See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/338138135

Quantum approximate optimization of the exact-cover problem on a

superconducting quantum processor

Preprint · December 2019

CITATIONS READS

0 159

15 authors, including:

Andreas Bengtsson Pontus Vikstål

Chalmers University of Technology Chalmers University of Technology
23 PUBLICATIONS   182 CITATIONS    4 PUBLICATIONS   1 CITATION   

SEE PROFILE SEE PROFILE

Marika Svensson Anton Frisk Kockum

Chalmers University of Technology Chalmers University of Technology
5 PUBLICATIONS   1 CITATION    54 PUBLICATIONS   2,275 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Giant atoms View project

Krantz NanoArt - Visualizes your research View project

All content following this page was uploaded by Anton Frisk Kockum on 01 January 2020.

The user has requested enhancement of the downloaded file.

 Quantum approximate optimization of the exact-cover problem on a superconducting
quantum processor

Andreas Bengtsson,1 Pontus Vikstål,1 Christopher Warren,1 Marika Svensson,2, 3 Xiu Gu,1
Anton Frisk Kockum,1 Philip Krantz,1 Christian Križan,1 Daryoush Shiri,1 Ida-Maria Svensson,1
Giovanna Tancredi,1 Göran Johansson,1 Per Delsing,1 Giulia Ferrini,1 and Jonas Bylander1, ∗
1
Microtechnology and Nanoscience, Chalmers University of Technology, SE-412 96, Göteborg, Sweden
2
Computer Science and Engineering, Chalmers University of Technology, SE-412 96, Göteborg, Sweden
3
Jeppesen, SE-411 03, Göteborg, Sweden
(Dated: December 24, 2019)
Present-day, noisy, small or intermediate-scale quantum processors—although far from fault-
tolerant—support the execution of heuristic quantum algorithms, which might enable a quantum
advantage, for example, when applied to combinatorial optimization problems. On small-scale
arXiv:1912.10495v1 [quant-ph] 22 Dec 2019

quantum processors, validations of such algorithms serve as important technology demonstrators.

We implement the quantum approximate optimization algorithm (QAOA) on our hardware platform,
consisting of two transmon qubits and one parametrically modulated coupler. We solve small
instances of the NP-complete exact-cover problem, with 96.6% success probability, by iterating the
algorithm up to level two.

Quantum computing promises exponential computa- instance, the exact-cover problem can provide feasible
tional speedup in a number of fields, such as cryptography, solutions to airline planning problems such as tail assign-
quantum simulation, and linear algebra [1]. Even though ment [17]. Currently, this is solved by well-developed op-
a large, fault-tolerant quantum computer is still many timization techniques in combination with heuristics. By
years away, impressive progress has been made over the leveraging heuristic quantum algorithms such as QAOA,
last decade using superconducting circuits [2–4], leading the current approach can be augmented and might provide
to the noisy intermediate-scale quantum (NISQ) era [5]. high-quality solutions while reducing the running time.
It was predicted that NISQ devices should allow for “quan- Applying QAOA to instances of the exact-cover problem
tum supremacy” [6], that is, solving a problem that is extracted from real-world data in the context of the tail
intractable on a classical computer in a reasonable time. assignment has been numerically studied with 25 qubits,
This was recently demonstrated on a 53-qubit processor by corresponding to 25 routes and 278 flights [18]. Other
sampling the output distributions of random circuits [7]. quantum algorithms for solving the exact-cover problem,
Two of the most prominent NISQ algorithms are the specifically quantum annealing, have been considered in
quantum approximate optimization algorithm (QAOA), Refs. [19–21].
for combinatorial optimization problems [8–10], and the All NP-complete problems can be formulated in terms
variational quantum eigensolver (VQE), for the calcula- of finding the ground state of an Ising Hamiltonian [22].
tion of molecular energies [11–13]. QAOA is a heuristic QAOA aims at finding this state by applying two non-
algorithm that could bring a polynomial speedup to the commuting Hamiltonians, B̂ and Ĉ, in an alternating
solution of specific problems encoded in a quantum Hamil- sequence (with length p) to an equal superposition state
tonian [14, 15]. Moreover, QAOA should produce output of n qubits [visualized in Fig. 1(a)],
distributions that cannot be efficiently calculated on a p h i  |0i + |1i ⊗n
classical computer [16].
Y
~ =
|~γ , βi e −iβi B̂ −iγi Ĉ
e , (1)
QAOA is a hybrid algorithm, as it is executed on both i=1
2
a classical and a quantum computer. The quantum part
consists of a circuit with p levels, where better approxima- where γi and βi are (real) variational angles. The first
tions to the solution of the encoded problem are generally Hamiltonian in the sequence is the Ising (cost) Hamilto-
achieved with higher p. In this Letter, we report on using nian specifying the problem,
our superconducting quantum processor to demonstrate n
QAOA with up to p = 2, enabled by adequately high
X X
Ĉ = hi σ̂iz + Jij σ̂iz σ̂jz , (2)
gate fidelities. We solve instances of the NP-complete i=1 i<j
exact-cover problem with 96.6% success probability. For
p > 1, the QAOA solution cannot be efficiently calculated and the second is a transverse field (mixing) Hamiltonian
on a classical computer, as the computational complexity defined by
scales doubly exponentially in p [8].
n
Our interest in solving the exact-cover problem orig-
X
B̂ = σ̂ix , (3)
inates from its use in many real-world applications, for i=1
 2

