Sampling problems on a Quantum Computer

Maximilian Balthasar Mansky∗, Jonas Nüßlein∗ , David Bucher†, Daniëlle Schuman∗, Sebastian Zielinski∗ ,
and Claudia Linnhoff-Popien∗
∗ LMUMunich
Munich, Germany
Email: [email protected]
† Aqarios GmbH

Munich, Germany
arXiv:2402.16341v1 [quant-ph] 26 Feb 2024

Abstract—Due to the advances in the manufacturing of quan- A novel approach in generating samples is the use of a quan-
tum hardware in the recent years, significant research efforts have tum computer. Quantum computing is a new computational
been directed towards employing quantum methods to solving paradigm with promises of significant computational speedup
problems in various areas of interest. Thus a plethora of novel
quantum methods have been developed in recent years. In this [4]. The technology is structurally different from classical
paper, we provide a survey of quantum sampling methods along- computing and relies on the effects of quantum mechanics
side needed theory and applications of those sampling methods to process information. Instead of representing information
as a starting point for research in this area. This work focuses in through a binary encoding, quantum computing relies on the
particular on Gaussian Boson sampling, quantum Monte Carlo superposition of states for encoding. The states have a physical
methods, quantum variational Monte Carlo, quantum Boltzmann
Machines and quantum Bayesian networks. We strive to provide representation depending on the underlying hardware and
a self-contained overview over the mathematical background, are typically different energy levels in a quantum-mechanical
technical feasibility, applicability for other problems and point system. Superposition means that the state is composed from
out potential areas of future research. a complex-valued superposition of base states. Information
can be changed on the quantum computer through operations
I. I NTRODUCTION
called gates and can be retrieved for further classical process-
Sampling from a population is a well established way to ing and interpretation with a measurement.
learn about the structure of large data or to learn about a The two approaches, quantum computing and sampling
distribution of properties in spaces that are too large for from the unknown, can be combined to shed light on distribu-
an exhaustive examination. Sampling is the act of drawing tions within quantum-mechanical systems that are otherwise
samples from some unknown distribution with the goal of difficult to calculate or to model classically. In the context
learning about the underlying distribution by using statistical of sampling, this means obtaining a sample via classical
reasoning. One samples to generate a smaller population simulation is difficult, but drawing a sample via a quantum
from a distribution because querying all possible elements is computer is easy. Many of the physical systems that one is
prohibitively expensive. For many problems sampling provides interested in form a high-dimensional space whose structure
a reasonable approximation of the total distribution and is an is difficult to model. Even simple composite two-state systems
important approximative method. Sampling is used to build an suffer from a combinatorial explosion in their complete state
understanding of complex systems with limited resources and representation, where the number of dimensions scales as
therefore an important tool in many areas of science. [1]. O(2n ), with n the number of qubits. Since the system size
For reliable statistical reasoning, one needs a sufficient of a quantum computer scales in the same way, a quantum
amount of samples. The number of samples depends on the computer is an appropriate tool to simulate these systems. This
problem at hand and the applicable statistical rigour and is is reflected in the approaches discussed in this paper, drawing
an important consideration when using a sampling approach. inspiration from physical systems.
Samples also need to be independent and identically dis- Classical distributions can also be modeled on a quantum
tributed (i.i.d.), meaning that subsequent samples should have computer, taking advantage of the larger representation space
no relationship with each other and should not affect each in order to represent more complex distributions and then draw
other’s result and that they need to come from the same total samples from them. The expressiveness of these systems can
distribution. In turn, this means that sampling processes can be higher than an equivalent classical formulation, again due
be easily parallelized for fast querying. to the larger available state space.
These benefits of sampling approaches to problem solving, The topic of quantum-based sampling has received signif-
namely fast approximative solutions to complex systems ex- icant attention over the past few years, with experimental
plain their ubiquity and applicability across a broad range of breakthroughs in the size of the experiments [5].
subjects [2], [3]. There is continuous effort to improve current In this work we introduce the state of the art of quantum
theory and provide better approaches to creating independent sampling techniques and applications and provide an overview
samples for individual problems. of their technical feasibility, current technical state and appli-
 B. Sampling on quantum circuits
A quantum computer can similarly be used for a) per-
forming calculations and b) providing a translation between
uniform randomness and a biased distribution. The necessary
randomness is however not easy to control, despite the fact that
a quantum computer is fundamentally a probabilistic machine
[4]. The obvious quantum circuit for generating randomness,
applying a Hadamard gate to each available qubit in the |0i
state, results in uniform independent random qubits when
measuring in the same h0| basis. In order to represent any
uniformly random state of the Hilbert space, one needs an
exponentially deep circuit in the number of qubits [7]. This is
in stark contrast to the classical world, where the randomness
of a bit string is fundamentally the same as the randomness of
an equivalently sized group of individually sampled bits [8].
Fig. 1. Visualization of a sampling process to determine the value of π Suitable construction of the quantum circuit can also gen-
via a sampling process. Black dots represent random (x, y) positions. If the
distance to the origin is smaller or equal to 1 (red area), it is counted towards
erate an appropriate sampling statistic without a source of
the red bin, otherwise it is only counted towards the total number of samples. randomness. This makes use of the probabilistic nature of
The fraction of the count of elements in the red bin to the total number of a quantum computer, where a measured mixed state returns
samples multiplied by 4 approximates π.
a probabilistic value based on the measurement basis. Mixed
states form a statistical ensemble that can be expressed through
a density matrix ρ [4].
cability as a solution to other problems. We present the topic as
For this discussion, we assume a perfect quantum computer
a self-contained as possible to allow a fast translation towards
without extrinsic noise. The noise of current imperfect quan-
a solution.
tum computers is an unwanted source of randomness and will
In the remainder of the paper we provide a background on in most cases bias the desired calculation.
sampling as an explicit approach on classical and quantum
systems. In section III we introduce all current approaches to III. E XAMPLES OF SAMPLING PROBLEMS
quantum-based sampling, including their mathematical basis, We provide a thorough coverage of different quantum com-
technological readiness, open questions and applicability as a puting approaches to sampling problems. We provide a self-
solution path. Lastly we provide a discussion of the presented contained background on the mathematical basis. Technical
methods. feasibility on current hardware is explored as well, especially
for the cases where an implementation has already been
II. BACKGROUND ON SAMPLING TECHNIQUES achieved. We also highlight the applicability as a solution path
to other areas of science and take note of open questions for
A sampling algorithm can be imagined as a machine that future work.
transforms uniform random bits into non-uniformly distributed We start with Gaussian Boson Sampling, an experiment
random bits [6] In the context of sampling from a population, on current hardware that explicitly samples on an unknown
it means taking independent and identical samples from the distribution and is very time-consuming to model classi-
distribution. This simple structure is a starting point into ques- cally. Quantum-Enhanced Markov Chain Monte-Carlo ex-
tions of statistics, probability modeling, conditional inference pands the classical MCMC structure to quantum computers
and much more. The sampling algorithm that transforms the with promises of significant speed-up when investigating quan-
bit strings is often opaque to the questions, if known at all. tum systems. Variational Monte Carlo is especially applicable
to material science and chemistry. Quantum Boltzmann Ma-
A. Sampling on a classical system chines are a sampling-based machine learning model that take
advantage of the quantum computing structure for faster and
A sampling algorithm turns a source of uniform randomness easier sampling from arbitrary distributions. Lastly, quantum
into a non-uniform one. Running the algorithm many times Bayesian Networks are a direct translation of classical models
generates a statistical sample with meaningful insights into for chained probabilities. There are indications that the quan-
the underlying problem. A classic example is sampling π by tum version has significantly higher expressive power.
generating uniform random x, y positions in the unit square.
