High-Fidelity Qutrit Entangling Gates for Superconducting Circuits

Noah Goss,1, 2, ∗ Alexis Morvan,2 Brian Marinelli,1, 2 Bradley K. Mitchell,1 Long B. Nguyen,2 Ravi K. Naik,1, 2
Larry Chen,1 Christian Jünger,2 John Mark Kreikebaum,1, 3
David I. Santiago,2 Joel J. Wallman,4 and Irfan Siddiqi1, 2, 3
1
Department of Physics, University of California, Berkeley, Berkeley CA 94720, USA.
2
Computational Research Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720, USA.
3
Materials Science Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720, USA.
4
Keysight Technologies Canada, Kanata, ON K2K 2W5, Canada.
(Dated: June 17, 2022)
arXiv:2206.07216v2 [quant-ph] 16 Jun 2022

Ternary quantum information processing in superconducting devices poses a promising alternative

to its more popular binary counterpart through larger, more connected computational spaces and
proposed advantages in quantum simulation and error correction. Although generally operated
as qubits, transmons have readily addressable higher levels, making them natural candidates for
operation as quantum three-level systems (qutrits). Recent works in transmon devices have realized
high fidelity single qutrit operation. Nonetheless, effectively engineering a high-fidelity two-qutrit
entanglement remains a central challenge for realizing qutrit processing in a transmon device. In
this work, we apply the differential AC Stark shift to implement a flexible, microwave-activated, and
dynamic cross-Kerr entanglement between two fixed-frequency transmon qutrits, expanding on work
performed for the ZZ interaction with transmon qubits. We then use this interaction to engineer
efficient, high-fidelity qutrit CZ† and CZ gates, with estimated process fidelities of 97.3(1)% and
95.2(3)% respectively, a significant step forward for operating qutrits on a multi-transmon device.

Quantum error correction (QEC) [1] is necessary for gates remains a major challenge in superconducting cir-
noisy intermediate-scale quantum (NISQ) [2] computers cuits.
to realize their full potential. The surface code [3, 4] us- The most commonly used qubit in superconducting cir-
ing qubits is considered the main route to fault toler- cuits [29], the transmon [30], is well suited to be oper-
ance [5–8], though its technical challenges have led the ated as a qutrit due to its weak anharmonicity. Techni-
community to explore other approaches that could have cal advancements in microwave engineering and improved
more favorable QEC schemes, such as storing a two level fabrication techniques have increased transmon coherence
system in the large Hilbert space of quantum oscillators times [31], enabling coherent control of the full qutrit
[9, 10]. Another alternative is to use d-dimensional quan- Hilbert space. Furthermore, dispersive readout can be
tum objects, or qudits, which mobilize a larger and more used for high-fidelity single shot qutrit readout [27]. In
connected computational space than their qubit counter- addition, high-fidelity single qutrit operations [32, 33],
parts. Qutrits, the simplest form of qudits, can provide quantum information scrambling [28], compact decompo-
advantages in QEC for magic state distillation [11, 12], sitions of multi-qubit gates [34–37], and improved qubit
compactly encoding qubits [13], and can be used to en- readout [38] have all been demonstrated using transmons
code logical qutrits themselves [14–16]. There are several as qutrits. Nonetheless, past qutrit entangling gates in
proposals utilizing qutrits to improve quantum algorithms transmons have been limited by relying on either a slow,
[17–20] and in the short term, qutrits can improve quan- static interaction or an interaction that restricts the en-
tum simulation [21] and other NISQ applications [22]. Re- tanglement to only a subspace of the qutrit.
alizing multi-qudit systems, however, is challenging due to In this work, we characterize the differential AC Stark
the complexities of the larger Hilbert space. Nonetheless, shift [39–42] on two fixed frequency transmon qutrits
coherent control of qudits has been performed in several with static coupling and leverage it to generate dynamic
physical platforms [23–28]. While state of the art exper- qutrit entangling phases. With this interaction, we en-
iments have demonstrated high-fidelity qudit entangling gineer the ternary controlled-Z gate (CZ) and its inverse
gates with trapped ions [23, 24] and photonic circuits [26], (CZ† ). Both gates performed in our work are univer-
generating high-fidelity, maximally entangling two-qudit sal for ternary computation and Clifford gates needed for
QEC in qutrits. We achieve an estimated process fidelity
of 97.3(1)% and 95.2(3)% for the CZ† and CZ respec-
tively, measured using cycle benchmarking [43] and our
∗ Correspondence should be addressed to [email protected] generalization of the cross-entropy benchmarking routine
2

where each term can be be calculated with perturbation

theory (see supplementary note 2). In this microwave-
activated case, the αij (ZZ-like) terms are given by:
αij = Aij (ωd )Ωa Ωb cos(φa − φb ), (2)
where Ωi and φi are respectively the amplitude and ph-
sae of the drive on transmon i. The coefficients Aij are
functions of the proximity of the microwave drive fre-
quency (ωd ) to nearby transitions. We note that this
Hamiltonian generates entanglement between the entire
two-qutrit Hilbert space, contrary to the CR case, where
the entanglement is mostly restrained to a subspace of the
qutrit [28, 48]. This dynamic, driven cross-Kerr interac-
tion is depicted schematically in Fig. 1. It is important to
note that the number of degrees of freedom in this inter-
action are not sufficient, in general, to realize a Clifford
two-qutrit gate like the CZ gate with a single round of
Figure 1. Microwave-activated cross-Kerr entangle- cross-Kerr entanglement, a difficulty discussed in further
ment. Two transmon qutrits with qubit frequency ωi , an-
detail in the next section.
harmonicity ηi , and coupling J, experience a dynamical cross-
Kerr (ZZ-like) entanglement when simultaneously driven by
Measuring the ZZ interaction in the qubit case
an off-resonant microwave drive. The strength of the cross- can be performed by a conditional Ramsey experi-
Kerr entangling terms (α11 , α12 , α21 , α22 ) is tuned by the pa- ment or through a dynamically decoupled JAZZ (Joint-
rameters of the microwave drive (ωd , Ω, φ). Amplification-of-ZZ) sequence that removes the low fre-
quency drift [49, 50]. In the larger Hilbert space of
two qutrits, we need to measure four of these entangling
[44]. The fidelity of the CZ† represents a factor of 4 re- phases with a rate of accumulation set by αij in Eq. 2.
duction in infidelity over previous two qutrit transmon To simplify the measurement and reduce the number of
gates [28, 32]. Finally, we numerically demonstrate that experiments needed, we generalize the controlled-Ramsey
our gate scheme is efficient for generating additional two- experiment to the full qutrit space with a pulse sequence
qutrit Clifford gates. presented in Fig. 2a. In this sequence, we apply simulta-
neous ternary Hadamard gates on both qutrits, execute
Differential AC Stark Shift the microwave drive, and subsequently perform the full
Recent works by Refs. [39–42] demonstrated that the two qutrit state tomography. Doing so for several dura-
architecture employed in the Cross-resonance (CR) en- tions of the Stark driving allow us to fully characterize
tangling gate can also realize a two-qubit CZ gate by the conditional and unconditional Stark shifts.
leveraging the conditional Stark shifts from simultane- We demonstrate in Fig. 2b the result of this measure-
ously driving a pair of coupled qubits off-resonantly. The ment scheme: the entangling phases increase linearly with
advantages of this method are two-fold: firstly, the CZ the duration of the Stark drive, where the proportional-
gate commutes with ZZ errors from the always-on dis- ity constant is set by αij as predicted by our cross-Kerr
persive coupling between the transmons. Secondly, unlike model in Eq. 1. In Fig. 2c-d, we present how the driven
in the CR gate, the frequency of the microwave drive can cross-Kerr interaction depends on the parameters of our
take a range of values. This flexibility in drive frequency entanglement scheme, specifically the phase and the am-
affords significant advantages in avoiding frequency col- plitude of the Stark drive. We note that the qualitative
lisions with other transmons or spurious two-levels sys- behavior is properly captured by the perturbation the-
tems [47]. The generalization of conditional Stark shifts ory in Eq. 2. We also explore the behavior as a function
to qutrits is straightforward. Working in the energy eigen- of the drive frequency in Fig. 2e. In this case, the per-
basis of our two qutrit Hilbert space, up to single-qutrit turbation theory fails, but an ab-initio master equation
phases, one’s system evolves under the cross-Kerr Hamil- simulation captures some of the response; additional de-
tonian: tails on the frequency dependence of all αij terms and
the master equation simulation can be found in the sup-
H = α11 |11i h11| + α12 |12i h12| plement. We expect the unaccounted features can be at-
(1)
+α21 |21i h21| + α22 |22i h22| , tributed to higher-order terms, frequency dependent clas-
 3

a b
Stark Drive Duration
Hadamard Tomography

ℋ(𝜔! , 𝜙, Ω)
Q ! |0⟩
%$ (% %$ (%
𝑋"/$ 𝑋"/$ 𝑋"/$ 𝑋"/$
Stark Drive Duration
Hadamard Tomography

ℋ(𝜔! , 𝜙, Ω)
Q " |0⟩
%$ (% %$ (%
𝑋"/$ 𝑋"/$ 𝑋"/$ 𝑋"/$

c d e

𝜔%$,'" /$"
𝜔%$,'! /$"

𝜔(%,'! /$"

