TYPE Original Research

PUBLISHED 07 February 2024

DOI 10.3389/frqst.2024.1282730

Empowering complex-valued
OPEN ACCESS data classification with the
EDITED BY
Jeffrey Larson,
Argonne National Laboratory (DOE),
variational quantum classifier
United States

REVIEWED BY Jianing Chen and Yan Li*

Zixuan Hu,
Purdue University, United States Department of Electrical Engineering, The Pennsylvania State University, University Park, PA,
Kathleen Hamilton, United States
Oak Ridge National Laboratory (DOE),
United States

*CORRESPONDENCE
Yan Li, The evolution of quantum computers has encouraged research into how to
[email protected]
handle tasks with significant computation demands in the past few years. Due to
RECEIVED 24 August 2023 the unique advantages of quantum parallelism and entanglement, various types
ACCEPTED 17 January 2024
PUBLISHED 07 February 2024
of quantum machine learning (QML) methods, especially variational quantum
classifiers (VQCs), have attracted the attention of many researchers and have
CITATION
Chen J and Li Y (2024), Empowering complex- been developed and evaluated in numerous scenarios. Nevertheless, most of the
valued data classification with the variational research on VQCs is still in its early stages. For instance, as a consequence of the
quantum classifier. mathematical constraints imposed by the properties of quantum states, the
Front. Quantum Sci. Technol. 3:1282730.
doi: 10.3389/frqst.2024.1282730 majority of research has not fully taken into account the impact of data
formats on the performance of VQCs. In this paper, considering a significant
COPYRIGHT
© 2024 Chen and Li. This is an open-access number of data in the real world exist in the form of complex numbers, i.e., phasor
article distributed under the terms of the data in power systems and the result of Fourier transform on image processing,
Creative Commons Attribution License (CC BY). we develop two categories of data encoding methods, including coupling data
The use, distribution or reproduction in other
forums is permitted, provided the original encoding and splitting data encoding. This paper features the coupling data
author(s) and the copyright owner(s) are encoding method to encode complex-valued data in a way of amplitude
credited and that the original publication in this encoding. By leveraging the property of quantum states living in a complex
journal is cited, in accordance with accepted
academic practice. No use, distribution or Hilbert space, the complex-valued data is embedded into the amplitude of
reproduction is permitted which does not quantum states to comprehensively characterize complex-valued information.
comply with these terms. Optimizers will be utilized to iteratively tune a parameterized ansatz, with the aim
of minimizing the value of loss functions defined with respect to the specific
classification task. In addition, distinct factors in VQCs have been explored in
detail to investigate the performance of VQCs, including data encoding methods,
loss functions, and optimizers. The experimental result shows that the proposed
data encoding method outperforms other typical encoding methods on a given
classification task. Moreover, different loss functions are tested, and the capability
of finding the minimum value is evaluated for gradient-free and gradient-based
optimizers, which provides valuable insights and guidelines for practical
implementations.

KEYWORDS

quantum machine learning, variational quantum classifier, complex-valued data,

coupling data encoding, splitting data encoding, loss function, optimizer

1 Introduction
Quantum computing, as a cutting-edge technology, is upending traditional computing
methods based on digital electronics, and has shown its unique advantages in various fields
(Arute et al., 2019; Hendrickx et al., 2021). The use of quantum superposition and
entanglement enables the possibility of solving intricate problems with exponential
speedup (Shor, 1994) or quadratical speedup (Grover, 1996). IBM Osprey, the latest

Frontiers in Quantum Science and Technology 01 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

Noisy Intermediate-Scale Quantum (NISQ) device used for general complex quantum circuit to implement and is highly based on
purposes, has up to 433 qubits, and its successor with more than the selection of basis functions. In reference to Havlíček et al. (2019),
1,000 qubits is about to be unveiled in 2023 (Gambetta, 2021). a VQC was constructed using a methodology akin to traditional
Despite the rapid growth in the number of qubits in quantum support vector machines. This approach not only enhances the
computers, the incorporation of quantum computing into practical classification performance but also opens up the possibility of
engineering problems still poses significant challenges, including exploiting quantum advantages. It introduces a sophisticated data
their accuracy, efficiency, and how to take quantum advantages. encoding technique designed to be computationally challenging for
Variational Quantum Algorithms (VQAs) hold great promise as classical evaluation, highlighting the potential for achieving
a viable technology applicable to near-term NISQ computers quantum advantages. Nevertheless, a comprehensive analysis of
(Cerezo de la Roca et al., 2021). By leveraging the inherent the individual factors influencing VQCs is not thoroughly
merits of classical optimization approaches and parameterized demonstrated. The relationship between feature maps and kernels
quantum circuits, VQAs possess the potential to mitigate the is implicitly analyzed in Schuld and Killoran (2019). The analysis
impact of noise stemming from NISQ devices, which is a result disclosed the fact that the process of data encoding is actually a
substantial advantage over other quantum algorithms. Quantum kernel method that maps data to quantum Hilbert space, and ansatz
machine learning (QML) serves as a popular candidate application with optimal parameters served as a hyperplane to separate unseen
in VQAs, where the VQA framework can be perfectly implemented. data. The performance of the proposed QML framework with more
The parameterized quantum circuits in VQAs can be optimized to qubits is still not clear, and further efforts are required to refine and
minimize/maximize the objective function by transforming a enhance the current approach.
general-purpose problem into a minimization or maximization task. In this paper, on the basis of the previous VQC framework, we
Recently, numerous research endeavors have been carried out develop a VQC that is used for handling the input data living in
on QML, including implementation on quantum processors complex domains, such as complex power and voltage in alternating
(Tacchino et al., 2019), network architecture (Cong et al., 2019), current power systems. By leveraging the property of n-qubit
and data encoding (Schuld et al., 2021). Tacchino et al. (2019) quantum states |ψ〉∈ (C2 )⊗n , the coupling data encoding method
proposed a perceptron quantum model utilized for an elementary exploits amplitude encoding to embed the complex-valued data into
classifier, which can be efficiently implemented on a real quantum quantum states. Then an optimizer is used to update the variational
computer. A state-of-the-art quantum conventional neural network, parameters in the parameterized quantum circuit until finding the
inspired by the traditional conventional neural network, was minimum of the loss function. Finally, the parameterized quantum
provided by Cong et al. (2019) under the assumption that the circuit with the set of optimal parameters is employed to process the
input can be prepared as a quantum state in a physical system. encoded quantum states, and further performs the classification
Schuld et al. (2021) comprehensively investigated the pros and cons problem, e.g., finding unstable equilibrium points of power system
of various data encoding methods in detail, which is an important from its operating dataset. Compared with other data encoding
part of variational quantum circuits, providing good guidance in the methods, the developed approach requires fewer qubits to
selection of data encoding methods. accomplish the classification, and it also gives the highest
VQCs are one of the commonly used algorithms in the field of accuracy. In addition, we investigate different factors in the VQC
QML. For a classification task, we use a set of training data DT and a construction, including data encoding methods, loss functions, and
set of test data DS with their labels, and DT ∪ DS
Ω ∈ Rd . The goal optimizers, which may have an effect on the performance of VQC.
of the classification task is to approximate a potential mapping of The results show that these factors directly affect the performance of
~
data features to labels on DT, which can predict labels L(s)
L(s) VQC, loss function l2, misclassification-based loss function, and
with high precision on DS. The loss function of VQCs is defined as gradient-based optimizers demonstrate better capabilities for
the error between the true labels and the expectation values of the the given task.
observations, and classical optimizers are employed to minimize it. The contributions of this work can be summarized into three
Parameterized circuits were popularized by authors such as Farhi main aspects.
et al. (2014), who studied parameterized circuits for solving MaxCut
and related problems. Since then, parameterized quantum circuits • We explore the applicability of amplitude encoding for complex-
have emerged for quantum classification algorithms. Farhi and valued data and validate its effectiveness in universal VQCs,
Neven (2018) formulated a general classification framework achieving high accuracy in classification tasks.
based on quantum computers. Initially, labeled input strings are • Different factors that may influence the result of VQCs are
mapped to computational quantum states. Classical optimizers train compared and analyzed in the experiment, including data
a quantum neural network to learn the mapping between strings and encoding methods, loss functions, and optimizers. The
labels. In this paper, only basis encoding was investigated, which was analysis and experimental results offer a reference for
only suited to the bit-strings z ∈ {0,1}n. Thus, methods for other types selecting the proper methods in VQCs.
of input data need to be further explored. Mitarai et al. (2018) • The power system data is utilized as the input data of VQCs.
proposed a novel and specialized data encoding method, designed to The experimental result of VQC shows the potential of QML
approximate analytical functions by the property of nonlinearity of in real-world engineering problems.
the designed tensor product. The numerical result shows the method
has the capability of representing elementary functions and has the The remainder of this paper is organized as follows. Section 2
potential to handle the non-linear classification task with a low- introduces the general scheme of VQCs. Section 3 introduces three
depth ansatz. However, this data encoding method requires a types of critical components in VQCs, including the coupling and

