field (SCF) method. The Fock operator, energy surface. As a result, static correlation often
fˆ, is given by (Helgaker et al., 2014) dominates in many systems of scientific interest, such
as excited states (Lischka et al., 2018), or transition
(hij + Vij )a†i aj ,
X
fˆ = metals (Vogiatzis et al., 2019).
i,j
(31)
The Slater determinant generated from a HF calcula-
X
Vij = (hikkj − hikjk ),
k∈occ
tion is typically taken as the reference state for post-HF
methods, such as configuration interaction and coupled
where Vij describes the effective potential, and occ is the cluster, which seek to capture some of the dynamic corre-
set of occupied orbitals. We see that the Fock operator lation energy by including additional determinants, de-
depends on the spatial form of the orbitals through hij , scribing excitations above the HF state. Although we
hikkj , and hikjk which are obtained by calculating the will discuss orbital basis sets in more detail in Sec. III.D,
integrals in Eq. (28). When performing a HF calcula- we note here that the HF orbitals are not suitable for
tion, we typically input a set of atomic orbitals, which describing the virtual orbitals that electrons are excited
are localised around each atom. These orbitals are used into. This is because the HF method only optimises the
to calculate the Fock operator, which is then diagonalised occupied orbitals in the single Slater determinant wave-
to obtain new orbitals (which are linear combinations of function. In order to obtain suitable virtual orbitals, we
 15

can instead perform correlated calculations on individual all of the determinants that could be generated from
atoms. The details of how to perform these correlated distributing a certain number of electrons in these or-
atomic calculations are beyond the scope of this review, bitals. MCSCF and CASSCF calculations are among
and are not relevant for quantum computing applications. the most effective classical methods at treating systems
We refer the interested reader to Section 8.3 of the text- with strong static correlation (Szalay et al., 2012). Re-
book by Helgaker et al. (2014) for additional details. The cent approaches, including replacing the CASSCF sub-
importance of each of the virtual atomic orbitals is deter- routine with tensor network methods, have enabled the
mined by their contribution to either the electron density treatment of even larger active spaces (Knecht et al.,
(atomic natural orbitals) or the atomic correlation energy 2016). However, while the methods described above are
(correlation-consistent basis sets). These two metrics are often effective at recovering the static correlation energy,
closely related, and while either can be used, the lat- they struggle to recover the dynamic correlation energy.
ter produces more compact basis sets, and hence is used This requires additional techniques, such as multirefer-
more frequently in practice (Helgaker et al., 2014). We ence configuration interaction (see Sec. III.C.3 and Sza-
will discuss the use of correlation consistent basis sets lay et al. (2012)) and multireference coupled cluster (see
further in Sec. III.D.2. Sec. III.C.4 and Lyakh et al. (2012)) .

2. Multiconfigurational self-consistent field 3. Configuration interaction

As discussed above, the Hartree-Fock method performs The configuration interaction (CI) method generates a
poorly for strongly correlated systems. Systems with correlated wavefunction by considering excitations above
strong static correlation are defined as those where mul- a reference state, typically the Hartree-Fock state. If
tiple Slater determinants are equally dominant. These all determinants are included, we recover the full con-
include excited states (Lischka et al., 2018), transition figuration interaction (FCI) wavefunction, generated by
states (Szalay et al., 2012), systems at the dissociation considering all excitations above the Hartree-Fock wave-
limit (Lyakh et al., 2012), and transition metals (Vo- function
giatzis et al., 2019). In order to account for static cor-
relation, we need to use a wavefunction which exhibits |ΨFCI i
 
the required multireference nature. One such approach
Ciα a†i aα + Cijαβ a†i a†j aα aβ + ... |ΨHF i,
X X
is the multiconfigurational self-consistent field (MCSCF) = I +
method. The MCSCF approach considers a wavefunc- i,α i>j,α>β
tion with several Slater determinants, and variationally (32)
optimises both the molecular orbitals, and the deter- where C are parameters to be optimised according to
minant amplitudes simultaneously (Roos et al., 1980). the Rayleigh-Ritz variational principle. The Rayleigh-
Mathematically,
P we write our MCSCF wavefunction as Ritz variational principle states that the energy expec-
|Ψi = e−κ f αf |f i, where again αf are determinant tation value of a parametrized wavefunction is greater
amplitudes, and |f i are Slater determinants. Here, κ is than or equal to the lowest energy eigenvalue of the
P †
an anti-Hermitian operator given by κ = i,j kij ai aj . Hamiltonian being measured. As considering all deter-
Exponentiating κ produces a unitary operator which ro- minants is classically intractable, the CI method is typ-
tates the orbital basis (Helgaker et al., 2014). We then ically limited to including a small number of excitations
variationally minimise the energy by optimising both the above the reference state: single excitations (CIS), dou-
amplitudes αf , and the entries kij of κ. ble excitations (CISD), and occasionally triple excita-
MCSCF can be considered the best approximation to tions (CISDT). However, as low energy excitations domi-
the exact wavefunction for a given number of determi- nate the ground state wavefunction in many physical sys-
nants (Wang et al., 2008). It is not possible to per- tems, these truncations can still produce good approxi-
form a complete MCSCF calculation on all possible de- mations to the ground state energy (Helgaker et al., 2014;
terminants for systems with more than a few electrons, Szabo and Ostlund, 2012). The CI method is effective at
as the number of determinants scales exponentially with recovering dynamic correlation, but less effective at re-
the number of electrons. We can attempt to use chem- covering static correlation (Helgaker et al., 2014). If the
ical intuition to select the most important Slater deter- reference state is a MCSCF state, the method is known
minants, and perform an MCSCF calculation on this as multireference configuration interaction, which seeks
restricted number of determinants. Alternatively, we to recover dynamic correlation on top of the static corre-
can use the complete active space self-consistent field lation described by the MCSCF state.
(CASSCF) method (Roos et al., 1980). This consid- The CI method suffers from two major limitations.
ers only the most important orbitals (an active space, The method converges slowly to the full configura-
see Sec. III.E) and performs an MCSCF calculation on tion interaction wavefunction, as a result of its linear
 16

parametrization. In addition, the energy obtained from determinants). Instead, coupled non-linear equations can
a truncated CI calculation does not scale correctly with be derived. The solution to these equations is the CCSD
the system size. A calculation is said to produce ‘size approximation to the ground state (Helgaker et al., 2014;
extensive’ results for an observable if the correct scaling Purvis and Bartlett, 1982). The time taken to solve these
behaviour of that observable is obtained as the system is equations scales as O((M −N )4 N 2 ) (Purvis and Bartlett,
scaled to the thermodynamic limit (Lyakh et al., 2012). 1982), while the memory needed to store the molecular
The energy obtained from a truncated CI calculation is integrals needed scales as O(M 4 ). For more accurate
not size extensive. Truncated CI calculations also fail results, the CCSD(T) method can be used, which treats
to satisfy a related property, known as ‘size consistency’. the triple excitations pertubatively, and scales in time ap-
Size consistency arises from multiplicative separability proximately as O(M 7 ). There has been significant work
of wavefunctions describing two non-interacting systems; to reduce these high computational costs, often introduc-
i.e. |ABi = |Ai |Bi for HAB = HA ⊕ HB . One can easily ing approximations which exploit the locality of dynami-
verify that this does not hold for truncated CI calcula- cal electron correlation in certain systems (Schutz, 2000;
tions. Size consistency is important for obtaining the Schutz and Werner, 2000). This has reduced the scaling
correct behaviour from calculations to determine reac- to be, in some cases, linear (Werner and Schutz, 2011).
tion rates, as reactants at the beginning of the process
are considered non-interacting.
Because of its product parametrization, the CC
method generates a trial wavefunction which includes
4. Coupled cluster all possible determinants, albeit with an incorrect
parametrization. It therefore provides faster conver-
The coupled cluster (CC) method also includes addi- gence than the configuration interaction method. The
tional determinants to recover the correlation energy, but product parametrization also ensures size extensivity
uses a product parametrization. The CC wavefunction is and size consistency. However, the CC method is not
given by without its own shortcomings. Most notably, the wave-
Y  function generated by the canonical CC method does not
|ΨCC i = I + Ciα a†i aα × obey the Rayleigh-Ritz variational principle (Helgaker
i,α et al., 2014). While it is possible to formulate alternative
  (33)
I + Cijαβ a†i a†j aα aβ × ...|ΨHF i.
Y
variants of the coupled cluster method, which are
i>j,α>β variational (Van Voorhis and Head-Gordon, 2000), these
are not as widely used by the computational chemistry
This formula can be recast in an exponential form, writ- community. Moreover, the conventional CC method
ten as described above is a single determinant reference state
method. Consequently, it does not tend to perform well
|ΨCC i = eT |ΨHF i, (34) when applied to multireference states, which are required
P to treat systems with strong static correlation (Lischka
where T = i Ti , et al., 2018; Lyakh et al., 2012). While there have
been efforts to develop multireference coupled cluster
tiα a†i aα ,
X
T1 = approaches, these have their own limitations, and are
i∈virt,α∈occ
not in widespread use, as discussed by Lyakh et al.
tijαβ a†i a†j aα aβ ,
X
T2 = (35) (2012). In Sec. V.B we will describe a modified form
i,j∈virt,α,β∈occ of the CC method, known as unitary coupled cluster
..., (UCC). This method is both variational and suitable for
multireference states. While it is exponentially costly
where occ denotes orbitals that are occupied in the to implement with a classical computer, this method is
Hartree-Fock state, virt denotes orbitals that are unoccu- efficient to implement using a quantum computer.
pied (virtual) in the Hartree-Fock state, and t are excita-
tion amplitudes. When all of the excitation operators Ti
are included, the CC method recovers the full configura-
tion interaction wavefunction – however, performing this This section has treated the inaccuracies which re-
calculation would be exponentially costly. As a result, sult from approximating the full configuration interaction
the method is normally truncated at a lower excitation wavefunction, while including all molecular orbitals. The
level, often single and double excitations (CCSD). The following section will discuss the converse case; we con-
canonical implementation of CCSD does not store the sider only a limited number of molecular orbitals, but
wavefunction, as this would be exponentially costly (as assume that we include all possible determinants that
the CCSD wavefunction has support on all possible Slater they can generate, unless explicitly stated.
 17

D. Chemical basis sets GTOs are less prominent in molecular calculations (Hel-
gaker et al., 2014).
In this section, we describe some of the conventional The most common basis sets construct approximate
orbital basis sets used in classical computational chem- STOs from linear combinations of GTOs. These ap-
istry. Throughout this section, we will refer to the ‘true’ proximate STOs are used as the basis functions for our
orbitals of the system. These can be obtained by numer- atomic orbitals. The number and type of orbitals defines
ically solving the Schrödinger equation using grid based the basis set. There is a compromise between the accu-
methods with a very fine grid spacing, which is only pos- racy obtained and the number of basis functions used.
sible for small atoms or simple molecules. The orbital The number of orbitals considered determines the run-
functions introduced in this section are approximations time and memory requirements of classical chemistry al-
of these true orbitals. gorithms. In the case of quantum computational chem-
Although the Schrödinger equation can be solved ex- istry, the number of basis functions determines the num-
actly for one electron atoms, the orbitals obtained be- ber of qubits and gate operations required to solve the
come diffuse too rapidly to accurately describe many- problem, which we will discuss explicitly in Sec. IV and
electron atoms, especially close to the nuclei (Helgaker Sec. V, respectively.
et al., 2014). A better basis can be obtained by consider-
ing parametrized functions known as Slater-type orbitals
(STO) 1. STO-nG and split-valence basis sets

RnSTO (r) ∝ (ζr)n−1 e−ζr , (36) Some of the most simple bases are the STO-nG ba-
sis sets (Slater Type Orbital-n Gaussians) (Hehre et al.,
where n is the energy level and ζ is a fitting parameter.
1969). In an STO-nG basis, each atomic orbital is consid-
By using different values of ζ for each orbital, we can
ered to be an approximate STO. The STOs are approxi-
generate a good basis (Helgaker et al., 2014). Unlike the
mated using n GTOs. STO-nG basis sets are often called
true atomic orbitals, these functions do not display oscil-
minimal basis sets, as they contain only the orbitals re-
latory behaviour. Consequently, linear combinations of
quired to write the Hartree–Fock (HF) state (and those
STOs are required to approximate the true orbitals. It is
orbitals of similar energy). Calculations using minimal
possible to only introduce a single basis function for each
basis sets are of limited accuracy, giving only a qualita-
considered orbital in the molecule, and give each basis
tive description of the system. It is important to note
function a different ζ value. This is known as a single-
that when carrying out a HF calculation in an STO-nG
zeta representation. Alternatively, we can introduce n
basis, the true HF energy (i.e. the energy obtained by
basis functions (where n is not the energy level of the
performing a HF calculation using a grid based method,
orbital, but a number defining the number of basis func-
on an infinitely precise grid) will not be obtained, as the
tions we wish to include), each with a different ζ value,
STO-nG basis sets only approximate the true HF or-
for each orbital. This is known as an n-zeta representa-
bitals. As an example of an STO-nG basis set we con-
tion. Introducing additional basis functions in this way
sider lithium, which has 3 electrons, of which 2 can reside
increases the radial flexibility of the wavefunction. While
in the 1s orbital, leaving 1 in the second energy level. We
the STO functions exhibit many desirable features, they
include in the minimal basis set {1s, 2s, 2px , 2py , 2pz } or-
make evaluating the two-electron integrals in Eq. (28)
bitals. We include both the 2s and 2p orbitals because
computationally expensive.
they are of the same energy level.
To simplify the two-electron integrals, we can instead
More accurate basis sets can be formed by adding in-
use Gaussian basis functions. The Gaussian basis func-
creased radial flexibility to the valence orbitals (the or-
tions are obtained by considering the Schrödinger equa-
bitals of the highest occupied energy level), by consider-
tion with a three dimensional Harmonic oscillator poten-
ing a double-zeta representation of the valence orbitals.
tial. The form of a Gaussian-type orbital (GTO) is given
This can be achieved using split-valence (Ditchfield et al.,
by
1971) basis sets, such as the 6-31G basis. These basis sets
√ 2
can be further improved by adding additional orbitals
GTO
Rnl (r) ∝ ( αnl r)l e−αnl r , (37)
with higher angular momenta, which make the angular
where αnl is a fitting parameter and l denotes the an- part of the wavefunction more flexible. These orbitals are
gular momentum quantum number of the orbital. Be- called ‘polarisation functions’, as they describe the polar-
cause of the dependence on r2 in the exponent, GTOs isation of atomic charge caused by bonding (for example,
are more localised than STOs. As a result, GTOs do not the 6-31G* basis).
approximate the atomic charge distribution as well as Although the aforementioned basis sets are too small
STOs, so more are required to describe a given orbital. for calculations of such high accuracy that quantum com-
However, this limitation is compensated by the ease of puters are warranted, it is important to discuss them
integral evaluation. Furthermore, the disadvantages of here, as they have been used extensively in the small ex-
 18

perimental demonstrations possible on today’s quantum

hardware.
STO-nG

Orbitals in basis (H2)

1s 6-31G

cc-PVDZ
2. Correlation-consistent basis sets
1s’
Additional accuracy can be obtained by using cc-PVnZ
basis sets (correlation consistent polarised valence n
zeta), introduced by Dunning Jr. (1989). These include
additional unoccupied (‘virtual’) orbitals to recover the
2px 2pz 2py
correlation energy. The virtual orbitals are generated
from correlated calculations on atoms. The core orbitals FIG. 4 The orbitals included in different basis sets for the
have a single-zeta representation, while the valence or- di-hydrogen molecule. The 1s0 orbital is often written as 2s.
bitals have an n-zeta representation. The virtual orbitals The plots show the radial probability distributions for the true
considered are polarisation functions, with higher angu- Hydrogenic orbitals, which the basis orbitals approximate.
lar momenta than the valence orbitals. The polarisation
functions are selected by the size of their contribution to
3. Plane wave basis sets
the correlation energy.
For atomic hydrogen in the cc-PVDZ (D = double, While the aforementioned basis sets have a long history
so n = 2) the highest occupied energy level (the valence of use in classical computational chemistry (and as a re-
level) is the first level, and so we take a double-zeta rep- sult, early work in quantum computational chemistry),
resentation of the 1s state, considering {1s, 1s0 } orbitals. they are not necessarily optimal basis sets for calcula-
The 1s0 orbital is often referred to as a 2s orbital. This tions performed on quantum computers. While these
is because the additional function chosen to describe the basis sets result in an accurate description of the sys-
valence orbital has the same angular momentum as the tem with relatively few basis functions, they also lead to
ordinary 1s orbital, but is more diffuse – so it resembles Hamiltonians containing up to O(M 4 ) terms. As we will
a 2s orbital. We then include polarisation functions, see in Sec. V, the number of terms in the Hamiltonian
which have a higher angular momentum value than the plays a key role in the cost of some quantum chemistry al-
valence functions. In total, there are five basis functions gorithms. It is therefore interesting to question whether
for cc-PVDZ hydrogen: {1s, 1s0 , 2px , 2py , 2pz }. For there are basis sets that are more useful for quantum
lithium in the cc-PVDZ basis, the core orbital is {1s}. computational chemistry. Two examples of such bases
The valence orbitals (which have a double-zeta repre- are the plane wave and plane wave dual basis sets intro-
sentation) are {2s, 2px , 2py , 2pz , 2s0 , 2p0x , 2p0y , 2p0z }, duced for quantum computing by Babbush et al. (2018c).
and the polarisation functions are {3dzz , 3dxz , 3dyz , The plane wave basis functions, φν (r), are given by
3dxy , 3dx2 −y2 }, which we write as {5 × 3d}. This yields r  
14 basis functions. For lithium in the cc-PVTZ basis 1 2πiνr
φν = exp , (38)
(T = triple so n = 3), we first include the 14 orbitals V L
above. As we consider a triple-zeta representation of the
for a plane wave with wavevector corresponding to the
valence orbitals, we need additional {2s00 , 2p00x , 2p00y , 2p00z }
ν th harmonic of the computational cell with length L
orbitals. We then include additional polarisation func-
and volume V . The plane wave dual basis is obtained by
tions; {5×3d0 , 7×4f }. This leads to a total of 30 orbitals.
taking the discrete Fourier transform of the plane wave
basis states, so are like a smooth approximation to a
cc-PVnZ basis sets with higher values of n contain grid (Babbush et al., 2018c). These basis sets diagonalise
orbitals that better approximate the true atomic orbitals the kinetic and potential operators, respectively. This re-
than those with lower n values. However, even large duces the number of Hamiltonian terms from O(M 4 ) to
(n = 5) basis sets struggle to exactly represent the true O(M 3 ) in the plane wave basis, and O(M 2 ) in the plane
HF orbitals of simple molecules such as N2 (Helgaker wave dual basis. This in turn leads to a reduction in the
et al., 2014). This limitation can be overcome by asymptotic scaling of quantum chemistry algorithms to
measuring the ground state energy in several different find the ground state energy of molecules and solid state
cc-PVnZ bases, and then extrapolating to the basis set systems (we will discuss the magnitude of this poten-
limit. tial speedup in Sec. V). These plane wave basis sets are
well suited to periodic systems, and have a long history
 19

of use in classical density functional theory calculations. occupied, respectively. Occupied orbitals are typically
However, to describe molecular systems, approximately referred to as ‘core’ orbitals, while the empty orbitals are
10 to 100 times as many plane wave basis functions are known as ‘virtual’ orbitals. As a result, we can reduce our
required as GTOs (Babbush et al., 2018c). problem by considering only the ambiguously occupied
A similar reduction in the number of Hamiltonian orbitals. In Sec. VII we provide an explicit example of
terms can be obtained using gausslet basis sets (White, how this method can be used to reduce the number of
2017), or ‘discontinuous Galerkin’ sets (McClean et al., orbitals required to simulate lithium hydride in an STO-
2019a), which can both require fewer functions to ac- 3G basis set.
curately describe individual molecules than plane wave Takeshita et al. (2019) showed how to re-integrate
basis sets. Creating efficient basis sets for quantum com- some of the energy contribution of the virtual orbitals
putational chemistry remains an open and fundamental into a quantum simulation, without requiring additional
area of research. qubits to represent the virtual orbitals. Their method
utilises an increased number of measurements, and
classical post-processing. We will describe this method
E. Reduction of orbitals in more detail in Sec. V.C.1. Those authors also
provide a technique to improve the energy estimate
It is sometimes the case that certain orbitals are very from a calculation by optimising the active space using
likely to be either occupied or virtual in all Slater deter- orbital rotations, as is done in MCSCF calculations
minants in the wavefunction. As calculating the ground (Sec. III.C.2).
state energy is essentially a question of distributing elec-
trons among orbitals, we can simplify our calculation by This section has introduced the concepts in classical
using this information. Specifically, we are able to re- computational chemistry necessary to understand how
move spin-orbitals from the calculation if their expected quantum computers can be used for chemical simulation.
occupation number is close to 0 or 1. Our calculation is The following sections introduce methods developed to
reduced to including only the most important (ambigu- solve chemistry problems using quantum computers. We
ously occupied) orbitals. This is known as performing return to classical computational chemistry methods in
the calculation in a reduced active space. Sec. VIII.A, where we assess the strengths, weaknesses,
In order to determine the occupation of orbitals, we and computational limits of the methods introduced here.
can use the reduced density matrices (RDMs) of the sys-
tem. The expectation value of any 1- or P2-electron Hermi-
tian operator, O, with a state |Ψi = f αf |f i, is given IV. QUANTUM COMPUTATIONAL CHEMISTRY
by (Helgaker et al., 2014) MAPPINGS
X X
hΨ| O |Ψi = Oij ρ1ij + Vijkl ρ2ijkl , In this section, we describe the techniques developed
i,j i,j,k,l (39) to enable quantum computers to represent problems in
ρij = hΨ| ai aj |Ψi , ρijkl = hΨ| a†i a†k al aj |Ψi ,
1 † 2 chemistry. In Sec. IV.A and Sec. IV.B we introduce
methods for encoding fermions into qubits (both in first
where ρ1 is the single-particle reduced density matrix (1- and second quantisation). We then describe methods
RDM), ρ2 is the two-particle reduced density matrix (2- which utilise knowledge of the structure of chemistry
RDM), and Oij and Vijkl are defined in a similar way to problems to reduce the resources required, in Sec. IV.C.
the coefficients in Eq. (28). When eliminating orbitals in As discussed in Sec. III.B the distinguishing feature be-
this way, the RDMs are defined with respect to a state tween first and second quantised methods is whether an-
which is an approximation of the ground state, which tisymmetry is enforced in the wavefunction directly (first
could be the results of a classically tractable configu- quantised), or in the behaviour of the operators which act
ration interaction or coupled cluster calculation. These on the wavefunction (second quantised). As in the pre-
RDMs contain all of the information required to evalu- vious section, we consider a system with M spin-orbitals
ate hOiΨ . From the definition above, we can see that the (when discussing basis set approaches) and N electrons.
diagonal elements of ρ1 are the expectation values of the We summarise the number of qubits required to store the
number operator for the corresponding orbitals. As ρ1 is wavefunction in each of the representations in Table I.
a Hermitian operator, we can diagonalise it with a uni- In order to make the mappings more clear, we show
tary transform. This is a basis change from the canonical how the wavefunction would look under each mapping for
orbitals to the ‘natural molecular orbitals’. The diago- a fictitious system. When considering a basis set map-
nal elements of the basis transformed ρ1 are called the ping of this system, we consider spin-orbitals |A↑ i, |A↓ i,
natural orbital occupation numbers (NOONs). |B↑ i, |B↓ i. We are free to arbitrarily define the Hartree-
Spin-orbitals with a NOON close to 0 or 1 (compared Fock state of our fictitious system, and choose it to be
to the other NOONs) can be assumed to be empty or both electrons in the |Ai orbital. We are interested in the
 20

In order to make the first quantised grid based map-

TABLE I The number of qubits required to store the wave-
function using the different encoding methods. M is the
ping more understandable, we consider three example
number of spin-orbitals used in a basis set simulation, N wavefunctions. Without loss of generality, we neglect the
is the number of simulated particles in the problem, and spin coordinate of the electrons, and consider only a sin-
m = log2 (P ), where P is the number of grid points per axis gle spatial dimension for each electron. The first example
for a grid based simulation. The number of qubits can often considers a single spinless electron, on a four point grid.
be further reduced, as discussed in Sec. IV.C.1. The ‘n’-shaped wavefunction considered is given by
Mapping Number of qubits
1 1 1 1
First quantised, basis sets N dlog2 (M )e |ϕi = √ |00i + √ |01i + √ |10i + √ |11i
6 3 3 6
(41)
First quantised, grid based (3m + 1)N 1 1 1 1
= √ |0i + √ |1i + √ |2i + √ |3i
Second quantised, basis sets M 6 3 3 6
and can be stored using two qubits. The second exam-
ple considers an antisymmetrised product state of two
wavefunction when the z component of the spin is zero.
electrons, where one electron is in the state |ϕi, and the
When considering a grid based approach, for the sake of
other has a ‘u’-shaped wavefunction
simplicity we consider spinless electrons on a 1D grid.
The two single particle wavefunctions considered are ap- 1 1 1 1
|φi = √ |0i + √ |1i + √ |2i + √ |3i . (42)
proximately ‘n’ shaped (see Eq. (41)) and ‘u’ shaped (see 3 6 6 3
Eq. (42)).
The pair wavefunction is then given by
1
A. First quantised encoding methods |Φi = √ (|ϕi1 |φi2 − |φi1 |ϕi2 ) ,
2
Here we give an overview of first quantised quantum 1
= √ (|1i1 |0i2 − |0i1 |1i2 + |1i1 |3i2 − |3i1 |1i2
simulation, which can be carried out using either a dis- 6 2
crete single-particle basis, or grid based methods. + |2i1 |0i2 − |0i1 |2i2 + |2i1 |3i2 − |3i1 |2i2 ),
(43)
where the subscripts label the electrons. In general, an
1. Grid based methods
entangled state of two spinless electrons on a P point 1D
grid can be written in this way as
As discussed in Sec. III.B.1, the wavefunction of an
N -particle system can be represented in real space on a P
X −1 P
X −1
discretised grid of P points per axis, and is given by |Ψi = ψij (|ii1 |ji2 − |ji1 |ii2 ) , (44)
X i=0 j=0
|Ψi = ψ(x1 , . . . , xN )A (|x1 , . . . , xN i) , (40)
x1 ,...,xN where again, the subscripts label the electrons.
Grid based methods were first introduced for the quan-
where |xi i = |ri i |σi i is a spatial and spin-coordinate, tum simulation of general quantum systems by Wiesner
|ri i = |xi i |yi i |zi i , ∀i ∈ {1, 2, . . . , N }, xi , yi , zi ∈ (1996) and Zalka (1998). They were then adapted for
{0, 1, . . . , P − 1}, and σ ∈ {0, 1}. We consider the simulating problems in chemistry by Lidar and Wang
case where P = 2m , where m is an arbitrary num- (1999) and Kassal et al. (2008). Physically relevant states
ber which determines the precision of our simulation. can then be prepared using the algorithms outlined by
While it is classically intractable to store the required Ward et al. (2009). Kassal et al. (2008) showed how to
P 3N ×2N = 2(3m+1)N complex amplitudes for large quan- time evolve the wavefunction under the electronic stuc-
tum systems, it is possible using a quantum computer. ture Hamiltonian. As we will discuss in Sec. V.A, time
If we write the basis vector |x = 2m − 1i in binary as evolution is a key subroutine of algorithms that find the
|11....11i, we can see that it only requires m bits. An ground states of chemical systems. Finally, the relevant
m qubit register can be in a superposition of 2m pos- observables can be measured (Kassal et al., 2008; Whit-
sible states. As a result, it only requires (3m + 1)N field, 2015). A thorough investigation of the resources
qubits to store the N electron wavefunction described by required to perform these simulations in a fault-tolerant
Eq. (40). Using a grid based method, rather than basis manner was carried out by Jones et al. (2012). The time
sets, means that the Born-Oppenheimer approximation evolution algorithm of Kassal et al. (2008) was subse-
is not required. As a result, we are able to treat the quently made more efficient by Kivlichan et al. (2017),
electrons and nuclei on an equal footing using grid based who also performed a more thorough analysis of both
methods, which can be important for systems undergoing gate counts and errors. We will discuss the method used
reactions. by Kivlichan et al. (2017) in more detail in Sec. V.A.4.
 21