𝜔(%,'" /$"
Figure 2. Characterizing the dynamical cross-Kerr entanglement. a, To study the accumulation of entangling phases
under the driven cross-Kerr interaction, we place two qutrits in a full superposition using ternary Hadamard gates (virtual Z
gates ommited in diagram), then study the evolution under the Stark drive scheme by performing state tomography. b, We
demonstrate fitting the accumulation of entangling phase found by tomography to our linear, driven cross-Kerr model, where
αij is the slope of the line and the uncertainty is from the linear fit. c-d, We match the behavior of the cross-Kerr entanglement
given relevant experimental parameters in our system to our Hamiltonian model for the relative phase of the driving, φ, and
amplitude of the driving, fixing Ω = Ωa = Ωb . e, We additionally compare the dependence of α12 on the frequency of the drive
ωd using an ab-initio master equation simulation in QuTiP [45, 46].

sical crosstalk, or parasitic two level systems (TLS) in and robust gate scheme and this “fine-tuned” approach,
our device. In an experimental setting, the flexibility of while still taking advantage of the dynamical nature of our
this entanglement allows us the freedom to set the drive cross-Kerr interaction. By employing the pulse scheme in
frequency far from any of these features. Fig. 3, where echo pulses in the {|1i , |2i} subspace shuffle
Qutrit CZ/CZ† gate entangling phases, we have the modified unitary evolution
(omitting the single-qutrit phases for brevity):
We next construct qutrit controlled-phase gates utiliz-
ing this entangling interaction. The qutrit CZ and CZ† U = exp(−i[(α11 + α22 )τ (|11ih11| + |22ih22|)+
gate are both maximally entangling and members of the (4)
(α12 + α21 )τ (|12ih21| + |21ih21|)]),
two-qutrit Clifford group making them particularly useful
gates for ternary computation. The CZ gate is defined as: Tuning the experimental knobs demonstrated in Fig. 2,
X we are able to find the relaxed conditions on our cross-
UCZ = ω ij |ijihij| , (3) Kerr evolution: (α11 + α22 ) ≈ −(α12 + α21 ) and (α11 +
i,j∈{0,1,2}2 α22 ) ≈ 2(α12 + α21 ). Under these conditions, at some
drive time τ , we will have approximately acquired the
with ω = e2iπ/3 , the third root of unity. Under simultane- desired entangling phases found in Eq. 3 to synthesize re-
ous Stark drives, the two-qutrit Hilbert space follows the spectively a CZ or CZ† unitary. We find that when using
unitary evolution U = exp(−i(H+φ1 I ⊗Z 01 +φ2 I ⊗Z 12 + flat-top cosine pulses for the cross-Kerr drive, the ramp
φ3 Z 01 ⊗I +φ4 Z 12 ⊗I)τ ) where H is given in Eq. 1. To per- features can lead to offsets on the linear accumulation
form a CZ gate with a single round of cross-Kerr driving, of entangling phase; for this reason, the highest fidelity
for example, one would need to find driving parameters implementation of our gates only uses an approximate
meeting the conditions: {α11 = α22 = 2α21 = 2α12 }, relationship on the αij terms with some adjustment nec-
a task that is not broadly feasible. Practically speak- essary. Finally, as outlined schematically in Fig. 3, one
ing, we desire a compromise between the most general can undo the local Z-like phases (found via tomography)
 †

τ τ a d
×𝒎
ℋ(𝜔$ , 𝜙, Ω) ℋ(𝜔$ , 𝜙, Ω)
Q! Z!" 𝜙" |0⟩ B P P B†

CZ !
𝑋!"# 𝑋!"# Z"#(𝜙#)

τ τ |0⟩ B P P B†
CB
Q" ℋ(𝜔$ , 𝜙, Ω) ℋ(𝜔$ , 𝜙, Ω) Z!" 𝜙$
Z"#(𝜙%) b ×𝒎
𝑋!"# 𝑋!"# †
|0⟩ G G

CZ!
XEB
Figure 3. Gate schematic. For the CZ and CZ† gate, we |0⟩ G G
perform two rounds of cross-Kerr entanglement for duration τ XEB
with interleaved echo pulses in the {|1i , |2i} subspace which c
shuffle the entangling phases. For proper conditions on the αij
terms in Eq. 1, the CZ† (CZ) is compiled with a total gate time
of 580(783) ns. The local Z terms in both two level subspaces

Ref. Purity Limit

CZ Purity Limit
of the qutrit are then undone using virtual Z gates.

†
CB
in both the {|0i,|1i} and {|1i,|2i} subspaces with virtual
Z gates [51]. In this work, we performed the CZ and CZ†
on two different pairs of transmon qutrits, demonstrating
the flexible nature of generating two-qutrit gates from this
driven cross-Kerr scheme. Figure 4. Benchmarking. a, Circuit schematic of cycle
benchmarking (CB). The errors of the CZ† are twirled via
Benchmarking random Weyl gates (red) to tailor errors into stochastic Weyl
channels. The initial state and measurement basis (blue) are
We first benchmark our two-qutrit gates with cycle selected to pick out the decay associated with specific Weyl
benchmarking (CB) [43] using True-Q [52]. While origi- channels. b, Circuit schematic of cross-entropy benchmarking
nally written in terms of qubits, CB naturally generalizes (XEB). The errors of the CZ† are twirled via random SU(3)
to qutrits [32]. We use CB instead of, e.g., interleaved gates (green) to tailor the noise to a simple depolarizing chan-
randomized benchmarking [53, 54], because it requires nel. c, An integrated histogram of CB for both the CZ† gate
significantly fewer multi-qutrit gates per circuit. We de- and a reference cycle, with the solid vertical lines giving the
scribe the generalization in the supplementary material. fidelities 0.936(1) and 0.966(1) respectively, yielding an esti-
mated process fidelity of 97.3(1)%. We extract an error bud-
With this technique, we estimate the Weyl (generalized-
get directly from CB, estimating a purity limited fidelity of
Pauli) error rate of the CZ† and CZ gate to be 2.7(1)% 0.973(9) and 0.989 (with negligible error) for the dressed CZ†
and 4.8(3)% respectively. By contrast, the highest fidelity, and reference cycles, yielding a purity limit 0.986(9) for the
two-qutrit gate performed previously with transmons had isolated CZ† gate. d, From XEB we estimate the depolarized
an error rate of 11.1% [28]. CB also allows us to construct fidelity as 0.933(3). Additionally, we estimate the speckle-
the Weyl-twirled error per channel of the unitary in Fig. 4. purity limited fidelity of the CZ† dressed with random SU(3)
This provides us with an estimate of the worst case sce- gates to be 0.961(3).
nario of less than 8% and demonstrates a relatively low
dispersion of our error channels.
As an added confirmation of the fidelity of the CZ† gate, Finally, we would like to be able to characterize what
we generalize the cross-entropy benchmarking (XEB) rou- fraction of the errors present in our two qutrit unitaries
tine [44, 55] to tailor all gate errors into a depolarizing are coherent on the time scale of multiple experiments,
channel, for work with qutrit unitaries. In the qutrit case, and thus could be removed by improved calibration. As
we find sufficient tailoring of our noise can be performed we show in supplementary note 6, under the depolariz-
by interleaving random SU(3) gates around our target ing unitary noise model, the variance of CB and XEB
gate. The circuit diagram for qutrit XEB is in Fig. 4b circuits both provide a robust method of estimating the
and the results can be found in Fig. 4e. We find that the purity limit [56, 57]. The corresponding estimates are
depolarized fidelity of the CZ† dressed with random SU(3) shown in Fig.4c with the CB estimate of 97.3(9)% exceed-
gates agrees with the estimate of the process fidelity from ing the speckle-purity limit of 96.1(3)% for the dressed
our Weyl twirled CB results within a standard error. Ad- CZ† gate. This disagreement is likely due to the fact
ditional discussion of the qutrit XEB method is provided that the CB data reveals that the noise is dominated by
in the supplementary materials. single-qutrit phase errors. As these errors are likely to
 5

a ×𝒎 c gates. The gate synthesis is performed by optimizing the

ansatz parameters to minimize the distance between V
2Q gate

R1 Ri
and U, i.e. the infidelity 1-F(V, U).
R2 Rj
We perform this numerical investigation on 1000 Haar
Ansatz (V)
random gates and 1000 Clifford gates. We find that all
1000 Haar random gates can be synthesized at depth 6
for CZ/CZ† , 7 for Cinc , and 9 for Cex . The synthesis
U success rate for all 3 unitaries in terms of target Clifford
Target (U) circuits are shown in Fig. 4b. Notably, almost all two
b qutrit Clifford gates were successfully compiled at depth
2 in CZ/CZ† , with 100% success at depth 3. By contrast,
Haar Cinc Cex and Cinc did not demonstrate as much improvement