Here, the sampling algorithm takes uniform randomness – A. Gaussian Boson Sampling
the position – and assigns a new value based on the points’ Boson Sampling is a simplified, non-universal model
distance from the origin. After generating many points, the of quantum computation first introduced by Aaronson and
fraction of points within unit distance of the origin provide an Arkipov [9] in which n Bosons, originally in an input
estimate of the value of π. This is visualized in Figure 1. arrangement k, are scattered by a passive, linear unitary
transformation U into m ≫ n output modes. The Boson Hamilton by adjusting the protocol to account for displaced
Sampling problem consists of producing a fair sample of squeezed states and higher-order photon number contributions.
the output probability distribution P (l|k, U ), where l is the Since the conception of GBS, a number of experimental
output arrangement [10], [11]. Aaronson and Arkipov argue implementations have advanced the study of the protocol.
that the existence of an efficient classical algorithm which Most notably, Zhong et al. [14] used a photonic quantum
accomplishes this given a random transformation U implies computer, Jiuzhang, to execute the GBS protocol with 50
the ability to estimate the permanent of an arbitrary complex indistinguishable single-mode squeezed states and a 100-mode
valued matrix, a problem lying in the class #P [9]. This means ultralow-loss interferometer with full connectivity and sampled
that the problem is in fact hard for classical computers to the output using 100 high-efficiency single-photon detectors.
solve and provides an argument for the superiority of quantum Jiuzhang has a 76-photon coincidence, with an output state
computers over classical ones, as the Boson Sampling problem space dimension of 1030 and outpaced classical state-of-the-
can be solved efficiently by the former, as well as evidence art simulation on supercomputers by a factor of 1014 .
against the Church-Turing thesis [11]. Significant progress has also been achieved in the classical
The primary hurdle in implementing Boson Sampling exper- simulation of GBS. Bulmer et al. [15] present a classical GBS
imentally lies in the fact that currently available single photon simulation method using threshold detectors, which demon-
sources are spontaneous, meaning that the cost of producing strates a nine-orders of magnitude speedup over previous
an input state with exactly n photons grows exponentially classical algorithms that employ photon number–resolving
in n. To combat this issue, Lund et al. [11] suggest using detectors. The novel GBS simulation using threshold detectors
Gaussian states, which can be produced deterministically with was applied to two separate sampling algorithms, a probability
high purity. They describe a Gaussian Boson Sampler, a chain rule method and Metropolis independence sampling,
quantum optical processor consisting of 2-mode squeezed and was able to simulate the GBS protocol with up to 92
input states and a non-adaptive linear optical network, which photons and 100 modes, reducing computation time from 600
produces photon number counting statistics as its output. They million years to a matter of months, a nine-orders of magnitude
argue that in one particular case, namely in the context of improvement. However, such an approach only proves useful
the generalized Boson Sampling problem, such a device can for verification purposes, as state-of-the-art GBS setups, such
efficiently sample distributions which are hard to sample for as Jiuzhang, require only minutes for the same computation.
classical counterparts. Furthermore, Lund et al. contend that
approximate Boson Sampling is also a hard problem, even in B. Quantum-Enhanced Markov Chain Monte Carlo
the generalized case [11]. Markov Chain Monte Carlo (MCMC) is a statistical ap-
In [12], Hamilton et al. formally introduce Gaussian Boson proach for generating random samples from a target probabil-
Sampling (GBS), which, unlike previous protocols involving ity distribution. The basic idea of Markov Chains is to start
Gaussian states, takes full advantage of the Gaussian nature from an initial state and repeatedly jump to new states ac-
of the states. In the GBS setup, Single Mode Squeezed States cording to a transition rule. This allows for a computationally
(SMSS) enter a linear interferometer, and the output patterns inexpensive estimation of various statistics (e.g., mean, vari-
are sampled in the photon number basis. They show that the ance) of the target distribution. In various parts of physics, they
probability of measuring a specific output distribution of a are widely used to estimate observables of statistical systems
Gaussian input state is related to the hafnian, a matrix function whose probability distributions are inaccessible through direct
more general than the permanent which resides in the #P computation [16]. Furthermore, they are used for sampling
complexity class [12], [13]. With this result, Hamilton et al. from Boltzmann distributions [17] (see also Sec. III-D) and
prove that the GBS protocol resides in #P along with the for combinatorial optimization using the simulated annealing
approximate sampling problem. This protocol differs above heuristic [18].
all due to the fact that the sampling matrix describes not only Sampling from the Boltzmann distribution of a classical
the action of the interferometer, but also the shape of the Ising spin-glass at low temperatures is known to be a hard
Gaussian input state. This implies that a coherent superposition problem [19] The probability of a certain classical spin string
of all n-photon patterns from the input can be used and no s = {±1}N is given by
exact input pattern must be heralded as in other protocols. As
1 −βE(s)
such, GBS increases photon generation probability relative to p(s) = e , (1)
standard boson sampling protocols which use single photon Z
Fock states [10]. Furthermore, GBS reduces the size of the where β is the reciprocal temperature given by β −1 = kB T ,
2

sampling space by a factor of NN compared to Scattershot with kB being the Boltzmann factor. The energy of the general
Boson sampling, thereby significantly advancing experimental Ising system [20] is given by
possibilities. While the classification of the Boson Sampling X X
problem with Gaussian states has not definitively been as- E(s) = − Jij si sj − hi si . (2)
i<j
signed a complexity class, it has been shown that the special
case of sampling from a multimode thermal state resides in The partition functionPZ is defined as the sum over the Boltz-
−βE(s)
BP P N P [12]. In [13], Kruse et al. built upon the work of mann factors Z = {s} e . Although the Boltzmann
 • Uniform updates: Now, update candidates s′ are chosen
randomly. In higher temperature simulations, this strategy
works fine and is able to traverse the whole state space
E(s) relatively quickly, which means fast convergence of the
Local Markov chain. However, the acceptance rate drops rapidly
o rm
Unif as the temperature decreases [24], see Fig. 2.
• Cluster updates: In the phase transition between the mag-
Quantum
s netized and disordered state of the Ising model, generally,
ordered patches emerge in the material. For this reason,
Fig. 2. Visualization of different candidate proposal techniques. The local one cluster update proposals have been introduced [25], [26].
does achieve relatively good acceptance rates while not exploring the state Being able to explore the state space rather quickly with
space. Uniform updating tries to explore the state space but struggles with
acceptance since the proposed state has most likely high energy. However,
high acceptance rates, they only work in critical phases
the discussed quantum proposal routine samples states that are far away in of the material and similarly lose their advantage as soon
the state space while having comparable energy, thus, also leading to high as the temperature falls significantly below the critical
acceptance rates.
point.
In general, sampling from low-temperature Ising spin glasses
factors are easy to compute independently, the partition func- with Markov chains is plagued by slow convergence and,
tion is not, because of the exponential number of summands. therefore, long runtimes.
Without specific domain knowledge and advanced analytical To counteract this problem, Layden et al. [24] suggested
methods, the calculation is intractable. a quantum routine to propose updates. The quantum register
The Metropolis-Hastings algorithm is an MCMC method is prepared in the state of the current MCMC chain and
to sample from many distributions, like this Boltzmann dis- undergoes an arbitrarily chosen unitary evolution. The au-
tribution. Essentially, we start at a random spin configuration thors of [24] chose to use a time evolution of the problem
and then propose update steps. These are accepted or rejected Hamiltonian paired with a mixer Hamiltonian known from the
according to the transition probabilities [21], [22]. The accep- Quantum Approximate Operator Ansatz (QAOA) [27], [28].
tance probability of a proposed update s → s′ is given by The joint Hamiltonian can be expressed through
p(s′ )Q(s|s′ )
  X X
′
A(s, s ) = min 1, , (3) Hint = (1 − γ)α E(s) |si hs| − γ σix , (4)
p(s)Q(s′ |s) s i
′
with Q(s|s ) being the transition probability of the system where α is a normalizing factor and γ ∈ [0, 1] controls the
moving from s to s′ . These transition probabilities are cho- strength of the quantum transitions by increasing the effect of
sen such that the samples from this procedure resemble the the mixer.
demanded probability distribution. The output of measuring the state
Commonly, the transition probabilities are chosen symmet-
T
rically, meaning that the probability is the same no matter if Y
exp(−iHint t) |si ≈ exp(−iHint t/T ) |si (5)
the state is moving from s → s′ or the other way round. In the
acceptance probability (3) the expression then reduces to just is the proposed spin configuration, where the Hamiltonian
a fraction of the state probabilities. Furthermore, note that due is time-evolved using Trotterization. Important to note is
to that ratio, we never need to calculate the partition function the symmetry of the transition probability | hs| U |s′ i | =
explicitly. A metric to measure how well the Markov chain | hs′ | U |si | using this approach. Yet, there are still two free
samples through the probability distribution is the acceptance parameters γ and t that need to be set. To circumvent the
ratio α, i.e., the number of accepted proposals compared to need for tuning them, [24] chose to sample the parameters of
the number of all trials. Theoretical investigations suggest an each Monte Carlo iteration randomly, decreasing the bias of
optimal value of α = 0.243 for random walk problems [23]. a constant setting. The achieved effect, compared to local and
However, most importantly, the acceptance ratio should not uniform updating, is visualized in Fig. 2.