Although the spatial resolution of grid based methods We can apply the first quantised basis set mapping
increases exponentially with the number of qubits used, to the fictitious system described above. We first la-
it is not possible to use this to exponentially improve the bel each of the orbitals: |A↑ i = |00i = |0i, |A↓ i =
accuracy of the calculation. This is because all known |01i = |1i, |B↑ i = |10i = |2i, |B↓ i = |11i = |3i. The
grid based algorithms have gate counts which scale poly- Hartree-Fock state has both electrons in the |Ai orbitals.
nomially with the inverse grid spacing. As a result, any An incorrectly symmetrised HF state would therefore
attempt to exponentially increase the simulation accu- be |A↑ i1 |A↓ i2 = |0i1 |1i2 , where the subscripts denote
racy by exponentially reducing the grid spacing causes which electron each orbital describes. The correctly an-
the gate count to increase exponentially. Kivlichan et al. tisymmetrised HF wavefunction would be
(2017) also showed that there exist systems where the 1
grid spacing must decrease exponentially with the num- |ΨHF i = √ (|0i1 |1i2 − |1i1 |0i2 ). (45)
2
ber of particles in the system to maintain constant accu-
racy. Consequently, these systems are not efficient to sim- If we now consider excitations above the HF state, then a
ulate using this method. However, those authors noted general wavefunction with sz = 0 that has been correctly
that such pathological cases can also exist for basis set antisymmetrised is given by
methods, but are typically dealt with efficiently using a α
|Ψi = √ (|0i1 |1i2 − |1i1 |0i2 )
clever choice of basis function. 2
The simulation of chemical systems using a grid based β
method can require considerably more qubits than in ba- + √ (|2i1 |3i2 − |3i1 |2i2 )
2
sis set approaches. For example, it would require 96 logi- γ (46)
cal qubits to store the position of a single spinless particle + √ (|0i1 |3i2 − |3i1 |0i2 )
2
to 32 bits of accuracy using a grid based approach. This
δ
can be contrasted with basis set approaches, where in- + √ (|1i1 |2i2 − |2i1 |1i2 ) .
teresting molecules or Fermi-Hubbard models could be 2
simulated with around 100 logical qubits (as we will dis- As we have N = 2 electrons, and M = 4 spin-orbitals,
cuss in Sec. VIII.B). Consequently, grid based approaches we can see that we only require N dlog2 (M )e = 2 ×
are typically considered unsuitable for near-term quan- dlog2 (4)e = 4 qubits to store the wavefunction.
tum computers, which will have relatively few qubits. The Hamiltonian can be obtained by projecting it onto
the single-particle basis functions
N M
X X −1
2. Basis set methods H= hαβ |φβ ii hφα |i
i=1 α,β=0
The original algorithm for simulating quantum systems (47)
N M −1
in the first quantisation using a discrete basis was given 1X X
+ hαβγδ |φα ii |φβ ij hφγ |j hφδ |i ,
by Abrams and Lloyd (1997). If we consider M single- 2
i6=j α,β,γ,δ
particle basis functions (such as the molecular orbitals
or lattice spin sites described in Sec. III.B.2), we can where
!
enumerate these from 0 to M −1. We can store these spin- Z
∇2 X ZI
orbitals using dlog2 (M )e qubits, denoting spin-orbital 0 hαβ = dxφ∗β (x) − − φα (x),
2 |r − RI |
as |0...00i, spin-orbital 1 as |0...01i and so on, such that I
φ∗α (x1 )φ∗β (x2 )φγ (x2 )φδ (x1 )
Z
spin-orbital M − 1 is represented as |1...11i. We then use
hαβγδ = dx1 dx2 .
N registers of these dlog2 (M )e qubits (one register for |r1 − r2 |
each electron) to describe the states of all of the electrons (48)
in the system. As a result, it requires N dlog2 (M )e qubits For example, if we denote terms in the first sum of
to store the wavefunction. Eq. (47) as Hφi α φβ , and consider our model system with
If we consider a product state generated by each elec- spin-orbitals φα = {|A↑ i, |A↓ i, |B↑ i, |B↓ i}, then the term
tron being in a single orbital, we observe that the wave- 1
HA ↑ B↑
(which acts on electron 1) is given by
function does not have the correct antisymmetry. As
such, it must be antisymmetrised. The original approach, hA↑ B↑ |B↑ ie1 hA↑ |e1 =
by Abrams and Lloyd (1997), accomplished this using hA↑ B↑ |10ie1 h00|e1 =
O(N 2 log22 (M )) gates and O(N log2 (M )) ancilla qubits.
hA↑ B↑ (|1iq3 h0|q3 ) ⊗ (|0iq2 h0|q2 ) =
This was improved by Berry et al. (2018), who used a     (49)
circuit with O(N log2c (N )log2 (M )) gates, with a depth of 1 1
hA↑ B↑ (Xq3 − iYq3 ) ⊗ (Iq2 − Zq2 ) =
O(log2c (N )log2 log2 (M )), where c > 1 and depends on the 2 2
choice of sorting network used, and O(N log2 (N )) ancilla hA↑ B↑
qubits. (Xq3 Iq2 − iYq3 Iq2 − Xq3 Zq2 + iYq3 Zq2 ),
4
 22

where ei denotes electron i, qi denotes the ith qubit the parity of the state’. Using the JW encoding, the
(counting from the right), X, Y, Z, I are the Pauli op- second quantised fermionic Hamiltonian is mapped to a
erators introduced in Sec. II.A, and hA↑ B↑ is given by linear combination of products of single-qubit Pauli op-
the first integral in Eq. (48). There are up to O(N 2 M 4 ) erators
possible 2-body terms, each leading to up to O(22log2 (M ) ) X X Y j
Pauli terms – meaning the Hamiltonian can contain up to H= hj Pj = hj σi , (52)
O(N 2 M 6 ) Pauli terms, which are up to 2log2 (M )-local. j j i
Once the Hamiltonian has been obtained, we can use it
to time evolve the wavefunction, which maintains the cor- where hj is a real scalar coefficient, σij represents one of
rect antisymmetry (Abrams and Lloyd, 1997). As men- I, X, Y or Z, i denotes which qubit the operator acts on,
tioned previously, and shown in Sec. V.A, time evolution and j denotes the term in the Hamiltonian. Each term
is a key subroutine of algorithms to find the ground state Pj in the Hamiltonian is typically referred to as a ‘Pauli
of chemical systems. string’, and the number of non-identity single-qubit Pauli
operators in a given string is called its ‘Pauli weight’. All
of the second quantised encoding methods discussed in
B. Second quantised basis set encoding methods this section produce Hamiltonians of this form. An ex-
ample JW mapping is shown in Table. II. In order to
To simulate chemical systems in the second quantised further clarify the second quantised JW encoding, we
representation on a quantum computer, we need to apply it to the fictitious system described earlier. As
map from operators which act on indistinguishable stated previously, we assume that the Hartree-Fock state
fermions to operators acting on distinguishable qubits. for this system has both electrons occupying the |Ai or-
An encoding method is a map from the fermionic Fock bitals. We store the occupations of the spin-orbitals |A↑ i,
space to the Hilbert space of qubits, such that every |A↓ i, |B↑ i, |B↓ i, which we order as |fB↓ , fB↑ , fA↓ , fA↑ i,
fermionic state can be represented by a qubit state. with fi = 0, 1. The Hartree-Fock state is then given by
There are multiple methods of encoding, which we
describe below. In the following section, we only discuss |ΨHF i = |0011i . (53)
second quantised basis set methods, as second quantised
grid based methods have only briefly been discussed in This state corresponds to the antisymmetrised Slater de-
the context of quantum computational chemistry (see terminant shown in Eq. (45). The sz = 0 wavefunction
Babbush et al. (2018c) Appendix A). is then

|Ψi = α |0011i + β |1100i + γ |1001i + δ |0110i . (54)

1. Jordan-Wigner encoding This state can be compared with the first quantised
basis set mapping shown in Eq. (46). Working in the
In the Jordan–Wigner (JW) encoding (Jordan and JW basis, it is easy to see the advantage that quantum
Wigner, 1928), we store the occupation number of a spin- computers have over their classical counterparts for
orbital in the |0i or |1i state of a qubit (unoccupied and chemistry problems. As discussed in Sec. III.B.2, the
occupied, respectively). More formally, full configuration interaction wavefunction contains
a number of determinants which scales exponentially
|fM −1 , fM −2 , . . . , f0 i → |qM −1 , qM −2 , . . . , q0 i , with the number of electrons, as roughly O(M N ). As
(50) such, it requires an amount of memory that scales
qp = fp ∈ {0, 1}.
exponentially with the system size. However, using
The fermionic creation and annhilation operators in- a quantum computer, we can instead store the full
crease or decrease the occupation number of a spin- configuration interaction (FCI) wavefunction using only
orbital by 1, and also introduce a multiplicative phase M qubits (Aspuru-Guzik et al., 2005). A register of M
factor (see Eq. (25)). The qubit mappings of the opera- qubits can be in a superposition of 2M computational
tors preserve these features, and are given by, basis states. In the JW basis, every Slater determinant
required for the FCI wavefunction can be written as one
ap = Qp ⊗ Zp−1 ⊗ · · · ⊗ Z0 , of these computational basis states. As such, quantum
(51)
a†p = Q†p ⊗ Zp−1 ⊗ · · · ⊗ Z0 , computers can efficiently store the FCI wavefunction.
This is also true for the other second quantised encodings.
where Q = |0i h1| = 12 (X + iY ) and Q† = |1i h0| =
1 †
2 (X −iY ). The Q or Q operator changes the occupation The primary advantage of the JW encoding is its sim-
number of the target spin-orbital, while the string of Z
Pp−1 plicity. However, while the occupation of a spin-orbital is
operators recovers the exchange phase factor (−1) i=0 fi . stored locally, the parity is stored non-locally. The string
We refer to the action of the Z operators as ‘computing of Z operators means that a fermionic operator mapped
 23

TABLE II Example mappings of a fermionic Fock state and its fermionic operators onto the corresponding qubit state, and
qubit operators. n̂i is the fermionic number operator.

Fermion Jordan-Wigner Parity Bravyi-Kitaev

a |0001i + b |0010i a |0001i + b |0010i a |1111i + b |1110i a |1011i + b |1010i

+c |0100i + d |1000i +c |0100i + d |1000i +c |1100i + d |1000i +c |1100i + d |1000i
a0 Q0 X3 X2 X1 Q0 X3 X1 Q0
   
a1 Q1 Z0 X3 X2 Q1 |0i h0|0 − Q†1 |1i h1|0 X3 Q1 |0i h0|0 − Q†1 |1i h1|0
 
a2 Q2 Z1 Z0 X3 Q2 |0i h0|1 − Q†2 |1i h1|1 X3 Q2 Z1
 
a3 Q3 Z2 Z1 Z0 Q3 |0i h0|2 − Q†3 |1i h1|2 1
2
Q3 (1 + Z2 Z1 ) − Q†3 (1 − Z2 Z1 )
n̂i = a†i ai |1i h1|i |1i h1|i=0 , 1
2
(1 − Zi Zi−1 )i=1,2,3 |1i h1|i=0,2 , 1
2
(1 − Z1 Z0 )i=1 , 1
2
(1 − Z3 Z2 Z1 )i=3

to qubits generally has a weight of O(M ) Pauli operators, where A is an (2x × 2x ) matrix of zeros, with the bottom
each acting on a different qubit. row filled with ones, and 0 is a (2x × 2x ) matrix of zeros.
An alternative to the JW mapping that has not yet As an example, when M = 4 (x = 1), the matrix βpq is
found particular use within the field, but is worth the  
reader being aware of, is the parity encoding. This ap- 1 0 0 0
proach stores the parity locally, and the occupation num-
 
 
1 1 0 0
ber non-locally. We use the pth qubit to store the parity β4 =  . (58)
of the first p modes,

0 0 1 0
 
|fM −1 , fM −2 , . . . , f0 i → |qM −1 , qM −2 , . . . , q0 i , 1 1 1 1
p
(55)
X 
qp = fi (mod 2). When the number of qubits is not a power of two, the
i=0
BK encoding is carried out by creating the BK matrix
for the next largest power of two, and only using the
The transformed creation and annihilation operators are first M rows. The qubit operators for the BK encoding
described by Seeley et al. (2012). An example of the are considerably more complicated than those in the JW
parity mapping is shown in Table. II. or parity encodings. We refer to Table II for an exam-
ple, and works by Seeley et al. (2012) and Tranter et al.
(2015) for a more detailed discussion. Applying the BK
2. Bravyi–Kitaev encoding
mapping to a fermionic operator results in a qubit op-
erator with a Pauli weight of O(log2 M ). A thorough
The Bravyi–Kitaev (BK) encoding (Bravyi and Ki-
comparison of the BK and JW mappings was performed
taev, 2002) is a midway point between the JW and parity
by Tranter et al. (2018) for 86 molecular systems. They
encoding methods, in that it compromises on the local-
found that the BK transform was at least as efficient, in
ity of occupation number and parity information. The
general, as the JW transform when finding the ground
orbitals store partial sums of occupation numbers. The
states of the molecular systems. In many cases using the
occupation numbers included in each partial sum are de-
BK transform made the calculations considerably more
fined by the BK matrix, βpq .
efficient.
|fM −1 , fM −2 , . . . , f0 i → |qM −1 , qM −2 , . . . , q0 i ,
" p # Another version of the BK encoding also exists in the
X (56) literature. This is referred to as the BK-tree method,
qp = βpq fq (mod 2).
q=0
as it takes its inspiration from a classical data structure
known as a Fenwick tree (Havlicek et al., 2017). We ex-
It is defined recursively (Bravyi and Kitaev, 2002; Seeley plicitly show how to use this mapping with molecules in
et al., 2012) via Sec. VII. As with the standard BK mapping, the BK-tree
encoding balances how it stores occupation and parity
β1 = [1], information. As a result, it too only requires O(log2 M )
 
β2x 0 (57) qubit operations to realise a fermionic operator, in gen-
β2x+1 =  , eral. However, there are subtle differences between the
A β2x two mappings. It has been noted that the BK-tree map-
 24

ping produces qubit operators with a greater weight than (2018a). It also preserves the locality of the underlying
the standard BK mapping (Sung, 2018). This would sug- model, and offers some protection from errors. We will
gest that it is less suitable for near-term quantum com- discuss this error detecting/correcting property in more
putation. However, the BK-tree mapping also possesses detail in Sec. VI.D.
advantages over the standard BK encoding. The BK-tree
mapping is uniquely defined even when the number of or-
bitals, M , is not a power of 2. As a result, when using the C. Resource reduction
BK-tree mapping we are always able to use the qubit re-
duction by symmetry technique, which we will discuss in In this section, we focus on general techniques that
Sec. IV.C. We have observed that it is only possible to use can be used to reduce the resources required for quantum
this technique with the standard BK mapping when the chemistry simulation. In particular, we focus on methods
number of orbitals is a power of two. As a result, it is im- to remove qubits from the simulation using symmetries,
portant to consider the benefits of both mappings, before and low rank decomposition methods for reducing the
choosing which one to use. A further generalisation of cost of quantum circuits.
the BK-tree encoding is to consider ternary trees (Jiang
et al., 2019), which leads to asymptotic reductions in the
Pauli weight of Hamiltonian terms. 1. Hamiltonian reduction

We focus on techniques to reduce the number of qubits

3. Locality preserving mappings required for the second quantised approach, using Z2
symmetries. More general qubit reduction schemes have
Mappings have been developed which endow the also been developed (Bravyi et al., 2017), but these have
qubit operators with the same locality as the underlying yet to be numerically or experimentally investigated.
fermionic Hamiltonian. These mappings typically In the JW, parity and BK encoding methods, the num-
require more qubits than the JW encoding. Verstraete ber of qubits is equal to the number of spin-orbitals
and Cirac (2005) developed a scheme to eliminate considered, M . However, as the Hamiltonian possesses
the strings of Z operators introduced by the JW symmetries, the wavefunction can be stored in a smaller
transform, resulting in qubit operators with the same Hilbert space. Here, we will describe the method by
locality as the fermionic operators. This is achieved Bravyi et al. (2017), which utilises two such symmetries:
by doubling the number of qubits. Similar ideas were conservation of electron number and spin. This method
introduced by Ball (2005) and Farrelly and Short (2014). enables the systematic reduction of two qubits when us-
These ideas were later generalised and expanded upon by ing the parity, BK (with the caveat that the number of
Steudtner and Wehner (2018) and Whitfield et al. (2016). orbitals is a power of two), or BK-tree encoding. For a
system with M spin-orbitals, we can arrange the orbitals
There is also another variant of the BK trans- such that the first M/2 spin-orbitals describe spin up
form, known as the Bravyi-Kitaev superfast transform states, and the last M/2 spin-orbitals describe spin down
(BKSF) (Bravyi and Kitaev, 2002). This mapping first states. For non-relativistic molecules, the total number
represents each spin-orbital by a vertex on a graph, and electrons and the total sz value are conserved. Exam-
each interaction term in the Hamiltonian as an edge on ining the BK matrix presented in Eq. (58), we see that
the graph. Qubits are then associated to the edges. In every element in the final row is one, and the first half
general, a graph will have more edges than vertices, so of the elements in the M/2th row are also one. Conse-
this increases the number of qubits required. However, quently, the final element of the vector encoded by this
the number of gates required to implement a fermionic matrix, qM −1 , is equal to the number of electrons (mod
operator will scale as O(d) where d is the degree of the 2). Similarly, the M/2th element in the encoded vector,
graph. Assuming fairly local interactions for a molecule, q M −1 , is equal to the number of spin up electrons (mod
2

the degree of the graph will be less than the number of 2). As the electron number and total sz value are con-
vertices. As a result, the BKSF transform may require served by the Hamiltonian, these qubits are only acted
fewer gates than the JW mapping. We refer the reader to on by the identity or Pauli Z operators. We can re-
the work of Chien et al. (2019) and Setia and Whitfield place these operators by their corresponding eigenvalues
(2018) for a detailed discussion of the BKSF transform, (+1 for the identity, +1 for ZM −1 if the total number
and comparison to the JW transform. The BKSF trans- of electrons is even, −1 for ZM −1 if the total number of
form has been generalised (Setia et al., 2018) to either: electrons is odd, +1 for Z M −1 if the number of spin up
2
1) Reduce the weight of each of the Pauli operators in the electrons is even, and −1 for Z M −1 if the number of spin
2
Hamiltonian to O(log d), or 2) Provide some protection up electrons is odd). The Hamiltonian then only acts
from errors. A related mapping, known as the majo- on (M − 2) qubits, so two qubits can be removed from
rana loop stabilizer code, was introduced by Jiang et al. the simulation. Exactly the same method can be used
 25

for the parity and BK-tree encodings. We will explicity V. QUANTUM COMPUTATIONAL CHEMISTRY
show how this method can be used to remove two qubits ALGORITHMS
from chemical Hamiltonians in Sec. VII. We remark that
while this transformation leaves the ground state of the In this section, we focus on methods used to solve the
system unchanged, it does alter the excited states that electronic structure problem with a quantum computer.
can be found. In particular, we are restricted to finding We describe the quantum phase estimation algorithm
those states with an electron number and total sz value and related methods in Sec. V.A. We then discuss the
equal to the values determined as described above. These variational quantum eigensolver (VQE) in Sec. V.B. Both
techniques have been extended by Setia et al. (2019) to of these sections are concerned with finding the ground
include molecular point group symmetries. state energies of chemical systems. We conclude this sec-
tion with a discussion of methods that can be used to
find excited states in Sec. V.C.
As mentioned in Sec. I, the techniques developed for
solving the electronic structure problem can often be gen-
2. Low rank decomposition techniques eralised to solve other problems in computational chem-
istry. For example, O’Brien et al. (2019) showed how
As discussed in Sec. III.A, the electronic structure to calculate the energy derivatives of molecular Hamil-
Hamiltonian in the canonical molecular orbital basis con- tonians (which can be used for geometry optimisation
tains O(M 4 ) terms, where M is the number of spin- and transition state discovery) by using algorithms based
orbitals in the molecule. This means that many quantum on phase estimation (Sec. V.A), or based on a linear re-
circuits implementing time evolution under the molecular sponse quantum subspace expansion (Sec. V.C.1). The
Hamiltonian naively scale in a similar way. As we shall same properties can also be calculated using the method
discuss in Sec. V.A and Sec. V.B, such circuits are a key of Mitarai et al. (2019), who leveraged the variational
component of many quantum algorithms. Motta et al. quantum eigensolver (Sec. V.B), with analytic gradient
(2018) utilised a low rank decomposition of the Hamilto- measurements (Sec. V.B.3), and by Parrish et al. (2019a)
nian in a Gaussian orbital basis to reduce the cost of time who used the contraction VQE method (Sec. V.C.3).
evolution. This decomposition is made possible by the It can be argued that the VQE and phase estimation,
structure present in the molecular Hamiltonian, which as presented herein, represent near-term and long-term
arises from the pairwise nature of the electron-electron methods (respectively) for solving chemistry problems
interactions. Those authors leveraged a doubly factorised with a quantum computer. However, in reality, aspects
Cholesky decomposition of the Hamiltonian, with an effi- of each algorithm can be incorporated into the other,
cient quantum circuit implementation. By carefully trun- creating new methods (Wang et al., 2018; Yung et al.,
cating the number of terms in the decomposition, they 2014) which occupy the intermediate region in the
were able to reduce the cost of time evolution, with a quantum computational chemistry timeline. Algorithms
controllable error. Similar decompositions were also in- to find the ground state using methods which differ
troduced for other quantum circuits of interest. We will from both phase estimation and the VQE have also
discuss the impact of these cost reductions in more detail been proposed such as techniques based on time series
in Sec. V.A and Sec. V.B. estimation (O’Brien et al., 2018a; Somma et al., 2002;
This low rank decomposition technique was also used Somma, 2019), or the method by Ge et al. (2017).
by Huggins et al. (2019b), to reduce the number of
circuit repetitions required for variational algorithms to
find the ground state energy of chemical systems. We
A. Quantum phase estimation
discuss this technique in more detail in Sec. V.B.2.
1. Implementation
We note that most of the work to date in quantum
computational chemistry has focused on second quan- Phase estimation (Kitaev, 1995) can be used to find
tised basis set methods. While first quantised basis set the lowest energy eigenstate, |E0 i, and excited states,
simulations require asymptotically fewer qubits than sec- |Ei>0 i, of a physical Hamiltonian (Abrams and Lloyd,
ond quantised basis set simulations, for the smallest sim- 1999). In the case of quantum computational chemistry,
ulable systems (such as small molecules in minimal basis this qubit Hamiltonian can encode a fermionic Hamilto-
sets), second quantised basis set methods require either nian, obtained using the methods discussed in Sec. IV.
fewer qubits and/or shorter gate sequences. This has The canonical phase estimation algorithm is described
caused second quantised basis set methods to become as follows (Nielsen and Chuang, 2002), and shown in
the de facto option for experimental demonstrations of Fig. 5.
quantum computational chemistry algorithms, due to the
limits of current quantum hardware. 1. We initialise the qubit register in state |Ψi, which
 26

|0 Had • ancilla qubits. Phase estimation has been experimentally

demonstrated in a variety of quantum architectures (Du
|0 Had • QFT−1 et al., 2010; Lanyon et al., 2010; Li et al., 2011; O’Malley
et al., 2016; Paesani et al., 2017; Santagati et al., 2018;
|0 Had •
Wang et al., 2015)
|Ψ e −2πiH20
e −2πiH21
e −2πiH22 To realise the standard phase estimation algorithm
given above, we sequentially need to time evolve the main
FIG. 5 The canonical quantum phase estimation circuit with register under the Hamiltonian H for times t0 = 2π, t1 =
three ancilla qubits. When the ancilla qubits are in state |xi, 4π, ..., tω−1 = 2ω π. The total coherent time evolution,
a control rotation e −2πiHx
is applied to the target state |Ψi. T , is then given by approximately T = 2ω+1 π. Using
QFT denotes the quantum Fourier transform (Coppersmith, Eq. (61), for a success probability of p = 0.5, we require
1994; Shor, 1994). By measuring the ancilla qubits in the ω = n + 2 ancilla qubits. The total evolution time can
computational basis, they collapse to an eigenvalue of H and be related to the binary precision PE = 1/2n , to show
the register qubits collapse to an estimate of the correspond-
that T = 8π/PE . Given that our success probability for
ing energy eigenstate.
this estimate is p = 0.5, we expect to have to repeat the
procedure twice to obtain a good estimate of the ground
has non-zero overlap with the true full configura- state. This is equivalent to a total of 16π/PE calls to the
tion interaction (FCI) target eigenstate state of the unitary e−iH (Reiher et al., 2017). In order to account
system. We require an additional register of ω an- for the fact that c0 < 1, we must multiply the number of
cilla qubits. We can expand the state |Ψi in terms repetitions of phase estimation by 1/|c0 |2 , on average, to
of energy eigenstates
P of the Hamiltonian, writing obtain the ground state.
that |Ψi = i ci |Ei i, where ci are complex coeffi- The basic phase estimation algorithm described above
cients. can be improved in many ways. It can be modified to use
only a single ancilla qubit, which is used to measure each
2. We apply a Hadamard gate to each ancilla qubit,
bit in the energy eigenvalue sequentially (Aspuru-Guzik
placing the ancilla register in the superposition
√1
P et al., 2005; Kitaev, 1995). It can also be made more
2ω x |xi, where x are all possible bit-strings that efficient (Kivlichan et al., 2019b; Svore et al., 2013),
can be constructed from ω bits. We then apply the parallelised (Knill et al., 2007; Reiher et al., 2017), or
controlled gates shown in Fig. 5: made more resilient to noise (O’Brien et al., 2018a).
We can further improve upon the asymptotic scaling
1 XX 1 X X −2πiEi x of phase estimation by using classically obtainable
√ |xi ci |Ei i → √ e ci |xi |Ei i .
ω
2 i x ω
2 i x knowledge about the energy gap between the ground
(59) and first excited states (Berry et al., 2018). The ultimate
limit for the number of calls required to e−iH is π/PE
3. We apply the inverse quantum Fourier transform (for a completely general Hamiltonian, H), which is ap-
to the ancilla qubits in order to learn the phase, proximately obtained using Bayesian approaches (Berry
which encodes the information about the energy et al., 2009; Higgins et al., 2007; Paesani et al., 2017;
eigenvalue: Wiebe and Granade, 2016), or entanglement based
1 X X QFT −1 X approaches (Babbush et al., 2018b). For the case of a
√ e−2πiEi x ci |xi |Ei i −−−−−→ ci |bin(Ei )i |Ei i . molecular Hamiltonian Reiher et al. (2017) showed that
2ω i x i a number of calls scaling as π/2PE will suffice.
(60)
4. We measure the ancilla qubits in the Z basis, which The finite precision, PE , obtained in E0 is not the
gives an estimate of the energy eigenvalue as a only source of error in the algorithm. There are also
binary bit-string, bin(Ei ), with probability |ci |2 . errors arising from imperfect implementation of the con-
This procedure collapses the main register into the trolled unitary evolutions applied to the main register,
corresponding energy eigenstate, |Ei i. which we denote as U . This error can arise, for exam-
ple, from a decomposition of e−iH into arbitrary single
The number of ancilla qubits, ω, required for the
and two qubit gates, as occurs during a Trotter decom-
method of phase estimation described above is deter-
position. There are also errors arising from constructing
mined by the desired success probability and precision in
arbitrary gates from a discrete set of gates – such as ap-
the energy estimate. Nielsen and Chuang (2002) showed
proximating single qubit rotations from multiple T and
that to obtain a binary estimate of the energy, precise to
Hadamard gates. These are typically referred to as cir-
n bits, with success probability p, requires
   cuit synthesis errors, CS , and can be quantified using
1 techniques such as the Solovay-Kitaev theorem (Dawson
ω = n + log2 2 + (61)
2p and Nielsen, 2005). For the specific case of a Trotter de-
 27

composition of e−iH , Reiher et al. (2017) showed that a state that is close to the ground state of Hs . The effi-
the error in the energy eigenvalue obtained from phase ciency of adiabatic state preparation depends on the gap
estimation is upper bounded by PE + U + CS . In between the ground state and the first excited state along
general, it is difficult to optimally allocate resources be- the path between H0 and Hs . For chemical systems, ASP
tween these error budgets in order to minimise the total may be achieved by initialising the system in the ground
error (Kivlichan et al., 2019a; Reiher et al., 2017). state of the Hartree-Fock Hamiltonian (H0 ), and interpo-
lating between the initial and final Hamiltonians using an
Regardless of which version of phase estimation is annealing schedule such as H(t) = (1−t/T )H0 +(t/T )Hs ,
used, there are two universal features. Firstly, it is where t is the time and T is the maximum desired simula-
necessary for the register to initially be in a state with tion time (Aspuru-Guzik et al., 2005). Alternative paths
a non-zero overlap with the target eigenstate. Secondly, that may be more efficient for problems of chemical inter-
we must have a way to coherently implement a unitary est have also been investigated (Veis and Pittner, 2014;
operator defined by an efficiently invertible function of Wecker et al., 2015b). The maximum annealing time, T ,
the Hamiltonian. This unitary operator is often (but not is given by
always) chosen to be the time evolution operator, e−iH ,
M4
 