Haar Cex
Haar CZ

at synthesizing Clifford gates over Haar random gates,

achieving 100% success for target Clifford gates at depth
6 for Cinc and 9 for Cex . Additionally, unique amongst
these gates, the CZ/CZ† can generate maximally entan-
gled two qutrit states with a single iteration of the gate.
Ansatz Depth
We demonstrate the power of this feature in the experi-
mentally reconstructed qutrit Bell-state density matrix in
Figure 5. Demonstration of gate expressability. a, A Fig. 5c. In summary, the maximally entangling CZ/CZ†
parameterized Ansatz circuit (V) is used to synthesize a tar-
gates have low intrinsic errors and can also can synthesize
get unitary (U), given some 2-qutrit gate and arbitrary SU(3)
gates. b, We study the Ansatz circuit (V) for the different a very important family of gates for QEC (the Cliffords)
two-qutrit gates discussed in the text for 1000 Haar random [1] at much lower depths than the two-qutrit gates which
and Clifford gates, minimizing the infidelity 1 − F (V,U). The only entangle a subspace of the qutrit.
dashed lines represent 100% numerical success for synthesiz-
Discussion and outlook
ing our set of Haar random gates, and the bars display the
success rate for synthesizing Clifford gates. We perform the We realized a microwave-activated, dynamic cross-Kerr
minimization until we find a 100% success rate for each two- entangling interaction that can be employed to engineer
qutrit gate between depths 0 ≤ m ≤ 9. c, An experimen- qutrit entangling phases with high precision. Leveraging
tally reconstructed density matrix of the two qutrit Bell state
this interaction, we generated two maximally entangling
|ψi = √13 (|00i + |11i + |22i) formed using a single CZ gate
and high-fidelity two-qutrit gates on two separate pairs
with state fidelity F = 0.952. The black outline is the target
density matrix.
of fixed-frequency transmon qutrits. We demonstrated
numerically that these two qutrit gates are efficient for
producing additional two-qutrit unitaries, especially other
fluctuate around a mean, they will add dephasing errors Clifford gates. Future work developing a systematic gate
that are not captured by the depolarizing unitary model tune up procedure may prove essential in improving the
used to analyze XEB. Another possible contributing fac- fidelity and scalability of our approach. Additionally,
tor is that the noise fluctuated between the XEB and CB a study of the effects of this gate scheme on spectator
experiments, which were performed in separate batches. qutrits will also be necessary for determining its scala-
bility. We expect that by being maximally entangling
Gate synthesis of two qutrit unitaries and a member of the two-qutrit Clifford group, the gates
To study the expressibility of the two-qutrit gates performed in this work will prove especially powerful in
in this work, we numerically explore the ability of the future efforts to employ qutrits for QEC, quantum simu-
ternary CZ/CZ† (localy equivalent to each other and the lation, and quantum computation. Perhaps most impor-
CSUM gate) to synthesize other two-qutrit gates, and tantly, all of this work was performed on multi-transmon
compare them to two-qutrit entangling gates that only quantum processors which are normally used for qubit ex-
entangle a subspace of the qutrit, such as the controlled- periments; the untapped potential of transmons as qutrits
exchange (Cex ) and controlled-increment (Cinc ) gates per- is only beginning to be explored. As a final note, the two-
formed on a trapped ion system in Ref. [24]. To this end, qutrit Hilbert space is larger than even the three qubit
we consider an Ansatz circuit V as in Fig. 5a, with depth Hilbert space; as the community continues to explore qu-
m, which we use to synthesize target circuits belonging to dits, we propose that metrics and benchmarks should be
either the two qutrit Clifford group or set of Haar random developed to reasonably compare qudit vs. qubit gates.
6

[1] Girvin, S. M. Introduction to quantum error correction damping. IEEE Transactions on Information Theory 64,
and fault tolerance (2021). URL https://arxiv.org/ 4674–4685 (2018).
abs/2111.08894. [17] Bocharov, A., Roetteler, M. & Svore, K. M. Fac-
[2] Preskill, J. Quantum computing in the NISQ era and toring with qutrits: Shor’s algorithm on ternary and
beyond. Quantum 2, 79 (2018). metaplectic quantum architectures. Phys. Rev. A 96,
[3] Bravyi, S. B. & Kitaev, A. Y. Quantum codes on a lattice 012306 (2017). URL https://link.aps.org/doi/10.
with boundary (1998). URL https://arxiv.org/abs/ 1103/PhysRevA.96.012306.
quant-ph/9811052. [18] Gedik, Z. et al. Computational speed-up with a single
[4] Dennis, E., Kitaev, A., Landahl, A. & Preskill, J. qudit. Scientific Reports 5, 14671 (2015). URL https:
Topological quantum memory. Journal of Mathematical //doi.org/10.1038/srep14671.
Physics 43, 4452–4505 (2002). [19] Bullock, S. S., O’Leary, D. P. & Brennen, G. K. Asymp-
[5] Google Quantum AI. Exponential suppression of bit totically optimal quantum circuits for d-level systems.
or phase errors with cyclic error correction. Nature Phys. Rev. Lett. 94, 230502 (2005). URL https://link.
595, 383 (2021). URL https://doi.org/10.1038/ aps.org/doi/10.1103/PhysRevLett.94.230502.
s41586-021-03588-y. [20] Pavlidis, A. & Floratos, E. Quantum-fourier-transform-
[6] Marques, J. F. et al. Logical-qubit operations in based quantum arithmetic with qudits. Phys. Rev. A
an error-detecting surface code. Nature Physics 18, 103, 032417 (2021). URL https://link.aps.org/doi/
80–86 (2021). URL http://dx.doi.org/10.1038/ 10.1103/PhysRevA.103.032417.
s41567-021-01423-9. [21] Gustafson, E. Noise improvements in quantum simula-
[7] Krinner, S. et al. Realizing repeated quantum error cor- tions of sqed using qutrits (2022). URL https://arxiv.
rection in a distance-three surface code (2021). URL org/abs/2201.04546.
https://arxiv.org/abs/2112.03708. 2112.03708. [22] Gokhale, P. et al. Asymptotic improvements to quan-
[8] Zhao, Y. et al. Realization of an error-correcting surface tum circuits via qutrits. In Proceedings of the 46th In-
code with superconducting qubits (2022). URL https: ternational Symposium on Computer Architecture, ISCA
//arxiv.org/abs/2112.13505. 2112.13505. ’19, 554–566 (Association for Computing Machinery, New
[9] Campagne-Ibarcq, P. et al. Quantum error correction of York, NY, USA, 2019). URL https://doi.org/10.1145/
a qubit encoded in grid states of an oscillator. Nature 3307650.3322253.
584, 368–372 (2020). URL https://doi.org/10.1038/ [23] Hrmo, P. et al. Native qudit entanglement in a trapped
s41586-020-2603-3. ion quantum processor (2022). URL https://arxiv.org/
[10] Grimm, A. et al. Stabilization and operation of a Kerr- abs/2206.04104.
cat qubit. Nature 584, 205–209 (2020). URL http://dx. [24] Ringbauer, M. et al. A universal qudit quantum proces-
doi.org/10.1038/s41586-020-2587-z. sor with trapped ions. arXiv preprint arXiv:2109.06903
[11] Campbell, E. T., Anwar, H. & Browne, D. E. Magic-state (2021). URL https://arxiv.org/abs/2109.06903.
distillation in all prime dimensions using quantum reed- [25] Lanyon, B. P. et al. Manipulating biphotonic qutrits.
muller codes. Phys. Rev. X 2, 041021 (2012). URL https: Phys. Rev. Lett. 100, 060504 (2008). URL https://link.
//link.aps.org/doi/10.1103/PhysRevX.2.041021. aps.org/doi/10.1103/PhysRevLett.100.060504.
[12] Campbell, E. T. Enhanced fault-tolerant quantum [26] Chi, Y. et al. A programmable qudit-based quantum pro-
computing in d-level systems. Phys. Rev. Lett. 113, cessor. Nature Communications 13, 1166 (2022). URL
230501 (2014). URL https://link.aps.org/doi/10. https://doi.org/10.1038/s41467-022-28767-x.
1103/PhysRevLett.113.230501. [27] Bianchetti, R. et al. Control and tomography of a three
[13] Kapit, E. Hardware-efficient and fully autonomous quan- level superconducting artificial atom. Phys. Rev. Lett.
tum error correction in superconducting circuits. Phys. 105, 223601 (2010). URL https://link.aps.org/doi/
Rev. Lett. 116, 150501 (2016). URL https://link.aps. 10.1103/PhysRevLett.105.223601.
org/doi/10.1103/PhysRevLett.116.150501. [28] Blok, M. S. et al. Quantum information scrambling on
[14] Majumdar, R., Basu, S., Ghosh, S. & Sur-Kolay, S. Quan- a superconducting qutrit processor. Phys. Rev. X 11,
tum error-correcting code for ternary logic. Phys. Rev. A 021010 (2021). URL https://link.aps.org/doi/10.
97, 052302 (2018). URL https://link.aps.org/doi/10. 1103/PhysRevX.11.021010.
1103/PhysRevA.97.052302. [29] Blais, A., Grimsmo, A. L., Girvin, S. M. & Wallraff,
[15] Muralidharan, S., Zou, C.-L., Li, L., Wen, J. & Jiang, A. Circuit quantum electrodynamics. Rev. Mod. Phys.
L. Overcoming erasure errors with multilevel systems. 93, 025005 (2021). URL https://link.aps.org/doi/10.
New Journal of Physics 19, 013026 (2017). URL https: 1103/RevModPhys.93.025005.
//doi.org/10.1088/1367-2630/aa573a. [30] Koch, J. et al. Charge-insensitive qubit design de-
[16] Grassl, M., Kong, L., Wei, Z., Yin, Z.-Q. & Zeng, rived from the cooper pair box. Phys. Rev. A 76,
B. Quantum error-correcting codes for qudit amplitude 042319 (2007). URL https://link.aps.org/doi/10.
 7

1103/PhysRevA.76.042319. 1760–1772 (2012). URL https://www.sciencedirect.