drop close to zero in order to maintain a good sampling quality. The authors of [24] have found a significant improvement
Small acceptance rates are an indication that the chain gets in convergence speed when compared to local and uniform
stuck in a trench in the energy landscape. updating procedures. Furthermore, with clever error mitigation
Typically, update-proposal strategies comprise in use, they have also been able to observe a performance
• Local Spin flips: Here, just a random spin is chosen to increase when running on quantum hardware. The gain was
be flipped. In many scenarios, mainly where the energy not as big as the simulations suggested, nevertheless, faster
landscape is very rugged, single spin flips cannot get the convergence than local and uniform updates has been achieved.
chain out of deep trenches, also depicted in Fig. 2. Since Remarkably, the proposed algorithm never miscalculates a
only a limited number of proposals are available, if all quantity based on quantum imperfections since the precise
of them are unlikely to be accepted due to higher energy, values are computed using the Metropolis algorithm. The
the chain gets stuck, and the acceptance rate drops. quantum routine only produces update proposals, which still
need to be accepted in order to be included to the computation. Hamiltonian efficiently using MC sampling:
The quantum routine only helps with increasing the acceptance
E(Ω) = hψΩ | H |ψΩ i (6)
rate and the exploration speed, leading to faster convergence. X X
∗
How the authors [24] deal with the parameters in the time = ψΩ (x)ψΩ (x′ ) hx| Ha |x′ i (7)
evolution surely removes bias but is not the ideal setting. Thus a {x},{x′ }

more effort can be devoted to parameter tuning. Furthermore, N MC

X 1 X
performing a time evolution on a sample is heuristically chosen = |ψΩ (x)|2 EΩ
loc
(x) ≈ loc
EΩ (xi ), (8)
NMC
and has no physical reason. Thus other quantum proposal {x} i
finding methods can be investigated, like a reverse quantum where xi is sampled from the probability distribution
annealing ansatz or a Quantum Phase Estimation (QPE)-based |ψΩ (x)|2 . This is only possible if EΩ loc
(x) is efficiently com-
method, as already mentioned in [24]. putable on classical computers,
X X ψΩ (x′ ) X hx| Ha |ψΩ i
E loc = hx| Ha |x′ i = . (9)
C. Variational Monte Carlo a
ψΩ (x) a
ψΩ (x)
x ′

Typically, this is the case for a sufficiently well-behaved ansatz

One of the most demanding challenges in modern physics, ψΩ [29].
chemistry, and material sciences is the description and simula- Usually, the samples are drawn from an MCMC simulation,
tion of many-body quantum systems and their ground state at but this simulation is prone to similar issues as previously dis-
zero or low temperatures [29]. Tensor network methods, like cussed in Sec. III-B. However, as the probability distribution
density matrix renormalization group (DMRG) [30], projected is of an entangled quantum state and not of a classical Ising
entangled pair states (PEPS) [31] or multi-scale renormaliza- spin-glass, we cannot utilize the previously presented method
tion ansatz (MERA) [32], have proven incredibly valuable for enhancing the MCMC. Instead, we propose to prepare the
for computing observables for many classes of materials and quantum state directly on the QC and measure in the z-basis,
chemical systems. Despite their success, especially in one- similar to [38]. This automatically produces samples that are
dimensional and, with limitations, in two-dimensional ma- distributed according to |ψΩ (x)|2 . However, for this approach
terials [33], tensor networks reach practical computational to be viable, we need a circuit UΩ comprised of a polynomial
boundaries in higher-dimensional lattices and highly entangled number of gates, such that UΩ |0i = |ψΩ i.
quantum states [32]. Furthermore, in order to train the parameters of the wave
On the other hand, Quantum Monte Carlo Methods (QMC) function ansatz, we need an efficient method of computing the
also emerged as a direct calculation of quantum mechanical gradient with respect to the parameters. The gradient must be
properties through approximation with Monte Carlo sampling. representable as a set of local Hermitian operators Oλ , where
Due to the infamous sign problem, they struggle to simulate
∂
fermionic and frustrated systems [34]–[37], such as high- Oλ (x) = log ψΩ (x) (10)
temperature superconductivity [35] or electronic band structure ∂Ωλ
computations [38]. Both are extremely interesting systems, not From that follows the energy gradient instantly, following [41]:
only for theoretical investigation but also for practical use. ∂E(Ω)
Recently, an QC-assisted QMC method has been proposed to = hOλ† Hi − hOλ† ihHi. (11)
∂Ωλ
control the effect of sign problem [39]. Notably, it achieves
comparable results to state-of-the-art classical methods. Finally, gradient-descent-based optimization can be pursued
after the gradient has been approximated with MC using the
Variational Monte Carlo (VMC) [29], [40] methods evade
same samples that already have been obtained for computing
the sign problem by using parametrized wave function ansätze.
the energy [41].
In comparison to the previously discussed unbiased ansätze,
Typically, physically motivated wave function ansätze are
VMC is inherently biased [32], as the wave function ansatz
used [40]–[42], but recently efforts have been made to unbias
has to be meticulously chosen to fit the problem one wants to
the ansatz through universal function approximators, also
solve. The parameters of the quantum wave function ansatz are
known as neural networks [29], [43], [44]. As neural networks
subsequently optimized using an MC evaluation of the energy
are well known for various other applications, they are a fitting
of the system.
choice, especially since classical computational acceleration is
Given a local Hamiltonian H, i.e., a sum of Hamiltonians possible.
Ha who act only on a subspace of complete system Hilbert At this point, it makes sense to draw a dividing line to
space Ha ⊂ H with dim Ha ∈ O(1). We typically consider the well-known Variational Quantum Eigensolver (VQE) [45],
nearest-neighbor Hamiltonians, whose local Hamiltonians act [46], which is likewise used to find the ground state of quan-
on two spins or qubits. tum many-body systems. Both methods rely on parametrized
P
Let |ψΩ i = {x} ψΩ (x) |xi be the wave function ansatz, circuits, which aim to encode a physical quantum state. VQE
with the parameters gathered in Ω and x ∈ {−1, +1}N . We computes the energy directly on the Quantum Device and has
can evaluate the expectation value of a local observable or to perform multiple measurements to obtain the gradient with
respect to the parameters [46]. On the other hand, the VMC x2
x1 h0 h1 h2 h3 h4
ansatz is only used for sampling, energy and gradients can x3
are computed classically. Furthermore, one can use higher x90
order methods, like Stochastic Reconfiguration [42] for faster x4
training, which are otherwise only possible with far more x8
circuit evaluations [47]. However, these benefits come with x5
the drawback that the ansatz wave functions are required to x6 x7 v0 v1 v2 v3 v4 v5
be efficiently prepared on the QC and simultaneously be effi-
ciently evaluated on classical computers. Restricted Boltzmann Fig. 3. Drawing of the Boltzmann Machine (left) and Restricted Boltzmann
Machine (RBM) based ansätze could be a fitting candidate, as Machine (right). Connections between the nodes indicate the Boltzmann
weights Jij , each node itself contains a bias weight hi . In the RBM, the nodes
they have been used for VMC [29] and polynomial size RBM are split up into two groups: Hidden and visible units. It has no connections
wave function construction circuits exist [38]. between two visible (hidden) units.
Finally, it is an open question whether it is beneficial to
restrict variational QC to states that are efficiently produced
with classical computers in order to make use of the mentioned To achieve better classical performance, one typically di-
advantages. Furthermore –– except for the RBM quantum state vides the in- and output units of the BM into two separate
further discussed in the next section –– the authors of this clusters, called visible and hidden units [50], denoted with v
paper are not aware of other suitable ansätze. and h, which can be seen in the right-hand drawing of Fig. 3.
