Exact Ising model simulation on a quantum computer

Alba Cervera-Lierta1,2
1
Barcelona Supercomputing Center (BSC), Barcelona, Spain
2
Institut de Ciències del Cosmos, Universitat de Barcelona, Barcelona, Spain
December 20, 2018

We present an exact simulation of a one- only company that has launched this kind of ser-
dimensional transverse Ising spin chain vice: Rigetti Computing also allows the use of
with a quantum computer. We construct its 19-qubits device on the cloud [3]. Although
an efficient quantum circuit that diagonal- both companies are betting for superconducting
izes the Ising Hamiltonian and allows to qubits, their respective device characterization is
arXiv:1807.07112v3 [quant-ph] 19 Dec 2018

obtain all eigenstates of the model by just not the same: basic gate sets and qubits connec-
preparing the computational basis states. tivity are some of the differences. As more quan-
With an explicit example of that circuit for tum devices are appearing, it is important to find
n = 4 spins, we compute the expected value some methods to test their quality when running
of the ground state transverse magneti- sophisticated quantum algorithms.
zation, the time evolution simulation and Several approaches on computer’s quantumness
provide a method to also simulate thermal have already been tried. The first published ar-
evolution. All circuits are run in IBM and ticle using an IBM device tested the violation of
Rigetti quantum devices to test and com- Bell inequalities by more than two qubits (Mer-
pare them qualitatively. min inequalities) [4], or a recent article tests if the
16-qubit IBM device can be fully entangled by
generating graph states [5]. Other works tried to
1 Introduction exploit different few qubit experiments, such as
In recent years the quantum computing has dived error correcting codes and quantum arithmetics
fully into the experimental realm. Control of [6].
quantum systems has improved so much that On the other hand, the community has not for-
quantum computing devices have become a near got Feynman’s original aim for the proposal of
term reality. These experimental advances rely construction of a quantum computer [7]: the sim-
on some criteria proposed in the 2000’s by DiVi- ulation of quantum systems. Many classical tech-
cenzo [1]: scalable physical system to characterize niques have been developed in that direction, for
the qubits, simple fiducial qubit state initializa- instance quantum Monte Carlo methods [8, 9, 10]
tion, long coherence times (longer than the gate or tensor networks algorithms [11]. However, the
implementation times), universal set of quantum first suffer from the well-known sign problem and
gates and qubit-specific measurement capability. the second are only efficient for slightly entangled
Although there are some candidates that can ful- systems [12]. In the end, very strongly correlated
fill the first criterion, the field is still in an early quantum systems, such as those displaying frus-
stage of development of this technology, where tration, will need a quantum computer to be ef-
the improvement of qubits control is crucial to ficiently simulated [13]. There are some works
accomplish the others. that propose quantum algorithms to construct
Private companies have also joined the field. arbitrary Slater determinants, both in one and
Since 2016, IBM offers cloud based quantum com- two dimensions, to simulate the dynamics of the
putation platform [2]. Any user can run quantum ground state of fermionic hamiltonians, in partic-
algorithms on their two five qubits devices, their ular the Hubbard model [14, 15]. Other propos-
16 qubits device, and their 20 qubits device which als introduce the concept of compressed quantum
is available for hubs and partners. It is not the computation, i.e. simulation of n-spin chain using
log n qubits [16]. This method has been tested in
Alba Cervera-Lierta: a.cervera.lierta@gmail.com
one of the IBM’s quantum computers also simu-

