Experimental Quantum Learning of a Spectral Decomposition

Michael R. Geller,1, ∗ Zoë Holmes,2 Patrick J. Coles,3, 4 and Andrew Sornborger2, 4, †

1
Center for Simulational Physics, University of Georgia, Athens, Georgia 30602, USA
2
Information Sciences, Los Alamos National Laboratory, Los Alamos, NM, USA.
3
Theoretical Division, Los Alamos National Laboratory, Los Alamos, NM, USA.
4
Quantum Science Center, Oak Ridge, TN, 37931, USA.
Currently available quantum hardware allows for small scale implementations of quantum machine
learning algorithms. Such experiments aid the search for applications of quantum computers by
benchmarking the near-term feasibility of candidate algorithms. Here we demonstrate the quantum
arXiv:2104.03295v1 [quant-ph] 7 Apr 2021

learning of a two-qubit unitary by a sequence of three parameterized quantum circuits containing

a total of 21 variational parameters. Moreover, we variationally diagonalize the unitary to learn
its spectral decomposition, i.e., its eigenvalues and eigenvectors. We illustrate how this can be
used as a subroutine to compress the depth of dynamical quantum simulations. One can view our
implementation as a demonstration of entanglement-enhanced machine learning, as only a single
(entangled) training data pair is required to learn a 4 × 4 unitary matrix.

Quantum simulation and machine learning are among lap with an input state that might itself be the output
the most promising applications of large-scale quantum of another algorithm [38–41]. Although a minimum of
computers. Of these, the discovery of algorithms with 2(2n − 1) real parameters are required to do this exactly
provable exponential speedup has been more challeng- in general, it is widely believed that for many cases of
ing in the machine learning domain, in part because it is interest a polynomial number of parameters will suffice.
harder to port established machine learning techniques to Beyond learning states, it is also useful to variation-
the quantum setting [1]. Notable exceptions include lin- ally learn a unitary channel ρ 7→ U ρU † . This is more
ear system solvers [2–4], kernel methods [5, 6], and Boltz- challenging because now all of the 4n − 1 matrix ele-
mann machines [7, 8]. But quantum simulation demon- ments must match for an exact replica V of an arbitrary
strations [9–13] appear to be ahead of machine learning U , up to a global phase. A hybrid protocol for learning
[14–20] in terms of the maximum problem sizes achieved, a unitary is provided by the quantum-assisted quantum
suggesting that quantum simulation might yield the ear- compiling algorithm [38]. This is a low-depth subroutine
liest applications with quantum advantage. appropriate for both near-term and fault-tolerant quan-
Variational quantum algorithms [21–23] will likely fa- tum computing.
cilitate near-term implementations for these applications. Given a target unitary U and parameterized unitary
Such algorithms employ a problem-specific cost function V (θ), both acting on n-qubits, quantum-assisted quan-
that is evaluated on a quantum computer, while a classi- tum compiling uses a maximally entangled Bell state on
cal optimizer trains a parameterized quantum circuit to 2n qubits to compute the Hilbert-Schmidt inner product,
minimize this cost. |Tr(U V (θ)† )|. Since this inner product is directly related
Some variational quantum algorithms have interest be- to the average fidelity between states acted on by U and
yond their ability to achieve quantum advantage on their V (θ) [42, 43], this allows the action of U on all possible
own, and can serve as subroutines for larger quantum al- input states to be studied using a single entangled input
gorithms. These include quantum autoencoders for data state. Consequently, with this entanglement-enhanced
compression [24, 25] and linear algebra methods [26– learning strategy, only a single training state is needed
28]. A common subroutine is the training of a varia- to fully learn U , in contrast to the ∼2n input-output pairs
tional quantum state to approximate the ground state of that are required in the absence of entanglement [44, 45].
a given n-qubit Hamiltonian, which can be the Hamilto- Although more challenging than state learning, learn-
nian of a simulated model [29] or some other optimiza- ing a unitary can be used for a wide variety of quantum
tion objective [30–32]. Variational quantum algorithms information applications, including circuit depth com-
to learn and diagonalize density matrices have also been pression, noise-tailoring, benchmarking, and the ‘black
developed [33–35], which is a fundamental subroutine box uploading’ of an unknown experimental system uni-
that will have many uses including principal component tary [38]. Quantum-assisted quantum compiling has been
analysis and estimation of quantum information quanti- demonstrated on 1+1 qubits, where a single-qubit vari-
ties [36, 37]. Another subroutine is the learning of one ational circuit V (θ) learned the value of a single-qubit
quantum state by a second state, where the output of unitary U [38].
a variational circuit is optimized to maximize the over- Quantum-assisted quantum compiling can be gener-
 2