When disallowing any connections between units of the same
D. Quantum Boltzmann Machines cluster, the emerging Restricted Boltzmann Machine (RBM)
Boltzmann Machines (BM) are generative machine learning gains some advantages. First, it is easier to sample using Gibbs
models that have been introduced by Hinton et al. [48]. sampling, another MCMC sampling method that updates hid-
By turning the interactions in a classical Ising spin-glass den and visible units in an alternating fashion [49], [56].
into parameters {J, h}, the Boltzmann distribution of the Furthermore, training using Contrastive Divergence learning
parameterized Ising energy becomes a learnable probability becomes way more efficient [57]. As the naming suggests,
distribution over a discrete domain [49]. The spin variables the visible units are the ones exposed to the outside, i.e.,
si ∈ {−1, 1} are the discrete inputs (and outputs) to the they need to have the same dimensionality as the input data.
network, see the left-hand side of Fig. 3, while the learnable Hidden units, on the other hand, are traced over and only
probability distribution has the formula: referenced internally. The probability distribution over visible
units is defined as [49]
1 −βEJ,h (x)
p(x) = e . (12) 1 X −βEJ,h ({v,h})
Z p(v) = e (13)
Z
{h}
Here, the inverse temperature β serves as a regularization
parameter, governing the ruggedness of the resulting proba- where J incorporates the required connectivity restrictions.
bility distribution. Z and EJ,h (x) are defined as described Despite the initial drawback of reduced expressive power due
in Sec. III-B. To learn the probability distribution of a set to the limitations in connectivity, one can still adapt the RBM’s
of data points, one can tune the parameters using gradient expressiveness by choosing an appropriate number of hidden
descent. This makes them applicable in a variety of Machine units [58]. Nevertheless, hidden units can likewise be added
Learning (ML) tasks, including unsupervised learning [49], to fully connected BMs.
[50], supervised learning [51] and even reinforcement learning Quantum Computing applications for Boltzmann machines
problems [52]. As such, application areas range from image are twofold: first, encoding the Boltzmann probability in
recognition and denoising [53] to cybersecurity [54] and quantum states, and second, creating a modified BM, called
medical diagnostics [55]. Quantum Boltzmann Machine (QBM), that adds a transverse
For training, the Kullback-Leibler-Divergence between the field to the BM’s Ising energy function.
BM’s distribution and the dataset’s distribution is chosen as the In thePfirst approach, we aim to prepare a quantum state
loss function. In order to obtain the gradients with respect to |ψi = {x} ψ(x) |xi on the computer. Wiebe et al. [59]
the parameters of the Boltzmann Machine, one needs sample assembled a circuit for preparing such states for RBMs that
from both these probability distributions [49]. Done classically, is based on a mean-field approximation with followed im-
exact sampling from these distributions can become exceed- portance sampling. They observed a quadratic reduction of
ingly time-consuming since one has to compute the partition training data access during the optimization process. Further
function. Similar to Sec. III-B, Metropolis sampling can be state preparation methods include an approximate imaginary
utilized to alleviate this issue. In fact, the BM problem is time evolution method based ansatz called VarQITE [60], and
exactly the same as the one previously discussed. Although Quantum Annealing (QA)-based approaches [61], [62]. The
proposed in [24], no experiments have so far been con- QA approaches need precise tuning of the inverse temperature
ducted on applying the quantum-enhanced method described parameter [63]. Otherwise, the final distribution will not be
in Sec. III-B to sampling from a BM. close to a Boltzmann distribution [61].
 Also belonging to the state preparation methods, but with q5 |ψ5 i U5
the alternative purpose of finding the ground state of specific
Hamiltonians are the VMC methods, where we can now close
the circle to Sec. III-D. Here, Xia et al. [38] proposed an q4 |ψ4 i
RBM quantum state circuit for electronic structure calcula-
tions. Instead of considering just amplitude information, they, q3 |ψ3 i U3
and related methods [64], [65], additionally include sign or
phase information through a separate network, i.e., |ψi = q2
P p |ψ2 i
{x} s(x) p(x) |xi. The circuit construction works similarly
to [59], however, no mean-field initialization is present. Yet,
simulations of the RBM approach produce good results on q1 |ψ1 i
small molecules, like H2 , LiH or H2 O, outperforming the
Hartree-Fock outcome. Despite having quadratic space and Fig. 4. A Bayesian network and its corresponding quantum circuit. Each
depth requirements, the NISQ applicability is criticized be- connection in the graph is translated to a control wire. The influence of
each parent node to its children is modeled through unitary gates on the
cause of a huge constant overhead in circuit construction and corresponding qubit. Adapted from [80]
the required ability to perform in-circuit measurements, which
is not necessarily given in NISQ devices [66]. Furthermore,
the expressive power of the sign part is also questioned, but This structure can be modelled through directed acyclic graphs
this is also a known issue in VMC literature [67]–[69], even (DAG), where each probability pi is assigned to a node and the
with different ansätze for the phase information. parental relations are mapped to edges between the nodes. The
The second approach augments the original Ising Hamil- directionality of the edges is a consequence of the parent-child
tonian with an additional transverse field into the spin-glass relationship between the probabilities, which also prohibits
Hamiltonian. The Quantum Boltzmann machine leverages cyclic relationships inside the network. Finding the correct
quantum effects to encode distributions that are believed to Bayesian network for a given set of data is generally difficult
be more expressive than ordinary BMs [70] (NP-complete in the general case [77]), but there are a number
X X X of approximate algorithms and heuristics that improve training
H(J, h, Γ) = − Jij σiz σjz − hi σiz − Γi σix . (14)
time [78].
ij i i
These networks can in principle be mapped to a quantum
Here, σix , σiz
denote the Pauli x- and z-operators at qubit computer [79], [80]. There is a direct mapping between nodes
i. In numerical studies, assuming perfect state preparation, and edges of the network and the operators on a quantum
QBMs seem to outperform classical BMs [70]. Nonetheless, computer [80, fig: 9]. Using k-controlled unitaries, where
one major advantage of BMs, that is direct access to the the controls to the ith qubit correspond to the edges on
analytical gradient, is unfortunately lost in QBMs due to the the equivalent graph formulation, a state is prepared such
non-commuting terms in the Hamiltonian. that |hxi |U (parents of xi )|xi i|2 = p(xi |parents of xi ). The
In conclusion, BMs on quantum computers pave interesting correspondence between the Bayesian network graph and its
directions. Advanced circuit generation for machine learning, the quantum circuit is shown in figure 4.
as well as VMC application, needs further investigation as The approach differs from the ones discussed in section
current methods are not satisfactory with regards to practical III-B in that rather than using a learning framework where the
applicability [66]. For QBMs, it is the question if they can system is changed to approximate some desired solution, the
withstand normal BMs in practical circumstances. A general Bayesian networks are constructed rather schematically as a
performance benefit has to our knowledge so far only been translation of a known system or an ansatz expression of the
demonstrated on a few small instances [70], [71]. DAG.
E. Bayesian networks Structurally, these models should be expandable to more
general sequential quantum models. The construction on the
Bayesian networks are used to represent chained prob-
quantum computer may be able to model any sequential
abilities [72]. They are used to represent certainty about
structure. Sequential and ordered structures are ubiquitous
knowledge and to model information insecurity. As they can be
across applications, from biological systems to natural lan-
used to calculate difficult probability distributions, they have
guage processing.
found wide application across a number of areas, including
Quantum natural language processing is already in develop-
agriculture [73], data mining [74], meteorology [75], software
quality control [76] and many other areas with distributed ment [81], [82]. Its structure is at first dissimilar to the chained
probability approach presented here but in practice follows
influence factors.
Mathematically, the expected probability over each edge of much the same structure of distributed meanings translated to
the network is given by quantum circuits that are then chained together to form a quan-
n
tum computation [83]. Bridging between the two approaches
Y can bring a meaningful contribution in understanding for both
p(x1 , . . . , xn ) = p(xi |parents of xi ) (15)
fields.
i=1
 Biological systems, especially where they model complex would perform in the place of the time-evolution remains an
interactions, are another field of application where Bayesian open research question.
networks find wide application [84]–[86]. Leveraging the The third approach considers Variational Monte Carlo struc-
additional structure of quantum Bayesian networks can lead tures. Here, parametrized wave function ansätze are used, ei-
to additional insights into the structure of relationships be- ther fitting the physical problem to be solved or represented by
tween different influence factors. This can be used to model a neural network. Observables and quantities can subsequently
disease vectors [87], disease recognition [88], neuroscientific be computed by sampling from the probability distribution
discoveries [89] or the analysis of genomes [90]. of the wave function, i.e., the absolute square of the wave
Beyond static Bayesian networks, dynamic Bayesian net- function factors. VMC has the drawback that the ansätze need
works that reflect a change of values and possibly structure to be computed both classically and on the quantum devices.
of the Bayesian network [91]–[93] may also be applicable To date, only a single ansatz has been used. Here, of course,
to quantum ansatz presented here. Their expressive power is research for new ansätze can be pursued. Yet it is still unclear
significantly higher and can be suited to the higher dimensional whether VMC are preferable to similar approaches such as
space of quantum computers. VQE.