Frontiers in Quantum Science and Technology 02 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

FIGURE 1
The general scheme of VQCs.

splitting data encoding methods, the construction of loss functions Similar to conventional supervised learning, VQCs also contain
in the framework of VQCs, and typical optimizers. The numerical two phases: the training phase and the classifying phase. First, in the
result is given in Section 4 to verify the effectiveness of the coupling training phase, a loss function evaluates the gap between true labels
amplitude encoding method used for complex-valued data and and predicted values, so a classical optimizer can be employed to
compares the influence of the above factors. Conclusions are tune the ansatz. The goal of the training phase is to find a set of
discussed in Section 5. parameters that maximizes the accuracy of the classification models.
Second, in the classifying phase, the encoded quantum state will
evolve through the well-trained ansatz W(θ) with optimized
2 General scheme of VQC for complex- variational parameters θ. Finally, the resulting quantum state will
valued data be measured, and a customized function f: {0,1}n → {−1, 1} is utilized
to map the distribution of the measurement results to the labels.
A VQC is a class of hybrid methods that combines quantum From the aforementioned workflow, we can find that there are
computing techniques with classical optimizers (Cerezo de la Roca several factors that can potentially impact the effectiveness of VQCs,
et al., 2021), which can be used for classification tasks. Utilizing including the data encoding method, the loss function, the type of
VQCs allows the finding of an ansatz with optimized parameters to optimizers, etc. Carefully selecting methods in VQCs tailored to
classify data as does classical machine learning methods. Moreover, specific problems is crucial to achieving improved classification
by leveraging the advantages of entanglement and superposition of results. In practice, furthermore, a large amount of data is
qubits, VQCs potentially offer enhanced capabilities for tackling presented in the form of complex numbers, e.g., phase data for
classification problems. power systems and Fourier transform results on image processing.
The workflow of VQCs is outlined in Figure 1, including the How to process and analyze such data by VQC remains a
following three steps: challenging problem. In the following analysis, we will delve into
these factors and provide the most suitable method specifically for
• Data Encoding: In order to run in a superconducting quantum the case of complex-valued data.
processor, the training data set DT and test data set DS are
mapped to quantum states |Φ(x)〉 on a basis state, e.g., |0〉⊗n
(Nielsen and Chuang, 2010). 3 Methods—VQCs for complex-
• Ansatz: Ansatz (parameterized quantum circuit) acts as a valued features
candidate quantum circuit with a set of variational
parameters. The ansatz with optimal parameters can lead to 3.1 Data encoding for complex-valued data
the best quantum state for given tasks.
• Loss Function and Optimizer: This step is implemented on Data encoding is a methodology that transforms classical data
classical computers. A loss function is expected to be properly into quantum information, which can be further processed by
designed and then used to evaluate the gap between true labels quantum computers. Choosing proper data encoding methods
and predicted values, and optimizers are classical algorithms can better characterize classical data, enhancing its compatibility
that are used to adjust the parameters of ansatz, thereby and interpretability within the QVC framework. Schuld (2021) has
minimizing loss functions of VQCs. proven that data encoding is equivalent to kernel methods for