Accepted in Quantum 2018-12-18, click title to verify 1

lating the transverse magnetization of the Ising and in section 6 we present the results of ground
model [17]: the main difference respect to the state transverse magnetization and the time evo-
work proposed in that paper is we have access to lution of | ↑↑↑↑i state and compare the three de-
the whole energy spectrum, which allows us to vices according to them. Finally, the conclusions
simulate time and temperature evolution as well. are exposed in section 7.
In this work, we implement a four-qubit experi-
ment that could be interesting both as a proposal
for testing and comparing devices quality and for
2 Quantum circuit for the Ising Hamil-
its implications in condensed matter physics. We
perform the exact simulation of a spin chain pro- tonian
posed in Ref.[18] with an Ising-type interaction.
The Ising model is one of the most famous ex- Let’s consider the existence of a quantum cir-
actly solvable models, i.e. those models that are cuit that disentangles a given Hamiltonian and
integrable. Actually, the steps to find a quan- transforms its entangled eigenstates into product
tum circuit that diagonalizes the Ising Hamilto- states. This circuit will be represented by an uni-
nian follow the same strategy than the analytical tary transformation Udis
solution of the model. Therefore, the method can
be extended to other integrable models like the e = U † HUdis ,
H dis (1)
Kitaev-honeycomb model, which a circuit has al-
ready been proposed [19]. As we are performing
an exact simulation, we have access to the whole where H is the model Hamiltonian and H e is a
spectrum and not only to ground state: time evo- noninteracting Hamiltonian that can be written
lution and thermal states can be simulated ex- as He = P i σ z . This diagonal Hamiltonian con-
i i
actly as well. This provides a new approach in tains the energy spectrum i of the original one
quantum simulation if an exact circuit is found and its eigenstates correspond to the computa-
for those non trivial models, such as Heisenberg tional basis states. Then, we will have access to
model, which have an ansatz to be solved. In the whole spectrum of the model by just prepar-
particular, for one-dimensional spin chains, the ing a product state and applying Udis .
Bethe ansatz [20] is the most successful method There exist infinite Udis gates for a given
and several proposals exist to simulate and ex- Hamiltonian. In general, to find these disentan-
tend it to two-dimensions using tensor network gling unitaries will be a hard task, probably as
techniques [21]. As the one-dimensional Ising hard as finding a method to diagonalize analyti-
model has analytic solutions for arbitrary num- cally the Hamiltonian. However, for some mod-
ber of spins and the circuit proposed in this paper els we can follow a kind of recipe to construct a
can be efficiently generalized to larger number of disentangling gate. For the case of Ising Hamilto-
qubits, the methods outlined in this work can be nian, the steps to obtain the Udis quantum gate
used to benchmark a quantum computer by see- are based on the analytical solution of the model
ing how this compares against known solutions. [22, 23]: i) Implement the Jordan-Wigner trans-
The paper is structured as follows. In section formation to map the spins into fermionic modes.
2 we describe the method proposed in Ref.[18] to ii) Perform the Fourier transform to get fermions
construct an efficient circuit that diagonalizes the to momentum space. iii) Perform a Bogoliubov
Ising Hamiltonian: the number of gates scales as transformation to decouple the modes with op-
n2 and the circuit depth as n log n. In section 3 posite momentum. Thus, the construction of the
we explain briefly the basic concepts of time evo- disentangling gate can be done by pieces:
lution in quantum mechanics and give a specific
example to be simulated using the circuit derived
Udis = UJW UF T UBog . (2)
in the previous section. In section 4, we propose
two methods to simulate the expected value of an
operator for finite temperature. Section 5 sum- In the following subsections, we derive the quan-
marizes the properties of the three devices used tum gates needed to implement the above trans-
for this work, two from IBM and one from Rigetti, formation.

Accepted in Quantum 2018-12-18, click title to verify 2

2.1 Jordan-Wigner transformation which corresponds with the usual SWAP gate fol-
lowed or preceded by a controlled-Z gate (see ap-
Let’s start with the antiferromagnetic Ising
pendix A).
Hamiltonian with transverse field
n n
+σ1y σ2z · · · σn−1
X X
H= σix σi+1
x z
σny +λ σiz , (3) 2.2 Fourier Transform
i=1 i=1
The next step to solve the Ising model consists on
where λ is the transverse field strength. The sec- getting the fermionic modes to momentum space
ond term has been added to cancel the periodic using the well-known quantum Fourier transform
boundary term, σnx σ1x , after the Jordan-Wigner
transformation in order to solve the system as 1 X n
2πj

n

n
it was infinite. This modified Hamiltonian will b†k = √ exp i k c†j , k = − +1, · · · , .
n j=1 n 2 2
have finite size effects that become negligible as
n grows. (8)
The Jordan-Wigner transformation corre- For n = 2m
for some integer m, this transfor-
sponds to transform the spin operators σ into mation can be implemented with a log-depth cir-
fermionic modes c [24]: cuit and using at most two-body quantum gates.
This method is called fast Fourier transform and
σj + iσjy σjx − iσjy
! x !
consists in two parallel Fourier transformations
c†j
Y Y
cj = σlz , = σlz , over n/2 sites, the even and the odd sites [25]:
l<j
2 2 l<j
(4) n
−1
n−1 2
where cj and c†j
are the fermionic annihilation X
e
2πik
n
j
c†j =
X 2πik 0
j
e n/2 c†2j 0 + e
2πik
n
2πik 0
j
e n/2 c†2j 0 +1 .
and creation operators acting on the vacuum |Ωc i, j=0 j 0 =0
ci |Ωc i = 0, and following the anticommutation (9)
rules {ci , cj } = 0 and {ci , c†j } = δij . After this
transformation the Hamiltonian reads To implement such a transformation we need
1n n a combination of a two-qubit gate, a ‘beam-
c†i ci+1+c†i+1 ci+ci ci+1+c†i c†i+1 +λ c†i ci .
X
X
Hc = splitter’ F2 , and one-qubit gate, the ‘phase-delay’
2 i=1 i=1 ωnk , which applies the so-called twiddle-factor
(5)
e2πik/n :
In terms of the wave function,  
1 0 0 0
X
|ψi = ψi1 ···in |i1 · · · in i  0 √1 √1
  !
0  1 0
F2 =  2 2 , ωnk = ,
 
i1 ,··· ,in =0,1 2πik
0 √1 − √1 0  0e n
ψi1 ···in (c†1 )i1 · · · (c†n )in |Ωc i. (6)
2 2
X
=
 
