Exponential suppression of bit or phase flip errors with repetitive error correction

Google Quantum AI∗

(Dated: February 12, 2021)
Realizing the potential of quantum computing will require achieving sufficiently low logical er-
ror rates [1]. Many applications call for error rates in the 10−15 regime [2–9], but state-of-the-art
quantum platforms typically have physical error rates near 10−3 [10–14]. Quantum error correction
(QEC) [15–17] promises to bridge this divide by distributing quantum logical information across
many physical qubits so that errors can be detected and corrected. Logical errors are then expo-
nentially suppressed as the number of physical qubits grows, provided that the physical error rates
are below a certain threshold. QEC also requires that the errors are local and that performance is
maintained over many rounds of error correction, two major outstanding experimental challenges.
Here, we implement 1D repetition codes embedded in a 2D grid of superconducting qubits which
demonstrate exponential suppression of bit or phase-flip errors, reducing logical error per round by
arXiv:2102.06132v1 [quant-ph] 11 Feb 2021

more than 100× when increasing the number of qubits from 5 to 21. Crucially, this error suppres-
sion is stable over 50 rounds of error correction. We also introduce a method for analyzing error
correlations with high precision, and characterize the locality of errors in a device performing QEC
for the first time. Finally, we perform error detection using a small 2D surface code logical qubit
on the same device [18, 19], and show that the results from both 1D and 2D codes agree with
numerical simulations using a simple depolarizing error model. These findings demonstrate that
superconducting qubits are on a viable path towards fault tolerant quantum computing.

I. INTRODUCTION because typical error models for QEC do not include ef-
fects such as crosstalk errors. Moreover, exponential er-
Many quantum error correction schemes can be classi- ror suppression has never previously been demonstrated
fied as stabilizer codes [20], where a single bit of quantum with cyclic stabilizer measurements, which are a key re-
information is encoded in the joint state of many physi- quirement for fault tolerant computing but put into play
cal qubits, which we refer to as data qubits. Interspersed error mechanisms such as state leakage, heating, and data
among the data qubits are measure qubits, which period- qubit decoherence during the measurement cycle [21, 29].
ically measure the parity of chosen combinations of data In this work, we focus on two stabilizer codes. First,
qubits. These projective measurements turn undesired in the repetition code, qubits are laid out in a 1D
perturbations to the data qubit states into discrete er- chain which alternates between measure qubits and data
rors which we track by looking for changes in the parity qubits. Each measure qubit checks the parity of its two
measurements. The history of parity measurements can neighbors, and all of the measure qubits check the same
then be decoded to determine the most likely correction basis so that the logical qubit is protected from either X
for such errors. The error rate on the logical qubit is or Z errors, but not both. In the surface code [3, 30], the
determined by the error rate on the physical qubits as qubits are laid out in a 2D grid which alternates between
well as the effectiveness of decoding. If physical error measure and data qubits in a checkerboard pattern. The
rates are below a certain threshold determined by the measure qubits further alternate between X and Z types,
decoder, then the probability of logical error per round allowing for protection against both types of errors. The
of error correction (L ) should scale as: repetition code will serve as a probe for exponential error
suppression with number of qubits, while a small (d = 2)
L = C/Λ(d+1)/2 , (1) primitive of the surface code will test the forward com-
patibility of our device with larger 2D codes.
where Λ is the exponential suppression factor, C is a fit-
ting constant, and d is the code distance, which is related
to the maximum number of physical errors allowed and II. QEC WITH THE SYCAMORE PROCESSOR
increases with the number of physical qubits [3, 21].
Many previous experiments have demonstrated the We implement QEC using a Sycamore processor [31],
principles of stabilizer measurements in various platforms consisting of a 2D array of transmon qubits [32] where
such as NMR [22, 23], ion traps [24–26], and supercon- each qubit is tunably coupled to four nearest neighbors
ducting qubits [19, 21, 27, 28]. However, achieving expo- - the connectivity required for the surface code. Com-
nential error suppression in large systems is not a given, pared to Ref [31], this device has an improved design of
the readout circuit, allowing for faster readout with less
crosstalk and a factor of 2 reduction in readout error per
qubit. While this processor has 54 qubits like its prede-
∗ Corresponding author (Z. Chen):chenjimmy@[Link]; cessor, we used at most 21. Figure 1a shows the layout of
Corresponding author (J. Kelly):juliankelly@[Link] the d = 11 (21 qubit) repetition code and d = 2 (7 qubit)
 2

a measure qubit data qubit b

qubit repetition code. Using simultaneous cross-entropy

Cumulative distribution
benchmarking [31], we find that the median Pauli error
H M
for the CZ gates is 0.62% (or an average error of 0.50%).
We focused our repetition code experiments on the
repetition code:
distance 3-11
phase flip code where data qubits occupy superposition
states and are sensitive to both energy relaxation and
surface code:
distance 2
dephasing, making it more challenging to implement and
more predictive of the performance of a surface code.
stabilizers: XX ZZ XXXX A 5-qubit unit of the phase flip code is shown in Fig. 1c.
c d stabilizer measurements
0 1 1 1 0 This stabilizer circuit maps the X-basis parity of the data
initialize repeat measure change detection events qubits onto the measure qubit, which is measured then
stabilizers (XX)
reset, and this circuit is repeated in both space (across
|+⟩ H H DD H M rounds
|0⟩ H H M R
the 1D chain) and time. During measurement and reset,
e
|-⟩ H H DD H M the data qubits are dynamically decoupled to protect the
|0⟩ H H M R data qubits from various sources of dephasing [35]. In
|+⟩ H H DD H M
a single shot of the experiment, we initialize the data
80 ns 880 ns
qubits into a random string of |+i or |−i on each qubit.
H Hadamard Controlled Z Then, we repeat stabilizer measurements across the chain
M Measurement R Reset over many rounds, and finally, we measure the state of
DD Dynamical decoupling the data qubits in the X basis.
Our first pass at analyzing the experimental data is to
turn measurements into error detection events, which we
FIG. 1. Stabilizer circuits on Sycamore. a, Layout find by comparing stabilizer measurements of the same
of distance-11 repetition code and distance-2 surface code in measure qubit between adjacent measurement rounds.
the Sycamore architecture. In the experiment, the two codes We refer to each possible spacetime location of a detec-
use overlapping sets of qubits, which are offset in the figure tion event (i.e. a specific measure qubit and measurement
for clarity. b, Pauli error rates for gates and identification round) as a detection node.
error rates for measurement. All benchmarks are for simul-
In Fig. 1e, for each detection node in a 50-round, 21-
taneous operation. c, Circuit schematic for the phase flip
code. Data qubits are randomly initialized into |+i or |−i, qubit phase flip code, we plot the fraction of experi-
followed by repeated application of XX stabilizer measure- ments (76,000 total) where a detection event was ob-
ments and finally X-basis measurements of the data qubits. served on that node, or the detection event fraction.
d, Illustration of error detection events which occur when a Overall, roughly 11% of measurements signaled a de-
measurement disagrees with the previous round. e, Fraction tection event, except in the first and last round. At
of measurements which detected an error versus measurement these two time boundary rounds, detections are deter-
round for the d = 11 phase flip code. The dark line is an av- mined by comparing the first (last) stabilizer measure-
erage of the individual traces (gray lines) for each of the 10 ment with data qubit initialization (measurement). Im-
measure qubits. The first (last) round also uses data qubit portantly, the time boundary rounds are not subject to
initialization (measurement) values to identify parity errors
errors accumulated by the data qubits during measure
and generate detection events.
qubit readout, illustrating the importance of running
QEC for multiple rounds to accurately extract perfor-
mance [35]. Aside from these boundary effects, we find
surface code in the Sycamore device, while Fig. 1b sum- that the detection event fraction is stable across all 50
marizes the error rates of the components which make up rounds of the experiment, a key finding for the feasibility
the stabilizer circuits. Additionally, the typical coherence of QEC. Previous experiments had observed rising detec-
times for each qubit are T1 = 15 µs and T2 = 19 µs. tion event fractions [21], and we attribute the stability of
We note here two advancements in gate calibration. our system to our use of reset to remove leakage in every
First, we use the reset protocol introduced in Ref. [33], round [33].
which removes population from excited states (includ-
ing non-computational states) by sweeping the transmon
past the readout resonator. This reset gate is appended III. CORRELATIONS IN ERROR DETECTION
after each measurement during QEC operation, and pro- EVENTS
duces the ground state within 280 ns with a typical error
below 0.5%. Second, we implement a 26 ns controlled- We next characterize the pairwise correlations between
Z gate using a direct swap between the states |11i and detection events. A Pauli error affecting any operation in
|02i, similar to the gates described in [14, 34]. As in the repetition code should produce exactly two detections
Ref. [31], the tunable qubit-qubit couplings allow these (except at the spatial boundaries of the code) which come
CZ gates to be executed with high parallelism, and up in three flavors [21]. First, an error on a data qubit usu-
to 10 CZ gates are executed simultaneously for the 21 ally produces a detection on the two neighboring measure
 3

a c pij: full-scale pij: small-scale d

0.3 raw
9 smoothed
time
meas q (space)

i 0.2
space
j

detection event fraction

space- 0.1

measure qubit index

time 6 space spacetime
round (time) 0.0
0 40 80
crosstalk: experiment index (103)
time
b q3 and q7
9 0.3 data removed in
3
later analysis
5 6 7 8 1 0 0.2
leakage
on q2 0.1
4 3 2
0
measure data 0 0123 3 6 9 0.0
0 100 200 300
Δround measure qubit index wall clock time (ms)

FIG. 2. Analysis of error detections. a, Detection event graph. Errors in the code trigger two detections (except at the ends
of the chain), each represented by a node, and edges represent the expected correlations due to data qubit errors (spacelike and
spacetimelike) and measure qubit errors (timelike) b, Ordering of the measure qubits in the repetition code. c, Measured two
point correlations (pij ) between detection events represented as a symmetric matrix. The axes correspond to possible locations
of detection events, with major ticks marking measure qubits (space) and minor ticks marking difference in rounds (time).
For the purposes of illustration, we have averaged together the matrices for 4-round segments of the 50-round experiment
shown in Fig. 1e, and also set pij = 0 if i = j. The upper triangle shows the full scale, where only the expected spacelike and
timelike correlations are apparent. The lower triangle shows a truncated color scale, highlighting unexpected correlations due
to crosstalk and leakage. Note that crosstalk errors are still local in the 2D array. d, (Top) Observed high energy event in a
time series of repetition code runs. (Bottom) Zoom in on high energy event, showing rapid rise and exponential decay of device
wide correlated errors, and data which is removed when computing logical error probabilities.

qubits in the same round - a spacelike error. The excep- find features which do not fit the expected categories. In
tion is an error during the CZ gates, which may cause the lower triangle, we plot the same data but with the
detection events offset by 1 unit in time and space - a scale truncated by nearly an order of magnitude. The
spacetimelike error. Finally, an error on a measure qubit next most prominent correlations are spacetimelike, as
which does not propagate to a data qubit will produce we expect, but we also find two additional categories
detections in two subsequent rounds - a timelike error. of correlations. First, we observe detection correlations
These rules are represented in the planar graph shown in between non-adjacent measure qubits in the same mea-
Fig. 2a, where expected correlations are drawn as graph surement round. While these non-adjacent qubits are far
edges between detection nodes. apart in the repetition code chain, they are in fact spa-
We check how well Sycamore conforms to these ex- tially close [35] since the 1D chain is embedded in a 2D
pectations by computing the correlation probabilities be- array, which suggests that while crosstalk exists in our
tween arbitrary pairs of detection nodes. Under the as- system, it is short range. Optimization of the frequencies
sumption that all correlations are pairwise and that er- in our system already mitigates crosstalk errors to a large
ror rates are sufficiently low, the probability of simulta- extent [35, 36], but further research is required to further
neously triggering two detection nodes i and j can be suppress these errors. Second, we find excess correlations
estimated as between measurement rounds that differ by more than 1.
We attribute these long lived correlations to the presence
hxi xj i − hxi ihxj i
pij ≈ , (2) of leakage on the data qubits, which may be generated by
(1 − 2hxi i)(1 − 2hxj i) a number of sources including gates [12], measurement,
where xi = 1 if there is a detection event and xi = 0 oth- and heating [37, 38]. For the observed crosstalk and leak-
erwise, and hxi denotes an average over all experiments age errors, the excess correlations are around 3 × 10−3 ,
[35]. The numerator can be understood as the covariance an order of magnitude below the measured spacelike and
between detections in i and j, while the denominator is timelike errors but well above the noise floor of the mea-
an adjustment factor. Note that pij is symmetric between surement of 2 × 10−4 .
i and j. In Fig. 2c, we plot the correlation matrix for the Having established that on average, the errors are
data shown in Fig. 1e. In the upper triangle, we show the mostly well-behaved, we now highlight a different kind of
full scale of the data, where the only visible correlations error correlation. In Fig. 2d, we plot a time series of de-
are either spacelike or timelike, demonstrating that error tection event fractions averaged over all measure qubits
correlations in the device behave mostly as expected. for each shot of an experiment. We clearly observe a
However, the sensitivity of this technique allows us to sharp spike in the errors at a specific point in time, fol-
 4

a Time (μs)
0 9.6 19.2 28.8 38.4 48.0 gorithm to determine which errors were most likely to
0.4 have occurred given the observed detection events, and
21 qubits
correct the final measured state of the data qubits in
post-processing. A logical error occurs if the corrected
Logical error probability

0.3 5 5 final state is not equal to the initial state. We repeat the
avg. qubit decoherence experiment and analysis while varying the number of de-
5 qubits
9 qubits tection rounds from 1 to 50 with a fixed number of qubits,
13 qubits
0.2 17 qubits 21. We determine logical performance of smaller code
21 qubits
sizes by analyzing spatial subsets of the 21-qubit data,
which reduces the amount of data required [35]. These
0.1 results are shown in Fig. 3a, where we clearly observe a
decrease in the logical error probability with increasing
code size. Figure 3b plots the same data on a semilog
0.0 scale and illustrates the exponential nature of the error
0 10 20
Round
30 40 50 reduction.
b c Code distance, d
To extract logical error per round (L ), we fit the data
for each number of qubits (averaged over spatial subsets)
Logical error per round (ϵL)