alized to learn not only a unitary, but also its Schur METHODS
decomposition W (θ)D(γ)W (θ)† , where W (θ) is a pa-
rameterized unitary and D(γ) is a parameterized diag- Learning task. We demonstrate the variational learn-
onal unitary. That is, one can use quantum-assisted ing of the spectral decomposition of a unitary by learning
quantum compiling to variationally diagonalize a uni- a diagonalization of the short-time evolution operator of
tary. This is useful for a variety of quantum information the 2-spin Ising model
science applications, since access to the spectral decom-
position of a unitary U enables arbitrary powers of U
X X
H=J Zi Zi+1 + B Xi . (1)
to be implemented using a fixed depth circuit. Specif- i=1,2 i=1,2
ically, suppose we learn the optimum parameters θopt
and γopt such that U = W (θopt )D(γopt )W (θopt )† . We Here J quantifies the exhange energy, B is the trans-
can then implement U k using the fixed depth circuit verse field strength and Zi and Xi are Pauli operators
W (θopt )D(kγopt )W (θopt )† . We stress that the parameter on the ith qubit. We approximate the short-time evolu-
k here can take any real value and hence this approach tion exp(−iH∆t) of the spin chain using a second order
can be used to implement non-integer and negative pow- Trotter approximation, that is we take
ers of U .
Y Y Y 2
One important application of variational unitary diag- U= Rx (θB )i × Rzz (θJ )i,i+1 × Rx (θB )i (2)
onalization is quantum simulation. Let U be a Trotter- i i i
ized (or other) approximation to a short-time evolution
e−iH∆t by some Hamiltonian H. We assume that H is where θB = B∆t/2 and θJ = 2J∆t/2. The simulated
local [46–49], sparse [50–52], or given by a linear combina- Ising model parameters are listed in Table I. The specific
tion of unitaries [53–55], permitting efficient simulation circuit we used for U is shown in Fig. 1.
with bounded error. Then W Dt/∆t W † is an approximate
fast-forwarded evolution operator with a circuit depth TABLE I. Ising model parameters.
that is independent of t. By contrast, most of the best Parameter Value
known Hamiltonian simulation algorithms [49, 52, 55] N number spins 2
have depths scaling at least linearly in t, inhibiting long ∆t short evolution time 0.2
time simulations on near-term hardware. J exchange energy 1
B transverse field 1
This low-depth algorithm, called variational fast-
forwarding [56], lies at an exciting intersection of machine
learning and quantum simulation and is an promising Ansatz. To learn the spectral decomposition of U we
approach in the burgeoning field of variational quantum variationally compile it to a unitary with a structure of
simulation [57–68]. Variational fast-forwarding has been the form
demonstrated on 1+1 qubits [56]. Refinements of varia-
tional fast-forwarding for simulating a given fixed initial V (θ, γ) = W (θ)D(γ)W (θ)† , (3)
state [67] and for learning the spectral decomposition of a
given Hamiltonian [68] have also been proposed. It is im- where W is an arbitrary unitary, D is a diagonal matrix
portant to note that the unitary being diagonalized need and θ and γ are vectors of parameters. After successful
not already be known. Therefore, variational unitary di- training, D will capture the eigenvalues of U , and W will
agonalization could also be used to perform a ‘black box capture the rotation matrix from the computational basis
diagonalization’ of the dynamics of an unknown experi- to the eigenbasis of U .
mental system. Thus, this approach provides a new al- The parameterized circuits we used as ansätze for
gorithmic tool for probing dynamics in an experimental the diagonal unitary D and basis transformation W are
setting. shown in Fig. 3. A general diagonal unitary D ∈ SU(2n )
on n qubits contains 2n − 1 variational parameters. In
In this work we use ibmq bogota to demonstrate the our experiment we implement a two-qubit version of this
successful learning of a spectral decomposition on 2+2 exact D containing 3 variables. In general an arbitrary
qubits. Specifically, we diagonalize the short time evolu- unitary W can be constructed from any general n-qubit
tion unitary for an Ising spin chain. After only 16 steps parameterized quantum circuit with poly(n) variables.
of training by gradient descent, the spectral decompo- The expressive power of different options has been in-
sition is used to fast-forward the evolution of the Ising vestigated in [69, 70]. The two-qubit circuit we use to
model, resulting in a dramatic reduction of simulation implement the arbitrary unitary W consists of 3 layers
error compared with Trotterized Hamiltonian simulation of X, Y rotations and phase gates on each qubit, sepa-
and a significant (∼ 10×) increase in the effective quan- rated by 2 CNOT gates.
tum volume of the simulation.
 3

FIG. 1. Circuit for the short-time evolution of the Ising model. The boxes indicate individual Trotter steps. Here
Rx (θ) = eiθX/2 is a single qubit Pauli X rotation, P (γ) = diag(1, eiγ ) is a single qubit phase gate and Rzz (θ) = eiθZ⊗Z/2 is a
two-qubit diagonal unitary.