Frontiers in Quantum Science and Technology 03 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730


supervised learning, where kernel methods are employed to project ~
x2
x1 2 + x2 2 + / + xN 2 + 1. We can find that the
original data into a high-dimensional space. Thus, in the field of information on the norm of the vector remains in ^ x.
QML, the original data is also mathematically mapped into a new Therefore, by changing the N-dimensional vector to a N + 1-
Hilbert space through data encoding, which serves as the dimensional vector, the original data will not lose information. In
mathematical framework where quantum states reside and the quantum phase, if an N + 1-dimensional vector is encoded
interact. In practice, a unitary quantum circuit is commonly into an n-qubit quantum state, the number of the quantum state n
employed to realize the data encoding process. Specifically, the is required to satisfy n ≥ log (N + 1). Generally, n = log (N + 1)
basis state |0〉⊗n evolves through the unitary quantum circuit to holds with a very tiny probability, so the dimension of the
become different quantum states with input features. encoded vector is usually less than log (N + 1). When the
So far, researchers have proposed many data encoding strategies dimension of the processed data in (3) is less than 2n, we can
(Biamonte et al., 2017), including basis encoding, amplitude always extend it into a new vector with a dimension of precisely
encoding, angle encoding, etc. Data encoding methods can be 2n by adding zero entries. Thus, the data can fit into an n-qubit
achieved by building unitary quantum circuits. Rotation gates quantum state.
and controlled gates are the main components in these quantum Power systems are a vast and intricate network that encompasses
circuits. By taking advantage of quantum entanglement, 2n features extensive energy interactions and data transmission. Phasor
can be encoded into quantum states with a n-qubit quantum circuit. measurement units (PMU) are a fundamental monitoring
In this paper, a unitary quantum circuit U Φ(x) , a feature map, component in modern alternating current power systems Guarnieri
represents the data encoding process, and the resulting states (2013). With the large-scale deployment of PMUs, modern power
processed by this map can be expressed as |Φ(x)〉
U Φ(x) |0〉⊗n . systems have become easier to monitor and control. In practice, PMUs
Next, two categories of data encoding strategies will be introduced quickly capture the samples from a sinusoidal waveform (including
for complex-valued data used for QML. voltage, current, and power) and compute the magnitude and phase
angle of an electrical phasor quantity.
3.1.1 Coupling encoding method for complex- For an operating power system, the complex power S in phasor
valued data form can be expressed as Eq. 4,
The coupling encoding method refers to encoding complex-valued
S
|S|∠ϕ
|S|ejϕ
|S| cos ϕ + j|S| sin ϕ
P + jQ, (4)
data as a whole entity without splitting it into individual components
and representing it in a quantum state when performing data encoding. where j represents the imaginary unit, |S| and ϕ are the magnitude and
Amplitude encoding is a type of quantum state preparation in which the phase angle of the complex power, P denotes the active power and Q
classical data are encoded as the amplitudes of quantum superposition denotes the reactive power. Similarly, this property exists in various
states, and also one of the common data encoding methods in the field electrical quantities. Note that S ∈ C in which the amplitudes of
of quantum computing is one of the common data encoding methods quantum states live. Hence, the raw complex data provided by PMUs
in the field of quantum computing (Harrow et al., 2009). For a classical can be directly embedded in quantum states by amplitude encoding. We
n n
vector x ∈ C2 that satisfies
2k
1 |xk |2
1, it can be represented by the can also use the Bloch sphere to visualize a normalized complex electrical
amplitudes of a quantum state |Φ(x)〉, which is defined by Eq. 1, quantity shown as Figure 2. For high-dimensional complex power data,
2n each feature can be embedded into the amplitude in (1) as long as the
|Φ(x)〉
 xk |k〉, (1) data encoding circuit with a sufficient number of qubits is available.
k
1
For a power system dataset with m complex-valued features, e.g.,
where |k〉 is the kth computational basis. complex power, it can be represented by a m-dimensional vector
If the original vector x ∈ CN is not normalized, in order to Eq. 5,
encode x through amplitude encoding, the initial step is to normalize s
P1 + jQ1 , P2 + jQ2 , . . . , Pm + jQm
T
(5)
it to ensure that it satisfies
N k
1 |xk |
1. This step implies that
2

quantum states represent the data in a dimension that is one less The quantum state after amplitude encoding is represented by
than its original dimension. To address this issue, one approach is to Eq. 6,
add an element 1 to the given feature vector and normalize the
^ 1 |1〉 + P^2 + jQ
|Φ(s)〉
P^1 + jQ ^ 2 |2〉+, . . . , + P^m + jQ
^ m |m〉
resulting vector. For instance, for any vector x

[x1 , x2 , . . . , xN ]T ∈ CN to be normalized, we can increase its 1
+ |m + 1〉,
dimension by one and set the added component as 1. The s22 + 1
augmented vector ~x ∈ CN+1 can be expressed as Eq. 2,
(6)
~x
[x1 , x2 , . . . , xN , 1]T , (2)
where s
2 is  the Euclidean  
norm of the vector s, and
^ i
Qi / s2 + 1.
P^i
Pi / s22 + 1, Q
The augmented vector after normalization is given by Eq. 3, 2
An inherent advantage of amplitude encoding is that it allows for
x1 x2 xN 1
T the exploitation of quantum superposition and entanglement
^x
 , ,..., ,  , (3) properties, so a large number of features can be input into the QML
~x ~x ~x ~x
algorithm with a limited number of qubits. Specifically, the complex
where ~x is the norm of the augmented vector ~x. When the vector x with 2n features can be encoded into a n-qubit quantum state by
Euclidean norm is employed for normalization, leveraging the superposition state of quantum systems. This means, by

Frontiers in Quantum Science and Technology 04 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

FIGURE 2
Data encoding for the phasor data in power systems.

FIGURE 3
An angle encoding architecture for complex-valued data with multiple repetitions. Ri,j represents the rotation gate of ith qubit in jth layer, different
colored blocks indicate the utilization of different rotation gates within the data encoding architecture. The yellow and green boxes correspond to Ry and
Rz rotation gates, respectively.

adopting amplitude encoding, data can potentially be processed in 3.1.2.1 Angle encoding
parallel, indicating a potential to reduce computational time and Angle encoding is a common data encoding method based on
improve efficiency compared to classical machine learning methods. rotation gates. It utilizes the angles of quantum gates as parameters
to encode classical information in quantum states and can be applied
3.1.2 Splitting encoding methods for complex- to different forms of data (Schuld, 2021), such as binary, real-valued,
valued data or categorical data. Therefore, angle encoding does not necessitate
Contrary to the coupling encoding method, the splitting classical data normalization, meaning that data preprocessing for
encoding method involves processing the real part and the angle encoding is not required. The basic angle encoding with a
imaginary part of complex-valued data separately. Many data single layer of Pauli rotation gates is given by Eq. 7,
encoding methods only accept real numbers as the parameters of
their quantum circuits, e.g., angle encoding and Pauli feature |Φ(x)〉
⊗ni
1 (Ri (xi )|0〉), (7)
map. For this type of data encoding method, we have to separate
the real parts and the imaginary parts, so we can further embed them where Ri represents the rotation gate of ith qubit, which can be expressed
into quantum states. as the exponential form of Pauli matrices σ, where σ ∈ {I, X, Y, Z}.

Frontiers in Quantum Science and Technology 05 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

FIGURE 4
Pauli feature map for complex-valued data.

