Improving the Performance of Deep Quantum Optimization Algorithmswith Continuous Gate Sets

Nathan Lacroix,1 Christoph Hellings,1 Christian Kraglund Andersen,1 Agustin Di Paolo,2

Ants Remm,1 Stefania Lazar,1 Sebastian Krinner,1 Graham J. Norris,1

Mihai Gabureac,1 Alexandre Blais,2, 3 Christopher Eichler,1 and Andreas Wallraff1

1Department of Physics, ETH Zurich, CH-8093 Zurich, Switzerland2Institut Quantique and Departement de Physique,

Universite de Sherbrooke, Sherbrooke J1K2R1 Quebec, Canada3Canadian Institute for Advanced Research, Toronto, ON, Canada

(Dated: May 12, 2020)

Variational quantum algorithms are believed to be promising for solving computationally hardproblems and are often comprised of repeated layers of quantum gates. An example thereof isthe quantum approximate optimization algorithm (QAOA), an approach to solve combinatorialoptimization problems on noisy intermediate-scale quantum (NISQ) systems. Gaining computationalpower from QAOA critically relies on the mitigation of errors during the execution of the algorithm,which for coherence-limited operations is achievable by reducing the gate count. Here, we demonstratean improvement of up to a factor of 3 in algorithmic performance as measured by the successprobability, by implementing a continuous hardware-efficient gate set using superconducting quantumcircuits. This gate set allows us to perform the phase separation step in QAOA with a single physicalgate for each pair of qubits instead of decomposing it into two CZ-gates and single-qubit gates.With this reduced number of physical gates, which scales with the number of layers employed in thealgorithm, we experimentally investigate the circuit-depth-dependent performance of QAOA appliedto exact-cover problem instances mapped onto three and seven qubits, using up to a total of 399operations and up to 9 layers. Our results demonstrate that the use of continuous gate sets may bea key component in extending the impact of near-term quantum computers.

I. INTRODUCTION

Quantum computers have the potential to outperformclassical computers on a range of computational problemssuch as prime factoring [1] and quantum chemistry [2].Although many of these applications will require quan-tum error correction [3] to provide a quantum advantage,there is an increasing interest in exploring quantum ap-plications on noisy intermediate-scale quantum (NISQ)devices [4] available in the near-term. Recent experimentshave demonstrated a computational advantage of quan-tum computers [5], explored many-body physics [6, 7] andsimulated small-scale quantum chemistry problems [8–10].Moreover, there is a significant interest in solving opti-mization problems on quantum computers, in particularwith the quantum approximate optimization algorithm(QAOA) [11–13]. This variational algorithm has beenused to study a range of discrete [11, 14–16] and continu-ous [17] optimization problems, and may have applicationsfor unstructured search [18]. While there is currently noproof that it can provide an asymptotic quantum advan-tage, QAOA is an emerging approach for benchmarkingquantum devices and is a candidate for demonstrating apractical quantum speed-up on near-term NISQ devices.

To find an approximate solution to a combinatorialproblem with QAOA, a problem Hamiltonian is formu-lated, whose ground state corresponds to the solution ofthe combinatorial problem. To approximate this groundstate, a quantum computer prepares an ansatz state witha parameterized gate sequence, whose parameters areiteratively updated by a classical optimizer. The gate

sequence consists of layers, each characterized by two vari-ational parameters, γq and βq, see Fig. 1. The numberof layers, p, sets the depth of the algorithm and QAOAcan reach the global optimum of any cost function forp → ∞ [11]. It is therefore expected that the compu-tational power of QAOA increases with p. In practice,however, the number of layers that can be executed reli-ably on near-term quantum computers is limited due tofinite gate errors induced by relaxation, dephasing andpulse imperfections [14, 19].

Small-scale implementations of QAOA, while restrictedto solving problems that can also be efficiently solvedon classical computers, provide crucial insights into thefeasibility and challenges related to the execution of thealgorithm on NISQ devices. Previous studies of QAOAwith superconducting qubits [12, 13, 19–21], photonics [22]and trapped ions [23] highlight the applicability of QAOAon a range of platforms and illustrate the breadth ofproblems that can be addressed with QAOA. The workpresented in Ref. [12] studied the MaxCut problem, whichis the canonical problem for QAOA [11], with up to 19qubits, Ref. [20] studied a channel decoding problem,Ref. [23] searched the eigenstate of all-to-all connectedIsing models with up to 40 qubits and Ref. [21] consideredan exact-cover problem with 2 qubits. Many of theseexperiments consider problems that can be solved withshallow QAOA circuits (p = 1 or 2). However, theseexamples may not be representative of the broad rangeof problems that can be addressed with QAOA. Indeed,studies of all-to-all connected Ising models show that deepcircuits may be needed [13].

arX

iv:2

005.

0527

5v1

[qu

ant-

ph]

11

May

202

0

2

When implementing quantum algorithms on a quan-tum device, it is common to decompose the gate sequenceinto a discrete set of gates available on the hardware. Toimprove performance, recent experiments have exploredcontinuous gate sets motivated by applications in quan-tum simulations [24, 25], quantum chemistry [26, 27] andfor QAOA using XY interactions [28]. In this work, webenchmark QAOA with a continuous hardware-efficientgate set. We present a controlled arbitrary-phase gate(C-ARB gate), which allows to execute each QAOA layerwith only one two-qubit gate per ZZ-term in problemHamiltonians formulated as Ising models, see Fig. 1(a).We demonstrate how our gate set shortens the QAOAsequence and, thus, leads to better performance for a fixedQAOA depth compared to a decomposed implementationof the algorithm with a discrete gate set. In particu-lar, we demonstrate with two concrete examples that thereduction in gate sequence duration outweighs errors orig-inating from the interpolation of parameters necessary forimplementing the continuous gate set. Taking advantageof this gain in performance, we investigate the trade-offbetween experimental noise, which favors shallow circuits,and increasing the number of layers, which is needed tosolve complex problem instances.

II. IMPLEMENTATION

The objective function of many NP-complete discreteoptimization problems can be mapped to an Ising Hamil-tonian [29, 30],

C =∑i<j

JijZiZj +

n∑i=1

hiZi, (1)

where Zi is the Pauli-Z operator for spin i. QAOA canfind the ground state of this Hamiltonian by minimiz-

ing the expectation value of C for the ansatz state |~γ, ~β〉where ~γ = (γ1, . . . , γp), ~β = (β1, . . . , βp) are variationalparameters. In particular, the quantum circuit prepar-

ing |~γ, ~β〉 consists of p layers each containing a phase-

separation operator UC = e−iγqC and a mixing operator

UB = e−iβqB, where B =∑iXi, with q = 1, . . . , p [11].

Since all terms of C commute, we can implement each

term U ijC = e−iΓij2 ZiZj separately, where Γij/2 = γqJij is

a continuous parameter.A common approach is to decompose U ijC into a gate

sequence consisting of two conditional phase rotations ofπ, i.e. standard CZ-gates, combined with several single-qubit gates [12, 21]. We present such a decomposition inFig. 1(b), where the dependence on the continuous param-eters Γij is introduced via an arbitrary-angle single-qubitZ-rotation. An alternative approach is to use a singlecontrolled arbitrary-phase gate (C-ARB gate), which canadd any desired phase factor e−iφ to the |11〉 state. Thisgate naturally applies the angle 2Γij and, together with

25 50 75Pulse length, l (ns)

0.9

1.0

Puls

e am

plitu

de, a

(V)

0 ¼ 2¼

Cond. phase, Á (rad.)

0.0 0.5 1.0

Population, Pe -60

-30

0

30 Det

unin

g, ¢

(MH

z)

(a) QAOA with C-ARB gates

(c) (d)

Qi

Qj

UijUBUijUC( ) px

Y-π/2 Z2β Y π/2

Y-π/2 Z2β Y π/2

ZijΓ

ZijΓ

Zij2Γ

Jij

H H H H

Qi

Qj

(b) QAOA with CZ gates

ZijΓ

Y-π/2 Z2β Y π/2

Y-π/2 Z2β Y π/2

Jij

UijUBUijUC( ) px

a l

FIG. 1. (a) Quantum circuit of a layer q of QAOA for the two-qubit subspace |QiQj〉, using the controlled arbitrary-phasegate (blue) to rotate the |11〉 state by an angle 2Γij whereΓij = 2γqJij . (b) A QAOA layer with the phase-separationunitary U ij

C decomposed into CZ gates (green) and additionalHadamard gates and single-qubit Z-gates. (c) Excited-statepopulation Pe of the control qubit Qi = A1 brought in inter-action with the target qubit Qj = B2 via a flux pulse. Weperform a two-dimensional sweep of flux pulse amplitude aand flux pulse length l, and indicate the maximum populationrecovery with blue dots. (d) Conditional phase for the dotsindicated in (c). The green diamond corresponds to the CZ-gate. The right axis indicates the detuning between |11〉 and|20〉. The inset depicts the pulse sequence used to measurethe conditional phase. Single-qubit π-pulses and π/2-pulsesare shown in dark blue and purple, respectively. The fluxpulse (light blue) of amplitude a and length l is applied to thecontrol qubit Qi, see Appendix B for more details.

two single-qubit Z-rotations, realizes the unitary U ijC , seeFig. 1(a).