Eq. (4) involves only local measures and hence mitigates

trainability issues associated with barren plateaus [71–
82]. For more details on the cost see Ref. [38].
Training. The parameters (θ1 , · · · , θ18 , γ1 , · · · , γ3 )
are initialized to minimize the cost (4) with the exchange
energy J in (1) set to zero, which can be simulated effi-
ciently and solved exactly.
The circuits are trained to minimise the cost using gra-
dient descent. At each step of the training, we measure
∂C ∂C
the cost and gradients ∂θ k
and ∂γ l
at a particular point
(θ, γ) in the parameter space, and use this to update
(θ, γ) according to
∂C
FIG. 2. Cost function circuit. Circuits to learn the spec- θk ← θk − η (5)
∂θk
tral decomposition W DW † of a two-qubit unitary U on a
four-qubit chain. Each cost function and gradient evaluation ∂C
γl ← γl − η , (6)
requires the estimation of Pr(00)1,2 and Pr(00)3,4 , obtained ∂γl
by measuring the qubits in the dashed boxes of (a) and (b).
∂C ∂C
where η is the learning rate. The gradients ∂θ k
and ∂γ l
are measured using the parameter shift rule derived in
Cost function. To compile the target unitary into a Ref. [56]. In each optimization step the 21 components
diagonal form we use the local Hilbert-Schmidt test cost of the gradient are measured, requiring the measurement
function defined in [38]. For learning a 4 × 4 unitary of 156 distinct circuits. For each cost evaluation we took
matrix, this cost can be written as 8000 measurement shots.
The learning rate was decreased as the optimization
1 progressed according to a schedule
C =1− (Pr(00)1,2 + Pr(00)3,4 ) (4)
2
η0
where Pr(00)1,2 and Pr(00)3,4 are the probabilities of ob- η(j) = , (7)
[1 + (j/δ)]κ
serving the outcome 00 on qubits (1,2) and (3,4) on run-
ning the circuits shown in Fig. 2(a) and Fig. 2(b) respec- where j ∈ {0, 1, 2, · · · , 16} is the optimization step num-
tively. ber. The hyperparameters η0 = 1.1, κ = 0.5, and δ = 12
The probabilities Pr(00)1,2 and Pr(00)3,4 are measures were optimized by classical simulation. Additional de-
of the entanglement fidelity of the unitary channel U V † tails about the training are provided in the Supplemen-
with V = W DW † . As a result, this cost is faithful, van- tary Information.
ishing if and only if the diagonalization W DW † matches
the target unitary U (up to a global phase). Furthermore,
the cost is operationally meaningful for non-zero values RESULTS
in virtue of upper bounding the average gate fidelity be-
tween U and V . Hence a small value of C guarantees To assess the quality of the optimization, the parame-
that the diagonalization W DW † is an accurate approxi- ters found at each step of the training were used to eval-
mation of the target unitary U . We note that this cost, uate C both on the quantum computer and classically.
 4

FIG. 3. Ansatz circuits. Implemented circuits for the diagonal unitary D and basis transformation W . Here Rzz (γ) =
eiγZ⊗Z/2 is a two-qubit diagonal unitary and P (γ) = diag(1, eiγ ) is a single qubit phase gate. D contains 3 variational
parameters. W has 3 layers and 18 variational parameters.

The results are shown in Fig. 4. The classical cost, which

2
we call the noise-free or ideal cost, reflects the true accu- 0.5
Total error
Eigenvalue error
racy of the optimized circuits. We successfully reduced

Error
the raw cost to 0.1099 corresponding to an ideal cost 1
0.4
of 0.013. The raw cost from the quantum computer is
higher than the ideal cost because of gate errors. 0

Cost
0.3 0 4 8 12 16
The inset of Fig. 4 confirms that the errors in D Iteration
and W DW † are both iteratively reduced as the cost 0.2
is trained. Here the Frobenius distance between U
and V is plotted, minimized over the arbitrary global 0.1
Raw
phase eiϕ . The ideal and learnt diagonals are also Ideal
compared, accounting for the global phase eiϕ and for 0.0
a possible reordering, specified by a permutation ma- 0 4 8 12 16
Iteration
trix χ. Specifically, we plot the Frobenius distance
minϕ,χ ||Dexact − eiϕ χDχ† || where Dexact is a diagonal
matrix with {λexact }, the ordered exact spectrum of U , FIG. 4. Training curves. In the main plot we show the
i
cost as a function of the iteration step during training. The
along the diagonal. P This is equivalent to the sum of the
blue curve is the measured cost function C and the green
eigenvalue errors i |λexact
i −λ i e iϕopt 2
| , where {λi } is the curve is its noise-free ‘ideal’ value. In the inset, the yellow
ordered learnt spectrum and ϕopt accounts for the global curve indicates the total error in the learnt unitary defined
phase shift. as Frobenius distances between U and V minimized over a
It is also interesting to monitor the training by using global phase eiϕ , i.e., the quantity minϕ,χ ||U −eiϕ V ||. The red
the measured gradient g meas at each step to calculate curve indicates the eigenvalue error, defined as the Frobenius
distance between the learnt and exact diagonal also minimized
the angle between g meas and the exact gradient simu-
over a permutation χ, i.e., minϕ,χ ||Dexact − eiϕ χDχ† ||.
lated classically. This is plotted in the Supplementary
Information. The data confirms that the optimization is
correctly moving downhill in the cost function landscape. |+i⊗2 , but with D0 s parameters (γ1 , γ2 , γ3 ) changed to
Having learned the spectral decomposition, we use the t
result to fast-forward the Hamiltonian simulation of the (γ1 , γ2 , γ3 ) × . (8)
∆t
Ising model (1) for 2 spins on ibmq bogota. In this ex-
periment we prepare an initial state |+i⊗2 and propagate The state fidelity with perfect evolution e−iHt |+i⊗2 is
it forward in time by both Trotterization and variational also measured in this case. The results for this exper-
fast-forwarding (with backend circuit optimization dis- imental fast-forwarding and experimental Trotter sim-
abled in both cases). The Trotterized evolution at times ulation are indicated by the green and blue solid lines
t = K∆t, is obtained by applying K = 0, 8, 16, · · · , 96 respectively in Fig. 5.
consecutive Trotter steps from Fig. 1. After each step We compare the experimental fast forwarding to the
we experimentally measure the fidelity of the Trotter- ideal classical fast-forwarding by also measuring the
ized evolution with the perfect evolution e−iHt |+i⊗2 , noise-free fidelities obtained by classical simulation. In
which contains no Trotter error. The variational fast- this ideal simulation, the initial state |+i⊗2 is prepared
forwarding evolution at time t is obtained by applying perfectly and the Trotterized evolution includes Trotter
the optimized variational circuits for W † , D∗ , and W to errors but no gate errors. The measurement is also as-
 5

