Realization of Three-Qubit Quantum Error Correction with Superconducting Circuits

M. D. Reed,1 L. DiCarlo,2 S. E. Nigg,1 L. Sun,1 L. Frunzio,1 S. M. Girvin,1 and R. J. Schoelkopf1

1
Departments of Physics and Applied Physics, Yale University, New Haven, Connecticut 06520, USA
2
Kavli Institute of Nanoscience, Delft University of Technology, Delft, The Netherlands
(Dated: September 26, 2011)

7–9
Quantum computers promise to solve certain and solid-state10 NMR and with trapped ions11,12
problems exponentially faster than possible clas- have demonstrated two possible strategies for using the
sically but are challenging to build because of error syndromes. The first is to measure the ancillas
their increased susceptibility to errors. Remark- and use a classical logic operation to correct the de-
ably, however, it is possible to detect and cor- tected error. This “feed-forward” capability is challeng-
arXiv:1109.4948v1 [quant-ph] 22 Sep 2011

rect errors without destroying coherence by us- ing in superconducting circuits as it requires a fast and
ing quantum error correcting codes1 . The sim- high-fidelity quantum non-demolition measurement, but
plest of these are the three-qubit codes, which is likely a necessary component to achieve scalable fault-
map a one-qubit state to an entangled three-qubit tolerance2,13 . The second strategy, as recently demon-
state and can correct any single phase-flip or bit- strated with trapped ions12 and used here, is to replace
flip error of one of the three qubits, depending the classical logic with a quantum CCNot gate which
on the code used2 . Here we demonstrate both performs the correction coherently, leaving the entropy
codes in a superconducting circuit by encoding associated with the error in the ancilla qubits. The CC-
a quantum state as previously shown3,4 , induc- Not performs exactly the action that would follow the
ing errors on all three qubits with some probabil- measurement in the first scheme: flipping the primary
ity, and decoding the error syndrome by revers- qubit if and only if the ancillas encode the associated
ing the encoding process. This syndrome is then error syndrome.
used as the input to a three-qubit gate which cor- The CCNot gate is also vital for a wide variety of ap-
rects the primary qubit if it was flipped. As the plications such as Shor’s factoring algorithm14 and has
code can recover from a single error on any qubit, attracted significant experimental interest with recent
the fidelity of this process should decrease only implementations in linear optics15 , trapped ions16 , and
quadratically with error probability. We imple- superconducting circuits17,18 . Here we use the circuit
ment the correcting three-qubit gate, known as a quantum electrodynamics architecture19 to couple four
conditional-conditional NOT (CCNot) or Toffoli transmon qubits20 to a single microwave cavity bus21 ,
gate, using an interaction with the third excited where each qubit transition frequency can be controlled
state of a single qubit, in 63 ns. We find 85±1% fi- on nanosecond timescales with individual flux bias lines22
delity to the expected classical action of this gate and collectively measured by interrogating transmission
and 78 ± 1% fidelity to the ideal quantum pro- through the cavity23 . (The details of the device can be
cess matrix. Using it, we perform a single pass of found in the Methods Summary and in Ref. 3.) Qubits
both quantum bit- and phase-flip error correction are tuned to 6, 7, and 7.85 GHz, with the fourth at
with 76 ± 0.5% process fidelity and demonstrate ∼ 13 GHz and unused (hereafter referred to as Q1 -Q4 ).
the predicted first-order insensitivity to errors. In this paper, we first demonstrate the three-qubit inter-
Concatenating these two codes and performing action used in the gate, which is the logical extension of
them on a nine-qubit device would correct arbi- interactions used in previous two-qubit gates3,22,24 , and
trary single-qubit errors. When combined with demonstrate how this interaction can be used to create
recent advances in superconducting qubit coher- the desired CCNot. We then characterize its classical and
ence times5,6 , this may lead to scalable quantum quantum action and finally use the gate to demonstrate
technology. three-qubit error correction.
Quantum error correction relies on detecting the pres- Our three-qubit gate employs an interaction with the
ence of errors without gaining knowledge of the encoded third excited state of one qubit. Specifically, it relies on
quantum state. In the three-qubit code, the subspace the unique capability among computational states (σz
of the two additional “ancilla” qubits uniquely encodes eigenstates) of |111i (the notation |abci refers to the ex-
which of the the four possible single-qubit errors has oc- citation level of Q1 -Q3 , respectively) to interact with the
curred, including the possibility of no flip. Critically, er- non-computational state |003i. As the direct interaction
rors consisting of finite rotations can also be corrected by of these states is first-order prohibited, we first transfer
projecting this syndrome, essentially forcing the system the quantum amplitude of |111i to the intermediate state
to “decide” if a full phase- or bit-flip occurred2 . Previ- |102i, which itself couples strongly to |003i. Calculated
ous works implementing error correcting codes in liquid- energy levels and time-domain data showing interaction
 2

a between |011i and |002i (which is identical to |111i and

|002 |102i except for a 6 GHz offset) as a function of the flux
bias on Q2 are shown in Fig. 1(a). Once the amplitude
|100 ⊗ |1 of |111i is transferred to |102i with a sudden swap inter-