(a) level i = 1 level i = p

H ... TABLE I. The four different exact-cover problems available
with two subsets, and their solutions and respective sets of
⊗n
H ...
Optimize coefficients in the Ising Hamiltonian Ĉ = h1 σ̂1z +h2 σ̂1z +J σ̂1z σ̂2z .
|0 e−iγ1 Ĉ e−iβ1 B̂ e−iγp Ĉ e−iβp B̂  Ĉ|γ , β

.. .. .. .. γ , β|
. . . . # Subsets h1 h2 J Solution
H ... A S1 = {x1 , x2 }, -1/2 0 1/2 |10i
S2 = {x1 }
 B S1 = {x1 , x2 }, -1 0 0 |10i or |11i
Update variational angles (γ , β)
S2 = {}
(b) level i
C S1 = {x1 }, -1/2 -1/2 0 |11i
Rz (2γi h1 ) Rx (2βi ) S2 = {x2 }
H Rx (2γi J) H Rz (2γi h2 ) Rx (2βi ) D S1 = {x1 , x2 }, 0 0 1 |10i or |01i
S2 = {x1 , x2 }

FIG. 1. (a) The quantum approximate optimization algorithm

(QAOA) for a problem specified by the Ising Hamiltonian
or under-rotations can be compensated for by a change
Ĉ. An alternating sequence of two Hamiltonians (Ĉ and
in the variational angles [12].
B̂) is applied to an equal superposition of n qubits. After
measurement of the qubit states, a cost is calculated, which
We apply QAOA to the exact-cover problem, which
a classical optimization algorithm minimizes by varying the reads: given a set X and several subsets Si containing
~ (b) Our implementation of one QAOA level with
angles ~γ , β. parts of X, which combination of subsets include all
n = 2 using controlled-phase and single-qubit gates. The elements of X just once? Mathematically speaking, this
background color of each gate identifies which part in (a) it combination of subsets should be disjoint, and their union
implements. should be X. This problem can be mapped onto an
Ising Hamiltonian, where the number of spins equals the
x(z)
number of subsets, while the size of X can be arbitrary.
where hi and Jij are real coefficients, and σ̂i are the Let us consider n = 2, for which the two-spin Ising
Pauli X (Z) operators applied to the ith qubit. Hamiltonian is
The ground state of Eq. (2) corresponds to the lowest-
energy state. We therefore define the energy expectation Ĉ = h1 σ̂1z + h2 σ̂2z + J σ̂1z σ̂2z . (5)
value of Eq. (1) as a cost function
The exact-cover problem is mapped onto this Hamiltonian
n
X X by choosing hi and J as follows [23]:
~ = h~γ , β|
F (~γ , β) ~ Ĉ|~γ , βi
~ = hi hσ̂iz i + Jij hσ̂iz σ̂jz i.
i=1 i<j J > min(c1 , c2 ), (6)
(4) h1 = J − 2c1 ,
This cost function is evaluated by repeatedly preparing
~ on a quantum processor. To find h2 = J − 2c2 ,
and measuring |~γ , βi
the state that minimizes Eq. (4), a classical optimizer is where ci is the number of elements in subset i, and J > 0
used to find the optimal variational angles ~γ ∗ , β~ ∗ . For if the two subsets share at least one element. We are
a high enough p, |~γ ∗ , β ~ ∗ i is equal to the ground state free to choose J, as long as it fulfills the criterion in
of Ĉ and hence yields the answer to the optimization Eq. (6). For example, consider X = {x1 , x2 } and two
problem [8]. However, for algorithms executed on real subsets S1 = {x1 , x2 } and S2 = {x1 }. This gives c1 = 2
hardware without error correction, noise will inevitably and c2 = 1, and we could choose J = 2, yielding h1 = −2
limit the circuit depth, implying that there is a trade-off and h2 = 0. It is easy to check that the corresponding
between algorithmic errors (too low p) and gate errors ground state is |10i (i.e., S1 is the solution). Finally, we
(too high p). Note that, in order to find the solution to normalize J and hi such that the Ising Hamiltonian has
the optimization problem, it is not necessary for |~γ ∗ , β~∗i integer eigenvalues, allowing us to restrict γi and βi to
to be equal to the ground state: as long as the ground- the interval [0, π[.
state probability is high enough, the quantum processor For two subsets, four different problems exist, which
can be used to generate a shortlist of potential solutions all yield different sets of hi and J. These are summarized
that can be checked efficiently (in polynomial time) on in Table I. Problem A is the example given above; it is
a classical computer. For instance, even if the success the most interesting, as the other problems are trivial.
probability of measuring the ground state is only 5%, we Problems B and C are trivial since they do not contain
could measure 100 instances and still attain a probability any qubit-qubit interaction (J = 0). Problem D is also
greater than 99% of finding the correct state. Moreover, trivial since both subsets are equal. Additionally, the
the angles ~γ ∗ , β~ ∗ themselves are not interesting, as long ground states are degenerate for problems B and D.
as they yield the lowest-energy state. This gives some We implement Eq. (1) on our quantum processor using
robustness against coherent gate errors, since any over- the circuit in Fig. 1(b). The circuit can be somewhat
 3