1.0 program under project number 20190065DR. Addition-

ally, PJC and AS acknowledge that (subsequent to the
above acknowledged funding) this material is based upon
0.8
work supported by the U.S. Department of Energy, Of-
Fidelity

fice of Science, National Quantum Information Science

0.6 Trotter Ideal Research Centers. ZH acknowledges support from the
VFF Raw LANL ASC Beyond Moore’s Law project. We would also
0.4
like to thank Andrew Arrasmith, Lukasz Cincio, and Joe
Gibbs for useful discussions.
0 2 4 6 8 10 12 14 16 18 20
Time

FIG. 5. Demonstration of variational fast-forwarding. ∗

mgeller@uga.edu
The plot shows the measured state fidelity versus evolution †
sornborg@lanl.gov
time. The data (circles) are obtained at multiples of 8 ∆t, out [1] Jacob Biamonte, Peter Wittek, Nicola Pancotti, Patrick
to t = 19.2. Rebentrost, Nathan Wiebe, and Seth Lloyd, “Quantum
machine learning,” Nature 549, 195–202 (2017).
[2] Aram W Harrow, Avinatan Hassidim, and Seth Lloyd,
sumed to be ideal. The fidelities of these ideal simulations “Quantum algorithm for linear systems of equations,”
are indicated by the dashed lines in Fig. 5. Physical review letters 103, 150502 (2009).
The ideal Trotterized evolution is nearly perfect in [3] B David Clader, Bryan C Jacobs, and Chad R
Fig. 5 due to the small value of ∆t and use of a second- Sprouse, “Preconditioned quantum linear system algo-
order Trotter step. The ideal variational fast-forwarding rithm,” Physical review letters 110, 250504 (2013).
[4] Andrew M Childs, Robin Kothari, and Rolando D
evolution is less accurate, due to the imperfect learning of Somma, “Quantum algorithm for systems of linear equa-
U ’s spectral decomposition. However, the real data taken tions with exponentially improved dependence on pre-
on ibmq bogota exhibits the opposite behavior. Whereas cision,” SIAM Journal on Computing 46, 1920–1950
the variational fast-forwarding evolution is only slightly (2017).
degraded by gate errors—because the circuit depth is in- [5] Maria Schuld and Nathan Killoran, “Quantum machine
dependent of t—the Trotterized evolution quickly loses fi- learning in feature hilbert spaces,” Physical review letters
delity. Namely, while the fidelity of the variationally fast 122, 040504 (2019).
[6] Vojtěch Havlı́ček, Antonio D Córcoles, Kristan Temme,
forwarded simulation remains above 0.7 for all 20 times Aram W Harrow, Abhinav Kandala, Jerry M Chow, and
steps, the fidelity of the Trotterized simulation is less Jay M Gambetta, “Supervised learning with quantum-
than 0.4 after only 6 times steps. Thus Figure 5 demon- enhanced feature spaces,” Nature 567, 209–212 (2019).
strates a striking example of fast-forwarding as well as [7] Nathan Wiebe, Ashish Kapoor, and Krysta M
provides further evidence that the spectral decomposi- Svore, “Quantum deep learning,” arXiv preprint
tion of U has been well learnt. arXiv:1412.3489 (2014).
[8] Guillaume Verdon, Michael Broughton, and Ja-
In conclusion, we have experimentally demonstrated cob Biamonte, “A quantum algorithm to train neu-
the entanglement-enhanced quantum learning of a 4 × 4 ral networks using low-depth circuits,” arXiv preprint
unitary using a diagonal ansatz containing a total of 6 arXiv:1712.05304 (2017).
CNOTs and 21 variational parameters. A single input- [9] Abhinav Kandala, Antonio Mezzacapo, Kristan Temme,
Maika Takita, Markus Brink, Jerry M Chow, and
output pair was used for training compared to the 22 = 4
Jay M Gambetta, “Hardware-efficient variational quan-
input-output pairs necessary for training without entan- tum eigensolver for small molecules and quantum mag-
gling ancillas. We both learn the unitary and its spectral nets,” Nature 549, 242–246 (2017).
decomposition, enabling the fast-forwarding of an Ising [10] Abhinav Kandala, Kristan Temme, Antonio D Córcoles,
model. This four-qubit experiment took 8000 shots for Antonio Mezzacapo, Jerry M Chow, and Jay M Gam-
each of the 156 independent circuit evaluations, at each betta, “Error mitigation extends the computational reach
of the 16 steps of the optimization algorithm, a total of a noisy quantum processor,” Nature 567, 491–495
(2019).
of 20 × 107 circuit evaluations. Thus, this experiment
[11] Google AI Quantum and Collaborators, Frank Arute,
is among the most complex quantum machine learning Kunal Arya, Ryan Babbush, Dave Bacon, Joseph C.
demonstrations to date [14–20] and constitutes an im- Bardin, Rami Barends, Sergio Boixo, Michael Broughton,
portant primitive in experimental quantum information Bob B. Buckley, David A. Buell, Brian Burkett, Nicholas
science. Bushnell, Yu Chen, Zijun Chen, Benjamin Chiaro,
Roberto Collins, William Courtney, Sean Demura, An-
Acknowledgments. PJC and AS acknowledge initial drew Dunsworth, Edward Farhi, Austin Fowler, Brooks
Foxen, Craig Gidney, Marissa Giustina, Rob Graff, Steve
support and MRG acknowledges support from LANL’s
Habegger, Matthew P. Harrigan, Alan Ho, Sabrina Hong,
Laboratory Directed Research and Development (LDRD)
 6