3 5 7 9 11
Logical error probability

10-1 10-2
ΛZ = 2.99±0.09 to 2Perror = 1 − (1 − 2L )nrounds , which expresses an ex-
10-2 ΛX = 3.18±0.08 ponential decay in logical fidelity with number of rounds.
10-3 10-3 In Fig. 3c, we show L for the phase flip and bit flip codes
Fit:
versus qubit number. The data clearly demonstrates ex-
10-4
ϵL = C/Λ(d+1)/2 ponential suppression of logical errors, with more than
10-4
100× suppression in L from 5 qubits (L = 8.7 × 10−3 )
0 25 50 5 9 13 17 21
Round Number of qubits to 21 qubits (L = 6.7 × 10−5 ). Additionally, we fit L
vs. code distance to Eqn. 1 to extract Λ, which we plot
FIG. 3. Logical errors in the repetition code. a, in Fig. 3c. We find ΛX = 3.18 ± 0.08 for the phase flip
Logical error probability versus number of detection rounds code and ΛZ = 2.99 ± 0.09 for the bit flip code [35].
and number of qubits for the phase flip code. Smaller code
sizes are subsampled from the 21 qubit code as shown in the
inset; small dots are data from subsamples and large dots are
averages. b, Semilog plot of the averages from a showing even V. ERROR BUDGETING AND PROJECTING
spacing in log(error probability) between the code sizes. Error QEC PERFORMANCE
bars are estimated standard error from binomial sampling.
The lines are exponential fits to data for rounds greater than To better understand our repetition code results and
10. c, Logical error per round (L ) vs. number of qubits, project surface code performance on the Sycamore archi-
showing exponential suppression of error rate for both bit
tecture, we simulated our experiments using a depolariz-
and phase flip, with extracted Λ factors. The fit excludes
nqubits = 3 to reduce the influence of spatial boundary effects ing noise model, meaning that we inject a random Pauli
[35]. error (X, Y , or Z) with some probability after each op-
eration [35]. The Pauli error probability for each type
of operation is computed using averages of the data in
lowed by an exponential decay. These types of events Fig. 1b and shown in Fig. 4a. We perform two different
introduce significant correlated errors for roughly 0.5% types of simulations to compare our model to the data.
of all data taken [35], and we attribute them to high en- First, we run a direct simulation using the error rates
ergy particles such as cosmic rays striking the quantum in Fig. 4a. to obtain a value of Λ which should corre-
processor, also recently observed in Ref. [39]. For the spond to our measured values. Second, we simulate the
purposes of understanding the typical behavior of our experiment while individually sweeping operational error
system, we remove data near these events (Fig. 2d.), but rates and observing how 1/Λ changes. The relationship
note that these errors will need to be understood and between 1/Λ and the component error rates is roughly
mitigated [40, 41] for large-scale fault-tolerant comput- linear [35], and the sensitivity coefficients obtained from
ers. the second simulation allow us to estimate how much each
operation in the circuit increases 1/Λ (decreases Λ). The
resulting error budgets for the phase and bit flip codes
are shown in Fig. 4b. Overall, measured values of Λ are
IV. LOGICAL ERRORS IN THE REPETITION roughly 20% lower than simulated values, which we at-
CODE tribute to mechanisms such as the leakage and crosstalk
errors which are shown in Fig. 2c and were not included
We decode detection events and determine logical error in the simulations. Of the modeled contributions to 1/Λ,
probabilities following the procedure outlined in Ref. [21]. the dominant sources of error are from the CZ gate and
Briefly, we use a minimum weight perfect matching al- decoherence of the data qubits during measurement and
 5

a c preparations; we discard 27% of runs each round, in good

Op. Rep. Z Rep. X Surface agreement with the model prediction. Logical errors can
X data
DD 5.1e-2 4.1e-2 4.4e-2 Z data still occur after post-selection, for example with two si-

Fraction of runs kept

CZ 6.6e-3 Model multaneous errors. Following post-selection, we compute
M 1.9e-2 the logical error probability in the final measured state of
R 5.0e-3 the data qubits, shown in Fig. 4d, where we find roughly
H 1.1e-3 2 × 10−3 error probability per round [35]. The model
I 8.4e-4 6.5e-4 7.1e-4 slightly underestimates the logical error, with stray error
similar to the repetition code case, giving us confidence
b d that our surface code projections are accurate up to small
Logical error (post-selected) corrections for crosstalk and leakage.
Threshold X data
Z data
I H DD X model
Projection

H M R Z model VI. CONCLUSION AND OUTLOOK

I DD
Stray error
In this work, we show that a system with 21 supercon-
ducting qubits is stable when undergoing many repetitive
stabilizer measurement cycles. By computing the proba-
bilities of detection event pairs, we find that the physical
Rep. Z Rep. X Surface
Round
errors detected on the device are localized in space and
time to the 3 × 10−3 level. Logical errors in the repeti-
tion code are exponentially suppressed when increasing
FIG. 4. Error budgeting repetition and surface codes. the number of qubits from 5 to 21, even after 50 rounds
a, Probability of depolarizing errors (bit flip errors for M and
of operation. Finally, we corroborate experimental re-
R) for various operations in the stabilizer circuit, derived from
averaging quantities in Fig. 1b. Note the idle gate (I) and dy-
sults on both 1D and 2D codes with depolarizing model
namical decoupling (DD) values depend on the code being simulations and show that the Sycamore architecture is
run because the data qubits occupy different states. b, Esti- within a striking distance of the surface code threshold.
mated error budgets for the bit flip and phase flip codes, and Nevertheless, many challenges remain on the path to-
projected error budget for the surface code, based on the de- wards scalable quantum error correction. In the short
polarizing errors from a. The repetition code budgets slightly term, our error budgets point to the salient research di-
underestimate the experimental errors, and the discrepancy rections required to reach the surface code threshold: re-
is labeled stray error. For the surface code, the estimated 1/Λ ducing the CZ gate error, and reducing data qubit er-
corresponds to difference in L between a d = 3 and d = 5 rors during the measurement and reset cycle. Reaching
surface code. c, For the d = 2 surface code, fraction of runs
this threshold will be an important milestone in quan-
that had no detection events versus number of rounds, plotted
with the prediction from a similar error model as the repeti-
tum computing, but practical quantum computation will
tion code (dashed line). Inset: physical qubit layout of the require Λ ∼ 10 for the physical qubit overhead to be rea-
d = 2 surface code, 7 qubits embedded in a 2D array. d, sonable [35]. Achieving this performance will require sig-
Surface code logical error probability among runs with no de- nificant reductions in operational error rates, and main-
tection events versus number of rounds. Simulations from the taining a stable system over the course of a computation
same model as c (dashed lines) show good agreement. Error will require further research into mitigation of novel error
bars for c (not visible) and d are estimated standard error mechanisms such as high energy particles.
from binomial sampling with 240,000 experimental shots, mi-
nus the shots removed by post-selection in d.
VII. AUTHOR CONTRIBUTIONS

reset. In the same plot, we show the projected error Z. Chen, K. Satzinger, H. Putterman, A. Fowler, A.
budget for a surface code, where we find that overall per- Korotkov and J. Kelly designed the experiment. Z. Chen,
formance must be improved to observe error suppression K. Satzinger, and J. Kelly performed the experiment,
in a d = 5 surface code compared to d = 3. and analyzed the data. C. Quintana, K. Satzinger, A.
Finally, we test our model against a distance-2 sur- Petukhov, and Y. Chen developed the controlled-Z gate.
face code logical qubit [19]. We use seven qubits in M. McEwen, D. Kafri, A. Petukhov, and R. Barends de-
the same Sycamore device to implement one weight-4 X veloped the reset operation. M. McEwen and R. Barends
stabilizer and two weight-2 Z stabilizers as depicted in performed experiments on leakage, reset, and high en-
Fig. 1a. This encoding can detect any single error, but ergy events in error correcting codes. D. Sank and Z.
contains ambiguity in what correction corresponds to a Chen developed the readout operation. A. Dunsworth,
given detection, so we discard any runs where we observe B. Burkett, S. Demura, and A. Megrant led the design
a detection event. We show the fraction of runs where no and fabrication of the processor. J. Atalya and A. Ko-
errors were detected in Fig. 4c for both logical X and Z rotkov developed and performed the pij analysis. C.
 6

Jones developed the 1/Λ model and performed the sim- revising the manuscript and writing the supplementary
ulations. A. Fowler and C. Gidney wrote the decoder information. All authors contributed to the experimental
and interface software. S. Hong, K. Satzinger, and J. and theoretical infrastructure to enable the experiment.
Kelly developed the dynamical decoupling protocols. P.
Klimov developed error mitigation techniques based on
system frequency optimization. Z. Chen, K. Satzinger,
S. Hong, P. Klimov and J. Kelly developed error correc- VIII. DATA AVAILABILITY
tion calibration techniques. Z. Chen, K. Satzinger, and
J. Kelly wrote the manuscript. S. Boixo, V. Smelyanskiy, The data that support the plots within this paper and
Y. Chen, A. Megrant, and J. Kelly coordinated the team- other findings of this study are available from the corre-
wide error correction effort. All authors contributed to sponding authors upon reasonable request.

Google Quantum AI

Zijun Chen1 , Kevin J. Satzinger1 , Juan Atalaya1 , Alexander N. Korotkov1, 4 , Andrew Dunsworth1 , Daniel Sank1 , Chris
Quintana1 , Matt McEwen1, 5 , Rami Barends1 , Paul V. Klimov1 , Sabrina Hong1 , Cody Jones1 , Andre Petukhov1 , Dvir
Kafri1 , Sean Demura1 , Brian Burkett1 , Craig Gidney1 , Austin G. Fowler1 , Harald Putterman1, † , Igor Aleiner1 , Frank
Arute1 , Kunal Arya1 , Ryan Babbush1 , Joseph C. Bardin1, 2 , Andreas Bengtsson1 , Alexandre Bourassa1, 3 , Michael
Broughton1 , Bob B. Buckley1 , David A. Buell1 , Nicholas Bushnell1 , Benjamin Chiaro1 , Roberto Collins1 , William
Courtney1 , Alan R. Derk1 , Daniel Eppens1 , Catherine Erickson1 , Edward Farhi1 , Brooks Foxen1 , Marissa Giustina1 ,
Jonathan A. Gross1 , Matthew P. Harrigan1 , Sean D. Harrington1 , Jeremy Hilton1 , Alan Ho1 , Trent Huang1 , William J.
Huggins1 , L. B. Ioffe1 , Sergei V. Isakov1 , Evan Jeffrey1 , Zhang Jiang1 , Kostyantyn Kechedzhi1 , Seon Kim1 , Fedor Kostritsa1 ,
David Landhuis1 , Pavel Laptev1 , Erik Lucero1 , Orion Martin1 , Jarrod R. McClean1 , Trevor McCourt1 , Xiao Mi1 , Kevin
C. Miao1 , Masoud Mohseni1 , Wojciech Mruczkiewicz1 , Josh Mutus1 , Ofer Naaman1 , Matthew Neeley1 , Charles Neill1 ,
Michael Newman1 , Murphy Yuezhen Niu1 , Thomas E. O’Brien1 , Alex Opremcak1 , Eric Ostby1 , Bálint Pató1 , Nicholas
Redd1 , Pedram Roushan1 , Nicholas C. Rubin1 , Vladimir Shvarts1 , Doug Strain1 , Marco Szalay1 , Matthew D. Trevithick1 ,
Benjamin Villalonga1 , Theodore White1 , Z. Jamie Yao1 , Ping Yeh1 , Adam Zalcman1 Hartmut Neven1 , Sergio Boixo1 , Vadim
Smelyanskiy1 , Yu Chen1 , Anthony Megrant1 , Julian Kelly1

1
Google Research
2
Department of Electrical and Computer Engineering, University of Massachusetts, Amherst, MA
3
Pritzker School of Molecular Engineering, University of Chicago, Chicago, IL
4
Department of Electrical and Computer Engineering, University of California, Riverside, CA
5
Department of Physics, University of California, Santa Barbara, CA
†
Present address: AWS Center for Quantum Computing, Pasadena, CA 91125, USA (Work was done prior to joining AWS)

[1] J. Preskill, Quantum computing in the nisq era and be- [7] C. Gidney and M. Ekerå, How to factor 2048 bit RSA
yond, Quantum 2, 79 (2018). integers in 8 hours using 20 million noisy qubits, arXiv
[2] P. W. Shor, Polynomial-time algorithms for prime factor- preprint arXiv:1905.09749 (2019).
ization and discrete logarithms on a quantum computer, [8] J. Lee, D. Berry, C. Gidney, W. J. Huggins, J. R. Mc-
SIAM review 41, 303 (1999). Clean, N. Wiebe, and R. Babbush, Even more efficient
[3] A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. quantum computations of chemistry through tensor hy-
Cleland, Surface codes: Towards practical large-scale percontraction, arXiv preprint arXiv:2011.03494 (2020).
quantum computation, Physical Review A 86, 032324 [9] J. Lemieux, G. Duclos-Cianci, D. Sénéchal, and
(2012). D. Poulin, Resource estimate for quantum many-body
[4] A. M. Childs, D. Maslov, Y. Nam, N. J. Ross, and ground state preparation on a quantum computer, arXiv
Y. Su, Toward the first quantum simulation with quan- preprint arXiv:2006.04650 (2020).
tum speedup, Proceedings of the National Academy of [10] C. Ballance, T. Harty, N. Linke, M. Sepiol, and D. Lu-
Sciences 115, 9456 (2018). cas, High-fidelity quantum logic gates using trapped-ion
[5] E. Campbell, A. Khurana, and A. Montanaro, Applying hyperfine qubits, Physical Review Letters 117, 060504
quantum algorithms to constraint satisfaction problems, (2016).
Quantum 3, 167 (2019). [11] W. Huang, C. Yang, K. Chan, T. Tanttu, B. Hensen,
[6] I. D. Kivlichan, C. Gidney, D. W. Berry, N. Wiebe, R. Leon, M. Fogarty, J. Hwang, F. Hudson, K. M. Itoh,
J. McClean, W. Sun, Z. Jiang, N. Rubin, A. Fowler, et al., Fidelity benchmarks for two-qubit gates in silicon,
A. Aspuru-Guzik, et al., Improved fault-tolerant quan- Nature 569, 532 (2019).
tum simulation of condensed-phase correlated electrons [12] M. Rol, F. Battistel, F. Malinowski, C. Bultink,
via trotterization, Quantum 4, 296 (2020). B. Tarasinski, R. Vollmer, N. Haider, N. Muthusubra-
 7