0 0 0 −1
i1 ,··· ,in =0,1
(10)
Notice that the coefficients ψi1 ···in do not change. where the fermionic anticommutation relation
Then it will not be necessary to implement any has been taken into account in the −1 element
gates on the quantum register to perform this of the F2 matrix.
transformation. However, for now on we should All together, the Fourier transform gate be-
take into account we are dealing with fermionic comes
modes, so any swap between two occupied modes
will carry a minus sign. In terms of quantum
 
1 0 0 0
gates, this is translated into the use of fermionic  √1 e 2πik
 
0 2 √n 0

SWAP gate (fSWAP) each time we exchange two n
Fk = 
 2

. (11)
2πik
modes:  0 √1 − e √n

 2 2
0 

  2πik
1 0 0 0 0 0 0 −e n
 0 0 1 0 
fSWAP =  , (7)
 
 0 1 0 0  The appendix shows the explicit decomposition
0 0 0 −1 of this gate.

Accepted in Quantum 2018-12-18, click title to verify 3

 After the Fourier transformation, the Hamilto-
nian becomes
B1† F1† F0†
n/2
2πk
   
b†k bk
X
Hb = 2 λ − cos
k=−n/2+1
n
2πk  † † F0† F0†
 
+i sin b−k bk + b−k bk , (12)
n
which it is not diagonal yet as modes with oppo-
Figure 1: Quantum circuit that transforms computa-
site momentum are still coupled.
tional basis states into transverse Ising eigenstates. The
two qubit gates F1† and F0† apply the inverse Fourier
2.3 Bogoliubov transformation transform and the B1† with the inverse Bogoliubov trans-
formation. Gates represented with crosses correspond
The last step will consist on finding a transfor-
with the fSWAP gates that take care of the fermion an-
mation which mixes the two modes according to ticommutation relations and can be removed depending
ak = uk bk + ivk c†−k , on the connectivity of the quantum chip.

a†k = uk c†k − ivk c−k . (13) Thus we first undo Bogoliubov transformation
This will be implemented by a two-qubit gate by applying (Bkn )† gates, followed by undoing
which acts over qubits that represent opposite the Fourier transform using the (Fkn )† gates and
momenta. For the case of Ising model, this gate finally undo the Jordan-Wigner transformation
is which, fortunately, does not need from any gate
     as has been explained in the previous section.
θk θk
cos 2 0 0 i sin 2 For n = 4, the Bogoliubov modes are ±3π/2
 
 0 10 0  and ±π/2, so we need two Bogoliubov gates. No-
Bkn = ,
 
 0 01 0  tice that we have removed the B0† gate from the
circuit of Figure 1; this gate corresponds with
    
i sin 2 0 0 cos θ2k
θk
  the identity for λ > 1 and exchange qubits in the
λ−cos( 2πk
) same state for λ < 1, i.e. |00i → −i|11i and
n
θk = arccos  q 2
 ,(14)
|11i → −i|00i. As the initial state for λ < 1
2πk
(λ−cos( )) +sin2 ( 2πk
n )
n
is the |0001i and B0† is applied over the last two
and its decomposition in basic gates is shown in qubits, it does not affect this state and we can
the appendix. avoid it. If we want to obtain an excited state
Then, we have finally arrived to the diagonal which eigenstate in the diagonal basis contains
Hamiltonian: |00i or |11i states, then we should only apply bit
n/2 flip gates over the last two qubits to implement
ωk a†k ak , the B0† gate.
X
H
e = Ha = (15)
k=−n/2+1 The circuit shown in Figure 1 also contains
r fSWAP gates represented with crosses. These will
2
be necessary if even and odd qubits are not phys-
  
2πk 2πk
where ωk = λ − cos n + sin2 n .
ically connected and, as much, they will increase
the total number of gates in n2 . We can eliminate
2.4 n = 4 spin chain them if the implementation is done in the ibmqx5
The explicit circuit for a n = 4 chain is shown in device, which allow us to save up to 16 gates of
Figure 1. First, we prepare the initial state as the depth according to IBM gate set, but they are in-
ground state for the diagonal Hamiltonian H: e dispensable for the implementation in the other
( IBM device, ibmqx4, as well as in Rigetti’s 19-
|0000i for λ > 1, qubit chip.
|gsi = (16)
|0001i for λ < 1.