Trent Huang, William J. Huggins, Lev Ioffe, Sergei V. Clean, Kosuke Mitarai, Xiao Yuan, Lukasz Cincio, and
Isakov, Evan Jeffrey, Zhang Jiang, Cody Jones, Dvir Patrick J. Coles, “Variational quantum algorithms,”
Kafri, Kostyantyn Kechedzhi, Julian Kelly, Seon Kim, arXiv preprint arXiv:2012.09265 (2020).
Paul V. Klimov, Alexander Korotkov, Fedor Kostritsa, [22] Suguru Endo, Zhenyu Cai, Simon C Benjamin, and Xiao
David Landhuis, Pavel Laptev, Mike Lindmark, Erik Yuan, “Hybrid quantum-classical algorithms and quan-
Lucero, Orion Martin, John M. Martinis, Jarrod R. tum error mitigation,” Journal of the Physical Society of
McClean, Matt McEwen, Anthony Megrant, Xiao Mi, Japan 90, 032001 (2021).
Masoud Mohseni, Wojciech Mruczkiewicz, Josh Mutus, [23] Kishor Bharti, Alba Cervera-Lierta, Thi Ha Kyaw,
Ofer Naaman, Matthew Neeley, Charles Neill, Hartmut Tobias Haug, Sumner Alperin-Lea, Abhinav Anand,
Neven, Murphy Yuezhen Niu, Thomas E. O’Brien, Eric Matthias Degroote, Hermanni Heimonen, Jakob S.
Ostby, Andre Petukhov, Harald Putterman, Chris Quin- Kottmann, Tim Menke, Wai-Keong Mok, Sukin Sim,
tana, Pedram Roushan, Nicholas C. Rubin, Daniel Sank, Leong-Chuan Kwek, and Alán Aspuru-Guzik, “Noisy
Kevin J. Satzinger, Vadim Smelyanskiy, Doug Strain, intermediate-scale quantum (nisq) algorithms,” arXiv
Kevin J. Sung, Marco Szalay, Tyler Y. Takeshita, Amit preprint arXiv:2101.08448 (2021).
Vainsencher, Theodore White, Nathan Wiebe, Z. Jamie [24] Jonathan Romero, Jonathan P Olson, and Alan Aspuru-
Yao, Ping Yeh, and Adam Zalcman, “Hartree-fock on a Guzik, “Quantum autoencoders for efficient compression
superconducting qubit quantum computer,” 369, 1084– of quantum data,” Quantum Science and Technology 2,
1089 (2020). 045001 (2017).
[12] Frank Arute, Kunal Arya, Ryan Babbush, Dave Ba- [25] Dmytro Bondarenko and Polina Feldmann, “Quantum
con, Joseph C Bardin, Rami Barends, Andreas Bengts- autoencoders to denoise quantum data,” Physical Review
son, Sergio Boixo, Michael Broughton, Bob B Buckley, Letters 124, 130502 (2020).
et al., “Observation of separated dynamics of charge [26] Carlos Bravo-Prieto, Ryan LaRose, Marco Cerezo, Yigit
and spin in the fermi-hubbard model,” arXiv preprint Subasi, Lukasz Cincio, and Patrick J Coles, “Variational
arXiv:2010.07965 (2020). quantum linear solver,” arXiv preprint arXiv:1909.05820
[13] Igor Aleiner, Frank Arute, Kunal Arya, Juan Atalaya, (2019).
Ryan Babbush, Joseph C Bardin, Rami Barends, An- [27] Xiaosi Xu, Jinzhao Sun, Suguru Endo, Ying Li, Simon C
dreas Bengtsson, Sergio Boixo, Alexandre Bourassa, Benjamin, and Xiao Yuan, “Variational algorithms for
et al., “Accurately computing electronic properties linear algebra,” arXiv preprint arXiv:1909.03898 (2019).
of materials using eigenenergies,” arXiv preprint [28] Hsin-Yuan Huang, Kishor Bharti, and Patrick Reben-
arXiv:2012.00921 (2020). trost, “Near-term quantum algorithms for linear systems
[14] Diego Riste, Marcus P da Silva, Colm A Ryan, An- of equations,” arXiv preprint arXiv:1909.07344 (2019).
drew W Cross, Antonio D Córcoles, John A Smolin, [29] Alberto Peruzzo, Jarrod McClean, Peter Shadbolt, Man-
