Boltzmann Machine and QMP

Boltzmann machines and quantum many-body
problems
Yusuke Nomura
Department of Applied Physics and Physico-Informatics, Keio University, 3-14-1
Hiyoshi, Kohoku-ku, Yokohama, 223-8522, Japan
arXiv:2306.16877v3 [cond-mat.str-el] 1 Nov 2023
E-mail: nomura@appi.keio.ac.jp
January 2023
Abstract. Analyzing quantum many-body problems and elucidating the

entangled structure of quantum states is a significant challenge common to a
wide range of fields. Recently, a novel approach using machine learning was
introduced to address this challenge. The idea is to “embed” nontrivial quantum
correlations (quantum entanglement) into artificial neural networks. Through
intensive developments, artificial neural network methods are becoming new
powerful tools for analyzing quantum many-body problems. Among various
artificial neural networks, this topical review focuses on Boltzmann machines and
provides an overview of recent developments and applications.
Boltzmann machines and quantum many-body problems 2
1. Introduction powerful variational ansatz.

Historically, human insight has played a crucial
The motions of particles in a quantum many-body role in the development of variational ansatzs. In
system are described by the following Schrödinger fact, a variety of good ansatzs have been proposed,
equation bringing innovation in our understanding of quantum
H|ψ⟩ = E|ψ⟩ (1) many-body phenomena. Well-acknowledged wave
functions include the Bardeen-Cooper-Schrieffer [2]
where H is the Hamiltonian, an operator that wave function for conventional superconductivity, the
represents the total energy of the system. The specific resonating valence bond wave function [3] for quantum
form of the Hamiltonian depends on the type of spin liquid, and the Laughlin wave function [4] for
particles constituting the quantum many-body system, fractional quantum Hall effect. Human-constructed
how they interact with each other, and the form of the wave functions have the advantage of being easy to
potential. interpret, but they have the disadvantage of being
Despite the simplicity of the equation, its quantum difficult to systematically improve the accuracy; it is
and many-body nature gives rise to diverse quantum highly nontrivial to identify the types of variational
many-body phenomena that cannot be explained parameters that should be added for this purpose.
by the motion of a single particle. From a Furthermore, for strongly correlated systems, as
numerical perspective, the difficulty lies in the fact various quantum phases compete within a small
that the dimensions of the Hilbert space increase energy scale, bias due to human insight may lead to
exponentially with the number of particles. Therefore, qualitatively incorrect predictions.
obtaining exact solutions of the quantum many- Given this situation, it is interesting to see how
body problems on a classical computer requires an machine learning techniques can contribute to this
exponentially large computational cost as the system field. Using “machine brains” instead of human
size increases. To understand quantum many-body brains may open up new routes to building good
phenomena and design functional materials, it is variational ansatz. The idea is to describe variational
necessary to develop powerful numerical methods wave functions by means of artificial neural networks,
that can obtain approximate but sufficiently accurate exploiting the flexible representability of artificial
solutions with tractable computational resources. neural networks. Notably, artificial neural networks
To date, various numerical methods have been have the property of universal approximation; in the
developed, including the quantum Monte Carlo limit of large networks, they can reproduce any wave
(QMC) method and the variational wave function function form with arbitrary accuracy [5,6]. Thus, one
method. The former applies the Monte Carlo method can systematically improve the representative power of
to estimate multidimensional integrals in quantum the variational ansatz by increasing the network size,
systems (for a review, see, e.g., Ref. [1]). If which marks an advantage over insight-based ansatzs.
the sampling weights of the Monte Carlo method Such a new approach emerged in 2017 when
are always positive, it is possible to estimate the Carleo and Troyer [7] introduced a variational wave
exact expectation values of physical quantities such function based on a restricted Boltzmann machine
as the energy and correlation functions within the (RBM) [8]. They showed that the RBM wave
margin of error. However, in the case of challenging function can accurately simulate the ground state
quantum many-body systems such as frustrated spin and real-time evolution of quantum spin systems
and fermion systems, a negative sign problem arises without frustration [7]. Since then, extensions
where some of the Monte Carlo sampling weights and improvements of the artificial neural network
become unavoidably negative. In the presence of a method have been continuously pursued all over
severe negative sign problem, an exponentially large the world. Currently, its applications have been
computation time is required to accurately determine extended to simulations of quantum spin systems with
the true integral value obtained by summing the geometrical frustration [?, 9–19, 21, 22], itinerant boson
samples with positive and negative weights. This systems [23, 24], fermion systems [9, 25–35], fermion-
limits the applicability of the QMC method. The boson coupled systems [36], topologically nontrivial
latter approximates the eigenstate wave function of quantum states [5, 6, 37–41], excited states [15, 32, 36,
a quantum many-body system by a functional form 42–44], real-time evolution [7, 45, 46], open quantum
described by, at most, a polynomial number of systems [47–50], and finite-temperature properties [51,
parameters. In this case, physical quantities such 52]. In the early stages of research, most studies
as energy and correlation functions can be obtained focused on benchmarks to check the accuracy of neural-
without explicit negative sign problems, but their network quantum states. More recently, calculations
accuracy depends on the quality of the variational wave beyond benchmarks have begun to appear, and neural
function. Therefore, the problem is how to construct a
network methods are being applied to the analysis of (a) (c) d1 d2 … dM 0 deep layer
unsolved problems, such as the physics of frustrated hidden layer
quantum spin systems [16, 17] (note that the sign h1 h2 h3 … hM
problem limits the applicability of the QMC method h1 h2 h3 … hM