compiled by simple identities (e.g., two Hadamard gates π (a) 1.00

(b)
equal identity). We stress that our implementation of
0.75
QAOA is scalable in that we do not use any exponentially

β1
costly pre-compilation (e.g., calculating the final circuit π/2 0.50
unitary and using Cartan decomposition to minimize the

Cost F (γ1 , β1 )
0.25
number of two-qubit gates). 0
Our quantum processor is fabricated using the same pro- π (c)
0.00
(d)
cesses as in Ref. [24] and consists of two fixed-frequency −0.25
transmon qubits with individual control and readout.
−0.50

β1
Both qubits are coupled to a common frequency-tunable π/2

coupler used to mediate a controlled-phase gate (CZ) −0.75

between the qubits. The CZ gate is realized by a full 0 −1.00
coherent oscillation between the |11i and |02i states. The 0 π/2 π 0 π/2 π
interaction is achieved by parametrically modulating the γ1 γ1
resonant frequency of the coupler at a frequency close to
the difference frequency between the |0i − |1i and |1i − |2i ~ for QAOA applied to four
FIG. 2. Cost functions F (~γ , β)
transitions of qubit 1 and 2, respectively [25, 26]. We have instances of the exact-cover problem with p = 1 and n = 2. (a-
benchmarked such a gate on the same device during the d) correspond to problems A-D in Table I. Each experimental
same cooldown to above 99%; however, the benchmark data point is evaluated from the average of 5000 measurements
performed closest in time to the experiments presented on our quantum processor. The dashed lines indicate the
here showed a fidelity of 98.6%. These kinds of fidelity positions of the linecuts in Fig. 3.
fluctuations might be related to fluctuations in the qubits’
coherence times [24]. Single-qubit X rotations are driven
by microwave pulses at the qubit transition frequencies computer without any noise. We see excellent agreement
with fidelities of 99.86% and 99.93% for the respective between measurement and theory: the measured positions
qubits. Z-rotations are implemented in software as a shift of each minimum and maximum are aligned with those
in drive phase and thus have unity fidelity [27]. All the of the theory, consistent with low coherent-error rates.
reported gate fidelities were measured by (interleaved) In addition, we observe excellent agreement between the
randomized benchmarking [28]. More experimental de- absolute values at the minima and maxima, indicating
tails, a measurement setup along with a device schematic, low incoherent-error rates as well. Even with high gate
and benchmarking results are found in the Supplemental fidelities, a high algorithmic fidelity is not guaranteed.
Material [29]. Randomized benchmarking gives the average fidelity over
For p = 1, we apply a simple grid (61 × 61) search of a large number of random gates, which transforms any
β1 , γ1 ∈ [0, π[ while recording 5000 measurements of each coherent errors into incoherent ones. For real quantum