manian, A. Bruno, B. Terhal, et al., Fast, high-fidelity Anderson, et al., Demonstration of the qccd trapped-
conditional-phase gate exploiting leakage interference in ion quantum computer architecture, arXiv preprint
weakly anharmonic superconducting qubits, Physical Re- arXiv:2003.01293 (2020).
view Letters 123, 120502 (2019). [30] S. B. Bravyi and A. Y. Kitaev, Quantum codes on a
[13] P. Jurcevic, A. Javadi-Abhari, L. S. Bishop, I. Lauer, lattice with boundary, arXiv preprint quant-ph/9811052
D. F. Bogorin, M. Brink, L. Capelluto, O. Günlük, (1998).
T. Itoko, N. Kanazawa, et al., Demonstration of quan- [31] F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin,
tum volume 64 on a superconducting quantum comput- R. Barends, R. Biswas, S. Boixo, F. G. Brandao, D. A.
ing system, arXiv preprint arXiv:2008.08571 (2020). Buell, et al., Quantum supremacy using a programmable
[14] B. Foxen, C. Neill, A. Dunsworth, P. Roushan, B. Chiaro, superconducting processor, Nature 574, 505 (2019).
A. Megrant, J. Kelly, Z. Chen, K. Satzinger, R. Barends, [32] J. Koch, M. Y. Terri, J. Gambetta, A. A. Houck,
et al., Demonstrating a continuous set of two-qubit gates D. Schuster, J. Majer, A. Blais, M. H. Devoret, S. M.
for near-term quantum algorithms, Physical Review Let- Girvin, and R. J. Schoelkopf, Charge-insensitive qubit
ters 125, 120504 (2020). design derived from the cooper pair box, Physical Re-
[15] P. W. Shor, Scheme for reducing decoherence in quantum view A 76, 042319 (2007).
computer memory, Physical Review A 52, R2493 (1995). [33] M. McEwen et al., Removing leakage-induced correlated
[16] A. R. Calderbank and P. W. Shor, Good quantum errors in superconducting quantum error correction, Sub-
error-correcting codes exist, Physical Review A 54, 1098 mitted (2020).
(1996). [34] Y. Sung, L. Ding, J. Braumüller, A. Vepsäläinen, B. Kan-
[17] B. M. Terhal, Quantum error correction for quantum nan, M. Kjaergaard, A. Greene, G. O. Samach, C. Mc-
memories, Reviews of Modern Physics 87, 307 (2015). Nally, D. Kim, et al., Realization of high-fidelity cz and
[18] C. Horsman, A. G. Fowler, S. Devitt, and R. Van Meter, zz-free iswap gates with a tunable coupler, arXiv preprint
Surface code quantum computing by lattice surgery, New arXiv:2011.01261 (2020).
Journal of Physics 14, 123011 (2012). [35] See Supplementary information.
[19] C. K. Andersen, A. Remm, S. Lazar, S. Krinner, [36] P. V. Klimov, J. Kelly, J. M. Martinis, and H. Neven, The
N. Lacroix, G. J. Norris, M. Gabureac, C. Eichler, and snake optimizer for learning quantum processor control
A. Wallraff, Repeated quantum error detection in a sur- parameters, arXiv preprint arXiv:2006.04594 (2020).
face code, Nature Physics 16, 875 (2020). [37] Z. Chen, J. Kelly, C. Quintana, R. Barends, B. Campbell,
[20] D. Gottesman, Stabilizer codes and quantum error cor- Y. Chen, B. Chiaro, A. Dunsworth, A. Fowler, E. Lucero,
rection, arXiv preprint quant-ph/9705052 (1997). et al., Measuring and suppressing quantum state leakage
[21] J. Kelly, R. Barends, A. G. Fowler, A. Megrant, E. Jef- in a superconducting qubit, Physical Review Letters 116,
frey, T. C. White, D. Sank, J. Y. Mutus, B. Campbell, 020501 (2016).
Y. Chen, et al., State preservation by repetitive error [38] C. J. Wood and J. M. Gambetta, Quantification and
detection in a superconducting quantum circuit, Nature characterization of leakage errors, Physical Review A 97,
519, 66 (2015). 032306 (2018).
[22] D. G. Cory, M. Price, W. Maas, E. Knill, R. Laflamme, [39] A. Vepsäläinen, A. H. Karamlou, J. L. Orrell, A. S. Do-
W. H. Zurek, T. F. Havel, and S. S. Somaroo, Experi- gra, B. Loer, F. Vasconcelos, D. K. Kim, A. J. Melville,
mental quantum error correction, Physical Review Let- B. M. Niedzielski, J. L. Yoder, et al., Impact of ioniz-
ters 81, 2152 (1998). ing radiation on superconducting qubit coherence, Na-
[23] E. Knill, R. Laflamme, R. Martinez, and C. Negrevergne, ture 584, 551 (2020).
Benchmarking quantum computers: the five-qubit error [40] K. Karatsu, A. Endo, J. Bueno, P. de Visser, R. Barends,
correcting code, Physical Review Letters 86, 5811 (2001). D. Thoen, V. Murugesan, N. Tomita, and J. Baselmans,
[24] O. Moussa, J. Baugh, C. A. Ryan, and R. Laflamme, Mitigation of cosmic ray effect on microwave kinetic in-
Demonstration of sufficient control for two rounds of ductance detector arrays, Applied Physics Letters 114,
quantum error correction in a solid state ensemble quan- 032601 (2019).
tum information processor, Physical review letters 107, [41] L. Cardani, F. Valenti, N. Casali, G. Catelani, T. Charp-
160501 (2011). entier, M. Clemenza, I. Colantoni, A. Cruciani, L. Gironi,
[25] D. Nigg, M. Mueller, E. A. Martinez, P. Schindler, L. Grünhaupt, et al., Reducing the impact of radioactiv-
M. Hennrich, T. Monz, M. A. Martin-Delgado, and ity on quantum circuits in a deep-underground facility,
R. Blatt, Quantum computations on a topologically en- arXiv preprint arXiv:2005.02286 (2020).
coded qubit, Science 345, 302 (2014). [42] C. Horsman, A. G. Fowler, S. Devitt, and R. V. Meter,
[26] L. Egan, D. M. Debroy, C. Noel, A. Risinger, D. Zhu, Surface code quantum computing by lattice surgery, New
D. Biswas, M. Newman, M. Li, K. R. Brown, M. Cetina, Journal of Physics 14, 123011 (2012).
et al., Fault-tolerant operation of a quantum error- [43] C. K. Andersen, A. Remm, S. Lazar, S. Krinner,
correction code, arXiv preprint arXiv:2009.11482 (2020). N. Lacroix, G. J. Norris, M. Gabureac, C. Eichler, and
[27] M. Takita, A. W. Cross, A. Córcoles, J. M. Chow, and A. Wallraff, Repeated quantum error detection in a sur-
J. M. Gambetta, Experimental demonstration of fault- face code, Nature Physics 16, 875 (2020).
tolerant state preparation with superconducting qubits, [44] A. Y. Kitaev, Fault-tolerant quantum computation by
Physical review letters 119, 180501 (2017). anyons, Annals of Physics 303, 2 (2003).
[28] J. R. Wootton, Benchmarking near-term devices [45] A. G. Fowler and C. Gidney, Low overhead quantum com-
with quantum error correction, arXiv preprint putation using lattice surgery (2019), arXiv:1808.06709
arXiv:2004.11037 (2020). [quant-ph].
[29] J. Pino, J. Dreiling, C. Figgatt, J. Gaebler, S. Moses, [46] D. Gottesman, Stabilizer Codes and Quantum Error Cor-
C. Baldwin, M. Foss-Feig, D. Hayes, K. Mayer, C. Ryan- rection, Ph.D. thesis, Pasadena, CA (1997).
 8

[47] S. Bravyi and A. Kitaev, Universal quantum computa- perfect quantum error correcting code, Physical review
tion with ideal clifford gates and noisy ancillas, Phys. letters 109, 100503 (2012).
Rev. A 71, 022316 (2005). [63] B. Bell, D. Herrera-Martı́, M. Tame, D. Markham,
[48] S. Anders and H. J. Briegel, Fast simulation of stabilizer W. Wadsworth, and J. Rarity, Experimental demonstra-
circuits using a graph-state representation, Phys. Rev. A tion of a graph state quantum error-correction code, Na-
73, 022334 (2006). ture communications 5, 1 (2014).
[49] R. Barends, C. M. Quintana, A. G. Petukhov, Y. Chen, [64] G. Waldherr, Y. Wang, S. Zaiser, M. Jamali, T. Schulte-
D. Kafri, K. Kechedzhi, R. Collins, O. Naaman, S. Boixo, Herbrüggen, H. Abe, T. Ohshima, J. Isoya, J. Du, P. Neu-
F. Arute, K. Arya, D. Buell, B. Burkett, Z. Chen, mann, et al., Quantum error correction in a solid-state
B. Chiaro, A. Dunsworth, B. Foxen, A. Fowler, C. Gid- hybrid spin register, Nature 506, 204 (2014).
ney, M. Giustina, R. Graff, T. Huang, E. Jeffrey, J. Kelly, [65] D. Riste, S. Poletto, M.-Z. Huang, A. Bruno, V. Vesteri-
P. V. Klimov, F. Kostritsa, D. Landhuis, E. Lucero, nen, O.-P. Saira, and L. DiCarlo, Detecting bit-flip errors
M. McEwen, A. Megrant, X. Mi, J. Mutus, M. Nee- in a logical qubit using stabilizer measurements, Nature
ley, C. Neill, E. Ostby, P. Roushan, D. Sank, K. J. communications 6, 1 (2015).
Satzinger, A. Vainsencher, T. White, J. Yao, P. Yeh, [66] A. D. Córcoles, E. Magesan, S. J. Srinivasan, A. W.
A. Zalcman, H. Neven, V. N. Smelyanskiy, and J. M. Cross, M. Steffen, J. M. Gambetta, and J. M. Chow,
Martinis, Diabatic gates for frequency-tunable supercon- Demonstration of a quantum error detection code using
ducting qubits, Phys. Rev. Lett. 123, 210501 (2019). a square lattice of four superconducting qubits, Nature
[50] A. G. Fowler, A. C. Whiteside, and L. C. Hollenberg, To- communications 6, 1 (2015).
wards practical classical processing for the surface code, [67] J. Cramer, N. Kalb, M. A. Rol, B. Hensen, M. S. Blok,
Physical review letters 108, 180501 (2012). M. Markham, D. J. Twitchen, R. Hanson, and T. H.
[51] S. Bravyi, M. Suchara, and A. Vargo, Efficient algorithms Taminiau, Repeated quantum error correction on a con-
for maximum likelihood decoding in the surface code, tinuously encoded qubit by real-time feedback, Nature
Phys. Rev. A 90, 032326 (2014). communications 7, 1 (2016).
[52] A. G. Fowler, Minimum weight perfect matching of fault- [68] N. Ofek, A. Petrenko, R. Heeres, P. Reinhold, Z. Leghtas,
tolerant topological quantum error correction in average B. Vlastakis, Y. Liu, L. Frunzio, S. Girvin, L. Jiang,
o (1) parallel time, Quantum Information & Computa- et al., Extending the lifetime of a quantum bit with error
tion 15, 145 (2015). correction in superconducting circuits, Nature 536, 441
[53] D. Szombati, A. G. Frieiro, C. Müller, T. Jones, (2016).
M. Jerger, and A. Fedorov, Quantum rifling: Protecting [69] N. M. Linke, M. Gutierrez, K. A. Landsman, C. Figgatt,
a qubit from measurement back action, Physical Review S. Debnath, K. R. Brown, and C. Monroe, Fault-tolerant
Letters 124, 070401 (2020). quantum error detection, Science advances 3, e1701074
[54] J. Bylander, S. Gustavsson, F. Yan, F. Yoshihara, (2017).
K. Harrabi, G. Fitch, D. G. Cory, Y. Nakamura, J.-S. [70] J. R. Wootton and D. Loss, Repetition code of 15 qubits,
Tsai, and W. D. Oliver, Noise spectroscopy through dy- Physical Review A 97, 052313 (2018).
namical decoupling with a superconducting flux qubit, [71] C. K. Andersen, A. Remm, S. Lazar, S. Krinner, J. Hein-
Nature Physics 7, 565 (2011). soo, J.-C. Besse, M. Gabureac, A. Wallraff, and C. Eich-
[55] H. Y. Carr and E. M. Purcell, Effects of diffusion on free ler, Entanglement stabilization using ancilla-based parity
precession in nuclear magnetic resonance experiments, detection and real-time feedback in superconducting cir-
Physical review 94, 630 (1954). cuits, npj Quantum Information 5, 1 (2019).
[56] S. Meiboom and D. Gill, Modified spin-echo method for [72] M. Gong, X. Yuan, S. Wang, Y. Wu, Y. Zhao, C. Zha,
measuring nuclear relaxation times, Review of scientific S. Li, Z. Zhang, Q. Zhao, Y. Liu, et al., Experimental
instruments 29, 688 (1958). verification of five-qubit quantum error correction with
[57] T. Gullion, D. B. Baker, and M. S. Conradi, New, superconducting qubits, arXiv preprint arXiv:1907.04507
compensated carr-purcell sequences, Journal of Magnetic (2019).
Resonance (1969) 89, 479 (1990). [73] L. Hu, Y. Ma, W. Cai, X. Mu, Y. Xu, W. Wang, Y. Wu,
[58] P. Klimov, J. Kelly, Z. Chen, M. Neeley, A. Megrant, H. Wang, Y. Song, C.-L. Zou, et al., Quantum error cor-
B. Burkett, R. Barends, K. Arya, B. Chiaro, Y. Chen, rection and universal gate set operation on a binomial
et al., Fluctuations of energy-relaxation times in super- bosonic logical qubit, Nature Physics 15, 503 (2019).
conducting qubits, Physical review letters 121, 090502 [74] C. Bultink, T. O’Brien, R. Vollmer, N. Muthusubrama-
(2018). nian, M. Beekman, M. Rol, X. Fu, B. Tarasinski, V. Os-
[59] P. Schindler, J. T. Barreiro, T. Monz, V. Nebendahl, troukh, B. Varbanov, et al., Protecting quantum entan-
D. Nigg, M. Chwalla, M. Hennrich, and R. Blatt, Ex- glement from leakage and qubit errors via repetitive par-
perimental repetitive quantum error correction, Science ity measurements, Science advances 6, eaay3050 (2020).
332, 1059 (2011). [75] Y.-H. Luo, M.-C. Chen, M. Erhard, H.-S. Zhong, D. Wu,
[60] J. Zhang, D. Gangloff, O. Moussa, and R. Laflamme, H.-Y. Tang, Q. Zhao, X.-L. Wang, K. Fujii, L. Li, et al.,
Experimental quantum error correction with high fidelity, Quantum teleportation of physical qubits into logical
Physical Review A 84, 034303 (2011). code-spaces, arXiv preprint arXiv:2009.06242 (2020).
[61] M. D. Reed, L. DiCarlo, S. E. Nigg, L. Sun, L. Frun- [76] P. Campagne-Ibarcq, A. Eickbusch, S. Touzard, E. Zalys-
zio, S. M. Girvin, and R. J. Schoelkopf, Realization of Geller, N. Frattini, V. Sivak, P. Reinhold, S. Puri,
three-qubit quantum error correction with superconduct- S. Shankar, R. Schoelkopf, et al., Quantum error cor-
ing circuits, Nature 482, 382 (2012). rection of a qubit encoded in grid states of an oscillator,
[62] J. Zhang, R. Laflamme, and D. Suter, Experimental im- Nature 584, 368 (2020).
plementation of encoded logical qubit operations in a
 9

Supplementary information for

“Exponential suppression of bit or phase flip errors with repetitive error correction”

a.
I. DATA FOR BIT FLIP CODE
0.14

In addition to the phase flip code that is primarily

detection event fraction

0.12
described in the main text, we also ran a bit flip code for
which the logical error rates are shown in Fig. 3c of the 0.10
main text. The experimental implementation of the bit
flip code is similar to the phase flip code except for the 0.08
following differences:
0.06
• Initialization and measurements are performed in
the Z basis instead of X.
0.04

• The stabilizers used are Z type instead of X type,

which means that the the data qubits do not have 0 10 20 30 40 50
round
Hadamards at the beginning and end of each stabi-
lizer round, and parity is measured in the Z basis full scale small scale
rather than X. b.
0
• We do not run dynamical decoupling pulses on the
data qubits during measurement.
• Finally, prior to measurement in every round, we measure qubit index 3
flip all of the data qubits with a π pulse to ensure
that the data qubits do not collapse into the ground
state and remain there, which would artificially re- 6
duce logical error probabilities.
In Fig. S1, we show detection fractions and two point
correlations for the 50 round bit flip code, and in Fig. S3, 9
we show the logical error probabilities for rounds 1-50 of 0 01 23 3 6 9
Δround
the bit flip code. measure qubit index

FIG. S1. a, Detection event fraction for a 50 round bit flip

II. LOGICAL ERROR PROBABILITIES code, similar to Fig. 1d of the main text. b, pij correlation
WITHOUT POST-SELECTION matrix for the 50 round bit flip code, similar to Fig. 2c of the
main text

Logical error probabilities shown in Fig. 3 of the main

text were computed while excluding device-wide corre-
lated error events which we attributed to high energy
particles. In Fig. S3, we show the fraction of data that eigenstates. Consider the case where the three stabilizer
was discarded for every number of rounds in the phase values are all +1. Then, the logical qubit exists in the
and bit flip codes, as well as the logical error probabili- two-dimensional ground state manifold of the Hamilto-
ties. To within the uncertainty from fitting, values of ΛX nian [44]
and ΛZ do not change when we do not discard data.