The circuit strategy consists in undoing the

steps that diagonalize the Ising Hamiltonian.

Accepted in Quantum 2018-12-18, click title to verify 4

3 Time evolution
Once we have the Udis circuit, we have access
to the whole Ising spectrum by only implement-
ing this gate over the computational basis states.
This allows us to perform exactly time evolution,
where the characterization of all states is needed.
The time evolution of a given state driven by a
time-independent Hamiltonian is described using
the time evolution operator U (t) ≡ e−itH :
e−iti |Ei ihEi |ψ0 i, (17)
X
|ψ(t)i = U (t)|ψ0 i = Figure 2: Transverse magnetization of n = 4 Ising spin
i chain as a function of temperature β = 1/(kB T ) and
transverse field λ. The system undergoes a transition,
where |ψ0 i is the initial state and i are the ener-
from antiferromagnetic to paramagnetic, as λ increases.
gies of the Hamiltonian states |Ei i. Then, if |ψ0 i For zero temperature (β → ∞) the transition point is
is an eigenstate of H there is no change in time located at λ = 1, whereas for finite temperature there
(steady state) and therefore the expected value of is a quantum critical region around λ = 1.
an observable O will be constant in time. On the
contrary, and if [H, O] 6= 0, the expected value
from which we can obtain the expected value of
will show an oscillation in time given by
transverse magnetization, Mz = 21 hσz i.
e−it(i −j ) hψ0 |Ej ihEj |O|Ei ihEi |ψ0 i.
X
hO(t)i =
i,j
(18) 4 Thermal simulation
We can take advantage from the fact that the
When a quantum system is exposed to a heat
eigenstates of the non-interacting Hamiltonian H e
bath its density matrix at thermal equilibrium
are the computational basis states and, as we
is characterized by thermally distributed popula-
have solved the model, we also know all energies
tions of its quantum states following a Boltzmann
i . Then, it is straightforward to construct the
distribution:
time evolution of a given state |ψ0 i by only ex-
pressing it in the computational basis and adding e−βH 1 X −βi
ρ(β) = = e |Ei ihEi |, (22)
the corresponding factors e−iti . After that, we Z Z i
only need to implement Udis gate over this state
where β = 1/(kB T ), Z = i e−βi is the parti-
P
to obtain the time evolution driven by the Ising
Hamiltonian. tion function and i and |Ei i are the energies and
As example, we compute the time evolution of eigenstates of the Hamiltonian H. The expected
the expected value of transverse magnetization. value of some operator O for finite temperature
We take all spins aligned in the positive z direc- is computed as
tion as initial state, i.e. | ↑↑↑↑i, which in the 1 X −βi
computational basis is the |0000i state. First, we hO(β)i = Tr[Oρ(β)] = e hEi |O|Ei i.
Z i
have to express this state in the H e basis, which
† (23)
using Udis become
† Simulate thermal evolution according to Ising
|ψ0 i = Udis |0000i = cos φ|0000i + i sin φ|1100i,
Hamiltonian is, again, straightforward once we
√ (19)
have Udis gate, since it consists on preparing the
with φ = arccos(λ/ 1 + λ2 )/2. Then, we apply
corresponding state in the He basis and apply Udis
the time evolution operator to obtain |ψ(t)i:
 √  circuit. In the case of thermal evolution, |Ei i
1+λ2
|ψ(t)i = cos φ|00i + ie4it sin φ|11i ⊗ |00i. states are the states of the computational basis,
(20) so no further gates are needed to initialize qubits
Analytically, apart from the corresponding combination of X
 √  gates.
1 + 2λ2 + cos 4t 1 + λ2 At that point, we can perform an exact simu-
hσz i = , (21) lation or sampling. In the first case, we run the
2 + 2λ2