qubit. From these, we calculate hσiz i, hσ1z σ2z i, the cost algorithm circuits, the gates are generally not random.
function F , and the occupation probability for each of Therefore, any coherent errors can quickly add up and
the four possible states, while accounting for the limited, yield algorithmic performance far lower than expected
but calibrated, readout visibility [29]. The grid search from randomized benchmarking fidelities alone [30].
allows us to explore the shape of the optimization land- To quantify the performance of QAOA with p = 1, we
scape, which may bring important understanding in the compare the highest-probability state at the minima of
difficulty of finding global minima for black-box optimiz- F with the solutions in Table I. Problem A [Fig. 3(a)]
ers. In Fig. 2, we show measured cost functions for the does not reach its ground state (F ≈ −0.5); however,
four problems in Table I. Due to the normalization of hi the probability of measuring the correct state (|10i) is
and J, the ground state for each problem corresponds to approximately 50%, which is still better than random
F = −1. In Fig. 2(a), the cost function for problem A guessing. For problem C [Fig. 3(c)], we see that F ≈ −1
never reaches below −0.5. To achieve costs approaching does indeed correspond to a probability close to unity
-1, additional levels (p > 1) are needed. Moreover, the ex- of measuring the ground state (|11i). Problems B and
istence of a local minimum around γ1 ≈ β1 ≈ 3π/4 could D [Fig. 3(b) and (d)] have degenerate ground states, in-
cause difficulties for optimizers trying to find the global dicated by two state probabilities close to 50% each at
minimum. For problems B-D [Fig. 2(b-d)], we see clear F ≈ −1.
minima where F ≈ −1, indicating that we have found the To increase the success probability for problem A, we
optimal variational angles |~γ ∗ , β~ ∗ i corresponding to the add an additional level (p = 2). For p > 1, a grid search
ground state. to map out the full landscape becomes unfeasible due to
In Fig. 3, we take linecuts along the dashed lines in the many parameters (equal to 2p). Therefore, we instead
Fig. 2 and benchmark our measured cost functions and use black-box optimizers to find the optimal variational
state probabilities against those of an ideal quantum angles. We try three different gradient-free optimizers:
 4

F P|00i P|10i P|01i P|11i TABLE II. Comparison between different optimizers. We
run QAOA for problem A over 200 iterations with random
1 (a) (b) starting parameters. We extract the convergence probability
for reaching a cost below -0.95, the average number of function
calls required to reach that level, and the highest achieved
F, P

0
probability of measuring the problem solution (P|10i ).
Optimizer Convergence Function calls P|10i
−1
BGP 61.5% 44 ± 16 96.5%
Nelder-Mead 20.0% 38 ± 13 96.3%
1 CMA-ES 49.5% 121 ± 46 96.6%
F, P