Similar to the previous approach, we cover Boltzmann
IV. D ISCUSSION Machines as wave function ansätze and as generative machine
In this paper, we review sampling applications as an im- learning models. BMs need sampling from both a model
portant application area for quantum computing. We start by and a data Boltzmann distribution in order to optimize their
defining sampling as a means of approximating complicated parameters to make the former match the latter as closely
distributions with queries from that distribution, and estab- as possible using gradient descent. Here, quantum computers
lishing the ubiquity of sampling problems across a broad are employed to model the Boltzmann distribution such that
range of subject areas. We note that a quantum computer sampling naturally becomes measuring in the z-basis. We
can be used as a sample source and thereby gain insights cover several methods in BM wave function construction,
into quantum-mechanical processes and distributions. After both using circuit-based constructions and quantum anneal-
introducing a sampling algorithm as a transformation of uni- ing. A handful of indications that faster training and better
formly distributed random bits into bits distributed according approximation can be achieved have been given. Yet, circuit
to a non-uniform probability distribution, we then briefly construction is not considered NISQ-friendly, despite having
compare the implementation of random number generation and quadratic complexity in space and time requirements. Further-
sampling algorithms in the classical versus the quantum realm, more, augmenting the Boltzmann Distribution by including
establishing that quantum computers need a different approach a transverse field Hamiltonian to the Ising Energy has been
due to their larger internal state space. discussed. Numerical studies with perfect circuit construction
Subsequently, we describe how the previously covered demonstrate higher expressibility, but a general advantage has
properties of quantum computing can be utilized to improve not been proven.
some well-known solutions approaches to sampling problems. Lastly, we describe Bayesian networks, being directed
The first of these solution approaches treats the problem of acyclic graphs that describe chained sets of probabilities,
Gaussian Boson Sampling, in which a coherent superposition where the probability of a child node depends directly on that
of input states is to be transformed into a potentially high- of his parents. These networks can be directly mapped to quan-
dimensional superposition following a multi-modal probability tum circuits, the probabilities being represented by the qubits
distribution, from which a fair sample is to be taken. While the and the edges by the control wires between them that steer
problem itself has not definitively been assigned a complexity unitaries acting on them. The usage of these quantum Bayesian
class, current protocols solving it using a photonic quantum networks might provide additional insights into the structure
computer have been proven to be in the “hard” complexity of relationships be- tween different influence factors and may
class #P and can be used to perform a probability chain rule be suitable to accommodate the higher expressive power of
method or Metropolis independence sampling within minutes, dynamic Bayesian networks using higher dimensional space
while their classical simulation takes months. of quantum computers.
In the second approach we explain quantum-enhanced To summarize, what most of these approaches have in
Markov Chain Monte Carlo methods, specifically a common is that they use the inherent randomness of a quantum
Metropolis-Hastings algorithm, which can be used, e.g., computer expressed by its ability to store basis states in a
for combinatorial optimization or to sample from probability superposition that collapses to each of them with a certain
distributions that are classically hard to access such as the probability when measured, as well as the high dimensionality
Boltzmann distribution. The quantum-enhanced version of of its state spaces, to model complicated, possibly classically
this algorithm uses time-evolution for the step of proposing intractable probability distributions, and subsequently sample
candidates for a state update, thereby achieving faster from them.
convergence than the classical version of the algorithm. However, for most of these approaches, the hardness in
Whether the tuning of the free parameters could further terms of theoretical complexity is unknown, as well as whether
improve the performance and how other quantum circuits certain quantum approaches might be working better for them
than others, and, if so, why. In a lot of cases, it is not even clear [12] C. S. Hamilton, R. Kruse, L. Sansoni, S. Barkhofen,
why exactly the presented quantum approach outperforms the C. Silberhorn, and I. Jex, “Gaussian boson sampling,” Physical
Review Letters, vol. 119, no. 17, oct 2017. [Online]. Available:
comparable classical approach from a theoretical perspective. https://doi.org/10.1103%2Fphysrevlett.119.170501
Furthermore, many details about how to practically implement [13] R. Kruse, C. S. Hamilton, L. Sansoni, S. Barkhofen, C. Silberhorn, and
the approaches on near-term quantum computers in an optimal, I. Jex, “Detailed study of gaussian boson sampling,” Physical
Review A, vol. 100, no. 3, sep 2019. [Online]. Available:
efficient manner, i.a. the optimal way to tune quantum-specific https://doi.org/10.1103%2Fphysreva.100.032326
hyperparameters, are not entirely clear yet. By presenting [14] H.-S. Zhong, H. Wang, Y.-H. Deng, M.-C. Chen, L.-C. Peng, Y.-H.
these open research questions, this paper lays the groundwork Luo, J. Qin, D. Wu, X. Ding, Y. Hu, P. Hu, X.-Y. Yang, W.-J. Zhang,
H. Li, Y. Li, X. Jiang, L. Gan, G. Yang, L. You, Z. Wang, L. Li, N.-L.
for further exploration regarding the applicability of quantum Liu, C.-Y. Lu, and J.-W. Pan, “Quantum computational advantage using
computing in sampling applications, as we and our readers, photons,” Science, vol. 370, no. 6523, pp. 1460–1463, 2020. [Online].
now more aware of them, can strive to answer them in our Available: https://www.science.org/doi/abs/10.1126/science.abe8770
[15] J. F. F. Bulmer, B. A. Bell, R. S. Chadwick, A. E. Jones,
future work. D. Moise, A. Rigazzi, J. Thorbecke, U.-U. Haus, T. V. Vaerenbergh,
R. B. Patel, I. A. Walmsley, and A. Laing, “The boundary
ACKNOWLEDGEMENTS for quantum advantage in gaussian boson sampling,” Science
Advances, vol. 8, no. 4, p. eabl9236, 2022. [Online]. Available:
The authors acknowledge funding from the German Fed- https://www.science.org/doi/abs/10.1126/sciadv.abl9236
eral Ministry of Education and Research under the funding [16] K. Huang, Statistical Mechanics, 2nd ed. John Wiley & Sons, 1987.
program ”Förderprogramm Quantentechnologien – von den [17] D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, “A
learning algorithm for boltzmann machines,” Cognitive Science,
Grundlagen zum Markt” (funding program quantum tech- vol. 9, no. 1, pp. 147–169, Jan. 1985. [Online]. Available:
nologies – from basic research to market), project BAIQO, https://www.sciencedirect.com/science/article/pii/S0364021385800124
13N16089 (Maximilian Balthasar Mansky) and project Q- [18] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi,
“Optimization by Simulated Annealing,” Science, vol. 220,
Grid, 13N16177 (David Bucher), as well as from the German no. 4598, pp. 671–680, May 1983. [Online]. Available:
Federal Ministry for Economic Affairs and Climate Action, https://www.science.org/doi/10.1126/science.220.4598.671
project PlanQK, 01MK20005I (Daniëlle Schuman). [19] F. Barahona, “On the computational complexity of Ising
spin glass models,” Journal of Physics A: Mathematical and
R EFERENCES General, vol. 15, no. 10, p. 3241, 1982. [Online]. Available:
https://dx.doi.org/10.1088/0305-4470/15/10/028
[1] R. S. Witte and J. S. Witte, Statistics. John Wiley & Sons, 2017. [20] E. Ising, “Beitrag zur theorie des ferromagnetismus,” Zeitschrift für
[2] A. Agresti and C. Franklin, “The art and science of learning from data,” Physik, vol. 31, no. 1, pp. 253–258, 1925. [Online]. Available:
Upper Saddle River, New Jersey, vol. 88, 2007. https://doi.org/10.1007/BF02980577
[3] S. K. Thompson, Sampling. John Wiley & Sons, 2012, vol. 755. [21] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and
[4] M. A. Nielsen and I. L. Chuang, Quantum computation and quantum E. Teller, “Equation of State Calculations by Fast Computing Machines,”
information, 10th ed. Cambridge ; New York: Cambridge University The Journal of Chemical Physics, vol. 21, no. 6, pp. 1087–1092, 1953.