used above. We will discuss techniques to satisfy both of T ≈O , (62)
these requirements in Sec. V.A.2 and Sec. V.A.3, below. mins ∆(t)
where ∆(t) = E1 (t)−E0 (t) and M is the number of spin-
orbitals in the molecule. Reiher et al. (2017) noted that
the true scaling may be closer to O(M 2 /mint ∆(t)). It is
2. State preparation difficult to know the size of the gap along the entire adi-
abatic path a priori, restricts our ability to perform ASP
Initialising the qubit register in a state which has a in the minimum amount of time. One possible method
sufficiently large overlap with the target eigenstate (typ- for reducing the annealing time required is to introduce
ically the ground state), is a non-trivial problem. This is additional ‘driving’ Hamiltonians, as was numerically in-
important, because a randomly chosen state would have vestigated by Matsuura et al. (2018). Although Eq. (62)
an exponentially vanishing probability of collapsing to does not explicitly depend on the initial state used, it
the desired ground state, as the system size increases. is intuitively preferable to start in a state that has good
Even more problematically, McClean et al. (2014) showed overlap with the target ground state. We would expect
that phase estimation can become exponentially costly, the anneal path to be shorter, and we may be more con-
by considering the imperfect preparation of eigenstates of fident that the gap between the ground and excited state
non-interacting subsystems. This highlights the necessity does not shrink exponentially. This view is supported by
of developing state preparation routines which result in the numerical simulations of Veis and Pittner (2014), who
at worst a polynomially decreasing overlap with the FCI found that annealing times for methylene (CH2 ) could be
ground state, as the system size increases. Several tech- reduced by up to four orders of magnitude by using an ini-
niques have been proposed for state preparation. One ap- tial state with larger overlap with the true ground state.
proach is to prepare reference states obtained from classi- We note however that if an initial state with sufficiently
cally tractable calculations, such as: configuration inter- large overlap with the ground state is available, we may
action states (Babbush et al., 2015; Wang et al., 2009), be able to forgo adiabatic state preparation entirely, and
open-shell spin symmetry-adapted states (Sugisaki et al., instead carry out phase estimation directly on that initial
2016, 2018), multireference states (Sugisaki et al., 2019), state. As discussed above, phase estimation only requires
or states produced by adaptive sampling configuration a non-negligible overlap with the target ground state.
interaction methods (Tubman et al., 2018). Alterna- There are a variety of methods that can be used to
tively, we can use: the variational methods discussed in evolve the system under this time-dependent Hamilto-
Sec. V.B (Yung et al., 2014), quantum algorithms for nian, which are discussed below.
imaginary time evolution (Motta et al., 2019), or adia-
batic state preparation (Aspuru-Guzik et al., 2005). We
focus here on adiabatic state preparation, an approach 3. Hamiltonian simulation
inspired by the adiabatic model of quantum computa-
tion (Farhi et al., 2000). As discussed above, both the canonical phase estima-
For any Hamiltonian Hs , we can prepare a state |Ψi tion algorithm and adiabatic state preparation require
that is close to its ground state via adiabatic state prepa- implementation of the time evolution operator, e−iHt ,
ration (Albash and Lidar, 2018). To do so, we first start where H may or may not be time dependent. There are
with a simple Hamiltonian H0 and prepare its ground several ways to do this, each with their own advantages
state. We then time evolve the system under a Hamilto- and disadvantages.
nian that changes slowly from H0 to Hs , thus preparing
 28

a. Product formulae the computation. One approach is the ‘linear combina-

The most simple method for time evolution has al- tions of unitaries’ (LCU) query model, which decomposes
ready been described in Sec. II.B; Trotterization. If a the Hamiltonian or time evolution operator into a linear
time-independent
P Hamiltonian H, can be decomposed as combination of unitary operators, which are then applied
H = i hi , where hi are local Hamiltonians, then a first in superposition using oracle circuits (which must be ex-
order Lie-Trotter-Suzuki approximation (Trotter, 1959) plicitly constructed for a given problem). As a linear
of the time evolution is combination of unitaries is itself not necessarily unitary,
!S these approaches may require additional techniques (such
Y as amplitude amplification (Berry et al., 2014)) to main-
e−iHt = e−ihi t/S + O(t2 /S). (63) tain a high probability of success. Alternatively, we can
i
use oracles to access the non-zero entries of Hamiltoni-
This approach is also referred to as the ‘product formula’ ans which are d-sparse (they have at most d non-zero
method. In practice, to achieve accuracy ε, the number of elements in each row and column, where d is a polyloga-
Trotter steps S = O(t2 /ε) should be large in order to sup- rithmic function of the matrix dimension).
press the errors in the approximation. This is effectively In addition the the product formula approach, the
a stroboscopic evolution under time evolution operators Taylor series and quantum signal processing + qubiti-
corresponding to each of the terms in the Hamiltonian. zation techniques have found the most use to date in
It is also possible to use higher order product formu- quantum computational chemistry algorithms. Both of
lae (Berry et al., 2007; Dür et al., 2008; Suzuki, 1976), the aforementioned query models can be used for the
which scale better with respect to the simulation error Taylor series and quantum signal processing + qubitiza-
than the first order method. Randomisation procedures tion approaches. The Taylor series and quantum signal
(such as randomly ordering the terms in the Trotter se- processing + qubitization algorithms scale exponentially
quence, or stochastically choosing which terms to include better with regards to the accuracy of the simulation
in the Hamiltonian), have been shown to improve the ac- than product formula based methods.
curacy obtained using product formulae (Campbell, 2018;
Childs et al., 2018; Ouyang et al., 2019). The Taylor series method expands the time evolution
Product formulae can also be used to simulate dynam- operator as a truncated Taylor series, where each term in
ics under a time dependent Hamiltonian, H(t). Wiebe the expansion is unitary. We can then use the aforemen-
et al. (2011) showed that the accuracy of such simulations tioned LCU oracles to implement these unitary opera-
depends on the derivatives of the Hamiltonian (although tors in superposition, thus realising time evolution under
this dependence may be alleviated by incorporating ran- the Hamiltonian. Variants of the Taylor series method
domisation procedures (Poulin et al., 2011)). for simulating time dependent Hamiltonians have also
As discussed above, the error in the simulation is been developed (Berry et al., 2019a; Kieferova et al.,
determined by the Trotter formula used, the number 2018). Qubitization provides another way of accessing
of Trotter steps, and the ordering of the terms. It is the Hamiltonian during quantum computation, using a
important to note that the gate counts of all product ‘block encoding’ of the Hamiltonian. Qubitization nat-
formula based methods scale as O(poly(1/)). Endo urally implements a quantum walk operator with eigen-
et al. (2018c) investigated using extrapolation to sup- values e−iarcsin(Ek /α) where Ek is the k th eigenvalue of
press the error arising from using a finite number of the Hamiltonian H and α is a normalisation factor. We
Trotter steps, as is often done in classical computations. can then use quantum signal processing to invert the
arcsin function, recovering time evolution with the de-
sired eigenvalues e−iEk t . We refer the readers to work by
Childs et al. (2017) and the review of Cao et al. (2019) for
b. Advanced Hamiltonian simulation methods a summary of recent progress in Hamiltonian simulation
Alternative methods have been introduced which and a comparison of the different methods.
may realise the time evolution operator more efficiently
than Trotterization, including: quantum walk based
methods (Berry et al., 2015b; Childs and Kothari, 2011), 4. Phase estimation for chemistry simulation
multiproduct formulae (Childs and Wiebe, 2012; Low
et al., 2019b), Taylor series expansions (Berry and Childs, Both product formulae and more advanced methods
2012; Berry et al., 2015a,c) or Chebyshev polynomial ap- of Hamiltonian simulation have been applied to solve
proximations (Subramanian et al., 2018), and qubitiza- problems in quantum computational chemistry. We
tion (Low, 2018; Low and Chuang, 2016) in conjunction discuss the asymptotic results obtained using these
with quantum signal processing (Low and Chuang, 2017; different methods for four classes of problems: molecules
Low et al., 2016). The cost of these methods depends in Gaussian orbital basis sets, systems treated with
on how the Hamiltonian is accessed or ‘queried’ during plane wave basis sets, the Fermi-Hubbard model, and
 29

first quantised grid based simulations. We will discuss to further increase the sparsity of the problem, such
the explicit gate and qubit counts required to implement as using symmetries present in the electronic structure
some of these methods in Sec. VIII.B. Once again, the Hamiltonian, have also been proposed; however these
number of spin-orbitals in the molecule, or spin-sites in have yet to be thoroughly investigated in the context of
a lattice is given by M , and the number of electrons Hamiltonian simulation (Gulania and Whitfield, 2019;
is given by N . We also refer the reader to the tables Whitfield, 2013).
produced by Babbush et al. (2018b,c), which chart de-
velopments in the asymptotic scaling of phase estimation As discussed above, qubitization (Low, 2018; Low and
based approaches to quantum computational chemistry. Chuang, 2016) is a technique originally introduced in
conjunction with quantum signal processing (Low and
Chuang, 2017; Low et al., 2016) to approximate the uni-
tary e−iHt , in order to perform time evolution of a given
a. Gaussian orbital basis sets state. However, it was later noted that for certain func-
As discussed in Sec. III.D, Gaussian orbitals can be tions f (H), qubitization could be used on its own to di-
used to compactly describe single molecules. They result rectly implement the unitary e−if (H)t without approx-
in a second quantised Hamiltonian with O(M 4 ) terms. imation errors (Berry et al., 2018; Poulin et al., 2018).
The overall cost of phase estimation using product for- One can then perform phase estimation on this unitary
mulae depends on the cost of implementing a single Trot- to obtain the energy eigenvalues of the Hamiltonian, pro-
ter step (which can be influenced by the number of terms vided f (H) is efficiently invertible. This technique was
in the Hamiltonian), and the number of Trotter steps re- used by Berry et al. (2019b), in conjunction with a low
quired to achieve the desired accuracy, and the accuracy rank decomposition of the Hamiltonian (Motta et al.,
desired (typically taken as a constant value, such as to 2018) (described in Sec. IV.C.2). This produced an algo-
within chemical accuracy). rithm with a scaling of O(M 1.5 λ), where λ is the 1-norm
Early works on finding the ground state of molecular of the Hamiltonian. While rigorous bounds on λ in a
systems in Gaussian basis sets used first and second Gaussian basis are not known, Berry et al. (2019b) noted
order product formalae (Aspuru-Guzik et al., 2005; that it usually scales as at least O(M 1.5 ).
Babbush et al., 2015; Hastings et al., 2015; Poulin
et al., 2015; Reiher et al., 2017; Wecker et al., 2014;
Whitfield et al., 2011). A series of improvements b. Plane wave basis sets
throughout these papers reduced the scaling of phase As discussed in Sec. III.D.3, plane wave basis sets
estimation for molecules from O(M 11 ) (Wecker et al., are particularly suitable for periodic systems, such as
2014) to (empirically) O(M 5 ) (Babbush et al., 2015). materials. While they can also be used to simulate sin-
A notable feature of these Trotter-based simulations is gle molecules, we require approximately 10 to 100 times
that rigorous bounds on the Trotter error are believed as many plane waves as Gaussian orbitals for accurate
to be loose by several orders of magnitude, which may simulations. The plane wave and plane wave dual basis
impact the scaling of these approaches. We discuss this sets reduce the number of terms in the Hamiltonian to
in more detail in Sec. V.A.5. O(M 3 ) and O(M 2 ), respectively. As a result, using the
plane wave basis sets can reduce the asymptotic scaling
As would be expected, the introduction of more of several approaches to chemistry simulation.
advanced Hamiltonian simulation algorithms led to a The best product based formulae approaches de-
reduction in the asymptotic scaling of chemistry algo- scribed above scale as O(M 1.67 N 1.83 ) in a plane wave
rithms. Using the Taylor series method, algorithms were basis (Babbush et al., 2018c). This can be further
developed which scale as O(M 5 ) (Babbush et al., 2016) improved for systems where it is appropriate to target
and O(N 2 M 3 ) (Babbush et al., 2017) for molecules in an extensive error in the energy (Kivlichan et al., 2019a).
Gaussian basis sets. The first result uses the second This means that the error in the energy scales with the
quantised representation of the Hamiltonian. The size of the system, and therefore with M . As the scaling
second, more efficient result is obtained by constructing of a Trotterized simulation is inversely proportional to
oracle circuits which access the non-zero elements of the the energy error desired, factors of M can be cancelled
Hamiltonian in a basis of Slater determinants (known from both the numerator and denominator of the algo-
as the CI matrix (Toloui and Love, 2013)). The Hamil- rithm scaling. For 3D materials, such as diamond, the
tonian has a sparsity of O(N 2 M 2 ) in this basis. These scaling of Trotter based approaches to phase estimation
algorithms calculated the molecular integrals on the in a plane wave dual basis was reduced to around O(M 2 ).
quantum computer, exploiting an analogy between the
discretisation of space in Riemannian integration and The plane wave basis also benefited the algorithms de-
the discretisation of time in time dependent Hamiltonian scribed above which utilise the Taylor series method for
simulation methods (Babbush et al., 2016). Techniques Hamiltonian simulation. The scaling of the CI matrix
 30

based algorithm of Babbush et al. (2017) was reduced Kivlichan et al. (2019a), who considered the case where
from O(N 2 M 3 ) in a Gaussian basis to O(M 2.67 ) in a an extensive error in the energy is targeted, and found
plane wave basis (Babbush et al., 2018c). that the cost of Fermi-Hubbard model simulations was
A time dependent form of the Taylor series approach effectively between O(1) and O(M 1/2 ) using Trotterisa-
was used for simulation in the ‘interaction picture’, tion.
which enables more efficient time evolution in a plane The qubitization algorithm of Babbush et al. (2018b),
wave dual basis, scaling as O(M 2 ) (Low and Wiebe, discussed above, was also applied to the Fermi-Hubbard
2018) (although this algorithm uses O(M log2 (M )) model. Those authors obtained a gate scaling of O(M )
qubits). A similar interaction picture method was for the Fermi-Hubbard model. When considering an
introduced by Babbush et al. (2018a), using plane waves intensive error in the energy, this approach outperforms
and first quantisation. This algorithm has a gate scaling the Trotter-based method of Kivlichan et al. (2019a).
of O(N 8/3 M 1/3 ) and requires O(N log2 (M )) qubits,
plus ancillas (Babbush et al., 2018a). This was the first New procedures have also been developed specifically
quantum computational chemistry algorithm to achieve for time evolution under lattice Hamiltonians. These
sublinear scaling in M ; a feature that can be used Hamiltonians have geometrically local interactions
to mitigate the inability of plane waves to compactly between qubits that are laid out on a lattice – such as
describe molecular systems. the Fermi-Hubbard model under a locality preserving
mapping (see Sec. IV.B.3). Haah et al. (2018) made use
Qubitization has also been used to produce highly effi- of arguments about the speed of information propaga-
cient algorithms for the plane wave and related basis sets. tion in lattice systems to obtain a simulation algorithm
Babbush et al. (2018b) used qubitization and a form of requiring O(M t) gates to simulate time evolution for
phase estimation which saturates the Heisenberg limit to time t under lattice Hamiltonians. Similar results were
construct algorithms for simulations of matter in a plane obtained by Childs and Su (2019), who proved that a
wave dual basis. Their algorithm has a gate scaling of k th order product formula can simulate time evolution
1
O(M 3 ), for condensed phase electronic structure prob- of an M qubit lattice Hamiltonian using O((M t)1+ k )
lems. elementary operations. These algorithms are almost
Continuing the trend of exploiting different chemical optimal in terms of asymptotic gate complexity.
representations to improve the efficiency of qubitization,
McClean et al. (2019a) investigated the use of discontin-
uous Galerkin basis sets. As discussed in Sec. III.D.3),
these basis sets provide a more compact description of d. Grid based methods
molecular systems than plane waves, but also reduce the As discussed in Sec. III.B.1, grid based methods are
number of terms in the Hamiltonian to O(M 2 ). Mc- particularly suitable for high accuracy calculations – in
Clean et al. (2019a) estimated from numerics on hydro- particular, calculations which do not make the Born-
gen chains that these basis sets could reduce the scaling Oppenheimer approximation, and so treat both the elec-
of qubitization to around O(M 2.6 ). Further work is re- trons and nuclei on an equal footing.
quired to better ascertain and optimise the scaling of this The first quantised grid based algorithm algorithm
technique. of Kassal et al. (2008) proceeds as follows. The qubits
Qubitization can also be used in first quantization with are used to create a discretised grid, as described
the plane wave basis to achieve a sublinear scaling with in Sec. IV.A. Physically relevant states can then be
the number of plane waves. Babbush et al. (2018a) ob- prepared using the algorithms outlined by Kassal et al.
tained a scaling of O(N 4/3 M 2/3 + N 8/3 M 1/3 ), similar (2008) and Ward et al. (2009). The state can be
to the scaling of their interaction picture algorithm dis- propagated in time by repeatedly using the quantum
cussed above. Fourier transform to move between the position and
momentum bases, so that the potential and kinetic
terms are diagonal (respectively) and therefore simple to
c. Fermi-Hubbard model apply. Kassal et al. (2008) did not calculate an explicit
As described in Sec. III.B.2, the Fermi-Hubbard scaling for their algorithm.
Hamiltonian contains O(M ) terms, many of which com-
mute with each other. This can improve the efficiency of Kivlichan et al. (2017) improved upon the method de-
many of the algorithms discussed above. scribed above, using the Taylor series method for time
Wecker et al. (2015b) investigated using adiabatic state evolution. Their method discretised the kinetic and po-
preparation and a Trotter based approach to phase esti- tential terms of the Hamiltonian, separated them into
mation to probe the phase diagram of the Fermi-Hubbard linear combinations of unitary operators, and applied
model. Their approach to finding the ground state re- the Taylor expansion method to simulate time evolution.
quired O(M 3 ) gates. This was subsequently improved by Those authors obtained a scaling of O((N/h2 +N 2 )t) (ex-
 31

cluding logarithmic factors), where h is the grid spacing 𝑈(𝜃⃗ )|000⟩ = |Ψ(𝜃⃗ )⟩
and t is the simulation time.

5. Outstanding problems
|0⟩ 𝜃%& 𝜃,&
M
𝜃0& E
While the advanced Hamiltonian simulation methods |0⟩ 𝜃-& A
described above are asymptotically more efficient than 𝜃/& S
U
Trotterization, the Trotter error bounds appear to be |0⟩ 𝜃.& R
loose by several orders of magnitude (Babbush et al., E
2015; Poulin et al., 2015). A study of spin Hamiltoni-
ans (Childs et al., 2017) found that the asymptotic scal- 𝜃%& → 𝜃%&2% 𝑂(𝜃⃗)
ing of Trotter methods was much worse than the qubiti-

…
zation + quantum signal processing, and Taylor series 𝜃0& → 𝜃0&2%
methods. However, when numerical simulations were
performed, Trotter methods required lower gate counts FIG. 6 A schematic of the variational quantum eigensolver
than the other methods. There are three main factors (VQE). The VQE attempts to find the ground state of a
given problem Hamiltonian, by classically searching for the
that may make the Trotter error bound loose in chemical
optimal parameters θ~ which minimise hΨ(θ)| ~ H |Ψ(θ)i.
~ The
simulations. Firstly, it is difficult to obtain a tight error state preparation and measurement subroutines (red, upper
bound for a Trotterized time evolution under a physi- left, and blue, right) are performed on the quantum computer.
cal Hamiltonian, as analytic proofs typically utilise mul- The measured observable O(θ) ~ and parameter values are fed
tiple triangle inequalities, each of which may be loose. into a classical optimisation routine (green, lower), which out-
Secondly, these error bounds can be understood as the puts new values of the parameters. The new parameters are
worst error that could be induced in any state which the then fed back into the quantum circuit. The gates acting on
unitary acts on. However, for chemistry simulation, we the qubits can be any parametrized gates, e.g. single qubit ro-
tations or controlled rotations. Non-parametrized gates (e.g.
are often only interested in a small subset of the pos-
X, Y, Z, CNOT) are also allowed. The circuit U (θ) ~ and trial
sible states, which again may reduce the Trotter error.
~
wavefunction it produces |Ψ(θ)i are both known as the VQE
Finally, we are often interested not in the error in the
ansatz. The process is repeated until the energy converges.
state, but in the error of some observable. This fact was
exploited by Heyl et al. (2019) who considered the er-
ror in local observables of spin chains, and obtained an alternative methods are required for near-future chem-
improved simulation complexity. istry simulation. We discuss one such approach in the
A challenge facing the Taylor series method is that it following section.
may require many elementary logic operations, resulting
in a large T gate count when considering fault-tolerant
approaches (although work by Sanders et al. (2019) may B. Variational algorithms
help to alleviate this problem). However, while the Tay-
lor series method is asymptotically more expensive than A promising algorithm for near-term quantum hard-
qubitization + quantum signal processing, there has not ware is the variational quantum eigensolver (VQE), first
yet been a variant of the latter technique formulated for proposed and experimentally realised by Peruzzo et al.
time dependent Hamiltonians. Time dependent tech- (2014), and elaborated on by McClean et al. (2016). The
niques are used in the interaction picture method that VQE aims to find the lowest eigenvalue of a given Hamil-
underpins the algorithms of Babbush et al. (2018a) and tonian, such as that of a chemical system. The VQE is a
Low and Wiebe (2018). In addition, it was noted by hybrid quantum-classical algorithm. This means that it
Childs et al. (2017) that implementing quantum signal uses the quantum computer for a state preparation and
processing required intensive classical pre-computation measurement subroutine, and the classical computer to
to obtain the parameters used in the quantum circuit. process the measurement results and update the quan-
As such, it is not yet possible to conclusively state which tum computer according to an update rule. This ex-
method will perform best for chemical systems. changes the long coherence times needed for phase esti-
Despite this progress, all of the methods discussed mation for a polynomial overhead due to measurement
above require circuits with a large number of gates. As repetitions and classical processing. To date, the VQE
a result, these methods are typically assumed to require has only been applied to second quantised basis set simu-
fault-tolerance. As near-term quantum computers will lations, and so our discussion of it will only be concerned
not have enough physical qubits for error correction, the with that scenario.
long gate sequences required by these algorithms make The VQE relies upon the Rayleigh-Ritz variational
them unsuitable for near-term hardware. Consequently, principle. This states that for a parametrized trial wave-
 32

~
function |Ψ(θ)i These state preparation and measurement steps should
be repeated many times in order to measure the expec-
~ H |Ψ(θ)i
hΨ(θ)| ~ ≥ E0 , (64) tation value of every term in the Hamiltonian to the
where E0 is the lowest energy eigenvalue of the Hamilto- required precision. As the quantum computer is reini-
nian H, and θ~ is a vector of independent parameters, tialised for each repetition, the required coherence time
θ~ = (θ1 , ..., θn )T . This implies that we can find the is reduced compared to quantum phase estimation. This
ground state wavefunction and energy by finding the val- is known as the Hamiltonian averaging method of cal-
ues of the parameters which minimise the energy expec- culating the energy (McClean et al., 2014), and requires
tation value. As classical computers are unable to effi- O(1/2 ) measurements to determine the energy to a pre-
ciently prepare, store and measure the wavefunction, we cision . We will discuss the measurement aspect of the
use the quantum computer for this subroutine. We then VQE in more detail in Sec. V.B.2.
use the classical computer to update the parameters us- Once the energy has been measured, it is passed to a
ing an optimisation algorithm. This sequence is shown classical optimisation routine, together with the current
in Fig. 6. The qubit register is initialised in the zero values of θ~k (other observables, such as the energy gradi-
state. We can optionally apply a non-parametrized set ent, could also be supplied to the optimisation routine).
of gates to generate a mean-field (Jiang et al., 2018b; The optimisation routine outputs new values of the cir-
Kivlichan et al., 2018; Wecker et al., 2015b) or multi- cuit parameters, θ~k+1 . These are used to prepare a new
reference state (Babbush et al., 2015; Dallaire-Demers trial state, |Ψ(θ~k+1 )i, which is ideally lower in energy.
et al., 2018; Sugisaki et al., 2019, 2016, 2018; Tubman These steps are repeated until the energy converges to
et al., 2018) describing the chemical system of interest a minimum. We will summarise previous investigations
into classical optimisation routines in Sec. V.B.3.
|Ψref i = Uprep |0̄i. (65) The VQE has been experimentally demonstrated on
most leading quantum architectures including: supercon-
A series of parametrized gates ~
U (θ) = ducting qubits (Colless et al., 2018; Kandala et al., 2017;
UN (θN ) . . . Uk (θk ) . . . U1 (θ1 ) are then applied to the O’Malley et al., 2016), trapped ions (Hempel et al., 2018;
qubits. Here, Uk (θk ) is the k th single or two qubit Kokail et al., 2019; Nam et al., 2019; Shen et al., 2017),
unitary gate, controlled by parameter θk . This circuit and photonic systems (Peruzzo et al., 2014), and shows
generates the trial wavefunction many desirable features. It appears to be robust against
~ = U (θ)|Ψ
~ ref i. (66) some errors (McClean et al., 2016; O’Malley et al., 2016),
|Ψ(θ)i
and capable of finding the ground state energies of small
~ as the ansatz state, and U (θ)
We refer to |Ψ(θ)i ~ as the chemical systems using low depth circuits (Kandala et al.,
ansatz circuit. However, the reader will find that in the 2017). While the VQE appears promising for near-term
literature the word ansatz is typically used interchange- chemistry simulation, several considerable challenges re-
ably to describe both. The set of all possible states that main. Firstly, the VQE is a heuristic method; it is cur-
can be created by the circuit U is known as the ‘ansatz rently unclear whether the quantum circuits proposed
space’. We must select an ansatz appropriate for both will yield better approximations of the ground state than
the available hardware, and chemical system being sim- the classical methods already available. This challenge
ulated. The merits, drawbacks and implementation of is exacerbated by the difficulty of optimising the wave-
common ansätze will be discussed in Sec. V.B.1. function, as optimisation routines could fail to find the
Once the wavefunction has been generated, we need global minimum. Secondly, the number of measurements
to measure the expectation value of the Hamiltonian. required to obtain the energy to the desired accuracy can
Chemical Hamiltonians in the second quantised basis set be large. Finally, the VQE is typically considered in the
approach can be mapped to a linear combination of prod- context of near-term quantum computers, without error
ucts of local Pauli operators, using the transformations correction. While techniques have been proposed to pro-
introduced in Sec. IV.B. We write that tect these calculations from the effects of noise (which we
X X Y j will discuss in Sec. VI), it is still possible that noise may
H= hj Pj = hj σi , (67) prevent us from implementing sufficiently long circuits to
j j i surpass classical methods.

where hj is a real scalar coefficient, σij represents one of

I , X , Y or Z, i denotes which qubit the operator acts 1. Ansätze
on, and j denotes the term in the Hamiltonian. We can
then use the linearity of expectation values to write that The parametrized circuits, or ‘ansätze’, for the VQE lie
N
X Y between two extremes: hardware-efficient and chemically
E(θ~k ) = hj hΨ(θ~k )| σij |Ψ(θ~k )i. (68) inspired. There has been relatively little work compar-
j i ing the effectiveness of different ansätze for anything but
 33
P
the smallest chemistry problems. As quantum comput- where T = Ti , and
i
ers with tens of qubits become more widely available, and
tiα a†i aα ,
X
quantum simulators become more powerful, there will be T1 =
greater scope to test different ansätze for larger problem i∈virt,α∈occ
sizes. tijαβ a†i a†j aα aβ ,
X
T2 = (70)
i,j∈virt,α,β∈occ