0
number of call is kept low. However, for more optimiza-
−1 (c) (d) tion parameters (higher p), the performance of BGP is
generally decreased due to an increasing need for classical
0 π/2 π 0 π/2 π
computation. CMA-ES, on the other hand, excels when
β1 β1
the number of parameters are high, and thus might be a
good optimizer for QAOA with tens or hundreds of pa-
FIG. 3. A comparison between experiment (open circles) and rameters. Here, with just four parameters, CMA-ES has a
theory (solid lines) for four exact-cover problems using QAOA convergence-probability similar to that of BGP, although
with p = 1. Each color (given at the top) corresponds to with a greater number of function calls on average.
either a state probability or the value of the cost function As Bayesian optimization performs best, we study its
F . The four panels (a-d) correspond to the four problems
(A-D) in Table I. The linecuts are taken at the vertical dashed
individual optimization trajectories in Fig. 4. For each
lines in Fig. 2. The theory curves are calculated assuming run, we plot the cost F and the success probability P|10i ,
an ideal quantum processor, whereas each experimental data along with their average over all converged runs. The first
point is derived from the average of 5000 measurements on function calls are random points, thus explaining why
our quantum processor. there is no improvement in the beginning. After that,
the cost function decreases rapidly, accompanied by an
expected increase in P|10i . We see an indication of a local
Bayesian optimization with Gaussian processes (BGP), minimum at F ≈ −0.5, and that the optimizer sometimes
Nelder-Mead, and covariance matrix adaptation evolution manages to escape from this minimum, showcasing the
strategy (CMA-ES). We choose BGP due to its ability to strength of Bayesian optimization. A comparison between
find global minima, Nelder-Mead due to it being common all three optimizers is found within the Supplementary
and simple, and CMA-ES due to its favorable scaling with Material [29].
the number of optimization parameters. At the end of the optimization, the highest recorded
We evaluate the optimizer performances by running 200 probability of generating the correct state is 96.6%. The
independent optimizations with random starting values success probability is limited by imperfect gates (we have
~ ∈ [0, π[) for each optimizer. We set a threshold
(~γ , β verified that an ideal quantum computer and p = 2 can
for convergence at F < −0.95 and count the number of achieve P|10i = 1). We compare our measured success
converged optimization runs as well as the number of probability to what we would expect from the randomized-
calls to the quantum processor (function calls) required benchmarking fidelities. The quantum circuit for p = 2
to converge. We also record the success probability of consists of 6 X, 4 Hadamard, 4 Z, and 3 CZ gates, which,
measuring the problem solution (P|10i ). The results are when multiplied together with the fidelities for each gate,
summarized in Table II. predicts a total fidelity of 96.3%, in good agreement with
We observe that the success probabilities after con- the measured fidelity considering experimental uncertain-
vergence are similar for all three optimizers. However, ties (e.g., fluctuations in qubit coherence and gate fideli-
there is a difference in convergence probability, of which ties). Note that p = 3 would not yield a higher success
BGP has the highest and Nelder-Mead has the lowest. probability, since adding more gates would lower the total
The lower performance of Nelder-Mead is most likely due fidelity further (predicted to be 94.2%).
to its sensitivity to local minima, a well-known problem In conclusion, we have implemented the quantum ap-
for most optimizers. In contrast, one of the strengths of proximate optimization algorithm with up to p = 2 levels.
Bayesian optimization is its ability to find global min- Using a superconducting quantum processor with state-
ima, which could explain why it performs better than of-the-art performance, we successfully optimized four
Nelder-Mead and CMA-ES. Additionally, Bayesian opti- instances of the exact-cover problem. For the non-trivial
mization is designed to handle optimization where the instance (problem A), we used p = 2 and black-box opti-
time of each function call is high (costly), such that the mization to reach a success probability to 96.6% (up from
 5

1.0 H. Neven, Nature Physics 14, 595 (2018).

F P|10i [7] F. Arute et al., Nature 574, 505 (2019).
[8] E. Farhi, J. Goldstone, and S. Gutmann, arXiv preprint
0.5 arXiv:1411.4028 (2014).
[9] J. Otterbach et al., arXiv preprint arXiv:1712.05771
F , P|10i