Jay M Gambetta, Jerry M Chow, and Blake R John- Hong Yung, Xiao-Qi Zhou, Peter J Love, Alán Aspuru-
son, “Demonstration of quantum advantage in machine Guzik, and Jeremy L O’brien, “A variational eigenvalue
learning,” npj Quantum Information 3, 1–5 (2017). solver on a photonic quantum processor,” Nature com-
[15] Maria Schuld, Mark Fingerhuth, and Francesco Petruc- munications 5, 4213 (2014).
cione, “Implementing a distance-based classifier with a [30] Edward Farhi, Jeffrey Goldstone, and Sam Gutmann, “A
quantum interference circuit,” EPL (Europhysics Let- quantum approximate optimization algorithm,” arXiv
ters) 119, 60002 (2017). preprint arXiv:1411.4028 (2014).
[16] Xi-Wei Yao, Hengyan Wang, Zeyang Liao, Ming-Cheng [31] Stuart Hadfield, Zhihui Wang, Bryan O’Gorman,
Chen, Jian Pan, Jun Li, Kechao Zhang, Xingcheng Lin, Eleanor G Rieffel, Davide Venturelli, and Rupak Biswas,
Zhehui Wang, Zhihuang Luo, et al., “Quantum image “From the quantum approximate optimization algorithm
processing and its application to edge detection: theory to a quantum alternating operator ansatz,” Algorithms
and experiment,” Physical Review X 7, 031041 (2017). 12, 34 (2019).
[17] He-Liang Huang, Xi-Lin Wang, Peter P Rohde, Yi-Han [32] Kosuke Mitarai, Makoto Negoro, Masahiro Kitagawa,
Luo, You-Wei Zhao, Chang Liu, Li Li, Nai-Le Liu, Chao- and Keisuke Fujii, “Quantum circuit learning,” Physical
Yang Lu, and Jian-Wei Pan, “Demonstration of topo- Review A 98, 032309 (2018).
logical data analysis on a quantum processor,” Optica 5, [33] Ryan LaRose, Arkin Tikku, Étude O’Neel-Judy, Lukasz
193–198 (2018). Cincio, and Patrick J Coles, “Variational quantum
[18] D. Zhu, N. M. Linke, M. Benedetti, K. A. Landsman, state diagonalization,” npj Quantum Information 5, 1–10
N. H. Nguyen, C. H. Alderete, A. Perdomo-Ortiz, N. Ko- (2019).
rda, A. Garfoot, C. Brecque, L. Egan, O. Perdomo, and [34] M Cerezo, Kunal Sharma, Andrew Arrasmith, and
C. Monroe, “Training of quantum circuits on a hybrid Patrick J Coles, “Variational quantum state eigensolver,”
quantum computer,” 5 (2019), 10.1126/sciadv.aaw9918. arXiv preprint arXiv:2004.01372 (2020).
[19] Kunal Kathuria, Aakrosh Ratan, Michael McConnell, [35] Carlos Bravo-Prieto, Diego Garcı́a-Martı́n, and José I
and Stefan Bekiranov, “Implementation of a hamming Latorre, “Quantum singular value decomposer,” Physical
distance–like genomic quantum classifier using inner Review A 101, 062310 (2020).
products on ibmqx2 and ibmq 16 melbourne,” Quantum [36] Marco Cerezo, Alexander Poremba, Lukasz Cincio, and
machine intelligence 2, 1–26 (2020). Patrick J Coles, “Variational quantum fidelity estima-
[20] Teague Tomesh, Pranav Gokhale, Eric R Anschuetz, and tion,” Quantum 4, 248 (2020).
Frederic T Chong, “Coreset clustering on small quantum [37] Jacob L Beckey, M Cerezo, Akira Sone, and Patrick J
computers,” arXiv preprint arXiv:2004.14970 (2020). Coles, “Variational quantum algorithm for estimat-
[21] M. Cerezo, Andrew Arrasmith, Ryan Babbush, Simon C ing the quantum fisher information,” arXiv preprint
Benjamin, Suguru Endo, Keisuke Fujii, Jarrod R Mc-
 7