Accepted in Quantum 2018-12-18, click title to verify 5

 ducting transmon qubits which are accessible on
the cloud, both interactively in their webpage
(Quantum Composer )[2] or using a software de-
velopment kit (QISKit). Currently, three quan-
tum chips are available for the general public: two
of 5 qubits, ibmqx2 and ibmqx4, and one of 16
qubits, ibmqx5 [27]. When we performed the ex-
periments ibmqx2 was offline, so we have only
Figure 3: IBM quantum chips used in this work. Up fig-
used ibmqx4 and ibmqx5.
ure corresponds to ibmqx4 and down figure to ibmqx5.
Arrows indicate the directionality of CNOT gates (con- All backends work with an universal gate set
trol Ý target). composed by one-qubit rotational and phase
gates
circuit to obtain the expected value of the ob-  
iλ
servable taking as initial state all states in the √1 − e√2
!
1 0 2
computational basis and average them with their U1 (λ) = , U2 (λ, φ) =  e√iφ ei(λ+φ)
,
0 eiλ √
corresponding energies. This is done classically 2 2
!
once we have the expected values of each state. cos(θ/2) −eiλ sin(θ/2)
On the other hand, we can perform a more re- U3 (θ, λ, φ) = ,
eiφ sin(θ/2) ei(λ+φ) cos(θ/2)
alistic simulation by sampling all states accord-
(24)
ing to Boltzmann distribution. First, we need to
prepare classically a random generator that re- and a two-qubit gate, the controlled-X or CNOT
turns one of the computational states following gate.
the distribution e−βi . Then, we run the circuit The differences between the devices, apart from
many times and compute the expected value of the number of qubits, come from the qubits con-
the operator by preparing as initial state the one nectivity or topology and the role that each qubit
returned by the generator each time. plays when applied a CNOT gate (control or tar-
The first method demands more runs of the get). Figure 3 shows the connectivity of the
experiment, N × 2n , needed for the computation used devices. Each qubit in the 5-qubit device is
of each expected value. No statistical errors come connected with other two except the central one
from the averaging part, as it is done classically. which is connected with the other four. Qubits
For the second method, with only N runs we will in the 16-qubit device are connected with three
obtain a value
√ for the observable with a statistical neighbors in a ladder-type geometry. Both, the
error of 1/ N . one-directionality of CNOT gate and the qubits
For n = 4 the transverse magnetization can be connectivity, are crucial for the quantum circuit
computed analytically and it is shown in Figure 2. implementation. If the circuit demands an inter-
At zero temperature, i.e. β → ∞, the system un- action between qubits that are not physically con-
dergoes to phase transition in the thermodynamic nected, we should implement SWAP gates which
limit, i.e. n → ∞, from antiferromagnetic to will increase our circuit depth and the probability
paramagnetic, at the corresponding critical point of errors in our final result. Moreover, each time
of λ = 1. As temperature increases (β decreases), we need to implement a CNOT gate using as a
the system have a critical region around λ = 1 control qubit a physical qubit which is actually
until the temperature is high enough to disor- a target, we have to invert the CNOT direction
der all spins, independently of the transverse field using Hadamard gates which, again, will increase
strength (limit β → 0) [26]. the circuit depth and the error probability.
For our propose, ibmqx5 is the best choice for
5 Implementation on a quantum com- the implementation of the n = 4 circuit. We can
use any of the squares and identify upper qubits
puter as 0 and 2 and lower qubits as 1 and 3: according
to the circuit of Figure 1, we will not need to
5.1 IBM Quantum Experience
use any fSWAP gates. We should only take into
Since 2016, IBM is providing universal quan- account which qubits are control or target to try
tum computer prototypes based on supercon- to reduce the times that we have to invert the

Accepted in Quantum 2018-12-18, click title to verify 6

 three devices.
√ All points contain a statistical er-
ror of 1/ N with N = 1024 which comes from
the average over all runs to compute the expected
value. The other error sources are discussed qual-
itatively in the following paragraphs.
The best performance come from the ibmqx5
device. This is an expected result as we do not
Figure 4: Rigetti’s 19-qubit processor ‘Acorn’. Lines in- need from fSWAP gates because the qubits con-
dicate the two-qubit connection ruled by a controlled-Z nectivity. On the other hand, Rigetti’s device
gate. For technical reasons, qubit 3 is offline. performs better than the ibmqx4, even though
the number of gates is very similar.
The simulation approaches better to the pre-
CNOT direction. diction for low λ. The explanation could come
from how affect the experimental error sources
to the magnetization. Assuming that two-qubit
5.2 Rigetti Computing: Forest
gates implementation take several hundreds of ns
At the end of 2017, Rigetti Computing launched and single qubit gates around one hundred of ns,
a 19-qubit processor, ‘Acorn’, that can be used errors coming from decoherence are expected to
in the cloud through a development environment be low, as these times are around 50 µs. On the
called Forest [3]. It includes a python toolkit, other hand, errors coming from the gate imple-
pyQuil, that allows the users to program, sim- mentation are cumulative and probably the most
ulate and run quantum algorithms in a similar important error source. It is not negligible nei-
way as IBM’s QISKit. The chip is made of of 20 ther errors coming from qubits readout, which
superconducting transmon qubits but for some can induce a bit flip.
technical reasons, qubit 3 is off-line and cannot The analysis of the results become more clear if
interact with its neighbors, so it is treated as a we look at the exact ground state wave function:
19-qubit device. 
 α(|0001i−|0010i+|0100i−|1000i)