|011 0 action, three-qubit phase is acquired by moving Q1 up in
|02 frequency adiabatically, near the avoided crossing with
|003i. Figure 1(b) shows the avoided crossing between
these states as a function of the flux bias on Q1 . This
crossing shifts the frequency of |102i relative to the sum
of |100i and |002i to yield our three-qubit phase. The
detailed procedure of the gate is shown in Fig. 2(a), tak-
ing a total of 63 ns. Further details can be found in the
Supplementary Information.
We first demonstrate the gate by measuring its classical
(mΦ0 ) action. The controlled-controlled-phase (CCPhase) gate,
which maps |111i to −|111i, has no effect on pure com-
b 1 |1
⊗| |012 ⊗ putational states so we implement a CCNot gate by con-
|110 0
|021 |20 catenating pre- and post-rotations on Q2 , as described
|003
in the Supplementary Information. Such a gate ideally
|030 |111 swaps |101i and |111i and does nothing to the remaining
 states. To verify this, we prepare the eight computational
|102 |120
 |20
1
0 0 states, perform the gate, and measure its output with
|21 |30
three-qubit state tomography3 to generate the classical
truth table. The intended state is reached with 85 ± 1%
fidelity on average. This measurement is only sensitive
to classical action, however, and a more thorough set of
measurements is needed to fully characterize the gate.
To complete our verification, we perform full quan-
tum process tomography (QPT) on the CCPhase gate.
In addition to detecting the action of the gate on quan-
(mΦ0 ) tum superpositions of computational states, QPT also
detects non-unitary time evolution due to spurious cou-
FIG. 1. Calculated energy spectra and time domain pling to the environment. It is done by preparing 64
measurements of the interactions used in the three- input states which span the computational Hilbert space
qubit gate. (a) The energy spectrum of doubly excited states and performing state tomography on the result of the
showing the avoided crossing between |011i and |002i (identi- gate’s action on each state. As shown in Fig. 3, the fi-
cal to that between |111i and |102i except for a 6 GHz offset)
delity is found to be 78 ± 1% to a process in which the
is shown with both (top) a numerical diagonalization of the
system Hamiltonian and (bottom) a time-domain measure- spurious two-qubit phase between Q1 and Q3 is set to the
ment as a function of the flux bias on Q2 . (top) The frequen- measured value of 57 degrees (see the Supplementary In-
cies for the involved eigenstates are blue and non-interacting formation for details on this phase and an explanation of
eigenstates of similar energy are grey. The notation |abci⊗|di why it is irrelevant here). Due to this extraneous phase,
indicates the excitation level of each qubit and the cavity pho- the phase gate is most accurately described as a CC-eiφ Z
ton number, respectively. When omitted, d = 0. (bottom) gate (Z is a Pauli operator2 ). The infidelity is consistent
The state |011i is prepared and a square flux pulse of dura-
with the expected energy relaxation of the three qubits
tion t and amplitude V2 is applied. Coherent oscillations pro-
duce a “chevron” pattern, with darker colors corresponding during the 85 ns measurement, with some remaining er-
to population left in |002i. (b) The spectrum of triply excited ror owing to qubit transition frequency drift during the
states showing the avoided crossing between |102i and |003i 90 minutes it takes to collect the full dataset.
as a function of the flux bias on Q1 is characterized in the With our CCPhase gate in hand, we now demonstrate
same way as above. |102i is prepared by first making |111i three-qubit error correction. Both the phase- and bit-flip
and then performing the swap as described in Fig. 2. Many
codes begin by encoding the quantum state to be pro-
additional eigenstates are close in energy but are irrelevant
because they do not interact with the populated states. A tected in a three-qubit entangled state2 by using condi-
large avoided crossing between the relevant eigenstates that tional phase (CPhase) gates, as shown in Fig. 4(a). The
is used to produce an adiabatic three-qubit interaction hap- two codes differ only by single-qubit gates applied after
pens near 28 mΦ0 . Extra lines near 31 mΦ0 and 29 mΦ0 are entanglement in the encoding step. For quantum states
due to third-order interactions predicted by the Hamiltonian, on the equator of the Bloch sphere, the resulting encod-
as is the larger first-order interaction at 25 mΦ0 , but their ing is a maximally entangled three-qubit GHZ state3,4,26
effect on the protocol in Fig. 2 is negligible.
 3

a which we independently measure to have a state fidelity

of 89 ± 1%. Once the state is encoded, a single error
of the chosen type on any of the qubits can be detected
and corrected. The error syndrome is decoded by revers-
ing the encoding sequence, leaving the ancilla qubits (Q1
and Q3 ) in a state indicating which error occurred. For
a full flip, they will be in a computational state. In par-
ticular, both ancillas will be excited if the primary qubit
(Q2 ) was flipped, and so the application of the CCNot
gate will correct it. As detailed in the Supplementary
b Information for the case of bit-flip errors, an arbitrary
1.0
rotation on any single qubit about the protected axis can
be also encoded, detected, and reversed.
|ψout |Ô|ψin |2