arXiv:2010.10488 (2020). ational fast forwarding for quantum simulation beyond

[38] Sumeet Khatri, Ryan LaRose, Alexander Poremba, the coherence time,” npj Quantum Information 6, 1–10
Lukasz Cincio, Andrew T Sornborger, and Patrick J (2020).
Coles, “Quantum-assisted quantum compiling,” Quan- [57] Andrew Arrasmith, Lukasz Cincio, Andrew T Sorn-
tum 3, 140 (2019). borger, Wojciech H Zurek, and Patrick J Coles, “Vari-
[39] Kunal Sharma, Sumeet Khatri, Marco Cerezo, and ational consistent histories as a hybrid algorithm for
Patrick J Coles, “Noise resilience of variational quantum quantum foundations,” Nature communications 10, 1–7
compiling,” New Journal of Physics 22, 043006 (2020). (2019).
[40] Tyson Jones and Simon C Benjamin, “Quantum compi- [58] Colin J Trout, Muyuan Li, Mauricio Gutiérrez, Yukai
lation and circuit optimisation via energy dissipation,” Wu, Sheng-Tao Wang, Luming Duan, and Kenneth R
arXiv preprint arXiv:1811.03147 (2018). Brown, “Simulating the performance of a distance-3 sur-
[41] Lukasz Cincio, Yiğit Subaşı, Andrew T Sornborger, and face code in a linear ion trap,” New Journal of Physics
Patrick J Coles, “Learning the quantum algorithm for 20, 043038 (2018).
state overlap,” New Journal of Physics 20, 113022 (2018). [59] Suguru Endo, Jinzhao Sun, Ying Li, Simon C Benjamin,
[42] M. Horodecki, P. Horodecki, and R. Horodecki, “General and Xiao Yuan, “Variational quantum simulation of gen-
teleportation channel, singlet fraction, and quasidistilla- eral processes,” Physical Review Letters 125, 010501
tion,” Physical Review A 60, 1888–1898 (1999). (2020).
[43] M. A. Nielsen, “A simple formula for the average gate [60] Yong-Xin Yao, Niladri Gomes, Feng Zhang, Thomas
fidelity of a quantum dynamical operation,” Physics Let- Iadecola, Cai-Zhuang Wang, Kai-Ming Ho, and Peter P
ters A 303, 249–252 (2002). Orth, “Adaptive variational quantum dynamics simula-
[44] Kyle Poland, Kerstin Beer, and Tobias J Osborne, “No tions,” arXiv preprint arXiv:2011.00622 (2020).
free lunch for quantum machine learning,” arXiv preprint [61] Marcello Benedetti, Mattia Fiorentini, and Michael
arXiv:2003.14103 (2020). Lubasch, “Hardware-efficient variational quantum
[45] Kunal Sharma, M. Cerezo, Zoë Holmes, Lukasz Cincio, algorithms for time evolution,” arXiv preprint
Andrew Sornborger, and Patrick J Coles, “Reformu- arXiv:2009.12361 (2020).
lation of the no-free-lunch theorem for entangled data [62] Kentaro Heya, Ken M Nakanishi, Kosuke Mitarai, and
sets,” arXiv preprint arXiv:2007.04900 (2020). Keisuke Fujii, “Subspace variational quantum simula-
[46] Seth Lloyd, “Universal quantum simulators,” Science , tor,” arXiv preprint arXiv:1904.08566 (2019).
1073–1078 (1996). [63] Kishor Bharti and Tobias Haug, “Quantum assisted sim-
[47] Andrew M Childs and Yuan Su, “Nearly optimal lattice ulator,” arXiv preprint arXiv:2011.06911 (2020).
simulation by product formulas,” Physical review letters [64] Jonathan Wei Zhong Lau, Kishor Bharti, Tobias Haug,
123, 050503 (2019). and Leong Chuan Kwek, “Quantum assisted simula-
[48] Jeongwan Haah, Matthew B Hastings, Robin Kothari, tion of time dependent hamiltonians,” arXiv preprint
and Guang Hao Low, “Quantum algorithm for simulat- arXiv:2101.07677 (2021).
ing real time evolution of lattice hamiltonians,” SIAM [65] Tobias Haug and Kishor Bharti, “Generalized quan-
Journal on Computing , FOCS18–250 (2021). tum assisted simulator,” arXiv preprint arXiv:2011.14737
[49] Guang Hao Low, Vadym Kliuchnikov, and Nathan (2020).
Wiebe, “Well-conditioned multiproduct hamiltonian sim- [66] Stefano Barison, Filippo Vicentini, and Giuseppe Carleo,
ulation,” arXiv preprint arXiv:1907.11679 (2019). “An efficient quantum algorithm for the time evolution of
[50] Dorit Aharonov and Amnon Ta-Shma, “Adiabatic quan- parameterized circuits,” arXiv preprint arXiv:2101.04579
tum state generation and statistical zero knowledge,” in .
Proceedings of the thirty-fifth annual ACM symposium on [67] Joe Gibbs, Kaitlin Gili, Zoë Holmes, Benjamin Com-
Theory of computing (2003) pp. 20–29. meau, Andrew Arrasmith, Lukasz Cincio, Patrick J
[51] Dominic W Berry, Andrew M Childs, Richard Cleve, Coles, and Andrew Sornborger, “Long-time simulations
Robin Kothari, and Rolando D Somma, “Exponential with high fidelity on quantum hardware,” arXiv preprint
improvement in precision for simulating sparse hamil- arXiv:2102.04313 (2021).
tonians,” in Proceedings of the forty-sixth annual ACM [68] Benjamin Commeau, Marco Cerezo, Zoë Holmes, Lukasz
symposium on Theory of computing (2014) pp. 283–292. Cincio, Patrick J Coles, and Andrew Sornborger,
[52] Guang Hao Low and Isaac L Chuang, “Optimal hamilto- “Variational hamiltonian diagonalization for dynamical
nian simulation by quantum signal processing,” Physical quantum simulation,” arXiv preprint arXiv:2009.02559
review letters 118, 010501 (2017). (2020).
[53] Dominic W Berry, Andrew M Childs, Richard Cleve, [69] Jirawat Tangpanitanon, Supanut Thanasilp, Ninnat
Robin Kothari, and Rolando D Somma, “Simulating Dangniam, Marc-Antoine Lemonde, and Dimitris G An-
hamiltonian dynamics with a truncated taylor series,” gelakis, “Expressibility and trainability of parameterized
Physical review letters 114, 090502 (2015). analog quantum systems for machine learning applica-
[54] Dominic W Berry, Andrew M Childs, and Robin tions,” arXiv preprint arXiv:2005.11222 (2020).
Kothari, “Hamiltonian simulation with nearly optimal [70] Sukin Sim, Peter D Johnson, and Alán Aspuru-Guzik,
dependence on all parameters,” in 2015 IEEE 56th An- “Expressibility and entangling capability of parameter-
nual Symposium on Foundations of Computer Science ized quantum circuits for hybrid quantum-classical al-
(IEEE, 2015) pp. 792–809. gorithms,” Advanced Quantum Technologies 2, 1900070
[55] Guang Hao Low and Isaac L Chuang, “Hamiltonian sim- (2019).
ulation by qubitization,” Quantum 3, 163 (2019). [71] Jarrod R McClean, Sergio Boixo, Vadim N Smelyanskiy,
[56] Cristina Cirstoiu, Zoe Holmes, Joseph Iosue, Lukasz Cin- Ryan Babbush, and Hartmut Neven, “Barren plateaus
cio, Patrick J Coles, and Andrew Sornborger, “Vari- in quantum neural network training landscapes,” Nature
 8

communications 9, 1–6 (2018).