Currently, Rigetti’s gate set is formed by three 

 +|0111i−|1011i+|1101i−|1110i


one-qubit rotational gates 1
|gsi= for λ<1,
θ θ θ
N

RX (θ) = ei 2 σx , RY (θ) = e−i 2 σy , RZ (θ) = ei 2 σz ,  α(|0011i−|0110i+|1001i+|1100i)




(25) 
+2|1111> for λ>1,
and a two-qubit gate, controlled-Z. This two- (26)
qubit gate has the advantage of bi-directionality √ √ √
as the result is the same independently of which where α = λ − 1 + λ2 and N = 2 2 1 + λα.
is the control qubit. For that reason, the connec- As λ increases, the amplitude for the states pro-
tivity of the device shown in Figure 4 does not portional to α goes to zero. That means that any
specify the direction of the two-qubit gate. error occurring for λ > 1 is dramatic as it will af-
The qubit topology is very different from IBM’s fect the state with higher probability amplitude,
devices: some qubits are connected with three the |1111i. Then, any error in that regime will
neighbors and others with two in a zigzag-type ge- inevitably cause a decrease in magnetization. On
ometry. Then, we can not do without the fSWAP the other hand, errors in some states for λ < 1
gates, which means that the circuit depth will be can be compensated in average for the other ele-
greater than the ibmqx5’s. On the other hand, it ments with the same probability amplitude.
will be comparable with the ibmqx4, which also Similar results are obtained for the time evolu-
needs from these gates. tion simulation. Figure 6 shows the results for the
simulation of the | ↑↑↑↑i state transverse magne-
tization as it was explained in Section 3. Since
6 Results and discussion for the preparation of the initial state it is neces-
sary to implement more gates, we only show the
Figure 5 shows the results of the exact simulation results for the ibmqx5 device, which is the one
of ground state transverse magnetization for the that can afford this extra circuit depth.

Accepted in Quantum 2018-12-18, click title to verify 7

 Figure 6: Time evolution simulation of transverse mag-
netization, hσz i, for the state | ↑↑↑↑i of a n = 4 Ising
spin chain. Left plot compares the exact result with the
Figure 5: Expected value of hσz i of the ground state of experimental run in the ibmqx5 chip for different val-
a n = 4 Ising spin chain as a function of transverse field ues of λ. Right plots detailed the results for each λ to
strength λ. Solid line represents the exact result in com- compare them with the theoretical values. Although the
parison with the experimental simulations represented by magnetization is lesser than expected, the oscillations
scatter points. The best simulation comes from ibmqx5 follow the same theoretical pattern.
device, which is an expected result since the number of
gates used is lesser than with the other devices because
of qubits connectivity. do so, we programmed the algorithm proposed in
Ref.[18] in IBM and Rigetti quantum chips. We
As expected from the previous result, points have simulated the expected value of ground state
that represent higher magnetization have more transverse magnetization as well as the time evo-
error respect the theoretical values. However, it is lution of the state of all spins aligned. We have
remarkable that the relations among the different also provided two methods to compute thermal
points for different values of transverse magnetic evolution of some operator using the same cir-
field are proportionally correct. The oscillations cuit: exact simulation or sampling.
take place for lower values of hσz i, have lower The circuit presented allows to compute all
amplitudes and are a little bit shifted to the left: eigenstates of the Ising Hamiltonian by just ini-
even though, they cross each other at the corre- tializing the qubits in one of the states of the com-
sponding points and increase and decrease pro- putational basis. It is then a implementation of
portionally to the exact result. That is a clear a Slater determinant with a quantum computer.
indicator that the error sources in the quantum Since the one-dimensional Ising model is an ex-
device are systematic, as the result does not de- actly solvable model, which means that we can
pend on the state preparation. compute analytically all the states and energies
Notice that in this work we have computed the for any number of spins, and the circuit is effi-
transverse magnetization instead of the staggered cient, the number of gates scales as n2 and the cir-
magnetization, i.e. Mx =
P i x cuit depth as n log n, it can represent a method to
i (−1) σi , which
is the order parameter for the antiferromagnetic test quantum computing devices of any size. As
Ising model. For our purpose, it is more nat- has been shown, it is also a hard test since this
ural to compute hσz i since the states obtained model is strongly correlated and both the simula-
with these quantum devices are expressed in the tion of the phase transition surrounding and time
σz basis. However, it will be straightforward to evolution require a high qubits control.
compute Mx as the only change needed appears The best performance has been obtained with
in the classical post-processing part. the ibmqx5 chip, although the error respect to the
theoretical prediction is large in the paramagnetic
phase of the model. A possible reason why this
7 Conclusions chip shows better results than the others comes
from the number of gates used in the quantum
In this work we have implemented the exact sim- circuit, as the qubits connectivity in that device
ulation of a one-dimensional Ising spin chain with allows us to save all the fSWAP gates. On the
transverse field in some quantum computers. To other hand, Rigetti’s chip performs better than