to frustrated spin systems).
v1 v2 … vN
Because the applications and architectures of
neural-network quantum states are diverse, it is far visible layer …
v1 v2 vN
beyond the scope of this article to cover all the
developments. Here, we focus on Boltzmann machines (d)
(b)
among various neural network architectures and review
recent developments in quantum state representation
using Boltzmann machines. The remainder of this
paper is organized as follows. In Sec. 2, we
introduce Boltzmann machines as generative models.
Then, in Sec. 3, we discuss how to use Boltzmann
machines to construct variational wave functions
to be applied to quantum many-body problems.
Secs. 4–6 are devoted to more concrete applications. Figure 1. Structures of (a) restricted Boltzmann
Sec. 4 discusses the variational approach using RBMs machine (RBM), (b) general Boltzmann machine, and (c) deep
Boltzmann machine (DBM). (d) Structures with more than two
for zero temperature calculations. Secs. 5 and 6 hidden layers can be attributed to structures with two hidden
introduce algorithms using deep Boltzmann machines layers by changing the position of the hidden units.
(DBMs) for zero-temperature and finite-temperature
simulations, respectively. Finally, we give a summary
and outlook in Sec. 7. Here, ai and bj are the magnetic fields on the visible
and hidden spins, respectively, and Wij is the classical
2. Boltzmann machines interaction between the visible and hidden spins. In
the machine learning field, the terms associated with
Here, we introduce the RBM and DBM as generative ai and bj are called bias terms. For the artificial
models. The generative model is a generic term for energy E(v, h), we can define the Boltzmann weight
mathematical models that are intended to model the −E(v,h)
p(v, h) = e Z , where Z is the partition function
structure and probability distribution inherent in data. P −E(v,h)
Z = v,h e . By tracing out the hidden-spin
In the next section (Sec. 3), we will discuss how
degrees of freedom, we obtain the marginal probability
these Boltzmann machines can be applied to modeling
distribution p̃(v):
“quantum data,” or quantum states. X X
p̃(v) = p(v, h) p̃(v) = 1 . (3)
2.1. Restricted Boltzmann machines h v
Thus, the RBM takes the visible spin configuration v
Fig. 1(a) shows the structure of the RBM. The RBM as the input and gives the marginal probability p̃(v)
is composed of visible units {vi } that make up the as the output. This p̃(v) is used to approximate
data space and extra degrees of freedom called hidden the probability distributions over the visible spin
units {hj }. The regions where each degree of freedom configurations.
exists are called visible and hidden layers. Both vi This Boltzmann machine is called “restricted”
and hj take two states; here we define the values of Boltzmann machine because of the restriction that
the two states as vi = ±1, hj = ±1 (another popular there is no intralayer interaction in the visible and
definition of the two states is vi , hj = 0, 1). When hidden layers [For comparison, we show the structure
discussing applications to physics, it is convenient to of a general Boltzmann machine in Fig. 1(b)]. Because
identify them with Ising spins; therefore, we will refer there is no interaction between hidden spins, one can
to them as visible and hidden spins in this paper. analytically trace out hidden spins, and p̃(v) in Eq. (3)
The RBM approximates probability distributions can be efficiently computed as
over visible spin configurations, as described below. X Y X
First, we define the artificial energy function E(v, h) p̃(v) ∝ exp ai vi × 2 cosh bj + Wij vi . (4)
over the visible spin configuration v = (v1 , . . . , vN ) and i j i
the hidden spin configuration h = (h1 , . . . , hM ) as As briefly mentioned in Sec. 1, artificial neural
X X X networks have the property of universal approximation.
E(v, h) = − ai v i − Wij vi hj − bj hj . (2)
In the case of the RBM, an arbitrary probability
i i,j j
distribution can be reproduced with arbitrary precision
by introducing an exponentially large number of where ψ(x) is quantum many-body wave function.
hidden spins (M ∼ 2N ) comparable to the number of As seen in the previous section (Sec. 2), Boltzmann
visible spin configuration patterns 2N [53, 54]. machines give a probability (positive real number) p̃(v)
for a visible spin configuration v. On the other hand,
2.2. Deep Boltzmann machines the amplitude of quantum many-body wave function
ψ(x) can be negative or complex in general.
The DBM has an additional hidden layer compared Therefore, we can apply Boltzmann machines
to the RBM [Fig. 1(c)]. The structure with two to quantum state representations by satisfying the
hidden layers is general, encompassing structures following conditions:
with an arbitrary number of hidden layers with only
neighboring interlayer couplings [Fig. 1(d)]. In this (i) One-to-one correspondence between the configura-
paper, the two hidden layers are referred to as the tion of physical degrees of freedom x and visible-
hidden layer and the deep layer for the sake of spin configuration v.
distinction. In the case of DBM, the artificial energy (ii) Extension to complex amplitude.
E(v, h, d) with respect to the spin configurations v = (iii) (For fermion systems) Incorporation of fermionic
(v1 , . . . , vN ), h = (h1 , . . . , hM ), and d = (d1 , . . . , dM ′ ) anti-commutation relations.
reads
X X X For condition (i), one needs to define the mapping
E(v, h, d) = − ai vi − bj hj − b′k dk depending on the nature of the physical degrees of
i j k freedom. For S = 1/2 quantum spin systems, the most
X X
′ natural mapping between x and v is to identify the z
− Wij vi hj − Wjk hj dk . (5)
i,j j,k component of i-th quantum spin σiz with vi .
Condition (ii) can be satisfied, e.g., by extending
In addition to the parameters in the RBM, the DBM
the parameters of Boltzmann machines to complex
has parameters for the magnetic field on the deep spins
variables or by decomposing the wave function as
b′k and the interaction between the hidden and deep
′ ψ(x) = |ψ(x)|eiϕ(x) and expressing |ψ(x)| and ϕ(x)
spins Wjk . As with the RBM, using the partition
P −E(v,h,d)
using independent Boltzmann machines with real
function Z = v,h,d e and the Boltzmann parameters. For example, when approximating the
−E(v,h,d)
weight p(v, h, d) = e Z , the marginal probability quantum states of the one-dimensional Heisenberg
distribution p̃(v) is given by model by the RBM, the former appears to perform
X X better [55]. We note however that, for other general
p̃(v) = p(v, h, d) p̃(v) = 1 . (6) Hamiltonians, it is unclear whether the former is
h,d v
always better. Hereafter, we will focus on the former
Compared to the RBM, the DBM has the advantage of approach.
having deep degrees of freedom, which greatly improves For condition (iii), for example, one can take
the ability to represent the probability distribution, an approach to make the wave function itself anti-
but since there is no efficient analytical formula for symmetric. Alternatively, one can consider an
p̃(v) ‡, numerical methods such as the Monte Carlo approach that maps fermionic systems into spin
method are required to obtain the p̃(v) value. systems using, e.g., Jordan-Wigner [56] or Bravyi-
Kitaev [57] transformations. This issue will be
3. Wave function based on Boltzmann discussed in more detail in Sec. 4.3.1.
machines
3.2. Property of Boltzmann machine wave functions
3.1. Application of Boltzmann machines to quantum and comparison to tensor networks
state representation
Here, we discuss several properties of Boltzmann ma-
Here, we introduce how Boltzmann machines can be chine wave functions using S = 1/2 quantum spin sys-
used to represent quantum states. A quantum state tems. By (i) identifying σiz with vi and (ii) extend-
|ψ⟩ can be expanded using a certain basis {|x⟩} as ing the parameters to complex variables, Boltzmann
|ψ⟩ =
X
|x⟩ψ(x), (7) machines can be applied to the representation of the
x
quantum states of quantum spin systems (Fig. 2). In
this case, the amplitude of the RBM wave function for
‡ Either hidden or deep spin degrees of freedom can be traced a spin configuration σ = (σiz , . . . σN
z
) is given by
out analytically, but tracing out spins of one layer will introduce X
X X X
effective interactions between the spins of the other layer. This
ψ(σ) = exp ai σiz + Wij σiz hj + bj hj
makes it difficult to analytically trace out the spin degrees of
freedom of the other layer. h i i,j j
sha1_base64="PPh9wTdOZeejIzBkhNfdqOYZGOw=">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</latexit>
<latexit
(a)
h1
sha1_base64="STaGrgTjNXcv1LcwXSK95Y0Wd4Q=">AAACbXichVHJSgNBEH0Z97gkKoKgSDC4nEJFQjQeRPDi0S0qbmFmbGPjbMxMAib4A54FD6KgICJ+hhd/wIOfIB48KHjxYGUSERG1mu6qfl2v6nW35hjS84keQkpdfUNjU3NLuLWtvSMS7exa9uyCq4usbhu2u6qpnjCkJbK+9A2x6rhCNTVDrGh7M5XzlaJwPWlbS/6+IzZNNW/JHamrPkNrG57Mm2ouuVXKReOUyGTSND4R+xkkExRYHDWbs6NX2MA2bOgowISABZ9jAyo8HutIguAwtokyYy5HMjgXOECYuQXOEpyhMrrHa5536zXU4n2lphewde5i8HSZGcMQ3dM1vdAd3dAjvf9aqxzUqGjZZ69VucLJRQ57F9/+ZZnsfex+sf7U7GMHE4FWydqdAKncQq/yi6Xjl8XJhaHyMF3QE+s/pwe65RtYxVf9cl4snCDMH/D5yrHfg+WxRDKdSM2n4tNTta9oRh8GMcrvPY5pzGIOWe5r4QinOAs9Kz1KvzJQTVVCNU43vpky8gH27o4p</latexit>
<latexit
1
ψ(σ) =
z
sha1_base64="O658+vz8yxBpRqpEF2TAnrGhDvI=">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</latexit>
<latexit
Figure 2.
sha1_base64="YXY1Mn7R932DLGnwqu//yGSj0T4=">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</latexit>
<latexit
h2
h,d
= exp
X
2
z
function reads
i
sha1_base64="OCCPdF6c24s87PZFEbHNlkkJ+wM=">AAACZnichVE9SwNBEH05v2L8iooo2ARDxCpMNOTDQgQbS5MYFVTC3blJDi93x90lEIN/QLA1hZWCiPgzbPwDFv4DxVLBxsLJJSIi0Vl2Z/btvJm3u4qla45L9OiTenr7+gf8g4Gh4ZHRseD4xJZjVm1V5FVTN+0dRXaErhki72quLnYsW8gVRRfbyuFa63y7JmxHM41Nt26J/YpcMrSipsouQ7lyYakQDFM0nU5QMhX6HcSi5FkYHdswg9fYwwFMqKiiAgEDLsc6ZDg8dhEDwWJsHw3GbI4071zgGAHmVjlLcIbM6CGvJd7tdlCD962ajsdWuYvO02ZmCBF6oBt6pXu6pWf66Fqr4dVoaamzV9pcYRXGTmZy7/+yKuxdlL9Zf2p2UUTK06qxdstDWrdQ2/zaUfM1t5yNNObpkl5Y/wU90h3fwKi9qVcZkT1HgD/g65VD3YOtxWgsEY1n4uHVlc5X+DGLOSzweyexinVsIM99SzjFGZq+J2lUmpKm26mSr8OZxA+TQp8qPosI</latexit>
<latexit
exp
X
h3
…
sha1_base64="9G/AsRCmW+TzN7XDsaqR5bfgoK4=">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</latexit>
<latexit
+
i
…
X
z
N
i,j
sha1_base64="YLLV2clXgOm6Mv6QaFkPjXLpld0=">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</latexit>
<latexit
X
ai σiz ×
hM
and entanglement entropy

Y
ai σiz +
j
3.2.1. Universal approximator

(b)
sha1_base64="PPh9wTdOZeejIzBkhNfdqOYZGOw=">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</latexit>
<latexit
tence of the deep layer [58, 59].
X
Wij σiz hj
h1
sha1_base64="TW8NU84FrhaF/GCPpel4B5jzi+w=">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</latexit>
<latexit
sha1_base64="STaGrgTjNXcv1LcwXSK95Y0Wd4Q=">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</latexit>
<latexit
+

d1
1
z
the limit of large bond dimensions.

sha1_base64="O658+vz8yxBpRqpEF2TAnrGhDvI=">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</latexit>
<latexit
j,k
X
h2
sha1_base64="YXY1Mn7R932DLGnwqu//yGSj0T4=">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</latexit>
<latexit
bj hj +
sha1_base64="XvcSdxak569LPkc2ZNqxz/B7BGY=">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</latexit>
<latexit
2 cosh bj +
2
z
d2
k
i
′
X
Wjk
sha1_base64="OCCPdF6c24s87PZFEbHNlkkJ+wM=">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</latexit>
<latexit
X
states with area-law entanglement entropy.

h3
…
…
sha1_base64="MfRjImHuvIglz3Wh6ilokpJ4tRM=">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</latexit>
<latexit
b′k dk
sha1_base64="9G/AsRCmW+TzN7XDsaqR5bfgoK4=">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</latexit>
<latexit
…
states of S = 1/2 spin systems: (a) RBM and (b) DBM.

dM0
z
N
hj dk .
sha1_base64="YLLV2clXgOm6Mv6QaFkPjXLpld0=">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</latexit>
<latexit

hM
known for Boltzmann machine wave functions (or,

On the other hand, no such good indicator is
tensor networks are good at approximating quantum
bond dimension of χ can exhibit is log χ. Therefore,
maximum entanglement entropy that an MPS with a
entropy. For example, in the case of MPS, the
bond dimensions obey the area law of entanglement
by entanglement entropy. Tensor networks with finite
representability of tensor networks is well described
In the case of finite bond dimensions, the
networks are also known as universal approximators in
matrix product states (MPSs). This is because tensor
matrix renormalization group (DMRG) [62, 63] using
tensor network method [60, 61], including the density
The neural network method is complementary to the
3.2.2. Comparison to tensor networks: representability
for quantum state approximation thanks to the exis-
curacy [5, 6]. The DBM has more flexible expressivity
represent an arbitrary wave function with arbitrary ac-
in the limit of a large number of hidden spins, it can
versal approximation for quantum state representation;
with complex parameters also has the property of uni-
butions by the RBM with real parameters, the RBM
As in the case of approximating probability distri-
(9)
simplicity. Similarly, the form of the DBM wave
Hereafter, we omit the normalization factor for
Wij σiz . (8)
Boltzmann machines for representing quantum
Boltzmann machines and quantum many-body problems
states.
volume-law entanglement entropy.

entropy follows the area law [37, 39].
between MPS and Boltzmann machines

approaches the limit of universal approximation.
the RBM can reproduce an arbitrary MPS in the