Furthermore, a more intricate quantum circuit with multiple where ϕk(x) is the encoding functions embedded in single-qubit-gate
layers and repetitions for complex-valued angle encoding can be unitary, and ϕm,n(x) is the encoding functions embedded in two-
constructed by incorporating more layers of Pauli rotation gates, qubit-gate unitary. The repetition of H and UΦ(x) can be adjusted to
i.e., Figure 3, where H is the Hadamard gate that maps |0〉 to pursue better performance for VQCs. For instance, given the trade-
√1|0〉 + √1|1〉. Nonetheless, for angle encoding, the number of off between the circuit depth and its performance, the provided
2 2
qubits is equal to the dimension of features as only rotation gates feature map in Havlíček et al. (2019) consists of two Hadamard
exist in the data encoding circuits, which limits the potential transforms and uses the Ising model, and it has shown good
exploitation of quantum entanglement benefits. performance for their classification tasks. When specifying to
For the dataset being the same as Eq. 5, the quantum state after power system applications, the Pauli feature map is given in Figure 4.
processing by the angle encoding in Figure 3 can be represented as Eq. 8, For the dataset being the same as (5), the quantum state after
c processing by the Pauli feature map with two repetitions can be
|Φ(s)〉
i
1 Ri,1 (Pi )Ri,2 (Pi ) ⊗ ⊗i
1 Rm+i,1 (Qi )Rm+i,2 (Qi ) ⊗ H
⊗m m ⊗2m
|0〉⊗2m , represented as Eq. 13,
k
1

(8) |Φ(s)〉
UΦ(sf )H⊗m UΦ(sf )H⊗m |0〉⊗m , (13)

where c is the number of repetitions in 3.
where sf
[P1 , Q1 , P2 , Q2 , . . . , Pm , Qm ]T ∈ R2m .
Pauli feature maps, compared with angle encoding, exhibit the
3.1.2.2 Pauli feature map
ability to handle diverse data types, but offer the potential for
Pauli feature map was first proposed in Havlíček et al. (2019),
quantum advantage through the exploitation of quantum
and it represents a family of data encoding methods. It can be
parallelism and entanglement. Nevertheless, the considerable
formulated as follows Eqs 9, 10,
number of quantum gates and qubits required can pose
U Φ(x)
UΦ(x) H⊗n . . . UΦ(x) H⊗n , (9) challenges. Moreover, its complex structure and the extensive
number of gates can result in a deep circuit, leading to the
j
ϕC (x)  σ i  encoding circuit being more susceptible to noise and errors in
UΦ(x)
e C⊆[n] i∈C
, (10)
current quantum hardware.
where [n] means the enumeration of n qubits, and C describes the
interconnections between individual qubits. Theoretically, C is very
large when the number of qubits is high, which renders the 3.2 Parameterized quantum circuits of VQCs
implementation of the Pauli feature map more intricate. for complex-valued data
Therefore, we have |C| ≤ 2 in this paper, so the Pauli feature
map can be efficiently implemented. ϕC(x) is a set of linear fl(x) Parameterized quantum circuits are typically referred to as
and the non-linear fnl(x) functions of the original data. Havlíček et al. ansatz in the quantum computing field. They are a critical part
(2019) provides examples of functions utilized for the Pauli feature of VQCs, whose structure plays a crucial role in determining the
map, which are defined as follows Eqs 11, 12, final parameters, thereby influencing the training process to achieve
optimal results. Among the various ansatzes, layered hardware-
efficient ansatzes have been prominently used for general-purpose
ϕk (x)
xk , (11) problems. What is even more noteworthy is that quantum circuits
ϕm,n (x)
(π − xm )(π − xn ), (12) can be constructed with fewer layers or parameters thanks to the

Frontiers in Quantum Science and Technology 06 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

FIGURE 5
The demonstration of layered parameterized quantum circuit.

symmetry and locality properties inherent in quantum circuits In Figure 5, the variational parameters only exist in the rotation
(Iblisdir et al., 2014). Generally, ansatz architecture usually gates. Experimental validation in Kandala et al. (2017) has
employs a kind of universal and ‘problem-agnostic’ demonstrated that entangling unitaries with fixed phases exhibit
parameterized quantum circuits, so they can be utilized as the superior performance compared to entangling unitaries with
training quantum circuit even in those situations where no variational parameters. Thus, the number of variational
relevant information is readily accessible. parameters in the ansatz comes to 3ln, where n is the number of
For a general parameterized quantum circuit, the training implementing qubits.
parameters are embedded in unitary operators W(θ) consisting of In order to reduce the complexity of the VQC, we only use Pauli
a collection of basic quantum gates, including some parameterized Y and Pauli Z rotations in the single-qubit-gate unitaries defined by
and unparameterized gates. By increasing the depth of a (15), which represents a set of rotation gates in a layer of W(θ)
parameterized quantum circuit, the number of variational in Figure 5,
parameters θ also increase, through which the function-fitting
capabilities of ansatz are expected to be improved. For W(θ), it U(locl) (θl )
⊗nm
1 Uθm,t , (15)
can be represented by the product of a series of unitaries Eq. 14,
where θl ∈ R2n , and θm,t ∈ R2 , and U(θ) is defined as Eq. 16,
W(θ)
Wl (θl ) . . . W2 (θ2 )W1 (θ1 ) (14)
1 zZ 1 yY
Uθm,t 
ej2θ m
ej2θ m
, (16)
where l is the number of layers in ansatz.
In this paper, an ansatz initially proposed in Kandala et al. where Ym means the mth Pauli Y matrix and Zm means the mth Pauli
(2017) is implemented for the classification issue of complex-valued Z matrix.
data, as shown in Figure 5, which contains single-qubit rotation To further simplify the construction of the ansatz, we can build
gates with variational parameters and entangling unitaries with fixed the entangling unitary based on the Ising model. The controlled-Z
phases. Each single-qubit-gate unitary is followed by an entangling gates are utilized as entanglement gates, which only entangle with
unitary, except for the final unitary Wl (θl). This ansatz architecture the neighboring qubits. The entanglement unitary is given by (17),
leads to better performance on superconducting quantum which represents the entangling unitary in a layer of W(θ)
processors, as summarized in Kandala et al. (2017). in Figure 5.