(2017).
0.0
[10] G. Pagano et al., arXiv preprint arXiv:1906.02700 (2019).
[11] A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q.
−0.5 Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien,
Nature Communications 5, 4213 (2014).
[12] P. J. O’Malley et al., Physical Review X 6, 031007 (2016).
−1.0 [13] A. Kandala, A. Mezzacapo, K. Temme, M. Takita,
0 20 40 60 80 100
M. Brink, J. M. Chow, and J. M. Gambetta, Nature
Function call 549, 242 (2017).
[14] Z. Jiang, E. G. Rieffel, and Z. Wang, Physical Review A
95, 062317 (2017).
FIG. 4. QAOA for problem A using p = 2 and Bayesian opti-
[15] M. Y. Niu, S. Lu, and I. L. Chuang, arXiv preprint
mization with Gaussian processes. We run the optimization
arXiv:1905.12134 (2019).
200 times with random starting parameters. Plotted as blue
[16] E. Farhi and A. W. Harrow, arXiv preprint
lines are the individual optimization trajectories for the cost
arXiv:1602.07674 (2016).
F . In orange and green are F and the success probability
[17] M. Grönkvist, The Tail Assignment Problem, Ph.D. the-
(P|10i ) averaged over the converged runs.
sis, Chalmers University of Technology and Göteborg
University (2005).
[18] P. Vikstål, M. Grönkvist, M. Svensson, M. Andersson,
50% with p = 1), in good agreement with a prediction G. Johansson, and G. Ferrini, To be published (2019).
from our gate fidelities. Even if many more qubits are [19] E. Farhi, J. Goldstone, S. Gutmann, J. Lapan, A. Lund-
needed to solve problems that are intractable for classical gren, and D. Preda, Science 292, 472 (2001).
computers, algorithmic performance serves as a critical [20] H. Wang and L.-A. Wu, Scientific reports 6, 22307 (2016).
[21] T. Graß, Physical Review Letters 123, 120501 (2019).
quantum-processor benchmark since performance can be
[22] A. Lucas, Frontiers in Physics 2, 5 (2014).
much lower than what individual gate fidelities predicts. [23] V. Choi, arXiv preprint arXiv:1004.2226 (2010).
Although further experiments with larger devices are [24] J. J. Burnett, A. Bengtsson, M. Scigliuzzo, D. Niepce,
needed to explore if QAOA can have an advantage over M. Kudra, P. Delsing, and J. Bylander, npj Quantum
classical algorithms, our results show that QAOA can be Information 5, 9 (2019).
used to solve the exact-cover problem. [25] D. C. McKay, S. Filipp, A. Mezzacapo, E. Magesan, J. M.
We are grateful to the Quantum Device Lab at ETH Chow, and J. M. Gambetta, Physical Review Applied 6,
064007 (2016).
Zürich for sharing their designs of sample holder and
[26] S. Caldwell et al., Physical Review Applied 10, 034050
printed circuit board, and Simon Gustavsson and Bruno (2018).
Küng for valuable support with the measurement infras- [27] D. C. McKay, C. J. Wood, S. Sheldon, J. M. Chow, and
tructure. We also thank Morten Kjaergaard, Devdatt J. M. Gambetta, Physical Review A 96, 022330 (2017).
Dubhashi, and Kevin Pack for insightful discussions. This [28] A. D. Córcoles, J. M. Gambetta, J. M. Chow, J. A. Smolin,
work was performed in part at Myfab Chalmers. We ac- M. Ware, J. Strand, B. L. Plourde, and M. Steffen,
knowledge financial support from the Knut and Alice Physical Review A 87, 030301 (2013).
[29] See Supplemental Material.
Wallenberg Foundation, the Swedish Research Council,
[30] K. Michielsen, M. Nocon, D. Willsch, F. Jin, T. Lippert,
and the EU Flagship on Quantum Technology H2020- and H. De Raedt, Computer Physics Communications
FETFLAG-2018-03 project 820363 OpenSuperQ. 220, 44 (2017).

∗
bylander@chalmers.se
[1] A. Montanaro, npj Quantum Information 2, 15023 (2016).
[2] G. Wendin, Reports on Progress in Physics 80, 106001
(2017).
[3] X. Gu, A. F. Kockum, A. Miranowicz, Y. X. Liu, and
F. Nori, Physics Reports 718-719, 1 (2017).
[4] M. Kjaergaard, M. E. Schwartz, J. Braumüller, P. Krantz,
J. I.-J. Wang, S. Gustavsson, and W. D. Oliver, arXiv
preprint arXiv:1905.13641 (2019).
[5] J. Preskill, Quantum 2, 79 (2018).
[6] S. Boixo, S. V. Isakov, V. N. Smelyanskiy, R. Babbush,
N. Ding, Z. Jiang, M. J. Bremner, J. M. Martinis, and
 6

Supplementary Material (a)