However, owing to its universal approximation ability,
is no general transformation rule from MPS to RBM.
transformation, to the best of our knowledge, there
From MPS to RBM/DBM. As for the inverse
dimension is required to represent quantum states with
bond dimension χ is log χ, an exponentially large bond
the maximum entanglement entropy of an MPS with
can exhibit volume-law entanglement entropy. Because
Sec. 3.2.2, quantum states represented by the RBM
in terms of entanglement entropy. As discussed in
the required bond dimension can also be understood
specific transformation formulas. The magnitude of
becomes smaller. Please refer to Ref. [64] for the
the RBM are local, the required bond dimension
size. However, when the coupling parameters of
large bond dimension χ with respect to the system
RBM, analytic conversion requires an exponentially
law from RBM to MPS. In the case of long-range
From RBM to MPS. Ref. [64] discussed a conversion
product states and Boltzmann machine quantum
shall consider the transformation rule between matrix
relationship between the two. As an example, we
vice versa, is an interesting question to clarify the
transformed into tensor network wave functions, or
Whether the neural network wave functions can be
3.2.3. Comparison to tensor networks: transformation
represented by a short-range RBM, the entanglement
parameters [37]. However, when the quantum state is
can be constructed by an RBM with O(N ) complex
one-dimensional N qubits (S = 1/2 quantum spins)
volume-law entangled state in a system consisting of
entanglement entropy [37, 39]. In fact, the maximal
by a “long-range RBM” can exhibit volume-law
Then, we can show that quantum states represented
thus, the coupling becomes long-ranged [short-ranged].
coupling parameters to O(N ) [O(1)] visible spins;
In the former [latter], the hidden spins have finite
of RBMs: long-range RBM and short-range RBM.
functions. For this purpose, we define two types
the entanglement entropy of Boltzmann machine wave
Nevertheless, several properties are known about
of hidden degrees of freedom and how it asymptotically
the approximation accuracy improves with the number
general theory (to the best of our knowledge) on how
finite hidden degrees of freedom. Therefore, there is no
more generally, neural network wave functions) with
5
limit of a large number of hidden spins. On the DBM method has the advantage of being exact except
other hand, in the case of the DBM having a higher for the Suzuki-Trotter error; however, its applicability
representation power than the RBM, the number of is limited by the negative sign problem as in the QMC
hidden spins required for the transformation has been method. Therefore, for systems where the sign problem
discussed [6, 58]. In Ref. [6], it was shown that MPS cannot be avoided, the former variational approach,
with bond dimension χ for N qubit systems has a DBM which can avoid the explicit sign problem, will work
representation with O(N χ2 ) hidden degrees of freedom better. Finally, we discuss an extension to finite-
(hidden and deep spins). temperature simulations in Sec. 6.
3.2.4. Comparison to tensor networks: short summary 4.1. Variational method

Because of the difference in the properties of
Among the eigenstates of the Hamiltonian, the ground
entanglement entropy, quantum states with volume law
state has the lowest energy and is the most stable
entanglement entropy may be efficiently represented
quantum state at zero temperature. Obtaining an
with artificial neural networks (for example, maximal
accurate representation of the ground state is a long-
volume-law entangled state in one-dimensional N qubit
standing challenge not only in condensed matter
systems, as discussed in Sec. 3.2.2), but not by tensor
physics but also in various fields, including particle
networks.
physics, nuclear physics, and quantum chemistry.
On the other hand, there are examples in
Here, we introduce a variational method using
which tensor networks have efficient representations
RBMs. Since the ground state is the lowest-energy
with fewer parameters. For example, whereas the
state, the energy can be considered as a loss function.
Affleck-Kennedy-Lieb-Tasaki (AKLT) [65] state can
Thus, the problem of finding the ground state by
be represented by an MPS with χ = 2 [61], in
minimizing energy can be viewed as a machine-learning
the case of the RBM using σ z basis, a long-range
task, i.e., optimization of a nonlinear function (RBM)
RBM with nonlocal coupling parameters is needed to
using the nonlinear loss function (energy) in a high-
represent the hidden antiferromagnetic order in the
dimensional parameter space [68].
AKLT state [64]. The number of parameters is O(N )
in the case of MPS and increases to O(N 2 ) when RBMs
are used [39] §. This example shows that the capacity 4.1.1. Comparison to conventional variational method
of entanglement entropy is not the only factor that Conventional variational methods often assume the
determines expressive power [64]. form of a variational wave function based on some
Thus, it is not the case that either tensor networks physical insights. While this approach has the
or artificial neural networks are always superior. advantage of requiring a small number of parameters
Rather, they are complementary. Further discussion of and being intuitively easy to understand, it is difficult
the relationship between the two will lead to a better to systematically improve its accuracy. On the
understanding of quantum state representation. other hand, the artificial neural network method, in
principle, can be systematically improved by increasing
the size of the network (see Sec. 3.2.1). It may be
4. RBM for zero temperature calculations possible to extract nontrivial quantum correlations
even in situations where physical intuition is difficult
In the following, we will discuss methods using
to obtain.
Boltzmann machines for zero-temperature and finite-
temperature calculations. In this section, we introduce
a variational method based on RBM for zero- 4.1.2. Optimization of RBM
temperature simulations. This can be understood as Here, we discuss how to optimize the RBM varia-
an approximate (but accurate) representation of the tional parameters to minimize energy (loss function).
ground state obtained by numerically optimizing the Among various optimization methods, we introduce
RBM parameters. In the next section (Sec. 5), taking the stochastic reconfiguration (SR) method [69]. Some
advantage of the DBM’s more flexible representability studies employ simpler methods such as stochastic gra-
than the RBM, we introduce an analytic construction dient descent and its refined variants (e.g., Adam [70]).
of the DBM that represents the ground state. As we We choose this method because it is closely related
will see in Sec. 5, this analytical construction connects to the imaginary-time Hamiltonian evolution of quan-
a well-known physics concept, path-integral formalism, tum states, and hence the SR method greatly sta-
with a machine-learning concept, deep learning. This bilizes the optimization ∥. Essentially the same op-
timization method is known as the natural gradient
§ More recently, it was shown that the RBM can represent the
AKLT state efficiently with O(N ) parameters by choosing the ∥ Note that e−τ H |ψ (0) ⟩ converges to the ground state as τ →
basis appropriately [66, 67]. ∞, as long as the initial state |ψ (0) ⟩ has a nonzero overlap with
the ground state.
method [71, 72] in the machine learning community. The move in Eq. (15) can also be interpreted as the
Both were developed independently, and their relation- steepest descent direction with respect to the quantum
ship has only recently been recognized. information geometry [73].
The SR method makes use of information about The quantities used in the SR method can
the geometry of quantum states (Fubini-Study metric). be estimated using the Monte Carlo method (see
Hence, we first introduce the Fubini-Study metric and Appendix A.2 for details). Efficient techniques to
move on to the explanation of the update rule of the compute the gradient with respect to the variational
SR method. The Fubini-Study metric is a metric on parameters, such as backpropagation [74] (reverse
the space of quantum states, whose form is given by mode of automatic differentiation), can also be
s combined.
⟨ψ|ϕ⟩ ⟨ϕ|ψ⟩
F[|ψ⟩ , |ϕ⟩] := arccos . (10)
⟨ψ|ψ⟩ ⟨ϕ|ϕ⟩
4.1.3. Calculation procedure
|ψ⟩ Here, we summarize the calculation procedure by
For normalized quantum states |ψ̄⟩ = √ and
⟨ψ|ψ⟩
focusing on the variational approach for ground-state
|ϕ̄⟩ = √|ϕ⟩ , the Fubini-Study metric reads calculations.
⟨ϕ|ϕ⟩