Frontiers in Quantum Science and Technology 07 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

2
Uent
CZi, j, (17) Ll2 (θ)
 yi − 〈Φ(xi )|W† (θ)OW(θ)|Φ(xi )〉 (20)
(i,j)∈E i

where E represents the collection of neighboring qubits, Another typical loss function replaces (·)2 with absolute value |·|,
and CZ (i, j) represents the controlled Z gates applied at which is named l1 loss.
qubit i and j.
In this setting Eq. 14 can be expanded in a more detailed manner (2) Cross-entropy loss function: The cross-entropy loss based on
Eq. 18, and the structure of the VQC used is shown in Figure 5, information theory is also popular in QML. It also exhibits the
property of convexity, akin to the square function, making it
W(θ)
U(locl) (θl )Uent . . . U(loc2) (θ2 )Uent U(loc1) (θ1 ), (18) well-suited for gradient descent optimization methods
(Murphy, 2012). The expression is defined as Eq. 21,
where U(l)
loc (θl ) denotes the lth single-qubit-gate unitary, and Uent
represents the entangling unitary. Lent (θ)
 −yi logpi − 1 − yi log1 − pi , (21)
As the number of input qubits increases, the depth of the i

proposed VQCs needs to be incrementally increased to meet the

where pi = 〈Φ(xi)|W†(θ)OW(θ)|Φ(xi)〉 in VQCs.
demand for better expressivity. However, considering the rapid
increase in the number of parameters arising from deeper
(3) Misclassification-based loss function: This loss function is
quantum circuits, the number of layers in the ansatz is set to
formulated to assess the performance of VQCs by evaluating
2 in this paper, as detailed in Section 4.
the probability of occurrence of misclassified samples.
Havlíček et al. (2019) introduces a loss function for binary
classification, which shows a good result in their classification
3.3 Loss functions and optimizers of VQCs tasks. Here we provide a brief overview of it. By evaluating the
for complex-valued data ratio of misclassified samples, the loss function can be defined
as Eq. 22,
3.3.1 Loss functions
For a complex-valued dataset with labels, such as power system 1 ⎛
Perr
⎝  Pr m ⎠,
~ T (s) ≠ y | s ∈ Ty ⎞ (22)
data with stable and unstable scenarios, the loss function is designed |T| y s∈Ty
in VQCs to learn the relationship between the power system data
and the stability characteristics, i.e., the label of each operational where |T| is the total number of samples. Ty is the sample subset
scenario. Thus, we can assess power system stability from test data. labeled by y. m~ T (s) is the measurement result for the samples of
Thus, the choice of the loss function has a significant effect on the the subset Ty.
optimization process, as it alters the landscape of the optimization The estimation of misclassifying ratio can be thought of as
problem. Therefore, the loss function needs to be carefully selected. carrying out experiments R times for each sample, which follow a
For classification tasks, in general, the loss function can be defined as binomial distribution. Thus, the probability of misclassification is
the error between the true label and the expectation value under an given by Eq. 23,
observable O. Basically, the observable O here serves as a bridge R
between the final quantum states and the data label {+1, −1}. R2
R−i
~ T (s) ≠ y | s ∈ Ty
 piy 1 − py
Pr m , (23)
Mathematically, the observable can be viewed as an operator and i
0
i
has eigenvalues either +1 or −1, so the expectation value 〈Φ(xi)|
W†(θ)OW(θ)|Φ(xi)〉 under the observable O can be bounded by [−1, where R/2 means that a sample is misclassified when less than half of the
+1]. It can be further utilized to assess the gap to labels by tuning the measurements in R experiments are not the state corresponding to the
parameters in the ansatz. In quantum physics, the observable O true label. The probability of classification labels y is expressed as Eq. 24,
measures the property of quantum states by means of successive 1  
py
1 + y〈Φ(x)W† (θ)OW(θ)Φ(x)〉 , (24)
operations, e.g., applying electromagnetic fields, and eventually 2
reading a value in [−1, +1]. The general form of the loss function r
can be given by Eq. 19, In practice, py can be estimated by py
Ry , where ry is the
number of occurrences of samples labeled by y.
L(θ)
 fyi , 〈Φ(xi )|W† (θ)OW(θ)|Φ(xi )〉, (19) When R is very large, the probability of misclassification can be
i
further expressed as Eq. 25. See (Havlíček et al., 2019) for details.
where f denotes the loss function, and yi is the label of ith sample.
⎜√ 1
− py ⎟
In the field of traditional machine learning, various loss ~ T (s) ≠ y | s ∈ Ty ≈ sig⎛
Pr m ⎜
⎝ R 
2 ⎞
⎟
⎠ (25)
functions with favorable capabilities have been proposed 2 1 − py py
(Alpaydin, 2010). Correspondingly, these loss functions can also
be utilized in VQCs to tackle classification problems effectively. Here
we provide examples for designing the loss function. 3.3.2 Optimzers for VQCs
In order to obtain the optimal parameters of ansatz, an
(1) l1 and l2 loss functions: The squared loss, also called l2 loss, is optimizer implemented in classical computers makes queries to
frequently selected as the preferred loss function, and Eq. 19 the quantum device repeatedly and searches for a set of parameters
can be reformulated as Eq. 20, better than the current one. Optimizers can be divided into two

Frontiers in Quantum Science and Technology 08 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

FIGURE 6
A microgrid system containing renewable energy sources and power loads.

FIGURE 7
Data visualization for active power and reactive power of stable and unstable scenarios. The label of the x-axis indicates 10 different sets of samples,
each containing 10 features. For each sample, it induces dynamic behavior in the system, leading to either convergence or divergence, similar to the
trajectories shown in the right panel.

Frontiers in Quantum Science and Technology 09 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

FIGURE 8
The designs of angle encoding: (A) Angle encoding A, (B) Angle encoding B.

TABLE 1 The number of quantum gates used for different data encoding methods.