...
a. Hardware efficient ansätze and occ are occupied orbitals in the reference state, and
Hardware efficient ansätze have been in use since the virt are orbitals that are initially unoccupied in the ref-
first VQE experiment by Peruzzo et al. (2014). These erence state. The UCC method is intractable on classical
ansätze are comprised of repeated, dense blocks of a lim- computers, but can be efficiently implemented on a quan-
ited selection of parametrized gates that are easy to im- tum computer. It was originally proposed for quantum
plement with the available hardware. They seek to build computational chemistry by Yung et al. (2014) and Pe-
a flexible trial state using as few gates as possible. As ruzzo et al. (2014). A comprehensive review of the UCC
such, they are well suited to the quantum computers cur- method for quantum computational chemistry is given
rently available, which have short coherence times and by Romero et al. (2019).
constrained gate topologies. Hardware-efficient ansätze The UCC method retains all of the advantages of the
have been used to find the ground state energies of sev- CC method, with the added benefits of being fully vari-
eral small molecules (Kandala et al., 2017, 2018). How- ational, and able to converge when used with multi-
ever they are unlikely to be suitable for larger systems, as reference initial states. The UCC ansatz is typically trun-
they take into account no details of the chemical system cated at a given excitation level – usually single and dou-
being simulated. Barkoutsos et al. (2018) attempted to ble excitations (known as UCCSD). We show a canonical
tackle this issue, by proposing hardware efficient ansätze implementation of the UCCSD ansatz in Sec. VII.A. The
which conserve particle number, and so permit the use canonical UCCSD implementation requires O(M 3 N 2 )
of a chemically motivated initial state. This proposal gates when using the Jordan-Wigner mapping (where M
has been experimentally demonstrated (Ganzhorn et al., is the number of spin-orbitals, and N is the number of
2018) on a superconducting system. electrons) (Romero et al., 2019). This has been improved
McClean et al. (2018) showed that using hardware to O(M 3 N ) gates, with a depth of O(M 2 N ) using swap
efficient ansätze with random initial parameters makes networks (O’Gorman et al., 2019), or O(M 3 ) − O(M 4 )
the energy gradient essentially zero among most direc- gates using a low-rank decomposition of the UCCSD ten-
tions of Hilbert space, making classical optimisation sor (Motta et al., 2018).
extremely difficult. This effect becomes exponentially These gate counts assume that a single Trotter step can
more prominent as the number of qubits and circuit be used to implement the UCC operator, which has pre-
depth increases. This suggests that randomly initialised viously been shown to yield accurate results (Barkoutsos
hardware efficient ansätze are not a scalable solution et al., 2018; Romero et al., 2019). However, this approach
for problems in quantum computational chemistry. has been questioned by Grimsley et al. (2019) who found
While techniques have been proposed to combat this that depending on how the operators in the Trotterized
problem (Grant et al., 2019), further work is needed to UCCSD ansatz are ordered, the optimised energies of sys-
determine their efficacy beyond small system sizes. tems displaying significant electron correlation can vary
by amounts larger than chemical accuracy. As a result,
determining a suitable (or even optimal) ordering of the
operators in the UCCSD ansatz may be an interesting
and important area of future research.
b. Chemically inspired ansätze In practice, the gate scaling is typically better than the
Chemically inspired ansätze result from adapting bounds given above, as many excitations are forbidden
classical computational chemistry algorithms to run ef- by the symmetry point groups of molecular orbitals.
ficiently on quantum computers. Most notably, the For example, the LiH molecule in an STO-3G basis
coupled cluster (CC) method discussed in Sec. III.C.4 naively has around 200 excitation operators to consider.
can be extended to produce the unitary coupled cluster However, taking into account symmetries and a reduced
(UCC) ansatz (Bartlett et al., 1989; Hoffmann and Si- active space, one can achieve accurate results while con-
mons, 1988). The UCC method creates a parametrized sidering only around 12 excitation operators (Hempel
trial state by considering excitations above the initial ref- et al., 2018). Moreover, we can use classically tractable
erence state, and can be written as methods to get initial approximations for the remaining
non-zero parameters (O’Malley et al., 2016; Romero
~ = eT −T † ,
U (θ) (69) et al., 2019), which makes the classical optimisation step
 34

of the VQE easier. et al. (2018) improved this asymptotic scaling using the
low rank decomposition method discussed in Sec. IV.C.
Alternative variants of the UCC ansatz have also They reduced the number of gates required to implement
been proposed for solving problems in quantum com- a Trotter step of the Hamiltonian to O(M 2 log2 (M )) with
putational chemistry. These include: the Bogoliubov- increasing molecular size, and O(M 3 ) for fixed molecu-
UCC ansatz (Dallaire-Demers et al., 2018) (a quasipar- lar size and increasing basis size. The electronic struc-
ticle variant of UCC that is suitable for more general ture Hamiltonian in a plane wave dual basis only con-
Hamiltonians than the UCC ansatz, potentially includ- tains O(M 2 ) terms. Kivlichan et al. (2018) showed that
ing the pairing terms present in superconductivity or for Hamiltonians of this form, we can implement Trotter
three body terms present in nuclear physics), the ‘low- steps in depth O(M ), using O(M 2 ) two qubit gates.
depth circuit ansatz’ (Dallaire-Demers et al., 2018) (a The Hamiltonian variational ansatz appears particu-
heuristic which attempts to mimic the aforementioned larly suitable for the Fermi-Hubbard model, which only
Bogoliubov-UCC ansatz using a lower depth circuit), or- has O(M ) terms in its Hamiltonian. Kivlichan et al.
bital optimised-UCCD (Mizukami et al., 2019; Sokolov (2018) showed that it is possible to implement Trotter √
et al., 2019) (which treats single excitations on the clas- steps of the Fermi-Hubbard Hamiltonian with O( M )
sical computer, rather than on the quantum computer, depth and O(M 1.5 ) gates on a linear array of qubits with
to reduce the depth and number of gates in the cir- nearest-neighbour connectivity. Jiang et al. (2018b) im-
cuit), k-UpCCGSD (Lee et al., 2019) and k-uCJ (Mat- proved this result for the case of a 2D array of qubits with
suzawa and Kurashige, 2019) (heuristic ansätze com- nearest-neighbour connectivity. They showed that it is
prised of repeated layers of selected UCC operators), possible to prepare initial states of the Fermi-Hubbard
adaptive-VQE (Grimsley et al., 2018; Tang et al., 2019) model using O(M 1.5 ) gates, and perform Trotter steps of
(which creates an adaptive ansatz designed to maximise the Hamiltonian using O(M ) gates for each Trotter step.
recovery of the correlation energy), and qubit coupled- It has also been noted (Kivlichan et al., 2018) that it is
cluster (Ryabinkin and Genin, 2019; Ryabinkin et al., possible to use locality preserving mappings (Jiang et al.,
2018) (a heuristic method which produced lower gate 2018a; Verstraete and Cirac, 2005) to perform Trotter
counts than the UCC ansatz when applied to several steps in constant depth. However, this comes at a cost
small molecules). of requiring additional qubits, and may have a large con-
stant factor gate overhead.

c. Hamiltonian variational ansatz

There are also variational ansätze that lie between 2. Measurement
the extremes of chemically inspired ansätze and hard-
ware efficient ansätze. One important example is the In Sec. V.B we described the Hamiltonian averaging
Hamiltonian variational ansatz (also commonly referred method, whereby the expectation value of each term
to as a Trotterized adiabatic state preparation ansatz), in the Hamiltonian is estimated through repeated state
proposed by Wecker et al. (2015a). This ansatz was in- preparation and measurement. This procedure can be
spired by adiabatic state preparation and the quantum used to calculate the 1-RDM and 2-RDM of the system.
approximate optimisation algorithm (a similar quantum- McClean et al. (2016) and Romero et al. (2019) showed
classical hybrid algorithm for combinatorial optimisation that the number of measurements Nm , required to esti-
problems (Farhi et al., 2014)). The idea is to Trotterize mate the energy to a precision , is bounded by
an adiabatic path to the ground state, using a number of P
2
Trotter steps that may be insufficient for accurate results. i |hi |
Nm = , (72)
One can then variationally optimise the Trotter evolution 2
times to create a variational ansatz for the ground state. where hi are the coefficients of each Pauli string in the
The number of parameters in this ansatz scales linearly Hamiltonian. This leads to a scaling of O(M 6 /2 ) in a
with the number of Trotter steps, S. Mathematically, we Gaussian orbital basis, and O(M 4 /2 ) for a plane wave
write that dual basis (Babbush et al., 2018c; Cao et al., 2019; Mc-
S Y
Clean et al., 2014).
These scalings may be problematic for large molecular
Y
U= exp (θis Pi ) (71)
s i
calculations. For example, Wecker et al. (2015a) found
that around 1013 samples would be required per energy
where Pi are the Pauli strings in the Hamiltonian. evaluation for a 112 spin-orbital molecule such as Fe2 S2 .
The efficiency of this ansatz is determined by the num- Fortunately, strategies have been proposed to reduce the
ber of terms in the Hamiltonian. This leads to a scaling cost of measurement. Several groups have proposed using
of approximately O(M 4 ) when considering a Gaussian heuristics to group together commuting terms (Izmaylov
molecular orbital basis (Wecker et al., 2015a). Motta et al., 2018; Kandala et al., 2017; McClean et al., 2016;
 35

Verteletskyi et al., 2019), which appears to reduce the optimisation algorithms used in conjunction with the
measurement cost by a constant factor. Some authors VQE. We also discuss methods to assist the classical
have also considered using additional unitary transforms optimisation procedure.
to enable more terms to be grouped, reducing the num-
ber of groups of terms from O(M 4 ) to O(M 3 ) (Crawford
et al., 2019; Gokhale et al., 2019; Gokhale and Chong,
2019; Jena et al., 2019; Yen et al., 2019b; Zhao et al., a. Previous optimisation studies
2019). By grouping terms at a fermionic level, rather Prior experimental VQE implementations have been
than a qubit level, Bonet-Monroig et al. (2019) devised a limited to small systems by the number of qubits avail-
method for measuring the fermionic 2-RDM using O(M 2 ) able. As a result, the parameter space to optimise over
circuits (an additional gate depth of O(M ) is required). is relatively small, so previous results may not be in-
Similar results were obtained by Jiang et al. (2019), by dicative of how these optimisation algorithms will per-
combining Bell basis measurements, and the ternary-tree form for large problems. However, these studies are able
mapping described in Sec. IV.B.2. Another interesting to demonstrate which methods can cope with the high
direction to explore may be to combine locality preserv- noise rates of current hardware. Direct search methods
ing fermion-to-qubit mappings with efficent results for (such as Nelder-Mead simplex (Hempel et al., 2018; Pe-
qubit tomography (Bonet-Monroig et al., 2019; Cotler ruzzo et al., 2014; Shen et al., 2017), simulated anneal-
and Wilczek, 2019; Paini and Kalev, 2019). Although ing (Hempel et al., 2018), particle swarm (Colless et al.,
we might expect that it would be optimal to divide the 2018; Santagati et al., 2018) covariance matrix adapta-
Hamiltonian into the fewest possible number of groups, tion (Sagastizabal et al., 2019), and the dividing rectan-
this is not always the case. McClean et al. (2016) showed gles algorithm (Kokail et al., 2019)) were found to exhibit
that it is important to consider the covariances between good convergence to the ground state energy of small sys-
Hamiltonian terms within groups, as if these covariances tems, despite relatively high noise rates.
are positive, then it may be better to measure the terms The simultaneous perturbation stochastic approxima-
separately. These covariances can be estimated using tion (SPSA) algorithm, has also been used to successfully
classical approximations of the target state. find the ground state energy of small molecules, despite
If we are only interested in measuring the energy ex- relatively large uncertainties due to shot noise, and phys-
pectation value, rather than the 1 and 2-RDMs, Hug- ical error rates (Ganzhorn et al., 2018; Kandala et al.,
gins et al. (2019b) showed that it is possible to use 2017, 2018). The SPSA algorithm is an approximate
the low rank Hamiltonian decomposition technique (dis- gradient based method, which steps ‘downhill’ along
cussed in Sec. IV.C.2) and orbital basis rotations to di- a randomly chosen direction in parameter space. In
vide the Hamiltonian into O(M ) measurement groups. contrast, canonical gradient descent based methods have
When taking into account covariances between commut- struggled to find the ground state, due to the high levels
ing terms, the total number of measurements required for of noise present in current quantum devices (Peruzzo
small molecules was reduced by several orders of magni- et al., 2014).
tude. This measurement scheme also reduces the locality
of strings in the qubit Hamiltonian, thus reducing the The number of numerical studies comparing a subset of
impact of qubit read-out error. optimisation algorithms for the VQE has grown rapidly –
hence, a full discussion of every study is beyond the scope
of this review. Instead, we summarise the major findings,
3. Classical optimisation and highlight the different approaches for optimisation.
In comparisons of ‘out-of-the-box’ optimisers (Mc-
As discussed above, classical optimisation is a crucial Clean et al., 2016; Romero et al., 2019), the Nelder-
aspect of the VQE. However, it is often difficult to min- Mead simplex algorithm utilised in experimental stud-
imise complicated functions in high dimensional parame- ies was outperformed by both other direct search ap-
ter spaces. Classical optimisation routines must be both proaches, and gradient based approaches, such as L-
fast and accurate. They also need to be robust to noise, BFGS-B. All methods were liable to becoming trapped
which will be significant in near-term quantum comput- in local minima, even for small systems. This prob-
ers. lem was partly mitigated by using a good initialisation
Classical optimisation algorithms can be broadly di- strategy, such as using a chemically motivated (Möller-
vided into two classes: direct search and gradient based Plesset) guess for UCC excitation amplitudes (Romero
methods. Direct search algorithms are considered more et al., 2019). Similar results were obtained using heuris-
robust to noise than gradient based methods, but may tic methods of optimisation and initialisation (Wecker
require more function evaluations (Kolda et al., 2006). et al., 2015a). Other investigations have considered op-
Below, we summarise the results of previous ex- timisation using algorithms related to gradient descent,
perimental and numerical investigations into classical such as: stochastic gradient descent (Harrow and Napp,
 36

2019; Sweke et al., 2019), variational imaginary time evo- |0a Had • • Rx ( π2 )
lution (McArdle et al., 2018a), and quantum natural gra-
dient descent (Stokes et al., 2019; Yamamoto, 2019). A |0 X Rx (θ) Y
limitation of the majority of these studies is that realis-
tic noise has rarely been considered, so it is still unclear FIG. 7 A quantum circuit to calculate the gradient of a toy
exactly how effective each method may be in practice. VQE simulation. In this toy problem, the ansatz used is
An alternative approach to optimising all of the pa- |Ψi = Rx (θ) |0i, and the Hamiltonian is H = Y . The energy
rameters simultaneously, is to optimise them sequentially is given by E(θ) = hΨ(θ)| H |Ψ(θ)i = h0| Rx† (θ)Y Rx (θ) |0i.
(Nakanishi et al., 2019; Ostaszewski et al., 2019; Parrish The energy gradient, ∂E ∂θ
= 2i (h0| XRx† (θ)Y Rx (θ) |0i −
†
et al., 2019c). One can analytically minimise the energy h0| Rx (θ)Y Rx (θ)X |0i). This is the expectation value gen-
with respect to a single parameter, and can then sequen- erated by the circuit above.
tially optimise the cost function over all of the parame-
ters. Repeated sequential optimisations are required to
find a good estimate of the energy minimum. solve. When s = 1, the Hamiltonian is equivalent to the
Recent numerical work has explored the possibility of problem Hamiltonian. The annealed variational method
using classical neural networks to optimise VQE circuits works by considering the S layers as separate, distinct
for small instances of the Fermi-Hubbard model (Ver- ansätze. The input state to the first layer of the ansatz
don et al., 2019; Wilson et al., 2019). This approach is the ground state of Hs=0 . We optimise this first layer
(known as ‘meta-learning’) works by training the neural to find the ground state of Hs=1/S . This state is then the
networks on many random instances of the system being input into layer 2, which is optimised to find the ground
studied. The networks were able to successfully ‘learn’ state of Hs=2/S . This process is repeated until the final
to optimise the VQE ansatz, and showed indications layer, which takes the ground state of Hs=(S−1)/S as its
that they may be able to generalise to larger systems input, and targets the ground state of Hs=1 . All of these
than those on which they were trained (Verdon et al., steps are then combined, and used as the starting point
2019). The approach also demonstrated some resilience for a standard VQE optimisation. A similar technique
to noise included in the simulations (Wilson et al., 2019). was proposed by Garcia-Saez and Latorre (2018).
Barkoutsos et al. (2018) introduced a transformation
of the Hamiltonian, such that it only measures the
energy of excitations above the Hartree-Fock state (the
b. Related methods of optimisation correlation energy). Because only the correlation energy
Methods which aid classical optimisation, but that is calculated, fewer measurements are required and
are not optimisation algorithms in their own right, have classical optimisation becomes easier. Overall, simulated
also been proposed. VQE calculations on small molecules were sped up by a
Quantum circuits have been proposed to calculate the factor of 2 to 4 (Barkoutsos et al., 2018).
analytic gradient of the energy with respect to one of the
parameters (Guerreschi and Smelyanskiy, 2017; Mitarai
and Fujii, 2018; Schuld et al., 2019). This avoids the use
of finite difference formulae, which restrict the accuracy C. Evaluation of excited states
of gradient evaluation, as the finite difference considered
is limited by the noise in the energy evaluation. The In this section, we discuss methods used to evaluate
quantum gradient method makes use of the differentia- the excited states of chemical Hamiltonians. These are of
bility of parametrized unitary operators. parametrized interest for calculations of spectral properties, such pho-
unitaries can be written as exponentials of the parame- todissociation rates and absorption bands. We are some-
ter and an anti-Hermitian operator, which are simple to times interested in obtaining excited states with specific
differentiate. A circuit to obtain the gradient of a toy properties, such as a certain spin, or charge (Ryabinkin
VQE simulation is shown in Fig. 7. and Genin, 2018; Ryabinkin et al., 2019; Yen et al.,
Several works have used concepts from adiabatic quan- 2019a). The techniques used to obtain these special
tum computing to aid the classical optimisation proce- states are similar in nature to the methods we discuss
dure. Wecker et al. (2015a) proposed an ‘annealed vari- for finding general excited states. There is still no clear
ational’ method alongside the Hamiltonian variational consensus as to which method may perform best for find-
ansatz. This technique can be generalised to any ansatz ing excited states. As such, we describe each of the lead-
with a repeating, layered structure. We assume that ing methods in turn, and discuss their advantages and
the ansatz is composed of S layers. We first decompose limitations. We note that this area has seen rapid recent
the Hamiltonian of interest into Hs = H0 + sH1 , where development. Consequently, we do not discuss initial pro-
H0 is a Hamiltonian that is efficiently solvable with a posals to calculate excited states, such as the folded spec-
classical computer, and H1 is a difficult Hamiltonian to trum method (McClean et al., 2016) and WAVES (San-
 37

tagati et al., 2018), which are comparatively less efficient As discussed above, the QSE has also been extended
than more recent proposals. to provide better estimates of the ground state energy.
Takeshita et al. (2019) showed how the QSE can be
used to recover the energy contribution of virtual or-
1. Quantum subspace expansion bitals, without requiring additional qubits to represent
the virtual orbitals. This was achieved by consider-
The quantum subspace expansion (QSE) method was ing a QSE with single and double excitations, where
originally introduced to find excited states (McClean the double excitations take two electrons from an active
et al., 2017a), but has proven to be one of the most useful space, to the virtual space. Naively, this would require
techniques introduced to near-term quantum computa- the measurement of matrix elements like Hαβγδ,ζηθ ijkl
=
tional chemistry, with benefits including better ground † †

† †

† †
state estimation (discussed below) and mitigation of hE0 | aα aβ aγ aδ ai aj ak al a aζ aη aθ |E0 i, of which there
hardware errors (discussed in Sec. VI.C). are O(M 12 ). However, because the double excitations
The original formulation of the QSE used a polynomial are restricted to exciting two electrons into the virtual
number of additional measurements to find the excited space, it is possible to use Wick’s theorem, and contract
states of a quantum system (McClean et al., 2017a). The over the operators in the virtual space, to express the ma-
motivation for this was that the higher order reduced trix element above in terms of matrix elements involving
density matrices (RDMs) can be obtained by expanding only 8 creation/annihilation operators (which defines the
the wavefunction in a subspace around the ground state. 4-RDM). These operators only act on the active space or-
These RDMs can then be used to find the linear response bitals, MA , which reduces the number of measurements
excited eigenstates. McClean et al. (2017a) considered required to O(MA8 ), and means that no additional qubits
the single excitation linear response subspace around the are required. Those authors demonstrate this method
fermionic ground state. This subspace is spanned by the by numerically simulating a cc-PVDZ calculation on H2
states a†i aj |E0 i for all possible i, j, which corresponds to (which normally requires 20 qubits), using just 4 qubits,
measuring the 3- and 4-RDMs. This is designed to target and the additional measurements described above.
the low-lying excited states, which are assumed to differ
from the ground state by a small number of excitations.
The excited states can be found by solving a gener- The QSE was also extended by Huggins et al. (2019a),
alised eigenvalue problem in fermionic Fock space who devised a low cost method to extend the subspace
from being excitations above a reference state, to being
H QSE C = S QSE CE, (73) any state that is efficiently preparable on the quantum
computer. The method thus allows for creation of flex-
with a matrix of eigenvectors C, and a diagonal matrix ible ansatz states, without a dramatic increase in the
of eigenvalues E. The Hamiltonian projected into the circuit depth. This approach was further refined by Stair
subspace is given by et al. (2019), who provided a more efficient approach to
QSE
Hij,kl = hE0 | ai a†j Ha†k al |E0 i . (74) realising this technique, when the ansatz circuit used is
composed of products of exponentiated Pauli operators.
The overlap matrix, required because the subspace states
are not orthogonal to each other, is given by
QSE
Sij,kl = hE0 | ai a†j a†k al |E0 i . (75) The main drawback of the QSE is the large number of
measurements that may be required to obtain the 4-RDM
The QSE was experimentally demonstrated in a two of the system. In general there are O(M 8 ) elements to
qubit superconducting system to find the ground and measure in the 4-RDM, compared to O(M 4 ) elements for
excited states of H2 (Colless et al., 2018). Ollitrault the Hamiltonian. The cost can be somewhat reduced by
et al. (2019) showed that the QSE with single and double approximating the 4-RDM using products of lower order
excitations can be understood as an approximation to RDMs, and perturbative corrections. In addition, using
the ‘equation of motion’ (EoM) method for finding the linear response excitation operators described above,
excited states, which was originally introduced as a we are limited in our description of excited states. This
classical method in quantum chemistry by Rowe (1968). can be problematic for systems which need higher order
Ollitrault et al. (2019) showed how the EoM method excitations to accurately describe the excited states (Lee
can be implemented on a quantum computer. They et al., 2019; Watson and Chan, 2012).
found the EoM method to be robust to noise when
demonstrated experimentally, and more accurate for
finding excitation energies than the QSE, in numerical We provide more information on the QSE method in
tests. Sec. VI, where we discuss how it can also be used to
mitigate the effects of errors.
 38

2. Overlap-based methods able to be initialised in any of these states. The orthog-

onality of these states is invariant under the application
Overlap based methods exploit the orthogonality of of a unitary ansatz circuit. The aim of the ansatz circuit
energy eigenstates. Once an eigenstate is found, we can is to generate linear combinations of these initial states,
find other eigenstates by ensuring that they are orthog- to form good approximations of the low energy subspace
onal to the original state (Endo et al., 2018a; Higgott of the system. This is achieved by optimising the ansatz
et al., 2018). After finding the ground state |E0 i of a over an ensemble cost function,
Hamiltonian H with a VQE calculation, we replace the
k
Hamiltonian with ~ =
X
~ ~ |φj i .
C(θ) hφj | U † (θ)HU (θ) (77)
0 j=0
H = H + α |E0 i hE0 | , (76)
Both of the contraction VQE methods then use fur-
where α is chosen to be sufficiently large compared to
ther processing to find the ordering of the energy eigen-
the energy scale of the system. The ground state of
states. The SSVQE method uses a similar quantum-
the updated Hamiltonian H 0 is no longer |E0 i, but
classical hybrid approach to the ordinary VQE. In con-
the first excited state |E1 i of the original Hamilto-
trast, the MCVQE method uses classical diagonalisation
nian H. This process can be repeated to obtain
of a Hamiltonian matrix obtained in the low energy sub-
higher energy eigenstates. The energy of the up-
~ H 0 |Ψ(θ)i
~ = hΨ(θ)|~ H |Ψ(θ)i
~ + space (similar to the way in which the quantum subspace
dated Hamiltonian, hΨ(θ)|
expansion works). These contraction VQE methods dif-
~ ~
α hΨ(θ)|E0 i hE0 |Ψ(θ)i can be obtained by measuring fer from the overlap based methods, in that they obtain
each term separately. We can measure the first term all of the eigenstates at the same time, rather than se-
using the Hamiltonian averaging procedure described in quentially. Consequently, all eigenstates should be ob-
Sec. V.B. The second term can be obtained from circuits tained to a similar accuracy. However, it may be diffi-
which calculate the overlap between the states, such as cult to find a unitary circuit that simultaneously prepares
the SWAP test. The SWAP test approach requires a many eigenstates on an equal footing, which may make
circuit that has twice as many gates as the ansatz used these contraction methods difficult to realise in practice.
(but is of the same depth), and has twice as many qubits.
We can also use a method which requires twice as many
gates and is twice as deep as the original ansatz, but VI. ERROR MITIGATION FOR CHEMISTRY
does not require additional qubits (Higgott et al., 2018).
The additional resources required are the main limitation As discussed in Sec. II.A, physical qubits accumulate
of this approach to calculating excited states. Overlap- errors during computation, due to their interaction with
based techniques were numerically investigated by Lee the environment, and our imperfect control. While the
et al. (2019), who also considered the propagation of er- effects of these errors can be suppressed using quantum
rors resulting from only obtaining an approximation of error correction, this requires a considerable increase in
lower lying eigenstates, rather than the true eigenstates. the number of qubits used for the computation. If these
errors are not dealt with, they will corrupt the results of
our algorithms, rendering the calculations meaningless.
This was confirmed for the case of chemistry calculations
3. Contraction VQE methods by Sawaya et al. (2016), who used numerical simulations
to show the impact that noise has on phase estimation
The Subspace-search variational quantum eigensolver based chemistry calculations. Phase estimation based
(SSVQE) (Nakanishi et al., 2018) and the multistate con- approaches (discussed in Sec. V.A) require long circuit
tracted variational quantum eigensolver (MCVQE) (Par- depths, and so implicitly assume the use of quantum er-
rish et al., 2019b) methods are related to the overlap ror correction. We will discuss the resources required to
based methods described above, in that they are driven carry out these calculations in Sec. VIII.B – however, it
by the orthogonality of energy eigenstates. The differ- suffices to say here that they are considerably greater
ences lie in how they enforce this orthogonality, as well than the resources we have available at time of writing.
as how the ordering of the eigenstates is determined. In We claimed in Sec. V.B that the reduced coherence time
the overlap based methods, orthogonality is enforced by requirements of the VQE make it more suitable for noisy
including the overlap between states in the cost function. quantum hardware, and may enable the extraction of
In contrast, the contraction VQE methods exploit con- accurate results without the use of quantum error cor-
servation of orthogonality between states under a unitary rection. It has been shown both theoretically (McClean
transform. To be more specific, a set of orthonormal ap- et al., 2016) and experimentally (O’Malley et al., 2016)
k
proximate eigenstates {|φi ii=0 } are chosen using an effi- that the VQE can be inherently robust to some coher-
cient classical approach. The quantum computer must be ent errors, such as qubit over-rotation. However, small
 39

experimental demonstrations of the VQE have shown

that noise can still prevent us from reaching the de-
sired levels of accuracy (Hempel et al., 2018; Kandala 𝑂"#$%&
et al., 2017). Consequently, it appears that additional
techniques are required, if we are to perform classically 𝑂" (𝜀# )
intractable chemistry calculations, without error correc-
tion. 𝑂" (𝜆𝜀# )
The methods described in this section have been
developed to mitigate errors, rather than correct them.
These techniques are only effective when used with
low depth circuits, such that the total error rate in the 𝜖# 𝜆𝜖#
circuit is low. However, the additional resources required Error rate 𝜖
are much more modest than for full error correction. In
general, these techniques only require a multiplicative FIG. 8 A comparison of linear (blue, lower) and exponential
overhead in the number of measurements required, if the (grey, upper) extrapolation. The horizontal axis is the error
error rate is sufficiently low. Many of these techniques rate of each gate and the vertical axis is the expectation value
were introduced for use in general near-term algorithms, of the measured observable, Ō. This figure was reproduced
and so can be applied to problems beyond chemistry from Endo et al. (2017), with permission.
simulation.
the dominant error rate, 0 , by a factor λ, and inferring
As we are dealing with errors, it becomes necessary to the error free result by extrapolation. We can increase
consider mixed states, rather than just pure states. As the error rate using the techniques described by Kandala
such, we now switch to the density matrix formalism of et al. (2018) and Li and Benjamin (2017). The technique
quantum mechanics. is based on Richardson extrapolation (Richardson et al.,
We consider a quantum circuit that consists of G uni- 1927), which to first order corresponds to linear extrap-
tary gates applied to an initial reference state |0̄i. The olation using two points. We could also take a linear or
output state if errors do not occur is given by higher order fit with several data points. For the former
ρ0 = UG ◦ · · · ◦ U2 ◦ U1 (|0̄i h0̄|), (78) case, the estimated value of the observable is given by

where for a density matrix ρ, U(ρ) = U ρU † . We extract λŌ(0 ) − Ō(λ0 )