F |ψ̄⟩ , |ϕ̄⟩ = arccos |⟨ψ̄|ϕ̄⟩| . (11) (i) Initialization: Construct a variational ansatz
based on the RBM and initialize RBM variational
This can be interpreted as the angle between the two parameters. We often put small random numbers.
complex vectors. Therefore, the Fubini-Study metric
is occasionally called the quantum angle. Now, we (ii) Optimization: Optimize the RBM parameters to
consider the infinitesimal form of this metric; for a minimize the loss function (total energy). In this
variational quantum state |ψθ ⟩ characterized by the paper, we employ the SR scheme for this purpose.
parameter set θ, it is given by (iii) Post-process: Once we obtain the optimized
X wave function, we can compute various physical
F 2 [|ψθ+δθ ⟩ , |ψθ ⟩] = Skl (θ)δθk δθl . (12) quantities such as the correlation function, as well
kl
as the total energy.
Skl (θ) = Re Qkl (θ) is the Fubini-Study metric tensor
and is given as the real part of so called quantum When random numbers are used for the initialization
geometric tensor Qkl (θ): of the parameters, multiple optimization calculations
are often performed using different sets of initial
⟨∂k ψθ |∂l ψθ ⟩ ⟨∂k ψθ |ψθ ⟩ ⟨ψθ |∂l ψθ ⟩ parameters. Then, for the calculation of physical
Qkl (θ) = − . (13)
⟨ψθ |ψθ ⟩ ⟨ψθ |ψθ ⟩ ⟨ψθ |ψθ ⟩ quantities in step (iii), we employ the lowest-energy
The Fubini-Study metric can be regarded as a quantum state among the optimized variational states.
version of the Fisher information metric and thus has
a close relationship with information geometry [73]. 4.1.4. Demonstration
As described above, the SR optimization is closely Here, we demonstrate the application of the RBM
related to the imaginary-time evolution of quantum wave function following the procedure described above.
states. The update rule can be derived by trying to As an example, we consider the one-dimensional
reproduce the infinitesimal imaginary-time evolution Heisenberg model on the 8-site spin chain with the
e−2δτ H |ψθ ⟩ by the variational state with updated periodic boundary condition. The Hamiltonian reads
parameters |ψθ+δθ ⟩: 8
− σiy σi+1
y
X
δ θe = arg min F[e−2δτ H |ψθ ⟩ , |ψθ+δθ ⟩] (14) H= (−σix σi+1
x
+ σiz σi+1
z
). (17)
δθ i=1
After some algebra (see Appendix A.1 for details), we This form is obtained by rotating the spin quantization
obtain the following update rule: axis by 180 degrees about the z axis for one of the
−1 (t) two sublattices
P (gauge transformation) from the usual
(t)
X
δθk = −δτ S (t) kl gl , (15)
l
form H = i (σix σi+1x
+ σiy σi+1
y
+ σiz σi+1
z
). With this
where t indicates the iteration number of optimiza- form, the amplitude of the ground-state wave function
(t) becomes positive ψ(σ) > 0 for all σ in the sector with
tion ¶. gl is the gradient of the total energy (loss P z
function) with respect to the variational parameters i σi = 0.
First, we construct a variational ansatz based on
(t) ∂Eθ the RBM. The system has translational symmetry;
gk = . (16)
∂θk θ=θ (t) therefore, we impose translational symmetry in the
¶ Replacing the S matrix with the identity matrix (Hessian
variational ansatz. For this purpose, we impose
matrix) results in an update corresponding to the stochastic translational symmetry on Wij parameters. In
gradient descent method (Newton’s method). addition, one way to preserve the symmetry between
0.8 −10
(a) (b) 0.4
(c) (↑↓↑↓↑↓↑↓) (↓↑↓↑↓↑↓↑)
0.4 −11
Exact
0.3
Total energy
0 −12 Initial RBM
Wi−j
Optimized
ψ(σ)
0.2 RBM
−0.4 −13
Initial
−0.8 Optimized −14 0.1
−1.2 −15 0
0 1 2 3 4 5 6 7 0 100 200 300 400
(↑↑↑↑↓↓↓↓) (↓↓↓↓↑↑↑↑)
i−j Optimization loop σ
Figure 3. Demonstration of the variational method based on the RBM wave function. The method is applied to the one-
dimensional Heisenberg model on the 8-site spin chain with the periodic boundary condition. (a) Initial and optimized RBM
variational parameters in Eq. (18). Because of the translational symmetry, the origin of i − j is arbitrary. We set the origin of i − j
so that |W0 | is the largest among |Wi−j |. (b) Total energy (= loss function) as a function of the optimization loop. (c) Initial and
optimized RBM wave functions and comparison to the exact diagonalization result.
the up and down spins is to set the bias terms to zero, 4.2. Results by Carleo and Troyer
i.e., ai = 0 and bj = 0. When the number of hidden
The variational method based on the RBM was first
spins is equal to the number of physical spins, the RBM
introduced by Carleo and Troyer [7]. They applied the
wave function reads
RBM wave function to the one-dimensional transverse-
8 8
field Ising model, the one-dimensional Heisenberg
Y X
ψ(σ) = 2 cosh Wi−j σiz . (18)
model, and the two-dimensional Heisenberg model
j=1 i=1
on a square lattice [7]. Since these models have
In the present case, the wave function is real and no geometrical frustration, the values of the ground-
positive ψ(σ) > 0; thus, the Wi−j parameters are taken state wave function can be made positive and
to be real. In this setup, the number of independent real by a proper gauge transformation (the gauge
parameters is 8 (Wi−j with i − j = 0, . . . , 7). transformation was also used in Sec. 4.1.4). Using
As the initialization of the variational parameters, this fact, Carleo and Troyer also set the variational
we put small random numbers [Fig. 3(a)]. We then parameters to real numbers, as in Sec. 4.1.4. They
optimize the variational parameters using the SR also imposed the same translational symmetry on W
method to minimize the total energy (loss function) parameters (so that the number of hidden spins M is
[Fig. 3(b)]. The optimized Wi−j parameters show a an integer multiple of the number of visible spins N ).
sign change between W0 where |Wi−j | is largest and Defining α = M/N , α is the parameter that controls
W1 where |Wi−j | is second largest, which indicates the accuracy of the RBM wave function (α = 1 was
that the RBM learns the antiferromagnetic correlation used in the demonstration in Sec. 4.1.4).
[Fig. 3(a)]. For the present system size (8-site spin When introducing a new method, it is essential
chain), we can easily perform exact diagonalization and to verify its accuracy. Fig. 4 shows the benchmark
list
P all wave function amplitudes in the sector with results comparing the RBM variational energies and
z
i σi = 0. While the initial RBM wave function numerically exact energies. We see that, for all
has no recognizable features, the optimized RBM Hamiltonians, increasing α improves the accuracy of
wave function reproduces the exact result quite well the RBM variational state and reduces the relative
[Fig. 3(c)]. Once we obtain the optimized RBM wave error in energy, as expected. The results for the
function, we can compute, for example, the spin-spin two-dimensional Heisenberg model shown in Fig. 4(c)
correlation function (not shown). show that the RBM can achieve better accuracy
When the system size becomes large, it is not than those obtained from tensor networks such
possible to compute the wave function amplitudes for as entangled-plaquette states (EPSs) [75, 77] and
all the spin configurations. In this case, physical projected entangled pair states (PEPSs) [78]+ .
quantities are estimated numerically using the Monte
+ The comparison was made under the periodic boundary
Carlo method (see Appendix A.2). If the number of
condition. We note that the performance of tensor networks
hidden spins is polynomial with respect to N , the is much better under the open boundary condition.
computational cost scales polynomially with respect to
N.
(a) 1D TFI model (b) 1D Heisenberg model (c) 2D Heisenberg model

10−2 10−2
10−2
10−3
10−3 Jastrow
10−4 Γ=1 EPS
rel
rel
rel
−4
10−5 10
10−3 PEPS
−6
10 Γ = 1/2
10−5
10−7 Γ=2
10−8 10−6 10−4
1 2 4 1 2 4 1 2 4 8 16 32
α α α
Figure 4. Results of variational calculations using RBM by Carleo andPTroyer [7]. Relative error ϵrel of the RBM variational
energy for (a) one-dimensional transverse-field Ising (TFI) model H = − i σiz σi+1 z − Γ i σix (periodic boundary condition, 80
P
P x x y y z z
sites) (b) one-dimensional Heisenberg model H = i (−σi σi+1 − σi σi+1 + σi σi+1 ) (periodic boundary condition, 80 sites), and
P x x y y z z
(c) two-dimensional Heisenberg model H = ⟨i,j⟩ (−σi σj − σi σj + σi σj ) (periodic boundary condition, 10 × 10 square lattice). α
(= M/N ) is a parameter that controls the accuracy of the RBM. ϵrel is given by ϵrel = (ERBM (α) − Eexact )/|Eexact |. “Jastrow” in
(b) shows the result of Jastrow ansatz, and EPS and PEPS in (c) are the results of tensor networks taken from Refs. [75] and [76],
respectively.
4.3. Extensions to strongly-correlated Hamiltonians (a) U/t = 4 (b) U/t = 8

×10−2 ×10−2
The application of the variational method based on 5.6 8
artificial neural networks started from frustration-free 5.2 7
spin systems as introduced above (Sec 4.2). When RBM+Fermi sea RBM+Fermi sea
relative error
relative error
geometrical frustration is absent, the QMC method can 4.8 6

be applied without the negative sign problem (see also 5
Sec. 1). 1.2
0.044
RBM+PP
4
0 0.1 0.2 0.3 0.4 0.5
However, such cases are special, and general 0.8

quantum many-body Hamiltonians exhibit unavoid-
RBM+PP
able negative sign problems. Constructing a good 0.4 1
RBM+symmetrized PP
ground-state variational hypothesis for such problems RBM+symmetrized PP
0 0
is an important task for variational methods. Here, 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
among many challenging unsolved quantum many- 1/α 1/α
body Hamiltonians, we consider fermion and frustrated
spin systems as examples. Figure 5. Relative errors of variational energies for the
Hubbard model on the 8 × 8 square lattice (periodic and anti-
periodic boundary conditions) at (a) U/t = 4 and (b) U/t = 8.
4.3.1. Fermion systems Numerically exact references are obtained by the QMC method.
Analyzing the electronic structures of strongly- α (= M/N ) is a parameter that controls the accuracy of the
correlated materials is one of the most fundamental RBM part. Adapted with permission from Ref. [25]. Copyright
2017 by the American Physical Society.
problems in physics. Because electrons are fermionic
particles, it is highly important to extend the method
to fermionic systems.
When we consider applications to fermion sys- in quantum chemistry and satisfies the anti-symmetric
tems, we must consider the anti-commutation rela- property. In this case, the RBM is employed as
tions of fermion operators. Studies on fermion sys- a generalization of correlation factors such as the
tems can be classified into two types (Table 1): One Gutzwiller [80] and Jastrow [81] factors. Alternatively,
takes account of the anti-symmetric property in wave we can reformulate the problem as approximating
functions [25–27, 29–31, 33–35], while the other incor- ψexact /ψPP using RBM. Therefore, when the combined
porates it in Hamiltonians [28, 32]. Among studies, wave function becomes sophisticated, it becomes easy
Refs. [25], [28], and [32] employed RBMs in a varia- for the RBM to learn the ground state. This trend was
tional ansatz, with Ref. [25] using the former approach confirmed by the benchmark results for the Hubbard
and Refs. [28] and [32] using the latter. model on the 8 × 8 square lattice (Fig. 5). See also
In Ref. [25], the RBM is combined with the Sec. 4.3.2 for a discussion of the possible importance
pair-product (PP) state [79]. The pair-product wave of the combination.
function is also known as the geminal wave function In Refs. [28] and [32], fermionic Hamiltonians are
Table 1. Applications to fermion systems. ANN stands for artificial neural network. When referring to molecules as a target, the
target also includes, e.g., atoms, diatoms, and hydrogen chains.
Work Method Main target Ref.