Accepted in Quantum 2018-12-18, click title to verify 8

the ibmqx4 chip, even though both implemented Appendix: Gate decomposition
circuits have the same gate depth.
The paramagnetic phase is difficult to simulate
7.1 Fermionic-SWAP
due to the fact that any error that can induce a The Jordan-Wigner transformation do not need
qubit bit flip will produce a decrease in magne- from any quantum gate, but as it transforms the
tization, as can be traced out from the ground spin operators σ into fermionic modes c, any swap
state wave function of Eq.(26). However, and between qubits should obey the fermionic anti-
taking into account this fact, the time evolution commutation relations, i.e. the exchange between
simulation is reasonably good since the expected two occupied modes carries a minus sign. This is
oscillations for different transverse magnetic field represented with the use of fermionic SWAP gate
strengths are shifted to the left and have lower (fSWAP), decomposed in basic gates in Figure 7.
amplitude and magnetization, but are also pro- From the point of view of IBM’s implemen-
portional each other as are the theoretical values. tation, at least one CNOT should be inverted
As a final remark, this circuit is also interest- to fit the circuit to the qubits connectivity; this
ing from a point of view of condensed matter can be easily done using the identity (H1 ⊗
−−−−→ ←−−−−
physics as specific methods to simulate exactly H2 )CNOT(H1 ⊗ H2 ) = CNOT. In addition,
time and thermal evolution are provided. This controlled-Z gate could be implemented using two
can open the possibility of simulate other inter- Hadamard gates and a CNOT, as it is also shown
esting models: integrable, like Kitaev Honeycomb in Figure 7.
model [19], or with an ansatz, like Heisenberg From Rigetti’s implementation point of view,
model. controlled-Z gates are part of their basic gate set
and, although CNOT gate and H are not, they are
included in pyquil language, so the quantum pro-
grammer should not care about decompose them
Notes in terms of the other gates (except to keep in
mind the circuit depth will increase with the use
The program used for this work for the IBM of non-basic gates).
quantum devices was awarded with the IBM
“Teach Me QISKit” award [28].
× • • • •
Recently, Rigetti computing changed the quan-
tum device to one of 8 qubits. The results shown
× Z ≡ • H H
in this work correspond to the previous device of
19 qubits.
Figure 7: Fermionic SWAP gate.

Acknowledgements 7.2 Fourier transform

The building blocks of Fourier transform gate
We acknowledge use of the IBM Q experi-
are represented by the quantum gate of Eq.(11),
ence for this work. The views expressed are
which decomposition is shown in Figure 8. Its im-
those of the authors and do not reflect the of-
plementation requires the controlled-Hadamard
ficial policy or position of IBM or the IBM
gate, decomposed in Figure 9. Phase gate Ph is
Q experience team. We also acknowledge
included in Rigetti’s pyquil language and it is
use of the Rigetti computing device as well
equivalent to IBM’s U1 gate (see Eq. (24)). In
as the help and availability of Rigetti staff.
particular, Ph(π/2) and Ph(π/4) correspond to
This work has been possible with the support
S and T gates respectively.
of FIS2015-69167-C2-2-P and FIS2017-89860-P
(MINECO/AEI/FEDER,UE) grants. Finally, we
would like to thank the discussions with Quantic
group members, in particular with José Ignacio
Latorre.

Accepted in Quantum 2018-12-18, click title to verify 9

   perform a bit-flip gate (Pauli-X or X gate), on
Ph 2πk
• H • • fourth qubit to initialize the circuit for λ < 1.
n
Fkn The example given for time evolution simula-
≡ • tion requires from the preparation as the initial
Z
state the one shown in Eq.(20). This can be done
by applying a RY (φ) gate on the first qubit to
Figure 8: Decomposition of the building block of Fourier introduce the φ angle, followed by a√phase gate
transform gate. The controlled-Hadamard gate is shown 2
to introduce the evolution phase e4it 1+λ and a
in Figure 9. CNOT gate between first and second qubits.