III. THE d = 2 SURFACE CODE H = −X0 X1 X2 X3 − Z0 Z1 − Z2 Z3 . (S1)

We implement a logical qubit in the distance-2 surface

code, the smallest non-trivial example of a surface code We can isolate specific logical states using the logical op-
logical qubit [42, 43]. The physical layout is depicted in erators ZL = Z0 Z2 and XL = X0 X1 shown in Fig. S4c.
Fig. S4a-b, consisting of a 2 × 2 array of data qubits, For example, |0L i (+1 eigenstate of ZL ) is the unique
indexed 0 to 3, subject to three stabilizer measurements ground state of H − ZL . An alternative way to identify
Z0 Z1 , X0 X1 X2 X3 , and Z2 Z3 . |0L i is to start with |ψ0 ψ1 ψ2 ψ3 i = |0000i, which is a +1
Since there are only four data qubits, it is straightfor- eigenstate of ZL and both Z stabilizers, and then project
ward to write explicit quantum states for the ZL and XL it into the X0 X1 X2 X3 = +1 subspace with the projec-
 10

a.
21 a

5 5
avg. qubit 2⋅T1
5 qubits
9 qubits
13 qubits
17 qubits
21 qubits

b. b

FIG. S2. a, Logical error probabilities vs number of detection

rounds for the bit flip code, similar to Fig. 3a of the main
text. b, Semilog plot of logical error probabilities, similar
to Fig. 3b of the main text. Lines depict fits to 2Perror =
1 − (1 − 2L )nrounds as in the main text for rounds greater
than 10.

tion operator (1 + X0 X1 X2 X3 )/2. The logical states are

√
|0L i = (|0000i + |1111i)/ 2
√
|1L i = XL |0L i = (|0011i + |1100i)/ 2
√
|+L i = (|0L i + |1L i)/ 2
√
= (|0000i + |1111i + |0011i + |1100i)/ 4
√
|−L i = (|0L i − |1L i)/ 2
√ FIG. S3. a, How much data was discarded for each run of
= (|0000i + |1111i − |0011i − |1100i)/ 4. the repetition code, in both X and Z bases b. Logical error
probabilities for the phase flip code if high energy events are
It is also possible for some stabilizer values to be −1. kept. Compare with Fig. 3b. of the main text. c, Logical
For example, if X0 X1 X2 X3 = −1 but the others√ are +1, error probabilities for the bit flip code if high energy events
then we identify |0L i = (|0000i − |1111i)/ 2, differing are kept. Compare with Fig. .
from the +1 case by Z0 (or any Zi ). Initializing to |0000i
and projectively measuring X0 X1 X2 X3 , this would be
the outcome half the time (also see Fig. S6a).
In our experiments, we explore all 8 stabilizer value
combinations, which is representative of stabilizer values
 11

a b c d a ZL = +1 b XL = +1 c XL = +1
XL = X 0 X 1 XL
A +1 ? ?
Z0 Z 1 0 0 + + + +
0 1 0 1
X0 X 1 ? +1 ?
B
X2 X 3
0 0 + + 0 0
2 3 2 3 +1 ? +1
Z2 Z 3
C 1st round: 1st round: 1st round:
ZL = Z 0 Z 2 Detect X0, X1, X2, X3 Detect Z0, Z1, Z2, Z3 Detect X2, X3
ZL

d ZL flip ZL flip
FIG. S4. Stabilizers and logical operators. a, Layout
of the distance-2 logical qubit as depicted in Fig. 1a, with X X
the data qubits labeled 0, 1, 2, 3, and the measure qubits
labeled A, B, C. b, The same logical qubit depicted in a more
standard lattice surgery surface code notation, as in Ref. [45]. X X
The Z stabilizers are light tiles (Z0 Z1 and Z2 Z3 ), and the X
stabilizer is a dark tile (X0 X1 X2 X3 ). c, The logical operators
XL = X0 X1 and ZL = Z0 Z2 , which cross at qubit 0, so e XL flip XL flip
[XL , ZL ] 6= 0. d, A distance-3 logical qubit and its logical
operators, analogous to c, with 9 data qubits and 8 stabilizers. Z Z

initialize repeat stabilizers measure

Z Z
|0⟩ H H M R
|ψ0⟩ H H DD H M
|ψ1⟩ H H DD H M
|0⟩ H H M R FIG. S6. Error detection. a, Example initialization to
|ψ2⟩ H H DD H M |0000i prior to the first round of stabilizer measurements.
|ψ3⟩ H H DD H M This is a +1 eigenstate of ZL and both Z stabilizers. In
|0⟩ H H M R the first round, any X error can be detected. However, the
160 ns 880 ns first X stabilizer measurement will be random, so no Z errors
can be detected. b, |++++i is a +1 eigenstate of XL and the
FIG. S5. Surface code quantum circuit. Quantum cir- X stabilizer. In the first round, any Z error can be detected,
cuit implementing repeated Z (green) and X (blue) stabiliz- but the two Z stabilizers will have random values. c, |++00i
ers, analogous to Fig. 1c. The stabilizer circuit is longer (four is a +1 eigenstate of XL and the lower Z stabilizer. As in a,
CZ layers) because of the weight-4 X stabilizer. For XL log- the first X stabilizer measurement will be random, so no Z
ical measurements, we include Hadamard gates on each data errors can be detected, risking a logical error XL = −1. d,
qubit prior to measurement, shown in gray; these are omitted Illustration of the detected syndrome for one X error. Note
for ZL logical measurements. X0 and X1 have the same syndrome, but X0 flips ZL while
X1 does not. X2 and X3 are similar. e, Illustration of the
detected syndrome for one Z error. All four have the same
that would be encountered by a long-lived logical qubit. syndrome, but Z0 and Z1 flip XL while Z2 and Z3 do not.
In d-e, there is an implicit decoding procedure: for flipped
In particular, we initialize the data qubits to each of the
X0 X1 X2 X3 , insert Z0 correction; for flipped Z0 Z1 , insert X0
16 possible bitstrings, such as |0111i. For experiments correction; and for flipped Z2 Z3 , insert X2 correction. When
in the logical Z basis, we proceed directly with stabilizer this correction is the wrong choice, which happens for about
measurements, and the Z stabilizers and ZL are already half of error events, we get logical errors.
well-defined (for |0111i, Z0 Z1 = −1, Z2 Z3 = +1, and
ZL = −1). The first X0 X1 X2 X3 measurement is ran-
domly ±1. For experiments in the logical X basis, we
perform Hadamards on all four data qubits before pro- We show standard Z and X initializations in Fig. S6a-b.
ceeding with the stabilizer measurements, so |0111i be- Alternatively, consider |++00i, shown in Fig. S6c, which
comes |+−−−i. Now the X stabilizer and XL are well- is employed in Ref. [43]. The first X0 X1 X2 X3 measure-
defined (for |+−−−i, X0 X1 X2 X3 = −1 and XL = −1), ment will be random, so no Z errors can be detected on
and the first Z stabilizer measurements are each ran- the first round, risking a logical error in XL . Moreover,
domly ±1. We show the specific quantum circuit for although |++00i is an eigenstate of XL = X0 X1 , it is
these experiments, analogous to Fig. 1c, in Fig. S5. not an eigenstate of XL0 = (X0 X1 X2 X3 )XL = X2 X3 , an
Note that to prepare a logical XL or ZL eigenstate, it equally valid logical operator.
is important to initialize all the data qubits in the same This encoding can detect any single error, but because
basis (X or Z) as the intended logical qubit state. Then, it is only distance-2, the code cannot be used to correct
the data qubit state is an eigenstate of all the stabilizers for errors, as shown in Fig. S6d-e. Any single error on
of the same type as the logical operator, and any errors a data qubit leads to an ambiguous syndrome, where it
of the opposite type can be detected in the first round. is unclear if a logical operator has been affected. This
 12

is distinct from the larger distance-3 logical qubit (see

Fig. S4d), where any single error can be corrected un-
ambiguously (distance-d can accommodate any (d − 1)/2
errors).
Consequently, any time we observe a detection event
in a run, we simply discard that run. As we increase the
number of rounds, we increase the probability that there
has been a detection event, so the fraction of runs we
keep decreases exponentially, as shown in Fig. 4c of the
main text. Empirically, we remove about 27% of runs
each round, which agrees well with simulations of the
experiment.
At the end of each run, we measure the data qubits in
the basis matching the logical basis of the experiment,
FIG. S7. Physical qubits per logical qubit. We estimate
either X or Z, and evaluate the appropriate logical op- the physical qubits required for one logical qubit to achieve an
erator. We identify a logical error if the logical mea- overall logical error suppression of 10−12 as a function of the
surement outcome differs from the value we initialized. inverse error suppression factor 1/Λ, marking Λ = 10 with a
By post-selecting only runs without detection events, we vertical line. Left: semi-log, right: log-log.
avoid most logical errors. However, two simultaneous er-
rors can be undetectable and lead to logical errors, such
as X0 X1 , which flips ZL . Following post-selection, the IV. QUANTIFYING LAMBDA
probability of a logical error is about 0.002 each round,
as shown in Fig. 4d. Specifically, for X basis, we ob-
serve 0.0016 ± 0.0001 error per round, and for Z basis, Accurately benchmarking the performance of quantum
0.0027±0.0001 (linear fit uncertainties). For comparison, error correction can be confounded by artifacts if exper-
in Ref. [43], about 60% of runs are removed each round, iments are not carefully designed. In particular, bound-
and the logical error probability is about 0.03 each round. ary effects can introduce different error characteristics
In Fig 4b, we project the error suppression factor Λ that must be understood. Here, we study two types of
for the surface code. Modest performance improvements boundary effects. The first is qubits at code boundaries,
will be needed to achieve Λ > 1, which would be a clear which interact with a reduced number of stabilizers and
demonstration of operating below threshold error rates, thus participate in a reduced number of entangling gates
where making the code larger makes it better (even if the and may decrease the number of physical errors present.
absolute error rate is worse than a physical qubit). How- Second, data qubits are subject to less errors in the first
ever, a practical surface code quantum computer would round of the code than in the steady-state, and data qubit
benefit from Λ ∼ 10, which vastly decreases the required measurement errors are only relevant in the final round
physical qubits per logical qubit for a given logical er- of measurements.
ror rate. For example, suppose we want an overall logi- In our analysis of the repetition code, we use the tech-
cal error suppression 1/Λ(d+1)/2 = 10−12 for a practical nique of subsampling outlined in the supplementary ma-
computation. For a given Λ, we can solve for distance d terials of [21]. In order to, for example, compare the
and estimate the required number of physical qubits per performance of a d=11 repetition code to a d=3 repeti-
logical qubit as roughly 2d2 , as shown in Fig. S7. For tion code, we take a single dataset for the d=11 code,
Λ = 10, this corresponds to roughly 1000 physical qubits perform matching analysis, then subsample this dataset
(distance-23). into a collection of d=3 datasets and perform matching
analysis on each sub-dataset. Generally, a repetition code

full dataset subsampling 1 subsampling 2 subsampling 3

Input 000000000 Input 00000 Input 00000 Input 00000
1 0 0 0 0 1 0 0 1 0 0 1 0 0
2 1 0 0 1 2 1 0 2 0 0 2 0 1
3 0 1 0 0 3 0 1 3 1 0 3 0 0
4 1 0 0 0 4 1 0 4 0 0 4 0 0
5 0 0 0 1 5 0 0 5 0 0 5 0 1
Out 0 1 0 1 0 Out 0 1 0 Out 1 0 1 Out 0 1 0
1
2
3

FIG. S8. Example of subsampling a d = 5 repetition code

dataset into 3 d = 3 repetition code datasets.
 13

TABLE S1. Pauli error rates (bit flip error rates for measure-
ment and reset) used in subsequent simulations.

Operation Error rate

H 1e-3

detection event fraction

CZ 5e-3
M 2e-3
R 5e-3
Idle (M + R) 4.4e-2
Idle (H) 7e-4

of distance ds can be subsampled from a larger code of

distance d, where n = d − ds + 1 is the number of unique
datasets one could produce. This can be understood by round
considering a line of 9 qubits (for d = 5), and uniquely
choosing a line of 5 qubits (for d = 3) along it, as shown FIG. S9. Simulated repetition code data for 10 QEC rounds
in Fig. S8. and 21 qubits. The plot shows detection event fraction as a
Subsampling has a number of practical advantages. function of round. We find a uniform behavior of detection
First and foremost, the experimental burden of acquir- fraction in the intermediate rounds, and different values at
ing data is reduced. In order to quantify the perfor- the first and last rounds of the code, which differ in circuit
mance of a distance d repetition code as well as all pos- structure.
sible configurations of smaller code distances, without
subsampling we would need to perform nexperiments =
P(d−1)/2, odd QEC, we seek to quantify the logical error rate in the
n=1 d − 2n. In the case of d = 11, subsampling steady-state, but these boundary effects indicate the er-
reduces the datasets needed by a factor of 25. Addi- ror rate is slightly different at the beginning and end of
tionally, by using only a single source dataset, we enforce the code. Due to this effect, the logical error probabili-
self-consistency in error rates between code distances and ties will deviate slightly from an exponential decay. To
reduce sensitivity to systematic errors and system drift mitigate this behavior, we choose to fit an exponential
that may occur between data acquisition runs. Alterna- decay to only experiments with a large number of rounds
tively, one could collect only a single dataset for each code (greater than 10), where this effect is minimized. This
distance. However, qubits typically have performance can be seen in Fig. S10, where in this simple model we
variations and the choice of which qubits for which code see logical error probabilities that deviate from an ex-
distance at what time will introduce bias or noise into ponential model (dashed, solid lines) at small numbers
benchmarking. of rounds. In this regime, the logical error probabilities
In order to understand boundary effects and their im- outperform the steady state and are not predictive of fu-
pact on repetition code data, we perform simulations us- ture QEC performance. This discrepancy, here up to a
ing an uncorrelated depolarized Pauli error model. Here, factor of 2, can vary depending on circuit construction
we use a simple error model described by Table S1, where and hardware.
every qubit shares identical error rates. Given these prob- In addition to time boundary effects, spatial boundary
abilities, we simulate 100,000 runs of a 21 qubit repetition effects also exist for qubits located at the edge of the
code over 10 QEC rounds. code, which participate in less entangling gates. This
We process this simulated data to explore the detec- can be seen in Fig. S11, where the measure qubits at
tion event fraction as a function of round, per qubit. We the edges of a simulated 21 qubit repetition code have
find that the first and last round deviate from the steady- lower detection event fraction. This introduces a small
state detection event round, as seen in Fig S9. This dis- but systematic difference in comparing subsampled data
crepancy comes from a difference in circuit structure as to experiments that are run in isolation.
well as initial conditions. Before initialization, all qubits
begin in the |0i state and suffer no Idling error during
the M + R operations that subsequent rounds do. In
the last round, the stabilizer outcomes are determined
from the final data qubit measurements, and require no
data qubit idling or entangling gates. These differences
manifest in smaller error rates and thus smaller detection
event fractions associated with these rounds.
This non-uniformity in detection event fraction must
be accounted for when analyzing Λ. In benchmarking
 14