In real physical systems, errors will occur at approx-

imately the same rate on all constituent qubits. The
0.5
correction scheme will succeed, therefore, when the sys-
tem projects to zero or one errors. The probability of
more than one error occurring is 3p2 − 2p3 , where p is
0 the single-qubit error rate2 , and so the fidelity of error
000 correction should be 1−3p2 +2p3 . For a scheme with gate
010 fidelity limited by decoherence, these coefficients will be
100 1
10 11
Inp 110 01 1 smaller, but crucially, any linear dependence on p will
ut 1 1
sta 00 11 0
11 01 be strongly suppressed. As shown in Fig. 4(b), we mea-
te 0 1 0 te
101 01 00 t sta
111 00 0 O utpu sure the process fidelity of the phase-flip error correction
0
scheme as a function of p by encoding four states which
span the single-qubit Hilbert space and performing state
FIG. 2. Three-qubit gate pulse sequence and classi-
cal action. (a) The frequency of the three qubits during the
tomography on the procedure’s output. We compare this
gate as a function of time. First, Q2 is moved suddenly into to the case of no error correction in which identical single-
resonance with the avoided crossing shown in Fig. 1(a) to qubit rotations are applied to Q2 but the ancillas are not
coherently transfer the population of |111i to |102i (and also involved (and with appropriate delays to have the same
|011i to |002i) in 7 ns. Fine adjustments in the first point of total procedure duration). Whereas without error cor-
the pulse compensates for finite pulse rise time and temporal rection we find a purely linear dependence on p, with
precision. Q2 is then moved suddenly further up in frequency, the correction applied the data is extremely well mod-
to where its two-qubit phase with Q3 is cancelled during the
gate by accumulating a multiple of 2π. Q1 is then moved
eled by only quadratic and cubic terms, demonstrating
up adiabatically to initiate the interaction between |102i and the desired first-order insensitivity to errors.
|003i. The duration and amplitude of this pulse is tuned to We have realized both bit- and phase-flip error cor-
acquire a three-qubit phase of exactly π. The population in rection in a superconducting circuit. In doing so, we
|102i is then transferred back to |111i by reversing the swap have tested both major conceptual components of the
procedure. Finally, the two-qubit phase between Q1 and Q2 nine-qubit Shor code1 , which can protect from arbitrary
is cancelled with an additional adiabatic interaction, which is
single-qubit errors by concatenating the bit- and phase-
sped up with a π pulse on Q2 at 37 ns. Here, this π pulse
is explicitly undone after the gate, but when it is used for flip codes. The implementation relies on our efficient
error correction the following pulse is compiled together with three-qubit gate which employs non-computational states
other post-rotations. The two-qubit phase between Q1 and in the third excitation manifold of our system, demon-
Q3 is uncontrolled, making this a CC-eiφ Z gate. (b) A CC- strating that the simple Hamiltonian of the system accu-
Not gate is made by appending to the phase gate pre- and rately predicts the dynamics even at these high excita-
post-rotations on Q2 as described in the Supplementary In- tion levels. The gate takes approximately half the time
formation. Its classical action is measured by preparing the
of an equivalent construction with one- and two-qubit
eight computational basis states and performing state tomog-
raphy on the result of applying the gate to them. The projec- gates. We expect it to work between any three nearest-
tion of these measurements with the computational basis is neighbor qubits in frequency regardless of the number
taken to generate the truth table and is plotted. The fidelity of qubits sharing the bus, as interactions involving other
to the expected action, where only the states |101i and |111i qubits will be first-order prohibited.
are swapped, is 85 ± 1%. We thank G. Kirchmair, M. Mirrahimi, I. Chuang,
and M. Devoret for helpful discussions. We acknowl-
edge support from LPS/NSA under ARO Contract No.
W911NF-09-1-0514 and from the NSF under Grants No.
DMR-0653377 and No. DMR-1004406. Additional sup-
 4

a b 0.6
0.6
0.5 0.5
0.4 0.4
|χthry |

|χexpt |
0.3 0.3
0.2 0.2
0.1 0.1
0 0
III III
YII YII
Z IXZ
IXZ YX ZZ YX
ZIY ZX Z ZIY ZX Z
Y I X Y
YYI YY
XY
X YY
YY Y X
XYX Z I XYX Z I
ZXY IX IY ZXY IX IY
Z Y Z XZ Y Z
YX I II Y I II
ZZZ II ZZZ II

FIG. 3. Quantum process tomography of the three-qubit phase gate. Absolute values of the elements of the (a) ideal
and (b) measured process matrices. Data is collected by preparing 64 input states which span the three-qubit Hilbert space,
applying the phase gate to them, and measuring the resulting density matrix with state tomography. The process matrix χ
P N
of the operator O is related to these data by ρout = O(ρin ) = 4m,n=1 χmn Am ρin A†n , where ρin is the density matrix of the
input state, ρout is the measured output, Ai is an operator basis spanning the three-qubit operator space, here chosen to be
the tensor products of three Pauli matrices, and N = 3 qubits2 . The operator basis is ordered as in Ref. 3 and is explicitly
written in the Supplementary Information. The ideal nonzero bars along the left edge are III, IIZ, IZI, ZII, IZZ, ZIZ, ZZI, and
ZZZ. The fidelity of the operation f = Tr[χexpt χthry ] = 78 ± 1%. The ideal process matrix is calculated with the uncorrected
phase between Q1 and Q3 set to its measured value of 57 degrees, which is irrelevant for our implementation of quantum error
correction because the ancilla qubits would be reset to their ground state25 for a repeated cycle of correction. The fidelity to
the “true” CCPhase gate, where the Q1 -Q3 phase is set to 0, is 69 ± 1%.