Ō0est = . (81)
information from the circuit by measuring a Hermitian λ−1
observable, O While this method can improve the accuracy of calcu-
Ō0 = Tr[ρ0 O]. (79) lations, it requires additional measurements in order to
keep the variance of the measured observable the same
If each gate is affected by a noise channel Ni , the pre- as the non-extrapolated case. The extrapolation method
pared state becomes has been demonstrated for VQE experiments in both
Y molecular chemistry (Kandala et al., 2018; Ollitrault
ρ= Ni ◦ Ui (|0̄i h0̄|), (80) et al., 2019), and nuclear physics (Shehab et al., 2019).
i
Exponential extrapolation was introduced by Endo
and the measurement result becomes Ō = Tr[ρO]. In et al. (2017) as a more appropriate extrapolation tech-
general, we cannot recover the noiseless state ρ0 from nique for large quantum circuits. A comparison between
the noisy state ρ without error correction. However, the the two extrapolation methods is shown in Fig. 8. Otten
error mitigation methods discussed below can approxi- and Gray (2018b) have also extended the extrapolation
mate the noiseless measurement result Ō0 from the noisy method to the scenario where the error rates of different
measurement result, Ō when the error rate is sufficiently gates are increased by different factors.
low. It is important to note that error mitigation schemes
are not a scalable solution to the problem of noise in
quantum hardware. In order to scale up computations B. Probabilistic error cancellation
to arbitrarily large sizes, fault-tolerant, error corrected
quantum computers are required. The probabilistic error cancellation method introduced
by Temme et al. (2017) works by effectively realis-
ing the inverse of an error channel, N −1 , such that
A. Extrapolation N −1 (N (ρ0 )) = ρ0 . Because realising the inverse channel
is in general an unphysical process, we use the scheme de-
The extrapolation method (Li and Benjamin, 2017; picted in Fig. 9 to effectively realise the inverse channel
Temme et al., 2017) works by intentionally increasing by focusing only on measurement results.
 40

many gates in the circuit. Instead, we can consider

Overhead 𝛾
only the most important terms, which result from a
Operations Probability Parity small number of errors occurring. If the error rate is
𝐼 𝑝( +1 low, then the other terms can be considered negligibly
𝑋 small. After each single qubit gate, we can apply X,
𝑌 𝑝) −1 Y or Z operators with probability p2 , or the identity
𝑍
gate with p1 . We repeat that circuit variant many
times to extract the expectation value, and multiply the
𝜌/ 𝒩 𝒩 -( expectation value by (−1)Gp , where Gp is the number
of additional X, Y or Z gates that were applied in that
Noise Recovery Measurement circuit iteration. We then sum up the values for several
(Unphysical) circuit variants and multiply by γ to obtain the error
mitigated result. This method can also be extended
FIG. 9 A schematic of the probabilistic error cancellation to multi-qubit gates. For example, for two qubit gates
method for a depolarising error resulting from a single qubit in the depolarising noise model, after each two qubit
gate. After the gate is applied, there is a noise channel N .
gate we insert one of the pairs: XI, IX, Y I, IY, ZI, IZ
The method works by effectively realising the inverse channel
N −1 . This is achieved by randomly applying one of the X, (parity −1) with probability p2 , one of the pairs
Y or Z operators with probability p2 , or the identity gate XX, Y Y, ZZ, XY, Y X, XZ, ZX, Y Z, ZY (parity +1)
with p1 . The expectation values resulting from the circuits with probability p2 and II (parity +1) with probability
are combined. If one of the Pauli matrices was applied to p1 .
realise the inverse channel, the resulting expectation value is
subtracted, rather than added (parity −1). The overhead γ The probabilistic error cancellation method described
determines the number of additional measurements required
to keep the variance of the error mitigated result equal to
above has been shown to work for general Markovian
the variance of the noisy result. This can be generalised to noise (Endo et al., 2017), and has also been extended to
multi-qubit gates as described in the main text. work with temporally correlated errors and low frequency
noise (Huo and Li, 2018). Probabilistic error cancella-
tion requires full knowledge of the noise model associated
As an example, we consider the case of a depolarising with each gate. This can be obtained from either process
error channel, tomography, or a combination of process and gate set
  tomography. The latter approach reduces the effect of
3 p errors due to state preparation and measurement (Endo
ρ = D(ρ0 ) = 1 − p ρ0 + (Xρ0 X + Y ρ0 Y + Zρ0 Z).
4 4 et al., 2017). The probabilistic error cancellation method
(82) has been experimentally demonstrated on both supercon-
The unphysical inverse channel is ducting (Song et al., 2018) and trapped ion (Zhang et al.,
2019) systems.
ρ0 = D−1 (ρ) = γ[p1 ρ − p2 (XρX + Y ρY + ZρZ)], (83)

where the coefficient γ = (p + 2)/(2 − 2p) ≥ 1, p1 =

C. Quantum subspace expansion
(4 − p)/(2p + 4), and p2 = p/(2p + 4) in this case.
We cannot directly realise D−1 due to the minus sign
The quantum subspace expansion (QSE) (McClean
before p2 . However, we can consider and correct its effect
et al., 2017a) described in Sec. V.C.1 can mitigate errors
on the expectation value Ō0
in the VQE, in addition to calculating the excited energy
Ō0 = Tr[OD−1 (ρ)], eigenstates. This method is most effective at correcting
systematic errors, but can also suppress some stochastic
= γ[p1 hOiρ − p2 (hOiXρX + hOiY ρY + hOiZρZ )],
errors. Suppose that we use the VQE to find an approx-
= γ[p1 hOiρ − p2 (hXOXiρ + hY OY iρ + hZOZiρ )], imate ground state |Ẽ0 i. Noise may cause this state to
(84) deviate from the true ground state |E0 i. For example,
where hOiρ = Tr[Oρ]. We can therefore measure O, if |Ẽ0 i = X1 |E0 i, we can simply apply an X1 gate to
XOX, Y OY , ZOZ, and linearly combine the measure- recover the correct ground state.
ment results to effectively realise the inverse channel, However, as we do not know which errors have oc-
thus obtaining the noiseless measurement result Ō0 . The curred, we can instead consider an expansion in the sub-
variance in our estimate of Ō0 is increased by a factor space {|Pi Ẽ0 i}, where Pi are matrices belonging to the
of γ G , where γ is the overhead coefficient, and G is the Pauli group. Then, one can measure the matrix repre-
number of gates in the circuit. sentation of the Hamiltonian in the subspace,
In practice, it is not possible to exactly measure all
QSE
of the possible terms resulting from errors if there are Hij = hẼ0 | Pi HPj |Ẽ0 i . (85)
 41

As the subspace states are not orthogonal to each other, for the procedure, the authors introduced a stochastic
we should also measure the overlap matrix sampling scheme for the code stabilisers.
QSE The choice of fermion-to-qubit mapping can also intro-
Sij = hẼ0 | Pi Pj |Ẽ0 i . (86)
duce additional symmetries. For example, both the gen-
By solving the generalised eigenvalue problem eralised Bravyi-Kitaev superfast transform (Setia et al.,
2018), and the majorana loop stabiliser code (Jiang et al.,
H QSE C = S QSE CE, (87) 2018a) (discussed in Sec. IV.B.3) introduce additional
qubits, whose values are constrained by the mappings.
with a matrix of eigenvectors C and diagonal matrix of
By performing suitable stabiliser checks on these qubits,
eigenvalues E, we can get the error mitigated spectrum
single qubit errors can be detected and corrected.
of the Hamiltonian. A small number of Pauli group op-
erators are typically considered, in order to minimise the
required number of measurements. The QSE has been
E. Other methods of error mitigation
experimentally demonstrated, using a two qubit super-
conducting system to measure the ground and excited
state energies of H2 (Colless et al., 2018). Other methods of error mitigation have been proposed,
but require further research in order to assess how they
can be best incorporated into chemistry calculations.
D. Symmetry based methods One such method is the quantum variational error cor-
rector (Johnson et al., 2017), which uses a variational al-
It is also possible to mitigate some errors by us- gorithm to construct noise-tailored quantum memories.
ing symmetry checks on a suitably constructed ansatz Another example is individual error reduction (Otten and
state (Bonet-Monroig et al., 2018; McArdle et al., 2018c). Gray, 2018a). This method uses error correction to pro-
A key concern for the VQE is preserving particle num- tect a single qubit, while leaving the rest of the physical
ber, as states with electron number far from the true qubits subject to noise. The process is repeated sev-
value appear to have a larger energy variance than those eral times, with each physical qubit being protected in
with smaller particle number errors (Sawaya et al., 2016). turn. Post-processing the results produces a more accu-
Consequently, we can perform ‘checks’ on quantities rate expectation value than would be obtained without
which should be conserved (such as number of electrons, the mitigation technique.
or sz value), discarding the results if the measured value
is not as expected.
This can be achieved in a number of ways: by using VII. ILLUSTRATIVE EXAMPLES
stabiliser checks with additional ancilla qubits (Bonet-
Monroig et al., 2018; McArdle et al., 2018c), by In this section we illustrate many of the techniques de-
taking additional measurements and performing post- scribed in the previous sections of this review, by explic-
processing (Bonet-Monroig et al., 2018) (this has itly demonstrating how to map electronic structure prob-
been experimentally demonstrated in superconducting lems onto a quantum computer. We do this in second
qubits (Sagastizabal et al., 2019)), by enforcing physi- quantisation for the Hydrogen molecule (H2 ) in the STO-
cally derived constraints on the form of the measured 3G, 6-31G and cc-PVDZ bases, and Lithium Hydride
1 and 2-RDMs (known as the n-representability con- (LiH) in the STO-3G basis (as described in Sec. III.D).
straints) (Rubin et al., 2018), or by using the low-rank + Across these examples, we showcase the Jordan–Wigner
orbital rotations measurement technique of Huggins et al. (JW), Bravyi–Kitaev (BK) and BK tree mappings (as de-
(2019b) (discussed in Sec. V.B.2). The latter method ap- scribed in Sec. IV.B), reduction of active orbitals using
pears to be the most powerful of these, requiring a mod- the Natural Molecular Orbital (NMO) basis (as described
est additional circuit depth, limited connectivity, and en- in Sec. III.E), reduction of qubits using symmetry conser-
abling effective post-selection on the electron number and vation (as described in Sec. IV.C.1) and the unitary cou-
sz value, rather than just the parities of these quantities. pled cluster ansatz (as described in Sec. V.B.1). These
These methods of error mitigation can be combined with examples are designed to familiarise the reader with the
some of the other techniques discussed above, such as key steps of formulating a quantum computational chem-
extrapolation. istry problem. Many of these techniques are applicable
A related extension of the quantum subspace expan- to both ground state and general chemical problems.
sion was developed by McClean et al. (2019b), who ef-
fectively engineered additional symmetries in the system
using an error correcting code. By post-processing mea- A. Hydrogen
surements from multiple iterations, they could detect
some errors, or effectively realise a limited form of er- The continuous space molecular Hamiltonian for H2 is
ror correction. In order to maintain a polynomial cost given by Eq. (17) (which we reproduce here), with two
 42

by a linear combination of three Gaussian type orbitals.

0
STO-3G Each hydrogen atom contributes one orbital, and there
6-31G are two possible spins for each orbital – resulting in a
-0.2 cc-pVDZ
total of 4 spin-orbitals for STO-3G H2 . We denote these
-1.11
orbitals as
-1.12
Energy (Hartree)

-0.4
Energy (Hartree)

-1.13 |1sA↑ i, |1sA↓ i, |1sB↑ i, |1sB↓ i, (91)

-1.14
-0.6 where the subscript A or B denotes which of the two
-1.15

-1.16 atoms the spin-orbital is centred on, and the ↑ / ↓ de-

-0.8 -1.17 notes the sz value of the electron in the spin-orbital. For
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
convenience, we work in the molecular orbital basis for
-1 H2 , which is simple to construct manually. These single
electron molecular spin-orbitals are given by
-1.2 1
0.5 1 1.5 2 2.5 |σg↑ i = p (|1sA↑ i + |1sB↑ i),
Ng
FIG. 10 Comparing the ground state dissociation curves of
1
|σg↓ i = p (|1sA↓ i + |1sB↓ i),
H2 for a range of basis sets. Ng
(92)
1
|σu↑ i = √ (|1sA↑ i − |1sB↑ i),
electrons and two nuclei Nu
X ∇2 X 1
i ZI 1X 1 |σu↓ i = √ (|1sA↓ i − |1sB↓ i),
HH2 = − − + . (88) Nu
i
2 |r i − R I | 2 |ri − rj |
i,I i6=j
where Ng/u are normalisation factors that depend on
To convert this Hamiltonian into the second quantised the overlap between the atomic orbitals, Ng = 2(1 +
representation, h1sA |1sB i), Nu = 2(1 − h1sA |1sB i). We can write a
X 1 X Slater determinant in the occupation number basis as
H= hpq a†p aq + hpqrs a†p a†q ar as , (89)
2
p,q p,q,r,s |ψi = |fσu↓ , fσu↑ , fσg↓ , fσg↑ i , (93)
with
! where fi = 1 if spin-orbital i is occupied, and fi = 0
∇2 X
Z
ZI if spin-orbital i is unoccupied. We can now calculate
hpq = dxφ∗p (x) − − φq (x),
2 |r − RI | the integrals given in Eq. (90) using these molecular or-
I
bitals. These integrals have been calculated for a large
φ∗p (x1 )φ∗q (x2 )φr (x2 )φs (x1 )
Z
hpqrs = dx1 dx2 , number of basis sets, and the results can be obtained by
|x1 − x2 | using a computational chemistry package (Frisch et al.,
(90) 2016; Muller, 2004; Parrish et al., 2017; Sun et al., 2017).
we need to select a basis set, φp (x). As discussed in We must then map the problem Hamiltonian from being
Sec. III.D, this is a discrete set of functions which are written in terms of creation and annihilation operators,
used to approximate the spin-orbitals of the molecule. to being written in terms of qubit operators. Using the
By considering a larger number of orbitals (and the Slater JW encoding, we can obtain the 4 qubit Hamiltonian for
determinants that they can generate), we are able to re- H2
cover a larger proportion of the correlation energy in a
molecule, resulting in a more accurate estimate of the H = h0 I + h1 Z0 + h2 Z1 + h3 Z2 + h4 Z3
true ground state energy. Fig. 10 shows the H2 ground + h5 Z0 Z1 + h6 Z0 Z2 + h7 Z1 Z2 + h8 Z0 Z3 + h9 Z1 Z3
state dissociation curves in the STO-3G, 6-31G and cc- + h10 Z2 Z3 + h11 Y0 Y1 X2 X3 + h12 X0 Y1 Y2 X3
PVDZ bases. We can see that the differences in energy
between the three minima are considerably larger than + h13 Y0 X1 X2 Y3 + h14 X0 X1 Y2 Y3 .
chemical accuracy (1.6 mHartree). This highlights that (94)
working in a suitably large basis set is crucial for obtain- While it is important to understand this procedure,
ing accurate results. every step from selecting a basis to producing an
encoded qubit Hamiltonian can be carried out using
a quantum computational chemistry package such
1. STO-3G basis as OpenFermion (McClean et al., 2017b), Qiskit
Aqua (IBM, 2018), or QDK-NWChem (Low et al.,
The STO-3G basis for H2 includes only the {1s} orbital 2019a).
for each hydrogen atom. The 1s orbital is represented
 43

In the JW encoding, it is simple to construct the value of the molecule when acting upon the HF state are:
Hartree-Fock (HF) state for the H2 molecule. The HF a†2 a0 , a†3 a1 , a†3 a†2 a1 a0 . Other valid operators are equiva-
state for H2 in the occupation number basis is given by lent to these operators, and can be combined with them,
such as a†3 a†0 a1 a0 = −a†3 a1 . As a result, the UCCSD
|ΨH
HF i = |0011i .
2
(95) operator takes the form
† † † † † † † †
This represents the Slater determinant U = et02 (a2 a0 −a0 a2 )+t13 (a3 a1 −a1 a3 )+t0123 (a3 a2 a1 a0 −a0 a1 a2 a3 ) .
(100)
1 We can split this operator using Trotterization with a
ΨHHF (r1 , r2 ) = √ (σg↑ (r1 )σg↓ (r2 ) − σg↑ (r2 )σg↓ (r1 )),
2

2 single Trotter step

(96) † † † †
where ri is the position of electron i. The most general U = et02 (a2 a0 −a0 a2 ) × et13 (a3 a1 −a1 a3 )
state for H2 (with the same sz value and electron number † † † †
(101)
× et0123 (a3 a2 a1 a0 −a0 a1 a2 a3 ) .
as the HF state) is given by
Using the JW encoding, we find that
|ΨH2 i = α |0011i + β |1100i + γ |1001i + δ |0110i , (97) i
(a†2 a0 − a†0 a2 ) = (X2 Z1 Y0 − Y2 Z1 X0 )
and the ground state of the H2 molecule at its equilibrium 2
i
bond distance is given by (Helgaker et al., 2014) (a†3 a1 − a†1 a3 ) = (X3 Z2 Y1 − Y3 Z2 X1 )
2
|ΨH
g i = 0.9939 |0011i − 0.1106 |1100i .
2
(98) (a†3 a†2 a1 a0 − a†0 a†1 a2 a3 ) =
i
The first determinant in the ground state wavefunction (X3 Y2 X1 X0 + Y3 X2 X1 X0 + Y3 Y2 Y1 X0 + Y3 Y2 X1 Y0
8
is the HF state for H2 , showing that a mean-field solu-
− X3 X2 Y1 X0 − X3 X2 X1 Y0 − Y3 X2 Y1 Y0 − X3 Y2 Y1 Y0 ).
tion is a good approximation for this molecule at this
interatomic distance. The second determinant represents (102)
the antibonding state, and accounts for dynamical cor- It was shown by Romero et al. (2019) that all Pauli terms
relation between the electrons due to their electrostatic arising from the same excitation operators commute. As
repulsion. While the HF determinant dominates at the a result, each of the exponentials in Eq. (101) can be
equilibrium separation, at large separation the two de- separated into a product of exponentials of a single Pauli
terminants contribute equally to the wavefunction. This string. For example
† † it02 −it02
is because the bonding and antibonding configurations et02 (a2 a0 −a0 a2 ) = e 2 X2 Z1 Y0
×e 2 Y2 Z1 X0
. (103)
become degenerate. We require both determinants to
accurately describe the state, ensuring that only one Hempel et al. (2018) simplified the UCCSD operator for
electron locates around each atom. This is an example H2 by implementing the single excitation terms as ba-
of static correlation, which can also be dealt with using sis rotations, and combining terms in the double excita-
multiconfigurational self-consistent field methods, as tion operator (by considering the effect of each term on
described in Sec. III.C.1. the HF state). This latter technique is only possible be-
cause there is only one double excitation operator for this
As discussed previously, in order to find the ground molecule, and so is not a scalable technique in general.
state of the H2 molecule (using either the VQE or phase The UCCSD operator is simplified to
estimation), we need to construct the state on the quan- U = e−iθX3 X2 X1 Y0 . (104)
tum computer. Here we explicitly derive the unitary cou-
pled cluster (UCC) ansatz with single and double excita- This can be implemented using the circuit (Whitfield
tions (UCCSD) for H2 . As discussed in Sec. V.B.1, the et al., 2011) shown in Fig. 11.
UCCSD operator we seek to realise is given by Applying the simplified UCCSD operator to the HF
state in Eq. (95) gives
† †
U = e(T1 −T1 )+(T2 −T2 ) , U |0011i = (cos(θ)I − isin(θ)X3 X2 X1 Y0 ) |0011i ,
(105)
tiα a†i aα ,
X
T1 = = cos(θ) |0011i − sin(θ) |1100i ,
i∈virt,α∈occ (99)
which can reproduce the ground state given by Eq. (98).
tijαβ a†i a†j aα aβ ,
X
T2 =
i,j∈virt,α,β∈occ
2. 6-31G basis
where occ are initially occupied orbitals in the HF state,
virt are initially unoccupied orbitals in the HF state, and As discussed in Sec. III.D, H2 in the 6-31G basis
tiα and tijαβ are variational parameters to be optimised. has a double-zeta representation of the valence elec-
For H2 , the only operators which do not change the sz trons. This means we have 8 spin-orbitals to consider
 44

3. cc-PVDZ basis
σg↑ : |1 Rx ( π2 ) • • Rx (− π2 )

σg↓ : |1 Had • • Had As discussed in Sec. III.D, the cc-PVDZ basis for H2
includes a double-zeta representation of the valence shell,
σu↑ : |0 Had • • Had and additional polarisation orbitals. Each atom con-
σu↓ : |0 Had Rz (2θ) Had tributes {1s, 1s0 , 2px , 2py , 2pz } orbitals, resulting in 20
spin-orbitals in total. In order to reduce our active
space, we first change to the natural molecular orbital
FIG. 11 The circuit for implementing the UCCSD operator
(NMO) basis, using the single particle reduced density
for H2 in the STO-3G basis, as given by Eq. (104). The Rx ( π2 )
and Had gates rotate the basis such that the exponentiated matrix (1-RDM), as discussed in Sec. III.E. We first ob-
operator applied to the corresponding qubit is either Y or X, tain the 1-RDM for H2 in the cc-PVDZ basis with a
respectively. Single excitation terms are implemented with a classically tractable configuration interaction singles and
change of basis (Hempel et al., 2018). double (CISD) calculation.
We perform a unitary diagonalisation of this ma-
trix, and rotate the orbitals by the same unitary ma-
trix. This constitutes a change of basis to the NMOs
in total; {1s↑ , 1s↓ , 1s0↑ , 1s0↓ } from each atom. Working of the molecule. The diagonalised 1-RDM is given by
in the canonical orbital basis, (obtained by performing Diag(1.96588, 0.00611, 0.02104, 0.0002, 0.00001, 0.00314,
a Hartree–Fock calculation) we show how to construct 0.00314, 0.00016, 0.00016, 0.00016). There are only 10
Bravyi-Kitaev encoded states of 6-31G H2 . The BK diagonal entries in this 1-RDM because the spin-up and
transform matrix for an 8 spin-orbital system is given spin-down entries for the same spatial orbitals have been
by combined. As discussed in Sec. III.E, the diagonal entries
are the natural orbital occupation numbers (NOONs).
  We can see that the 5th orbital has a NOON that is 20
1 0 0 0 0 0 0 0 times smaller than the next smallest NOON. As a result,



 we consider this spatial orbital to always be empty, and so
1 1 0 0 0 0 0 0 remove all terms involving it from the Hamiltonian. This
 

0 0 1 0 0 0 0 0
 leaves a Hamiltonian acting on M = 18 spin-orbitals. We



 now map these into qubits using the BK-tree method.
1 1 1 1 0 0 0 0
 . (106) To map fermionic orbitals to qubits we follow a similar
procedure to that shown in Fig. 12 for the LiH molecule.
 
0 0 0 0 1 0 0 0
The reader will find that the 9th and 18th orbitals store
 
 
0 0 0 0 1 1 0 0
  the number of spin up electrons and total number of elec-
trons, respectively. As a result, they can be removed.
 
0 0 0 0 0 0 1 0
This reduces the problem to one of 16 qubits.
 
1 1 1 1 1 1 1 1
The lowest energy computational basis state of cc-
PVDZ H2 in the Jordan-Wigner encoding (18 qubits)
is |000000001000000001i. The corresponding BK-tree
We order the spin-orbitals such that the first M/2 spin- mapped state (16 qubits) is given by |0001011100010111i.
orbitals are spin up, and the final M/2 spin-orbitals are We stress that while this procedure may seem complex,
spin down. When the spin-orbitals are ordered in this it can in fact be easily implemented using the functions
way, the 4th entry in the BK encoded vector is the sum available in the aforementioned quantum computational
(mod 2) of the spin up occupancies, which sums to the chemistry packages.
number of spin up electrons. Moreover, the 8th entry is
the sum (mod 2) of all of the orbital occupancies, which
sums to the number of electrons. As these quantities are
conserved, we can remove these two qubits from the sim- B. Lithium Hydride STO-3G basis
ulation, following the procedure of Sec. IV.C. If the spin-
orbitals are arranged ‘up-down, up-down’, then while the For LiH in the STO-3G basis, we consider
8th entry is still equal to the number of electrons, the {1s, 2s, 2px , 2py , 2pz } functions for lithium, and a
4th entry is no longer necessarily equal to a conserved single {1s} orbital for hydrogen. This gives a total of
quantity. The JW mapped HF state (8 qubits) is given 12 spin-orbitals. We can reduce this problem to one of
by |00010001i. Using the matrix given above, the BK six qubits, as illustrated in Fig. 12. The 1-RDM in the
mapped HF state (8 qubits) is |00111011i. When the canonical orbital basis from a CISD calculation on LiH
two conserved qubits are removed, the BK mapped HF (at an internuclear separation of 1.45 Ȧ) is given by
state (6 qubits) is |011011i.
 45

Fermionic orbitals: STO-3G LiH

0 1 2 3 4 5 6 7 8 9 10 11
↑ ↓
1.99992 1.96206 0.03454 0.00005 0.00171 0.00171
Set occupied Set unoccupied

(All sums mod 2) BK-tree

transformation
Spin up ↑ Spin down ↓
Re- 2 4 8 10 3 5 9 11
order Re-
0 1 2 3 4 5 6 7 label

𝑥 * = 𝑛* + 𝑥 ) + 𝑥 '

𝑥 - = 𝑛- + 𝑥 .+ 𝑥 ,+ 𝑥 *
𝑥 . = 𝑛.
𝑥 $ = 𝑛$

𝑥 ' = 𝑛' + 𝑥 $

𝑥 ) = 𝑛)

𝑥 + = 𝑛+

𝑥 , = 𝑛, + 𝑥 +
FIG. 12 A pictorial representation of the fermion-to-qubit mapping procedure for LiH in the STO-3G basis. The fermionic nat-
ural molecular orbitals (NMO) are initially arranged ‘spin up, spin down, spin up, spin down, ...’, and have their corresponding
natural orbital occupation number (NOON) below. As the NOON of orbitals 6 and 7 is so small, they can be assumed unfilled,
and removed from the Hamiltonian. As the combined NOON of orbitals 0 and 1 is close to 2, they can be assumed filled, and
removed from the Hamiltonian. We then rearrange the remaining orbitals to be ‘all spin up, all spin down’, and re-label them
from 0 to 7. We then perform the BK-tree mapping by constructing the Fenwick tree, Fen(0,7), as described in Fig. 13. The
value xi is the value of the ith qubitPunder the BK-tree mapping, while P
ni is the value of the ith qubit under the JW mapping.
We see that qubit 3 stores the sum i=0 ni , and qubit 7 stores the sum 7i=0 ni . As these sums are conserved quantities, these
3

qubits do not flip throughout the simulation, and so can be removed from the Hamiltonian as described in Sec. IV.C.

 

1.9999 −0.0005 0.0006 0.0000 0.0000 −0.0010

from the Hamiltonian. In contrast, the fourth orbital
−0.0005 1.9598 0.0668

0.0000 0.0000

0.0084  has a very small NOON. As a result, we assume that this
orbital is never occupied, and so remove the two corre-
 
−0.0138
 
 0.0006 0.0668 0.0097 0.0000 0.0000


.
 sponding fermion operators from the Hamiltonian. This
 0.0000 0.0000 0.0000 0.0017 0.0000 0.0000 




leaves a Hamiltonian acting on 8 spin-orbitals. As the
 0.0000 0.0000 0.0000 0.0000 0.0017 0.0000  number of orbitals is now a power of 2, we can use ei-
 
−0.0010 0.0084 −0.0138 0.0000 0.0000 0.0273 ther the BK or BK-tree mappings to remove the 2 qubits
associated with conservation symmetries. We use the
There are only 6 rows/columns in this 1-RDM because
BK-tree mapping in order to provide an explicit example
the spin-up and spin-down entries for the same spatial or-
of Fenwick tree construction. The Fenwick tree tells us
bitals have been combined. We diagonalise this 1-RDM,
which qubits store which orbitals in the BK-tree map-
moving to the NMO basis. The NMO 1-RDM is given by
ping. We denote the Fenwick tree for the M orbitals as
  Fen(0, M − 1). We can obtain this data structure using
1.99992 0.00000 0.00000 0.00000 0.00000 0.00000




an iterative algorithm, which we reproduce from Havlicek
 0.00000 1.96206 0.00000 0.00000 0.00000 0.00000  et al. (2017) below. The generation of the Fenwick tree
 
for the LiH molecule using this algorithm is shown in
 
 0.00000 0.00000 0.03454 0.00000 0.00000 0.00000 
 .

 0.00000 0.00000 0.00000 0.00005 0.00000

0.00000  Fig. 13.
 
 
 0.00000 0.00000 0.00000 0.00000 0.00171 0.00000 
 
0.00000 0.00000 0.00000 0.00000 0.00000 0.00171

The first orbital has a NOON close to two, and so we con- Our final Hamiltonian acts on 6 qubits, but differs
sider it to always be doubly occupied. We can then re- in energy from the full 12 qubit Hamiltonian by only
move any terms containing a†0 , a0 , a†1 , a1 (the spin-orbitals 0.2 mHartree. A similar procedure is described by
corresponding to the first spatial orbital in the 1-RDM) Hempel et al. (2018).
 46

F(0,7): 7→ 3
F(0,3) 0 1 2 3 4 5 6 7
F(4,7)

F(0,3): 3 → 1 F(4,7): 7 → 5
F(0,1) F(4,5) 0 1 2 3 4 5 6 7
F(2,3) F(6,7)

F(0,1): F(2,3): F(4,5): F(6,7):

1→0 3→2 5→4 7→6 0 1 2 3 4 5 6 7
F(0,0): Fin F(2,2): Fin F(4,4): Fin F(6,6): Fin
F(1,1): Fin F(3,3): Fin F(5,5): Fin F(7,7): Fin

0 1 2 3 4 5 6 7

FIG. 13 A pictorial representation of the Fenwick tree construction for LiH, shown in Fig. 12. We carry out the BK-tree
mapping by constructing the Fenwick tree, Fen(0,7), as described in Algorithm. 1. The algorithmic steps are shown on the left
hand side of the figure, while the corresponding actions are shown on the right hand side. The notation X → Y means connect
spin-orbital X to spin-orbital Y with an arrow. ‘Fin’ means that the corresponding branch of the Fenwick tree is finished. The
finished Fenwick tree Fen(0,7) is shown at the bottom of the figure.