Nomura et al. (2017) Jastrow-type lattice model [25]
Luo et al. (2019) backflow-type lattice model [26]
Han et al. (2019) anti-symmetrized ANN continuous-space, molecules [27]
Choo et al. (2020) mapping to spin model continuous-space, molecules [28]
Pfau et al. (2020) backflow-type continuous-space, molecules [29]
Hermann et al. (2020) backflow+Jastrow-type continuous-space, molecules [30]
Stokes et al. (2020) Jastrow-type lattice model [31]
Yoshioka et al. (2021) mapping to spin model continuous-space, periodic [32]
Inui et al. (2021) anti-symmetrized ANN lattice model [33]
Moreno et al. (2022) augmented Hilbert space+projection lattice model [34]
Cassella et al. (2023) backflow-type continuous-space, periodic [35]
first mapped onto quantum spin Hamiltonians. For

Table 2. Comparison of ground-state energy among various
this transformation, one can employ, e.g., the Jordan- variational methods for the J1 –J2 Heisenberg model with
Wigner [56] or Bravyi-Kitaev [57] transformations. J2 /J1 = 0.5 on the 10 × 10 square lattice (periodic boundary
The RBM is then employed to approximate the ground condition). The wave functions in bold font employ artificial
neural networks. Ref. [84] applies p-th order Lanczos step on
state (excited states) of the mapped quantum-spin
top of the optimized VMC state. CNN and GCNN stand for
Hamiltonians. convolutional neural network and group convolutional neural
network, respectively.
4.3.2. Frustrated quantum-spin systems
Energy per site Wave function Ref.
Quantum spin systems with geometrical frustration
−0.49516(1) CNN [11]
are fascinating systems that may host an exotic state of
−0.49521(1) VMC(p=0) [84]
matter, such as quantum spin liquid [82, 83]. However,
−0.495530 DMRG [85]
at the same time, it is numerically challenging to
−0.49575(3) RBM-fermionic w.f. [12]
analyze frustrated quantum spin systems. Extensions
−0.49717 CNN [18]
to frustrated quantum spin systems are one of the
−0.497437(7) GCNN [?]
major topics in the field of neural-network quantum
−0.497549(2) VMC(p=2) [84]
states.
−0.497629(1) RBM+PP [16]
Benchmarks
Among various frustrated quantum-spin Hamiltoni- projections). Unfortunately, the performance de-
ans, the J1 –J2 Heisenberg model on square lattices has pends on the symmetrization technique (see, e.g.,
been most intensively studied. Therefore, we focus on Ref. [21]); therefore, careful consideration of the
the J1 –J2 Heisenberg model here. symmetrization is necessary to obtain a good per-
Table 2 shows a comparison of the ground-state formance ∗ .
energy among various variational methods for the J1 –
• Combination. We can think of combining
J2 Heisenberg model with J2 /J1 = 0.5 on the 10 × 10
different ansatzs to construct a single variational
square lattice (periodic boundary condition). We leave
state. This is practically important because
several remarks from this comparison.
there exists a tradeoff between trainability and
• Symmetry. To achieve high accuracy, it is impor- representability in neural-network ansatzs. The
tant to employ symmetrization. Indeed, all highly combination helps to reduce the number of
accurate neural-network variational energies were parameters in the neural network part and
obtained by a certain symmetrization of the wave mitigates the problem of trainability (see also the
function. Symmetrization can be achieved in sev- discussion related to Fig. 5 in Sec 4.3.1). Indeed,
eral ways. For example, Ref. [7], which intro- for 10 × 10 square lattice at J2 /J1 = 0.5, the
duced the RBM wave function, employed the sym- best accuracy among the compared methods is
metrization of variational parameters. Recently, a obtained by the combination of RBM and pair-
projection onto a particular quantum number sec- ∗ Another advantage of symmetrization is that excited states
tor [15] has often been employed (see Fig. 6 for can be calculated by considering symmetry sectors that are
how the accuracy is improved by quantum-number different from those of the ground state.
10−1 such, the phase of research is shifting from develop-

Ψ(σ) ment/benchmarking to “true applications” (of course,
the development of methods must continue at the same
(ERBM − EED )/EED
−2
10
time).
ΨK=0 (σ)
−3 S+
10 ΨK=0 (σ) 5. DBM for zero-temperature calculations
Here, we describe a quantum state representation

10−4 1
ΨK=0+A ,S
(σ)
using DBMs, connecting deep learning and a well-
known physics concept, path-integral formalism. In the
10−5
0.40 0.45 0.50 0.55
previous section (Sec. 4), to represent the ground states
of quantum many-body systems with high accuracy,
J2 /J1 we described the numerical optimization of the RBM
parameters using the SR method to reproduce the
Figure 6. Effects of symmetrization on the accuracy imaginary-time Hamiltonian evolution as accurately
of RBM wave function with α = M/N = 2 for the J1 –
J2 Heisenberg model on the 6 × 6 square lattice (periodic as possible within the representability of the RBM
boundary condition). EED is the ground-sate energy obtained variational ansatz. Compared to the shallow RBM,
by exact diagonalization. Compared to the results of Ψ(σ) the DBM has more flexible representation capability
without any symmetrization, the results with quantum-number thanks to the existence of the deep layer (Fig. 2) [58,
projections show much better accuracy. Applied quantum-
number projections are as follows: total momentum (K = 0), 59]. Ref. [87] focused on the improved representability
spin-parity (even parity, represented as S+ ), and A1 irreducible and showed that the imaginary-time Hamiltonian
representation of the C4v point group. Reproduced from evolution can be reproduced with arbitrary precision
Ref. [15] (© 2021 IOP Publishing Ltd) by permission of IOP
within the DBM framework (see also Ref. [88] for
Publishing. All rights reserved.
a similar attempt). This leads to an analytic
construction of the DBM representation of ground
states.
product (PP) state that is mapped onto the spin In more detail, the procedure of this method is as
Hilbert space [16]. follows: First, the initial state is prepared as a DBM.
• Trainability. Although deep neural network We apply the imaginary-time Hamiltonian evolution
architectures should be better than shallow with the Suzuki-Trotter decomposition e−τ H ≃
Nτ
architectures to obtain good variational results, e−δτ H1 e−δτ H2 (H = H1 + H2 , δτ = τ /Nτ ) to
the problem of trainability often causes problems the DBM state. Each short-imaginary-time evolution
in achieving good performance. Introducing huge can be reproduced exactly by changing the DBM
computational resources is a brute-force way to structure (an analytical expression for the parameter
mitigate this problem, obtaining better accuracy can be derived). In the limit of long imaginary-time
using CNN compared to early results [18]. A evolution, we can obtain the DBM representation of
careful symmetric construction also helps [?]. the ground state as long as the initial DBM state is not
Improving the optimization algorithm is also an orthogonal to the ground state. Once the analytical
important issue (see, e.g., Ref. [22] for the latest DBM representation of the ground state is obtained,
attempts). physical quantities can be computed using the Monte
Carlo method. Since there is no efficient way to trace
Beyond benchmarks out the hidden and deep spins analytically, for this
Once a novel accurate variational method is devel- purpose, Monte Carlo sampling must be performed on
oped, it is important to apply it to study the physics the hidden or deep spin (or both) degrees of freedom
of unsolved quantum many-body Hamiltonians. Re- in addition to the visible spins (Sec. 2.2).
cently, such applications have started to emerge. For As an example, Fig. 7 shows how the DBM can
example, Ref. [16] applied the RBM+PP wave func- exactly reproduce the short imaginary-time evolution
tion to the J1 –J2 Heisenberg model on a square lattice of the transverse-field Ising model. To reproduce the
and revealed a quantum spin liquid phase. Ref. [17] evolution, new hidden and deep spins are introduced.
employed RBM, CNN, and many-variable variational Therefore, the size of the DBM network grows as
Monte Carlo (mVMC) [86] methods to investigate the a result of imaginary-time evolution; the number of
Heisenberg model on a pyrochlore lattice and con- hidden and deep spins is proportional to the system
cluded that symmetry-broken state showing long-range size and length of the imaginary time. The benchmark
dimer order is realized as a ground state instead of calculations in Ref. [87] for several quantum-spin
quantum spin liquid without any long-range order. As models confirmed that the constructed DBM indeed
Table 3. Comparison between the RBM and DBM methods described in Sec. 4 ad Sec. 5, respectively.
RBM (variational method) DBM (quantum-to-classical mapping)

Optimization and accuracy numerical and approximate analytical and exact (up to Trotter error)
Number of hidden units smaller compared to DBM ∝ (imaginary time)×(system size)
Sampling objects visible spins visible spins + hidden and/or deep spins
Negative sign problem applicable to frustrated systems unavoidable in frustrated case
6. DBM for finite-temperature calculations

Before imaginary-time evolution After imaginary-time evolution
So far, we have discussed calculations at zero

temperature, where the quantum effect governs
the property. An extension to finite temperature
calculations is challenging because one needs to
z z
l m take into account thermal and quantum fluctuations
simultaneously. Finite-temperature simulations are
also important for comparison with experiments
(note that experiments are performed at finite
temperatures).
Here we describe two approaches for performing
finite-temperature simulations using DBM [52]. Both
z
l methods use the concept of “purification”, in which a
mixed state of a system is expressed in terms of a pure
state of an extended system.
Figure 7. Evolution of DBM quantum state by
short imaginary-time propagators, e−H1 δτ and e−H2 δτ , of
As a concrete example, we consider a quantum
the H = H1 + H2 , H1 = spin-1/2 system with N spins. Considering an
P transverse-field Ising model:
z z x
extended system with N system spins and N ancilla
P
l<m Vlm σl σm , H2 = − l Γl σ l . The relevant visible,
hidden, and deep spins are highlighted in color. spins, the infinite-temperature state of the original
system can be expressed as the following pure
state of the extended system: |Ψ(T = ∞)⟩ =
NN ′ ′
reproduces the ground state of each model. i=1 (| ↑↓ ⟩ + | ↓↑ ⟩)i ††. From this state, the
Finally, we briefly compare this DBM method imaginary-time evolution by the Hamiltonian of the
with the RBM variational method (see Table 3). system leads to a pure state of the extended system
This DBM method provides a way to perform corresponding to the thermal mixed state of the
quantum-to-classical mapping (note that DBMs are original system. More explicitly, the pure state at
based on a classical spin model), and in a special temperature T reads |Ψ(T )⟩ = e−βH/2 ⊗ 1′ |Ψ(T = ∞)⟩
case, the mapping becomes equivalent to path- (β = 1/T ). When the ancilla spin degrees of freedom
integral formalism ♯. While it has the advantage of are traced out, it gives the density matrix of the
not requiring cumbersome numerical optimization of thermal equilibrium state.
parameters, unfortunately, the negative sign problem Interestingly, the finite-temperature pure state
can not be avoided for, e.g., frustrated quantum- |Ψ(T )⟩ can be reproduced by the DBM shown in
spin systems, as in the path-integral formalism; one Fig. 8(a). First, it can be shown analytically
can construct a DBM representation of the ground that the infinite-temperature pure state |Ψ(T =
state, but negative Monte Carlo weights appear in ∞)⟩ can be reproduced using the DBM shown in
the computation of physical quantities. Nevertheless, Fig. 8(b) (theQcorresponding DBMwave function reads
this method is of great significance in that it connects Ψ(σ, σ ′ ) = i 2 cosh i π4 (σi + σi′ ) ). The imaginary-