Data encoding method Single-qubit gates Two-qubit gates Circuit depth

Angle encoding A 40 0 4

Angle encoding B 40 0 4

Pauli feature map 40 40 53

PCA + Angle encoding A 12 0 4

PCA + Angle encoding B 12 0 4

PCA + Pauli feature map 18 12 22

Splitting amplitude encoding 15 14 26

Coupling amplitude encoding 21 6 21

categories based on whether or not analytical gradients are required, analytical expression of gradients, it approximates the gradient by
as gradients are usually used to speed up finding the minimum value perturbing the parameters of the optimization problem in a
of loss functions. stochastic way, thereby adjusting the direction of optimization.
By utilizing random and simultaneous perturbation, SPSA can
(1) The first category of optimizers is the gradient-free process noisy and non-differentiable objective functions with a
optimizers, e.g., the Simplex algorithm and simultaneous relatively fast convergence speed.
perturbation stochastic approximation (SPSA). The
Simplex algorithm is an optimizer designed to find the (2) Another category of optimizers is the gradient-based
minimum value when the derivative of the objective optimizers. Gradient-based optimizers require the
function is unknown, making it a gradient-free and easy- calculation of analytical derivatives to tune the ansatz in
to-implement method. Constrained optimization by linear VQCs. The most common algorithm is the gradient
approximations (COBYLA) is one of the most widely used descent, which is given by Eq. 26,
approaches based on the Simplex method, especially in the
θ^
θ^ − αk ∇Lθ^ ,
k+1 k k
optimization of ansatz (Powell, 2007). The COBYLA is a (26)
sequential trust-region optimization technique that relies on
where θ^ means the current parameters, αk is the learning rate and
k
linear approximations. These approximations are constructed
∇L(θ^ ) is the estimated gradient at kth iteration.
k
using linear interpolation at n + 1 points in the variable space,
ensuring the maintenance of a well-formed simplex Computing the analytical gradients of a loss function always
throughout the iterations. It is particularly suitable for involves the derivatives of quantum circuit operators. The phase
solving non-smooth and nonlinear optimization problems shift strategy can be further utilized to compute derivatives with little
with a moderate number of variables. computing resources (Schuld et al., 2019). Among the popular
gradient-based optimizers in the quantum domain, well-known
There is an alternative optimization approach that does not rely ones include Conjugate Gradient Optimizer (CG) and Analytic
on accessing the gradient. If the derivative of the cost function is Quantum Gradient Descent (AQGD), etc.
difficult to access, numerical methods can be used to approximate We compared the performance of different optimization
the gradient. A typical example of such an algorithm is SPSA algorithms in the data set of the power system in the following
introduced in Spall (2005). Although it does not require the numerical examples.

Frontiers in Quantum Science and Technology 10 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

TABLE 2 Numerical result for different data encoding methods.

Data encoding method Qubit count Accuracy (%) Precision (%) Recall (%) F1-score (%) Loss function value
Angle encoding A 10 66.5 63.8 98.2 77.3 0.367

Angle encoding B 10 73.2 73.6 84.1 78.5 0.315

Pauli feature map 5 85.1 80.0 99.1 88.5 0.250

PCA + Angle encoding A 3 86.6 81.3 100 89.7 0.268

PCA + Angle encoding B 3 82.5 77.6 98.2 86.7 0.292

PCA + Pauli feature map 3 81.4 77.3 96.5 85.8 0.302

Splitting amplitude encoding 4 88.6 85.8 96.5 90.8 0.250

Coupling amplitude encoding 3 92.3 89.5 98.2 93.7 0.185

FIGURE 9
Stable and unstable samples distribution processed by PCA.

4 Results were employed to identify the stability of the power system

operating at different equilibrium points (Milano, 2005). A total
In this section, we obtain complex-valued data from a microgrid of 366 stable samples and 602 unstable samples were created for the
system to validate the effectiveness of the VQC developed in this experiment. Each equilibrium point contains various features (e.g.,
paper, as shown in Figure 6. All power system simulations are bus voltage, frequency, active power flow, reactive power flow, etc.)
conducted using the Power System Analysis Toolbox (PSAT) in the under both stable and unstable scenarios. In this paper, only active
MATLAB environment. The microgrid system contains 10 buses, power and reactive power (a total of 10 real-valued features) in the
5 renewable energy sources, and 5 power loads. Since changing renewable energy sources buses are selected as the key features to
generator power and load power may lead to power system input into the VQC, which can be combined to form complex power
instability, this system can operate at various equilibrium points as the input of complex-valued amplitude encoding. The label of
by setting different active and reactive power demands of loads and unstable samples is set to −1 and the label of stable samples is set to 1.
generations of renewable energy sources. Time-domain simulations The generated dataset is divided into the training set and the test set,

Frontiers in Quantum Science and Technology 11 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

TABLE 3 Numerical result for different loss functions on testing dataset.

Loss function Accuracy (%) Precision (%) Recall (%) F1-score (%) Loss function value
l1 loss function 88.1 58.9 100 74.1 0.645

l2 loss function 92.3 89.5 98.2 93.7 0.185

Cross-entropy loss function 89.1 85.4 98.2 91.4 0.501

Misclassification-based loss function 92.3 89.5 98.2 93.7 0.109

FIGURE 10
Optimization landscapes of the second two parameters θ2,1 and θ2,2 in Figure 5 when other parameters are optimal: (A) l1 loss function, (B) l2 loss
function, (C) cross-entropy loss function, (D) misclassification-based loss function.

and the training set data accounts for 80% of the total dataset. We objective of the VQC is to directly distinguish the unstable samples
selected 10 stable samples and 10 unstable samples to visualize their of equilibrium points from those stable ones based on the features
features, shown in Figure 7. For each sample, active power and generated. This section investigates the impact of data encoding
reactive power in 5 power source buses (total 10 features) are methods, loss functions, and optimizers on the final classification
selected for data visualization. The unstable samples will lead to results of VQCs when applied to the power system dataset.
power oscillations with increasing amplitude, which could make the
power system crash eventually. The stable samples will cause power
oscillations with decreasing amplitude, and the oscillations will 4.1 Comparison of data encoding methods
disappear finally and the power system will be running at a new
equilibrium point normally. Different data encoding methods could have different abilities to
In our experimental study, all related quantum circuits characterize the features in datasets. In this section, the coupling
associated with the data encoding methods and the VQCs are data encoding strategy and the splitting data encoding strategy were
formulated using the Qiskit simulator. A random initialized VQC explored on the microgrid dataset. The active power and the reactive
built in Section 3.2 is employed to execute the classification task. The power generation of renewable energy sources were used as the input

Frontiers in Quantum Science and Technology 12 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

FIGURE 11
Training process of different optimizers: (A) COBYLA, (B) SPSA, (C) CG, (D) AQGD.