Press, 2010. [Online]. Available: https://aip.scitation.org/doi/10.1063/1.1699114
[5] L. S. Madsen, F. Laudenbach, M. F. Askarani, F. Rortais, T. Vincent, [22] W. K. Hastings, “Monte Carlo sampling methods using Markov chains
J. F. F. Bulmer, F. M. Miatto, L. Neuhaus, L. G. Helt, M. J. and their applications,” Biometrika, vol. 57, no. 1, pp. 97–109, 1970.
Collins, A. E. Lita, T. Gerrits, S. W. Nam, V. D. Vaidya, [Online]. Available: https://doi.org/10.1093/biomet/57.1.97
M. Menotti, I. Dhand, Z. Vernon, N. Quesada, and J. Lavoie, [23] A. Gelman, W. R. Gilks, and G. O. Roberts, “Weak convergence
“Quantum computational advantage with a programmable photonic and optimal scaling of random walk Metropolis algorithms,”
processor,” Nature, vol. 606, no. 7912, pp. 75–81, Jun. 2022, The Annals of Applied Probability, vol. 7, no. 1, Feb. 1997,
number: 7912 Publisher: Nature Publishing Group. [Online]. Available: publisher: Institute of Mathematical Statistics. [Online]. Available:
https://www.nature.com/articles/s41586-022-04725-x https://doi.org/10.1214/aoap/1034625254
[6] A. P. Lund, M. J. Bremner, and T. C. Ralph, “Quantum [24] D. Layden, G. Mazzola, R. V. Mishmash, M. Motta, P. Wocjan, J.-S.
sampling problems, BosonSampling and quantum supremacy,” npj Kim, and S. Sheldon. Quantum-enhanced Markov chain Monte Carlo.
Quantum Information, vol. 3, no. 1, pp. 1–8, Apr. 2017, [Online]. Available: http://arxiv.org/abs/2203.12497
number: 1 Publisher: Nature Publishing Group. [Online]. Available: [25] U. Wolff, “Collective Monte Carlo Updating for Spin Systems,”
https://www.nature.com/articles/s41534-017-0018-2 Physical Review Letters, vol. 62, no. 4, pp. 361–364, 1989. [Online].
[7] J. Haferkamp, P. Faist, N. B. T. Kothakonda, J. Eisert, and Available: https://link.aps.org/doi/10.1103/PhysRevLett.62.361
N. Yunger Halpern, “Linear growth of quantum circuit complexity,” [26] J. Houdayer, “A Cluster Monte Carlo Algorithm for 2-Dimensional Spin
Nature Physics, vol. 18, no. 5, pp. 528–532, May 2022. [Online]. Glasses,” The European Physical Journal B, vol. 22, no. 4, pp. 479–484,
Available: https://www.nature.com/articles/s41567-022-01539-6 2001. [Online]. Available: http://arxiv.org/abs/cond-mat/0101116
[8] M. Herrero-Collantes and J. C. Garcia-Escartin, “Quantum random [27] E. Farhi, J. Goldstone, and S. Gutmann. A Quantum Approximate Opti-
number generators,” Reviews of Modern Physics, vol. 89, no. 1, p. mization Algorithm. [Online]. Available: http://arxiv.org/abs/1411.4028
015004, Feb. 2017, publisher: American Physical Society. [Online]. [28] S. Hadfield, Z. Wang, B. O’Gorman, E. G. Rieffel, D. Venturelli, and
Available: https://link.aps.org/doi/10.1103/RevModPhys.89.015004 R. Biswas, “From the Quantum Approximate Optimization Algorithm
[9] S. Aaronson and A. Arkhipov, “The computational complexity of to a Quantum Alternating Operator Ansatz,” Algorithms, vol. 12, no. 2,
linear optics,” ser. STOC ’11. New York, NY, USA: Association p. 34, 2019. [Online]. Available: http://arxiv.org/abs/1709.03489
for Computing Machinery, 2011, p. 333–342. [Online]. Available: [29] G. Carleo and M. Troyer, “Solving the Quantum Many-Body Problem
https://doi.org/10.1145/1993636.1993682 with Artificial Neural Networks,” Science, vol. 355, no. 6325, pp.
[10] H.-S. Zhong, L.-C. Peng, Y. Li, Y. Hu, W. Li, J. Qin, D. Wu, W. Zhang, 602–606, Feb. 2017, arXiv:1606.02318 [cond-mat, physics:quant-ph].
H. Li, L. Zhang, Z. Wang, L. You, X. Jiang, L. Li, N.-L. Liu, [Online]. Available: http://arxiv.org/abs/1606.02318
J. P. Dowling, C.-Y. Lu, and J.-W. Pan, “Experimental gaussian boson [30] U. Schollwöck, “The density-matrix renormalization group in the age of
sampling,” Science Bulletin, vol. 64, no. 8, pp. 511–515, apr 2019. matrix product states,” Annals of Physics, vol. 326, no. 1, pp. 96–192, jan
[Online]. Available: https://doi.org/10.1016%2Fj.scib.2019.04.007 2011. [Online]. Available: https://doi.org/10.1016%2Fj.aop.2010.09.012
[11] A. Lund, A. Laing, S. Rahimi-Keshari, T. Rudolph, J. O’Brien, [31] R. Orús, “A practical introduction to tensor networks: Matrix
and T. Ralph, “Boson sampling from a gaussian state,” Physical product states and projected entangled pair states,” Annals of
Review Letters, vol. 113, no. 10, sep 2014. [Online]. Available: Physics, vol. 349, pp. 117–158, Oct. 2014. [Online]. Available:
https://doi.org/10.1103%2Fphysrevlett.113.100502 https://www.sciencedirect.com/science/article/pii/S0003491614001596
[32] Z.-A. Jia, B. Yi, R. Zhai, Y.-C. Wu, G.-C. Guo, and microstructure of cognition, vol. 1: foundations. Cambridge, MA, USA:
G.-P. Guo, “Quantum Neural Network States: A Brief MIT Press, Jan. 1986, pp. 282–317.
Review of Methods and Applications,” Advanced Quantum [49] G. E. Hinton, “A Practical Guide to Training Restricted Boltzmann
Technologies, vol. 2, no. 7-8, p. 1800077, 2019, eprint: Machines,” in Neural Networks: Tricks of the Trade: Second Edition,
https://onlinelibrary.wiley.com/doi/pdf/10.1002/qute.201800077. [On- ser. Lecture Notes in Computer Science, G. Montavon, G. B. Orr, and
line]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/qute.201800077 K.-R. Müller, Eds. Berlin, Heidelberg: Springer, 2012, pp. 599–619.
[33] J. Eisert, M. Cramer, and M. B. Plenio, “Colloquium: Area laws for the [Online]. Available: https://doi.org/10.1007/978-3-642-35289-8 32
entanglement entropy,” Reviews of Modern Physics, vol. 82, no. 1, pp. [50] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for
277–306, Feb. 2010, publisher: American Physical Society. [Online]. deep belief nets,” Neural Computation, vol. 18, no. 7, pp. 1527–1554,
Available: https://link.aps.org/doi/10.1103/RevModPhys.82.277 Jul. 2006.
[34] E. Y. Loh, J. E. Gubernatis, R. T. Scalettar, S. R. White, D. J. [51] H. Larochelle, M. Mandel, R. Pascanu, and Y. Bengio, “Learning
Scalapino, and R. L. Sugar, “Sign problem in the numerical simulation algorithms for the classification restricted Boltzmann machine,” The
of many-electron systems,” Physical Review B, vol. 41, no. 13, pp. Journal of Machine Learning Research, vol. 13, no. 1, pp. 643–669,
9301–9307, May 1990, publisher: American Physical Society. [Online]. Mar. 2012.