including repetition codes and surface codes, the bulk

of the encoded operations consist only of gates from the
Clifford group [46]; the exception is the need to enact
logical non-Clifford gates, such as through magic-state
simulated data distillation [47], which is needed in a fault-tolerant quan-
exponential fit
tum computer but beyond the scope of logical memory
experiments like this work. A circuit composed entirely
fitted / simulated error

of Clifford gates can be simulated efficiently using the

Gottesman-Knill theorem [48], and this description in-
cludes noisy circuits where the noise is a probability dis-
tribution for randomly inserting a Pauli operator after
each gate. Moreover, for stabilizer codes [46], the sta-
bilizers are Pauli operators which can be measured by
Clifford gates, so it is convenient to represent errors as a
distribution of Pauli errors. We employ this model here
— Clifford circuits with Pauli errors — because the sim-
FIG. S10. Logical error probabilities fitted to an exponential ulations can easily scale to modeling large surface codes,
model of logical error rate (dashed, solid lines) (distance-3, such as a distance-23 surface code requiring at least 1057
repetitions = 10,000). At low rounds, we see deviations from qubits.
the exponential fit due to boundary effects at the start and
We employ circuit simulations to attempt to under-
end of the code, where error rates are reduced as compared to
the steady-state of the experiment. This can be seen in the stand the relative contributions of errors from different
lower graph, where we plot fitted error over simulated error. operations, also known as error budgeting. This proceeds
At low rounds, we find up to nearly a factor of 2 discrepancy. in two stages. First, we run simulations of the repetition
To mitigate this effect, we fit the exponential only to rounds codes with circuit-noise parameters informed by bench-
greater than 10. Similar fits can be seen in Fig. 3 of the main marking component operations, such as CZ gate error
text, and in Fig. S3 and Fig. S3. from cross-entropy benchmarking and idling qubit error
from measuring T1 and T2 . We compare the logical error
rate in the simulations with the logical errors in the ex-
periment, and see close agreement. We also discuss pos-
sible explanations for the gap between experiment and
simulation.
Second, we use simulations to estimate the relative
detection event fraction

contributions of component errors to the logical error

rate. We construct an error budget for Λ (see Eqn. (1)
of the main text) by attempting to represent its inverse
Λ−1 as a linear function of the component errors, which
we motivate by arguing that Λ−1 is approximately linear
in the component errors. For such a model, the frac-
tion budgeted to each component is simply given by the
weighted contribution of the component error, divided
by quantity Λ−1 . However, Λ−1 is not a perfectly linear
function, and we discuss our approach to dealing with
measure qubit index this. Our intent with the error budgeting is to determine
what component error rates are necessary to implement
a working demonstration of a surface code. We can fore-
FIG. S11. Detection event fraction vs measure qubit index
cast how a small surface code might perform if run on a
for a 21 qubit repetition code. Detection event fraction for
measure qubits at the edge of the code (index = 0, 9) have
device with current error rates, and we can use the er-
lower detection event fraction, as data qubits on the boundary ror budget to compare tradeoffs in component errors and
participate in fewer entangling gates. make design decisions for future devices.

V. CIRCUIT SIMULATIONS WITH PAULI A. A Description of a Component-Error Model for

NOISE Simulations

This section describes simulations that approximate We simulate the repetition and surface code experi-
errors in the experiment as Pauli errors sampled from ments in a simplified “circuit noise” model. A circuit is
probability distributions and inserted into a circuit of constructed from component operations, including Clif-
Clifford gates. In many quantum error-correcting codes, ford gates and related operations like initialization or
 15

measurement in the eigenbasis of a Pauli operator. A

TABLE S2. Error rates used in bit and phase flip simulations
circuit composed of these components can be simulated
efficiently, and this set of instructions is sufficient to im- Component Bitflip Phaseflip
plement stabilizer codes such as repetition codes and sur- DD 5.1e-2 4.1e-2
face codes.
CZ 6.6e-3 6.6e-3
Noise in the circuit is simulated by sampling random M 1.9e-2 1.9e-2
Pauli errors and inserting them into the circuit according
R 5.0e-3 5.0e-3
to the following probability model. For each component,
there is a “Pauli error channel,” which is a distribution H 1.1e-3 1.1e-3
over the possible Pauli errors to insert, including iden- I 8.4e-4 5.8e-4
tity for no error (e.g. the distribution has 4 elements
for single-qubit operation, or 16 for a two-qubit opera-
tion). For each component in the circuit, a Pauli error model, which are listed in Tab. S2. Since the error chan-
is sampled according to the distribution associated with nel on each component has a single parameter, the noise
that component, and this Pauli operator is inserted after in the simulator has six parameters. We refer to these
the component. Measurement errors are treated slightly parameters collectively as a vector denoted x, which we
differently, as follows. The binary measurement result is use to relate the component-error probabilities to per-
flipped with a probability p, i.e. it goes through a classi- formance measures of the repetition and surface codes,
cal binary symmetric channel instead of a Pauli channel. such as logical-error probability or Λ, the ratio by which
For the circuits used in this work, when a qubit is mea- logical error improves when code distance is increased by
sured, it is always reset before being used again; this 2.
means we do not assume that a measured qubit is left in
the state consistent with a measurement result, because
we unconditionally reset that qubit before using it again.
B. Comparing Component-Error Simulations to
The effect of the randomly sampled Pauli errors that the Experiments
are injected into the simulated circuit is to change some
of the measurement outcomes from their expected values.
To reproduce experimental conditions in the simplified
For example, an X (bitflip) error that occurs on a data
simulator, we try to approximate the error rate in each
qubit will be detected by the next syndrome circuits that
component with data from benchmarking of those com-
interrogate this data qubit. We collect the syndrome
ponents. The methods for characterizing error are:
measurements and final data-qubit measurements in the
simulation, and process them in the same way as the • Single- and two-qubit gates: cross-entropy
experiment using minimum-weight matching to infer a benchmarking [49], averaged over the gates used
most likely location of errors. in the experiment. Averages treat one-qubit and
Our simulations make some simplifying assumptions two-qubit gates separately.
about the Pauli error channels. First, we assume that
each use of a component of the same type (e.g. every CZ • Idle operations: modeled as memoryless depolar-
gate) has the same error channel. Of course, it would izing channel with decay time constant given the
be straightforward to simulate different error channels by relevant experiment, meaning “T1 decay” for
for each gate in the circuit. This would also be com- the bitflip code and “T2 decay” for the phaseflip
putationally efficient, but we opt to keep the number of code. T1 decay means initializing |1i and measur-
parameters in the simulation relatively small. Second, we ing probability of the state being |1i as a function of
further simplify error channels to be parameterized by a time; T2 decay meanings initializing |+i and mea-
single scalar parameter. The error channel for each gate suring decay of this state to the mixed state with
or idle is a depolarizing channel parametrized by a sin- time, while doing CPMG echoing to remove low-
gle probability p for any error to occur; for a single-qubit frequency phase noise (this dynamical decoupling
depolarizing channel, each of X, Y, or Z errors has proba- is also done during idle operations in the phaseflip
bility p/3 to occur; for a two-qubit depolarizing channel, experiments).
each of the 15 non-identity Paulis has probability p/15
to occur. Each reset operation is followed by a quantum • Reset and measurement: These errors are dif-
bitflip channel (random insertion of Pauli X), and each ficult to distinguish; measurement error presents
measurement operation is followed by a classical bitflip a noise floor for reset characterization. However,
channel (random flip of the measurement bit). All com- for simulation purposes, only the sum of the two
ponents (e.g. every CZ gate) have the same error channel, error probabilities is important. We characterize
but different components can have different error proba- reset by performing the reset gate between mea-
bilities (i.e. measurement error pm can be distinct from surement pulses, preparing the qubit in |0i or |1i;
the CZ error pCZ ). the error is the probability of finding |1i after re-
There are six types of component operations in our set. For measurement, we benchmark individual
 16

a 2
d=3 10 bitflip
0.3 d=5 bitflip fit: = 3.342
d=7

Logical error per round

phaseflip
Logical error

0.2 d=9 phaseflip fit: = 3.781

d = 11 3
10
0.1

0.0
4
10
b 1
10
2 4 6 8 10
Logical error

10 Code distance
3 d=3
10 d=5
4 d=7 FIG. S13. Logical error vs. code distance for the repetition
10 d=9 codes, and a fit to estimate Λ for the two codes.
10
5 d = 11

c 0.3 only simulating Markovian Pauli channels. The associ-

d=3 ated probability distributions are independent and iden-
d=5 tically distributed for each type of component. Other
d=7
Logical error

0.2 important physical effects that we suspect to be present

d=9 are not included in the model, such as leakage, cross-talk
d = 11 during gates, cosmic rays, parameter drift with time, or
0.1 any other non-Markovian noise source. The reason for
choosing such a limited noise model is that it scales to
large problem sizes and allows us to make forecasts of sur-
0.0
face codes. In future work, we will improve the simula-
d tions to incorporate approximations to effects like leakage
1 that are still computationally efficient at large numbers
10
of qubits.
2
Logical error

10 The simulation conditions mirror the experiments in

3 d=3 simulating bitflip and phaseflip error-correcting codes
10 d=5 with the following parameters. The values of component-
4 d=7 error probabilities are those given in the main text, Fig.
10 d=9 4a. The syndrome circuits are executed nrounds times, for
10
5 d = 11 nrounds being every integer in the range [1,50]. At each
value of nrounds , the simulation is executed M = 160,000
0 10 20 30 40 50
times. A logical error has occurred if the logical mea-
Syndrome rounds
surement at the end of an error-correction circuit gives
an encoded qubit state different from the initial encoded
FIG. S12. Simulations of logical-error probability for repeti- state. We count the number of simulated logical errors
tion codes using Pauli-channel noise calibrated to component me (nrounds ) at each value of nrounds , and the logical error
errors measured in the device. a, Logical error vs. number probability is calculated as
of syndrome rounds for the bit flip code. b, Same data as
panel a (bit flip code), plotted on a log-scaled vertical axis. Perror (nrounds ) = me (nrounds )/M. (S2)
c, Logical error vs. number of syndrome rounds for the phase
flip code. d, Same data as panel c (phase flip code), plotted For each value of code distance d ∈ {3, 5, 7, 9, 11}, we
on a log-scaled vertical axis. determine the logical error rate logical by fitting
nrounds
Perror (nrounds ) = 0.5 [1 − (1 − 2logical ) ] (S3)
qubits by preparing |0i or |1i and immediately mea-
suring, identifying the error probability. We also to the sampled data. This fitting ansatz has the prop-
benchmark simultaneous readout on all the mea- erties that Perror (nrounds = 0) = 0, it saturates as
sure qubits and all the qubits, as in Ref. [31]. Perror (nrounds → ∞) = 0.5, and the error after one round
Perror (1) = logical . As in the main text, we calculate Λ as
It is important to note that the model is limited to the ratio by which logical error improves when increasing
 17

the code distance by 2: function in its arguments, we could calculate the gradi-
ent g = ∇f anywhere to determine f exactly. However,
Λ(d) = logical (d)/logical (d + 2). (S4) numerical simulations show that this is not the case, and
the gradient changes for different choices of the point to
The simulated logical error vs. number of syndrome linearize around. Since we desire a linear model to form
rounds, and fits to this data, are shown in Fig. S12. The an error budget, we need to make a choice of how to do
simulated logical error rates match well but not perfectly so; since f (x) is not linear, there is no single “correct”
to the experimental results. Figure S13 shows the fit- answer.
ted logical error per round vs. code distance and fits Our approach is to treat f (x) as if it was a second-order
to determine Λ. The error rates are lower, and Λ val- function in its arguments,
ues are higher, than what is seen in the experiments.
We attribute this discrepancy to one of the assumptions f (x) ≈ gx + 0.5x| Hx, (S5)
of the simulator not holding in experiment. For exam-
ple, Section VI discusses evidence for cross-talk errors where g is the gradient of f , (H)ij = ∂ 2 f /∂xi ∂xj is the
happening during the experiment as well as long time Hessian matrix of f , and both are evaluated at x → 0+ .
correlations in detection events due to presence of leak- By doing so, we are saying that the second-order terms
age states in the data qubits. Another possibility is that would capture enough of the nonlinearity in f to pro-
parameter drift during the experiment leads to higher vide a good approximation in the domain of interest. We
error rates when running error correction than during then exploit the following property. For any second-order
the component benchmarking that determines the com- function f with f (0) = 0, there is a linear function given
ponent error probabilities used in the simulation. Said by the first-order Taylor series evaluated at a point a/2
another way, this method of forecasting Λ accounts for such that this linear function coincides with the second-
about 85% of the error, because it predicts Λ−1 values order function at a:
that are about 0.85 of the experimentally measured val-
∇f |x=a/2 a = ga + 0.5a| Ha = f (a)


ues, leaving weighted error contributions of about 15% of (S6)
the total not accounted for. This method was also used
To make an error budget for the experimental
to simulate the d=2 surface code, producing the “model”
component-error vector x (values in Fig. 4a of the main
traces in Fig. 4c-d of the main text.
text), we use simulations to numerically evaluate the gra-
dient of f at x/2, which determines the weights on the
error components. From the weights in this linear model,
C. Error Budgeting: Constructing a Linear Model we can produce an estimate of f = Λ−1 that shows the
Relating Component Errors to Inverse of Lambda weighted contribution of each component error. These
results are summarized in Tab. S3 and Tab. S4.
The quantity Λ is used to forecast logical error rate We see in these tables that the major source of logi-
for a quantum code of a given size, so we extend this cal error (more than 50% of the budget) is idling error
reasoning to determine what component error rates are during the measurement and reset process. This is sim-
needed to realize a target Λ value. We use the conven- ply due to T1 decay times around 15 µs and idle times
tion that Λ is the factor by which logical error is sup- (880 ns during measurement and reset), leading to an er-
pressed by increasing code size, where Λ > 1 means log- ror probability of 4–5% during each such operation. CZ
ical error decreases when code size increases. As a ra- gates and the combined effect of reset and measurement
tio, its inverse Λ−1 has the same meaning (the factor by account for most of the remaining errors, with very small
which logical error changes when code size increases one contributions from one-qubit gates and idle operations
step). Moreover, we argue that Λ−1 is approximately during gates.
a linear function of component errors. As in the main
text, we say that logical error rate is related to code dis-
tance d by logical (d) ∝ Λ−[(d+1)/2] for d odd. It has
been seen in numerical simulations with Pauli-channel
noise [50, 51] that for a single physical-error parameter p,
logical ∝ (p/pth )[(d+1)/2] , where pth is the threshold error
rate for the chosen code and error model parameterized
by p. Hence, a naive comparison of the two approximate
expressions would have Λ−1 = p/pth , meaning that Λ−1
is (approximately) linear in p.
For notational simplicity, denote the vector of compo-
nent error rates as x and let there be a function of compo-
nent error rates f (x) such that Λ−1 = f . We will assume
throughout that f (0) = 0, meaning Λ approaches ∞ in
the limit errors go to zero. If f (x) were a truly linear
 18

TABLE S3. Error Budget for bit flip code.