300 K
Measurement setup 3K

The experimental measurement setup used here is a 20 dB 20 dB 20 dB 20 dB

standard circuit-QED setup, see schematic in Fig. 5. The 800 mK
quantum processor consists of two xmon-style transmon
qubits coupled via a frequency-tunable anharmonic os- 10 dB 10 dB 10 dB 10 dB
cillator. The tunability is provided by two Josephson
junctions in a SQUID configuration. The two qubits 100 mK
are capacitively coupled to individual control lines and
quarter-wavelength resonators for readout. There is also 0 dB 10 dB 0 dB 0 dB
a readout resonator for the coupler, which is only used
as a debugging tool (i.e., it is not involved during any 10 dB 20 dB 10 dB 0 dB
algorithm execution). The SQUID for the tunable coupler 7 mK
is inductively coupled to a waveguide to allow for both
8 GHz 8 GHz 540
static and fast modulation of the resonant frequency. 20 dB 20 dB
MHz
The processor is fabricated on a high-resistivity intrin-
8 GHz 20 dB 8 GHz
sic silicon substrate. After initial chemical cleaning, an
aluminium film is evaporated. All features except the
Josephson junctions are patterned by direct-write laser Readout resonators
lithography and etched with a warm mixture of acids.
The Josephson junctions are patterned by electron-beam
lithography, and evaporated from the same target as pre-
viously. A third lithography and evaporation step (with
in-situ ion milling) is performed to connect the Josephson
junctions to the rest of the circuit. Finally, the wafer is
diced into individual dies and subsequently cleaned by a Qubit 1 Qubit 2
combination of wet and dry chemistry. (b) Coupler
Then, a die is selected and packaged in a copper box
and wire bonded to a palladium- and gold-plated printed
circuit board with 16 non-magnetic coaxial connectors.
For the present device, we use 5 of these connectors, 2
for readout, 2 for single-qubit control, and 1 for control
of the magnetic flux through the SQUID loop of the
coupler. These are connected to filtered and attenuated
coaxial lines leading up to room temperature. We point
out that the DC current for the static flux bias is also
provided through the coaxial line. Finally, the processor is
attached to the mixing chamber of a Bluefors LD250 cryo-
free dilution refrigerator. There, it is shielded from stray
magnetic fields by two shields of cryoperm/mu-metal and
two shields of superconductors.
We perform multiplexed readout by using a Zurich In-
struments UHFQA for generating and detecting the read-
out signals, together with a Rohde & Schwarz SGS100A
continuous-wave signal generator and two Marki IQ mix- FIG. 5. (a) Cryogenic setup and electrical circuit of the
ers for up- and down-conversion. The single-qubit pulses quantum processor. All lines are attenuated and filtered to
are synthesized by a Zurich Instruments HDAWG and minimize the amount of noise reaching the qubits. The readout
output contains cryogenic isolators and a high electron mobility
upconverted using Rohde & Schwarz SGS100A vector
transistor amplifier. (b) False-colored micro-graph of the
signal generators. The flux drive is generated directly by processor. The colors match the circuit elements in (a). The
the HDAWG since the modulation frequency is within the three waveguides at the bottom are for control over the qubits
bandwidth of the instrument. Finally, all instruments are and the coupler.
controlled and orchestrated by the measurement and au-
tomation software Labber. Labber also does cost-function
 7

TABLE III. Device parameters. Readout-resonator frequency

fR and qubit transition frequencies fij . g is the coupling 1.0
between qubit and resonator, and j is the coupling between
qubit and coupler. T1 and T2∗ are the relaxation and free
0.8
induction decay times measured over 14 hours. F1q , and FCZ
are the single-qubit and CZ fidelities, respectively.

p0
0.6
Parameter Qubit 1 Qubit 2 Qubit 1
fR 6.17 GHz 6.04 GHz Qubit 2
f01 3.82 GHz 4.30 GHz 0.4 Reference
f12 − f01 -229 MHz -225 MHz Interleaved
j 29.1 MHz 33.0 MHz 0.2
g 53.2 MHz 56.9 MHz 100 101 102 103
T1 77 µs 55 µs
Number of cliffords
T2∗ 49 µs 82 µs
F1q 0.9986 0.9993
FCZ 0.986 FIG. 6. Randomized benchmarking of single- and two-qubit
gates. Plotted are the probability of measuring the ground
state as a function of number of Clifford gates applied. Circles
evaluations and calls external Python packages for the are data, and lines are fits to extract the gate fidelity. For
three different optimizers. All three optimizers were run qubit 1 and 2, the extracted single-qubit fidelities (averaged
over all possible single-qubit Cliffords) are 0.9986 and 0.9993.
using publicly available packages: Scikit-Optimize for
For benchmarking of the two-qubit gate, we take a reference
BGP, scipy for Nelder-Mead, and pycma for CMA-ES. (random Clifford gates) and an interleaved (a CZ gate between
each Clifford) trace to extract the CZ fidelity (0.986).