Available: https://link.aps.org/doi/10.1103/PhysRevB.41.9301 [52] D. Crawford, A. Levit, N. Ghadermarzy, J. S. Oberoi, and
[35] Z.-X. Li and H. Yao, “Sign-Problem-Free Fermionic Quantum P. Ronagh, “Reinforcement Learning Using Quantum Boltzmann
Monte Carlo: Developments and Applications,” Annual Review Machines,” Jan. 2019, arXiv:1612.05695 [quant-ph]. [Online]. Available:
of Condensed Matter Physics, vol. 10, no. 1, pp. 337–356, http://arxiv.org/abs/1612.05695
Mar. 2019, arXiv:1805.08219 [cond-mat]. [Online]. Available: [53] Y. Tang, R. Salakhutdinov, and G. Hinton, “Robust Boltzmann Machines
http://arxiv.org/abs/1805.08219 for recognition and denoising,” in 2012 IEEE Conference on Computer
[36] M. Troyer and U.-J. Wiese, “Computational Complexity and Vision and Pattern Recognition, Jun. 2012, pp. 2264–2271, iSSN: 1063-
Fundamental Limitations to Fermionic Quantum Monte Carlo 6919.
Simulations,” Physical Review Letters, vol. 94, no. 17, p. 170201, [54] I. H. Sarker, “Deep Cybersecurity: A Comprehensive Overview from
May 2005, publisher: American Physical Society. [Online]. Available: Neural Network and Deep Learning Perspective,” SN Computer
https://link.aps.org/doi/10.1103/PhysRevLett.94.170201 Science, vol. 2, no. 3, p. 154, Mar. 2021. [Online]. Available:
[37] D. Ceperley, G. V. Chester, and M. H. Kalos, “Monte Carlo simulation https://doi.org/10.1007/s42979-021-00535-6
of a many-fermion study,” Physical Review B, vol. 16, no. 7, pp. [55] G. van Tulder and M. de Bruijne, “Learning Features for Tissue
3081–3099, Oct. 1977, publisher: American Physical Society. [Online]. Classification with the Classification Restricted Boltzmann Machine,”
Available: https://link.aps.org/doi/10.1103/PhysRevB.16.3081 in Medical Computer Vision: Algorithms for Big Data, ser. Lecture
[38] R. Xia and S. Kais, “Quantum Machine Learning for Electronic Notes in Computer Science, B. Menze, G. Langs, A. Montillo, M. Kelm,
Structure Calculations,” Nature Communications, vol. 9, no. 1, p. 4195, H. Müller, S. Zhang, W. T. Cai, and D. Metaxas, Eds. Cham: Springer
2018. [Online]. Available: http://arxiv.org/abs/1803.10296 International Publishing, 2014, pp. 47–58.
[39] W. J. Huggins, B. A. O’Gorman, N. C. Rubin, D. R. Reichman, [56] S. Geman and D. Geman, “Stochastic Relaxation, Gibbs Distributions,
R. Babbush, and J. Lee, “Unbiasing fermionic quantum Monte Carlo and the Bayesian Restoration of Images,” IEEE Transactions on Pattern
with a quantum computer,” Nature, vol. 603, no. 7901, pp. 416–420, Analysis and Machine Intelligence, vol. PAMI-6, no. 6, pp. 721–741,
Mar. 2022, number: 7901 Publisher: Nature Publishing Group. [Online]. Nov. 1984, conference Name: IEEE Transactions on Pattern Analysis
Available: https://www.nature.com/articles/s41586-021-04351-z and Machine Intelligence.
[57] M. A. Carreira-Perpinan and G. Hinton, “On contrastive divergence
[40] W. L. McMillan, “Ground State of Liquid ${\mathrm{He}}ˆ{4}$,”
learning,” in Proceedings of the tenth international workshop on
Physical Review, vol. 138, no. 2A, pp. A442–A451, Apr.
artificial intelligence and statistics, ser. Proceedings of machine
1965, publisher: American Physical Society. [Online]. Available:
learning research, R. G. Cowell and Z. Ghahramani, Eds.,
https://link.aps.org/doi/10.1103/PhysRev.138.A442
vol. R5. PMLR, Jan. 2005, pp. 33–40. [Online]. Available:
[41] S. Sorella, M. Casula, and D. Rocca, “Weak binding between https://proceedings.mlr.press/r5/carreira-perpinan05a.html
two aromatic rings: Feeling the van der Waals attraction by
[58] G. Montufar, J. Rauh, and N. Ay, “Expressive Power and Approximation
quantum Monte Carlo methods,” The Journal of Chemical Physics, Errors of Restricted Boltzmann Machines,” Jun. 2014, arXiv:1406.3140
vol. 127, no. 1, p. 014105, Jul. 2007. [Online]. Available: [math, stat]. [Online]. Available: http://arxiv.org/abs/1406.3140
https://pubs.aip.org/aip/jcp/article/906474
[59] N. Wiebe, A. Kapoor, and K. M. Svore, “Quantum Deep
[42] S. Sorella, “Green Function Monte Carlo with Stochastic Learning,” May 2015, arXiv:1412.3489 [quant-ph]. [Online]. Available:
Reconfiguration,” Physical Review Letters, vol. 80, no. 20, pp. http://arxiv.org/abs/1412.3489
4558–4561, May 1998, publisher: American Physical Society. [Online]. [60] C. Zoufal, A. Lucchi, and S. Woerner, “Variational quantum Boltzmann
Available: https://link.aps.org/doi/10.1103/PhysRevLett.80.4558 machines,” Quantum Machine Intelligence, vol. 3, no. 1, p. 7, Feb.
[43] K. Choo, T. Neupert, and G. Carleo, “Study of the Two-Dimensional 2021. [Online]. Available: https://doi.org/10.1007/s42484-020-00033-7
Frustrated J1-J2 Model with Neural Network Quantum States,” Physical [61] M. Amin, “Searching for quantum speedup in quasistatic quantum
Review B, vol. 100, no. 12, p. 125124, 2019. [Online]. Available: annealers,” Physical Review A, vol. 92, no. 5, p. 052323, Nov. 2015,
http://arxiv.org/abs/1903.06713 arXiv:1503.04216 [cond-mat, physics:quant-ph]. [Online]. Available:
[44] F. Ferrari, F. Becca, and J. Carrasquilla, “Neural Gutzwiller-projected http://arxiv.org/abs/1503.04216
variational wave functions,” Physical Review B, vol. 100, no. 12, p. [62] M. Benedetti, J. Realpe-Gómez, R. Biswas, and A. Perdomo-Ortiz,
125131, Sep. 2019, publisher: American Physical Society. [Online]. “Estimation of effective temperatures in quantum annealers for
Available: https://link.aps.org/doi/10.1103/PhysRevB.100.125131 sampling applications: A case study with possible applications
[45] J. R. McClean, J. Romero, R. Babbush, and A. Aspuru- in deep learning,” Physical Review A, vol. 94, no. 2, p.
Guzik, “The theory of variational hybrid quantum-classical 022308, Aug. 2016, arXiv:1510.07611 [quant-ph]. [Online]. Available:
algorithms,” New Journal of Physics, vol. 18, no. 2, p. http://arxiv.org/abs/1510.07611
023023, Feb. 2016, publisher: IOP Publishing. [Online]. Available: [63] S. H. Adachi and M. P. Henderson, “Application of Quantum Annealing
https://dx.doi.org/10.1088/1367-2630/18/2/023023 to Training of Deep Neural Networks,” Oct. 2015, arXiv:1510.06356
[46] J. Tilly, H. Chen, S. Cao, D. Picozzi, K. Setia, Y. Li, E. Grant, [quant-ph, stat]. [Online]. Available: http://arxiv.org/abs/1510.06356
L. Wossnig, I. Rungger, G. H. Booth, and J. Tennyson, “The Variational [64] M. Sajjan, S. H. Sureshbabu, and S. Kais, “Quantum Machine-Learning
Quantum Eigensolver: A review of methods and best practices,” for Eigenstate Filtration in Two-Dimensional Materials,” Journal of
Physics Reports, vol. 986, pp. 1–128, Nov. 2022. [Online]. Available: the American Chemical Society, vol. 143, no. 44, pp. 18 426–
https://www.sciencedirect.com/science/article/pii/S0370157322003118 18 445, Nov. 2021, arXiv:2105.09488 [physics]. [Online]. Available:
[47] J. Stokes, J. Izaac, N. Killoran, and G. Carleo, “Quantum Natural http://arxiv.org/abs/2105.09488
Gradient,” Quantum, vol. 4, p. 269, May 2020, arXiv:1909.02108 [65] S. H. Sureshbabu, M. Sajjan, S. Oh, and S. Kais, “Implementation
[quant-ph, stat]. [Online]. Available: http://arxiv.org/abs/1909.02108 of Quantum Machine Learning for Electronic Structure Calculations
[48] G. E. Hinton and T. J. Sejnowski, “Learning and relearning in Boltz- of Periodic Systems on Quantum Computing Devices,” Journal of
mann machines,” in Parallel distributed processing: explorations in the Chemical Information and Modeling, vol. 61, no. 6, pp. 2667–2674,
 Jun. 2021, publisher: American Chemical Society. [Online]. Available: Mar. 2010, arXiv:1003.4394 [cs, math]. [Online]. Available:
https://doi.org/10.1021/acs.jcim.1c00294 http://arxiv.org/abs/1003.4394
[66] F. Paul, M. Falkenthal, and S. Feld, “Clever Design, Unexpected [84] C. J. Needham, J. R. Bradford, A. J. Bulpitt, and D. R. Westhead,
Obstacles: Insights on Implementing a Quantum Boltzmann “Inference in Bayesian networks,” Nature biotechnology, vol. 24, no. 1,
Machine,” Jan. 2023, arXiv:2301.13705 [quant-ph]. [Online]. Available: pp. 51–53, 2006, publisher: Nature Publishing Group US New York.