In QAOA, the number of unitaries U ijC grows linearly

with the number of two-qubit terms in C and with thenumber of QAOA layers, p. Thus, it is essential that eachU ijC is implemented with high fidelity. The direct imple-mentation we present in this work significantly reducesboth the physical gate count and the sequence duration.Thus, this approach is expected to find correct solutionsto complex problems with higher probability.

We run QAOA on a quantum device with 7 super-conducting transmon qubits, see Appendix A for deviceparameters and a false-colored micrograph of the de-vice. The qubits are pairwise connected as illustratedin Fig. 2(a). Single-qubit X and Y -rotations are im-plemented with microwave pulses, while Z-rotations are

3

performed as virtual gates [31] which take zero time asthey are implemented through a redefinition of the ref-erence frame. To realize CZ-gates, we use a standardapproach relying on a flux pulse which shifts the tran-sition frequency of one of the qubits to bring the |11〉state of a pair of coupled qubits in resonance with thenon-computational |20〉 state [32–34]. The resulting hy-bridization leads to a coherent population oscillation be-tween the two states. The frequency detuning betweenthe |11〉 and |20〉 states, ∆, is 0 during the gate, andafter an interaction of the duration corresponding to oneoscillation period, the population returns to the |11〉 statewith an added phase of π, see green diamond in Fig. 1(c)and (d). We generalize the CZ-gate to a C-ARB gate onour device by exploiting near-resonant interactions of the|11〉 and |20〉 states [24], i.e. ∆ 6= 0, to acquire conditionalphase angles ranging from 0 to 2π, see Fig. 1(d). We vary∆ by sweeping the flux pulse amplitude and simultane-ously adapting the pulse length to maximize populationrecovery in the computational subspace, see blue dotsin Fig. 1(c). Details about the gate implementation areprovided in Appendix B.

We compare the performance of both approaches ontwo example instances of the NP-complete exact-coverproblem [29]. The aim of exact cover is to decide whetherit is possible to cover all elements in a set S exactly onceby an appropriate selection of subsets {Vi} from a givencollection of subsets V . In the example visualized inFig. 2(b), each row corresponds to an element of a three-element set S, while each column corresponds to a subsetVi out of three given subsets. The dots visualize whichelements (rows) are included in a subset (column). In thispicture, the task is to find a selection of columns such thateach row is covered by exactly one dot. This condition isfulfilled by two solutions: selecting the first two columns orselecting the last column. In a mathematical formulationof the exact-cover problem (see Appendix C), the gridin Fig. 2(b) corresponds to a visual representation of anincidence matrix K, where a dot in row ` and column iindicates an entry K`i = 1 while empty cells in the gridindicate entries equal to 0. When mapping an instanceof exact cover to an Ising Hamiltonian [30, 35], the i-thqubit encodes whether a subset Vi is selected or not, seeAppendix C. In the visualization in Fig. 2(b), the qubitthat represents a subset is indicated by the label abovethe column and by the color used for the dots. Fig. 2(c)shows an example of a larger instance of exact cover withseven subsets, requiring seven qubits.

To focus on the comparison between the two methodsfor realizing the two-qubit unitaries U ijC , these two prob-lem instances are chosen such that the resulting IsingHamiltonians respect the hardware connectivity graphof our device, see Fig. 2(a), and that all single-qubitterms vanish, i.e. hi = 0. The three-qubit probleminstance depicted in Fig. 2(b) yields an Ising Hamilto-nian with JA1B2

= 0.5 and JA2B2= 1. In the basis

|A1A2B2〉, the two possible selections of columns cov-ering all rows, namely A = {A1, A2} and B = {B2},

(a)

(b)

Solutions{| ⟩1 1 0| ⟩0 0 1

01

01

1 10 0

A2A1

A3 A4

B1 B2 B3

(c) A1 A2 B2A3 A4 B1 B3

| ⟩1 1 0| ⟩0 0 1

A1 A2 B2

{

FIG. 2. (a) Hardware connectivity graph of the quantumdevice. Dots correspond to qubits and edges indicate betweenwhich pairs of qubits two-qubit gates can be realized. Thegrey dashed line indicates the subset of qubits used for thethree-qubit problem instance depicted in (b). (b) Visual rep-resentation of the incidence matrix K (dots indicating entriesK`i = 1) for a chosen three-qubit exact-cover problem instance.The labels above the columns (and the colors) indicate whichphysical qubits are used to represent the corresponding subset.The two solution states are indicated below the grid. (c) Visualrepresentation of the incidence matrix for a chosen seven-qubitproblem instance.

are encoded with the states |110〉 and |001〉, respec-tively, where a 1 in position i indicates that the i-thcolumn of K is included in the selection of subsets. Forthe seven-qubit problem instance of Fig. 2(c), we haveJA3B2 = 0 and Jij = 0.5 for all other physically connectedqubit pairs. This instance also possesses two solutions,A = {A1, A2, A3, A4} and B = {B1, B2, B3}, correspond-ing to the states |1111000〉 and |0000111〉, respectively,using the basis |A1A2A3A4B1B2B3〉. Note that we havelabeled the qubits in Fig. 2(a) such that the solutionsalways correspond to either selecting the qubits labeledwith A or the qubits labeled with B, see Appendix C.

The QAOA circuit solving the three-qubit probleminstance consists of 15 (32) operations per layer while thecorresponding circuit for the seven-qubit problem instanceconsists of 42 (98) operations per layer, for the direct(decomposed) implementation, respectively. Thus, theseven-qubit problem instance, which requires deep QAOAcircuits, yields a circuit comprising 259 (399) operationsin total for the direct (decomposed) implementation atp = 6 (p = 4), see Appendix G for more details.

III. PERFORMANCE OF QAOA

A single-layer QAOA implementation (p = 1) is a use-ful intermediate benchmark towards implementing multi-layer QAOA circuits since there are only two variational

parameters ~γ = (γ1) and ~β = (β1), hereafter referred to asγ and β for ease of notation, which allows us to map outthe full optimization landscape experimentally. As furtherdiscussed in Appendix D, when p = 1, the cost-function

4

landscape is π/2-periodic in β for a problem withoutsingle-qubit terms. Moreover, since all eigenvalues ofC are odd multiples of 1

2 , the landscape is 2π-periodicin γ. Finally, the landscape is always point-symmetricaround the center point of a period. We can thus reduceour considerations to γ ∈ [0, π[ and β ∈ [0, π/2[. Foreach pair of parameters, we prepare the state |γ, β〉 20000times, see Appendix G for the full pulse sequence, and weevaluate the cost function C(γ, β) = 〈γ, β| C |γ, β〉. Weuse a three-level readout scheme discussed in Appendix A,which allows us to discard the measurement outcomeswith leakage outside of the computational space, see Ap-pendix H. In the context of QAOA, discarding leakageevents corresponds to reducing the effective number ofshots available for evaluating the cost function by reject-ing outcomes that are not valid bit-strings. In this regard,leakage is different from other undetectable errors, forwhich such a post-selection cannot be done.

We observe that the resulting cost-function landscapes,see Fig. 3, are odd functions of β with a line symmetryaxis at β = π/4, see Appendix D. The locations of allextrema in the measured landscape, see Fig. 3(a), are ingood agreement with noise-free simulations, see Fig. 3(b),which suggests that the coherent errors are small in ourimplementation. Errors due to decoherence mostly affectthe contrast of the landscapes [36], see Fig. 3(c) andAppendix E. The distortions of the local extrema locatedat γ > π/2 are attributed to the residual ZZ-couplingbetween the qubits [37], which we confirm with master-equation simulations, see Appendix E.

By embedding the evaluation of C(γ, β) measured onthe quantum device into a classical Nealder-Mead opti-mizer, we demonstrate that the landscape is suitable ascost function for a classical optimizer. The closed-loopclassical optimizer finds the optimal parameters for mostrandom initialization parameters, see Fig. 3(d-f), how-ever, some convergence traces get trapped in local minima.Note that in this single-layer implementation, the costnever reaches the ground-state energy Cgs = −1.5, nei-ther in the measurement nor in the noise-free simulation,which indicates that QAOA circuits of larger depth areindeed required for this problem.

To obtain better approximate solutions to the com-binatorial problem instance, we execute QAOA circuitswith additional layers and study the effect of the depthp on the output state distribution. To investigate theperformance of the quantum part of QAOA rather thanthe performance of the classical optimizer, we initializethe algorithm with optimal parameters obtained fromnoise-free simulations. We then optimize these parame-ters locally to correct for small coherent errors, and weestimate the resulting state distribution as a function ofdepth from 20000 single-shot measurements, see Fig. 4 forthe three-qubit case. Three layers are required in noise-free simulations (black wire-frames) to fully concentratethe probability distribution on the two solution states|110〉 and |001〉 corresponding to the selection of subsetsA and B, respectively.