[31] Siddiqi, I. Engineering high-coherence superconducting com/science/article/pii/S0010465512000835.
qubits. Nature Reviews Materials 6, 875–891 (2021). URL [46] Johansson, J., Nation, P. & Nori, F. Qutip 2: A
https://doi.org/10.1038/s41578-021-00370-4. python framework for the dynamics of open quantum sys-
[32] Morvan, A. et al. Qutrit randomized benchmarking. Phys. tems. Computer Physics Communications 184, 1234–
Rev. Lett. 126, 210504 (2021). URL https://link.aps. 1240 (2013). URL https://www.sciencedirect.com/
org/doi/10.1103/PhysRevLett.126.210504. science/article/pii/S0010465512003955.
[33] Yurtalan, M. A., Shi, J., Kononenko, M., Lupascu, A. [47] Morvan, A., Chen, L., Larson, J. M., Santiago, D. I.
& Ashhab, S. Implementation of a walsh-hadamard & Siddiqi, I. Optimizing frequency allocation for fixed-
gate in a superconducting qutrit. Phys. Rev. Lett. 125, frequency superconducting quantum processors. Phys.
180504 (2020). URL https://link.aps.org/doi/10. Rev. Research 4, 023079 (2022). URL https://link.
1103/PhysRevLett.125.180504. aps.org/doi/10.1103/PhysRevResearch.4.023079.
[34] Lanyon, B. P. et al. Simplifying quantum logic using [48] Cervera-Lierta, A., Krenn, M., Aspuru-Guzik, A. &
higher-dimensional hilbert spaces. Nature Physics 5, 134– Galda, A. Experimental high-dimensional greenberger-
140 (2009). horne-zeilinger entanglement with superconducting trans-
[35] Fedorov, A., Steffen, L., Baur, M., da Silva, M. P. & mon qutrits. Phys. Rev. Applied 17, 024062
Wallraff, A. Implementation of a toffoli gate with super- (2022). URL https://link.aps.org/doi/10.1103/
conducting circuits. Nature 481, 170–172 (2012). URL PhysRevApplied.17.024062.
https://doi.org/10.1038/nature10713. [49] Takita, M., Cross, A. W., Córcoles, A. D., Chow, J. M.
[36] Song, C. et al. Continuous-variable geometric phase and & Gambetta, J. M. Experimental demonstration of fault-
its manipulation for quantum computation in a supercon- tolerant state preparation with superconducting qubits.
ducting circuit. Nature Communications 8, 1061 (2017). Phys. Rev. Lett. 119, 180501 (2017). URL https://link.
URL https://doi.org/10.1038/s41467-017-01156-5. aps.org/doi/10.1103/PhysRevLett.119.180501.
[37] Hill, A. D., Hodson, M. J., Didier, N. & Reagor, M. J. [50] Garbow, J. R., Weitekamp, D. P. & Pines, A. Bilin-
Realization of arbitrary doubly-controlled quantum phase ear rotation decoupling of homonuclear scalar interac-
gates (2021). URL https://arxiv.org/abs/2108.01652. tions. Chemical Physics Letters 93, 504–509 (1982).
[38] Jurcevic, P. et al. Demonstration of quantum volume 64 URL <GotoISI>://WOS:A1982PW86100020. Chem Phys
on a superconducting quantum computing system. Quan- LettPw861Times Cited:358Cited References Count:22.
tum Science and Technology 6, 025020 (2021). URL [51] McKay, D. C., Wood, C. J., Sheldon, S., Chow, J. M. &
https://doi.org/10.1088/2058-9565/abe519. Gambetta, J. M. Efficient z gates for quantum computing.
[39] Mitchell, B. K. et al. Hardware-efficient microwave- Phys. Rev. A 96, 022330 (2017). URL https://link.
activated tunable coupling between superconducting aps.org/doi/10.1103/PhysRevA.96.022330.
qubits. Phys. Rev. Lett. 127, 200502 (2021). URL [52] Beale, S. J. et al. True-Q (2020). URL https://doi.org/
https://link.aps.org/doi/10.1103/PhysRevLett. 10.5281/zenodo.3945250.
127.200502. [53] Magesan, E., Gambetta, J. M. & Emerson, J. Scal-
[40] Wei, K. X. et al. Quantum crosstalk cancellation for fast able and robust randomized benchmarking of quan-
entangling gates and improved multi-qubit performance tum processes. Phys. Rev. Lett. 106, 180504
(2021). URL https://arxiv.org/abs/2106.00675. (2011). URL https://link.aps.org/doi/10.1103/
[41] Xiong, H. et al. Arbitrary controlled-phase gate on fluxo- PhysRevLett.106.180504.
nium qubits using differential ac stark shifts. Phys. Rev. [54] Magesan, E. et al. Efficient measurement of quantum gate
Research 4, 023040 (2022). URL https://link.aps.org/ error by interleaved randomized benchmarking. Phys.
doi/10.1103/PhysRevResearch.4.023040. Rev. Lett. 109, 080505 (2012). URL https://link.aps.
[42] Noguchi, A. et al. Fast parametric two-qubit gates with org/doi/10.1103/PhysRevLett.109.080505.
suppressed residual interaction using the second-order [55] Mullane, S. Sampling random quantum circuits: a pedes-
nonlinearity of a cubic transmon. Phys. Rev. A 102, trian’s guide (2020). arXiv:2007.07872.
062408 (2020). URL https://link.aps.org/doi/10. [56] Wallman, J., Granade, C., Harper, R. & Flammia, S. T.
1103/PhysRevA.102.062408. Estimating the coherence of noise. New Journal of Physics
[43] Erhard, A. et al. Characterizing large-scale quantum 17, 113020 (2015). URL https://doi.org/10.1088/
computers via cycle benchmarking. Nature Communica- 1367-2630/17/11/113020.
tions 10, 5347 (2019). URL https://doi.org/10.1038/ [57] Arute, F. et al. Quantum supremacy using a pro-
s41467-019-13068-7. grammable superconducting processor. Nature 574,
[44] Boixo, S. et al. Characterizing quantum supremacy in 505–510 (2019). URL https://doi.org/10.1038/
near-term devices. Nature Physics 14, 595–600 (2018). s41586-019-1666-5.
URL https://doi.org/10.1038/s41567-018-0124-x.
[45] Johansson, J., Nation, P. & Nori, F. Qutip: An open-
source python framework for the dynamics of open quan-
tum systems. Computer Physics Communications 183,
8

Acknowledgements
We are grateful to A. Hashim, W. Livingston, and
I. Hincks for conversations and insights. This work
was supported by the Quantum Testbed Program of
the Advanced Scientific Computing Research for Basic
Energy Sciences program, Office of Science of the U.S.
Department of Energy under Contract No. DE-AC02-
05CH11231.
 Supplementary Information: High-Fidelity Qutrit Entangling Gates for
Superconducting Circuits
Noah Goss,1, 2 Alexis Morvan,2 Brian Marinelli,1, 2 Bradley K. Mitchell,1, 2 Long B. Nguyen,2 Ravi K. Naik,1, 2
Larry Chen,1 Christian Jünger,2 John Mark Kreikebaum,1, 3
David I. Santiago,2 Joel J. Wallman,4 and Irfan Siddiqi1, 2, 3
1
Department of Physics, University of California, Berkeley, Berkeley CA 94720, USA.
2
Computational Research Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720, USA.
3
Materials Science Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720, USA.
arXiv:2206.07216v2 [quant-ph] 16 Jun 2022

4
Keysight Technologies Canada, Kanata, ON K2K 2W5, Canada.
(Dated: June 17, 2022)

SUPPLEMENTARY NOTE 1 . DEVICE PARAMETERS

Both gates and cross-Kerr characterization data was taken using an 8 transmon ring with fixed-frequency transmons
and fixed-frequency coupling mediated by a coplanar waveguide resonator. The CZ and CZ† gate are performed on
two different chips, using two different pairs of transmon qutrits. We give the relevant single-qutrit parameters and
coherences for the CZ and CZ† gate in Table S1 and Table S2 respectively. Here, T2r denotes coherence statistics
taken using a Ramsey experiment and T2e using an echo pulse.

Q3 Q4
Qubit freq. (GHz) 5.436 5.327
Anharm. (MHz) -260.20 -262.94
T101 (µs) 125(37) 78(16)
T112 (µs) 63(9) 47(5)
01
T2e (µs) 190(28) 138(25)
12
T2e (µs) 61(13) 45(7)
02
T2e (µs) 75(19) 62(6)
01
T2r (µs) 114(47) 99(24)
12
T2r (µs) 17(8) 17(9)
02
T2r (µs) 20(16) 21(9)

Table S1. Single-qutrit parameters for the pair of transmons used to perform the CZ gate.

Q5 Q6
Qubit freq. (GHz) 5.362 5.523
Anharm. (MHz) -275 -271.35
T101 (µs) 45(7) 58(7)
T112 (µs) 33(3) 28(3)
01
T2e (µs) 63(7) 84(6)
12
T2e (µs) 28(3) 30(3)
02
T2e (µs) 37(3) 35(3)
01
T2r (µs) 36(9) 76(8)
12
T2r (µs) 10(6) 18(6)
02
T2r (µs) 11(6) 21(8)

Table S2. Single-qutrit parameters for the pair of transmons used to perform the CZ† gate.
2

SUPPLEMENTARY NOTE 2 . PERTURBATION THEORY

6* 6*
|3,1⟩ |2,2⟩ |1,3⟩
Ω" Ω!
"!

""
3* 4*
|3,0⟩
|2,1⟩ |1,2⟩ 3* |1,0⟩ 𝐽

Δ
|0,3⟩
|0,1⟩
"" +2$

Ω"
Ω!
Energy |", $⟩

"! +2$
2* |1,1⟩
|2,0⟩ 2*
|0,2⟩
"" + $

𝜔! + 𝜂
𝜔!

𝜔"
Ω"

"! + $

𝜔" + 𝜂
Ω! Ω! Ω"
|1,0⟩ *
Δ

|0,1⟩
|0,0⟩
""

""

"" + $

"! + $

Ω" Ω!
"!