port provided by CNR-Istituto di Cibernetica, Pozzuoli, and ECq /h ≈ 330 MHz. The measured qubit lifetimes
Italy (LF) and the Swiss NSF (SEN). for Q1 -Q3 are T1 = (1.3, 0.9, 0.7) µs and coherence times
T2∗ = (0.5, 0.6, 1.3) µs respectively.

METHODS

Qubit rotations and gate calibration

Hamiltonian parameters

The Tavis-Cummings Hamiltonian describing our sys- Arbitrary qubit rotations around the x- and y-axis of
tem with four transmon qubits is the Bloch sphere are performed with pulse-shaped reso-
nant microwave tones. Rotations around the z-axis are
H = ~ωc a† a + done by rotating the reference phase of subsequent x and
4 X
X N N
X  y pulses. One-qubit dynamical phases resulting from flux
(q) (q)
~ ω0j |jiq hj|q + (a + a† ) gjk |jiq hk|q . excursions are measured with modified Ramsey exper-
q=1 j=0 j,k=0 iments comparing the phase difference between an un-
Here, ~ is Planck’s reduced constant, ωc is the bare cav- modified prepared state and that same state after a flux
(q)
ity frequency, ω0j is the transition frequency for trans- pulse and are cancelled with z rotations. Two- and three-
(q) qubit phases are measured with a similar Ramsey exper-
mon q from ground to excited state j, and gjk = gq njk , iment comparing the phase difference acquired when a
with gq a bare qubit-cavity coupling and njk a cou- control qubit is in its ground and excited state. For ex-
(q)
pling matrix element. ω0j and njk depend on trans- ample, the two-qubit phase between Q2 and Q3 is mea-
mon charging (ECq ) and Josephson (EJq ) energies27 . sured by preparing Q3 along the y-axis and Q2 either in
max
Flux dependence comes from EJq = EJq |cos(πΦq /Φ0 )|, its ground or excited state and then performing the flux
with Φq the flux through the transmon SQUID loop pulse in both cases. The single-qubit phase of Q3 is the
and Φ0 is thePflux quantum. A linear flux-voltage re- same for both states, and so the two-qubit phase is di-
4
lation Φq = i=1 αqi Vi + Φq,0 describes crosstalk and rectly measurable as their phase difference. All phases
offsets. Spectroscopy and transmission data as a func- are initially tuned to within one degree, limited by the
max
tion of flux bias gives ωc /2π = 9.070 GHz, EJq /h = resolution of control equipment and drifts of system pa-
{33, 35, 26, 57} GHz (from Q1 to Q4 ), gq /2π ≈ 220 MHz, rameters such as the qubit transition frequencies.
 5