Component Error rate Model weight Contribution to Λ−1 Error-budget percentage

DD 5.1e-2 3.5 0.179 58%
CZ 6.6e-3 11.7 0.077 25%
M 1.9e-2 1.6 0.030 10%
R 5.0e-3 1.6 0.008 3%
H 1.1e-3 3.4 0.004 1%
I 8.4e-4 6.6 0.006 2%

TABLE S4. Error budget for phase flip code. *Note that “I” gates are assigned zero weight. The term in the gradient
of Λ−1 for this component is actually a small negative number that depends on code distance, for example about -1 for Λ
between d=3 and d=5. The reason this is negative is that “I” gates only appear on data qubits at the endpoints of the linear
chain, and not across the data qubits like the other components. This is why the derivative of Λ−1 with respect to “I”-gate
probability is negative: errors in this component affect d=3 more than d=5, and the trend continues to higher distances. For
the experimentally measured error rate in this component, it has negligible contribution to logical error and hence Λ−1 , so we
choose to set its weight to zero for the purposes of an error budget.

Component Error rate Model weight Contribution to Λ−1 Error-budget percentage

DD 4.1e-2 3.5 0.144 54%
CZ 6.6e-3 11.9 0.079 29%
M 1.9e-2 1.5 0.029 11%
R 5.0e-3 1.5 0.008 3%
H 1.1e-3 8.0 0.009 3%
I 5.8e-4 0* 0 0%

VI. PROBABILITY pij OF ERROR-PAIRED

DETECTION EVENTS

In this section, we discuss a technique that allows

dq4
mq3
B
measure qubits (s)

us to characterize error processes in repetition code ex- dq3

data qubits

periments using correlations between detection events. mq2

We refer to this technique as the pij correlation matrix dq2
mq1
S ST
method. We use it to estimate the probability pij of con-
ventional (e.g., bit or phase flips) and unconventional dq1 T
(e.g., leakage and crosstalk) error processes that produce mq0
pairs of detection events at the error graph nodes i and dq0
j. We use this technique to produce in-situ diagnostics 0 1 2 3 4 5 6 7 8
for QEC operation, and because it extracts detailed error
information, it can also inform weights to the decoder.
rounds (t)
FIG. S14. Error graph and main edges. An example of
the error graph for Nmq = 4 measure qubits (5 data qubits)
A. Error graph and correlation matrix pij and Nr = 8 time rounds. The horizontal axis shows num-
bering of rounds (t coordinate), the vertical axis shows num-
bering of measure qubits mq0–mq3 (s coordinate). The dots
Figure S14 shows an example of the error graph of a denote the graph nodes; red dots indicate detection events.
quantum bit-flip or phase-flip repetition code. It con- The vertical, horizontal and diagonal edges are denoted as
tains (Nr + 1)Nmq nodes (vertices), where Nr is the num- Spacelike (S) [including the Boundary (B)], Timelike (T) and
ber of rounds (0, 1, ...Nr − 1) and Nmq is the number Spacetimelike (ST) edges. Positions of data qubits dq0–dq4
of measure qubits (the number of data qubits is then (not used in the error graph) are indicated at the left.
d = Nmq + 1, which is also the code distance). Each node
i corresponds to readout of a measure qubit (except for
the last column of nodes – see below) and can be asso-
ciated with a pair of error graph coordinates: i = {s, t}, sure qubit index) and t = 0, 1, ...Nr is the time-coordinate
where s = 0, 1, ...Nmq − 1 is the space-coordinate (mea- (round number). The nodes can also be counted, e.g., in
 19

the “time-first” manner, not only when a pair of nodes is connected by a conven-
tional edge on the error graph; therefore, we treat i and
i = t + (Nr + 1)s, (S7) j as arbitrary nodes. However, we still assume that such
pairs (edges) are uncorrelated with each other. In real-
or in the “space-first” manner, ity, sometimes there is a correlation between the edges
(discussed later); so the assumption of the absence of
i = s + Nmq t. (S8) correlation is a first approximation.
As mentioned above, pij denotes the probability that
In each experiment, some of the nodes experience error
two nodes i and j flip color simultaneously. These nodes
detection events [21] (or simply “detection events”) de-
can also flip color because of other edges connected to
noted by red dots in Fig. S14 (black dots denote absence
i and j separately. However, it is important that these
of detection events). By definition, a detection event at
additional flips are independent (uncorrelated) for i and
node i = {s, t} occurs when the corresponding measure-
j because they are caused by different physical errors.
ment result m{s,t} is different from the previous measure-
Therefore, we can consider three uncorrelated processes:
ment of the same qubit, x{s,t} = m{s,t} ⊕m{s,t−1} , where
node i flips color (xi → 1 − xi ) with some probability
xi = 1 means a detection event at node i, while xi = 0
pi , similarly node j flips color with probability pj , and
means no detection event (here ⊕ denotes XOR). There
both nodes flip color with probability pij . Since we start
are two exceptions to this rule. First, for the column with
with the black color (xi = xj = 0), the joint probabilities
t = 0, instead of non-existing m{s,−1} we use the parity
P (xi , xj ) of detection or no detection events at nodes i
of two neighboring data qubits in the initial state (if there
and j are
is no error, we are supposed to get x{s,0} = 0). The sec-
ond special case is for the last column of nodes, t = Nr ,
which does not correspond to a physical round (physi- P (0, 0) = (1 − pij )(1 − pi )(1 − pj ) + pij pi pj , (S9a)
cal rounds are t = 0, 1, ...Nr − 1); in this case, instead
P (0, 1) = (1 − pij )(1 − pi )pj + pij pi (1 − pj ), (S9b)
of non-existing m{s,Nr } , we use the parity of neighboring
data qubit readouts at the end (after the round Nr − 1), P (1, 0) = (1 − pij )pi (1 − pj ) + pij (1 − pi )pj , (S9c)
so that x{s,Nr } = 0 again indicates the expected no-error P (1, 1) = (1 − pij )pi pj + pij (1 − pi )(1 − pj ). (S9d)
situation.
A decoder’s task is to use detection events on the error These formulas have obvious meaning, describing com-
graph to choose one of two given complementary initial binations of the three processes occurring or not occur-
states of data qubits (initial parities of neighboring data ring. Note that P (0, 0) + P (0, 1) + P (1, 0) + P (1, 1) = 1.
qubits are given, so the decoder needs to determine only The relations (S9) can also be expressed via the fractions
one bit of information). The decoder for this experi- of the detection events (often abbreviated as DEF: detec-
ment used minimum-weight perfect matching algorithm tion event fraction) for each node, hxi i = P (1, 0)+P (1, 1)
[21, 50, 52], which connects detection events to each other and hxj i = P (0, 1) + P (1, 1), and the probability of both
(pairwise) or to a space-boundary. detection events, hxi xj i = P (1, 1), which gives
In the conventional Pauli error model assumed by the
decoder [21], the detection events can be produced only hxi i = pi (1 − pij ) + (1 − pi ) pij , (S10a)
in pairs, corresponding to the edges of the error graph
(for the space-boundary edges, only one detection event hxj i = pj (1 − pij ) + (1 − pj ) pij , (S10b)
near the boundary is produced). There are 3 types of hxi xj i = pij (1 − pi ) (1 − pj ) + (1 − pij ) pi pj . (S10c)
such edges – see Fig. S14. Spacelike (S) edges connect
nodes {s, t} and {s + 1, t} (the boundary S-edges connect Solving these equations for pij , pi , and pj , we obtain
nodes {0, t} and {Nmq − 1, t} to the corresponding space-
boundaries), timelike (T) edges connect nodes {s, t} and s
1 1 4 (hxi xj i − hxi ihxj i)
{s, t + 1}, and spacetimelike (ST, “diagonal”) edges con- pij = − 1− , (S11)
nect nodes {s, t} and {s + 1, t + 1}. In the conventional 2 2 1 − 2hxi i − 2hxj i + 4hxi xj i
Pauli error model, a single physical error corresponds to hxi i − pij hxj i − pij
an edge of the error graph. pi = , pj = . (S12)
1 − 2pij 1 − 2pij
Note that if two physical errors occur in edges sharing
a node (see Fig. S14), then there will be no detection We can think about pij as a symmetric matrix, pji =
event at this node: two detection events at the same pij , with indices corresponding to the nodes ordered ei-
node cancel each other. Therefore it is better to say that ther in the “time-first” way (S7) or in the “space-first”
a physical error flips color (black↔red, xi → 1 − xi ) of way (S8) – see Figs. S15 and S16 discussed later. For-
two nodes, instead of producing two detection events. mally, in Eqn. (S11) the diagonal elements are the detec-
Now let us discuss how to find the probability pij of a tion fractions, pii = hxi i; however, we usually set them
physical error, which flips colors of both nodes i and j, to zero, pii ≡ 0, for clarity of graphical presentation.
using experimental statistics of detection events. From Note that in the experimentally relevant case when
experimental data we see that such processes may occur pij  1/4, Eqn. (S11) can be approximated as (i 6= j)
 20

unconventional edges on the error graph reported by pij ).

To study correlations between edges, we have generalized
hxi xj i − hxi ihxj i
pij ≈ . (S13) the method of pij to 3-point and 4-point correlators (es-
(1 − 2hxi i)(1 − 2hxj i) sentially the “hyperedges”), extending the approach of
Equation (S13) for pij is Eqn. (2) of the main text. Eq. (S9) to account for more nodes and more error pro-
This form shows a clear relation of pij to the covariance cesses. This generalization will be described in a future
hxi xj i − hxi ihxj i; however, the correction due to the de- publication.
nominator is typically quite significant. For example, for
hxi i ' hxj i ' 0.11 (see Fig. 1 of the main text), the
B. Fluctuations of the pij elements
denominator in Eqn. (S13) is about 0.6. The approxima-
tion (S13) slightly overestimates Eqn. (S11), the correc-
tion factor is roughly (1 − 3pij ). When evaluating Eqn. (S11) using experimental data,
Equation (S11) allows us to find accurate individual er- the pij values exhibit statistical fluctuations because the
ror probabilities for S, T, and ST edges of the error graph, averages hxi xj i, hxi i, and hxj i are estimated from a large
which are needed for the minimum-weight decoder. How- but finite number Nexpt of experimental realizations (typ-
ever, there is an important exception: the error proba- ical values of Nexpt are between 103 and 105 ). In this sec-
bility for a boundary S-edge cannot be obtained in this tion we estimate the standard deviation σpij of statistical
way because it contains only one node. To find the er- fluctuations of the pij elements.
ror probability pi B for a boundary edge from node i, we For the estimate, let us use the approximation (S13)
use Eqn. (S12), pi,Σ = (hxi i − pi B )/(1 − 2pi B ), in which and assume the usual experimental case when hxi i  1,
the “individual flip” probability pi,Σ is calculated from hxj i  1, and pij  1. Then the effect of the denomina-
already calculated error probabilities for S, T, and ST tor fluctuations is negligible in comparison with fluctua-
edges connected to the node i. We essentially sum up tions of the numerator (covariance Cij ), so
the known error probabilities of the connected edges and
σCij
find the missing error probability (due to the boundary σpij ≈ , Cij = hxi xj i − hxi ihxj i.
edge) to bring the sum to the DEF hxi i. Note, however, (1 − 2hxi i)(1 − 2hxj i)
that it is not a simple sum of the probabilities because (S17)
of the “color flipping” procedure, so that the errors pij1 , Using the form Cij = h(xi − hxi i)(xj − hxj i)i and using
pij2 , ... pijk due to k connected edges produce the total in it true averages hxi i and hxj i instead of averages over
flip probability Nexpt realizations (the effect of the change is negligible),
we find
q
pi,Σ = g(pijk , ... g(pij3 , g(pij2 , pij1 ))...), (S14) σCij = Var[(xi − hxi i)(xj − hxj i)]/Nexpt . (S18)
g(p, q) ≡ p(1 − q) + (1 − p)q = p + q − 2pq. (S15)
The variance here is h(xi − hxi i)2 (xj − hxj i)2 i − Cij 2
, in
Thus, after finding pi,Σ , we calculate the boundary S- which the first term can be rewritten after some algebra
edge probability as as Cij (1 − 2hxi i)(1 − 2hxj i) + hxi ihxj i(1 − hxi i)(1 − hxj i),
hxi i − pi,Σ using the properties x2i = xi and x2j = xj . Inserting
pi B = . (S16) this form into Eqn. (S17) and using Cij /[(1 − 2hxi i)(1 −
1 − 2pi,Σ 2hxj i)] ≈ pij , we obtain
Note that this procedure for boundary edges assumes
that error processes corresponding to different edges are r
hxi ihxj i(1 − hxi i)(1 − hxj i)
uncorrelated. In reality this is not a very good assump- pij (1 − pij ) +
tion (this is why we are actually using a slightly different (1 − 2hxi i)2 (1 − 2hxj i)2
σpij ≈ p .
procedure for boundary edges). A natural way to esti- Nexpt
mate the effect of correlation between the edges is to use (S19)
Eqn. (S14) for a node i not close to a boundary, sum- Note that the first and second terms in the numerator
ming up the contributions from all connected edges and of Eqn. (S19) have a clear meaning and can be obtained
then comparing the result with the DEF hxi i. Doing separately. When pij is well above the statistical noise
this test for the phase-flip experiment, we typically find floor, σpij mainly comes from fluctuation of the number of
a relative inaccuracy of about 4% (median value), which realizations, in which the edge
p error (color flipping event)
indicates a reasonably small but still nonzero correlation has occurred: Nexpt pij ± Nexpt pij (1 − pij ), as follows
between the main edges (for the bit-flip experiment the from the binomial statistics. It is easy to see that this
median relative inaccuracy is about 9%). A natural way leads to the first term in Eqn. (S19). The second term
of thinking about positive correlations between the edges is the noise floor, coming from the fluctuations of hxi i,
is to assume that some error processes flip color of 4, 6, hxj i, and hxi xj i when pij = 0. It can be obtained, e.g.,
... nodes on the error graph, so that the same process in- by considering the numberp of realizations with xi = 1:
creases pij for several pairs of nodes (this also produces Nxi =1 = Nexpt hxi i ± Nexpt hxi i(1 − hxi i), number of
 21