0

¼/4

¼/2

Mixi

ng a

ngle

, ¯ (a) ¯ 00

¯ 0

(b)

0

¼/4

¼/2

Mixi

ng a

ngle

, ¯(e)

0 ¼/2 ¼ Separation angle, °

-1

0

1

Cost

, C(°;¯) (c)

¯ 00 = 38¼

¯ 0 = 18¼

0 ¼/2 ¼0

20

40

fe

(d)

0 20 40Func. eval., fe

-1

0

1

Cost

, C(°;¯)(f)

-1 0 1Cost, C(°; ¯)

FIG. 3. Cost function evaluated for p = 1 on the three-qubitproblem instance, using C-ARB gates. (a) Cost-function land-scape as a function of variational parameters measured withdirect implementation of C-ARB gates. (b) Cost-function land-scape obtained from noise-free simulations. (c) Experimentalevaluation (blue) and simulation (black) of the cost functionfor two horizontal line cuts of (a) and (b), with β′ = π/8(dotted lines) and β′′ = 3π/8 (dashed lines) respectively. (d,e)10 convergence traces of the separation angle and the mixingangle, respectively, for end-to-end optimization starting fromrandom parameter initialization. (f) Average energy (solidblue line) and individual convergence traces (faded lines) ofthe energy corresponding to parameters shown in (d,e).

We quantify the experimental outcomes using the clas-sical fidelity [38] between the output state probabilitydistribution arising from the measurements, P , and fromnoise-free simulations, P ,

F(P, P ) =

(∑i

√Pi

√Pi

)2

(2)

where Pi and Pi correspond to the probabilities of the i-thbasis state in the Hilbert space. Note that 0 ≤ F(P, P ) ≤1 with F(P, P ) = 1 if and only if P = P . The statedistributions of the implementation using C-ARB gates,see filled bars in Fig. 4(a-c), have fidelities of 98.93 %,95.93 %, and 86.20 % for p = 1, 2, and 3 respectively,with respect to the corresponding distribution obtainedwith noise-free simulations. As expected, the reductionof the fidelity with the number of layers p illustrates theaccumulation of errors in circuits of increasing depth.However, the concentration of probability on solutionstates as p increases is stronger than the detrimental effectof the additional errors, such that overall, the probabilityof measuring a solution increases with p. By contrast,in the implementation using CZ-gates, see Fig. 4(d-f),the concentration of probability on solution states onlycompensates the additional errors for p = 2 while theerrors outweigh the gain of an additional layer for p = 3.This is also reflected by lower fidelities of 96.37 %, 87.43 %,and 64.52 % for p = 1, 2, and 3 respectively, and is

5

explained by the fact that decoherence and residual ZZ-coupling accumulate over the longer gate sequence.

Master-equation simulations (red wire-frames) are inexcellent agreement with the measured distributions, seeAppendix E for details. We confirm from these simulationsthat decoherence is the main limitation in this experimentwhile residual ZZ-coupling cause additional errors, inparticular for the decomposed implementation.

The landscape and state probability distributions of theseven-qubit problem instance presented in Appendix Flead to a similar conclusion, i.e. the direct implementationexploiting C-ARB gates is in better agreement with noise-free simulations than the decomposed implementation.

It is expected that additional layers in QAOA circuitsincrease the number of reachable states, thereby leadingto better approximate solutions in the absence of experi-mental noise. To gain further insight into the trade-offbetween the extended reachable state space and the addi-tional noise resulting from increased depth, we determinethe enhancement of the success probability provided bythe output state distribution over a uniform state distri-bution as a function of p for both problem instances, seeFig. 5. We define the enhancement as Ps/Pu, where Ps isthe success probability, i.e. the sum of the probabilities ofall solution states, and Pu = 2/(2N − 1) is the probabilityof sampling a solution from a uniform probability distri-bution over all possible states. Note that we exclude the

state |0〉⊗N , which is never a solution in the context ofexact cover. We indicate the sequence duration for bothimplementations with additional axes in Fig. 5, wheresequence duration is defined as the time between the start

0.00

0.25

0.50

Stat

e pr

obab

ility,

p|A

1A2B

2⟩

(a)p=1

(b)p=2

(c)p=3

|000

⟩|0

01⟩

|010

⟩|0

11⟩

|100

⟩|1

01⟩

|110

⟩|1

11⟩0.00

0.25

0.50(d)

p=1

|000

⟩|0

01⟩

|010

⟩|0

11⟩

|100

⟩|1

01⟩

|110

⟩|1

11⟩

(e)p=2

|000

⟩|0

01⟩

|010

⟩|0

11⟩

|100

⟩|1

01⟩

|110

⟩|1

11⟩

(f)p=3

State label, |A1A2B2⟩

FIG. 4. Output state probability distribution for the three-qubit problem instance implemented with controlled arbitraryphase gates (a,b,c), and decomposed using CZ-gates (d,e,f).States are measured at optimal parameters for depth of p = 1(a,d), p = 2 (b,e), and p = 3 (c,f). The filled bars correspondto the measured state probabilities in which we highlightthe problem solutions in blue (direct implementation) andgreen (decomposed implementation), respectively. The blackwire-frames are the expected QAOA outcome from noise-freesimulations and the red wire-frames are from master-equationsimulations.

of the initialization pulse and the start of the readout.

In the three-qubit case, Fig. 5(a), the direct (decom-posed) implementation shows a maximal enhancement ofsuccess probability of 3 (2.4) at p = 3 (p = 2). The directimplementation (blue dots) provides a higher enhance-ment than the decomposed implementation (green dots)because the sequence duration is shorter for a fixed p,and the problem instance requires at least p = 3 to reachmaximal enhancement in an ideal setting (black squares).

When the problem increases in complexity, the numberof layers required to reach maximal enhancement in anoise-free scenario also increases. For the seven-qubitinstance, we find that p = 6 is required to reach a successprobability above 90%, see Fig. 5(b). Consequently, theability to execute more layers in shorter time provides aneven more pronounced advantage. In particular, for thedirect implementation, we find that increasing p from 1to 2 increases the enhancement of the success probabilityto 9.5. However, when further increasing to p = 3, theextended reachable state space does not compensate theadditional noise arising from the increased sequence du-ration. For the decomposed version, going beyond p = 1does not provide any benefits. Thus, for the seven-qubitproblem we only benefit from adding layers when tak-ing advantage of the directly implemented C-ARB gates,which improves the performance by a factor of 3 comparedto the decomposed implementation.

Finally, to emphasize that the limitations for deepercircuits are directly related to the increased sequenceduration rather than the depth itself, we notice that for afixed sequence duration of L = 5µs, both implementations

0 2 4 6QAOA depth, p

1.0

10

N=7

0.1

1.0

Succ

ess p

rob.

, Ps

0 2 4Seq. duration, L (¹s)

0 4 8 12

0 5QAOA depth, p

1.0

10

Enha

ncem

ent, Ps=P

u

N=3

0.5

1.0

0 2Seq. duration, L (¹s)

0 3 6 9

(a) (b)

FIG. 5. Performance of QAOA for (a) three qubits and (b)seven qubits. The blue points are implemented with the di-rect controlled arbitrary-phase gate and the green points areimplemented with the gate sequences decomposed into CZgates. The black squares indicate the highest success probabil-ities found with noise-free simulations. The top axes indicatethe sequence duration for the direct implementation and thedecomposed implementation in green and blue, respectively.The gray areas indicate success probabilities above 1.

6

of the seven-qubit instance show similar enhancement ofsuccess probability despite being of depth p = 6 (direct)and p = 2 (decomposed).

IV. DISCUSSION

In this work, we show that controlled arbitrary phasegates (C-ARB gates) enable a significant reduction of thenumber of physical gates required to implement QAOAcircuits of any depth on quantum hardware. We demon-strate the advantage of this approach by comparing itto a standard QAOA decomposition on two problem in-stances of the exact-cover problem, with three and sevenqubits, respectively. Despite a more demanding calibra-tion scheme requiring interpolation of gate parameters,C-ARB gates in QAOA circuits systematically outper-form the decomposed alternative for a fixed depth and areable to benefit from the extended reachable state spaceof more layers.

We foresee an even more pronounced advantage forlarger-scale combinatorial optimization problems becausethe number of layers required to solve problems withQAOA is expected to scale with the number of qubitsinvolved in the experiment [39, 40], in particular for denseproblem graphs. In addition, the number of physicaltwo-qubit gates saved within each layer also scales withthe number of two-qubit terms in the cost Hamiltonian.Our results demonstrate that hardware-efficient gate setsare key components in extending the impact of near-termquantum applications, which may become even more rele-vant when solving problem instances that do not matchthe connectivity of the hardware. For example, it hasrecently been observed that the need for swap-gates cansignificantly reduce the performance of a QAOA imple-mentation if a decomposed implementation of swap gatesis used [13]. A direct, hardware-efficient implementationcombining a controlled arbitrary phase and a swap-gatemay therefore be another key component to improve theperformance in these cases, and should be considered infuture research.

ACKNOWLEDGMENTS

The authors are grateful for valuable feedback on themanuscript by B.R. Johnson and D. Schuster and forvaluable discussions with A. Choquette-Poitevin andP. Vikstal. The authors acknowledge J. Heinsoo for earlycontributions towards the realization of C-ARB gates andS. Storz, F. Swiadek and T. Zellweger for their contribu-tions to the measurement setup.