a
π/2

Phase flip errors

|0 Rxπ/2 R−x

Tomography
111
11 11
|ψ Rxπ/2 π/2
R−x Rxπ/2 R−x
π/2 1. Shor, P. W. Scheme for reducing decoherence in quan-
11 11 tum computer memory. Phys. Rev. A 52, R2493–R2496
|0 Rxπ/2 π/2
R−x (1995).
2. Nielsen, M. A. & Chuang, I. L. Quantum computation and
quantum information. Cambridge Series on Information
b and the Natural Sciences (Cambridge University Press,
2000).
3. DiCarlo, L. et al. Preparation and measurement of three-
|+Z qubit entanglement in a superconducting circuit. Nature
|+X 467, 574–578 (2010).
|+Y  4. Neeley, M. et al. Generation of three-qubit entangled
|−Z
states using superconducting phase qubits. Nature 467,
570–573 (2010).
5. Paik, H. et al. How coherent are Josephson junctions?
arXiv:1105.4652 (2011).
6. Kim, Z. et al. Decoupling a Cooper-Pair Box to Enhance
the Lifetime to 0.2 ms. Phys. Rev. Lett. 106 (2011).
7. Cory, D. et al. Experimental Quantum Error Correction.
Phys. Rev. Lett. 81, 2152–2155 (1998).
8. Knill, E., Laflamme, R., Martinez, R. & Negrevergne,
C. Benchmarking Quantum Computers: The Five-Qubit
(p = sin2 (θ/2)) Error Correcting Code. Phys. Rev. Lett. 86, 5811–5814
(2001).
FIG. 4. Three-qubit phase-flip error correction 9. Boulant, N., Viola, L., Fortunato, E. & Cory, D. Ex-
scheme and demonstration of first-order insensitivity perimental Implementation of a Concatenated Quantum
to errors. (a) The error correction protocol starts by entan- Error-Correcting Code. Phys. Rev. Lett. 94 (2005).
gling the two ancilla qubits with the primary qubit through 10. Moussa, O., Baugh, J., Ryan, C. A. & Laflamme, R.
the use of two CPhase gates (vertical lines terminating in solid Demonstration of sufficient control for two rounds of
circles). The number adjacent to each indicates which state quantum error correction in a solid state ensemble quan-
receives a phase shift. A π/2 rotation on the primary qubit tum information processor. arXiv:1108.4842 (2011).
is then performed, making this a phase-flip error correction 11. Chiaverini, J. et al. Realization of quantum error correc-
code. If we wished to protect from bit flips, the two ancilla tion. Nature 432, 602–605 (2004).
qubits would instead be rotated2 . We perform errors on all 12. Schindler, P. et al. Experimental Repetitive Quantum
three qubits simultaneously with z-gates of known rotation Error Correction. Science 332, 1059–1061 (2011).
angle, which is equivalent to phase-flip errors with probability 13. Shor, P. W. Fault-tolerant quantum computation.
p = sin2 (θ/2). The encoding is then reversed, leaving the an- arXiv:9605011 (1996).
cillas in a state indicating which single-qubit error occurred. If 14. Shor, P. W. Polynomial-Time Algorithms for Prime Fac-
an error has occurred on the primary qubit, the CCNot gate torization and Discrete Logarithms on a Quantum Com-
implemented with our CCPhase gate (represented by three puter. SIAM J. Sci. Statist. Comput. 26, 1484 (1995).
solid circles linked by a vertical line) at the end of the code 15. Lanyon, B. P. et al. Simplifying quantum logic using
will reverse it. We then perform three-qubit state tomogra- higher-dimensional Hilbert spaces. Nature Phys. 5, 134–
phy to measure the result. (b) The fidelity of the protected 140 (2008).
qubit process matrix to the identity operation is plotted as 16. Monz, T. et al. Realization of the quantum Toffoli gate
a function of p. As the code corrects only single-qubit er- with trapped ions. Phys. Rev. Lett. 102, 040501– (2009).
rors, it will fail if more than one error occurs, which happens 17. Fedorov, A., Steffen, L., Baur, M. & Wallraff, A. Im-
with probability 3p2 − 2p3 . These coefficients are reduced for plementation of a Toffoli Gate with Superconducting Cir-
processes with finite fidelity. The process fidelity is fit with cuits. arXiv:1108.3966 (2011).
f = (0.76 ± 0.005) − (1.46 ± 0.03)p2 + (0.72 ± 0.03)p3 . If a 18. Mariantoni, M. et al. Implementing the quantum von
linear term is allowed, its best-fit coefficient is 0.03±0.06. We neumann architecture with superconducting circuits. Sci-
compare this to the case of no error correction to simulate the ence (2011).
improvement. (insets) The constituent state fidelities of the 19. Wallraff, A. et al. Strong coupling of a single photon to
four basis states used to produce the process fidelity data for a superconducting qubit using circuit quantum electrody-
the case of (right) error correction and (left) no correction. namics. Nature 431, 162–167 (2004).
The state |+Y i is immune to errors because its encoded state 20. Schreier, J. A. et al. Suppressing charge noise decoher-
is an eigenvector of two-qubit phase flips. ence in superconducting charge qubits. Phys. Rev. B 77,
180502 (2008).
21. Majer, J. et al. Coupling superconducting qubits via a
cavity bus. Nature 449, 443–447 (2007).
22. DiCarlo, L. et al. Demonstration of two-qubit algorithms
with a superconducting quantum processor. Nature 460,
240–244 (2009).
 6

23. Reed, M. et al. High-Fidelity Readout in Circuit Quan- 203110 (2010).

tum Electrodynamics Using the Jaynes-Cummings Non- 26. Greenberger, H. M. Z. A., D. M. In Kafatos, M. (ed.)
linearity. Phys. Rev. Lett. 105, 173601 (2010). Bell’s Theorem, Quantum Theory and Conceptions of the
24. Strauch, F. et al. Quantum Logic Gates for Coupled Su- Universe (Kluwer Academic, 1989).
perconducting Phase Qubits. Phys. Rev. Lett. 91, 167005 27. Koch, J. et al. Charge-insensitive qubit design derived
(2003). from the Cooper pair box. Phys. Rev. A 76, 042319
25. Reed, M. D. et al. Fast reset and suppressing spontaneous (2007).
emission of a superconducting qubit. Appl. Phys. Lett. 96,
 Supplementary Information for “Realization of Three-Qubit Quantum Error
Correction with Superconducting Circuits”

M. D. Reed,1 L. DiCarlo,2 S. E. Nigg,1 L. Sun,1 L. Frunzio,1 S. M. Girvin,1 and R. J. Schoelkopf1

1
Departments of Physics and Applied Physics, Yale University, New Haven, Connecticut 06520, USA
2
Kavli Institute of Nanoscience, Delft University of Technology, Delft, The Netherlands
(Dated: September 26, 2011)