• • References
H ≡ S† H T† T H S [1] D. P. DiVincenzo, Fortschritte der Physik
48, 771 (2000).
[2] IBM Quantum Experience, https://www.
Figure 9: Controlled-Hadamard gate.
research.ibm.com/ibm-q/.
[3] R. Smith, M. J. Curtis and W. J. Zeng,
7.3 Bogoliubov transformation
arXiv:1608.03355 [quant-ph] (2016).
Bogoliubov transformation is implemented using [4] D. Alsina and J. I. Latorre, Phys. Rev. A 94,
Bkn gates written in Eq. (14). The explicit de- 012314 (2016).
composition is shown in Figure 10, where the
controlled-RX gate (shown in Figure 11 has been [5] Y. Wang, Y. Li, Z. Yin and B. Zeng, npj
decomposed using the methods of Ref.[29]). Ro- Quantum Information 4, 46 (2018).
tational gates are part of the basic Rigetti gate set [6] J. S. Devitt, Phys. Rev. A 94, 032329 (2016).
and are equivalent to IBM’s gates RX ≡ U3 (φ = [7] R. P. Feynman, Int. J. Theor. Phys. 21, 467
0, λ = π), RY ≡ U3 (φ = λ = 0) and RZ ≡ U1 . (1982).
[8] M. H. Kalos, Phys. Rev. 128, 1791 (1962).
• [9] B.L. Hammond, W. A. Lester Jr. and P.J.
Bkn Reynolds, MonteCarlo Methods in Ab Initio
≡ X • RX (θk ) • X Quantum Chemistry, World Scientific, Sin-
gapore (1994).
Figure 10: Bogoliubov gate decomposition. Controlled- [10] N. S. Blunt, T. W. Rogers, J. S. Spencer and
RX gate needed is shown in Figure 11. W. M. C. Foulkes, Phys. Rev. B 89, 245124
(2014).
• • •
[11] R. Orús, Ann. Phys. 349, 117 (2014).
RX (θk ) ≡ A B C [12] G. Vidal, Phys. Rev. Lett. 91, 147902
(2003).
[13] G. Ortiz, J. E. Gubernatis, E. Knill, and R.
Figure 11: Controlled-RX gate decomposition
in terms
of the rotational gates A = RZ π2 RY θ2 , B =

Laflamme, Phys. Rev. A 64, 022319 (2001).
RY − θ2 and C = RZ − π2 .



[14] D. Wecker, M. B. Hastings, N. Wiebe, B. K.
Clark, C. Nayak and M. Troyer, Phys. Rev.
7.4 Initial state preparation A 92, 062318 (2015).

The ground state of the n = 4 Ising model in [15] Z. Jiang, K. J. Sung, K. Kechedzhi, V. N.
the diagonal basis is |0000i for λ > 1 and |0001i Smelyanskiy and S. Boixo, Phys. Rev. Appl.
for λ < 1. Qubits are always initialized in the 9, 044036 (2018).
|0i state both in IBM and Rigetti devices. Then, [16] B. Kraus, Phys. Rev. Lett. 107, 250503
to compute the ground state, we only need to (2011).

Accepted in Quantum 2018-12-18, click title to verify 10

[17] M. Hebenstreit, D. Alsina, J. I. Latorre and
B. Kraus, Phys. Rev. A 95, 052339 (2017).
[18] F. Verstraete, J. I. Cirac and J. I. Latorre,
Phys. Rev. A 79, 032316 (2008).
[19] P. Schmoll and R. Orús, Phys. Rev. B 95,
045112 (2017).
[20] H. Bethe, Z. Phys. 71, 205 (1931).
[21] V. Murg, V. E. Korepin and F. Verstraete,
Phys. Rev. B 86, 045125 (2012).
[22] E. Lieb, T. Schultz and D. Mattis, Ann.
Phys. 16, 407 (1961).
[23] S. Katsura, Phys. Rev. 127, 1508 (1962).
[24] P. Jordan and E. Wigner, Z. Phys. 47, 631
(1928).
[25] A. J. Ferris, Phys. Rev. Lett. 113, 010401
(2014).
[26] S. Sachdev, Quantum Phase Transitions,
Cambridge University Press, Cambridge
(1999).
[27] Device specifications: https:
//github.com/Qiskit/
qiskit-backend-information/tree/
master/backends.
[28] Official announce of IBM “Teach
Me QISKit" award winnerhttps:
//www.ibm.com/blogs/research/2018/
06/teach-qiskit-winner/.
[29] A. Barenco, C. H. Bennett, R. Cleve, D.
P. DiVincenzo, N. Margolus, P. Shor, T.
Sleator, J. Smolin and H. Weinfurter, Phys.
Rev. A 52 3457 (1995).

Accepted in Quantum 2018-12-18, click title to verify 11