The authors acknowledge financial support by theEU Flagship on Quantum Technology H2020-FETFLAG-2018-03 project 820363 OpenSuperQ, by the Office ofthe Director of National Intelligence (ODNI), IntelligenceAdvanced Research Projects Activity (IARPA), via theU.S. Army Research Office grant W911NF-16-1-0071, by

the National Centre of Competence in Research QuantumScience and Technology (NCCR QSIT), a research instru-ment of the Swiss National Science Foundation (SNSF),by the SNFS R’equip grant 206021-170731 and by ETHZurich. This work was undertaken thanks in part to fund-ing from NSERC, Canada First Research Excellence Fundand ARO W911NF-18-1-0411. The views and conclusionscontained herein are those of the authors and should notbe interpreted as necessarily representing the official poli-cies or endorsements, either expressed or implied, of theODNI, IARPA, or the U.S. Government.

Appendix A: Experimental Setup and DeviceParameters

The experiments described in this manuscript are per-formed in a cryogenic setup [41, 42], the wiring scheme ofwhich is summarized in Fig. 6. Each qubit is controlled bya flux line for frequency tuning which enables two-qubitgates, and a microwave drive line for realizing single-qubitgates. The pulses are generated with arbitrary waveformgenerators (AWGs). The drive pulses are generated at anintermediate frequency of 100 MHz and upconverted tomicrowave frequencies. Multiplexed readout is performedvia two feedlines [42, 43] with the readout pulses generated

UHFQA

300 K

Control AWG

UC

4 K

0.1 K

9 mK

HE

MT

UC DC

Flux AWG

WA

MP

Flux x

7

Dri

ve x

7

TW

PA

RO x 2

Quantum Device

IQ mixer

Control/Readoutline

MW generator

Attenuator (20, 10, 3 dB)

Bias-T

Amplifier

Filter (BP, LP, Eccosorb)

Voltagesource

(a)

FIG. 6. Experimental setup. The experiment is controlledby AWGs whose signals are routed to the quantum devicethrough a series of bandpass filters (BP), lowpass filters (LP)and Eccosorb filters. The flux pulses are combined with avoltage source using a bias-T. The IQ signal from the controlAWG is upconverted (UC) to a microwave signal using an IQmixer. The readout signal is generated by the UHFQA, andthe output from the quantum device is amplified by a chainof amplifiers before being downconverted (DC) and analyzedby the UHFQA.

7

TABLE I. Measured parameters of the seven qubits.

A1 A2 A3 A4 B1 B2 B3

Qubit frequency, ωq/2π (GHz) 5.462 5.684 4.077 4.195 4.825 4.920 5.165Lifetime, T1 (µs) 12.9 6.7 24.5 21.1 16.7 14.8 14.3Ramsey decay time, T ∗2 (µs) 18.1 12.2 7.4 4.6 27.2 24.5 13.3Readout frequency, ωr/2π (GHz) 6.611 6.836 5.832 6.063 6.255 6.042 6.300Readout linewidth, κeff/2π (MHz) 7.5 10.6 6.0 7.2 17.3 10.9 11.0Dispersive shift, χ/2π (MHz) -2.5 -2.5 -0.75 -1.0 -1.25 -2.4 -2.0Thermal population, Pth (%) 0.04 0.01 0.2 0.8 0.3 0.04 0.2|0〉 readout assignment prob. (%) 99.98 99.97 96.54 96.85 99.47 99.92 99.97|1〉 readout assignment prob. (%) 98.06 96.19 90.39 88.94 97.70 94.45 98.20|2〉 readout assignment prob. (%) 96.63 89.68 78.90 80.86 95.18 94.89 96.96

A1 A2

A3 A4

B1 B2 B3

500 um

FIG. 7. Optical micrograph of the device. Each qubit is coloredcorresponding to the Fig. 2 and connected to neighboringqubits by coupling resonators (white). Each qubit is connectedto a readout resonator (red), a flux line (green) and a driveline (pink).

by an ultra-high frequency quantum analyzer (UHFQA).The measurement signals at the output ports of the sampleare first amplified with a wide-bandwidth near-quantum-limited traveling-wave parametric amplifier (TWPA) [44],then with a high-electron-mobility transistor (HEMT)amplifiers and finally with low-noise, room-temperatureamplifiers (WAMP). Thereafter, the signals are downcon-verted and processed using the weighted integration unitsof the UHFQAs.

The quantum device [42] shown in Fig. 7, is fabri-cated on a high-resistivity intrinsic silicon substrate. Pho-tolithography and reactive ion etching are used to de-fine resonators, signal lines and qubit structures in a150 nm thin niobium film sputtered onto the substrate.We also add air bridges to the device to establish a well-defined ground plane and for cross-overs in signal lines.The Al/AlOx/Al Josephson junctions of the transmonqubits are fabricated using electron-beam lithography andshadow evaporation.

The parameters of the device listed in Table I, aremeasured using standard spectroscopy and time-domainmethods. For the readout, we characterize the ability toidentify the correct qubit state as well as the second ex-cited state of each transmon qubit. In particular, we usetwo weighted-integration units per qubit to distinguish |0〉from |1〉 and |1〉 from |2〉, respectively. A standard Gaus-sian mixture model is then used to classify the resulting

A2A1

A3 A4

B1 B2 B3

3.09%1.70%

2.61%2.97%

0.18%

0.20%0.34%

0.25% 0.26%

0.37%

0.52%

3.17% 2.09%4.95%

-44kHz

-80kHz-141kHz

-53kHz-153kHz

-26kHz-28kHz

-34kHz

FIG. 8. Single-qubit error per gate (vertices) and two-qubiterror per gate (edges) in percent, measured with randomizedbenchmarking. We indicate the residual ZZ-coupling betweeneach pair of coupled qubits below the corresponding two-qubitgate infidelity. The gate between A3 and B2 is not needed forthe problem instances considered in this work (dashed line).

integrated weights of each single shot measurement.To characterize the gate performance, we perform ran-

domized benchmarking on all qubits to find the error persingle-qubit Clifford, see Fig. 8. For the two-qubit gates,we only characterize with a fixed conditional phase of πsuch that the two-qubit gate is in the Clifford group andwe measure the error per gate from interleaved random-ized benchmarking, see infidelities next to lines indicatingthe coupling elements in Fig. 8.

We also characterize the residual ZZ-coupling, αij ,between pairs of coupled qubits, as the frequency shift ofqubit i when qubit j is in the excited state. We verifythat αij = αji. The experimentally determined residualZZ-couplings are shown below each error rate in Fig. 8.

Appendix B: Controlled Arbitrary Phase Gate

The goal of a C-ARB gate is to apply a unitary in atwo-qubit subspace which adds a desired phase φ to the|11〉 state,

U(φ) =

1 0 0 00 1 0 00 0 1 00 0 0 e−iφ

(B1)

8

We realize this unitary by exploiting near-resonant inter-actions of the |11〉 and |20〉 states. In particular, to spana conditional phase in the range [0, 2π[, we vary the fre-quency detuning ∆ between the |11〉 and |20〉 states witha flux pulse of amplitude a and length l. During the fluxpulse, the population of the |11〉 state is transferred tothe |20〉 state and after a time τg, the population returnsto the |11〉 state with a phase

φ = π

(1 +

∆√4J2 + ∆2

), (B2)

where J is the constant coupling strength between the|11〉 and |20〉 states.

To calibrate the flux pulse amplitudes and lengths, westart by measuring the population in the |11〉 state as afunction of pulse amplitude and length, yielding a charac-teristic Chevron pattern (Fig. 1(c)). For each measuredamplitude, we fit the |11〉 population to a cosine to ex-tract the pulse length maximizing the population recoveryin the computational subspace. Next, we measure theconditional phase for 45 flux pulse amplitudes. For eachamplitude, the control qubit is brought to the excitedstate with a π-pulse while the target qubit is brought to asuperposition state with a π/2-pulse. Then, a flux pulseis applied to the control qubit and finally a π-pulse and aπ/2-pulse are applied to the control and the target qubit,respectively. We extract the conditional phase by varyingthe phase of the second π/2-pulse and comparing thephase of the target qubit with and without the initial π-pulse applied to the control qubit. We use the calibratedpulse lengths to ensure high population recovery in thecomputational subspace for each amplitude. We interpo-late the pulse length linearly in between calibration pointsif required to reach all phases in the range 0 to 2π. Notethat on our device, we acquire phase from 0 to −2π butreverse the sign in Fig. 1(d) for convenience. Finally, theflux-biased qubit also acquires a dynamic phase φD [33].We compensate for this single-qubit phase shift with avirtual Z-gate after the C-ARB gate. To calibrate φD,we compare the phase of the qubit with and without theflux pulse, for amplitudes in the range of interest. Thenumber of calibration points is gate-dependent and is setsuch that we can unwrap the dynamic phase as a functionof flux pulse amplitude without ambiguity. This allows usto interpolate (with cubic splines) the dynamic phase be-tween calibration points. Gates between specific pairs ofqubits also include additional flux pulses on neighboringqubits to avoid undesired interactions, see Appendix G.We calibrated the dynamic phase acquired by these neigh-boring qubits simultaneously to the dynamic phase of thequbit directly involved in the gate.

The calibration procedures are automated such thathuman interaction is only required to verify the quality ofthe fits. The approach is thus scalable to larger devices.Note that gate architectures allowing the acquisition ofconditional phase as a linear function of the flux pulselength could further simplify and speed up the calibrationprocedure [45].