deep learning with the well-known physics concept and time evolution from this state can be reproduced either
demonstrates the remarkable representative power of (i) analytically as described in Sec. 5 or (ii) numerically
deep neural networks. and approximately using the SR method described in
♯ When the mapping becomes equivalent to the path-integral
Sec. 4.
formalism, deep spins play a role as auxiliary spins along In Method (i), the imaginary-time evolution is
imaginary time slices, and hidden spins mediate the interaction decomposed using Suzuki-Trotter decomposition as in
between the deep spins. For more details, please refer to Ref. [87].
†† A pure state that reproduces the infinite-temperature state is
not unique. Here, we chose one of the simplest forms.
(a) 0 d (c) 0
<latexit sha1_base64="ph28gk994A05ueUSYBs/s/XKPz8=">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</latexit>
<latexit sha1_base64="N8yLYOoj5AYHix+9dd0XZoYC7iA=">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</latexit>
<latexit
(a)
0 0 … 0 −0.2
d1 d2 … dM0
E/N
<latexit sha1_base64="5qe38tm1TIukrQEolRuHkxHfY4Y=">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</latexit>
1 <latexit sha1_base64="M4ykC8KE+otofYC5jK3CCMgw3dQ=">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</latexit>
2 <latexit sha1_base64="QsXf1ipPjJGx/F/O8Y+WtvqhhZE=">AAACenicjVFLSgNBEH0Z//E3KoLgJhoiihB6RFBciW5cSUyMCiaGmbGNjfNjphPQIRfwAi5cKYiox3DjBVzkCOIyghsXViYBceGnmu6qfl2v6nW34VkikIzVY0pHZ1d3T29fvH9gcGhYHRndCdyKb/K86Vquv2foAbeEw/NSSIvveT7XbcPiu8bJevN8t8r9QLjOtjz1eNHWy444EqYuCSqp44VAlG39ICx4vrB5rRQWpjZrJTWppVlkiZ+DJNqWcdVbFHAIFyYqsMHhQFJsQUdAYx8aGDzCiggJ8ykS0TlHDXHiViiLU4ZO6AmtZdrtt1GH9s2aQcQ2qYtF0ydmAin2zO5Ygz2xB/bCPn6sFUY1mlpOyRstLvdKw+cTufc/WTZ5ieMv1q+aJY6wHGkVpN2LkOYtzBa/enbRyK1kU+EMu2avpP+K1dkj3cCpvpk3Wzx7ifj/PmBnIa2xtLa1mFxda39FLyYxjVl67yWsYgMZ5KlviCvc4yH2oUwrc8p8K1WJtTlj+GbK4ieTXZLa</latexit>
N
−0.4
<latexit sha1_base64="f/Om9ns79U5UW+fm9ajx31GhhCo=">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</latexit>
<latexit sha1_base64="LpGwqDGskY8RlKlZ/RxIEPm+gfE=">AAACaHicjVG7TgJBFD2sL8QHqIUaGwLBWJGBmGisiDaWPOSRICG7y6ATlt3N7kKChB+wsVRjpYkxxs+w8Qcs+AS1xMTGwstCYixQ72Tm3jlzz71nZhRTE7bDWNcjjY1PTE55p30zs3Pz/sDCYs42GpbKs6qhGVZBkW2uCZ1nHeFovGBaXK4rGs8rtb3+eb7JLVsY+oHTMnmpLh/poipU2SEoWym3451yIByLMteCo4MwhpY0Anc4RAUGVDRQB4cOh2INMmwaRcTAYBJWQpswiyLhnnN04CNug7I4ZciE1mg9ol1xiOq079e0XbZKXTSaFjGDiLBnds967Ik9sBf2ObJW263R19Iirwy43Cz7T1cyH3+y6uQdHH+zftXsoIptV6sg7aaL9G+hDvjNk/NeZicdaa+zG/ZG+q9Zlz3SDfTmu3qb4ukr+P73Abl4NMaisdRmOLE7/Aov1hDCBr33FhLYRxJZ6itwhgtcel6lgLQsrQ5SJc+Qs4QfJoW+AH/Ci5g=</latexit> <latexit sha1_base64="dAFsKvFSAd0iQ6bVMTIjKicCRhI=">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</latexit>
J2 = 0.0
−0.6 J2 = 0.5
h1 h2 h3 … hM
<latexit sha1_base64="tJh5Y4EOy0khgpdHA8GfC+C5xxw=">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</latexit> <latexit sha1_base64="xjZw70dzhBvlMJQHTJCdyj8n8zk=">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</latexit>
<latexit sha1_base64="44BhsYRns2Ez3NOpJghtIjhSdmw=">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</latexit> <latexit sha1_base64="k8EwEfcd9EDn46nQu0mHWr2jzkg=">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</latexit>
… 0.6
(b)
J2 = 0.0
1 2 … N 0.4 J2 = 0.5
C/N
<latexit sha1_base64="wfUMsIEjDz+6gKLu+Aa5wab6ExI=">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</latexit>
<latexit sha1_base64="MOj0b593i4d3Zu62OGy19r/cfxI=">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</latexit> <latexit sha1_base64="VETmeNG9s1dGOXBA1h1Dku2oEkw=">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</latexit>
…
0.2
(b)
0
… ⌧
<latexit sha1_base64="9YTcekeKfiGc69JnOmn+YYIt4k4=">AAACZ3ichVE9SwNBEH05v2L8SFSQgE00KFZhEhTFKmCjXfyIBqLI3bmaI5e7424TiME/YGEbwUpBRPwZNv4BC39CsIxgY+HcJSIi6iy7M/t23szbXc0xDU8SPYeUnt6+/oHwYGRoeGQ0Ghsb3/HsqquLvG6btlvQVE+YhiXy0pCmKDiuUCuaKXa18qp/vlsTrmfY1rasO2K/oh5bxpGhq9KH9qRaPYglKZVZJLbEzyCdosCS6FrOjt1iD4ewoaOKCgQsSI5NqPB4FJEGwWFsHw3GXI6M4FzgFBHmVjlLcIbKaJnXY94Vu6jFe7+mF7B17mLydJmZwCw90R216ZHuqUXvv9ZqBDV8LXX2WocrnIPoWXzr7V9Whb1E6Yv1p2aJIywHWg3W7gSIfwu9w6+dNNtbK5uzjTm6phfWf0XP9MA3sGqv+s2G2LxEhD/g85UTvwc7mVSaUumNhWR2vfsVYUxhBvP83kvIYg055LlvCedo4iLUUqLKpBLvpCqhLmcC30yZ/gDvzYtg</latexit>
<latexit
(c)
3 J2 = 0.0
S z (π, π)
2 J2 = 0.5
Figure 8. (a) Structure of DBM to represent pure states 1

of the extended system during finite temperature calculations. 0
(b) DBM representing the pure state of the extended system 0 1 2 3
|Ψ(T = ∞)⟩ corresponding to the infinite temperature of a T
quantum spin-1/2 system. (c) DBM representing the pure
state of the extended system corresponding to the thermal
equilibrium state Figure 9. Finite-temperature simulations of the J1 –J2
P of the one-dimensional transverse-field Ising
Heisenberg model (J1 = 1) on the 6 × 6 square lattice with
model H = −J i σiz σi+1 z − Γ i σix . As described in Fig. 1(d),
P
the structure in (c) can be redrawn as the structure in (a). The periodic boundary condition. (a) Energy, (b) specific heat, and
magnitudes of the light blue, orange, and purple bonds in (b) (c) spin structure factor at momentum (π, π). Symbols indicate
results obtained by Method (ii) with M/N = 8, which show
and (c) are given by i π4 , 1
2
arcosh 1
tanh(Γδτ )
, 1
2
arcosh(e2Jδτ ), good agreement with numerically exact references (solid curves).
respectively. Adapted with permission from Ref. [52]. Copyright Shaded regions show the size of the error bars of the exact
2021 by the American Physical Society. references. Adapted with permission from Ref. [52]. Copyright
2021 by the American Physical Society.
Sec. 5, and each short-time evolution is reproduced

exactly and analytically within the DBM framework. (i) has the advantage of not requiring cumbersome
As an example, we consider the one-dimensional numerical optimization of DBM parameters, it requires
transverse-field Ising model. Following the procedure Monte Carlo sampling for σ, σ ′ , h, and d spin
shown in Fig. 7, we can reproduce the imaginary-time configurations to calculate physical quantities, which
evolution applied to the infinite-temperature DBM may cause negative sign problems in general quantum
state |Ψ(T = ∞)⟩ shown in Fig. 8(b). Fig. 8(c) systems. On the other hand, in Method (ii), the
shows the constructed DBM structure representing analytical form of the wave function of the extended
|Ψ(T )⟩. Using this DBM structure, we can calculate system Ψ(σ, σ ′ ) can be obtained, and the physical
physical quantities at finite temperatures by Monte quantities can be calculated by Monte Carlo sampling
Carlo sampling over configurations of physical spins using only the σ and σ ′ configurations with the square
σ, ancilla spins σ ′ , hidden spins h, and deep spins d. of the wave function |Ψ(σ, σ ′ )|2 as the weight. This
In Method (ii), we consider the DBM structure allows Method (ii) to be applied to frustrated spin
without deep spins d [i.e., we omit d spins from systems. In the benchmark calculations in Ref. [52],
the structure shown in Fig. 8(a)]. In this case, it was shown that numerically exact solutions can be
the DBM wave function describing the pure state of obtained using Method (i) for frustration-free quantum
the extended spin models that do not suffer from the negative
Q system canPbe written analytically as
Ψ(σ, σ ′ ) = j 2 cosh bj + i (Wji σi + Wji ′ ′ sign problem, and that accurate finite-temperature

σi ) . It has
a form similar to that of Eq. (8) in Sec. 3, and this calculations can be performed using Method (ii) for
analytical form can be used as a variational ansatz. frustrated quantum spin models (Fig. 9).
By numerically optimizing the DBM parameters using
the SR method as described in Sec. 4, the imaginary- 7. Summary and Outlook
time evolution can be approximated as accurately
as possible within the expressive capabilities of the This paper presents a method for embedding quantum
variational wave function. correlations of quantum many-body systems at zero
Methods (i) and (ii) have a relationship similar and finite temperatures in Boltzmann machines. Using
to that of the DBM and RBM algorithms for zero- DBMs, one can analytically reproduce the imaginary-
temperature calculations (Table 3). While Method time Hamiltonian evolution of quantum states, thereby
proving the versatile representative power of the (Grant Number JPMJPF2221).