"!
|0,0⟩

Figure S1. Drive scheme for conditional stark-induced cross-Kerr Hamiltonian . We present an example energy level
diagram in the rotating frame of the drive where we place the off resonant microwave drive on both qutrits between their two
respective |1i → |2i transitions. The simultaneous off resonant microwave drives induce conditional Stark shifts that generate
entangling phases on four states in our two qutrit Hilbert space.

In the frame of the drive at frequency ωd and after making a rotating wave approximation (RWA), the system
Hamiltonian is
Xh ηi  i  
H= (ωi − ωd ) a†i ai + a†i a†i ai ai + Ωi eiϕi ai + e−iϕi a†i + J a†c at + ac a†t (S1)
i=c,t
2

with ~ = 1 and the transmons approximated as Duffing oscillators with qubit frequency ωi , anharmonicity ηi , capacitive
coupling J, and where we define ai as the bosonic annihilation operator. The parameters of the drive are given as
amplitude Ωi , and phase ϕi . Since only the relative drive phase is physical, we choose a basis where ϕc = 0 and
ϕd = ϕt − ϕc . The detuning of transmon i from the drive is ∆i = ωi − ωd . We analyze the system perturbatively in the
limit Ωi , J  |ηi |, |∆i |. In this limit the bare transmon Hamiltonians serve as the unperturbed system, H0 , and the
perturbation, V , is composed of the single qubit drive terms and the coupling term. Time independent perturbation
theory applied to Eq. S1 will yield energies Eij , the approximate diagonal elements of the Hamiltonian in the basis
labelled by the transmon occupation numbers |iji with i, j = 0, 1, 2, . . . , d − 1 the state of the “control” (c) and “target”
(t) transmons respectively. The dimension where we truncate the Hamiltonian is d. In the present case of qutrits
P2
d = 3 and H 0 ≈ i,j=0 Ẽij |iji hij| where we define Ẽij = Eij − E00 , performing a global shift of the energies to set
the energy of the |00i state to zero.
We isolate the entangling cross-Kerr terms by transforming H 0 according to the unitary
n h io
U = exp − it Ẽ01 I ⊗ |1i h1| + Ẽ10 |1i h1| ⊗ I + Ẽ02 I ⊗ |2i h2| + Ẽ20 |2i h2| ⊗ I , (S2)

where I = |0i h0| + |1i h1| + |2i h2| is the single qutrit identity operator. The transformation eliminates the single qutrit
energies Ẽi0 and Ẽ0j , i, j = 1, 2 which simply result in local phases that can be eliminated by virtual single qutrit
phase gates. The transformed Hamiltonian

H 00 = α11 |11i h11| + α21 |21i h21| + α12 |12i h12| + α22 |22i h22| (S3)
 3

is written in terms of the cross-Kerr rates αij which describe the rates at which the entangling phases on the states
|iji are accumulated. These αij , explicitly given by

αij = Eij + E00 − E0j − Ei0 , (S4)

generalize the qubit ZZ rate ζ to the case of qutrits [1].

(0) (1) (n) (n) (n) (n) (n) (n)
In perturbation theory we expand αij = αij + αij + · · · where αij = Eij + E00 − (Ei0 + E0j ) and Eij is the
(0) (1)
nth order correction to the energy of state |iji. The first two terms αij and αij are zero because in the unperturbed
system, H0 , the transmons are not interacting and the perturbation V only maps states |iji to states |i0 j 0 i that are
orthogonal to |iji. The first contribution comes at second order where the static coupling of strength J between the
transmons generates a static cross-Kerr interaction with rates

(2) 4ηJ 2
α11 = (S5)
(η − ∆)(η + ∆)
(2) 2ηJ 2 (5η − 4∆)
α21 = (S6)
∆(η − ∆)(2η − ∆)
(2) 2ηJ 2 (5η + 4∆)
α12 =− (S7)
∆(η + ∆)(2η + ∆)
(2) 16ηJ 2
α22 = (S8)
(η + ∆)(η − ∆)

where ∆ = ωc − ωt and we have set ηc = ηt = η in order to arrive at more compact expressions. In most systems
this is a reasonable approximation and in particular the pair of transmons used in this work have anharmonicities
ηc = 272 MHz and ηt = 270 MHz.
As in the qubit case the driven cross-Kerr interaction contributes starting at third order when both transmons are
driven (see the discussion in the Supplementary Materials of [1]) and we find rates

8 cos φη 2 ΩC ΩT J
α11 = (S9)
∆C ∆T (η − ∆C )(η − ∆T )

∆2C (η − 2∆T ) + η(2η 2 − 2η∆T − ∆2T )) + ∆C (−3η 2 + 2η∆T + 2∆2T )

α12 /α11 = (S10)
(2η − ∆C )(2η − ∆T )

∆2T (η − 2∆C ) + η(2η 2 − 2η∆C − ∆2C ) + ∆T (−3η 2 + 2η∆C + 2∆2C )

α21 /α11 = (S11)
(2η − ∆C )(2η − ∆T )

(η − 2∆C )(η − 2∆T )

α22 /α11 = (S12)
(2η − ∆C )(2η − ∆T )

The expressions are complicated but we now point out some important features. In typical systems the drive
(3) (2)
strengths Ωi are larger than the coupling strength J and as a result αij > αij . For example, the CZ gate described
in the main text is performed with Ωc , Ωt ≈ 11 MHz on a pair of transmons with estimated coupling J = 2.7 MHz.
Therefore, the static cross-Kerr can in principle be cancelled by the driven cross-Kerr. The result also gives the
leading order linear dependence of the rates αij on the drive strengths Ωi and the coupling J as well as their sinusoidal
dependence on the relative drive phase ϕ (see Figure 2 in the main text).
4

SUPPLEMENTARY NOTE 3 . THE WEYL AND GELL-MANN BASES

When analyzing qubits, we use tensor products of the single-qubit Pauli operators P = {I, X, Y, Z}. These operators
have the following helpful properties:

1. they form a projective group under matrix multiplication;

2. they are unitary;

3. they are a trace-orthogonal basis for the space of operators; and

4. they correspond to natural Hamiltonians.

Unfortunately, no set of operators with the same properties exist for higher dimensions. We thus need to separate
some of the properties when analyzing qudits.
There are two sets of operators, namely, the Weyl and Gell-Mann operators that, taken together, satisfy all four
properties and also coincide with the Pauli operators in 2D. We now review these two sets of operators.
The Gell-Mann operators are obtained by embedding the familiar Pauli matrices into two-dimensional subspaces of
a higher dimensional space. Recall that the standard single-qubit Pauli operators are

X = |1ih0| + |0ih1|
Y = i |1ih0| − i |0ih1|
Z = |0ih0| − |1ih1| . (S13)

We can embed these operators in a d-dimensional subspace by defining

X jk = |jihk| + |kihj|
Y jk = i |jihk| − i |kihj|
Z jk = |jihj| − |kihk| (S14)

for 0 ≤ j < k < d. The X jk and Y jk operators are trace-orthogonal Hermitian operators and also correspond to
transitions between two levels, which is the natural way to control a harmonic oscillator. Specifically, we can drive the
X 01 and X 12 Hamiltonians by applying tones at ω01 and ω12 , which generate Rabi oscillations in the corresponding
qubit subspaces of the qutrit, from which we can generate the single qutrit gates
   
1 1 −i 0 1 1 0 0
01
Xπ/2 = √ −i 1 0 , Xπ/2 12
= √ 0 1 −i . (S15)
2 0 0 1 2 0 −i 1

Similarly, we can perform virtual Z gates within these two qubit subspaces of the qutrit, which natively yield the
continuous gates
 −iφ   
e 0 0 1 0 0
Z 01 (φ) =  0 1 0 , Z 12 (φ) = 0 1 0 (S16)
0 0 1 0 0 eiφ

From these four gates, Xπ/2

01 12
, Xπ/2 , Z 01 (φ), Z 12 (φ), we can compile an arbitrary unitary in U(3) using at most 6 of
these gates, with a decomposition given in Ref [2].
However, while the Z jk are Hermitian operators, they are not linearly independent and so do not provide a suitable
basis for quantum process tomography. We thus extend the set {X jk , Y jk : 0 ≤ j < k < d} to a trace-orthogonal
basis by adding a trace-orthogonal set of diagonal operators. A natural choice would be the projectors onto the
 5

computational basis, however, this obscures the fact that all density operators have unit trace. We thus use the
operators
X
Dj = −j |jihj| + |kihk| (S17)
0≤k<j

for 1 ≤ j < d, together with the identity operator to obtain the Gell-Mann basis. Here the Gell-Mann matrices λi
plus the identity I3 span SU(3) and are a natural choice for qutrit Pauli transfer matrices (PTMs). For convenience,
we index these elements for a single qutrit as follows:
         
1 0 0 0 1 0 0 −i 0 1 0 0 0 0 1
I3 = 0 1 0 , λ1 = 1 0 0 , λ2 =  i 0 0 , λ3 = 0 −1 0 , λ4 = 0 0 0 ,
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
        (S18)
0 0 −i 0 0 0 0 0 0 1 1 0 0
λ5 = 0 0 0  , λ6 = 0 0 1 , λ7 = 0 0 −i , λ8 = √ 0 1 0 
i 0 0 0 1 0 0 i 0 3 0 0 −2