http://arxiv.org/abs/2301.13705 [85] Y. Liu and J.-D. J. Han, “Application of Bayesian networks on large-
[67] M. Bukov, M. Schmitt, and M. Dupont, “Learning the ground scale biological data,” Frontiers in Biology, vol. 5, pp. 98–104, 2010,
state of a non-stoquastic quantum Hamiltonian in a rugged neural publisher: Springer.
network landscape,” SciPost Physics, vol. 10, no. 6, p. 147, Jun. [86] T. E. Schaapveld, S. L. Opperman, and S. Harbison, “Bayesian networks
2021, arXiv:2011.11214 [cond-mat, physics:physics, physics:quant-ph]. for the interpretation of biological evidence,” Wiley Interdisciplinary
[Online]. Available: http://arxiv.org/abs/2011.11214 Reviews: Forensic Science, vol. 1, no. 3, p. e1325, 2019, publisher:
[68] C.-Y. Park and M. J. Kastoryano, “Expressive power of complex- Wiley Online Library.
valued restricted Boltzmann machines for solving non-stoquastic [87] C. L. Lau and C. S. Smith, “Bayesian networks in infectious disease
Hamiltonians,” Physical Review B, vol. 106, no. 13, p. 134437, eco-epidemiology,” Reviews on environmental health, vol. 31, no. 1, pp.
Oct. 2022, arXiv:2012.08889 [cond-mat, physics:quant-ph]. [Online]. 173–177, 2016, publisher: De Gruyter.
Available: http://arxiv.org/abs/2012.08889 [88] F. L. Seixas, B. Zadrozny, J. Laks, A. Conci, and D. C. M. Saade, “A
[69] T. Westerhout, N. Astrakhantsev, K. S. Tikhonov, M. I. Katsnelson, and Bayesian network decision model for supporting the diagnosis of demen-
A. A. Bagrov, “Generalization properties of neural network approxi- tia, Alzheimer’s disease and mild cognitive impairment,” Computers in
mations to frustrated magnet ground states,” Nature Communications, biology and medicine, vol. 51, pp. 140–158, 2014, publisher: Elsevier.
vol. 11, no. 1, p. 1593, Mar. 2020. [89] C. Bielza and P. Larrañaga, “Bayesian networks in neuroscience: a
[70] M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, and R. Melko, survey,” Frontiers in computational neuroscience, vol. 8, p. 131, 2014,
“Quantum Boltzmann Machine,” Physical Review X, vol. 8, no. 2, p. publisher: Frontiers Media SA.
021050, May 2018, arXiv:1601.02036 [quant-ph]. [Online]. Available: [90] P. Sebastiani, M. Abad, and M. F. Ramoni, “Bayesian networks for
http://arxiv.org/abs/1601.02036 genomic analysis,” Genomic signal processing and statistics, vol. 2, pp.
[71] M. Benedetti, J. Realpe-Gómez, R. Biswas, and A. Perdomo-Ortiz, 281–320, 2005, publisher: EURASIP Book Series on Signal Processing
“Quantum-assisted learning of hardware-embedded probabilistic graph- and Communications, Hindawi . . . .
ical models,” Physical Review X, vol. 7, no. 4, p. 041052, 2017. [91] K. P. Murphy, “Dynamic bayesian networks,” Probabilistic Graphical
Models, M. Jordan, vol. 7, p. 431, 2002.
[72] T. A. Stephenson, Ed., An Introduction to Bayesian Network Theory and
[92] Z. Ghahramani, “Learning dynamic Bayesian networks,” Adaptive Pro-
Usage. IDIAP, 2000.
cessing of Sequences and Data Structures: International Summer School
[73] B. Drury, J. Valverde-Rebaza, M.-F. Moura, and A. de An- on Neural Networks “ER Caianiello” Vietri sul Mare, Salerno, Italy
drade Lopes, “A survey of the applications of Bayesian September 6–13, 1997 Tutorial Lectures, pp. 168–197, 2006, publisher:
networks in agriculture,” Engineering Applications of Artificial Springer.
Intelligence, vol. 65, pp. 29–42, Oct. 2017. [Online]. Available: [93] L. Song, M. Kolar, and E. Xing, “Time-varying dynamic bayesian
https://www.sciencedirect.com/science/article/pii/S0952197617301513 networks,” Advances in neural information processing systems, vol. 22,
[74] D. Heckerman, “Bayesian Networks for Data Mining,” Data Mining 2009.
and Knowledge Discovery, vol. 1, no. 1, pp. 79–119, Mar. 1997.
[Online]. Available: https://doi.org/10.1023/A:1009730122752
[75] R. Cano, C. Sordo, and J. M. Gutiérrez, “Applications of Bayesian
Networks in Meteorology,” in Advances in Bayesian Networks, ser.
Studies in Fuzziness and Soft Computing, J. A. Gámez, S. Moral, and
A. Salmerón, Eds. Berlin, Heidelberg: Springer, 2004, pp. 309–328.
[Online]. Available: https://doi.org/10.1007/978-3-540-39879-0 17
[76] A. Tosun, A. B. Bener, and S. Akbarinasaji, “A systematic literature
review on the applications of Bayesian networks to predict software
quality,” Software Quality Journal, vol. 25, no. 1, pp. 273–305, Mar.
2017. [Online]. Available: https://doi.org/10.1007/s11219-015-9297-z
[77] D. M. Chickering, “Learning Bayesian Networks is NP-Complete,”
in Learning from Data: Artificial Intelligence and Statistics V,
ser. Lecture Notes in Statistics, D. Fisher and H.-J. Lenz, Eds.
New York, NY: Springer, 1996, pp. 121–130. [Online]. Available:
https://doi.org/10.1007/978-1-4612-2404-4 12
[78] D. Koller and N. Friedman, Probabilistic Graphical Models: Prin-
ciples and Techniques. MIT Press, Jul. 2009, google-Books-ID:
7dzpHCHzNQ4C.
[79] G. H. Low, T. J. Yoder, and I. L. Chuang, “Quantum inference on
Bayesian networks,” Physical Review A, vol. 89, no. 6, p. 062315,
Jun. 2014, publisher: American Physical Society. [Online]. Available:
https://link.aps.org/doi/10.1103/PhysRevA.89.062315
[80] X. Gao, E. R. Anschuetz, S.-T. Wang, J. I. Cirac, and
M. D. Lukin, “Enhancing Generative Models via Quantum
Correlations,” Physical Review X, vol. 12, no. 2, p. 021037,
May 2022, publisher: American Physical Society. [Online]. Available:
https://link.aps.org/doi/10.1103/PhysRevX.12.021037
[81] W. Zeng and B. Coecke, “Quantum Algorithms for Compositional
Natural Language Processing,” Electronic Proceedings in Theoretical
Computer Science, vol. 221, pp. 67–75, Aug. 2016. [Online]. Available:
http://arxiv.org/abs/1608.01406
[82] K. Meichanetzidis, A. Toumi, G. de Felice, and B. Coecke, “Grammar-
aware sentence classification on quantum computers,” Quantum
Machine Intelligence, vol. 5, no. 1, p. 10, Feb. 2023. [Online].
Available: https://doi.org/10.1007/s42484-023-00097-1
[83] B. Coecke, M. Sadrzadeh, and S. Clark, “Mathematical
Foundations for a Compositional Distributional Model of Meaning,”