of VQC. In order to reflect which data encoding method is more the complex power. The number of quantum gates used for different
suitable for power system data, here we adopted the squared error as data encoding methods is shown in Table 1.
the loss function and the COBYLA as the optimizer for each In this section, to better evaluate the classification model, we use
training. Besides, we used the same two layers in the four types of indexes, including accuracy, precision, recall, and F1-
parameterized quantum circuit to compare the experimental score, which are widely used in binary classification. They are
results of different data encoding methods. defined as Eqs 27–30,
Five types of data encoding methods were applied to encode the
TP + TN
power system features, including two angle encoding methods, Pauli Accuracy
, (27)
TP + TN + FP + FN
feature map, splitting amplitude encoding, and coupling amplitude
TP
encoding. The designs of the angle encoding circuits adopted are Precision
, (28)
TP + FP
shown in Figure 8, and the Pauli feature maps are constructed as the
TP
structure in Section 3.1.2. They all consist of the Hadamard gates Recall
, (29)
TP + FN
and the Pauli rotation gates, and they have two same repetitions to
2*Precision*Recall
strike a balance between performance and circuit depth. In addition, F1 − Score
, (30)
Precision + Recall
in order to achieve better performance on angle encoding methods
and the Pauli feature map, we used a data dimensionality reduction where TP denotes the number of true positives, TN the number of
method to reduce the dimension of the features before performing true negatives, FP the number of false negatives, and FN the number
the classification of VQCs. Hence, in this section, we utilized both of false negatives. The results of VQCs on the test dataset for
the original data and the data processed by principal component different data encoding methods are shown in Table 2. The data
analysis (PCA) as the input features of VQCs to obtain distribution processed by PCA is shown in Figure 9.
comprehensive experimental results. The principal components From Table 2, we can see that,
corresponding to the largest three eigenvalues were selected as
the input features. Considering the active power and the reactive • The coupling amplitude encoding can achieve 92.3% accuracy
power can be represented as a vector on the complex plane, the on the microgrid dataset, which performs better than the other
proposed complex-valued amplitude encoding is used for encoding data encoding methods. Angle encoding A, Angle encoding B,

Frontiers in Quantum Science and Technology 13 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

TABLE 4 Numerical result for different optimizers.

Data encoding method Angle encoding A (%) Angle encoding B (%) Pauli feature map (%)
COBYLA 84.5 86.6 80.4

SPSA 78.9 83.0 79.9

CG 84.5 88.7 80.4

AQGD 82.9 88.1 85.6

the Pauli feature map and the splitting amplitude encoding 4.3 Comparison of optimizers
can not achieve over 90% accuracy on the test data.
• Angle encoding A and Angle encoding B yield notably Optimizers play a pivotal role in optimization problems. In
unfavorable outcomes, a consequence attributed to the this section, we examined four common types of optimizers used
increase of qubits count in VQCs, thereby rendering for VQC optimization, aiming to investigate the performance of
optimization challenging. various optimization methods. The tested optimizers not only
• After applying PCA to the data, the performance of the include the gradient-free methods, i.e., COBYLA and SPSA, but
method has been improved overall, due to the shallow also gradient-based methods, i.e., AQGD and CG. We also utilize
depth of the parameterized quantum circuit. The barren the data processed by PCA as the input features. Since we found
plateau issue would become more intractable with that the difference in results between different optimizers was
increasing size of qubit count. We find that there is also a small when using the coupling amplitude encoding, the two kinds
decrease in the accuracy of the VQC model when the qubit of angle encoding methods and the Pauli feature maps are
count has to be increased due to different coding methods employed as the data encoding methods in the test. The l2
used in the numerical results. loss function is utilized as the loss function. Figure 11 shows
the training process of different optimizers by using the Angle
encoding A. Uniform distribution was employed to realize the
4.2 Comparison of loss functions random initialization of the VQC training. More specifically,
random numbers from −π to π are drawn uniformly as the initial
The choice of loss function directly influences the optimization parameters of the VQC. The maximum number of iterations is set
landscape of VQCs. Choosing appropriate loss functions can aid in to 2000, 500, 100 and 70, corresponding to COBYLA, SPSA, CG
the development of a more effective and efficient classification and AQGD respectively. The gradient norm tolerance of CG is set
model. In this section, four types of loss functions are to 10–5 and the learning rate of AQGD is set to 1.0. Table 4 shows
investigated, including l1 loss, l2 loss, cross-entropy loss, and the the accuracy of the proposed VQC under different optimizers
misclassification-based loss function proposed in Havlíček et al. and encoding methods.
(2019). To showcase which loss function performs better for this From Figure 11 and Table 4, we can find that,
problem, we use the dimensional-reduced microgrid system data
processed by PCA, and fix the data encoding method as the coupling • The tested optimizers have comparable capabilities to find the
amplitude encoding and employed the COBYLA as the optimizer. optimal parameters of the VQC. The gradient-based methods
The numerical results for the four loss functions are shown in slightly outperform the gradient-free methods in terms of
Table 3, and optimization landscapes are shown in Figure 10. classification accuracy. Specifically, the gradient-based
From Table 3, we can find that, method AQGD and CG show better performance than
gradient-free methods for searching the optimal parameters
• The l1 loss function exhibits poor optimization performance of VQCs. Furthermore, the training process of SPSA
for the proposed VQC. Figure 10A shows the minimum of the fluctuates, leading to a slight accuracy decrease compared
landscape is close to 0.5. Compared to l2 loss function in other three optimizers.
Figure 10B, the l1 loss function yields a significantly higher • The COBYLA took less time to find the optimum because the
minimum value. parameter count is limited in this experiment. With the
• Among the tested loss functions, the performance of the l2 loss growth of the number of qubits and variational parameters,
function and the misclassification-based loss function excels in the gradient-based methods are expected to exhibit
the classification tasks, both achieving the highest accuracy higher efficiency.
92.3%. The l1 loss function and the cross-entropy loss function
yield comparatively low accuracy.
• These loss functions may exhibit barren plateaus. If gradient- 5 Conclusion
based optimizers are utilized to find the minimum of the loss
function, barren plateaus may be a daunting challenge for This study introduces a new type of VQC designed
optimization. Furthermore, they all exhibit landscapes with specifically for data in the complex domain. Research focuses
multiple local optimums that could hinder the optimizer from on a comparative analysis of various factors that have the
finding the global optimum. potential to impact VQC performance. Leveraging the