Having defined the Gell-Mann basis, we now define the Heisenberg-Weyl operators, which are a unitary generalization
of the familiar single-qubit Pauli operators to higher dimensional spaces (qudits) that enable the Clifford group to be
generalized. Let d be a positive integer and Zd = {0, . . . , d−1} denote the set of integers modulo d. The generalizations
of the X and Z operators to qudits are
X
X= |j ⊕d 1ihj|
j∈Zd
X  
2πi
Z= exp j |jihj| , (S19)
d
j∈Zd

where ⊕d denotes addition modulo d. The Weyl basis for Cd×d is the set

Wd = {Wxz = X x Z z : x, z ∈ Zd } , (S20)

which is a trace-orthogonal basis for Cd×d as it satisfies

tr W † V = dδW,V ∀ W, V ∈ Wd . (S21)

Let n be a positive integer and D = dn . Then we define the n-qudit Weyl basis to be the set Wd,n = W⊗n d , which
is a trace-orthogonal basis for CD×D . Note that the Weyl basis is not a proper group as it is not closed under
multiplication. However, every element of the closure of the Weyl basis is proportional to an element of the Weyl
basis up to an overall phase. This overall phase vanishes in all cases as we only consider the adjoint action of Weyl
operators and so we treat the Weyl basis as a projective group. The n-qutrit Clifford group is then defined to be the
normalizer of the extended Weyl group EWd,n = U(1)Wd,n (which is a proper group), that is,

Cd,n = {U ∈ U(dn ) : U EWd,n U † = EWd,n }. (S22)

6

SUPPLEMENTARY NOTE 4 . ASSUMPTIONS FOR CHARACTERIZATION

We now define the assumption of time-dependent Markovian noise that we use to analyze the results of our charac-
terization protocols. At a high level, we assume that each operation applied to the system corresponds to a linear map
that is independent of what other maps have been applied but may depend on the number of operations that have
been applied since the system was initialized. We allow this time dependence primarily to facilitate post-processing
of time-stationary noise processes wherein the physical process does not depend on the number of operations that
have been applied since the system was initialized but we are performing a weighted average over the operation that
is applied in the jth time step.
Formally, for a vector space V let V∗ denote the dual of V and L(V) denote the set of linear maps from V to itself.
Then we assume the following.
1. The state space of a physical implementation of an n-qudit system is some fixed vector space V.
2. Preparing the system in the state ρ corresponds to setting the state of the quantum system to some Θρ ∈ V.
3. Applying some unitary operation U to the system in the jth time step after preparing the system in a state
corresponds to applying some linear map, Θ(j, U ) ∈ L(V) to the state of the system.
4. The expectation value of an observable Q is obtained by applying some fixed ΘQ ∈ V∗ . (Note that we will ignore
finite measurement statistics for now.)
With some abuse of notation, we refer to the functions Θj : N × U(dn ) → L(V → V) and the vectors Θρ and ΘQ
together as the implementation map Θ [3].
The assumption of time-dependent Markovian noise allows hidden Markovianity as the implementation map can
include a coupling to an environment. This hidden Markovianity is frequently referred to as non-Markovianity in
the quantum information community. However, we allow it in the general setting because the additional assumptions
(such as positivity and complete positivity) are more cumbersome to define and typically are only helpful in the final
steps of an analysis. Indeed, it is conceptually useful to include post-processing steps that depend only on one time
step into an “effective” implementation map.
An ideal isolated implementation is one wherein there is a linear isomorphism between V and CD×D . For concrete-
ness, we define an isomorphism |∗iiB : CD×D → V relative to a trace-orthonormal basis B ⊂ CD×D as follows. As B
is a trace-orthonormal basis, we can write any A ∈ CD×D in terms of B as
X

A= tr B † A B. (S23)
B∈B

Therefore we can choose {|BiiB : B ∈ B} to be an orthonormal basis of V and extend it to an isomorphism by defining
X

|AiiB = tr B † A |BiiB . (S24)
B∈B

We will typically suppress the subscript B as it will be clear from the context (either the normalized Weyl basis or the
normalized Gell-Mann basis). To avoid having to define normalized versions of the bases explicitly, we define
√
B̂ = B/ tr B † B. (S25)

Moreover, defining hhA| = |Aii† and hhA|Bii = hhA||Bii and using the fact that B is a trace-orthonormal basis and
{|Bii : B ∈ B} is an orthonormal basis, we have
X



tr A† T = tr A† B tr C † T tr B † C
B,C∈B
X


= tr A† B tr B † T
B∈B
= hhA|T ii. (S26)
 7

With the above isomorphism, we can define the ideal implementation of a unitary operator U ∈ U(D) to be
X
φB (U ) = |U BU † iihhB|, (S27)
B∈B

which, by linearity, will satisfy

φB (U )|Aii = |U AU † ii. (S28)

As above, we will also suppress the B on φ when the basis is clear from the context. Moreover, as U BU † is a trace-
orthonormal basis, one can readily verify that φ(U )φ(V ) = φ(U V ) for all U, V ∈ U(D), that is, φ is a representation
of U(D).
8

SUPPLEMENTARY NOTE 5 . CROSS ENTROPY BENCHMARKING

a b c

d e f

Figure S2. Cross entropy benchmarking (XEB) for qutrit, a, For the XEB circuits, the entropy difference H(p, q)−H(p, u)
is plotted against the ideal H(p, p) − H(p, u) at each cycle depth, with 30 randomizations at each depth. The linear fit gives
the fidelity at that particular depth. b, We show the exponential decay of the depolarized CZ† fidelity obtained from the linear
fits in a. c, We plot an example of the speckle purity decay as a function of cycle depth for a representative set of four states in
the two-qutrit Hilbert space. The probabilities of measuring the given tritstring are shown as a function of cycle depth across
all 30 randomizations. The bright “speckle” pattern characteristic of the Porter-Thomas distribution at low circuit depths is
smoothed out at larger depths. d-f, We show that the distribution of tritstrings transitions from the Porter-Thomas ditribution
at low depth, to a uniform distribution at deeper depths. The CDF is emphasized for a representative tritstring, |21i, while the
CDFs of the other tritstrings are shown in grey to show typical variation of the CDF across different states.

As discussed in the main text, performing randomized benchmarking (RB) on a two qutrit gate is prohibitively
expensive [2]. We utilize cross entropy benchmarking (XEB) as a second SPAM (state preparation and measurement)
free benchmarking protocol to corroborate the fidelity obtained from cycle benchmarking (CB). Recently XEB played
a central role in the quantum supremacy experiment in [4] and was used to benchmark the non-Clifford, three qubit
iToffoli gate in [5]. XEB theory has been discussed in detail in previous works so we present only a brief review of the
method before outlining more explicitly how it can be used to benchmark the two-qutrit CZ gate demonstrated here.
For XEB purposes a convenient definition of a random quantum circuit (RQC) is a quantum circuit randomly
selected from an ensemble of circuits such that the distribution of probabilities (across the ensemble of circuits) of
observing a particular ditstring, x, follows the Porter-Thomas distribution (for all possible ditstrings x) [6]. In the
present case of quantum circuits involving two qutrits we consider the probabilities of tritstrings 00, 01, . . . , 21,
22. After averaging over an ensemble of random circuits errors are tailored to be purely depolarizing so an error
corresponds to the circuit outputting a fully mixed state. In a mixed state the tritstring distribution is uniform
with each outcome xi having equal probability of 1/3n for an n qutrit system. Thus, intuitively the circuit fidelity
(under this error model) can be thought of as the deviation of the measured tritstring distribution from the uniform
distribution. The XEB fidelity makes this relationship precise as we now outline.
Denote the possible tritstrings xi for i = 1, . . . , 3n and let p(xi ) be the ideal tritstring distribution for the output
of a particular quantum circuit. Then q(xi ) is the measured distribution. The linear cross entropy of two probability
 9

distributions p1 (x) and p2 (x) with the same support is defined as

X
H(p1 , p2 ) = p1 (x)p2 (x) (S29)
x

where the sum runs over the full support of the probability distributions and the self entropy is written compactly as
H(p1 ) ≡ H(p1 , p1 ). Under the depolarizing error model it can be shown straightforwardly that the circuit fidelity is

H(p, q) − H(p, u)
FXEB = (S30)
H(p, p) − H(p, u)

where u(xi ) is the uniform probability distribution [4]. This is ultimately the difference in the ideal to measured and
ideal to uniform cross entropies, normalized by the difference if the measured distribution were to perfectly match the
ideal distribution (p(xi ) = q(xi ) for all i).
We follow the XEB protocol for benchmarking gates/cycles outlined in Ref. [4]. This consists of measuring two
qutrit circuits of varying cycle depths M where cycle i consists of two single qutrit gates Gi,j where i = 1, . . . , M
labels the cycle and j = 1, 2 labels the qutrit, followed by the two qutrit CZ gate† (see Figure S2a). The single qutrit
gates are randomly selected unitaries from SU (3) and decomposed according to the decomposition given in [7]. For
each cycle depth M we generate N random circuits to be run and compute the ideal probability distributions of the
output tritstrings for each circuit. The XEB fidelity at cycle depth M , FXEB,M is determined by performing a least
squares fit of the linear relationship between H(pi , qi )−H(pi , u) and H(pi , pi )−H(pi , u) with i = 1, . . . , N labeling the
random circuit at the given depth (see Figure). The cycle infidelity, cycle , is estimated from the exponential decay of
FXEB,M as a function of cycle depth M . The error rate extracted in this way agrees well with CB (see Benchmarking
section in the main text).
Next we validate our implementation of XEB by verifying that the distribution of probabilities of a given tritstring
across the ensemble of random circuits does indeed approach the Porter-Thomas distribution,

