Article

Logical quantum processor based on

reconfigurable atom arrays
­­­­­
https://doi.org/10.1038/s41586-023-06927-3 Dolev Bluvstein1, Simon J. Evered1, Alexandra A. Geim1, Sophie H. Li1, Hengyun Zhou1,2,
Tom Manovitz1, Sepehr Ebadi1, Madelyn Cain1, Marcin Kalinowski1, Dominik Hangleiter3,
Received: 21 October 2023
J. Pablo Bonilla Ataides1, Nishad Maskara1, Iris Cong1, Xun Gao1, Pedro Sales Rodriguez2,
Accepted: 1 December 2023 Thomas Karolyshyn2, Giulia Semeghini4, Michael J. Gullans3, Markus Greiner1,
Vladan Vuletić5 & Mikhail D. Lukin1 ✉
Published online: 6 December 2023

Open access
Suppressing errors is the central challenge for useful quantum computing1,
Check for updates
requiring quantum error correction (QEC)2–6 for large-scale processing. However,
the overhead in the realization of error-corrected ‘logical’ qubits, in which
information is encoded across many physical qubits for redundancy2–4, poses
substantial challenges to large-scale logical quantum computing. Here we report the
realization of a programmable quantum processor based on encoded logical qubits
operating with up to 280 physical qubits. Using logical-level control and a zoned
architecture in reconfigurable neutral-atom arrays7, our system combines high
two-qubit gate fidelities8, arbitrary connectivity7,9, as well as fully programmable
single-qubit rotations and mid-circuit readout10–15. Operating this logical processor
with various types of encoding, we demonstrate improvement of a two-qubit logic
gate by scaling surface-code6 distance from d = 3 to d = 7, preparation of colour-code
qubits with break-even fidelities5, fault-tolerant creation of logical Greenberger–
Horne–Zeilinger (GHZ) states and feedforward entanglement teleportation, as well
as operation of 40 colour-code qubits. Finally, using 3D [[8,3,2]] code blocks16,17, we
realize computationally complex sampling circuits18 with up to 48 logical qubits
entangled with hypercube connectivity19 with 228 logical two-qubit gates and
48 logical CCZ gates20. We find that this logical encoding substantially improves
algorithmic performance with error detection, outperforming physical-qubit
fidelities at both cross-entropy benchmarking and quantum simulations of fast
scrambling21,22. These results herald the advent of early error-corrected quantum
computation and chart a path towards large-scale logical processors.

Quantum computers have the potential to substantially outperform Recent experiments have achieved milestone demonstrations of
their classical counterparts for solving certain problems1. However, two logical qubits and one entangling gate5,6 and explorations of new
executing large-scale, useful algorithms on quantum processors encodings25–28.
requires very low gate error rates (generally below about 10−10)23, far One specific challenge for realizing large-scale logical processors
below those that will probably ever be achievable with any physical involves efficient control. Unlike modern classical processors that can
device2. The landmark development of QEC theory provides a con- efficiently access and manipulate many bits of information29, quan-
ceptual solution to this challenge2–4. The key idea is to use entangle- tum devices are typically built such that each physical qubit requires
ment to delocalize a logical qubit degree of freedom across many several classical control lines. Although suitable for the implementa-
redundant physical qubits, such that, if any given physical qubit tion of physical qubit processors, this approach poses a substantial
fails, it does not corrupt the underlying logical information. In prin- obstacle to the control of logical qubits redundantly encoded over
ciple, with sufficiently low physical error rates and sufficiently many many physical qubits.
qubits, a logical qubit can be made to operate with extremely high Here we describe the realization of a programmable quantum
fidelity, providing a path to realizing large-scale algorithms4. How- processor based on hardware-efficient control over logical qubits in
ever, in practice, useful QEC poses many challenges, ranging from reconfigurable neutral-atom arrays7. We use this logical processor to
large overhead in physical qubit numbers23 to highly complex gate demonstrate key building blocks of QEC and realize programmable
operations between the delocalized logical degrees of freedom24. logical algorithms. In particular, we explore important features of

Department of Physics, Harvard University, Cambridge, MA, USA. 2QuEra Computing Inc., Boston, MA, USA. 3Joint Center for Quantum Information and Computer Science, NIST/University of
1

Maryland, College Park, MD, USA. 4John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA. 5Department of Physics and Research Laboratory
of Electronics, Massachusetts Institute of Technology, Cambridge, MA, USA. ✉e-mail: [email protected]

58 | Nature | Vol 626 | 1 February 2024

 a b 80 1
Logical qubit storage Ancilla qubit reservoir
Storage zone

Site index

Qubit state
Addressed site Neighbour 40
1

Qubit state
Logical 1Q gate
0 0
0 10 20 30 40 50 60 0 20 40
Time (μs) Time (μs)
c
With local imaging
1.0
Entangling zone

No local imaging

Qubit state
Rydberg
laser Feedforward
AOD 0.5

Syndrome
Logical 2Q gate extraction FPGA 0
Logical qubit 0 0.5 1.0
decoding Ramsey phase (2π)
Readout zone