realizations with xi = xj = 1: Nxi =xj =1 = Nxi =1 hxj i ± these lines are around 0.03; they are shown in Fig. S17
discussed in more detail below. There is also a less visible
p
Nxi =1 hxj i(1 − hxj i) (with uncorrelated ±), and re-
alizations with xi = 0 and xj = 1: Nxi =0, xj =1 = line in Fig. S15 next to the S-line (one pixel farther, Nr +
p
(Nexpt − Nxi =1 )hxj i ± (Nexpt − Nxi =1 )hxj i(1 − hxj i) 2, from the main diagonal), which corresponds to ST
(also with uncorrelated ±). Then calculating the appar- edges. The typical values of pij for the ST-line are around
ent value of the covariance Cij and using it in Eqn. (S17), 0.004. Another well-visible feature in Fig. S15 is a reddish
we obtain the noise floor, which gives the second term in “dirt” near S and T lines for qubits mq1 and mq2 and to a
Eqn. (S19). less extent for some other qubits; we attribute this feature
As a final simplification, let us neglect the factors (1 − to leakage to state |2i in a data qubit. One more feature is
pij ) and (1 − hxi i)(1 − hxj i) in Eqn. (S19) (this slightly short lines (“scars”) parallel to the main diagonal, which
increases σpij , so we are on the safe side), thus obtaining we attribute to crosstalk. The leakage and crosstalk are
discussed later.
s S, T, and ST edges. In the conventional theory of
1 hxi ihxj i the repetition QEC code, the errors are associated only
σpij ≈ p pij + . (S20)
Nexpt (1 − 2hxi i)2 (1 − 2hxj i)2 with S, T, and ST edges. The elements of pij show the
probabilities of these errors individually for each edge on
In our repetition phase-flip code experiments, we have the error graph. We emphasize that these probabilities
Nexpt = 76, 000 realizations and the detection error frac- are obtained in situ, during the actual operation of the
tions are hxi i ' hxi i ' 0.11 (slightly bigger, ' 0.12 in code, in contrast to estimates based on qubit coherence
the bit-flip experiments). Thus, the standard deviation and gate fidelities.
of the experimental pij values that are nominally zero As expected from the conventional theory, S, T, and
(noise floor) is roughly ST edges are the main features in Fig. S15. The values of
pij elements for these edges are shown in Fig. S17 by blue
σpij ' 6 × 10−4 . (S21) markers for S-edges, red markers for T-edges, and green
markers for ST-edges; the lines are a guide for the eye.
In particular, this is the noise floor seen in the pij
The S-edge error probabilities for the boundary edges
matrix plots shown in Figs. S15 and S16. Additional
(denoted dq0 and dq10 in Fig. S17) are calculated using
averaging over the rounds leads to even smaller noise floor
Eqs. (S14)–(S16); we see that their values are consistent
(< 2 × 10−4 ) in Fig. 2(c) of the main text.
with other S-edges. Each block of blue markers corre-
sponds to a particular data qubit (indicated at the top),
C. Experimental results for pij markers within a block correspond to time rounds (from
0 to 30, see the horizontal axis). Note that S-edge prob-
abilities for rounds t = 0 and t = 30 are significantly
Figure S15 shows the correlation matrix pij for a phase-
smaller than for other rounds (emphasizing the need of
flip code experiment with 21 qubits (Nmq = 10 measure
many rounds in an experiment). This is because S-edge
qubits and 11 data qubits) and Nr = 30 rounds. In
errors in our phase-flip code are mainly due to dephas-
this particular experiment, no cosmic rays events were
ing of data qubits during readout and reset (or due to
detected, so no data was discarded from Nexpt = 76, 000
energy relaxation for a bit-flip code), while the special
runs. The error graph nodes i and j are ordered in the
rounds t = 0 and t = Nr do not have these parts of the
“time-first” way given by Eqn. (S7). Figure S15 contains
cycle. For other rounds, the error probability pij can
310×310 pixels, with the color of each pixel determined
be crudely estimated as τ /2T2 , where τ is the readout-
by the value of the corresponding pij element. Each axis and-reset time (expected contribution from CZ gates is
contains Nmq = 10 blocks (see grid lines) corresponding significantly smaller). In our experiment, τ = 0.88 µs and
to 10 measure qubits indicated on the axes; each block on average T2 ' 16 µs, which gives τ /2T2 ' 0.028. We
contains Nr + 1 = 31 points (small ticks on the axes) see that pij values for S edges (blue symbols) are close to
corresponding to time rounds. this estimate, though they are different for different data
We see that most pixels in Fig. S15 (which are away qubits, mostly reflecting variation in T2 times and also
from the features discussed below) have values close to having contributions from gate errors. The integrated
zero. The fluctuations are consistent with the expected histogram for the S-edges is shown by the blue line in
noise floor given by Eqn. (S21). The figure is symmetric the left panel of Fig. S18; the median pij value is
across the main diagonal (which runs bottom-left to top-
right) because pji = pij . The values on the main diagonal S−edge, median
pij ≈ 3.0 × 10−2 . (S22)
are set to zero.
The most visible features are 4 diagonal lines (2 from The T-edge errors (red symbols in Fig. S17) are
each side of the main diagonal), which correspond to S grouped in blocks corresponding to measure qubits in-
and T edges of the error graph: the T-edge line contains dicated below the red symbols. The T-edge errors are
pixels next to the main diagonal, while S-edge line is Nr + expected to come mainly from the readout errors, but
1 pixels away from the main diagonal. The color scale for there are also contributions from the gate errors and re-
S and T lines is saturated because the values of pij for set error. Our median readout error is around 0.018;
 22

however, the pij values are considerably higher, with the to 2T, 3T, etc. edges. The figure clearly shows that tem-
median value (see the integrated histogram in Fig. S18) poral correlations can survive for over 5 rounds.
of Leakage to state |2i. We attribute the detection-
T−edge, median
event correlations lasting for several rounds, as seen in
pij ≈ 2.7 × 10−2 . (S23) Fig. S16, to the leakage to state |2i in data qubits. The
same effect causes the “dirt” in Fig. S15 close to S and T
The error probabilities for ST edges (green symbols in lines, with the magnitude of the correlations for several
Fig. S17) are much lower than for S or T edges; they edge types shown in the right panel of Fig. S18. Note that
are supposed to come mainly from CZ gate errors. The measure qubits are reset to |0i at every round, so non-
integrated histogram in Fig. S18 (green line) shows for computational states can survive only in data qubits. For
ST edges the median value of a typical qubit energy relaxation time of T1 ' 15 µs and
ST−edge, median
the round duration of 960 µs, we would expect that state
pij ≈ 3.7 × 10−3 . (S24) |2i should survive on a data qubit for about 8 rounds.
Examining Figs. S18 and S16, we see that this estimate
Unconventional edges. Figure S15 clearly shows is in the right ballpark, but the actual decay of the state
that in contrast to what is expected from the conven- |2i can be significantly faster due to hopping of leakage,
tional QEC theory, some correlations between the detec- a subject of ongoing research.
tion events correspond to error graph edges different from We have found that the amount of leakage is sensitive
the S, T, and ST types. In particular, there are signif- to minor experimental details. The pij technique can be
icantly non-zero pij values near the lines corresponding used for a fast diagnostic to estimate the level of leakage
to T and S edges, separated from them by a few rounds. and to find which qubits suffer a bigger leakage. Spe-
The integrated histogram for some types of these edges cialized experiments have shown [33] that a typical prob-
is shown in the right panel of Fig. S18). As illustrated ability of state |2i in a data qubit is around 4 × 10−3 .
by the inset, with ST0 we denote the “diagonal” edges This magnitude is consistent with the values we extract
similar to the ST edges, but going into the other direc- from the pij analysis. While this analysis is somewhat
tion. With 2T, 3T, etc. we denote the edges spanning 2, involved, we note that ST0 edges have a somewhat sim-
3, etc. rounds for the same measure qubit. We see that ilar (though smaller) pij values due to leakage. For our
out of the unconventional edges, 2T edges have the high- phase-flip code experiment, the median value for ST0 -
est typical probability (the median of 1.7 × 10−3 ), which edge errors is 1.3 × 10−3 , while the biggest value (aver-
is still more than twice smaller than the typical ST-edge aged over rounds) is 3.3 × 10−3 for data qubit dq2 (as
probability. A relatively small probability of unconven- can be seen from Fig. S15, dq2 has the biggest leakage).
tional edges indicates a high quality of the experiment. So, as a crude proxy for leakage, we can use
Note that before the qubit reset [33] was implemented,
the unconventional-edge probabilities were much higher, ST0 −edge (leakage)
pij . 3 × 10−3 . (S25)
with 2T probabilities exceeding ST probabilities.
The negative values of pij for a small fraction of un- The 2T edges can also be used to estimate leakage; the
conventional edges shown in Fig. S18 are consistent with biggest 2T-edge value (averaged over rounds) is 3.6×10−3
the statistical noise level (S21). Note, however, that in for measure qubit mq2. (All these values are for the
some cases, for example, for 2T edges in a high-quality phase-flip code; for a bit-flip code there is an additional
bit-flip experiment, the pij values can actually be slightly contribution from “odd-even correlations” due to energy
negative. This can be understood using Eqn. (S13) as a relaxation of data qubits).
negative correlation. Indeed, a negative correlation be- Note that during several rounds while a data qubit
tween the nodes can be caused by a negative correla- is in state |2i, there is a relatively high probability of
tion between the edges. An example is the second-order detection events at the neighboring measure qubits [33].
anticorrelation due to data qubit energy relaxation (an This leads to a significant correlation between S-edges
energy relaxation event cannot be immediately followed (and also T-edges), which negatively affects performance
by another relaxation event), which may cause slightly of the minimum-weight-matching decoder. This is why
negative pij in a bit-flip repetition code experiment [33]. leakage is dangerous for quantum error correction even
Figure S16 shows the same data as Fig. S15 but with for a relatively low leakage probability.
the different ordering of nodes: here we use the “space- Crosstalk features. Short parallel lines (“scar” fea-
first” ordering from Eqn. (S8). Then each axis contains tures) in Fig. S15 far away from the main diagonal in-
Nr + 1 = 31 blocks corresponding to time rounds (grid dicate the presence of correlations between detection
lines), while Nmq = 10 points within each block corre- events at qubits, which are far apart along the 1D line of
spond to measure qubits. The S-edges are next to the qubits used in the experiment. However, they are actu-
main diagonal, the T-edges are the diagonal lines sepa- ally close to each other on the Sycamore chip – see the
rated by 10 pixels from the main diagonal, and the ST top panel of Fig. S19, which shows 10 pairs of measure
edges are on the next diagonal line (11 pixels from the qubits (indicated by arrows), for which there are visible
main diagonal). The parallel lines in Fig. S15 separated scars in Fig. S15. We attribute these scar features to the
by 20, 30, etc. pixels from the main diagonal correspond crosstalk.
 23

The lower panel of Fig. S19 shows the values of same-

round pij elements averaged over the rounds, for all pairs
of measure qubits except nearest neighbors. While most
values are within the statistical noise level, the elements
corresponding to the scar features are significantly above
the noise floor (bigger values are indicated by orange and
green cells). We see that the magnitude of the crosstalk
correlations is

pcrosstalk
ij . 2 × 10−3 . (S26)

For the crosstalk pairs shifted in time by one round we

find crudely twice smaller edge probabilities.
The long-range correlation between detection events
caused by crosstalk are dangerous to the code operation
because they can effectively reduce the code distance.
However, we see that in our device the crosstalk is quite
small and, most importantly, local in physical distance on
the chip. Therefore, we expect that in the future it will
not present a serious problem in a surface code operation.
 24

0.006
mq0 mq1 mq2 mq3 mq4 mq5 mq6 mq7 mq8 mq9

0.004
cross-talk

0.002
node index j

dg es s
T-e dg ge

0.000
S-eT-ed
es
S

leakage
0.002

0.004

minor ticks correspond to rounds

0.006
mq0 mq1 mq2 mq3 mq4 mq5 mq6 mq7 mq8 mq9
node index i (time-first ordering)

FIG. S15. Correlation matrix pij . A graphical representation of the 310×310 symmetric matrix pij [Eqn. (S11)] for a
phase-flip repetition code experiment with Nmq = 10 measure qubits (11 data qubits) and Nr = 30 rounds. The color of
each pixel depicts the probability pij for an error process involving error graph nodes i and j. The nodes are ordered in the
“time-first” fashion, Eqn. (S7), with 10 blocks (separated by grid lines) corresponding to measure qubits (mq0, mq1, ... mq9)
and 31 ticks within each block corresponding to time rounds (from t = 0 to t = Nr ). The main features are the diagonal lines
corresponding to T, S, and ST edges, which are shifted from the main diagonal by 1, 31, and 32 pixels, respectively (ST line is
more faint than T and S lines). Additional features are reddish (“dirty”) patches near S and T lines, which are due to leakage
to state |2i in data qubits, and also short parallel lines (“scars”) due to crosstalk. Note that the color bar ranges to 0.007,
while probabilities for S and T edges are above this truncation.
 25

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

0.006

0.004

2r 3 4 5
t= t= t= t=
ou
nd
s
0.002
node index j

0.000

0.002

ST-edges 0.004
T-edges
S-edges
0.006
minor ticks: measure qubits mq0, mq1,...mq9

rounds: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
node index i (space-first ordering)
FIG. S16. Matrix pij in space-first node ordering. The figure shows the same data as in Fig. S15, but with the nodes
ordered in the “space-first” fashion of Eqn. (S8). Each axis contains Nr + 1 = 31 blocks with Nmq = 10 points (ticks) within
each block. The lines for S, T, and ST edges are shifted from the main diagonal by 1, 31, and 32 pixels, respectively. Short
dashed lines correspond to 2T, 3T, ... edges, which connect nodes separated by ∆t = 2, 3, ... rounds. The well-visible diagonal
stripes indicate the presence of long-time correlations in detection events lasting for over 5 rounds.
 26

dq0 dq1 dq2 dq3 dq4 dq5 dq6 dq7 dq8 dq9 dq10
0.04

S-edges
Edge error probability pij

0.02 T-edges
mq0 mq1 mq2 mq3 mq4 mq5 mq6 mq7 mq8 mq9
ST-edges
0.00
0 300 300 300 300 300 300 300 300 300 300 30
node index (time-first ordering)
FIG. S17. S, T, and ST errors. The plot shows the error probabilities pij for S (spacelike), T (timelike), and ST (spacetime-
like) edges for the data in Fig. S15 (phase-flip code, 10+11 qubits, 30 rounds). For S-edges (blue symbols) the corresponding
data qubits dq0–dq10 are indicated at the top, 31 points within each block correspond to rounds. The S-edge probabilities
for boundary data qubits dq0 and dq10 are calculated using Eqs. (S14)–(S16). For T-edges (red symbols), the corresponding
measure qubits mq0–mq9 are indicated below the red symbols, each block contains 30 points. ST-edges (green symbols) are
positioned in the same way as S-edges (without boundaries), with 30 points per block. Lines are a guide for the eye.