P(p) = (D − 1)(1 − p)D−2 ≈ De−Dp (S31)

where again D is the Hilbert space dimension, D = 32 for two qutrits and the approximate equality holds in the limit
of large Hilbert space dimension, D  1. Next we consider the evolution of this distribution as a function of cycle
depth. At short to intermediate cycle depths, the circuit is sufficiently random and the distribution approaches Porter-
Thomas. At larger cycle depths, the distribution begins to converge to P(p) → δ(p−1/D) since the depolarizing errors
dominate and the tritstring distributions approach the uniform distribution for all circuits. We plot our experimental
observation of this behavior in Fig. S2d-f.
The method of Speckle Purity Benchmarking (SPB) is based on this observation. Denoting the state purity as γ, it
can be estimated at depth M from the raw results of the XEB protocol outlined above by the relationship

D2 (D + 1)
γ(M ) = Var(pM ) (S32)
(D − 1)

where pM is the set of measured probabilities of a given tritstring x across the N random circuits at cycle depth M
[4]. Thus, once we have demonstrated that the distribution does indeed converge to the Porter-Thomas distribution
we can estimate the decay of the state purity from just the variance of the distribution, without the need for state
tomography which requires an exponential number of measurements, 3n(d−1) where n is the number of qudits and
d the dimension of each qudit. For our present case of two qutrits we would need to perform 81 measurements per
circuit to determine the state purity by state tomography.
10

SUPPLEMENTARY NOTE 6 . EXTRACTING AN ERROR BUDGET ESTIMATION FROM CYCLE

BENCHMARKING RESULTS

 &= †
5HIHUHQFH

,QILGHOLW\








: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 

:

:

:
:

:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:

:
:
:
:
:
:
:
:

:

:
:
:
:
:
:
:
:
:

:H\O'HFD\7HUPV

Figure S3. Cycle benchmarking results with Weyl decay terms of CZ† gate. The cycle benchmarking results shown
analyze a subset of 53 Weyl channels at depths m ∈ {0, 3, 6, 15}. In order to provide a SPAM free estimation of the error, we
compare our CB results to a reference CB experiment without the two qutrit gate to isolate and subtract the error due to the
Weyl twirling itself. Here we plot only the shared decay channels analyzed for both the CZ† and reference CB experiments.

We now outline how data from cycle benchmarking [8] can be used to break down the error budget for a multi-qudit
Clifford gate. The following analysis also generalizes that of [8] to gate-dependent noise and allows some non-Markovian
errors (namely, couplings to an environment).

A. The cycle benchmarking protocol

We now outline the cycle benchmarking protocol. Let d, m, and n be positive integers, ψ be a quantum state, and
Q be an observable. Then a cycle benchmarking circuit consists of the following:

1. Prepare the system in a state ψ;

2. For each j ∈ Zm ,

(a) Choose an n-qudit Weyl operator Wj ∈ W⊗n

d uniformly at random;
(b) Apply Wj to the system; and
(c) Apply C to the system.

3. Choose an n-qudit Weyl operator Wm ∈ W⊗n

d uniformly at random;

4. Apply Wm to the system; and

5. Measure the expectation value of an operator Q.

 11

For time-dependent Markovian noise, the expectation value for the cycle benchmarking protocol for a fixed choice
~ = (W0 , . . . , Wm ), preparation ψ and operator Q relative to an implementation map Θ is
of random Weyl operators W
 
Y
µΘ,ψ,W~ ,Q = ΘQ
 Θ(2j, Wj )Θ(2j − 1, C) Θ(0, W0 )Θψ . (S33)
j=m→1

The Weyl operators are sampled uniformly and independently and are uncorrelated in eq. (S33). The average over all
choices of random Weyl operators is then
 
Y
EW~ µΘ,ψ,W~ ,Q = ΘQ
 Θ(j)Θ(2j − 1, C) Θ(0)Θψ , (S34)
j=m→1

where for an implementation map Θ we define

Θ(j) = EW ∈Wd,n Θ(2j, W ). (S35)
That is, we have factorized the average of the expectation values into a product of independent averages.

B. CB decay

Before turning to analyze the variance, we first show how the above method is equivalent to that of Ref. [8] under
equivalent assumptions. First, let Λ denote the physical implementation map and assume that for all W ∈ Wd,n , we
have
Λ(W ) = Lφ(W )R (S36)
for some linear maps L and R. We then redefine
Λψ → RΛψ
Λ(2j − 1, C) → RΛ(2j − 1, C)L,
ΛQ → ΛQ L, (S37)
so that without loss of generality we can set L and R to the identity, that is, we can assume that the implementation
of the random operations is effectively ideal.
One can readily verify that the ideal implementation of a Weyl operator W satisfies
X X
φ(W ) = |W V̂ W † iihhV̂ | = χV (W )|V̂ iihhV̂ | (S38)
V ∈Wd,n V ∈Wd,n

where
χV (W ) = hhV̂ |W V̂ W † ii = D−1 tr V † W V W † , (S39)
which is a character of the projective Weyl group. By Schur’s orthogonality relations, for any U ∈ Wd,n we have
X
EW ∈Wd,n χ∗U (W )χV (W )|V̂ iihhV̂ | = |Û iihhÛ |, (S40)
V ∈Wd,n

so that if we set Θ(2j, Wj ) = χ∗Uj (Wj )φ(Wj ) for Uj ∈ Wd,n , which can be accomplished by multiplying µΘ,W
~ by
χ∗Uj (Wj ) for each j, we find that eq. (S33) simplifies to
Y
ΛQ |Ûm iihhÛ0 |Λψ hhÛj+1 |Θ(2j − 1, C)|Ûj ii, (S41)
j=m→1

which is now a product of scalars so we can reorder terms as desired. When C is a Clifford operator, choosing
Uj = C j U0 C −j for some fixed U0 ∈ Wd,n reduces eq. (S41) to the expression in Ref. [8].
12

C. CB variance

We now analyze the variance to lowest order to show how the unitarity [9] can be estimated. The analysis closely
parallels that of Ref. [9], except that because we are using a weaker twirl (namely, over the Heisenberg-Weyl group
instead of over the full multi-qudit Clifford group), the matrix that governs the decay rates has more eigenvalues. As
in Ref. [9], We use the fact that for any scalar λ, we have |λ|2 = λ⊗λ∗ and that multiplication distributes across tensor
products, so that we will consider the implementation map Θ = Λ ⊗ Λ∗ , where the phases added when analyzing a
decay cancel. From eq. (S33), we then have
2
EW ~ ,Q | = EW
~ |µΘ,ψ,W ~ µΘ,ψ,W
~ ,Q
 
Y
= ΘQ  Θ(j)Θ(C) Θ(0)Θψ . (S42)
j=m→1

Applying Schur’s orthogonality relations gives

Θ(j) = EW ∈Wd,n φ(W ) ⊗ φ∗ (W )

X
= |Û ⊗ V̂ iihhÛ ⊗ V̂ |EW ∈W⊗n χU (W )χ∗V (W )
d
U,V ∈Wd,n
X
= |Û ⊗ V̂ iihhÛ ⊗ V̂ |δU,V
U,V ∈Wd,n
X
= |Û ⊗ Û iihhÛ ⊗ Û |. (S43)
U ∈Wd,n

In particular, Θ(j)2 = Θ(0) for all j and so eq. (S42) can be written as
2 m
EW
~ |µΛ,ψ,W
~ ,Q | = ΘQ M Θψ , (S44)

where
X
M = Θ(0)Θ(C)Θ(0) = |Û ⊗ Û iihhV̂ ⊗ V̂ ||ΛU V (C)|2 . (S45)
U,V ∈Wd,n

For simplicity, we assume that Λ(C) is unital and trace preserving, that is, that for all U ∈ Wd,n we have

ΛU,I (C) = ΛI,U (C) = δI,U . (S46)

Then we can rewrite eq. (S44) as

EW
~ |µΛ,ψ,W
2 m ˆ⊗2 iihhIˆ⊗2 |Θψ
~ ,Q | = ΘQ Mu Θψ + ΘQ |I

ˆ I|Λ
= ΘQ Mum Θψ + |ΛQ |Iiihh ˆ ψ |2 , (S47)

where we define the unital block of M to be

X
Mu = Θ(0)Θ(C)Θ(0) = |Û ⊗ Û iihhV̂ ⊗ V̂ ||ΛU V (C)|2 (S48)
U,V ∈W∗
d,n

and W∗d,n = Wd,n − {I}. Further, note that the constant term is can be directly estimated from eq. (S41) for any C
by setting U0 = I, and so for unital noise the variance over random Weyl operators satisfies
2 2 m
EW
~ |µΛ,ψ,W
~ ,Q | − |EW ~ ,Q | = ΘQ Mu Θψ .
~ µΛ,ψ,W (S49)
 13

We now show how the unitarity of a channel [9] can be estimated from eq. (S49) under assumptions similar to those
in XEB. First, note that if Mu is diagonalizable, eq. (S44) can be written as
X
2
EW~ |µΛ,W~ | = αj λm
j (S50)
j

where the λj are the nonzero eigenvalues of Mu and the αj are the overlaps of the corresponding eigenvectors of Mu
with the SPAM vectors. We thus want to prove that the unitarity of a channel corresponds to an eigenvalue of Mu .
The unitarity u of a quantum channel is defined to be [9]
1 X
u2 = |ΛU V (C)|2 . (S51)
D2 −1
U,V ∈W∗
d,n

As in XEB, we now assume that the error model consists only of a unitary error and global depolarization, so that

Λ(C) = φ(T )Dp (S52)

for some unknown T ∈ U(D) where

Dp (ρ) = pρ + (1 − p)I/D (S53)

is the global depolarizing channel. Under this assumption, we have u = p, so that the goal is to learn an effective
depolarizing rate of the noise process. We now show that
X
v= |U ii (S54)
U ∈W∗
d,n

is an eigenvector of Mu with eigenvalue p2 . First note that for all U ∈ W∗d,n , we have
X
|ΛU,V |2 = p2 . (S55)
V ∈W∗
d,n