DBM. This provides a novel method of quantum-
to-classical mapping, but Monte Carlo sampling for Appendix A. Details of SR method
the calculation of physical quantities suffers from
negative sign problems when the method is applied Appendix A.1. Derivation of parameter update rule
to, e.g., frustrated spin systems. When negative signs
become severe in this quantum-to-classical mapping Here, we derive the equation for the parameter update
method, the variational approach using RBM becomes [Eq. (15)] from Eq. (14). Eq. (14) is equivalent to
useful. In the variational approach, at the moment, δ θe = arg max F 2 [e−2δτ H |ψθ ⟩ , |ψθ+δθ ⟩]. (A.1)
one needs to pay attention to how to construct a δθ
variational ansatz. With proper setup, the machine Here, F 2 [e−2δτ H |ψθ ⟩ , |ψθ+δθ ⟩] is the fidelity between
learning method can achieve state-of-the-art accuracy. e−2δτ H |ψθ ⟩ and |ψθ+δθ ⟩:
This state-of-the-art variational method has begun to
F 2 [e−2δτ H |ψθ ⟩ , |ψθ+δθ ⟩]
contribute to solving challenging quantum many-body
problems. ⟨ψθ |e−2δτ H |ψθ+δθ ⟩⟨ψθ+δθ |e−2δτ H |ψθ ⟩
= . (A.2)
In this way, we start to understand that the ⟨ψθ |e−4δτ H |ψθ ⟩ ⟨ψθ+δθ |ψθ+δθ ⟩
machine learning methods (+ appropriate setups) are Expanding this equation to the second order for minute
useful also in quantum physics. In addition to the quantities yields
obvious need for further method improvement, a key X X
question for the future is whether unknown and non- F2 = 1 − δθk Skl δθl + 2δτ gk δθk + 4δτ2 Evar .
trivial physical insights can be drawn from neural k,l k
network methods, beyond a black box that somehow (A.3)
yields good results. It will be important to explore the
boundary between the neural network methods and the Here, Skl is the Fubini-Study metric tensor [see
physical intuition that essential physics is described by Eqs. (12) and (13)]. gk is the gradient of energy with
a small number of parameters. respect to the variational parameter [Eq. (16)], whose
Another important direction is cross-checking detailed form reads
⟨ψθ | H |∂k ψθ ⟩ ⟨ψθ | H |ψθ ⟩ ⟨ψθ |∂k ψθ ⟩

using various variational algorithms, including neural gk = 2Re − .
network methods. For example, in the case of ⟨ψθ |ψθ ⟩ ⟨ψθ |ψθ ⟩ ⟨ψθ |ψθ ⟩
the two-dimensional J1 –J2 Heisenberg model, several (A.4)
studies using different numerical algorithms show at
Evar is the variance of the total energy
least qualitative agreement in the ground-state phase
2
diagram [16, 89–91]. This makes the results more ⟨ψθ | H2 |ψθ ⟩ ⟨ψθ | H |ψθ ⟩

reliable. Such cross-checking will become increasingly Evar = − = ⟨H2 ⟩−⟨H⟩2 .
⟨ψθ |ψθ ⟩ ⟨ψθ |ψθ ⟩
important when the analytical solution is not available. (A.5)
To make comparison among different numerical
algorithms meaningful, a consistent accuracy metric By considering the stationary condition in Eq. (A.3),
is desired. With this in mind, attempts are being we obtain the following equation:
made to build a platform for comparing different
X
−1
δθk = −δτ Skl gl . (A.6)
algorithms [92]. Further development of variational l
approaches, including both classical and quantum
This equation defines the parameter update in the SR
algorithms, and reliable comparisons between methods
scheme.
will certainly contribute to a deeper understanding of
quantum many-body physics.
Appendix A.2. Monte Carlo sampling
Acknowledgments The numerical cost of exactly computing physical
quantities in the SR scheme grows exponentially with
We acknowledge fruitful discussions with A. S. Dar- respect to the number of degrees of freedom. In
mawan, Y. Yamaji, M. Imada, G. Carleo, N. Yosh- practical calculations, we estimate the expectation
ioka, and F. Nori. This work was supported by values using the Monte Carlo method as in the
Grant-in-Aids for Scientific Research (JSPS KAK- variational Monte Carlo (VMC) method [93–96].
ENHI) [Grant Numbers JP23H04869, JP23H04519, We can expand a variational state |ψθ ⟩ with a
JP23K03307, and JP21H01041], MEXT as “Program certain basis {|x⟩} as
for Promoting Researches on the Supercomputer Fu- X
gaku” (Grant Number JPMXP1020230411), and JST |ψθ ⟩ = |x⟩ψθ (x). (A.7)
x
Within this basis, the energy expectation value ⟨H⟩ is [16] Nomura Y and Imada M 2021 Phys. Rev. X 11(3) 031034
given by [17] Astrakhantsev N, Westerhout T, Tiwari A, Choo K, Chen
A, Fischer M H, Carleo G and Neupert T 2021 Phys.
∗ ′
P
⟨ψθ | H |ψθ ⟩ ′ ψ (x)Hxx′ ψθ (x ) Rev. X 11(4) 041021
= xx Pθ 2 (A.8) [18] Li M, Chen J, Xiao Q, Wang F, Jiang Q, Zhao X, Lin R,
⟨ψθ |ψθ ⟩ x |ψθ (x)| An H, Liang X and He L 2022 IEEE Transactions on
P 2 Parallel and Distributed Systems 33 2846–2859
|ψθ (x)| Eloc (x)
= xP 2 (A.9) [19] Rath Y and Booth G H 2022 Phys. Rev. Res. 4(2) 023126
x |ψθ (x)| [20] Roth C, Szabó A and MacDonald A 2023 High-accuracy
′ variational monte carlo for frustrated magnets with deep
with Hxx′ = ⟨x| H |x′ ⟩ and Eloc (x) = x′ Hxx′ ψψθθ(x )
P
(x) . neural networks arXiv:2211.07749
Thus, we can estimate ⟨H⟩ by the Monte Carlo average [21] Reh M, Schmitt M and Gärttner M 2023 Phys. Rev. B
107(19) 195115
of Eloc (x), where Monte Carlo samples are generated [22] Chen A and Heyl M 2023 Efficient optimization of
using the weight proportional to |ψθ (x)|2 . deep neural quantum states toward machine precision
By defining the diagonal operator Ok = arXiv:2302.01941
[23] Saito H 2017 J. Phys. Soc. Jpn. 86 093001
Okloc (σ) ⟨σ| with Okloc (σ) = ∂kψψθ (σ)
θ (σ)
P
σ |σ⟩ , we can [24] Saito H and Kato M 2018 J. Phys. Soc. Jpn. 87 014001
rewrite the equations for Skl and gk as [25] Nomura Y, Darmawan A S, Yamaji Y and Imada M 2017
Phys. Rev. B 96(20) 205152
Skl = Re ⟨Ok† Ol ⟩ − ⟨Ok† ⟩⟨Ol ⟩ , (A.10) [26] Luo D and Clark B K 2019 Phys. Rev. Lett. 122(22) 226401
[27] Han J, Zhang L and Weinan E 2019 Journal of Computa-
and tional Physics 399 108929
[28] Choo K, Mezzacapo A and Carleo G 2020 Nat. Commun.
gk = 2Re⟨HOk ⟩ − 2⟨H⟩Re⟨Ok ⟩, (A.11) 11 2368
[29] Pfau D, Spencer J S, Matthews A G D G and Foulkes
respectively. Similarly with the energy expectation W M C 2020 Phys. Rev. Research 2(3) 033429
value, all the expectation values in Eqs. (A.10) and [30] Hermann J, Schätzle Z and Noé F 2020 Nat. Chem. 12
(A.11) can be estimated using the Monte Carlo method 891–897
with the weight proportional to |ψθ (x)|2 . For example, [31] Stokes J, Moreno J R, Pnevmatikakis E A and Carleo G
2020 Phys. Rev. B 102(20) 205122
⟨Ok ⟩ given by [32] Yoshioka N, Mizukami W and Nori F 2021 Commun. Phys.
P 2 4 106
⟨ψθ |∂k ψθ ⟩ |ψθ (x)| Oloc (x)
⟨Ok ⟩ = = xP 2 (A.12) [33] Inui K, Kato Y and Motome Y 2021 Phys. Rev. Research
⟨ψθ |ψθ ⟩ x |ψθ (x)|
3(4) 043126
[34] Moreno J R, Carleo G, Georges A and Stokes J 2022 Proc.
is computed by the Monte Carlo average of Oloc (x). Natl. Acad. Sci. USA 119 e2122059119
[35] Cassella G, Sutterud H, Azadi S, Drummond N D, Pfau D,
Spencer J S and Foulkes W M C 2023 Phys. Rev. Lett.
References 130(3) 036401
[36] Nomura Y 2020 J. Phys. Soc. Jpn. 89 054706
[1] Becca F and Sorella S 2017 Quantum Monte Carlo Ap- [37] Deng D L, Li X and Das Sarma S 2017 Phys. Rev. X 7(2)
proaches for Correlated Systems (Cambridge University 021021
Press) [38] Deng D L, Li X and Das Sarma S 2017 Phys. Rev. B 96(19)
[2] Bardeen J, Cooper L N and Schrieffer J R 1957 Phys. Rev. 195145
108(5) 1175–1204 [39] Glasser I, Pancotti N, August M, Rodriguez I D and Cirac
[3] Anderson P 1973 Materials Research Bulletin 8 153 – 160 J I 2018 Phys. Rev. X 8(1) 011006
ISSN 0025-5408 [40] Lu S, Gao X and Duan L M 2019 Phys. Rev. B 99(15)
[4] Laughlin R B 1983 Phys. Rev. Lett. 50(18) 1395–1398 155136
[5] Clark S R 2018 Journal of Physics A: Mathematical and [41] Kaubruegger R, Pastori L and Budich J C 2018 Phys. Rev.
Theoretical 51 135301 B 97(19) 195136
[6] Huang Y and Moore J E 2021 Phys. Rev. Lett. 127(17) [42] Choo K, Carleo G, Regnault N and Neupert T 2018 Phys.
170601 Rev. Lett. 121(16) 167204
[7] Carleo G and Troyer M 2017 Science 355 602–606 ISSN [43] Hendry D and Feiguin A E 2019 Phys. Rev. B 100(24)
0036-8075 245123
[8] Smolensky P 1986 Parallel Distributed Processing: Explo- [44] Vieijra T, Casert C, Nys J, De Neve W, Haegeman J,
rations in the Microstructure of Cognition: Foundations Ryckebusch J and Verstraete F 2020 Phys. Rev. Lett.
(Cambridge: MIT Press) 124(9) 097201
[9] Cai Z and Liu J 2018 Phys. Rev. B 97(3) 035116 [45] Czischek S, Gärttner M and Gasenzer T 2018 Phys. Rev. B
[10] Liang X, Liu W Y, Lin P Z, Guo G C, Zhang Y S and He 98(2) 024311
L 2018 Phys. Rev. B 98(10) 104426 [46] Schmitt M and Heyl M 2020 Phys. Rev. Lett. 125(10)
[11] Choo K, Neupert T and Carleo G 2019 Phys. Rev. B 100503
100(12) 125124 [47] Nagy A and Savona V 2019 Phys. Rev. Lett. 122(25) 250501
[12] Ferrari F, Becca F and Carrasquilla J 2019 Phys. Rev. B [48] Hartmann M J and Carleo G 2019 Phys. Rev. Lett. 122(25)
100(12) 125131 250502
[13] Westerhout T, Astrakhantsev N, Tikhonov K S, Katsnelson [49] Vicentini F, Biella A, Regnault N and Ciuti C 2019 Phys.
M I and Bagrov A A 2020 Nat. Commun. 11 1593 Rev. Lett. 122(25) 250503
[14] Szabó A and Castelnovo C 2020 Phys. Rev. Research 2(3) [50] Yoshioka N and Hamazaki R 2019 Phys. Rev. B 99(21)
033075 214306
[15] Nomura Y 2021 J. Phys.: Condens. Matter 33 174003 [51] Irikura N and Saito H 2020 Phys. Rev. Research 2(1)
013284 Parcollet O, Pohle R, Romero I, Schmid M, Silvester