Appendix C: Exact Cover to Ising

The exact-cover problem is mathematically formu-lated as follows [29]. Given a collection of subsetsV = {Vi}i∈1,...,n with Vi ⊆ S, the task is to verify whetherthere exists a set of indices I ⊆ {1, . . . , n} such that{Vi}i∈I forms a partition of S, i.e., the sets in {Vi}i∈I aredisjoint and their union equals S. This is the case if

0 = min(b1,...,bn)∈{0,1}n

∑`

(1−

∑i

K`ibi

)2

(C1)

where the element K`i of the incidence matrix K is 1if the `-th element of S is contained in subset Vi and 0otherwise. A bit value bi = 1 indicates that the subset Vi isselected. Using spins zi ∈ {±1} instead of bits bi = zi+1

2 ,multiplying out, and dropping additive constants, theoptimization problem can be formulated as [30, 35, 46]

min(z1,...,zn)∈{±1}n

∑i<j

Jijzjzi +∑i

hizi (C2)

where

Jij =∑`

K`iK`j

2(C3)

hi =∑`

K`i

−1 +1

2

∑j

K`j

. (C4)

Solving this optimization problem is equivalent to findingthe ground-state energy of the Ising Hamiltonian, seeEq. (1), with Jij and hi values given by Eq. (C3) andEq. (C4), respectively. In the visual representations ofthe incidence matrices depicted in Fig. 2(b) and (c), thebullets represent the entries with K`i = 1 while emptycells correspond to K`i = 0. By substituting these valuesof K`i into the above equations, we see that hi = 0 for alli in both problem instances, and we obtain the values ofJij given in Section II.

To run QAOA without requiring swaps, we need all-to-all physical connectivity between qubits that occur jointlyin any row of the incidence matrix K. For the physi-cal connectivity graph shown in Fig. 2(a), this meansthat each row can contain only up to two nonzero en-tries. The positive sign in Eq. (C3) reveals that the spinscorresponding to a row with two nonzero entries havean antiferromagnetic coupling. This is in line with theexact-cover constraint, which requires that exactly one ofthem is selected in a valid solution, but not both. Thus,any problem instance that does not require swaps on ourdevice and that does not decompose into a set of isolatedsubgraphs must correspond to a lattice of antiferromag-netically coupled spins. Then, either all qubits labeledwith A or all qubits labeled with B have to be in anexcited state in a valid solution. In the presence of a row(or rows) with a single nonzero entry, some external fieldterm(s) hi of the Ising Hamiltonian become(s) nonzero

9

and the solution that fulfills the exact-cover conditionalso for this row (these rows) is favored. Otherwise, bothsolutions are valid, which is the case for the probleminstances considered in our experiments.

Appendix D: Properties of QAOA Landscapes

Following Ref. [11], the parameter γq can be restricted

to [0, 2π[ if the problem Hamiltonian C has integer eigen-values, while βq can always be restricted to [0, π[. In thisappendix, we discuss further periodicity and symmetryproperties of the QAOA cost function, which enable usto reduce the parameter space and better understandthe cost-function landscapes we observe. To this end, weconsider the cost of a p-layer QAOA circuit,

C(~γ′, ~β′) = 〈+|U(~γ′, ~β′)† C U(~γ′, ~β′) |+〉 (D1)

in which U(~γ′, ~β′) =∏q e−iβ′

qBe−iγ′qC is the p-layer

QAOA unitary with ~γ′ = (γ′1, . . . , γ′p) and ~β′ =

(β′1, . . . , β′p).

If the eigenvalues of C are integer multiples of α, thenby setting γ′q = γq + 2π

α in Eq. (D1) and noting that

e±i 2πα C = I is the identity, we find that C(~γ′, ~β′) is (2π/α)-

periodic in γq.

In addition, if all eigenvalues of C are odd multiples

of α, we have e±i πα C = −I, where the minus sign is aglobal phase, so that the cost is (π/α)-periodic. For bothproblem instances considered in this work, the eigenvaluesof C are odd multiples of 1/2, so that the landscapes are2π-periodic.

Inserting β′q = βq + π into Eq. (D1) and noting that

e±iπB =∏i e±iπXi =

∏i(−I) yields the π-periodicity in

βq mentioned in [11]. Moreover, since e−iπ2 B =∏i e−iπ2Xi

corresponds to an Xπ rotation of all qubits, setting β′p =βp + π

2 in the last layer p corresponds to flipping the sign

of all spins before estimating the energy of C. If theIsing Hamiltonian C does not contain single-qubit terms(hi = 0 for all i), this sign flip does not change the energy,and the cost landscape is π

2 -periodic in βp. As this appliesto the examples considered in this paper, we measure thelandscapes for p = 1 only up to β = β1 = π

2 .By simultaneously setting γ′q = −γq and β′q = −βq in

Eq. (D1), and noting that C, B, and |+〉 are real-valued,

we have C(−~γ′,−~β′) = (C(~γ, ~β))† = C(~γ, ~β). Therefore,the cost landscape is point-symmetric with respect to theorigin, which implies that it is also point-symmetric withrespect to the center point of a period. When measuringa landscape, we can thus restrict either β or γ to half aperiod without losing information about the landscape.In the examples shown in this paper, we restrict γ to halfa period, i.e. to the interval [0, π[.

Finally, when choosing γ′q = −γq and β′q = βq inEq. (D1), we obtain

− C(−~γ, ~β) = 〈+|U ′(~γ, ~β)†(−C)U ′(~γ, ~β) |+〉 (D2)

where U ′(~γ, ~β) =∏q e−iβqBe−iγq(−C). This is equivalent

to the QAOA cost function for a problem HamiltonianC ′ = −C. Thus, in cases for which running QAOA with Cand with −C leads to the same landscape, the landscapeis an odd function of ~γ. In particular, this occurs for bothproblem instances considered in this work.

Due to the point-symmetry observed above, the land-

scape is also an odd function of ~β if it is an odd functionof ~γ. In the landscape plots for p = 1, this manifestsas line symmetries (with a change of the sign of the en-ergy) about both coordinate axes and with respect to thecenter line of each period. Within the chosen range ofβ, we observe this type of symmetry with respect to thehorizontal line β = π

4 .

Appendix E: Master-Equation Simulations

We model the dynamics of our system by a master-equation given by

ρ = − i

h[H(t), ρ] +

∑k

[ckρc

†k −

1

2

(c†k ckρ+ ρc†k ck

)],

(E1)

where ρ is the density matrix describing the system attime t and H(t) is the Hamiltonian, the time-dependenceof which models the applied gate sequence. The collapseoperators ck model incoherent processes. We solve themaster equation numerically [47] in the rotating frameof qubits. Incoherent errors are described by Lindbladterms in Eq. (E1) with

cT1,i =

√1

T1,iσ−,i, (E2)

cTφ,i =

√1

2

( 1

T2,i− 1

2T1,i

)σz,i, (E3)

where T1,i and T2,i are the lifetime and decoherence time(Ramsey decay time) of qubit i as listed in Table I.

In addition to the incoherent errors introduced by theLindblad terms, it is important to also consider the impactof coherent errors on the algorithm. In our experiment,the main source of coherent errors is residual ZZ-couplingbetween neighboring qubits [37]. To model this couplingin the numerical simulations, we include the Hamiltonian

HZZ/h =∑(i,j)

αij |11〉ij 〈11| (E4)

where the sum is over connected pairs of qubits with theresidual ZZ-couplings listed in Appendix A. We noticefrom simulations of the full QAOA circuit that the residualZZ-couplings give rise to the distortions observed in thecost landscapes, see Fig. 9. The main effect of decoherenceis to reduce the overall contrast of the landscape. In par-ticular, for the direct implementation we find a minimum

10

0

¼/4

¼/2M

ixin

g an

gle,

¯(a) (b) (c)

0 ¼/2 ¼0

¼/4

¼/2 (d)

0 ¼/2 ¼Separation angle, °

(e)

0 ¼/2 ¼

(f)-1

0

1

Cost

, C(°;¯)

ZZ only ZZ+decoh. Exp.

FIG. 9. Simulated and experimental cost-function landscapesof the three-qubit problem instance for p = 1. The direct im-plementation is displayed in (a), (b) and (c). The decomposedversion is shown in (d), (e) and (f). The master-equationsimulations are performed including errors from residual ZZ-coupling only (a, d), and with both residual ZZ-coupling anddecoherence (b, e). The experimental data is shown in (c) and(f).

energy of −1.04, −0.99 and −0.98 for simulations includ-ing only residual ZZ-couplings, for simulations includingboth residual ZZ-couplings and decoherence, and for ex-periments, respectively. In comparison, the minimumenergy in noise-free simulations is −1.06, see Fig. 3(b).

Appendix F: Seven-qubit problem instance

We use a single-layer QAOA circuit with C-ARB gatesto measure the cost-function landscape of the seven-qubitproblem instance, see Appendix G for the full pulse se-quence. The measured landscape, see Fig. 10(a), is ingood qualitative agreement with noise-free simulations,see Fig. 10(b). Due to decoherence, the absolute values ofthe global extrema are smaller than in noise-free simula-tions, see Fig. 10(c). Starting from random initialization,the convergence traces of the separating angle, the mix-ing angle and the corresponding cost are displayed inFig. 10(d), (e) and (f) respectively.