DETAILS OF THREE-QUBIT PHASE GATE The three-qubit phase of our CCPhase gate arises from
an interaction between |111i and |003i, in analogy to the
two-qubit case. In the same way that the CPhase gate
To understand the physical mechanism behind our requires two excited qubits to access |02i, the CCPhase
three-qubit gate, it is useful to first review how two-
arXiv:1109.4948v1 [quant-ph] 22 Sep 2011

gate requires three excited qubits to access |003i. The

qubit CPhase gates are commonly implemented in the direct interaction between these states is first-order pro-
cQED architecture. The avoided crossing between the hibited because the states differ by a change of two exci-
first excited state of two transmon qubits (|11i) and the tations, so we achieve an effective interaction by using a
second excited state of one (|02i) can be employed both state which couples strongly to both: |102i. As described
adiabatically1 or suddenly2,3 to produce a CPhase gate. in the main text, the gate is initiated with a full coherent
If the system is adiabatically tuned to the vicinity of transfer of the population of |111i to |102i (and |011i to
the avoided crossing, the system will remain fully in the |002i) by suddenly approaching the crossing and oscillat-
eigenstate that maps to the computational subspace away ing between the two dressed eigenstates for exactly half
from the crossing. However, the interaction causes the the splitting period before moving the qubit suddenly
energy level of the |11i state to differ from the sum of further up in frequency. Small errors due to imperfect
the |01i and |10i states, advancing the quantum phase timing and finite pulse rise time are corrected by shaping
of |11i relative to the quantum phases of |10i and |01i, the qubit’s trajectory to minimize residual |002i popula-
and entangling the qubits1 . Alternatively, if the avoided tion after swapping into the state and back. Thanks to
crossing is small enough, the gate can be executed sud- the strong coupling (67 MHz) of these states, the trans-
denly. At the avoided crossing, the eigenbasis consists of fer takes only 7 ns. Once the amplitude of the target
the symmetric (|+i) and antisymmetric (|−i) superpo- state is transferred to |102i, the three-qubit phase is ac-
sitions of |11i and |02i. Starting from |11i and moving quired by moving Q1 up in frequency, near the avoided
suddenly, the wavefunction will no longer be in an en- crossing with |003i. The interaction strength between
ergy eigenstate, but rather an equal superposition of |+i these states is large (121 MHz) so we choose to acquire
and |−i whose relative phase will advance with the inter- this three-qubit phase adiabatically. As in the case of
action strength, oscillating between the undressed states two qubits described above, |102i experiences a frequency
|11i and |02i. Waiting one full period will return the state shift relative to its constituents (|100i plus |002i) due to
to the computational basis, but with a phase difference the avoided crossing, yielding the three-qubit phase.
of π, at which point it can be moved suddenly away from
Again extending the two-qubit case, a three-qubit
the avoided crossing to return to the undressed eigenba-
phase gate can be parametrized with seven unique
sis.
phases. Three are one-qubit phases (φ001 , φ010 , and
Any two-qubit phase gate can be described in terms φ100 ), three are two-qubit phases (φ011 , φ101 , and φ110 ),
of one- and two-qubit phases, which is a helpful ab- and one is a three-qubit phase (φ111 ). Ideally, our gate
straction to understand the different sources of phase procedure would provide full control over them all, but
delay. In this language, a two-qubit phase gate maps we make a simplifying assumption based on the intended
|00i → |00i, |10i → eiφ10 |10i, |01i → eiφ01 |01i, and application of the gate. During error correction, errors
|11i → ei(φ10 +φ01 +φ11 ) |11i. Here φ10 and φ01 are single- are transferred from the protected qubit to the ancil-
qubit phases given by the time-integrated detuning of a las, which are then reset either through measurement-
qubit transition frequency from its nominal bias point, conditioned pulses or by coupling them to a dissipative
and φ11 is a two-qubit phase which can be generated as bath5 . As their final state does not matter, we are free to
described above. Note that the state |11i suffers phases choose any one two-qubit phase to remain uncorrected:
from all of these sources, and so measuring the two-qubit here, the non-nearest neighbor interaction given by φ101 ,
interaction can only be done with separate measurements which is most challenging to control. This implies that
of φ01 and φ10 . Combining an interaction generating Q2 will be the target of our error correction scheme, with
φ11 = π with appropriate one-qubit phases can yield a Q1 and Q3 acting as the ancillas. The remaining two-
CPhase gate conditioned on any target computational qubit phases φ011 and φ110 , however, must be set to zero.
state1 . The former can be easily corrected via fine-tuning of the
 2

|000! → |000! |001! → |001!

|010! → |010! |011! → |011!

Average measurement

Average measurement
|100! → |100! |101! → |111!

|110! → |110! |111! → |101!