[52] Nomura Y, Yoshioka N and Nori F 2021 Phys. Rev. Lett. J M, Sorella S, Tocchio L F, Wang L, White S R,
127(6) 060601 Wietek A, Yang Q, Yang Y, Zhang S and Carleo G
[53] Roux N L and Bengio Y 2008 Neural Comput. 20 1631–1649 2023 Variational benchmarks for quantum many-body
[54] Montufar G and Ay N 2011 Neural Comput. 23 1306–1319 problems arXiv:2302.04919
[55] Viteritti L L, Ferrari F and Becca F 2022 SciPost Phys. 12 [93] McMillan W L 1965 Phys. Rev. 138(2A) A442–A451
166 [94] Ceperley D, Chester G V and Kalos M H 1977 Phys. Rev.
[56] Jordan P and Wigner E 1928 Z. Physik 47 631–651 B 16(7) 3081–3099
[57] Bravyi S B and Kitaev A Y 2002 Annals of Physics 298 [95] Yokoyama H and Shiba H 1987 J. Phys. Soc. Jpn. 56 1490–
210–226 1506
[58] Gao X and Duan L M 2017 Nat. Commun. 8 662 [96] Yokoyama H and Shiba H 1987 J. Phys. Soc. Jpn. 56 3582–
[59] Levine Y, Sharir O, Cohen N and Shashua A 2019 Phys. 3592
Rev. Lett. 122(6) 065301
[60] Verstraete F, Murg V and Cirac J 2008 Advances in Physics
57 143–224
[61] Orús R 2014 Annals of Physics 349 117 – 158
[62] White S R 1992 Phys. Rev. Lett. 69(19) 2863–2866
[63] White S R 1993 Phys. Rev. B 48(14) 10345–10356
[64] Chen J, Cheng S, Xie H, Wang L and Xiang T 2018 Phys.
Rev. B 97(8) 085104
[65] Affleck I, Kennedy T, Lieb E H and Tasaki H 1987 Phys.
Rev. Lett. 59(7) 799–802
[66] Lu S, Gao X and Duan L M 2019 Phys. Rev. B 99 ISSN
2469-9969
[67] Pei M Y and Clark S R 2021 Entropy 23 ISSN 1099-4300
[68] Melko R G, Carleo G, Carrasquilla J and Cirac J I 2019
Nat. Phys. 15 887–892
[69] Sorella S 2001 Phys. Rev. B 64(2) 024512
[70] Kingma D P and Ba J 2014 Adam: A method for stochastic
optimization arXiv:2009.01821
[71] Amari S I, Kurata K and Nagaoka H 1992 IEEE
Transactions on Neural Networks 3 260–271
[72] Amari S I 1998 Neural Comput. 10 251–276
[73] Stokes J, Izaac J, Killoran N and Carleo G 2020 Quantum
4 269 ISSN 2521-327X
[74] Rumelhart D E, Hinton G E and Williams R J 1986 Nature
323 533–536
[75] Mezzacapo F, Schuch N, Boninsegni M and Cirac J I 2009
New Journal of Physics 11 083026
[76] Lubasch M, Cirac J I and Bañuls M C 2014 Phys. Rev. B
90(6) 064425
[77] Gendiar A and Nishino T 2002 Phys. Rev. E 65(4) 046702
[78] Verstraete F and Cirac J I 2004 Renormalization algorithms
for quantum-many body systems in two and higher
dimensions (Preprint cond-mat/0407066)
[79] Tahara D and Imada M 2008 J. Phys. Soc. Jpn. 77 114701
[80] Gutzwiller M C 1963 Phys. Rev. Lett. 10(5) 159–162
[81] Jastrow R 1955 Phys. Rev. 98(5) 1479–1484
[82] Balents L 2010 Nature 464 199
[83] Zhou Y, Kanoda K and Ng T K 2017 Rev. Mod. Phys. 89
025003
[84] Hu W J, Becca F, Parola A and Sorella S 2013 Phys. Rev.
B 88 060402(R)
[85] Gong S S, Zhu W, Sheng D N, Motrunich O I and Fisher
M P A 2014 Phys. Rev. Lett. 113(2) 027201
[86] Misawa T, Morita S, Yoshimi K, Kawamura M, Motoyama
Y, Ido K, Ohgoe T, Imada M and Kato T 2019 Comput.
Phys. Commun. 235 447–462
[87] Carleo G, Nomura Y and Imada M 2018 Nat. Commun. 9
5322
[88] Freitas N, Morigi G and Dunjko V 2018 Int. J. Quantum
Inf. 16 1840008
[89] Wang L and Sandvik A W 2018 Phys. Rev. Lett. 121 107202
[90] Ferrari F and Becca F 2020 Phys. Rev. B 102(1) 014417
[91] Liu W Y, Gong S S, Li Y B, Poilblanc D, Chen W Q and
Gu Z C 2022 Sci. Bull. 67 1034–1041
[92] Wu D, Rossi R, Vicentini F, Astrakhantsev N, Becca F,
Cao X, Carrasquilla J, Ferrari F, Georges A, Hibat-Allah
M, Imada M, Läuchli A M, Mazzola G, Mezzacapo A,
Millis A, Moreno J R, Neupert T, Nomura Y, Nys J,

Boltzmann Machine and QMP

Uploaded by

Copyright:

Available Formats

Boltzmann Machine and QMP

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Boltzmann Machine and QMP

Uploaded by

Copyright:

Available Formats

Boltzmann machines and quantum many-body

Abstract. Analyzing quantum many-body problems and elucidating the

1. Introduction powerful variational ansatz.

problem limits the applicability of the QMC method h1 h2 h3 … hM

and entanglement entropy

3.2.1. Universal approximator

the limit of large bond dimensions.

states with area-law entanglement entropy.

states of S = 1/2 spin systems: (a) RBM and (b) DBM.

known for Boltzmann machine wave functions (or,

volume-law entanglement entropy.

between MPS and Boltzmann machines

the RBM can reproduce an arbitrary MPS in the

3.2.4. Comparison to tensor networks: short summary 4.1. Variational method

(a) 1D TFI model (b) 1D Heisenberg model (c) 2D Heisenberg model

4.3. Extensions to strongly-correlated Hamiltonians (a) U/t = 4 (b) U/t = 8

geometrical frustration is absent, the QMC method can 4.8 6

However, such cases are special, and general 0.8

Work Method Main target Ref.

first mapped onto quantum spin Hamiltonians. For

10−1 such, the phase of research is shifting from develop-

Here, we describe a quantum state representation

RBM (variational method) DBM (quantum-to-classical mapping)

6. DBM for finite-temperature calculations

So far, we have discussed calculations at zero

Figure 8. (a) Structure of DBM to represent pure states 1

Sec. 5, and each short-time evolution is reproduced

proving the versatile representative power of the (Grant Number JPMJPF2221).

013284 Parcollet O, Pohle R, Romero I, Schmid M, Silvester

You might also like