The output state distribution at optimal parametersfor the direct and decomposed implementation of C-ARBgates are shown in Fig. 11(a) and (b), respectively. Wedisplay the distributions yielding highest success proba-bility for each implementation, i.e. p = 2 for the directimplementation and p = 1 for the decomposed imple-mentation. For the direct implementation, the two mostlikely measured states are |1111000〉 and |0000111〉, corre-sponding to the respective selections of subsets A and Bforming exact covers of the considered problem instance.Conversely, the solution states are not the two most likelymeasured states for the decomposed implementation. Forboth implementations, the measured data matches wellwith expectations from master-equation simulations (redwire-frame).

Appendix G: QAOA gate sequences

Fig. 12(a) shows the pulse sequence generated by theAWGs for a single layer of QAOA in the seven-qubit in-stance using the direct implementation of C-ARB gates.Since the length of the flux pulses depends on the re-quired phase, see Fig. 1(c) and (d), the pulse sequenceis shown for a representative flux pulse length (close tothe average) that we obtain for γ = π

5 . After an initial

π/2-pulse on each qubit to prepare a |+〉⊗7state, the

phase-separation operator UC of the first QAOA layerstarts with two parallel C-ARB gates corresponding tothe couplings JA3B1 and JA2B2 , while qubit B3 is de-tuned by an additional flux pulse to avoid an unwantedinteraction when the ef-transition frequency of A2 crossesthe parking frequency of qubit B3. Since the additionalZΓij rotations, see Fig. 1(a), are implemented as virtualgates [31] through a redefinition of the reference frame,they are not shown in the pulse sequence. After the lastround of flux pulses, the final two π/2-pulses for eachqubit (plus a virtual gate between them) implement themixing operator UB = e−iβB , where we have decomposedeach term e−iβXi as shown in Fig. 1(a). After the end ofthe shown pulse sequence, we perform qubit readout.

In the significantly longer pulse sequence shown inFig. 12(b), the controlled arbitrary phase gates are de-composed as described in Fig. 1(b). Each Hadamardgate is implemented by a π/2-pulse and a Zπ rotation

0

¼/4

¼/2

Mixi

ng a

ngle

, ¯ (a) ¯ 00

¯ 0

(b)

0

¼/4

¼/2

Mixi

ng a

ngle

, ¯(e)

0 ¼/2 ¼ Separation angle, °

-1

0

1

Cost

, C(°;¯) (c)

¯ 00 = 38¼

¯ 0 = 18¼

0 ¼/2 ¼0

20

40

fe

(d)

0 20 40Func. eval., fe

-1

0

1

Cost

, C(°;¯)(f)

-1 0 1Cost, C(°; ¯)

FIG. 10. Cost function evaluated for p = 1 on the seven-qubitproblem instance, using C-ARB gates. (a) Cost-function land-scape as a function of variational parameters measured withdirect implementation of C-ARB gates. (b) Cost-function land-scape obtained from noise-free simulations. (c) Experimentalevaluation (blue) and simulation (black) of the cost functionfor two horizontal line cuts of (a) and (b), with β′ = π/8(dotted lines) and β′′ = 3π/8 (dashed lines), respectively. (d,e)10 convergence traces of the separation angle and the mixingangle, respectively, for end-to-end optimization starting fromrandom parameter initialization. (f) Average energy (solidblue line) and individual convergence traces (faded lines) ofthe energy corresponding to parameters shown in (d,e).

11

10-2

10-1

Stat

e pr

ob., Pjî p=2

j0000000®

j0000111®

j0001110®

j0010101®

j0011100®

j0100011®

j0101010®

j0110001®

j0111000®

j0111111®

j1000111®

j1001110®

j1010101®

j1011100®

j1100011®

j1101010®

j1110001®

j1111000®

j1111111®

State, jA1A2A3A4B1B2B3®

10-2

10-1

Stat

e pr

ob., Pjî p=1

(a)

(b)

FIG. 11. Output state probability distribution for the seven-qubit problem instance implemented with C-ARB gates (a) anddecomposed using CZ-gates (b). States are measured at optimal parameters for depth of p = 2 for (a) and p = 1 for (b). Thefilled bars correspond to the measured problem solutions, while the black (red) wire-frames are the expected QAOA outcomefrom noise-free (master-equation) simulations.

via a virtual gate, and the ZΓij in the center of the gatedecomposition is another virtual gate. Pulse sequencesfor the direct and the decomposed implementation of thethree-qubit problem instance are shown in Fig. 12(c) and(d), where analogous explanations apply.

To implement additional layers, the pulses betweenthe end of the initialization pulses and the start of thereadout are repeated p− 1 times. For the configurationsconsidered in Fig. 5, this leads to the gate counts shownin Table II.

TABLE II. Number of two-qubit gates (first row), single-qubit gates (second row), and virtual gates (third row) in theimplemented QAOA sequences.

# of layers, p 1 2 3 4 5 6 7 8 9

3 qb direct297

41514

62121

82728

103335

123942

144549

165156

185763

3 qb decomposed41713

83126

124539

165952

207365

248778

2810191

32115104

36129117

7 qb direct72121

143542

214963

286384

3577105

4291

126

7 qb decomposed144942

289184

42133126

56175168

Time, t ( s)

(b)

A1

A2

B1

B2

B3

A3

A4

(a)

A1

A2

B1

B2

B3

A3

A4

(c) A1

A2

B2

(d)A1

A2

B2

0.00 0.25 0.50 0.75 1.00

0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25

-halfpulsefluxpulse

FIG. 12. Pulse sequences for implementing QAOA with p = 1.The shaded area around flux pulses illustrates which interac-tion they implement and how long buffer times before andafter the flux pulse are chosen. (a) Direct implementation ofthe seven-qubit instance. (b) Decomposed implementation ofthe seven-qubit instance. (c) Direct implementation of thethree-qubit instance. (d) Decomposed implementation of thethree-qubit instance.

Appendix H: Post-selection

For both QAOA implementations, we discard all mea-sured states containing at least one leakage event. Weshow the percentage of single-shot measurements we keep

12

TABLE III. (First 4 rows) Percentage of data kept after dis-carding all measured states containing at least one leakageevent. (Bottom 4 rows) Corresponding average leakage pertwo-qubit gate in percent.

# of layers, p 1 2 3 4 5 6 7 8 9

3 qb direct 99.2 97.1 97.9 97.1 96.4 97.1 96.6 94.1 93.63 qb dec. 98.6 98.3 96.4 92.1 95.8 94.0 94.8 94.4 93.57 qb direct 87.1 87.7 80.6 55.9 79.2 77.57 qb dec. 80.2 82.4 61.6 62.13 qb direct 0.4 0.7 0.3 0.4 0.4 0.2 0.2 0.4 0.43 qb dec. 0.4 0.2 0.3 0.5 0.2 0.3 0.2 0.2 0.27 qb direct 1.9 0.9 1.0 2.1 0.7 0.67 qb dec. 1.6 0.7 1.1 0.8

as a function of the number of layers in the top half ofTable III. We estimate the corresponding average leakage

per gate as λ ≈ 1 − P 1/ng

post , where Ppost is the fractionof data left after post-selection and ng is the number oftwo-qubit gates in the sequence. All average leakage pergate values lie between 0.2% and 2.1%.

[1] P. W. Shor, “Algorithms for quantum computation: Dis-crete logarithms and factoring,” in Proceedings, 35th An-nual Symposium on Foundations of Computer Science,Santa Fe (IEEE Computer Society Press, 1994) p. 124.

[2] Yudong Cao, Jonathan Romero, Jonathan P. Olson,Matthias Degroote, Peter D. Johnson, Mria Kieferov,Ian D. Kivlichan, Tim Menke, Borja Peropadre, NicolasP. D. Sawaya, Sukin Sim, Libor Veis, and Aln Aspuru-Guzik, “Quantum chemistry in the age of quantum com-puting,” Chemical Reviews 119, 10856–10915 (2019),pMID: 31469277.

[3] Daniel A. Lidar and Todd A. Brun, Quantum Error Cor-rection (Cambridge University Press, 2013).

[4] John Preskill, “Quantum Computing in the NISQ era andbeyond,” Quantum 2, 79 (2018).