I I I I I I XY Z I I I I I I I I I XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z I I I I I I XY Z I I I I I I I I I XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z XY Z
I I I XY Z I I I XY Z XY Z XY Z I I I I I I I I I XXXYYY Z Z Z XXXYYY Z Z Z XXXYYY Z Z Z XXXYYY Z Z Z I I I XY Z I I I XY Z XY Z XY Z I I I I I I I I I XXXYYY Z Z Z XXXYYY Z Z Z XXXYYY Z Z Z XXXYYY Z Z Z
XY Z I I I I I I XXXYYY Z Z Z XXXYYY Z Z Z I I I I I I I I I XXXXXXXXXYYYYYYYYY Z Z Z Z Z Z Z Z Z XY Z I I I I I I XXXYYY Z Z Z XXXYYY Z Z Z I I I I I I I I I XXXXXXXXXYYYYYYYYY Z Z Z Z Z Z Z Z Z
Pauli operator Pauli operator

FIG. S1. Reconstructed density matrices of the result of applying the CCNot gate to computational states.
Each computational state is prepared and the CCNot gate described below is acted on it. Ideally, all states would have
nothing happen to them except for |101i and |111i, which swap. The computational state prepared and what it should map
to is indicated on each tomogram, which is visualized as the Pauli set as found in Ref. 2. The Pauli set of the ideal state is
superposed (open bars) and is reached with fidelity 95.4, 95.2, 87.1, 85.2, 84.4, 82.2, 78.1, and 75.7% for each of the states,
respectively. Note that the lack of spurious two- and three-qubit correlations indicates that there is no significant loss of
population from the computational subspace.

frequency of Q2 while in the |x02i state (x = 0, 1), where very sensitive to even phase errors, however, and so true
it acquires that phase very rapidly and so can be made to quantum process fidelity should be used whenever pos-
be an integer multiple of 2π. The latter phase, however, sible. For example, the phase fidelity of an ideal (e.g.
must be explicitly corrected with an additional adiabatic decoherence-free) CCPhase gate with a maximal single
phase gate. Because this angle is small, it is advanta- qubit phase error of φ100 = π to a CCPhase in which
geous to π-pulse Q2 prior to the adiabatic interaction, φ100 = 0 is 85.7% despite the fact that the quantum pro-
reversing the direction of phase evolution and reducing cess fidelity between those gates reveals them to be nearly
the overall correction time. The action of the gate is to orthogonal, with a fidelity of 6.3%. The real and imagi-
set φ100 , φ010 , φ001 , φ110 , and φ011 to zero, φ111 to π, nary parts of the measured process matrix are shown in
with φ101 measured to be 57 degrees. Fig. S2.
A CCNot gate is constructed by appending a π/2 and
a −π/2 pulse on Q2 before and after the phase gate de-
scribed in Fig. 2(a) of the main text. For the two input BIT-FLIP ERROR CORRECTION
states where both ancillas (Q1 and Q3 ) are excited, there
is an effective π phase shift applied to the second pulse, In order to illuminate the error projection process, we
and so the two pulses add together to a full rotation. demonstrate bit-flip error correction with errors on only
Other states are not shifted so the pulses cancel. The one qubit at a time. As in the phase-flip case described
full state tomograms of the output of CCNot with the in the main text, bit-flip correction begins by encoding
eight computational states as input are shown in Fig. S1. a quantum state in a three-qubit entangled state7 using
These tomograms are used to derive the classical truth two sudden CPhase gates paired with appropriate single-
table shown in Fig. 2(b) of the main text. qubit rotations2 . The difference between the phase- and
All relevant phases are controlled to one degree or bet- bit-flip codes lies only in the rotations performed after
ter, with their accuracy set by the voltage resolution of the entanglement. Instead of rotating the primary qubit
our arbitrary waveform generator. This implies a “quan- as in phase-flip correction, the ancillas are π/2 pulsed.
tum phase fidelity” as defined in a recent work6 in excess For simplicity, here we only measure state tomography
of 99% to the six relevant phases. This metric is not (rather than process tomography) of the state most sen-
 3

Re[χexpt ]
Re[χthry ]

III
YII Z
IXZ YX ZZ
ZIY Z Z
YYI XY XY
YY X
XYX Z I
ZXY IX IY
YX Z YI Z
I I
ZZZ II

Im[χexpt ]
Im[χthry ]

FIG. S2. Real and imaginary parts of ideal and measured process tomography χ matrix. These data were collected
as described in Fig. 3 of the main text. There only the absolute value is shown, but here the full real and imaginary parts
are reproduced. The order of operators here and in Fig 3. of the main text is as follows: III, IIX, IIY, IIZ, IXI, IYI, IZI, XII,
YII, ZII, IXX, IYX, IZX, IXY, IYY, IZY, IXZ, IYZ, IZZ, XIX, YIX, ZIX, XIY, YIY, ZIY, XIZ, YIZ, ZIZ, XXI, YXI, ZXI,
XYI, YYI, ZYI, XZI, YZI, ZZI, XXX, YXX, ZXX, XYX, YYX, ZYX, XZX, YZX, ZZX, XXY, YXY, ZXY, XYY, YYY, ZYY,
XZY, YZY, ZZY, XXZ, YXZ, ZXZ, XYZ, YYZ, ZYZ, XZZ, YZZ, and ZZZ. We do not make use of the maximum-likelihood
estimator commonly used to require the physicality of χ matrix so that the reported elements of χ and the fidelity are linearly
related to the raw measurements4 . The uncertainty of the fidelities reported in the main text is given by the standard deviation
of six repeated measurements of the full process matrix.