Algorithm.1 : Fenwick tree generation calculations on smaller system sizes. We conclude this
review in Sec. VIII with a blueprint for future investiga-
Define Fen(L, R) tions.
If L 6= R:
• Connect R to Floor R+L


2
; A. Classical limits
R+L



• Fen L, Floor 2 ;
R+L


As discussed in Sec. III, there are many different meth-
• Fen Floor 2 + 1, R . ods used in classical computational chemistry, which all
Else: seek to approximate the true ground state energy of the
• End the current Fenwick tree. system of interest, to varying degrees of accuracy. In gen-
eral, the cost of applying these methods grows with the
accuracy of the results that they produce, and the size
of the system simulated. In this section, we discuss the
VIII. DISCUSSION AND CONCLUSIONS system sizes that one can typically accurately treat using
some of the commonly used classical methods. This will
We have now reviewed the key concepts in both clas- help us to elucidate where, and when, quantum comput-
sical, and quantum, computational chemistry. In partic- ers may become useful for chemistry simulation. As in
ular, we have discussed and shown how to map chemi- previous sections, M denotes the number of spin-orbitals
cal problems onto quantum computers, and how to solve considered in basis set approaches, and N denotes the
them to obtain both the ground and excited states. We number of electrons in the system.
now turn our attention to how these techniques com- At one end of the spectrum are density functional the-
pare to the established classical methods discussed in ory (DFT), and the Hartree-Fock (HF) method. These
Sec. III. We will first review the applicability and lim- calculations are often very efficient to run, and can treat
itations of the various classical methods, in Sec. VIII.A. large systems. As a result, they are used widely in
This will highlight the problems for which quantum com- chemistry and materials science. However, these tech-
puters may one day be useful. We discuss the resources niques can struggle to achieve highly accurate results for
required for such calculations in Sec. VIII.B. We will strongly correlated systems, and are not systematically
see that the resources required are considerably greater improvable. Consequently, they are often used for ob-
than what we currently have available, at time of writing. taining qualitative results for large system sizes. We do
Consequently, in Sec. VIII.C we consider routes towards not expect these calculations to be replaced by those on
these calculations. This will include both heuristic calcu- quantum computers, given the large system sizes that are
lations on classically intractable system sizes, and exact simulated.
 47

At the other end of the spectrum, are exact cal- functions (Booth and Alavi, 2010)), as well as state of the
culations – by which we mean the exact energy that art results in Fermi-Hubbard systems that are accurate
can be obtained from the model of the system. It is to around 100 sites (LeBlanc et al., 2015), and less accu-
important to note that these ‘exact’ calculations are rate results in other, larger systems (Austin et al., 2012).
rarely performed. Moreover, the degree of accuracy However, Monte Carlo methods are not without their
depends on the details included in the calculation, such own shortcomings, including the infamous ‘sign problem’,
as: the inclusion of relativistic corrections, whether the that affects fermionic simulations (Austin et al., 2012;
Born-Oppenheimer approximation is used, or whether Ortiz et al., 2001).
nuclear vibrational and rotational contributions are Tensor network methods, such as density matrix renor-
included. Grid-based simulations (as described in malisation group (DMRG), have proven effective for
Sec. III.B.1) provide the most accurate results – but can dealing with systems displaying strong static correla-
only be carried out for a very small number of particles, tion. They provide an alternative approach to CASSCF
because of the large number of grid points required. (see Sec. III.C.2) approaches (Knecht et al., 2016), al-
Exact results (albeit with a less accurate model of the lowing the treatment of larger active spaces, including
system) can also be obtained by carrying out basis those of metalloenzyme complexes with active spaces of
set, full configuration interaction (FCI) calculations over 70 spin-orbitals (Kurashige et al., 2013; Sharma
(Sec. III.B.2), which are less computationally expensive et al., 2014). This is larger than the roughly 30 to 40
than grid based approaches. However, the cost of these spin-orbitals that can be treated with a CASSCF ap-
calculations still scales exponentially with the system proach (Lischka et al., 2018). Tensor network methods
size, so they are only applicable to small systems, like are also useful for treating systems in condensed matter
the dinitrogen molecule (Rossi et al., 1999). In the physics, including Fermi-Hubbard models that are accu-
context of condensed matter physics, these calculations rate to around 100 sites (LeBlanc et al., 2015). While ten-
are referred to as ‘exact diagonalisation’ and are possible sor network methods are best suited to dealing with sys-
up to system sizes of around 20 to 30 lattice sites, in tems with strong static correlation, recent work has inves-
the case of Fermi-Hubbard models (Jiang et al., 2018b; tigated post-DMRG methods to recover dynamic corre-
Yamada et al., 2005). lation (Knecht et al., 2016; Yanai et al., 2015). For more
information on the use of DMRG in quantum chemistry,
The vast majority of calculations carried out by the we refer the reader to the reviews by Olivares-Amaya
computational chemistry community do not achieve this et al. (2015) and Szalay et al. (2015).
level of accuracy. Instead, approximate, less costly We can see from the above discussion that there ap-
methods are used, such as: configuration interaction pears to be an untreated ‘sweet spot’, of systems with
(Sec. III.C.3), coupled cluster (Sec. III.C.4) multiconfig- around 100 to 200 spin-orbitals, that require high ac-
urational self-consistent field (Sec. III.C.2), tensor net- curacy calculations. These systems are too strongly
work methods, or quantum Monte Carlo. An exhaustive correlated to be tackled with methods like HF, DFT
comparison of these methods is beyond the scope of this or even CCSD. They are also too large to be reliably
review, as is attempting to catalogue the ever-evolving dealt with using DMRG or quantum Monte Carlo, and
list of the largest calculations performed. However, we much too large for classical FCI methods. Interestingly,
will briefly highlight some of the system sizes where these many problems of scientific interest fit this description,
methods have been successfully applied. including: transition metal catalysts (Podewitz et al.,
Coupled cluster methods (often CCSD) are some of 2011; Vogiatzis et al., 2019) and the Fermi-Hubbard
the most widely used high-accuracy methods. They model (LeBlanc et al., 2015). As we have discussed
are applicable to large systems (hundreds of spin- throughout this review, a small quantum computer, with
orbitals (Bartlett and Musial, 2007)) which do not around 100 perfect qubits, would be able to calculate the
display strong static correlation. Examples include: FCI energy of a system with around 100 spin-orbitals in
the DNA base guanine (C5 H5 N5 O) in a cc-PVTZ ba- polynomial time. This would imply that these problems
sis (Hobza and Sponer, 2002), or the hydrocarbon octane are among the best targets for quantum computers.
(C8 H18 ) also in a cc-PVTZ basis (Yamazaki et al., 2018). It is important to note that being able to accurately
While CC methods can also be applied to strongly cor- predict the ground state energy of 100 spin-orbital
related systems (LeBlanc et al., 2015; Watson and Chan, systems still leaves us far from our long-term goal of
2012), higher excitation degrees are often required, mak- designing new medicines and materials with simulations.
ing the method more costly to implement. For example, Yamazaki et al. (2018) noted that over
Quantum Monte Carlo has many variants, and has 95 % of the approved drug molecules in DrugBank 5.0
been used to obtain results comparable to FCI in rel- are larger than these 100-200 spin-orbital systems that
atively small systems (the Cr2 molecule with 24 active we might aim to simulate with a small, error-corrected
electrons in 30 spin-orbitals (Tubman et al., 2016), or quantum computer. However, in practice it is not always
the fluorine atom in a cc-PV5Z basis with additional basis necessary to perform highly accurate calculations on the
 48

entirety of a large molecule or enzyme. Instead, problem and Gidney, 2018; Gidney and Fowler, 2019; Litinski,
decomposition approaches can be utilised, whereby 2019a,b; Trout and Brown, 2015). These improvements,
the most important part of the system is accurately combined with the algorithmic advances described
simulated, and then integrated with a potentially less throughout this review, have contributed to a signifi-
accurate simulation of the less challenging parts of the cant reduction in the resources required for chemistry
system. This approach has been investigated in the calculations, compared to the initial estimates. In order
context of quantum computing by Bauer et al. (2016); to distinguish algorithmic advances from fault-tolerance
Kreula et al. (2016); Reiher et al. (2017); Rubin (2016); improvements, we list both the number of T and/or
and Yamazaki et al. (2018). Toffoli gates required for the simulation, as well as the
corresponding number of physical qubits, obtained using
the best fault-tolerance procedures available at that
work’s time of writing.
B. Quantum resources: medium to long-term
Reiher et al. (2017) carried out a fault-tolerant re-
As discussed in Sec. IV.B, quantum computers can source estimation for the problem of biological dinitro-
store the FCI wavefunction of M spin-orbitals using only gen fixation, as described in Sec. III.A.1. Those authors
M qubits in second quantisation. However, as discussed calculated the resources required to perform an FCI cal-
in Sec. II, we must also take into consideration the qubit culation on an active space of 54 electrons in 108 spin-
overhead of error correction. Initial work suggested that orbitals for FeMo-co, using a Trotter-based approach to
around 1018 gates would be necessary to perform phase phase estimation. They found that this would require
estimation on a system of around 100 spin-orbitals (ex- around 1014 T gates. Assuming the best physical error
cluding the overhead of error correction) (Wecker et al., rates (10−3 ) at our time of writing, this would require
2014). This estimate was subsequently reduced through around 200 million physical qubits, and take on the order
a series of algorithmic optimisations, as described in of weeks (10 ns to implement a T gate, including surface
Sec V.A.4.a. code decoding) or months (100 ns per T gate) (Reiher
These initial estimates did not focus on specific prob- et al., 2017). We note that those authors were considering
lems of interest, and neglected the overhead of quantum a targeted majorana fermion-based quantum computer,
error correction, which is necessitated by the large with physical error rates 103 times lower than has been
number of gates needed. Fault-tolerant resource estima- demonstrated in trapped ion or superconducting qubits,
tions have since been carried out for two main systems: at our time of writing.
small transition metal molecules, and condensed phase A similar resource estimation (although using the more
materials (including 2D Fermi-Hubbard models, 2D and accurate FeMo-co active space of Li et al. (2018), with
3D electron gases, and solid materials, such as lithium 113 electrons in 152 spin-orbitals) was carried out by
or diamond). When performing fault-tolerant resource Berry et al. (2019b), who used the algorithm based on
estimates, one must specify details of the problem, the qubitization and low-rank decompositions of the Hamil-
hardware considered, and the error correcting code tonian, described in Sec. V.A.4. This approach reduced
used. All resource estimates to date have focused on the resources required to around 1011 Toffoli gates (which
the 2D surface code, due to its high threshold, and are currently the bottleneck for this approach). While a
suitability for architectures with a 2D nearest-neighbour complete fault-tolerant resource analysis for this new ap-
connectivity. In the standard model of surface code proach has not yet been performed, those authors showed
resource estimation, Clifford gates (such as Pauli gates that the cost of Toffoli gate distillation is equivalent to
and the CNOT gate) are considered to be of negligible around 1 million physical qubits, working for 2 months
cost, while non-Clifford gates (such as the T gate, or (assuming 10−3 error rates and surface code cycle times
Toffoli gate) are more costly. This is because these of 1 µs), at our time of writing.
non-Clifford gates cannot be natively implemented in a
fault-tolerant way in the surface code, but instead are Other works have conducted similar resource esti-
typically implemented using magic state distillation and mations for equivalently sized problems (100-200 spin-
teleportation (Bravyi and Kitaev, 2005), which is often orbitals), but have focused on matter in the condensed
expensive (Campbell et al., 2017). As a result, algorithm phase. This enables the use of the plane wave dual ba-
complexities are measured in terms of the number of T sis (Sec. III.D.3), which as discussed in Sec. V.A.4.b can
and/or Toffoli gates that they contain, as these are often reduce the costs of simulations. Babbush et al. (2018b)
the dominant contribution to the cost of the algorithm. used the algorithm based on qubitization in a plane dual
There has been considerable work to reduce the cost wave basis (discussed in Sec. V.A.4.b) to obtain resource
of operations in the surface code (including magic estimates of around 2 × 108 T gates for a 128 spin-
state distillation), which has reduced the overhead of orbital 3D homogeneous electron gas (with similar re-
error correction by several orders of magnitude (Fowler sults for other 3D materials), and around 7.1 × 108 T
 49

gates for the 2D Fermi-Hubbard model with 100 lattice

sites (200 spin-orbitals). Assuming error rates of 10−3 , One might assume that grid based methods will
1 µs to implement a T gate (including surface code de- require considerably larger quantum computers than
coding), and 10 µs feedforward, this led to resources of basis set approaches, given the former’s non-compact
around 2-3 million qubits, running for tens of hours (us- description of the wavefunction. However, as described
ing the best fault-tolerance protocols available when that above, the figure of merit for fault-tolerant calculations
work was completed). Subsequent improvements in fault- is often the circuit depth – in particular the number of T
tolerance protocols have further reduced these physical or Toffoli gates. While an initial investigation into fault-
resources (Fowler and Gidney, 2018; Gidney and Fowler, tolerant grid based simulation was performed (Jones
2019; Litinski, 2019a,b). et al., 2012), it did not calculate the total number of T
Kivlichan et al. (2019a) also performed resource es- gates required, or the number of qubits, for systems of
timations for condensed matter and solid state prob- interest. As such, it is not directly comparable to the
lems, using the Trotter based algorithm discussed in methods described above. The algorithm investigated
Sec. V.A.4.b. They considered simulations targeting a in that work (the algorithm of Kassal et al. (2008))
size-extensive error in the energy (which is appropriate has since been surpassed by the algorithm of Kivlichan
when considering scaling to the thermodynamic limit). et al. (2017). Moreover, there have been many im-
They found that simulations of a 100 site Fermi-Hubbard provements in fault-tolerant circuit design and magic
model, and a 128 spin-orbital homogeneous electron gas state distillation since the work of Jones et al. (2012).
would both require on the order of 106 Toffoli gates and As classical computers are limited to small grid based
107 − 108 T gates. Assuming error rates of 10−3 , surface calculations, it would be an interesting direction of
code error detection times of 1 µs, and surface code er- future research to establish the quantum resources re-
ror decoding times of 10 µs, they require around 400,000 quired to surpass these small, high accuracy calculations.
– 600,000 physical qubits, running for a couple of hours.
As these simulations consider a loose-but-rigorous bound These results suggest that certain calculations with
on the energy error (as discussed in Sec. V.A.5), these around 100 spin-orbitals may be better suited to early
resource estimates may be overly pessimistic. Kivlichan quantum computers than others. In particular, materials
et al. (2019a) found that if an intensive (e.g. absolute) in the condensed phase and Fermi-Hubbard simulations
error in the energy is targeted, then their algorithm was so far require considerably fewer resources than simula-
less efficient than the qubitization algorithm of Babbush tions of individual molecules. Despite these promising re-
et al. (2018b). sults, and recent, rapid improvements, we see that it still
requires on the order of 100, 000 physical qubits to sur-
All of the resource estimation papers above focus on pass classical techniques. Current quantum computers
the cost of phase estimation, and assume that the system possess only around 100 physical qubits, and we have yet
is prepared in an initial state that has a sufficiently to demonstrate a fully protected logical qubit. It may be
large overlap with the true ground state. As discussed many years before we possess a quantum computer with
in Sec. V.A.2, there are many techniques for state the resources required to implement these algorithms,
preparation, including adiabatic state preparation, or given the challenges in scaling up hardware and perform-
variational approaches. Tubman et al. (2018) presented ing quantum error correction (Gambetta et al., 2017).
an algorithm which can prepare a suitable initial state by In order to attempt to solve classically intractable chem-
leveraging a classical adaptive configuration interaction istry problems before that time, different approaches are
method. This algorithm was numerically shown to required. We discuss the potential paths towards these
provide good estimates for the ground states of many classically intractable simulations below.
systems of interest in chemistry, physics and materials
science. This stems from the fact that the dominant
Slater determinants in the wavefunction typically con- C. Quantum resources: near to medium-term
verge much more quickly than the correlation energy.
Tubman et al. (2018) showed that the cost of this state As discussed above, existing estimates for surpassing
preparation algorithm is considerably lower than the classical calculations require on the order of 100,000 phys-
cost for phase estimation detailed above, and hence, ical qubits, in order to implement quantum error correc-
may be neglected for many systems of interest. This tion. The first generations of quantum computers will be
technique can also be combined with that of Berry et al. significantly smaller than this. Nevertheless, there are
(2018), which used classically obtained knowledge of many interesting avenues to pursue with these first ma-
the energy eigenvalues to reduce the number of times chines.
that phase estimation must be repeated. This reduced As discussed in Sec. V.B, the variational quantum
the necessity of having a large overlap with the ground eigensolver (VQE) has received significant attention in
state. recent years, due to its short required circuit depth (com-
 50

pared to phase estimation). It has been speculated that neglects constant factors, and so the number of gates
the VQE may enable small quantum computers with 100 required would likely be higher. More thorough resource
to 200 physical qubits to surpass classical methods. Con- estimates for the same problem on a silicon quantum
siderable further work is required to demonstrate that architecture were performed by Cai (2019). Our estimate
this will be possible. The VQE is a heuristic approach, of 40,000 two-qubit gates is approximately equal to the
which attempts to generate an approximation to the number of gates that we were limited to by noise in
ground state wavefunction that is better than classical our back-of-the-envelope calculation above. There are
methods, using a short circuit. It is difficult to prove a number of routes to try and overcome the issue of
that a given circuit will be able to obtain a good esti- noise in such a calculation. We may be able to combine
mate for the ground state, especially when the difficulty existing error mitigation techniques, or try to develop
of classical optimisation is considered. new ones. We could also utilise ansätze which appear
In general, the longer the circuit is, the better it can robust to noise (Borregaard et al., 2019; Kim, 2017; Kim
approximate the ground state wavefunction. However, and Swingle, 2017).
the length of circuit that we can implement without
error correction is heavily limited by noise. A simple A less widely discussed approach is to perform
calculation demonstrates the limited number of gates error corrected VQE simulations. The aim would be
that we have available. If we assume a discrete error to suppress the error rate to a value low enough to
model for our circuit, such that error events happen obtain chemically accurate energies from the simulation.
probabilistically and independently following each gate For example, we could use fermion-to-qubit mappings
in the circuit, then even with an optimistic two qubit which enable the detection and/or correction of single
gate error rate of 0.01 % (10 times better than the error qubit errors (Jiang et al., 2018a; Setia et al., 2018)
rates achieved to date), we could only carry out around (discussed in Sec. IV.B.3 and Sec. VI.D). Alternatively,
10,000 gates before we expect one or more errors to occur we could explore using small error correcting codes. An
in the circuit with high probability. While the error initial foray into this area was conducted by Urbanek
mitigation techniques discussed in Sec. VI may enable et al. (2019), who experimentally implemented a VQE
us to recover accurate results from a circuit deeper than calculation on the H2 molecule encoded in the [[4,2,2]]
10,000 gates, it seems unlikely that these methods alone error detecting code. This calculation showed improved
would enable more than a small multiplicative increase accuracy over an un-encoded calculation, due to the use
in the circuit depth. of post-selection. Nevertheless, the use of error correct-
ing codes is complicated by the difficulty of producing
It is unclear whether we would be able to surpass error protected T gates. As such, it is important to ask
classical methods for any chemistry problems with this if variational algorithms can be implemented with fewer
number of gates. The Fermi-Hubbard model is one T gates than their phase estimation based counterparts.
of the leading candidates for such a simulation. As As an example, the phase estimation based algorithm of
described in Sec. V.B.1, the Hamiltonian variational Kivlichan et al. (2019a) is already only a constant factor
ansatz is particularly efficient for this problem. We less efficient (in terms of T/Toffoli gates) than just
can prepare initial states of the Fermi-Hubbard model implementing time evolution under the Fermi-Hubbard
using O(M 1.5 ) gates, and perform Trotter steps of Hamiltonian directly (as is required for a Hamiltonian
the Hamiltonian using O(M ) gates for each Trotter variational ansatz). Moreover, synthesizing an arbitrary
step (Jiang et al., 2018b). Previous work has shown angle rotation gate can require at least 10 to 100 T
that the Hamiltonian variational ansatz performs well gates. For the hypothetical O(M 2 ) scaling variational
for the Fermi-Hubbard model, although it is not yet algorithm described above, we may still need around
known how many Trotter steps may be required for 4 × 105 to 4 × 106 T gates, neglecting constant factors.
accurate results. Wecker et al. (2015a) achieved good This is comparable to the T gate counts required
convergence for a 12 site problem with 20 Trotter steps. for phase estimation based approaches described in
Further work has shown promising results for both Sec. VIII.B. In addition, the long duration of such a
the ground state problem (Reiner et al., 2018a), and computation could be problematic, given the potentially
dynamics simulation (Reiner et al., 2018b), both in large number of measurements required by the VQE.
the presence of realistic noise rates. These results were
obtained using less efficiently scaling circuits than those An alternative approach to doggedly pursuing classi-
described above. Nevertheless, even if the number of cally intractable problems is to use chemistry calculations
Trotter steps required to find the ground state scaled whose results we do know as a benchmark of our tech-
only linearly with the number of spin-orbitals, the total nology. This effort has arguably already begun, follow-
algorithm scaling would be O(M 2 ), which is around ing the publication of many VQE experiments on small
40,000 two-qubit gates for a 10 × 10 site Fermi-Hubbard molecules like H2 and LiH, in a range of different hard-
model (which requires 200 qubits). This rough estimate ware systems. This proposal has recently been formalised
 51

by McCaskey et al. (2019) and Nam et al. (2019). We posals, and determine which look most promising. This
cannot expect to surpass classical methods without first review is a first step in this direction, as is the growing
reproducing classically known results. In a similar vein, availability of quantum computational chemistry pack-
once we are able to experimentally demonstrate error cor- ages, and software libraries to emulate quantum comput-
rected logical qubits, the next step will be to perform ers. Fast numerical simulations may enable us to test
small, error corrected demonstrations of the algorithms variational algorithms on systems with up to around 30
described throughout this review. This approach was to 40 qubits. This may begin to show which methods
recommended by Love (2012), who charted the evolution are most suitable for near-term hardware. These cal-
of classical computational chemistry milestones since the culations should be performed both with, and without
1930’s, and selected target problems to emulate. Equiv- noise, in order to ascertain the all-round performance of
alent targets would be small Fermi-Hubbard models, or the various techniques. As quantum hardware develops,
the G1 set of molecules (Pople et al., 1989). This is a this effort can be migrated onto real systems, in order to
small set of molecules whose energy is known extremely test whether our algorithms are as effective as we expect
accurately. For many of the molecules in the G1 set, an them to be. We expect that this more focused research
FCI calculation on a sufficiently accurate basis set would program will lead to new developments, as well as the
be classically intractable (although they are typically not optimisation of existing methods.
necessary due to highly accurate approximate methods It is more difficult to construct a road-map for
and experimental results). As a result, this may be an phase estimation based approaches to solving chemistry
excellent test case for future quantum computers. problems, given the higher degree of sophistication of
these methods, and the fact that there is less variation
between the different approaches. It is also difficult
D. Summary and outlook to anticipate breakthroughs that can lead to a large
reduction in required resources, such as the introduction
This review has sought to be accessible to both scien- of the qubitization technique or tightening of Trotter
tists working on quantum computing, and those work- error bounds. One potential route to new developments
ing on computational chemistry. We have discussed the is to investigate areas that appear to be less well ex-
key methods used in classical computational chemistry, plored. One example may be error correction procedures
and how these have been incorporated into quantum al- that are tailored specifically for chemistry problems.
gorithms. Emphasis has been placed on the key differ- Alternatively, one could import ideas that are well
ences between quantum and classical methods of chem- established in classical computational chemistry. This is
istry simulation, and the resulting benefits that quan- how transformative ideas like the plane wave basis sets
tum computing is widely predicted to bring to the field and low rank Hamiltonian decomposition entered the
of computational chemistry. field.
However, we have also shown that quantum methods
still face many challenges, not least the high error rates Successful exploration of these future directions may
and low qubit counts of existing hardware. Ultimately, only prove possible through close collaboration between
the success of quantum computational chemistry will chemists and quantum information scientists. We hope
depend on our ability to construct larger and better that this review helps to develop a common language for
controlled quantum computers. The question of how these two groups, facilitating this important collabora-
large, and how well controlled these machines must be tion.
will be determined by the quality of the procedures that
we have developed to carry out calculations of interest.
It is therefore crucial that we continue to develop and
optimise new: algorithms, mappings, error correction ACKNOWLEDGMENTS
codes and procedures, basis sets, and error mitigation
techniques. Below, we highlight potential research This work was supported by BP plc and by the EPSRC
directions to aid in this goal. National Quantum Technology Hub in Networked Quan-
tum Information Technology (EP/M013243/1). A.A.G.
In the realm of variational algorithms, a wide range acknowledges Anders G Froseth for his generous support,
of ansätze, chemistry mappings, classical optimisation as well as the Vannevar Bush Faculty Fellowship program
routines, and error mitigation techniques have been pro- of the US Department of Defense. We thank R. Babbush
posed in recent years. However, the vast majority of these for insightful comments. S.M. and X.Y. thank L. Lindoy
proposals: were tested on small system sizes, performed a for initial discussions on basis sets. SE is supported by
limited number of comparisons to other techniques, and Japan Student Services Organization (JASSO) Student
were not optimised for maximum efficacy. We suggest Exchange Support Program (Graduate Scholarship for
that future work should begin to collate the existing pro- Degree Seeking Students).
 52

REFERENCES Babej, T., C. Ing, and M. Fingerhuth (2018),

arXiv:1811.00713.
Abrams, D. S., and S. Lloyd (1997), Phys. Rev. Lett. 79, Ball, R. C. (2005), Phys. Rev. Lett. 95, 176407.
2586. Ballance, C. J., T. P. Harty, N. M. Linke, M. A. Sepiol, and
Abrams, D. S., and S. Lloyd (1999), Phys. Rev. Lett. 83, D. M. Lucas (2016), Phys. Rev. Lett. 117, 060504.
5162. Banchi, L., M. Fingerhuth, T. Babej, C. Ing, and J. M.
Aharonov, D., and M. Ben-Or (1997), in Proceedings of the Arrazola (2019), arXiv:1902.00462.
twenty-ninth annual ACM symposium on Theory of com- Barends, R., J. Kelly, A. Megrant, A. Veitia, D. Sank, E. Jef-
puting (ACM) pp. 176–188. frey, T. C. White, J. Mutus, A. G. Fowler, B. Campbell,
Aharonov, D., W. Van Dam, J. Kempe, Z. Landau, S. Lloyd, et al. (2014), Nature 508 (7497), 500.
and O. Regev (2008), SIAM review 50 (4), 755. Barkoutsos, P. K., J. F. Gonthier, I. Sokolov, N. Moll,
Albash, T., and D. A. Lidar (2018), Rev. Mod. Phys. 90, G. Salis, A. Fuhrer, M. Ganzhorn, D. J. Egger, M. Troyer,
015002. A. Mezzacapo, S. Filipp, and I. Tavernelli (2018), Phys.
Anderson, P. W. (2002), Physica Scripta T102 (1), 10. Rev. A 98, 022322.
Argüello-Luengo, J., A. González-Tudela, T. Shi, P. Zoller, Bartlett, R. J., S. A. Kucharski, and J. Noga (1989), Chem-
and J. I. Cirac (2018), arXiv:1807.09228. ical Physics Letters 155 (1), 133 .
Arute, F., K. Arya, R. Babbush, D. Bacon, J. C. Bardin, Bartlett, R. J., and M. Musial (2007), Rev. Mod. Phys. 79,
R. Barends, R. Biswas, S. Boixo, F. G. S. L. Brandao, 291.
D. A. Buell, B. Burkett, Y. Chen, Z. Chen, B. Chiaro, Bauer, B., D. Wecker, A. J. Millis, M. B. Hastings, and
R. Collins, W. Courtney, A. Dunsworth, E. Farhi, B. Foxen, M. Troyer (2016), Phys. Rev. X 6, 031045.
A. Fowler, C. Gidney, M. Giustina, R. Graff, K. Guerin, Bernstein, E., and U. Vazirani (1997), SIAM Journal on com-
S. Habegger, M. P. Harrigan, M. J. Hartmann, A. Ho, puting 26 (5), 1411.
M. Hoffmann, T. Huang, T. S. Humble, S. V. Isakov, E. Jef- Berry, D. W., G. Ahokas, R. Cleve, and B. C. Sanders (2007),
frey, Z. Jiang, D. Kafri, K. Kechedzhi, J. Kelly, P. V. Communications in Mathematical Physics 270 (2), 359.
Klimov, S. Knysh, A. Korotkov, F. Kostritsa, D. Land- Berry, D. W., and A. M. Childs (2012), Quantum Info. Com-
huis, M. Lindmark, E. Lucero, D. Lyakh, S. Mandrà, J. R. put. 12 (1-2), 29.
McClean, M. McEwen, A. Megrant, X. Mi, K. Michielsen, Berry, D. W., A. M. Childs, R. Cleve, R. Kothari, and
M. Mohseni, J. Mutus, O. Naaman, M. Neeley, C. Neill, R. D. Somma (2014), in Proceedings of the Forty-sixth An-
M. Y. Niu, E. Ostby, A. Petukhov, J. C. Platt, C. Quin- nual ACM Symposium on Theory of Computing, STOC ’14
tana, E. G. Rieffel, P. Roushan, N. C. Rubin, D. Sank, K. J. (ACM, New York, NY, USA) pp. 283–292.
Satzinger, V. Smelyanskiy, K. J. Sung, M. D. Trevithick, Berry, D. W., A. M. Childs, R. Cleve, R. Kothari, and R. D.
A. Vainsencher, B. Villalonga, T. White, Z. J. Yao, P. Yeh, Somma (2015a), Phys. Rev. Lett. 114, 090502.
A. Zalcman, H. Neven, and J. M. Martinis (2019), Nature Berry, D. W., A. M. Childs, and R. Kothari (2015b), in Foun-
574 (7779), 505. dations of Computer Science (FOCS), 2015 IEEE 56th An-
Aspuru-Guzik, A., A. D. Dutoi, P. J. Love, and M. Head- nual Symposium on (IEEE) pp. 792–809.
Gordon (2005), Science 309 (5741), 1704. Berry, D. W., A. M. Childs, and R. Kothari (2015c), in 2015
Aspuru-Guzik, A., R. Lindh, and M. Reiher (2018), ACS IEEE 56th Annual Symposium on Foundations of Com-
Central Science 4 (2), 144. puter Science, pp. 792–809.
Aspuru-Guzik, A., and P. Walther (2012), Nature Physics Berry, D. W., A. M. Childs, Y. Su, X. Wang, and N. Wiebe
8 (4), 285. (2019a), arXiv:1906.07115.
Austin, B. M., D. Y. Zubarev, and W. A. Lester Berry, D. W., C. Gidney, M. Motta, J. R. McClean, and
(2012), Chemical Reviews 112 (1), 263, pMID: 22196085, R. Babbush (2019b), arXiv:1902.02134.
https://doi.org/10.1021/cr2001564. Berry, D. W., B. L. Higgins, S. D. Bartlett, M. W. Mitchell,
Babbush, R., D. W. Berry, I. D. Kivlichan, A. Y. Wei, P. J. G. J. Pryde, and H. M. Wiseman (2009), Phys. Rev. A 80,
Love, and A. Aspuru-Guzik (2016), New Journal of Physics 052114.
18 (3), 033032. Berry, D. W., M. Kieferová, A. Scherer, Y. R. Sanders, G. H.
Babbush, R., D. W. Berry, J. R. McClean, and H. Neven Low, N. Wiebe, C. Gidney, and R. Babbush (2018), npj
(2018a), arXiv:1807.09802. Quantum Information 4 (1), 22.
Babbush, R., D. W. Berry, Y. R. Sanders, I. D. Kivlichan, Blatt, R., and C. F. Roos (2012), Nature Physics 8 (4), 277.
A. Scherer, A. Y. Wei, P. J. Love, and A. Aspuru-Guzik Boixo, S., S. V. Isakov, V. N. Smelyanskiy, R. Babbush,
(2017), Quantum Science and Technology 3 (1), 015006. N. Ding, Z. Jiang, M. J. Bremner, J. M. Martinis, and
Babbush, R., C. Gidney, D. W. Berry, N. Wiebe, J. McClean, H. Neven (2018), Nature Physics 14 (6), 595.
A. Paler, A. Fowler, and H. Neven (2018b), Phys. Rev. X Bonet-Monroig, X., R. Babbush, and T. E. O’Brien (2019),
8, 041015. arXiv:1908.05628.
Babbush, R., P. J. Love, and A. Aspuru-Guzik (2014), Sci- Bonet-Monroig, X., R. Sagastizabal, M. Singh, and T. E.
entific reports 4, 6603. O’Brien (2018), arXiv:1807.10050.
Babbush, R., J. McClean, D. Wecker, A. Aspuru-Guzik, and Booth, G. H., and A. Alavi (2010), The Jour-
N. Wiebe (2015), Phys. Rev. A 91, 022311. nal of Chemical Physics 132 (17), 174104,
Babbush, R., A. Perdomo-Ortiz, B. O’Gorman, https://doi.org/10.1063/1.3407895.
W. Macready, and A. Aspuru-Guzik (2012), Borregaard, J., M. Christandl, and D. S. Frana (2019),
10.1002/9781118755815.ch05, arXiv:1211.3422. arXiv:1909.04786.
Babbush, R., N. Wiebe, J. McClean, J. McClain, H. Neven, Braunstein, S. L., and P. van Loock (2005), Rev. Mod. Phys.
and G. K.-L. Chan (2018c), Phys. Rev. X 8, 011044. 77, 513.
Bravyi, S., J. M. Gambetta, A. Mezzacapo, and K. Temme
 53

(2017), arXiv preprint arXiv:1701.08213. A. Aspuru-Guzik (2018), arXiv preprint arXiv:1801.01053.