[5] Frank Arute, Kunal Arya, Ryan Babbush, Dave Bacon,Joseph C. Bardin, Rami Barends, Rupak Biswas, SergioBoixo, Fernando G. S. L. Brandao, David A. Buell, BrianBurkett, Yu Chen, Zijun Chen, Ben Chiaro, RobertoCollins, William Courtney, Andrew Dunsworth, Ed-ward Farhi, Brooks Foxen, Austin Fowler, Craig Gidney,Marissa Giustina, Rob Graff, Keith Guerin, Steve Habeg-ger, Matthew P. Harrigan, Michael J. Hartmann, AlanHo, Markus Hoffmann, Trent Huang, Travis S. Humble,Sergei V. Isakov, Evan Jeffrey, Zhang Jiang, Dvir Kafri,Kostyantyn Kechedzhi, Julian Kelly, Paul V. Klimov,Sergey Knysh, Alexander Korotkov, Fedor Kostritsa,David Landhuis, Mike Lindmark, Erik Lucero, DmitryLyakh, Salvatore Mandr, Jarrod R. McClean, MatthewMcEwen, Anthony Megrant, Xiao Mi, Kristel Michielsen,Masoud Mohseni, Josh Mutus, Ofer Naaman, MatthewNeeley, Charles Neill, Murphy Yuezhen Niu, Eric Os-tby, Andre Petukhov, John C. Platt, Chris Quintana,Eleanor G. Rieffel, Pedram Roushan, Nicholas C. Ru-bin, Daniel Sank, Kevin J. Satzinger, Vadim Smelyanskiy,Kevin J. Sung, Matthew D. Trevithick, Amit Vainsencher,Benjamin Villalonga, Theodore White, Z. Jamie Yao,Ping Yeh, Adam Zalcman, Hartmut Neven, and John M.Martinis, “Quantum supremacy using a programmablesuperconducting processor,” Nature 574, 505–510 (2019).

[6] J. Zhang, G. Pagano, P. W. Hess, A. Kyprianidis,P. Becker, H. Kaplan, A. V. Gorshkov, Z.-X. Gong, andC. Monroe, “Observation of a many-body dynamical phasetransition with a 53-qubit quantum simulator,” Nature551, 601 (2017).

[7] Hannes Bernien, Sylvain Schwartz, Alexander Keesling,Harry Levine, Ahmed Omran, Hannes Pichler, Soon-won Choi, Alexander S. Zibrov, Manuel Endres, MarkusGreiner, Vladan Vuletic, and Mikhail D. Lukin, “Probingmany-body dynamics on a 51-atom quantum simulator,”Nature 551, 579–584 (2017).

[8] P. J. J. O’Malley, R. Babbush, I. D. Kivlichan, J. Romero,J. R. McClean, R. Barends, J. Kelly, P. Roushan, A. Tran-ter, N. Ding, B. Campbell, Y. Chen, Z. Chen, B. Chiaro,A. Dunsworth, A. G. Fowler, E. Jeffrey, E. Lucero,A. Megrant, J. Y. Mutus, M. Neeley, C. Neill, C. Quin-tana, D. Sank, A. Vainsencher, J. Wenner, T. C. White,P. V. Coveney, P. J. Love, H. Neven, A. Aspuru-Guzik,and J. M. Martinis, “Scalable quantum simulation ofmolecular energies,” Phys. Rev. X 6, 031007 (2016).

[9] Abhinav Kandala, Antonio Mezzacapo, Kristan Temme,Maika Takita, Markus Brink, Jerry M. Chow, and Jay M.Gambetta, “Hardware-efficient variational quantum eigen-solver for small molecules and quantum magnets,” Nature549, 242–246 (2017).

[10] F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin,R. Barends, S. Boixo, M. Broughton, B. B. Buck-ley, D. A. Buell, B. Burkett, N. Bushnell, Y. Chen,Z. Chen, B. Chiaro, R. Collins, W. Courtney, S. De-mura, A. Dunsworth, E. Farhi, A. Fowler, B. Foxen,C. Gidney, M. Giustina, R. Graff, S. Habegger, M. P. Har-rigan, A. Ho, S. Hong, T. Huang, W. J. Huggins, L. Ioffe,S. V. Isakov, E. Jeffrey, Z. Jiang, C. Jones, D. Kafri,K. Kechedzhi, J. Kelly, S. Kim, P. V. Klimov, A. Ko-rotkov, F. Kostritsa, D. Landhuis, P. Laptev, M. Lind-mark, E. Lucero, O. Martin, J. M. Martinis, J. R. Mc-Clean, M. McEwen, A. Megrant, X. Mi, M. Mohseni,W. Mruczkiewicz, J. Mutus, O. Naaman, M. Neeley,C. Neill, H. Neven, M. Y. Niu, T. E. O’Brien, E. Ostby,A. Petukhov, H. Putterman, C. Quintana, P. Roushan,N. C. Rubin, D. Sank, K. J. Satzinger, V. Smelyan-skiy, D. Strain, K. J. Sung, M. Szalay, T. Y. Takeshita,A. Vainsencher, T. White, N. Wiebe, Z. J. Yao, P. Yeh,and A. Zalcman, “Hartree-fock on a superconductingqubit quantum computer,” arXiv:2004.04174 (2020).

[11] Edward Farhi, Jeffrey Goldstone, and Sam Gut-mann, “A quantum approximate optimization algorithm,”arXiv:1411.4028 (2014), 1411.4028v1.

[12] J. S. Otterbach, R. Manenti, N. Alidoust, A. Bestwick,M. Block, B. Bloom, S. Caldwell, N. Didier, E. Schuyler

13

Fried, S. Hong, P. Karalekas, C. B. Osborn, A. Papa-george, E. C. Peterson, G. Prawiroatmodjo, N. Rubin,Colm A. Ryan, D. Scarabelli, M. Scheer, E. A. Sete,P. Sivarajah, Robert S. Smith, A. Staley, N. Tezak, W. J.Zeng, A. Hudson, Blake R. Johnson, M. Reagor, M. P.da Silva, and C. Rigetti, “Unsupervised Machine Learn-ing on a Hybrid Quantum Computer,” arXiv:1712.05771[quant-ph] (2017), arXiv: 1712.05771.

[13] F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin,R. Barends, S. Boixo, M. Broughton, B. B. Buck-ley, D. A. Buell, B. Burkett, N. Bushnell, Y. Chen,Z. Chen, B. Chiaro, R. Collins, W. Courtney, S. Demura,A. Dunsworth, E. Farhi, A. Fowler, B. Foxen, C. Gid-ney, M. Giustina, R. Graff, S. Habegger, M. P. Harrigan,A. Ho, S. Hong, T. Huang, L. B. Ioffe, S. V. Isakov, E. Jef-frey, Z. Jiang, C. Jones, D. Kafri, K. Kechedzhi, J. Kelly,S. Kim, P. V. Klimov, A. N. Korotkov, F. Kostritsa,D. Landhuis, P. Laptev, M. Lindmark, M. Leib, E. Lucero,O. Martin, J. M. Martinis, J. R. McClean, M. McEwen,A. Megrant, X. Mi, M. Mohseni, W. Mruczkiewicz,J. Mutus, O. Naaman, M. Neeley, C. Neill, F. Neukart,H. Neven, M. Y. Niu, T. E. O’Brien, B. O’Gorman,E. Ostby, A. Petukhov, H. Putterman, C. Quintana,P. Roushan, N. C. Rubin, D. Sank, K. J. Satzinger,A. Skolik, V. Smelyanskiy, D. Strain, M. Streif, K. J.Sung, M. Szalay, A. Vainsencher, T. White, Z. J. Yao,P. Yeh, A. Zalcman, and L. Zhou, “Quantum approximateoptimization of non-planar graph problems on a planarsuperconducting processor,” arXiv:2004.04197 (2020).

[14] Leo Zhou, Sheng-Tao Wang, Soonwon Choi, Hannes Pich-ler, and Mikhail D. Lukin, “Quantum approximate op-timization algorithm: Performance, mechanism, and im-plementation on near-term devices,” arXiv:1812.01041(2018).

[15] S. Hadfield, Z. Wang, B. OGorman, E.G. Rieffel, D. Ven-turelli, and R. Biswas, “From the quantum approximateoptimization algorithm to a quantum alternating operatoransatz,” Algorithms 12, 34:1–34:45 (2019).

[16] Edward Farhi, Jeffrey Goldstone, Sam Gutmann, and LeoZhou, “The quantum approximate optimization algorithmand the Sherrington-Kirkpatrick model at infinite size,”arXiv:1910.08187 (2019).

[17] Guillaume Verdon, Juan Miguel Arrazola, KamilBradler, and Nathan Killoran, “A quantum approxi-mate optimization algorithm for continuous problems,”arXiv:1902.00409 (2019).

[18] Zhang Jiang, Eleanor G. Rieffel, and Zhihui Wang, “Near-optimal quantum circuit for Grover’s unstructured searchusing a transverse field,” Phys. Rev. A 95, 062317 (2017).

[19] M. Alam, A. Ash-Saki, and S. Ghosh, “Analysis of quan-tum approximate optimization algorithm under realis-tic noise in superconducting qubits,” arXiv:1907.09631(2019).

[20] Toshiki Matsumine, Toshiaki Koike-Akino, and Ye Wang,“Channel decoding with quantum approximate optimiza-tion algorithm,” Proc. 2019 IEEE Int. Symp. Inf. Theory(ISIT) , 2574–2578 (2019).

[21] A. Bengtsson, P. Vikstal, C. Warren, M. Svensson, X. Gu,A. F. Kockum, P. Krantz, C. Krizan, D. Shiri, I. Svens-son, G. Tancredi, G. Johansson, P. Delsing, G. Ferrini,and J. Bylander, “Quantum approximate optimization ofthe exact-cover problem on a superconducting quantumprocessor,” arXiv:1912.10495 (2019).

[22] Xiaogang Qiang, Xiaoqi Zhou, Jianwei Wang, Callum M.