sitive to bit-flip errors of the chosen type: the positive cillas, this y-rotation would normally be placed between
eigenstate of σx . The state is now in a protected subspace a positive and negative π/2 x-rotation associated with
which can recover from any single spurious y rotation on turning a CPhase into a CNot, and so we compile these
any of the three qubits. three single-qubit gates into one z-gate. After the error
has occurred, we disentangle the three-qubit state, effec-
We perform intentional rotations on one of the qubits
tively encoding an error code in the ancillas. This code
with a varying rotation angle θ. In the framework of
will leave both ancillas excited if and only if a bit flip has
error correction7 , errors are decomposed as probabilis-
occurred on the primary qubit, which is then reversed by
tic full bit flips rather than continuous rotations, and
our CCNot gate. At this point, the entropy of the error
so our partial rotations can be instead seen as varying
is stored in the ancillas, which should be reset via cou-
the probability of a full bit flip. In the case of the an-
 4

a pling to a cold bath5 if we were to loop the code. In Fig.

Single bit flip err.

π/2 π/2
|0 Rxπ/2 Rxπ/2 R−x R−x S3(b), we plot the fidelity of the protected qubit to its

Tomography
111
11 11 prepared state after the error correction procedure has
|0 Ryπ/2 Rxπ/2 R−x
π/2
been applied to errors on all three qubits and also if no
11 11
π/2 π/2 error correction is done. Ideally, these curves would be
|0 Rxπ/2 Rxπ/2 R−x R−x
flat and with unit fidelity, but because of qubit decay and
b the varying excitation level of the qubits depending on
the error performed, the curves show a small oscillation
centered at 75.7% fidelity.

We also measure the density matrices of the ancilla

qubits after the four possible full bit-flip errors (no er-
ror and bit flips on each of the three qubits). The
code should ideally encode each of those four possible
errors as one of the four computational states of the two-
qubit ancilla subspace. As shown in Fig. S3(c), the
measured density matrices of the ancillas do indeed en-
code the error syndrome as expected, albeit with finite
fidelity. The measured state fidelity to the ideal syn-
−2π −π 0 π 2π drome is fsyndrome = (81.3%, 69.7%, 73.1%, 61.2%) to the
states (|00i, |10i, |01i, |11i) encoding no error and errors
c on Q3 , Q1 , and Q2 respectively. In the case of a finite
No error Error on Q3 rotation, the ancillas will instead be in a superposition
1 1
of error and no error, and so the gate will coherently
Re[ρ]

correct the primary qubit, acting only on the subspace

where the implicated error occurred. This action is free
0 0 to be done by measurement and conditional feed-forward
00
II 00II pulses (that is, by projecting the qubits with measure-
01
IZ 11 01
IZ 11 ments), however, and indeed simplifies the requirements
10 10 ZZ 10 10 ZZ
ZI ZI ZI ZI
11 01
IZ 11 01
IZ for fault-tolerance if it is done that way7 .
ZZ 00
II ZZ 00II

1 Error on Q1 1 Error on Q2
Re[ρ]

1. DiCarlo, L. et al. Demonstration of two-qubit algorithms

0 0 with a superconducting quantum processor. Nature 460,
00
II 00
II 240–244 (2009).
01
IZ 11 01
IZ 11 2. DiCarlo, L. et al. Preparation and measurement of three-
10 10 ZZ 10 10 ZZ
ZI ZI ZI ZI
11 01
IZ 11 01
IZ qubit entanglement in a superconducting circuit. Nature
ZZ 00
II ZZ 00
II 467, 574–578 (2010).
3. Strauch, F. et al. Quantum Logic Gates for Coupled Su-
FIG. S3. Bit-flip error correction. (a) Bit-flip error cor- perconducting Phase Qubits. Phys. Rev. Lett. 91, 167005
rection gate sequence. This differs from Fig. 4(a) of the (2003).
main text by only single-qubit rotations. Note also that for 4. Chow, J. et al. Detecting highly entangled states with a
simplicity here we do only state tomography on the output joint qubit readout. Phys. Rev. A 81 (2010).
of the process on the maximally affected state, so the state 5. Reed, M. D. et al. Fast reset and suppressing spontaneous
preparation is as indicated for Q2 . (b) State fidelity to the emission of a superconducting qubit. Appl. Phys. Lett. 96,
created state after performing a single error on only one of 203110 (2010).
the qubits, with and without error correction. Ideally, the 6. Mariantoni, M. et al. Implementing the quantum von
curves would be flat lines at unit fidelity. Finite excited-state neumann architecture with superconducting circuits. Sci-
lifetimes cause oscillations and displacement down as the er- ence (2011).
rors change the excitation level of the system. (c) Two-qubit 7. Nielsen, M. A. & Chuang, I. L. Quantum computation and
density matrices of the ancillas after each of the four possible quantum information. Cambridge Series on Information
errors has occurred. The fidelity of each of these states to the and the Natural Sciences (Cambridge University Press,
ideal error syndromes are (81.3%, 69.7%, 73.1%, 61.2%). 2000).