Frontiers in Quantum Science and Technology 14 frontiersin.org

Chen and Li 10.3389/frqst.2024.1282730

inherent information of quantum states, the coupling amplitude Funding

encoding approach exhibits a greater ability to capture the
intrinsic nature of complex numbers, surpassing the efficacy of The author(s) declare that no financial support was received for
traditional splitting data encoding methods. To operationalize the research, authorship, and/or publication of this article.
these insights, the research proceeds to implement the VQC
framework, employing the coupling data encoding method to
achieve a highly accurate classification of labeled complex-valued Acknowledgments
data. To further elevate the performance of VQCs, this study
evaluates and compares several techniques, including data encoding, This work is supported by the Office of Naval Research under
loss function, and optimizer, all tailored to the demands of power the award N00014-22-1-2504. The authors thank Dr. Phil Lotshaw
system tasks. This in-depth assessment not only offers valuable insights at Oak Ridge National Laboratory and Dr. Hong Chen at PJM
but also serves as a guide for selecting optimal methods to effectively Interconnection for their important feedback on the manuscript.
train the model. The comprehensive evaluations conducted in this study
consistently underscore the advance of coupling amplitude encoding in
complex-valued data classification within the VQC framework. The Conflict of interest
results highlight its superiority over alternative data encoding methods,
reinforcing its potential to significantly enhance VQC performance. The authors declare that the research was conducted in the
absence of any commercial or financial relationships that could be
construed as a potential conflict of interest.
Data availability statement
The raw data supporting the conclusions of this article will be Publisher’s note
made available by the authors, without undue reservation.
All claims expressed in this article are solely those of the authors and
do not necessarily represent those of their affiliated organizations, or
Author contributions those of the publisher, the editors and the reviewers. Any product that
may be evaluated in this article, or claim that may be made by its
JC: Writing–original draft. YL: Writing–review and editing. manufacturer, is not guaranteed or endorsed by the publisher.

References
Alpaydin, E. (2010). Introduction to machine learning. Adaptive computation and Kandala, A., Mezzacapo, A., Temme, K., Takita, M., Brink, M., Chow, J. M., et al.
machine learning. 2nd ed. Cambridge, Mass: MIT Press. OCLC: ocn317698631. (2017). Hardware-efficient variational quantum eigensolver for small molecules and
quantum magnets. Nature 549, 242–246. doi:10.1038/nature23879
Arute, F., Arya, K., Babbush, R., Bacon, D., Bardin, J. C., Barends, R., et al. (2019).
Quantum supremacy using a programmable superconducting processor. Nature 574, Milano, F. (2005). An open source power system analysis toolbox. IEEE Trans. Power
505–510. doi:10.1038/s41586-019-1666-5 Syst. 20, 1199–1206. doi:10.1109/TPWRS.2005.851911
Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe, N., and Lloyd, S. (2017). Mitarai, K., Negoro, M., Kitagawa, M., and Fujii, K. (2018). Quantum circuit learning.
Quantum machine learning. Nature 549, 195–202. doi:10.1038/nature23474 Phys. Rev. A 98, 032309. doi:10.1103/physreva.98.032309
Cerezo de la Roca, M. V. S., Arrasmith, A. T., Babbush, R., Benjamin, S. C., Endo, S., Murphy, K. P. (2012). Machine learning: a probabilistic perspective. Adaptive
Fujii, K., et al. (2021). Variational quantum algorithms. Nat. Rev. Phys. 3, 625–644. computation and machine learning series. Cambridge, MA: MIT Press.
doi:10.1038/s42254-021-00348-9
Nielsen, M. A., and Chuang, I. L. (2010). Quantum computation and quantum
Cong, I., Choi, S., and Lukin, M. D. (2019). Quantum convolutional neural networks. information: 10th anniversary edition. Cambridge, MA: Cambridge University Press.
Nat. Phys. 15, 1273–1278. doi:10.1038/s41567-019-0648-8 doi:10.1017/CBO9780511976667
Farhi, E., Goldstone, J., and Gutmann, S. (2014). A quantum approximate Powell, M. (2007). A view of algorithms for optimization without derivatives. Math.
optimization algorithm. arXiv: Quantum Physics TODAY 43.
Farhi, E., and Neven, H. (2018). Classification with quantum neural networks on near Schuld, M. (2021). Supervised quantum machine learning models are kernel methods.
term processors. arXiv: Quantum Physics
Schuld, M., Bergholm, V., Gogolin, C., Izaac, J., and Killoran, N. (2019). Evaluating
Gambetta, J. (2021). Ibm quantum roadmap to build quantum-centric supercomputers. analytic gradients on quantum hardware. Phys. Rev. A 99, 032331. doi:10.1103/
physreva.99.032331
Grover, L. K. (1996). “A fast quantum mechanical algorithm for database search,” in
Proceedings of the twenty-eighth annual ACM symposium on Theory of computing - stoc ’96 Schuld, M., and Killoran, N. (2019). Quantum machine learning in feature hilbert
(Philadelphia, Pennsylvania, United States: ACM Press), 212–219. doi:10.1145/237814.237866 spaces. Phys. Rev. Lett. 122, 040504. doi:10.1103/physrevlett.122.040504
Guarnieri, M. (2013). The alternating evolution of dc power transmission [historical]. Schuld, M., Sweke, R., and Meyer, J. J. (2021). Effect of data encoding on the
IEEE Ind. Electron. Mag. 7, 60–63. doi:10.1109/MIE.2013.2272238 expressive power of variational quantum-machine-learning models. Phys. Rev. A
103, 032430. doi:10.1103/PhysRevA.103.032430
Harrow, A. W., Hassidim, A., and Lloyd, S. (2009). Quantum algorithm for linear
systems of equations. Phys. Rev. Lett. 103, 150502. doi:10.1103/PhysRevLett.103.150502 Shor, P. (1994). “Algorithms for quantum computation: discrete logarithms and factoring,”
in Proceedings 35th annual symposium on foundations of computer science (Santa Fe, NM,
Havlíček, V., Córcoles, A. D., Temme, K., Harrow, A. W., Kandala, A., Chow, J. M.,
USA: IEEE Comput. Soc. Press), 124–134. doi:10.1109/SFCS.1994.365700
et al. (2019). Supervised learning with quantum-enhanced feature spaces. Nature 567,
209–212. doi:10.1038/s41586-019-0980-2 Spall, J. C. (2005). “Simultaneous perturbation stochastic approximation,” in
Introduction to stochastic search and optimization (Hoboken, NJ, USA: John Wiley
Hendrickx, N. W., Lawrie, W. I. L., Russ, M., Van Riggelen, F., De Snoo, S. L.,
and Sons, Inc.), 176–207. doi:10.1002/0471722138.ch7
Schouten, R. N., et al. (2021). A four-qubit germanium quantum processor. Nature 591,
580–585. doi:10.1038/s41586-021-03332-6 Tacchino, F., Macchiavello, C., Gerace, D., and Bajoni, D. (2019). An artificial neuron
implemented on an actual quantum processor. npj Quantum Inf. 5, 26. doi:10.1038/
Iblisdir, S., Cirio, M., Boada, O., and Brennen, G. K. (2014). Low depth quantum
s41534-019-0140-4
circuits for Ising models. Ann. Phys. 340, 205–251. doi:10.1016/j.aop.2013.11.001

Frontiers in Quantum Science and Technology 15 frontiersin.org