Bravyi, S., and A. Kitaev (2005), Phys. Rev. A 71, 022316. Dallaire-Demers, P.-L., and F. K. Wilhelm (2016a), Phys.
Bravyi, S. B., and A. Y. Kitaev (2002), Annals of Physics Rev. A 93, 032303.
298 (1), 210 . Dallaire-Demers, P.-L., and F. K. Wilhelm (2016b), Phys.
Brown, K. L., W. J. Munro, and V. M. Kendon (2010), En- Rev. A 94, 062304.
tropy 12 (11), 2268. Daskin, A., A. Grama, and S. Kais (2014), Quantum Infor-
Burgess, B. K., and D. J. Lowe (1996), Chem- mation Processing 13 (2), 333.
ical Reviews 96 (7), 2983, pMID: 11848849, Dawson, C. M., and M. A. Nielsen (2005), arXiv:quant-
https://doi.org/10.1021/cr950055x. ph/0505030.
Cai, Z. (2019), arXiv:1910.02719. Devitt, S. J., W. J. Munro, and K. Nemoto (2013), Reports
Campbell, E. (2018), arXiv preprint arXiv:1811.08017. on Progress in Physics 76 (7), 076001.
Campbell, E. T., B. M. Terhal, and C. Vuillot (2017), Nature Dirac, P. (1929), Proceedings of the Royal Society of London
549 (7671), 172. 123, 10.1098/rspa.1929.0094.
Cao, Y., J. R. Fontalvo, and A. Aspuru-Guzik (2018), IBM Ditchfield, R., W. J. Hehre, and J. A. Pople (1971), The
Journal of Research and Development , 1. Journal of Chemical Physics 54 (2), 724.
Cao, Y., J. Romero, J. P. Olson, M. Degroote, P. D. John- DiVincenzo, D. P. (1995), Phys. Rev. A 51, 1015.
son, M. Kieferov, I. D. Kivlichan, T. Menke, B. Per- Du, J., N. Xu, X. Peng, P. Wang, S. Wu, and D. Lu (2010),
opadre, N. P. D. Sawaya, S. Sim, L. Veis, and A. Aspuru- Phys. Rev. Lett. 104, 030502.
Guzik (2019), Chemical Reviews 119 (19), 10856, pMID: Dunning Jr., T. H. (1989), The Journal of Chemical Physics
31469277, https://doi.org/10.1021/acs.chemrev.8b00803. 90 (2), 1007.
Chen, L.-K., Z.-D. Li, X.-C. Yao, M. Huang, W. Li, H. Lu, Dür, W., M. J. Bremner, and H. J. Briegel (2008), Phys.
X. Yuan, Y.-B. Zhang, X. Jiang, C.-Z. Peng, L. Li, N.-L. Rev. A 78, 052325.
Liu, X. Ma, C.-Y. Lu, Y.-A. Chen, and J.-W. Pan (2017), Endo, S., S. C. Benjamin, and Y. Li (2017), arXiv preprint
Optica 4 (1), 77. arXiv:1712.09271.
Chen, M.-C., M. Gong, X.-S. Xu, X. Yuan, J.-W. Wang, Endo, S., T. Jones, S. McArdle, X. Yuan, and S. Benjamin
C. Wang, C. Ying, J. Lin, Y. Xu, Y. Wu, et al. (2019), (2018a), arXiv preprint arXiv:1806.05707.
arXiv preprint arXiv:1905.03150. Endo, S., Y. Li, S. Benjamin, and X. Yuan (2018b), arXiv
Chien, R. W., S. Xue, T. S. Hardikar, K. Setia, and J. D. preprint arXiv:1812.08778.
Whitfield (2019), Phys. Rev. A 100, 032337. Endo, S., Q. Zhao, Y. Li, S. Benjamin, and X. Yuan (2018c),
Childs, A. M., and R. Kothari (2011), in Theory of Quantum arXiv preprint arXiv:1808.03623.
Computation, Communication, and Cryptography, edited Evans, M. G., and M. Polanyi (1935), Trans. Faraday Soc.
by W. van Dam, V. M. Kendon, and S. Severini (Springer 31, 875.
Berlin Heidelberg, Berlin, Heidelberg) pp. 94–103. Eyring, H. (1935), The Journal of Chemical Physics 3 (2),
Childs, A. M., D. Maslov, Y. Nam, N. J. Ross, and Y. Su 107, https://doi.org/10.1063/1.1749604.
(2017), arXiv preprint arXiv:1711.10980. Farhi, E., J. Goldstone, and S. Gutmann (2014), arXiv
Childs, A. M., A. Ostrander, and Y. Su (2018), arXiv preprint preprint arXiv:1411.4028.
arXiv:1805.08385. Farhi, E., J. Goldstone, S. Gutmann, and M. Sipser (2000),
Childs, A. M., and Y. Su (2019), arXiv:1901.00564. arXiv:quant-ph/0001106.
Childs, A. M., and N. Wiebe (2012), Quantum Info. Comput. Farrelly, T. C., and A. J. Short (2014), Phys. Rev. A 89,
12 (11-12), 901. 012302.
Chin, S., and J. Huh (2018), arXiv preprint Feynman, R. P. (1982), International Journal of Theoretical
arXiv:1803.10002. Physics 21 (6), 467.
Chow, J. M., J. M. Gambetta, A. D. Córcoles, S. T. Merkel, Fingerhuth, M., T. Babej, and C. Ing (2018), “A quantum
J. A. Smolin, C. Rigetti, S. Poletto, G. A. Keefe, M. B. alternating operator ansatz with hard and soft constraints
Rothwell, J. R. Rozen, M. B. Ketchen, and M. Steffen for lattice protein folding,” arXiv:1810.13411.
(2012), Phys. Rev. Lett. 109, 060501. Fowler, A. G., and C. Gidney (2018), arXiv:1808.06709.
Christiansen, O. (2012), Physical Chemistry Chemical Fowler, A. G., M. Mariantoni, J. M. Martinis, and A. N.
Physics 14 (19), 6672. Cleland (2012a), Physical Review A 86 (3), 032324.
Cirac, J. I., and P. Zoller (1995), Phys. Rev. Lett. 74, 4091. Fowler, A. G., A. C. Whiteside, and L. C. L. Hollenberg
Clements, W. R., J. J. Renema, A. Eckstein, A. A. Valido, (2012b), Phys. Rev. Lett. 108, 180501.
A. Lita, T. Gerrits, S. W. Nam, W. S. Kolthammer, Fradkin, E., S. A. Kivelson, and J. M. Tranquada (2015),
J. Huh, and I. A. Walmsley (2017), arXiv preprint Rev. Mod. Phys. 87, 457.
arXiv:1710.08655. Frisch, M. J., G. W. Trucks, H. B. Schlegel, G. E. Scuseria,
Colless, J. I., V. V. Ramasesh, D. Dahlen, M. S. Blok, M. E. M. A. Robb, J. R. Cheeseman, G. Scalmani, V. Barone,
Kimchi-Schwartz, J. R. McClean, J. Carter, W. A. de Jong, G. A. Petersson, H. Nakatsuji, X. Li, M. Caricato, A. V.
and I. Siddiqi (2018), Phys. Rev. X 8, 011021. Marenich, J. Bloino, B. G. Janesko, R. Gomperts, B. Men-
Coppersmith, D. (1994), IBM Research Report RC 19642. nucci, H. P. Hratchian, J. V. Ortiz, A. F. Izmaylov, J. L.
Cotler, J., and F. Wilczek (2019), arXiv:1908.02754. Sonnenberg, D. Williams-Young, F. Ding, F. Lipparini,
Crawford, O., B. van Straaten, D. Wang, T. Parks, E. Camp- F. Egidi, J. Goings, B. Peng, A. Petrone, T. Hender-
bell, and S. Brierley (2019), arXiv:1908.06942. son, D. Ranasinghe, V. G. Zakrzewski, J. Gao, N. Rega,
Cubitt, T., and A. Montanaro (2016), SIAM Journal on Com- G. Zheng, W. Liang, M. Hada, M. Ehara, K. Toy-
puting 45 (2), 268, https://doi.org/10.1137/140998287. ota, R. Fukuda, J. Hasegawa, M. Ishida, T. Nakajima,
Dagotto, E. (1994), Rev. Mod. Phys. 66, 763. Y. Honda, O. Kitao, H. Nakai, T. Vreven, K. Throssell,
Dallaire-Demers, P.-L., J. Romero, L. Veis, S. Sim, and J. A. Montgomery, Jr., J. E. Peralta, F. Ogliaro, M. J.
 54

Bearpark, J. J. Heyd, E. N. Brothers, K. N. Kudin, Hempel, C., C. Maier, J. Romero, J. McClean, T. Monz,
V. N. Staroverov, T. A. Keith, R. Kobayashi, J. Normand, H. Shen, P. Jurcevic, B. P. Lanyon, P. Love, R. Babbush,
K. Raghavachari, A. P. Rendell, J. C. Burant, S. S. Iyengar, A. Aspuru-Guzik, R. Blatt, and C. F. Roos (2018), Phys.
J. Tomasi, M. Cossi, J. M. Millam, M. Klene, C. Adamo, Rev. X 8, 031022.
R. Cammi, J. W. Ochterski, R. L. Martin, K. Morokuma, Heya, K., K. M. Nakanishi, K. Mitarai, and K. Fujii (2019),
O. Farkas, J. B. Foresman, and D. J. Fox (2016), “Gaus- arXiv e-prints , arXiv:1904.08566arXiv:1904.08566 [quant-
sian16 Revision B.01,” Gaussian Inc. Wallingford CT. ph].
Gaebler, J. P., T. R. Tan, Y. Lin, Y. Wan, R. Bowler, A. C. Heyl, M., P. Hauke, and P. Zoller (2019), Science advances
Keith, S. Glancy, K. Coakley, E. Knill, D. Leibfried, et al. 5 (4), eaau8342.
(2016), Physical review letters 117 (6), 060505. Higgins, B. L., D. W. Berry, S. D. Bartlett, H. M. Wiseman,
Gambetta, J. M., J. M. Chow, and M. Steffen (2017), npj and G. J. Pryde (2007), Nature 450, 393 EP .
Quantum Information 3 (1), 2. Higgott, O., D. Wang, and S. Brierley (2018), arXiv preprint
Ganzhorn, M., D. J. Egger, P. K. Barkoutsos, P. Ollitrault, arXiv:1805.08138.
G. Salis, N. Moll, A. Fuhrer, P. Mueller, S. Woerner, I. Tav- Hobza, P., and J. Sponer (2002), Journal of the Ameri-
ernelli, and S. Filipp (2018), arXiv:1809.05057. can Chemical Society 124 (39), 11802, pMID: 12296748,
Garcia-Saez, A., and J. I. Latorre (2018), arXiv:1806.02287. https://doi.org/10.1021/ja026759n.
Ge, Y., J. Tura, and J. I. Cirac (2017), arXiv:1712.03193. Hoffman, B. M., D. Lukoyanov, Z.-Y. Yang, D. R. Dean, and
Genin, S. N., I. G. Ryabinkin, and A. F. Izmaylov (2019), L. C. Seefeldt (2014), Chemical Reviews 114 (8), 4041,
arXiv:1901.04715. pMID: 24467365, https://doi.org/10.1021/cr400641x.
Georgescu, I. M., S. Ashhab, and F. Nori (2014), Rev. Mod. Hoffmann, M. R., and J. Simons (1988), The
Phys. 86, 153. Journal of Chemical Physics 88 (2), 993,
Gidney, C., and M. Eker (2019), arXiv:1905.09749. https://doi.org/10.1063/1.454125.
Gidney, C., and A. G. Fowler (2019), Quantum 3, 135. Houck, A. A., H. E. Türeci, and J. Koch (2012), Nature
Gokhale, P., O. Angiuli, Y. Ding, K. Gui, T. Tomesh, Physics 8 (4), 292.
M. Suchara, M. Martonosi, and F. T. Chong (2019), Hu, L., Y.-C. Ma, Y. Xu, W.-T. Wang, Y.-W. Ma, K. Liu, H.-
arXiv:1907.13623. Y. Wang, Y.-P. Song, M.-H. Yung, and L.-Y. Sun (2018a),
Gokhale, P., and F. T. Chong (2019), arXiv:1908.11857. Sci. Bull. 63 (5), 293.
Gottesman, D. (1998), Physical Review A 57 (1), 127. Hu, L., Y.-C. Ma, Y. Xu, W.-T. Wang, Y.-W. Ma, K. Liu, H.-
Grant, E., L. Wossnig, M. Ostaszewski, and M. Benedetti Y. Wang, Y.-P. Song, M.-H. Yung, and L.-Y. Sun (2018b),
(2019), arXiv:1903.05076. Science Bulletin 63 (5), 293 .
Grimsley, H. R., D. Claudino, S. E. Economou, E. Barnes, Hubbard, J. (1963), Proceedings of the Royal
and N. J. Mayhall (2019), arXiv:1910.10329. Society of London. Series A. Mathemati-
Grimsley, H. R., S. E. Economou, E. Barnes, and N. J. May- cal and Physical Sciences 276 (1365), 238,
hall (2018), arXiv:1812.11173. https://royalsocietypublishing.org/doi/pdf/10.1098/rspa.1963.0204.
Guerreschi, G. G., and M. Smelyanskiy (2017), Huggins, W. J., J. Lee, U. Baek, B. O’Gorman, and K. B.
arXiv:1701.01450. Whaley (2019a), arXiv:1909.09114.
Gulania, S., and J. D. Whitfield (2019), arXiv:1904.10469. Huggins, W. J., J. McClean, N. Rubin, Z. Jiang, N. Wiebe,
Haah, J., M. B. Hastings, R. Kothari, and G. H. Low (2018), K. B. Whaley, and R. Babbush (2019b), arXiv:1907.13117.
arXiv preprint arXiv:1801.03922. Huh, J., G. G. Guerreschi, B. Peropadre, J. R. McClean, and
Hanson, R., L. P. Kouwenhoven, J. R. Petta, S. Tarucha, and A. Aspuru-Guzik (2015), Nature Photonics 9 (9), 615.
L. M. K. Vandersypen (2007), Rev. Mod. Phys. 79, 1217. Huh, J., and M.-H. Yung (2017), Scientific reports 7 (1),
Harrow, A., and J. Napp (2019), arXiv:1901.05374. 7462.
Harrow, A. W., and A. Montanaro (2017), Nature Huo, M., and Y. Li (2018), arXiv:1811.02734.
549 (7671), 203. IBM, (2018), “Ibm qiskit,” https://qiskit.org/.
Harty, T., D. Allcock, C. J. Ballance, L. Guidoni, H. Janacek, Izmaylov, A. F., T.-C. Yen, and I. G. Ryabinkin (2018), arXiv
N. Linke, D. Stacey, and D. Lucas (2014), Physical review preprint arXiv:1810.11602.
letters 113 (22), 220501. Jameson, A. (1999), Journal of Aircraft 36 (1), 36,
Hastings, M. B., D. Wecker, B. Bauer, and M. Troyer (2015), https://doi.org/10.2514/2.2412.
QIC 15. Jena, A., S. Genin, and M. Mosca (2019), arXiv:1907.07859.
Hatcher, R., J. A. Kittl, and C. Bowen (2019), Jiang, Z., A. Kalev, W. Mruczkiewicz, and H. Neven (2019),
arXiv:1903.05550. arXiv:1910.10746.
Hattig, C., W. Klopper, A. Khn, and D. P. Tew Jiang, Z., J. McClean, R. Babbush, and H. Neven (2018a),
(2012), Chemical Reviews 112 (1), 4, pMID: 22206503, arXiv:1812.08190.
https://doi.org/10.1021/cr200168z. Jiang, Z., K. J. Sung, K. Kechedzhi, V. N. Smelyanskiy, and
Havlicek, V., M. Troyer, and J. D. Whitfield (2017), Phys. S. Boixo (2018b), Phys. Rev. Applied 9, 044036.
Rev. A 95, 032332. Johnson, P. D., J. Romero, J. Olson, Y. Cao, and A. Aspuru-
Hehre, W. J., R. F. Stewart, and J. A. Pople Guzik (2017), arXiv preprint arXiv:1711.02249.
(1969), The Journal of Chemical Physics 51 (6), 2657, Jones, N. C., J. D. Whitfield, P. L. McMahon, M.-H. Yung,
https://doi.org/10.1063/1.1672392. R. Van Meter, A. Aspuru-Guzik, and Y. Yamamoto (2012),
Helgaker, T., S. Coriani, P. Jrgensen, K. Kristensen, J. Olsen, New Journal of Physics 14 (11), 115023.
and K. Ruud (2012), Chemical Reviews 112 (1), 543, Jordan, P., and E. Wigner (1928), Zeitschrift für Physik
pMID: 22236047, https://doi.org/10.1021/cr2002239. 47 (9), 631.
Helgaker, T., P. Jorgensen, and J. Olsen (2014), Molecular Joshi, S., A. Shukla, H. Katiyar, A. Hazra, and T. S. Mahesh
electronic-structure theory (John Wiley & Sons). (2014), Phys. Rev. A 90, 022303.
 55

Jozsa, R. (2005), arXiv:quant-ph/0508124. Lanyon, B. P., J. D. Whitfield, G. G. Gillett, M. E. Goggin,

Kais, S., S. A. Rice, and A. R. Dinner (2014), Quantum M. P. Almeida, I. Kassal, J. D. Biamonte, M. Mohseni,
information and computation for chemistry (John Wiley & B. J. Powell, M. Barbieri, et al. (2010), Nature chemistry
Sons). 2 (2), 106.
Kandala, A., A. Mezzacapo, K. Temme, M. Takita, M. Brink, LeBlanc, J. P. F., A. E. Antipov, F. Becca, I. W. Bulik, G. K.-
J. M. Chow, and J. M. Gambetta (2017), Nature L. Chan, C.-M. Chung, Y. Deng, M. Ferrero, T. M. Hen-
549 (7671), 242. derson, C. A. Jiménez-Hoyos, E. Kozik, X.-W. Liu, A. J.
Kandala, A., K. Temme, A. D. Corcoles, A. Mezzacapo, J. M. Millis, N. V. Prokof’ev, M. Qin, G. E. Scuseria, H. Shi,
Chow, and J. M. Gambetta (2018), arXiv:1805.04492. B. V. Svistunov, L. F. Tocchio, I. S. Tupitsyn, S. R. White,
Kane, B. E. (1998), Nature 393 (6681), 133. S. Zhang, B.-X. Zheng, Z. Zhu, and E. Gull (Simons Col-
Kassal, I., S. P. Jordan, P. J. Love, M. Mohseni, and laboration on the Many-Electron Problem) (2015), Phys.
A. Aspuru-Guzik (2008), Proceedings of the National Rev. X 5, 041041.
Academy of Sciences 105 (48), 18681. Lee, J., W. J. Huggins, M. Head-Gordon, and K. B. Wha-
Kassal, I., J. D. Whitfield, A. Perdomo-Ortiz, M.-H. Yung, ley (2019), Journal of Chemical Theory and Computation
and A. Aspuru-Guzik (2011), Annual review of physical 15 (1), 311, https://doi.org/10.1021/acs.jctc.8b01004.
chemistry 62, 185. Lee, P. A., N. Nagaosa, and X.-G. Wen (2006), Rev. Mod.
Keen, T., T. Maier, S. Johnston, and P. Lougovski (2019), Phys. 78, 17.
arXiv:1910.09512. Leibfried, D., R. Blatt, C. Monroe, and D. Wineland (2003),
Kempe, J., A. Kitaev, and O. Regev (2006), Rev. Mod. Phys. 75, 281.
SIAM Journal on Computing 35 (5), 1070, Li, Y., and S. C. Benjamin (2017), Phys. Rev. X 7, 021050.
https://doi.org/10.1137/S0097539704445226. Li, Z., J. Li, N. S. Dattani, C. J. Umrigar, and G. K.-L. Chan
Kieferova, M., A. Scherer, and D. Berry (2018), arXiv (2018), arXiv:1809.10307.
preprint arXiv:1805.00582. Li, Z., M.-H. Yung, H. Chen, D. Lu, J. D. Whitfield, X. Peng,
Kim, I. H. (2017), arXiv:1703.00032. A. Aspuru-Guzik, and J. Du (2011), Scientific reports 1,
Kim, I. H., and B. Swingle (2017), arXiv:1711.07500. 88.
Kitaev, A. Y. (1995), Preprint at http://arxiv. Lidar, D. A., and T. A. Brun (2013), Quantum Error Cor-
org/abs/quant-ph/9511026. rection (Cambridge University Press).
Kitaev, A. Y. (1997), Russian Mathematical Surveys 52 (6), Lidar, D. A., and H. Wang (1999), Phys. Rev. E 59, 2429.
1191. Lischka, H., D. Nachtigallov, A. J. A. Aquino, P. G. Sza-
Kivlichan, I. D., C. Gidney, D. W. Berry, N. Wiebe, lay, F. Plasser, F. B. C. Machado, and M. Barbatti
J. McClean, W. Sun, Z. Jiang, N. Rubin, A. Fowler, (2018), Chemical Reviews 118 (15), 7293, pMID: 30040389,
A. Aspuru-Guzik, R. Babbush, and H. Neven (2019a), https://doi.org/10.1021/acs.chemrev.8b00244.
arXiv:1902.10673. Litinski, D. (2019a), Quantum 3, 128.
Kivlichan, I. D., C. E. Granade, and N. Wiebe (2019b), Litinski, D. (2019b), arXiv:1905.06903.
arXiv:1907.10070. Lloyd, S. (1996), Science 273 (5278), 1073.
Kivlichan, I. D., J. McClean, N. Wiebe, C. Gidney, Lloyd, S., and S. L. Braunstein (1999), Phys. Rev. Lett. 82,
A. Aspuru-Guzik, G. K.-L. Chan, and R. Babbush (2018), 1784.
Phys. Rev. Lett. 120, 110501. Loss, D., and D. P. DiVincenzo (1998), Phys. Rev. A 57,
Kivlichan, I. D., N. Wiebe, R. Babbush, and A. Aspuru- 120.
Guzik (2017), Journal of Physics A: Mathematical and The- Love, P. J. (2012), arXiv:1208.5524.
oretical 50 (30). Low, G. H. (2018), arXiv preprint arXiv:1807.03967.
Knecht, S., E. D. Hedegaard, S. Keller, A. Kovyrshin, Y. Ma, Low, G. H., N. P. Bauman, C. E. Granade, B. Peng, N. Wiebe,
A. Muolo, C. J. Stein, and M. Reiher (2016), CHIMIA E. J. Bylaska, D. Wecker, S. Krishnamoorthy, M. Roet-
International Journal for Chemistry 70 (4). teler, K. Kowalski, M. Troyer, and N. A. Baker (2019a),
Knill, E., R. Laflamme, and G. J. Milburn (2001), Nature arXiv:1904.01131.
409 (6816), 46. Low, G. H., and I. L. Chuang (2016), arXiv preprint
Knill, E., R. Laflamme, and W. Zurek (1996), . arXiv:1610.06546.
Knill, E., G. Ortiz, and R. D. Somma (2007), Phys. Rev. A Low, G. H., and I. L. Chuang (2017), Phys. Rev. Lett. 118,
75, 012328. 010501.
Kokail, C., C. Maier, R. van Bijnen, T. Brydges, M. K. Joshi, Low, G. H., V. Kliuchnikov, and N. Wiebe (2019b),
P. Jurcevic, C. A. Muschik, P. Silvi, R. Blatt, C. F. Roos, arXiv:1907.11679.
and P. Zoller (2019), Nature 569 (7756), 355. Low, G. H., and N. Wiebe (2018), arXiv:1805.00675.
Kolda, T. G., R. M. Lewis, and V. Torczon (2006), SIAM Low, G. H., T. J. Yoder, and I. L. Chuang (2016), Phys. Rev.
Rev. 45, 385482. X 6, 041067.
Kreula, J., S. R. Clark, and D. Jaksch (2016), Scientific re- Lu, D., B. Xu, N. Xu, Z. Li, H. Chen, X. Peng, R. Xu, and
ports 6, 32940. J. Du (2012), Phys. Chem. Chem. Phys. 14, 9411.
Kurashige, Y., G. K.-L. Chan, and T. Yanai (2013), Nature Lu, L.-H., and Y.-Q. Li (2019), arXiv:1906.09184.
chemistry 5 (8), 660. Lyakh, D. I., M. Musia, V. F. Lotrich, and R. J. Bartlett
Ladd, T. D., F. Jelezko, R. Laflamme, Y. Nakamura, C. Mon- (2012), Chemical Reviews 112 (1), 182, pMID: 22220988,
roe, and J. L. O’Brien (2010), Nature 464, 45 EP , review https://doi.org/10.1021/cr2001417.
Article. Macridin, A., P. Spentzouris, J. Amundson, and R. Harnik
Landauer, R. (1995), Philosophical Transactions of the Royal (2018a), Phys. Rev. A 98, 042312.
Society of London A: Mathematical, Physical and Engi- Macridin, A., P. Spentzouris, J. Amundson, and R. Harnik
neering Sciences 353 (1703), 367. (2018b), Phys. Rev. Lett. 121, 110504.
 56