Therefore
X
Mu v = |Û ⊗2 ii|ΛU,V (C)|2 hhV̂ ⊗2 |Ŵ ⊗2 ii
U,V,W ∈W∗
d,n
X
= |Û ⊗2 iip2
U ∈W∗
d,n

= p2 v (S56)

that is, we have an eigenvector of Mu the purity matrix whose eigenvector is exactly p2 . Moreover, by eq. (S55),
Mu /p2 is a stochastic matrix and so its largest eigenvalue is 1. Therefore, under the assumption of global depolarizing
and unitary noise, the largest eigenvalue of the unital part of Mu is u2 .
We now show how we can estimate the eigenvalues of Mu . Let s be the order of C, that is, the smallest positive
integer such that C s ∝ I and assume that s divides m. Then in the limit of high fidelity, the eigenvectors of Mus
are |Û ⊗2 ii, and so we can find the largest eigenvalue of Mu (and hence the purity under the assumption of global
depolarizing and unitary noise) by finding the slowest decay rate of eq. (S49).
14

SUPPLEMENTARY NOTE 7 . QUTRIT CZ GATE BENCHMARKING

Here we provide the cycle benchmarking results of the 783 ns qutrit CZ gate from the main text. As with the CZ† ,
we estimate the process fidelity of the isolated CZ gate by comparing our CB results to a reference cycle, and using
the formula:
D − 1 FCZ 
eF = 1− (S57)
D FReference
where D is the dimension of the Hilbert space of one’s system. In our benchmarking of the CZ, we find a Weyl
infidelity of 0.0861 for the dressed CZ cycle and 0.034 for the reference cycle. Ultimately isolating the errors of the
CZ gate, we find an estimated process fidelity of 95.2(3)%.
‘


&=
 5HIHUHQFH


,QILGHOLW\






: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 
: 

:

:

:
:
:
:
:
:
:

:
:
:
:
:
:
:
:
:
:
:

:
:
:
:

:
:
:
:
:
:
:
:
:
:
:
:
:
:

:H\O'HFD\7HUPV

Figure S4. Cycle benchmarking results with Weyl decay terms of the qutrit CZ gate. The cycle benchmarking results
analyze a subset of 54 Weyl channels at depths m ∈ {0, 3, 6} for the CZ cycle. In order to provide a SPAM free estimation of the
error, we again compare our CB results to a reference CB experiment. Here we plot only the shared decay channels analyzed
for both the CZ and reference CB experiments.
 15

SUPPLEMENTARY NOTE 8 . FREQUENCY DEPENDENCE OF DRIVEN CROSS-KERR

𝜔!",$! /2𝜋

𝜔!",$" /2𝜋

𝜔%!,$! /2𝜋
𝜔%!,$" /2𝜋

𝜔%!,$! /2𝜋
𝜔!",$! /2𝜋

𝜔!",$" /2𝜋

𝜔%!,$" /2𝜋
𝜔!",$" /2𝜋

𝜔!",$! /2𝜋
𝜔!",$! /2𝜋

𝜔%!,$" /2𝜋
𝜔%!,$! /2𝜋

𝜔%!,$! /2𝜋
𝜔!",$" /2𝜋
𝜔%!,$" /2𝜋

Figure S5. Frequency dependence of the full microwave-activated cross-Kerr Hamiltonian. We compare the depen-
dence of all of the αij in the driven cross-Kerr Hamiltonian on the frequency of the drive ωd using an ab-initio master equation
simulation in QuTiP.

In Figure 2 in the main text, we showed the results of fitting the frequency dependence of the α12 term in the
driven cross-Kerr Hamiltonian to an ab-initio master equation simulation. Here we provide some additional details
on how this simulation was performed in QuTiP [10, 11] and present the results for measuring and characterizing the
frequency dependence of all four of the αij in the driven cross-Kerr Hamiltonian. We note that transient TLS features
and higher transitions meant that some of the data did not fit to a linear model, we therefore only plot data where the
uncertainty on our linear fit (the source of the error bars) was less than 300 KHz. Similarly, our simulation at times
produced unphysical results near transitions, with very large cross-Kerr; in the interest of readability, we therefore
also omitted points where the magnitude of the simulated cross-Kerr was larger than 3 MHz.
To perform the master equation simulation, we considered the Hamiltonian of two fixed frequency transmons, with
a fixed capacitive coupling from a coplanar waveguide resonator. For the frequencies and anharmonicities of the pair
of transmons, we used the experimental parameters from our chip (as can be found in Table S1). We first found
the strength of the capacitive coupling, J, by adjusting it until the simulated parameters for the always on, static
αij best matched our experimental measurements of these parameters. We then fixed a single drive frequency, ωd ,
and performed a master equation simulation of the simultaneous Stark drives at that frequency until our data best
matched the values found for that point in Figure S5. After this, we extrapolated those parameters to be the same
across all frequencies of the Stark driving, and simulate using them for the rest of the frequencies in Figure S5.
16

SUPPLEMENTARY NOTE 9 . QUANTUM PROCESS TOMOGRAPHY

We analyze the Pauli transfer matrix (PTM) of the CZ† gate in the Gell-Mann basis. To construct the PTM in
this basis, we prepare 81 two-qutrit input states by applying an informationally complete set of native gates on
(01) (01) (12) (01) (12) (01) (12) (01) (12) (01)
each qutrit: {I, Xπ/2 , Yπ/2 , X(01) (12) (01)
π , Xπ Xπ , Yπ Xπ/2 , Xπ/2 Xπ , Yπ/2 Xπ , Xπ Xπ/2 }. We measure the output
tomography using the same set of native gets, and reconstruct the PTM from the data using maximum likelihood
estimation method. Figure S6 shows the reconstructed PTM of the CZ† gate Eexp in the Gell-Mann basis, and in
† †
Figure S7 we plot Eideal Eexp . The process fidelity is calculated from the PTM to be FPTM = Tr[Eideal Eexp ]/D2 = 93.2%.
We note that as process tomography does not decouple our characterization of our gate errors from state preparation
and measurement errors, it is disfavored as a benchmarking technique when compared to CB or XEB.
 17

I3I3

I3 8


1 7



2 6


3 5
2XWSXW*HOO0DQQ2SHUDWRU



4 4 


5 3


6 2



7 1


8I3

8 8
I3I3 I3 8 1 7 2 6 3 5 4 4 5 3 6 2 7 1 8I3 8 8
,QSXW*HOO0DQQ2SHUDWRU

Figure S6. CZ† process matrix. Experimentally reconstructed Process Matrix of Qutrit CZ† (Eexp ) gate with process fidelity
of 93.2%.
18

I3I3

I3 8


1 7



2 6


3 5
2XWSXW*HOO0DQQ2SHUDWRU



4 4 


5 3


6 2



7 1


8I3

8 8
I3I3 I3 8 1 7 2 6 3 5 4 4 5 3 6 2 7 1 8I3 8 8
,QSXW*HOO0DQQ2SHUDWRU

†
Figure S7. Comparing the experimentally reconstructed process matrix to the ideal case. We plot Eideal Eexp , from
†
which we can estimate the process fidelity of the CZ gate.
 19

[S1] Mitchell, B. K. et al. Hardware-efficient microwave-activated tunable coupling between superconducting qubits. Phys.
Rev. Lett. 127, 200502 (2021). URL https://link.aps.org/doi/10.1103/PhysRevLett.127.200502.
[S2] Morvan, A. et al. Qutrit randomized benchmarking. Phys. Rev. Lett. 126, 210504 (2021). URL https://link.aps.org/
doi/10.1103/PhysRevLett.126.210504.
[S3] Helsen, J., Roth, I., Onorati, E., Werner, A. H. & Eisert, J. A general framework for randomized benchmarking. arXiv
(2020).
[S4] Arute, F. et al. Quantum supremacy using a programmable superconducting processor. Nature 574, 505–510 (2019).
URL https://doi.org/10.1038/s41586-019-1666-5.
[S5] Kim, Y. et al. High-fidelity three-qubit itoffoli gate for fixed-frequency superconducting qubits. Nature Physics (2022).
URL https://doi.org/10.1038/s41567-022-01590-3.
[S6] Mullane, S. Sampling random quantum circuits: a pedestrian’s guide (2020). arXiv:2007.07872.
[S7] Dita, P. Factorization of unitary matrices. Journal of Physics A: Mathematical and General 36, 2781–2789 (2003). URL
https://doi.org/10.1088/0305-4470/36/11/309.
[S8] Erhard, A. et al. Characterizing large-scale quantum computers via cycle benchmarking. Nature Communications 10,
5347 (2019). URL https://doi.org/10.1038/s41467-019-13068-7.
[S9] Wallman, J., Granade, C., Harper, R. & Flammia, S. T. Estimating the coherence of noise. New Journal of Physics 17,
113020 (2015). URL https://doi.org/10.1088/1367-2630/17/11/113020.
[S10] Johansson, J., Nation, P. & Nori, F. Qutip: An open-source python framework for the dynamics of open quantum systems.
Computer Physics Communications 183, 1760–1772 (2012). URL https://www.sciencedirect.com/science/article/
pii/S0010465512000835.
[S11] Johansson, J., Nation, P. & Nori, F. Qutip 2: A python framework for the dynamics of open quantum systems. Com-
puter Physics Communications 184, 1234–1240 (2013). URL https://www.sciencedirect.com/science/article/pii/
S0010465512003955.