Wilkes, Thomas Loke, Sean OGara, Laurent Kling, Gra-ham D. Marshall, Raffaele Santagati, Timothy C. Ralph,Jingbo B. Wang, Jeremy L. OBrien, Mark G. Thomp-son, and Jonathan C. F. Matthews, “Large-scale siliconquantum photonics implementing arbitrary two-qubit pro-cessing,” Nature Photonics 12, 534 (2018).

[23] G. Pagano, A. Bapat, P. Becker, K. S. Collins, A. De,P. W. Hess, H. B. Kaplan, A. Kyprianidis, W. L. Tan,C. Baldwin, L. T. Brady, A. Deshpande, F. Liu, S. Jor-dan, A. V. Gorshkov, and C. Monroe, “Quantum approxi-mate optimization with a trapped-ion quantum simulator,”arXiv:1906.02700 (2019).

[24] R. Barends, L. Lamata, J. Kelly, L. Garcia-Alvarez, A. G.Fowler, A Megrant, E Jeffrey, T. C. White, D. Sank,J. Y. Mutus, B. Campbell, Yu Chen, Z. Chen, B. Chiaro,A. Dunsworth, I.-C. Hoi, C. Neill, P. J. J. O/’Malley,C. Quintana, P. Roushan, A. Vainsencher, J. Wenner,E. Solano, and John M. Martinis, “Digital quantumsimulation of fermionic models with a superconductingcircuit,” Nat Commun 6, – (2015).

[25] P. Roushan, C. Neill, A. Megrant, Y. Chen, R. Bab-bush, R. Barends, B. Campbell, Z. Chen, B. Chiaro,A. Dunsworth, A. Fowler, E. Jeffrey, J. Kelly, E. Lucero,J. Mutus, P. J. J. O[rsquor]Malley, M. Neeley, C. Quin-tana, D. Sank, A. Vainsencher, J. Wenner, T. White,E. Kapit, H. Neven, and J. Martinis, “Chiral ground-statecurrents of interacting photons in a synthetic magneticfield,” Nat Phys advance online publication, – (2016).

[26] M. Ganzhorn, D.J. Egger, P. Barkoutsos, P. Ollitrault,G. Salis, N. Moll, M. Roth, A. Fuhrer, P. Mueller, S. Wo-erner, I. Tavernelli, and S. Filipp, “Gate-efficient simu-lation of molecular eigenstates on a quantum computer,”Phys. Rev. Applied 11, 044092 (2019).

[27] B. Foxen, C. Neill, A. Dunsworth, P. Roushan, B. Chiaro,A. Megrant, J. Kelly, Z. Chen, K. Satzinger, R. Barends,F. Arute, K. Arya, R. Babbush, D. Bacon, J. C.Bardin, S. Boixo, D. Buell, B. Burkett, Y. Chen,R. Collins, E. Farhi, A. Fowler, C. Gidney, M. Giustina,R. Graff, M. Harrigan, T. Huang, S. V. Isakov, E. Jeffrey,Z. Jiang, D. Kafri, K. Kechedzhi, P. Klimov, A. Ko-rotkov, F. Kostritsa, D. Landhuis, E. Lucero, J. McClean,M. McEwen, X. Mi, M. Mohseni, J. Y. Mutus, O. Naaman,M. Neeley, M. Niu, A. Petukhov, C. Quintana, N. Rubin,D. Sank, V. Smelyanskiy, A. Vainsencher, T. C. White,Z. Yao, P. Yeh, A. Zalcman, H. Neven, and J. M. Marti-nis, “Demonstrating a continuous set of two-qubit gatesfor near-term quantum algorithms,” arXiv:2001.08343(2020).

[28] D. M. Abrams, N. Didier, B. R. Johnson, M. P.da Silva, and C. A. Ryan, “Implementation of thexy interaction family with calibration of a single pulse,”arXiv:1912.04424 (2019).

[29] Richard M. Karp, “Reducibility among combinatorialproblems,” in Complexity of Computer Computations:Proceedings of a symposium on the Complexity of Com-puter Computations, held March 20–22, 1972, at the IBMThomas J. Watson Research Center, Yorktown Heights,New York, and sponsored by the Office of Naval Research,Mathematics Program, IBM World Trade Corporation,and the IBM Research Mathematical Sciences Department ,edited by Raymond E. Miller, James W. Thatcher, andJean D. Bohlinger (Springer US, Boston, MA, 1972) pp.85–103.

[30] Andrew Lucas, “Ising formulations of many NP problems,”

14

Front. Phys. 2, 5:1–5:15 (2014).[31] D. C. McKay, C. J. Wood, S. Sheldon, J. M. Chow, and

J. M. Gambetta, “Efficient z-gates for quantum comput-ing,” arXiv:1612.00858 (2016).

[32] Frederick W. Strauch, Philip R. Johnson, Alex J. Dragt,C. J. Lobb, J. R. Anderson, and F. C. Wellstood, “Quan-tum logic gates for coupled superconducting phase qubits,”Phys. Rev. Lett. 91, 167005 (2003).

[33] L. DiCarlo, J. M. Chow, J. M. Gambetta, Lev S. Bishop,B. R. Johnson, D. I. Schuster, J. Majer, A. Blais, L. Frun-zio, S. M. Girvin, and R. J. Schoelkopf, “Demonstrationof two-qubit algorithms with a superconducting quantumprocessor,” Nature 460, 240–244 (2009).

[34] R. Barends, J. Kelly, A. Megrant, A. Veitia, D. Sank,E. Jeffrey, T. C. White, J. Mutus, A. G. Fowler, B. Camp-bell, Y. Chen, Z. Chen, B. Chiaro, A. Dunsworth, C. Neill,P. OMalley, P. Roushan, A. Vainsencher, J. Wenner, A. N.Korotkov, A. N. Cleland, and John M. Martinis, “Super-conducting quantum circuits at the surface code thresholdfor fault tolerance,” Nature 508, 500–503 (2014).

[35] P. Viksttal, M. Grnkvist, M. Svensson, M. Andersson,G. Johansson, and G. Ferrini, “Applying the quantumapproximate optimization algorithm to the tail assignmentproblem,” arXiv:1912.10499 (2019).

[36] Cheng Xue, Zhao-Yun Chen, Yu-Chun Wu, and Guo-PingGuo, “Effects of quantum noise on quantum approximateoptimization algorithm,” Arxiv (2019), 1909.02196v2.

[37] S. Krinner, S. Lazar, A. Remm, C. K. Andersen,N. Lacroix, G. J. Norris, C. Hellings, M. Gabureac,C. Eichler, and A. Wallraff, “Benchmarking coherenterrors in controlled-phase gates due to spectator qubits,”(in preparation).

[38] A. Bhattacharyya, “On a measure of divergence betweentwo statistical populations defined by their probability dis-tribution,” Bulletin of the Calcutta Mathematical Society35, 99110 (1943).

[39] Sergey Bravyi, Alexander Kliesch, Robert Koenig, andEugene Tang, “Obstacles to state preparation andvariational optimization from symmetry protection,”arXiv:1910.08980 (2019).

[40] E. Farhi, D. Gamarnik, and S. Gutmann, “The quan-tum approximate optimization algorithm needs to see thewhole graph: A typical case,” arXiv:2004.09002 (2020).

[41] S. Krinner, S. Storz, P. Kurpiers, P. Magnard, J. Hein-soo, R. Keller, J. Lutolf, C. Eichler, and A. Wallraff,“Engineering cryogenic setups for 100-qubit scale super-conducting circuit systems,” EPJ Quantum Technology6, 2 (2019).

[42] Christian Kraglund Andersen, Ants Remm, StefaniaLazar, Sebastian Krinner, Nathan Lacroix, Graham J.Norris, Mihai Gabureac, Christopher Eichler, and An-dreas Wallraff, “Repeated quantum error detection in asurface code,” arXiv:1912.09410 (2019).

[43] Johannes Heinsoo, Christian Kraglund Andersen, AntsRemm, Sebastian Krinner, Theodore Walter, YvesSalathe, Simone Gasparinetti, Jean-Claude Besse, An-ton Potocnik, Andreas Wallraff, and Christopher Eichler,“Rapid high-fidelity multiplexed readout of superconduct-ing qubits,” Phys. Rev. Applied 10, 034040 (2018).

[44] C. Macklin, K. O’Brien, D. Hover, M. E. Schwartz,V. Bolkhovsky, X. Zhang, W. D. Oliver, and I. Sid-diqi, “A near-quantum-limited Josephson traveling-waveparametric amplifier,” Science 350, 307–310 (2015).

[45] Michele C. Collodo, Johannes Herrmann, Nathan Lacroix,

Christian Kraglund Andersen, Ants Remm, StefaniaLazar, Jean-Claude Besse, Theo Walter, Andreas Wall-raff, and Christopher Eichler, “Implementation ofa conditional-phase gate by using in-situ tunable zz-interactions,” (in preparation).

[46] Vicky Choi, “Adiabatic quantum algorithms for the NP-complete maximum-weight independent set, exact coverand 3SAT problems,” arXiv:1004.2226 (2010).

[47] J. R. Johansson, P. D. Nation, and Franco Nori, “QuTiP2: A Python framework for the dynamics of open quan-tum systems,” Comput. Phys. Commun. 184, 1234–1240(2013).