Malik, M. R., and D. M. Bushnell (2012), NASA Technical M. Keesan, S. M. Kreikemeier, J. Mizrahi, P. Solomon,
Report . M. Williams, J. D. Wong-Campos, C. Monroe, and J. Kim
Matsuura, S., T. Yamazaki, V. Senicourt, L. Huntington, and (2019), arXiv:1902.10171.
A. Zaribafiyan (2018), arXiv:1810.11511. Nielsen, M. A., and I. Chuang (2002), “Quantum computa-
Matsuzawa, Y., and Y. Kurashige (2019), arXiv:1909.12410. tion and quantum information,”.
McArdle, S., S. Endo, Y. Li, S. Benjamin, and X. Yuan O’Brien, T. E., B. Senjean, R. Sagastizabal, X. Bonet-
(2018a), arXiv preprint arXiv:1804.03023. Monroig, A. Dutkiewicz, F. Buda, L. DiCarlo, and L. Viss-
McArdle, S., A. Mayorov, X. Shan, S. Benjamin, and X. Yuan cher (2019), arXiv:1905.03742.
(2018b), arXiv:1811.04069. O’Brien, T. E., B. Tarasinski, and B. M. Terhal (2018a),
McArdle, S., X. Yuan, and S. Benjamin (2018c), arXiv:1809.09697.
arXiv:1807.02467. O’Brien, T. E., P. Rożek, and A. R. Akhmerov (2018b), Phys.
McCaskey, A. J., Z. P. Parks, J. Jakowski, S. V. Moore, Rev. Lett. 120, 220504.
T. Morris, T. S. Humble, and R. C. Pooser (2019), O’Gorman, B., W. J. Huggins, E. G. Rieffel, and K. B. Wha-
arXiv:1905.01534. ley (2019), arXiv:1905.05118.
McClean, J., S. Boixo, V. Smelyanskiy, R. Babbush, and O’Gorman, J., and E. T. Campbell (2017), Phys. Rev. A 95,
H. Neven (2018), Nature Communications 9, 4812. 032338.
McClean, J. R., R. Babbush, P. J. Love, and Olivares-Amaya, R., W. Hu, N. Nakatani, S. Sharma, J. Yang,
A. Aspuru-Guzik (2014), The Journal of Physical and G. K.-L. Chan (2015), The Journal of Chemical Physics
Chemistry Letters 5 (24), 4368, pMID: 26273989, 142 (3), 034102, https://doi.org/10.1063/1.4905329.
https://doi.org/10.1021/jz501649m. Ollitrault, P. J., A. Kandala, C.-F. Chen, P. K. Barkoutsos,
McClean, J. R., F. M. Faulstich, Q. Zhu, B. O’Gorman, A. Mezzacapo, M. Pistoia, S. Sheldon, S. Woerner, J. Gam-
Y. Qiu, S. R. White, R. Babbush, and L. Lin (2019a), betta, and I. Tavernelli (2019), arXiv:1910.12890.
arXiv:1909.00028. O’Malley, P. J. J., R. Babbush, I. D. Kivlichan, J. Romero,
McClean, J. R., Z. Jiang, N. C. Rubin, R. Babbush, and J. R. McClean, R. Barends, J. Kelly, P. Roushan, A. Tran-
H. Neven (2019b), arXiv preprint arXiv:1903.05786. ter, N. Ding, B. Campbell, Y. Chen, Z. Chen, B. Chiaro,
McClean, J. R., M. E. Kimchi-Schwartz, J. Carter, and W. A. A. Dunsworth, A. G. Fowler, E. Jeffrey, E. Lucero,
de Jong (2017a), Phys. Rev. A 95, 042308. A. Megrant, J. Y. Mutus, M. Neeley, C. Neill, C. Quin-
McClean, J. R., I. D. Kivlichan, K. J. Sung, D. S. Steiger, tana, D. Sank, A. Vainsencher, J. Wenner, T. C. White,
Y. Cao, C. Dai, E. S. Fried, C. Gidney, B. Gimby, P. V. Coveney, P. J. Love, H. Neven, A. Aspuru-Guzik,
P. Gokhale, T. Hner, T. Hardikar, V. Havlek, C. Huang, and J. M. Martinis (2016), Phys. Rev. X 6, 031007.
J. Izaac, Z. Jiang, X. Liu, M. Neeley, T. O’Brien, I. Ozfi- Ortiz, G., J. E. Gubernatis, E. Knill, and R. Laflamme
dan, M. D. Radin, J. Romero, N. Rubin, N. P. D. Sawaya, (2001), Phys. Rev. A 64, 022319.
K. Setia, S. Sim, M. Steudtner, Q. Sun, W. Sun, F. Zhang, Ostaszewski, M., E. Grant, and M. Benedetti (2019), arXiv
and R. Babbush (2017b), arXiv:1710.07629. e-prints , arXiv:1905.09692arXiv:1905.09692 [quant-ph].
McClean, J. R., J. Romero, R. Babbush, and A. Aspuru- Otten, M., and S. Gray (2018a), arXiv:1804.06969.
Guzik (2016), New Journal of Physics 18 (2), 023023. Otten, M., and S. Gray (2018b), arXiv:1806.07860.
Mitarai, K., and K. Fujii (2018), arXiv:1901.00015. Ouyang, Y., D. R. White, and E. Campbell (2019),
Mitarai, K., Y. O. Nakagawa, and W. Mizukami (2019), arXiv:1910.06255.
arXiv:1905.04054. Paesani, S., A. A. Gentile, R. Santagati, J. Wang, N. Wiebe,
Mizukami, W., K. Mitarai, Y. O. Nakagawa, T. Yamamoto, D. P. Tew, J. L. O’Brien, and M. G. Thompson (2017),
T. Yan, and Y. ya Ohnish (2019), arXiv:1910.11526. Phys. Rev. Lett. 118, 100503.
Monroe, C., and J. Kim (2013), Science 339 (6124), 1164, Paini, M., and A. Kalev (2019), arXiv:1910.10543.
https://science.sciencemag.org/content/339/6124/1164.full.pdf.Parrish, R. M., L. A. Burns, D. G. A. Smith, A. C. Sim-
Monz, T., P. Schindler, J. T. Barreiro, M. Chwalla, D. Nigg, monett, A. E. DePrince, E. G. Hohenstein, U. Bozkaya,
W. A. Coish, M. Harlander, W. Hänsel, M. Hennrich, and A. Y. Sokolov, R. Di Remigio, R. M. Richard, J. F.
R. Blatt (2011), Phys. Rev. Lett. 106, 130506. Gonthier, A. M. James, H. R. McAlexander, A. Ku-
Motta, M., C. Sun, A. T. K. Tan, M. J. O. Rourke, E. Ye, mar, M. Saitow, X. Wang, B. P. Pritchard, P. Verma,
A. J. Minnich, F. G. S. L. Brandao, and G. K.-L. Chan H. F. Schaefer, K. Patkowski, R. A. King, E. F.
(2019), arXiv:1901.07653. Valeev, F. A. Evangelista, J. M. Turney, T. D. Craw-
Motta, M., E. Ye, J. R. McClean, Z. Li, A. J. Minnich, R. Bab- ford, and C. D. Sherrill (2017), Journal of Chemical
bush, and G. K.-L. Chan (2018), arXiv:1808.02625. Theory and Computation 13 (7), 3185, pMID: 28489372,
Mueck, L. (2015), Nature chemistry 7 (5), 361. https://doi.org/10.1021/acs.jctc.7b00174.
Muller, R. P. (2004), “Pyquante,” http://pyquante. Parrish, R. M., E. G. Hohenstein, P. L. McMahon, and T. J.
sourceforge.net/, accessed: 24.7.2018. Martinez (2019a), arXiv:1906.08728.
Nakamura, Y., Y. A. Pashkin, and J. S. Tsai (1999), Nature Parrish, R. M., E. G. Hohenstein, P. L. McMahon, and T. J.
398 (6730), 786. Martinez (2019b), arXiv preprint arXiv:1901.01234.
Nakanishi, K. M., K. Fujii, and S. Todo (2019), arXiv Parrish, R. M., J. T. Iosue, A. Ozaeta, and P. L. McMahon
preprint arXiv:1903.12166. (2019c), arXiv preprint arXiv:1904.03206.
Nakanishi, K. M., K. Mitarai, and K. Fujii (2018), arXiv Perdomo, A., C. Truncik, I. Tubert-Brohman, G. Rose, and
preprint arXiv:1810.09434. A. Aspuru-Guzik (2008), Phys. Rev. A 78, 012320.
Nam, Y., J.-S. Chen, N. C. Pisenti, K. Wright, C. Delaney, Perdomo-Ortiz, A., N. Dickson, M. Drew-Brook, G. Rose,
D. Maslov, K. R. Brown, S. Allen, J. M. Amini, J. Apis- and A. Aspuru-Guzik (2012), Scientific reports 2, 571.
dorf, K. M. Beck, A. Blinov, V. Chaplin, M. Chmielewski, Peruzzo, A., J. McClean, P. Shadbolt, M.-H. Yung, X.-Q.
C. Collins, S. Debnath, A. M. Ducore, K. M. Hudek, Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. Obrien
 57

(2014), Nature communications 5. Sanders, Y. R., G. H. Low, A. Scherer, and D. W. Berry

Podewitz, M., M. T. Stiebritz, and M. Reiher (2011), Faraday (2019), Phys. Rev. Lett. 122, 020502.
Discuss. 148, 119. Santagati, R., J. Wang, A. A. Gentile, S. Paesani, N. Wiebe,
Pople, J. A., M. HeadGordon, D. J. Fox, K. Raghavachari, J. R. McClean, S. Morley-Short, P. J. Shadbolt, D. Bon-
and L. A. Curtiss (1989), The Journal of Chemical Physics neau, J. W. Silverstone, D. P. Tew, X. Zhou, J. L. O’Brien,
90 (10), 5622, https://doi.org/10.1063/1.456415. and M. G. Thompson (2018), Science Advances 4 (1),
Poulin, D., M. B. Hastings, D. Wecker, N. Wiebe, A. C. Do- 10.1126/sciadv.aap9646.
herty, and M. Troyer (2015), QIC 15. Sawaya, N. P. D., and J. Huh (2018), arXiv:1812.10495.
Poulin, D., A. Kitaev, D. S. Steiger, M. B. Hastings, and Sawaya, N. P. D., T. Menke, T. H. Kyaw, S. Johri, A. Aspuru-
M. Troyer (2018), Phys. Rev. Lett. 121, 010501. Guzik, and G. G. Guerreschi (2019), arXiv:1909.12847.
Poulin, D., A. Qarry, R. Somma, and F. Verstraete (2011), Sawaya, N. P. D., M. Smelyanskiy, J. R. McClean,
Phys. Rev. Lett. 106, 170501. and A. Aspuru-Guzik (2016), Journal of Chemical The-
Preskill, J. (2018), arXiv preprint arXiv:1801.00862. ory and Computation 12 (7), 3097, pMID: 27254482,
Purvis, G. D., and R. J. Bartlett (1982), The https://doi.org/10.1021/acs.jctc.6b00220.
Journal of Chemical Physics 76 (4), 1910, Schneider, C., D. Porras, and T. Schaetz (2012), Reports on
https://doi.org/10.1063/1.443164. Progress in Physics 75 (2), 024401.
Raussendorf, R. (2012), Philosophical Transactions of Schuld, M., V. Bergholm, C. Gogolin, J. Izaac, and N. Kil-
the Royal Society of London A: Mathematical, Phys- loran (2019), Phys. Rev. A 99, 032331.
ical and Engineering Sciences 370 (1975), 4541, Schutz, M. (2000), The Journal of Chemical Physics 113 (22),
http://rsta.royalsocietypublishing.org/content/370/1975/4541.full.pdf.
9986, https://doi.org/10.1063/1.1323265.
Raussendorf, R., and H. J. Briegel (2001), Phys. Rev. Lett. Schutz, M., and H.-J. Werner (2000), Chemical Physics Let-
86, 5188. ters 318 (4), 370 .
Raussendorf, R., D. E. Browne, and H. J. Briegel (2003), Seeley, J. T., M. J. Richard, and P. J. Love (2012), The
Phys. Rev. A 68, 022312. Journal of Chemical Physics 137 (22), 224109.
Reiher, M., N. Wiebe, K. M. Svore, D. Wecker, and M. Troyer Senjean, B. (2019), “Openfermion-dirac,”
(2017), Proceedings of the National Academy of Sciences . https://github.com/bsenjean/Openfermion-Dirac.
Reiner, J.-M., F. Wilhelm-Mauch, G. Schn, and M. Marthaler Setia, K., S. Bravyi, A. Mezzacapo, and J. D. Whitfield
(2018a), arXiv:1811.04476. (2018), arXiv:1810.05274.
Reiner, J.-M., S. Zanker, I. Schwenk, J. Leppkangas, Setia, K., R. Chen, J. E. Rice, A. Mezzacapo, M. Pistoia, and
F. Wilhelm-Mauch, G. Schn, and M. Marthaler (2018b), J. Whitfield (2019), arXiv:1910.14644.
Quantum Science and Technology 3 (4), 045008. Setia, K., and J. D. Whitfield (2018), The Journal of Chem-
Richardson, L. F., B. J Arthur Gaunt, et al. (1927), Phil. ical Physics 148 (16), 164104.
Trans. R. Soc. Lond. A 226 (636-646), 299. Sharma, S., K. Sivalingam, F. Neese, and G. K.-L. Chan
Robert, A., P. K. Barkoutsos, S. Woerner, and I. Tavernelli (2014), Nature chemistry 6 (10), 927.
(2019), arXiv:1908.02163. Shehab, O., K. A. Landsman, Y. Nam, D. Zhu, N. M. Linke,
Romero, J., R. Babbush, J. R. McClean, C. Hempel, P. J. M. J. Keesan, R. C. Pooser, and C. R. Monroe (2019),
Love, and A. Aspuru-Guzik (2019), Quantum Science and arXiv:1904.04338.
Technology 4 (1), 014008. Shen, Y., Y. Lu, K. Zhang, J. Zhang, S. Zhang, J. Huh, and
Roos, B. O., P. R. Taylor, and P. E. Sigbahn (1980), Chemical K. Kim (2018), Chem. Sci. 9, 836.
Physics 48 (2), 157 . Shen, Y., X. Zhang, S. Zhang, J.-N. Zhang, M.-H. Yung, and
Rossi, E., G. L. Bendazzoli, S. Evangelisti, and D. Maynau K. Kim (2017), Phys. Rev. A 95, 020501.
(1999), Chemical Physics Letters 310 (5), 530 . Shnirman, A., G. Schön, and Z. Hermon (1997), Phys. Rev.
Rowe, D. J. (1968), Rev. Mod. Phys. 40, 153. Lett. 79, 2371.
Rubin, N. C. (2016), arXiv:1610.06910. Shor, P. W. (1994), in Foundations of Computer Science, 1994
Rubin, N. C., R. Babbush, and J. McClean (2018), New Proceedings., 35th Annual Symposium on (Ieee) pp. 124–
Journal of Physics 20 (5), 053020. 134.
Rungger, I., N. Fitzpatrick, H. Chen, C. H. Alderete, H. Apel, Shor, P. W. (1996), in Foundations of Computer Science,
A. Cowtan, A. Patterson, D. M. Ramo, Y. Zhu, N. H. 1996. Proceedings., 37th Annual Symposium on (IEEE) pp.
Nguyen, E. Grant, S. Chretien, L. Wossnig, N. M. Linke, 56–65.
and R. Duncan (2019), arXiv:1910.04735. Smirnov, A. Y., S. Savel’ev, L. G. Mourokh, and F. Nori
Ryabinkin, I. G., and S. N. Genin (2018), arXiv:1812.09812. (2007), EPL (Europhysics Letters) 80 (6), 67008.
Ryabinkin, I. G., and S. N. Genin (2019), arXiv:1906.11192. Sokolov, I., P. K. Barkoutsos, P. J. Ollitrault, D. Green-
Ryabinkin, I. G., S. N. Genin, and A. F. Iz- berg, J. Rice, M. Pistoia, and I. Tavernelli (2019),
maylov (2019), Journal of Chemical Theory arXiv:1911.10864.
and Computation 15 (1), 249, pMID: 30512959, Somma, R., G. Ortiz, J. E. Gubernatis, E. Knill, and
https://doi.org/10.1021/acs.jctc.8b00943. R. Laflamme (2002), Phys. Rev. A 65, 042323.
Ryabinkin, I. G., T.-C. Yen, S. N. Genin, and Somma, R. D. (2019), arXiv:1907.11748.
A. F. Izmaylov (2018), Journal of Chemical The- Song, C., J. Cui, H. Wang, J. Hao, H. Feng, and Y. Li (2018),
ory and Computation 14 (12), 6317, pMID: 30427679, arXiv:1812.10903.
https://doi.org/10.1021/acs.jctc.8b00932. Song, C., K. Xu, H. Li, Y. Zhang, X. Zhang, W. Liu, Q. Guo,
Sagastizabal, R., X. Bonet-Monroig, M. Singh, M. A. Rol, Z. Wang, W. Ren, J. Hao, et al. (2019), arXiv preprint
C. C. Bultink, X. Fu, C. H. Price, V. P. Ostroukh, arXiv:1905.00320.
N. Muthusubramanian, A. Bruno, M. Beekman, N. Haider, Song, C., K. Xu, W. Liu, C.-p. Yang, S.-B. Zheng, H. Deng,
T. E. O’Brien, and L. DiCarlo (2019), arXiv:1902.11258. Q. Xie, K. Huang, Q. Guo, L. Zhang, P. Zhang, D. Xu,
 58

D. Zheng, X. Zhu, H. Wang, Y.-A. Chen, C.-Y. Lu, S. Han, https://doi.org/10.1021/acs.jctc.8b00450.

and J.-W. Pan (2017), Phys. Rev. Lett. 119, 180511. Tranter, A., S. Sofia, J. Seeley, M. Kaicher, J. McClean,
Sparrow, C., E. Martı́n-López, N. Maraviglia, A. Neville, R. Babbush, P. V. Coveney, F. Mintert, F. Wilhelm, and
C. Harrold, J. Carolan, Y. N. Joglekar, T. Hashimoto, P. J. Love (2015), International Journal of Quantum Chem-
N. Matsuda, J. L. OBrien, et al. (2018), Nature 557 (7707), istry 115 (19), 1431.
660. Trotter, H. F. (1959), Proceedings of the American Mathe-
Stair, N. H., R. Huang, and F. A. Evangelista (2019), matical Society 10 (4), 545.
arXiv:1911.05163. Trout, C. J., and K. R. Brown (2015), International
Stephens, A. M. (2014), Phys. Rev. A 89, 022321. Journal of Quantum Chemistry 115 (19), 1296,
Steudtner, M., and S. Wehner (2018), arXiv:1810.02681. https://onlinelibrary.wiley.com/doi/pdf/10.1002/qua.24856.
Stokes, J., J. Izaac, N. Killoran, and G. Carleo (2019), Tubman, N. M., J. Lee, T. Y. Takeshita, M. Head-Gordon,
arXiv:1909.02108. and K. B. Whaley (2016), The Journal of Chemical Physics
Subramanian, S., S. Brierley, and R. Jozsa (2018), arXiv 145 (4), 044112, https://doi.org/10.1063/1.4955109.
preprint arXiv:1806.06885. Tubman, N. M., C. Mejuto-Zaera, J. M. Epstein, D. Hait,
Sugisaki, K., S. Nakazawa, K. Toyota, K. Sato, D. Shiomi, D. S. Levine, W. Huggins, Z. Jiang, J. R. McClean,
and T. Takui (2019), ACS Central Science 5 (1), 167, R. Babbush, M. Head-Gordon, and K. B. Whaley (2018),
https://doi.org/10.1021/acscentsci.8b00788. arXiv:1809.05523.
Sugisaki, K., S. Yamamoto, S. Nakazawa, K. Toyota, Unruh, W. G. (1995), Phys. Rev. A 51, 992.
K. Sato, D. Shiomi, and T. Takui (2016), The Journal Urbanek, M., B. Nachman, and W. A. de Jong (2019),
of Physical Chemistry A 120 (32), 6459, pMID: 27499026, arXiv:1910.00129.
https://doi.org/10.1021/acs.jpca.6b04932. Van Voorhis, T., and M. Head-Gordon (2000),
Sugisaki, K., S. Yamamoto, S. Nakazawa, K. Toyota, K. Sato, The Journal of Chemical Physics 113 (20), 8873,
D. Shiomi, and T. Takui (2018), Chemical Physics Letters: https://doi.org/10.1063/1.1319643.
X , 100002. Veis, L., and J. Pittner (2012), arXiv:1203.6204.
Sun, Q., B. T. C., B. N. S., B. G. H., G. Sheng, L. Zhendong, Veis, L., and J. Pittner (2014), The Jour-
L. Junzi, M. J. D., S. E. R., S. Sandeep, W. Sebastian, nal of Chemical Physics 140 (21), 214111,
and C. G. Kin-Lic (2017), Wiley Interdisciplinary Re- https://doi.org/10.1063/1.4880755.
views: Computational Molecular Science 8 (1), e1340, Veis, L., J. Višňák, T. Fleig, S. Knecht, T. Saue, L. Visscher,
https://onlinelibrary.wiley.com/doi/pdf/10.1002/wcms.1340. and J. c. v. Pittner (2012), Phys. Rev. A 85, 030304.
Sung, K. (2018), “https://github.com/quantumlib/ Verdon, G., M. Broughton, J. R. McClean, K. J. Sung,
OpenFermion/issues/259,” Accessed: 24.7.2018. R. Babbush, Z. Jiang, H. Neven, and M. Mohseni (2019),
Suzuki, M. (1976), Communications in Mathematical Physics arXiv:1907.05415.
51 (2), 183. Verstraete, F., and J. I. Cirac (2005), Journal of Statistical
Svore, K. M., M. B. Hastings, and M. Freedman (2013), Mechanics: Theory and Experiment 2005.
arXiv:1304.0741. Verteletskyi, V., T.-C. Yen, and A. F. Izmaylov (2019),
Sweke, R., F. Wilde, J. Meyer, M. Schuld, P. K. arXiv:1907.03358.
Fhrmann, B. Meynard-Piganeau, and J. Eisert (2019), Vogiatzis, K. D., M. V. Polynski, J. K. Kirkland,
arXiv:1910.01155. J. Townsend, A. Hashemi, C. Liu, and E. A.
Szabo, A., and N. S. Ostlund (2012), Modern quantum chem- Pidko (2019), Chemical Reviews 119 (4), 2453,
istry: introduction to advanced electronic structure theory https://doi.org/10.1021/acs.chemrev.8b00361.
(Courier Corporation). Wang, C. S., J. C. Curtis, B. J. Lester, Y. Zhang, Y. Y.
Szalay, P. G., T. Mller, G. Gidofalvi, H. Lischka, and Gao, J. Freeze, V. S. Batista, P. H. Vaccaro, I. L. Chuang,
R. Shepard (2012), Chemical Reviews 112 (1), 108, pMID: L. Frunzio, L. Jiang, S. M. Girvin, and R. J. Schoelkopf
22204633, https://doi.org/10.1021/cr200137a. (2019), arXiv:1908.03598.
Szalay, S., M. Pfeffer, V. Murg, G. Barcza, F. Ver- Wang, D., O. Higgott, and S. Brierley (2018),
straete, R. Schneider, and . Legeza (2015), Interna- arXiv:1802.00171.
tional Journal of Quantum Chemistry 115 (19), 1342, Wang, D. S., A. G. Fowler, and L. C. L. Hollenberg (2011),
https://onlinelibrary.wiley.com/doi/pdf/10.1002/qua.24898. Phys. Rev. A 83, 020302.
Takeshita, T., N. C. Rubin, Z. Jiang, E. Lee, R. Babbush, Wang, H., S. Ashhab, and F. Nori (2009), Phys. Rev. A 79,
and J. R. McClean (2019), arXiv:1902.10679. 042335.
Tang, H. L., E. Barnes, H. R. Grimsley, N. J. Mayhall, and Wang, H., S. Kais, A. Aspuru-Guzik, and M. R. Hoffmann
S. E. Economou (2019), arXiv:1911.10205. (2008), Phys. Chem. Chem. Phys. 10, 5388.
Temme, K., S. Bravyi, and J. M. Gambetta (2017), Phys. Wang, H., L.-A. Wu, Y.-x. Liu, and F. Nori (2010), Phys.
Rev. Lett. 119, 180509. Rev. A 82, 062303.
Teplukhin, A., B. K. Kendrick, and D. Babikov (2018), Wang, X.-L., L.-K. Chen, W. Li, H.-L. Huang, C. Liu,
arXiv:1812.05211. C. Chen, Y.-H. Luo, Z.-E. Su, D. Wu, Z.-D. Li, H. Lu,
Terhal, B. M. (2015), Rev. Mod. Phys. 87, 307. Y. Hu, X. Jiang, C.-Z. Peng, L. Li, N.-L. Liu, Y.-A. Chen,
Toloui, B., and P. J. Love (2013), arXiv:1312.2579. C.-Y. Lu, and J.-W. Pan (2016), Phys. Rev. Lett. 117,
Torrontegui, E., A. Ruschhaupt, D. Gury-Odelin, and J. G. 210502.
Muga (2011), Journal of Physics B: Atomic, Molecular and Wang, Y., F. Dolde, J. Biamonte, R. Babbush, V. Bergholm,
Optical Physics 44 (19), 195302. S. Yang, I. Jakobi, P. Neumann, A. Aspuru-Guzik, J. D.
Tranter, A., P. J. Love, F. Mintert, and P. V. Whitfield, et al. (2015), ACS nano 9 (8), 7769.
Coveney (2018), Journal of Chemical Theory and Ward, N. J., I. Kassal, and A. Aspuru-Guzik (2009), The
Computation 14 (11), 5617, pMID: 30189144, Journal of Chemical Physics 130 (19), 194105.
 59

Watson, M. A., and G. K.-L. Chan (2012), Journal of Zhang, S., Y. Lu, K. Zhang, W. Chen, Y. Li, J.-N. Zhang,
Chemical Theory and Computation 8 (11), 4013, pMID: and K. Kim (2019), arXiv:1905.10135.
26605568, https://doi.org/10.1021/ct300591z. Zhao, A., A. Tranter, W. M. Kirby, S. F. Ung, A. Miyake,
Wecker, D., B. Bauer, B. K. Clark, M. B. Hastings, and and P. Love (2019), arXiv:1908.08067.
M. Troyer (2014), Phys. Rev. A 90, 022305. Zhong, H.-S., Y. Li, W. Li, L.-C. Peng, Z.-E. Su, Y. Hu, Y.-M.
Wecker, D., M. B. Hastings, and M. Troyer (2015a), Phys. He, X. Ding, W. Zhang, H. Li, L. Zhang, Z. Wang, L. You,
Rev. A 92, 042303. X.-L. Wang, X. Jiang, L. Li, Y.-A. Chen, N.-L. Liu, C.-Y.
Wecker, D., M. B. Hastings, N. Wiebe, B. K. Clark, C. Nayak, Lu, and J.-W. Pan (2018), Phys. Rev. Lett. 121, 250505.
and M. Troyer (2015b), Phys. Rev. A 92, 062318.
Wendin, G. (2017), Reports on Progress in Physics 80 (10),
106001.
Werner, H.-J., and M. Schutz (2011), The Jour-
nal of Chemical Physics 135 (14), 144116,
https://doi.org/10.1063/1.3641642.
White, S. R. (2017), The Journal of Chemical Physics
147 (24), 244102.
Whitfield, J., V. Havlicek, and M. Troyer (2016), Physical
Review A 94 (3).
Whitfield, J. D. (2013), The Journal of Chemical Physics
139 (2), 021105, https://doi.org/10.1063/1.4812566.
Whitfield, J. D. (2015), arXiv preprint arXiv:1502.03771.
Whitfield, J. D., J. Biamonte, and A. Aspuru-Guzik (2011),
Molecular Physics 109 (5), 735.
Whitfield, J. D., P. J. Love, and A. Aspuru-Guzik (2013),
Phys. Chem. Chem. Phys. 15, 397.
Whitfield, J. D., M.-H. Yung, D. G. Tempel, S. Boixo, and
A. Aspuru-Guzik (2014), New Journal of Physics 16 (8),
083035.
Wiebe, N., D. W. Berry, P. Høyer, and B. C. Sanders
(2011), Journal of Physics A: Mathematical and Theoreti-
cal 44 (44), 445308.
Wiebe, N., and C. Granade (2016), Phys. Rev. Lett. 117,
010503.
Wiesner, S. (1996), arXiv preprint quant-ph/9603028.
Wilson, M., S. Stromswold, F. Wudarski, S. Hadfield, N. M.
Tubman, and E. Rieffel (2019), arXiv:1908.03185.
Wu, L.-A., M. S. Byrd, and D. A. Lidar (2002), Phys. Rev.
Lett. 89, 057904.
Xia, R., T. Bian, and S. Kais (2018), The Journal of Physical
Chemistry B 122 (13), 3384.
Xia, R., and S. Kais (2018), Nature Communications 9 (1),
4195.
Yamada, S., T. Imamura, and M. Machida (2005), in SC
’05: Proceedings of the 2005 ACM/IEEE Conference on
Supercomputing, pp. 44–44.
Yamamoto, N. (2019), arXiv:1909.05074.
Yamazaki, T., S. Matsuura, A. Narimani, A. Saidmuradov,
and A. Zaribafiyan (2018), arXiv:1806.01305.
Yanai, T., Y. Kurashige, W. Mizukami, J. Chalupsk,
T. N. Lan, and M. Saitow (2015), International
Journal of Quantum Chemistry 115 (5), 283,
https://onlinelibrary.wiley.com/doi/pdf/10.1002/qua.24808.
Yen, T.-C., R. A. Lang, and A. F. Izmaylov (2019a),
arXiv:1905.08109.
Yen, T.-C., V. Verteletskyi, and A. F. Izmaylov (2019b),
arXiv:1907.09386.
Yung, M.-H., J. Casanova, A. Mezzacapo, J. McClean,
L. Lamata, A. Aspuru-Guzik, and E. Solano (2014), Sci-
entific reports 4, 3589.
Yung, M.-H., J. D. Whitfield, S. Boixo, D. G. Tempel, and
A. Aspuru-Guzik (2012), arXiv:1203.1331.
Zalka, C. (1998), Proceedings of the Royal Society of Lon-
don A: Mathematical, Physical and Engineering Sciences,
454 (1969), 313.