Characterization and tune-up

is 2.3 µs long, well below our relaxation times (several
Initially, we perform basic spectroscopy and decoher- tens of microseconds). After finding the optimal readout
ence benchmarking of each qubit individually. This allows parameters, a voltage threshold is used to differentiate
us to extract readout frequencies, qubit frequencies and between |0i or |1i of the measured qubit.
anharmonicities, relaxation and dephasing times, and To accurately extract state probabilities in the pres-
static couplings between qubit and resonator, as well as ence of limited readout fidelity, we collect statistics of
between qubit and coupler. The extracted parameters are the measured qubit population as a function of qubit
found in Table III. drive amplitude (Rabi oscillations). Since the measured
After the initial characterization, we tune up high- population increases monotonically with the expected
fidelity single-qubit gates. The drive pulses have cosine population, we can renormalize the populations. This
envelopes together with first-order DRAG components calibration allows us to accurately measure the average
to compensate for the qubit frequency shift due to the quantities hσiz i, hσi1 σi2 i and state probabilities even in the
driving. Our rather long (50 ns) pulses makes leakage presence of limited readout fidelities.
from |1i to |2i minimal even without DRAG. To find Our two-qubit gate of choice is the controlled phase
optimal pulse amplitudes and DRAG coefficients, we use (CZ). This interaction is induced by parametrically modu-
error amplification by applying varying lengths of trains of lating the resonant frequency of the coupler at a frequency
π pulses. Qubit drive frequencies are measured accurately close to the difference of |11i and |02i. For our device,
by detuned Ramsey fringes. this frequency is 255 MHz. However, due to the frequency
Next, we calibrate our readout fidelities. By collecting modulation and the non-linear relationship between flux
raw voltages of the readout signals (as measured by the and frequency, the transition frequencies are slightly low-
digitizer in the UHFQA), with and without a calibrated pi- ered. This frequency shift will also induce deterministic
pulse applied to the qubit (|0i and |1i states, respectively), phase shifts on the individual qubits, which we compen-
and as a function of readout frequency and amplitude, sate for by applying Z gates on both qubits after each CZ
we can find the optimal readout parameters. Due to gate. We choose a static bias point and a modulation am-
our rather low coupling strengths, we cannot achieve plitude that yield a moderate effective coupling strength
short readout times in this device. However, QAOA of 5 MHz between the two states. From here, we find the
does not require any measurement feedback, so a long modulation frequency and time that yield a full oscillation
readout time is not an issue as long as the time is shorter between the |11i and |02i states. We then fine-tune the
than the relaxation times of the qubits. Also, longer frequency and time such that the controlled phase is π
readout times give greater signal-to-noise ratios, which and the leakage to |02i is minimal. Here, the final gate
allows us to achieve high readout fidelities even in the frequency and duration were 253 MHz and 271 ns.
absence of a quantum-limited amplifier. Here, the readout We benchmark our single and two-qubit fidelities using
 8

randomized benchmarking. A sequence of random gates

drawn from the Clifford group is applied together with
a final recovery gate which should take the system back
to the ground state. The number of random gates is
varied and the probability of measuring the ground state
is recorded. In Fig. 6, we plot these probabilities for
each qubit individually and for the two-qubit case. In
the single-qubit case, it is important to note that it was
done simultaneously for both qubits. Generally, the gate
fidelities are higher if they are done in isolation. However,
to reduce the total run time of algorithms, we usually run
single-qubit gates in parallel. Therefore, simultaneous
randomized benchmarking fidelities are more relevant
metrics than isolated ones.
Optimizer comparison

To quantify the optimization further, we study the

trajectories of each optimization run (Fig. 7). The tra-
jectories for BGP and Nelder-Mead [Fig. 7(a-b)] corrobo-
rate the indications about local minima. We see groups
of horizontal lines corresponding to different local min-
ima, especially clear at F ≈ −0.55 for both BGP and
Nelder-Mead. We also see that BGP tries, and sometimes
succeeds, to escape these local minima, which is one of
the advantages of Bayesian optimization. In comparison,
FIG. 7. Optimization of variational angles for problem A using
Nelder-Mead rarely gets out of a local minimum once it
p = 2 iterations of the algorithm and three different black-
box and gradient-free optimizers: (a) Bayesian optimization has found it. For the third optimizer, CMA-ES [Fig. 7(c)],
with Gaussian processes, (b) Nelder-Mead, and (c) covariance it is hard to draw any conclusions from the trajectories
matrix adaptation evolution strategy. We run the optimization other than that the convergence is slower than for the
200 times with random starting parameters. Plotted as blue other optimizers. However, we include the CMA-ES tra-
lines are the individual optimization trajectories for F . In jectories for completeness. For each optimizer, we also
orange and green are the cost (F ) and success probability plot the averaged trajectories for F and the probability
(P|10i ) averaged over the converged runs.
of finding the solution state P|10i .

View publication stats