3T
Integrated histograms
Integrated histograms

2T
space

ST’
ST
T S

time (rounds)

Edge error probability Edge error probability

FIG. S18. Integrated histograms of edge error probabilities. Left panel: Integrated histograms for the error probabilities
pij of the conventional edges: S (blue line), T (red line), and ST (green line). Median values are indicated by vertical dashed
lines and shown in the legend. Right panel: Integrated histograms for ST edges and several unconventional edge types: other-
direction spacetimelike edges ST0 and long timelike edges 2T, 3T, 4T, and 5T spanning 2, 3, 4, and 5 rounds. The inset
illustrates the edge types on the error graph. Median values (dashed lines) are listed in the legend.
 27

9 mq

dq
5 6 7 8 1 0

4 3 2

pij × 103 0.11 mq7

0.26 -0.30 mq6

0.54 -0.00 -0.14 mq5

2.21 1.17 0.25 -0.05 mq4

0.09 1.33 1.97 1.21 -0.05 mq3

0.37 -0.19 -0.07 1.12 1.49 0.13 mq2

0.24 0.05 -0.08 0.05 0.08 0.80 0.75 mq1

0.15 0.03 0.03 0.18 0.21 -0.07 0.28 0.83 mq0

mq2 mq3 mq4 mq5 mq6 mq7 mq8 mq9

FIG. S19. Crosstalk error probabilities. Top panel:

Layout of 10 measure qubits (black circles with integer labels)
and 11 data qubits (gray-filled circles) on the Sycamore de-
vice. Arrows indicate the pairs of measure qubits that exhibit
stronger (red arrows) and weaker (orange arrows) detection-
event correlations due to crosstalk. Bottom panel: Effective
crosstalk probabilities between pairs of measure qubits (ex-
cept for nearest neighbors). We show the values of pij × 103
for same-round pij elements averaged over rounds. Cells are
colored according to the values: yellow and green indicate
a significant crosstalk, blue indicates statistical noise. The
biggest crosstalk of 2.2 × 10−3 is between mq4 and mq6 (left-
most arrow in the top panel).
 28

VII. COMPARISON OF EDGE WEIGHTING a ​Stabilizer circuit schematic

METHODS FOR MATCHING

To decode the error detections obtained in the exper-

iment, we use a minimum weight perfect matching algo-
rithm to determine which physical errors were most likely
given the observed directions. A key component of this
algorithm is the weighting of the edges in the error graph
which correspond to the expected correlated probabilities b ​Stabilizer circuit rendered waveforms
of pairs of nodes. The weight of a particular edge (W )
and the expected probability for that edge (p) are related
by

W = − log p (S27)

which satisfies the property that adding the weights of

two edges corresponds to multiplying their probabilities.
We considered four candidate strategies for determining
expected edge probabilities and weights: FIG. S20. Stabilizer Circuit. a, Circuit schematic repre-
sentation of the stabilizer circuit. Layers of single qubit and
1. Uniform weighting - assume that all edges in the two qubit gates highlighted in blue. Measurement, reset, and
matching graph are equally likely dynamical decoupling operations highlighted in yellow to cor-
respond to the waveforms in b, Rendered waveforms to show
that the majority of the time spent during the stabilizer is
2. Bootstrapping - Run matching on a training during the measurement and reset operations. Lines repre-
dataset with uniform weights, then for a given edge, sent microwave control (XY), flux control (Z), and readout
count the number of times it was matched and di- for the stabilizer circuit for one data qubit (blue) and one
vide by the number of total experiments to compute measure qubit (red).
the expected probability for future matches.

3. Node correlations (pij ) - Use the node correla-

tion technique described in Section VI to determine
the correlated probabilities for edges from a train-
ing dataset.

4. First principles - From the measured gate, mea-

surement, and reset error probabilities, compute
the edge probabilities by propagating possible er-
rors through the circuit. FIG. S21. Dynamical Decoupling Sequences. The four
multi-pulse sequences used during the measurement and reset
portions of the stabilizer circuit. Each sequence has the same
For methods 2 and 3, we use the data at 50 rounds total idle time and executes the same number of gates. The
to determine the matching weights for all other datasets. distinction between these four sequences during the execution
While these methods can in general produce a unique of the circuit is only the phase of the microwave pulses, a
weight for each edge in the 50 round graph, we average technique used to compensate for cumulative pulse errors.
together all rounds so that the edge weights used during
matching are uniform in time. Phase flip and bit flip
edge weights, as well as weights for each of the smaller VIII. DYNAMICAL DECOUPLING OF DATA
subsampled codes, are determined separately. QUBITS
In Tab. S5, we show the fitted values of Λ using the
different weighting methods, for both the bit and phase The measurement and reset operations take 880 ns
flip codes. To within the uncertainty from fitting, we to complete and account for approximately 92% of the
find that methods 2, 3, 4 all give the same result for Λx time spent for the duration of the phase flip code (see
and Λz , while uniform weighting reduces Λx to 2.7 and Fig. S20). Leaving data qubits to idle during these opera-
Λz to 2.5. The primary effect of the more sophisticated tions, we undergo energy relaxation processes in addition
weighting methods is to increase the weights of spacetime to dephasing processes, accounting for a large portion
edges relative to spacelike and timelike edges. of the total error budget. The process of measurement
 29

TABLE S5. Error suppression factors (Λx , Λz for phase and bit flip) and multiplicative constants (Cx and Cz ) fit to logical
error rates vs code distance (Eqn. 1 of the main text) for the four different edge weighting methods.

Weighting method Cx λx Cz λz
Uniform 0.056 ± 0.005 2.79 ± 0.056 0.066 ± 0.007 2.75 ± 0.06
Bootstrapping 0.068 ± 0.008 3.18 ± 0.08 0.078 ± 0.01 3.01 ± 0.09
Correlation (pij ) 0.067 ± 0.008 3.18 ± 0.08 0.077 ± 0.011 3.01 ± 0.09
First principles 0.067 ± 0.007 3.17 ± 0.08 0.0756 ± 0.011 2.99 ± 0.09

a Detection Event Fraction

Idle CP CPMG XY4 XY8
0.5
Detection Fraction

0.4

0.3
0.2
0.1
0.0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Detection Round

b Logical Error Rates c Error Suppression Factors

3.25

3.00

2.75
10-2
2.50
Logical Error Rate

Lambda
2.25

10-3 Idle 2.00

CP 1.75
CPMG
1.50
XY4
10-4 XY8 1.25

Idle CP CPMG XY4 XY8

6 8 10 12 14 16 18 20
Number of Qubits

FIG. S22. Benchmarking phase flip performance with and without dynamical decoupling. b Detection event
fractions vs qubit and round for each of the data qubit Idle, CP, CPMG, XY4, and XY8 operations during measure qubit
readout and reset. Median detection event fraction by round plotted in black. b, Logical error rate vs number of qubits,
showing exponential suppression of error rate in all cases. c, Boxplot of extracted error suppression factors (Λ) from fits like
those shown in b, for five iterations of the experiment for each decoupling scheme. Overall, we see an 1.7x increase in Λ for
all decoupling schemes. The performance between the various decoupling schemes is comparable.

and reset on the measure qubits introduces additional av- verified that we were able to effectively decouple the
enues of error including measurement-induced dephasing qubits from the noise sources listed above. Using CPMG,
from photon crosstalk between readout resonators [53], we verified independently via phase coherence measure-
as well as frequency detuning errors incurred from any ments with and without adversarial readout tones, as
flux crosstalk between qubits. While energy relaxation well as with and without large frequency excursions on
is irreversible and cannot be mitigated here, dephasing neighboring qubits, that we are able to effectively decou-
can be mitigated using dynamical-decoupling techniques. ple away the intrinsic low-frequency noise, measurement-
We employ multi-pulse sequences developed within the induced dephasing on the data qubits caused by crosstalk
field of NMR which have been shown to mitigate low- from measure, as well as any flux crosstalk effects. We
frequency noise in superconducting qubits [54]: Carr- then evaluated the performance of each dynamical decou-
Purcell (CP) [55], Car-Purcell-Meiboom-Gill (CPMG) pling protocol within the context of the repetition code.
[56], XY4, and XY8 [57]. For all of the decoupling sequences, we fix the time be-
tween pulses such that every sequence has the same total
With independent phase coherence measurements, we
 30

idle time and executes the same number of gates (see our Snake optimizer [36].
Fig. S21). The fixed idle time was set such that each To illustrate the performance of our error mitigation
sequence performed eight gates. Using decoupling, we strategy, we conduct a qubit-crosstalk mitigation experi-
see an ∼1.7× increase in the error suppression factor, ment (see Fig. S23). In this experiment, we first optimize
Λ (Fig. S22). To compare the performance of the dif- our processor employing one of three qubit-crosstalk mit-
ferent decoupling schemes, the experiment was run and igation strategies. We then calibrate the processor and
analyzed a total of five times for each of the schemes run the bit-flip repetition code. The three mitigation
(Idle, CP, CPMG, XY4, and XY8). The performance strategies are labelled “none”, “partial”, and “full”, ac-
between schemes was comparable with the CPMG and cording to the expected degree of crosstalk protection. In
XY4 sequences slightly outperforming the CP and XY8 the “none” strategy, we do not penalize for crosstalk. In
sequences. the “partial” strategy, we penalize for crosstalk according
to the cross-entropy benchmarking (XEB) circuit [31],
which we often use in calibration. Although XEB and the
IX. QUBIT FREQUENCY OPTIMIZATION repetition code have different circuits and serve different
purposes, their respective circuits have similar gate pat-
terns (see Fig. S25 of Ref. [31]). Because of this similar-
Our processor employs frequency-tunable qubits [31].
ity, penalizing for crosstalk according to XEB should also
Quantum logic gates are executed at two distinct types
offer partial crosstalk protection for the repetition code.
of frequencies: idle and interaction frequencies, which are
Finally, in the “full” strategy, we penalize for crosstalk
collectively referred to as gate frequencies. Qubits idle
according to the repetition code circuit that we run.
and execute single-qubit gates at their respective idle
To quantify the efficacy of the three mitigation strate-
frequencies. Neighboring qubit-pairs execute CZ gates
gies, we inspect bit-flip repetition-code detection event
at their respective interaction frequencies. All gate fre-
fractions (DEF). We see that by increasing the degree of
quencies are explicitly or implicitly interdependent due
crosstalk mitigation from “none” to “partial” to “full”,
to engineered interactions and/or crosstalk according to
the median DEF is reduced by 33% and 7%, respectively.
the repetition code circuit and its mapping onto our pro-
Furthermore, the DEF standard-deviation is reduced by
cessor. Since many error mechanisms are frequency de-
82% and 51%, respectively. In total, this amounts to a
pendent, we can mitigate errors by constructing and op-
38% reduction in median DEF and a 91% reduction in
timizing an error model with respect to gate frequencies.
the DEF standard-deviation, representing a significant
To construct an error model, we combine error contri-
performance boost. We delegate error mitigation data
butions from Z pulse-distortion, relaxation, dephasing,
for other error mechanisms to a future publication.
and qubit crosstalk. The Z pulse-distortion model pe-
nalizes CZ gates for large frequency excursions. The re-
laxation and dephasing models penalize SQ and CZ gates
for approaching relaxation and dephasing hotspots, while
incorporating coupler physics, qubit hybridization, state-
dependent transitions, and hardware-accurate frequency
trajectories. Finally, the qubit-crosstalk model penalizes
for frequency collisions between nearest-neighbor (NN)
and diagonal next-nearest-neighbor (NNN) qubits, while
incorporating qubit hybridization and the mapping of the
repetition code circuit onto our processor. These con-
stituent models are determined via theory and/or exper-
iment, consolidated, and then trained to be predictive of
experimentally measured error benchmarks via machine
learning.
To determine a frequency configuration that mitigates
error, we optimize the error model with respect to gate
frequencies. Optimization is complex since the error
model spans 41 frequency variables, is non-convex, and
time-dependent [58]. Furthermore, since each frequency
variable is constrained to ∼102 values by the control
hardware and qubit-circuit parameters, the optimiza-
tion search space is ∼2272 , which significantly exceeds
the Hilbert-space dimension 221 . Given the optimization
complexity, exhaustive search is intractable and global
optimization is too slow and inefficient. To quickly and
efficiently find locally optimal gate-frequency configura-
tions and maintain them in the presence of drift, we use
 31

a repetition code b active gates c crosstalking gates d crosstalk mitigation strategies e crosstalk mitigation performance

bit-flip repetition code

full
full

time CZ SQ crosstalk partial

partial

none

H H none

0.0 0.1 0.2 0.3

detection event fraction

FIG. S23. Qubit-crosstalk mitigation. a, The repetition code, with three distinct temporal slices indicated by dashed
boxes. The empty boxes in the lowest temporal slice are either H or I depending on whether we run the bit- or phase-flip code.
b, Simultaneously active SQ (H or I) and CZ gates (blue nodes and edges, respectively) at each temporal slice. The geometry
of active gates is determined by the repetition code circuit and its mapping onto our processor. Simultaneously active gates
can crosstalk due to parasitic interactions between NN and NNN qubits. c, Crosstalking SQ and CZ gates (orange nodes and
edges, respectively) for one active SQ or CZ (blue nodes and edges, respectively) gate at each temporal slice. We mitigate
crosstalk and other error mechanisms by constructing and optimizing an error model with respect to gate-frequencies. d, Three
crosstalk mitigation strategies illustrated for one active CZ gate in the upper temporal slice in a - c. The strategies are labelled
“full”, “partial”, and “none”, according to the degree of expected crosstalk protection. Each strategy can be characterized by
domains (red) in which crosstalk is penalized. e, Bit-flip repetition code benchmarks for each mitigation strategy. The points
and error bars represent the DEF median and standard-deviation, respectively. By increasing the mitigation strength from
“none” to “full”, the DEF median and standard-deviation are reduced by 38% and 91%, respectively.

X. OVERVIEW OF ERROR CORRECTION

EXPERIMENTS

In Table S6, we list experimental implementations of

quantum error correction as a reference.
 32

TABLE S6. Various error correction and error detection experiments. Experiments using “classical” codes (i.e. codes that only
detect one type of error e.g. only phase flips or only bit flips) use classical [n, k, d] code notation instead of quantum [[n, k, d]]
code notation. Entries with an N/A are experiments related to embedding error correction into the physical qubits as opposed
to layering the error correction on top of the physical qubits. Note that there is, as of yet, no experiment exploring a range of
rounds and a range of code distances using a non-classical code.

Paper Year Code name [[#data,#logical,distance]] Physical qubits Rounds Physical qubit type
[22] 1998 Repetition Code [3,1,3] 3 single shot NMR
[23] 2001 Perfect Code [[5,1,3]] 5 single shot NMR
[59] 2011 Repetition Code [3,1,3] 3 3 Ion trap
[24] 2011 Repetition Code [3,1,3] 3 2 NMR
[60] 2011 Repetition Code [3,1,3] 3 single shot NMR
[61] 2012 Repetition Code [3,1,3] 3 single shot Superconducting
[62] 2012 Perfect Code [[5,1,3]] 5 single shot NMR
[63] 2014 Surface Code [[4,1,2]] 4 single shot Photons
[21] 2014 Repetition Code [3,1,3]-[5,1,5] 9 8 Superconducting
[25] 2014 Color Code [[7,1,3]] 7 single shot Ion trap
[64] 2014 Repetition Code [[3,1,3]] 4 single shot NV center
[65] 2015 Repetition Code [3,1,3] 5 single shot Superconducting
[66] 2015 Bell State [[2,0,2]] 4 single shot Superconducting
[67] 2016 Repetition Code [3,1,3] 4 1-3 Superconducting
[68] 2016 Cat States N/A 1 1-6 3D cavity
[27] 2017 Color Code [[4,2,2]] 5 single shot Superconducting
[69] 2017 Color Code [[4,2,2]] 5 single shot Ion trap
[70] 2018 Repetition Code [3,1,3]-[8,1,8] 15 single shot Superconducting
[71] 2019 Bell State [[2,0,2]] 3 1-12 Superconducting
[72] 2019 Perfect Code [[5,1,3]] 5 single shot Superconducting
[73] 2019 Binomial Bosonic States N/A 1 1-19 3D cavity
[28] 2020 Repetition Code [3,1,3]-[22,1,22] 5-43 single shot Superconducting
[43] 2020 Surface Code [[4,1,2]] 7 1-11 Superconducting
[74] 2020 Bell State [[2,0,2]] 3 1-26 Superconducting
[26] 2020 Bacon-Shor Code [[9,1,3]] 15 single shot Ion trap
[75] 2020 Bacon-Shor Code [[9,1,3]] 11 single shot Photons
[76] 2020 GKP States N/A 1 1-200 3D cavity
This work 2020 Repetition Code [3,1,3]-[11,1,11], [[4,1,2]] 5-21 1-50 Superconducting