[72] M Cerezo, Akira Sone, Tyler Volkoff, Lukasz Cincio,
and Patrick J Coles, “Cost function dependent barren
plateaus in shallow parametrized quantum circuits,” Na-
ture Communications 12, 1–12 (2021).
[73] Alexey Uvarov and Jacob Biamonte, “On barren plateaus
and cost function locality in variational quantum algo-
rithms,” arXiv preprint arXiv:2011.10530 (2020).
[74] Samson Wang, Enrico Fontana, Marco Cerezo, Kunal
Sharma, Akira Sone, Lukasz Cincio, and Patrick J Coles,
“Noise-induced barren plateaus in variational quantum
algorithms,” arXiv preprint arXiv:2007.14384 (2020).
[75] Marco Cerezo and Patrick J Coles, “Impact of barren
plateaus on the hessian and higher order derivatives,”
arXiv preprint arXiv:2008.07454 (2020).
[76] Arthur Pesah, M Cerezo, Samson Wang, Tyler Volkoff,
Andrew T Sornborger, and Patrick J Coles, “Absence
of barren plateaus in quantum convolutional neural net-
works,” arXiv preprint arXiv:2011.02966 (2020).
[77] Zoë Holmes, Andrew Arrasmith, Bin Yan, Patrick J
Coles, Andreas Albrecht, and Andrew T Sornborger,
“Barren plateaus preclude learning scramblers,” arXiv
preprint arXiv:2009.14808 (2020).
[78] Andrew Arrasmith, M Cerezo, Piotr Czarnik, Lukasz
Cincio, and Patrick J Coles, “Effect of barren
plateaus on gradient-free optimization,” arXiv preprint
arXiv:2011.12245 (2020).
[79] Carlos Ortiz Marrero, Mária Kieferová, and Nathan
Wiebe, “Entanglement induced barren plateaus,” arXiv
preprint arXiv:2010.15968 (2020).
[80] Taylor L Patti, Khadijeh Najafi, Xun Gao, and Su-
sanne F Yelin, “Entanglement devised barren plateau
mitigation,” arXiv preprint arXiv:2012.12658 (2020).
[81] Zoë Holmes, Kunal Sharma, M Cerezo, and Patrick J
Coles, “Connecting ansatz expressibility to gradi-
ent magnitudes and barren plateaus,” arXiv preprint
arXiv:2101.02138.
[82] Edward Grant, Leonard Wossnig, Mateusz Ostaszewski,
and Marcello Benedetti, “An initialization strategy for
addressing barren plateaus in parametrized quantum cir-
cuits,” Quantum 3, 214 (2019).
 1

Supplementary Information for

“Experimental Quantum Learning of a Spectral Decomposition”

This document provides additional details about the 2. TRAINING

experimental results. In Sec. 1 we describe the on-
line superconducting qubits used in the experiment and
give calibration results (gate errors, coherence times, and The variational circuits for W and D were trained by
single-qubit measurement errors) provided by the back- gradient descent
end. In Sec. 2 we provide additional details about the
circuit training.
∂CLHST
θk ← θk − η
1. QUBITS ∂θk
(S2)
∂CLHST
γl ← γl − η ,
Data was taken on the IBM Q processor ibmq santiago ∂γl
using the BQP software package developed by Geller
and colleagues. BQP is a Python package developed to
design, run, and analyze complex quantum computing where θk and γl are the variational parameters for the W
and quantum information experiments using commercial and D circuits, respectively, and η is the learning rate.
backends. We demonstrate the learning of a spectral de- We used the variable learning rate plotted in Fig. S2a),
composition using the qubits shown in Fig. S1. Calibra- which was optimized by classical simulation. At each
tion data supplied by the backend is summarized in Ta- step of the training, we measure the cost CLHST and gra-
ble I. Here T1,2 are the standard Markovian decoherence dients ∂CLHST /∂θk and ∂CLHST /∂γl at a particular point
times, and (θ1 , · · · , θ18 , γ1 , · · · , γ3 ) in parameter space, and use this
to update (θ1 , · · · , θ18 , γ1 , · · · , γ3 ) according to (S2). We
T (0|1) + T (1|0) also calculate the angle between the measured gradient
= (S1)
2 and the exact gradient simulated classically, which is
is the single-qubit state-preparation and measurement plotted in Fig. S2b). This data shows that the mea-
(SPAM) error, averaged over initial states. The U2 er- sured gradient is correctly pointing uphill until the end
ror column gives the single-qubit gate error measured by of the training when a local minimum is reached and the
randomized benchmarking. The CNOT errors are also gradient becomes small and noisy.
measured by randomized benchmarking.

TABLE I. Calibration data provided by IBM Q for the

ibmq bogota chip during the period of data acquisition.
Qubit T1 (µs) T2 (µs) SPAM error  U2 error
Q0 136.6 178.0 0.056 3.70e-4
Q1 132.3 117.4 0.045 2.01e-4
Q2 65.4 132.4 0.024 2.48e-4
Q3 103.6 158.3 0.020 4.74e-4
CNOT gates CNOT error
CNOT0,1 CNOT1,0 1.04e-3 FIG. S1. Layout of IBM Q device ibmq bogota. In this work
CNOT1,2 CNOT2,1 8.12e-3 we use qubits Q0 , Q1 , Q2 and Q3 for the training and qubits
CNOT2,3 CNOT3,2 3.81e-2 Q1 and Q2 for the Trotter and VFF comparison.
 2

1.2

Gradient Error, rad Learning Rate, η

a)
0.8

0.4

0.0
π/2
b)

−π/2
0 4 8 12 16
Iteration

η0
FIG. S2. (a) Learning rate versus optimization step during training by gradient descent. This rate obeys η(j) = [1+(j/δ)] κ

where j ∈ {0, 1, 2, · · · , 16} is the optimization step number and the hyperparameters η0 = 1.1, κ = 0.5, and δ = 12 were optimized
by classical simulation. (b) Error in the measured gradient direction during training. Here we plot the angle between gradient
measured on ibmq bogota during training, and true gradient calculating classically for the same set of parameters, for each
optimization step.