Number of events
1,000
Camera Mid-circuit
measurement 500
Local imaging

0
1,750 2,000 2,250
Camera counts

Fig. 1 | A programmable logical processor based on reconfigurable atom rotations are implemented using Raman excitation through a 2D AOD;
arrays. a, Schematic of the logical processor, split into three zones: storage, parallel grid illumination delivers the same instruction to multiple atomic
entangling and readout (see Extended Data Fig. 1 for detailed layout). Logical qubits. c, Mid-circuit readout and feedforward. The imaging histogram shows
single-qubit and two-qubit operations are realized transversally with efficient, high-fidelity state discrimination (500 μs imaging time, readout fidelity
parallel operations. Transversal CNOTs are realized by interlacing two logical approximately 99.8%; Methods) and the Ramsey fringe shows that qubit
qubit grids and performing a single global entangling pulse that excites atoms coherence is unaffected by measuring other qubits in the readout zone (error
to Rydberg states. Physical qubits are encoded in hyperfine ground states of probability p ≈ 10 −3; Methods). The FPGA performs real-time image processing,
87
Rb atoms trapped in optical tweezers. b, Fully programmable single-qubit state decoding and feedforward (Fig. 4).

logical operations and circuits, including scaling to large codes, fault large, dynamically programmable grids of light. Fully programmable
tolerance and complex non-Clifford circuits. local single-qubit rotations are realized through qubit-specific, paral-
lel Raman excitation through an additional 2D AOD (ref. 34) (Fig. 1b
and Extended Data Fig. 2). Mid-circuit readout is enabled by moving
Logical processor based on atom arrays selected qubits about 100 μm away to a readout zone and illuminating
Our logical processor architecture, illustrated in Fig. 1a, is split into with a focused imaging beam7,35, resulting in high-fidelity imaging, as
three zones (see also Extended Data Fig. 1). The storage zone is used well as negligible decoherence on stored qubits (Fig. 1c and Extended
for dense qubit storage, free from entangling-gate errors and featuring Data Fig. 3). The mid-circuit10–15 image is collected with a CMOS cam-
long coherence times. The entangling zone is used for parallel logical era and sent to a field-programmable gate array (FPGA) for real-time
qubit encoding, stabilizer measurements and logical gate operations. decoding and feedforward.
Finally, the readout zone enables mid-circuit readout of desired logical The central aspect of our logical processor is the control of individual
or physical qubits, without disturbing the coherence of the computa- logical qubits as the fundamental units, instead of individual physi-
tion qubits still in operation. This architecture is implemented using cal qubits. To this end, we observe that, during most error-corrected
arrays of individual 87Rb atoms trapped in optical tweezers, which can operations, the physical qubits of a logical block are supposed to real-
be dynamically reconfigured in the middle of the computation while ize the same operation, and this instruction can be delivered in paral-
preserving qubit coherence7,9. lel with only a few control lines. This approach naturally multiplexes
Our experiments make use of the apparatus described previously with optical techniques. For example, to realize a logical single-qubit
in refs. 7,8,30, with key upgrades enabling universal digital operation. gate2, we use the Raman 2D AOD (Fig. 1b) to create a grid of light beams
Physical qubits are encoded in clock states within the ground-state and simultaneously illuminate the physical qubits of the logical block
hyperfine manifold (T2 > 1s (ref. 7)) and stored in optical tweezer arrays with the same instruction. Such a gate is transversal2, meaning that
created by a spatial light modulator (SLM)30,31. We use systems of up to operations act on physical qubits of the code block independently.
280 atomic qubits, combining high-fidelity two-qubit gates8, enabled This transversal property further implies that the gate is inherently
by fast excitation into atomic Rydberg states interacting through robust fault-tolerant2, meaning that errors cannot spread within the code block
Rydberg blockade32, with arbitrary connectivity enabled by atom trans- (see Methods), thereby preventing a physical error from spreading into
port by means of 2D acousto-optic deflectors (AODs)7. Central to our a logical fault. Crucially, a similar approach can realize logical entan-
approach of scalable control, AODs10–15,31,33 use frequency multiplex- gling gates2,4. Specifically, we use the grids generated by our moving
ing to take in just two voltage waveforms (one for each axis) to create 2D AOD to pick up two logical qubits, interlace them in the entangling

Nature | Vol 626 | 1 February 2024 | 59

Article
a b Physical error propagation challenge is to improve entangling operations with code distance. Thus
|+L〉 X/Z
|\1L〉 X X motivated, we realize a transversal CNOT gate using logical qubits
|0L〉
X/Z |\2L〉 X encoded in two surface codes (Fig. 2). Surface codes have stabilizers
|+L〉 |0L〉 Control Target that are used for detecting and correcting errors without disrupting
stabilizers stabilizers the logical state4,24. The stabilizers form a 2D lattice of four-body pla-
1. Encode quettes of X and Z operators, which commute with the XL (ZL) logical
logical qubits
operators that run horizontally (vertically) along the lattice (Fig. 2d).
By measuring stabilizers, we can detect the presence of physical qubit
?
errors, decode (infer what error occurred) and correct the error simply
2. Parallel move Hyperedge by applying a software ZL/XL correction24. Such a code can detect and
(interlace) correct a certain number of errors determined by the linear dimension
c d=7
XX ZZ of the system (the code distance d).
0.5
To test the performance of our logical entangling gate, we first initial-
Probability ize the logical qubits by preparing physical qubits of two blocks in |+⟩
3. Pulse global and |0⟩ states, respectively, and performing a single round of stabilizer
CNOT
laser (transversal
CNOT)
measurements with parallel operations7. Although this state prepara-
tion is non-fault-tolerant (nFT) beyond d = 3, we are still able to study
error suppression of the transversal CNOT (Methods). Specifically, we
0
++ +– –+ –– 00 01 10 11 prepare the two logicals in state |+L⟩ and |0L⟩, perform the transversal
d Number of physical qubits per Bell pair Logical bitstring
CNOT and then projectively measure to evaluate the logical Bell-state
26 74 146
0.4 d = 3 (four pairs) stabilizers XL1XL2 and ZL1ZL2 (Fig. 2c). For decoding and correcting the
Logical Bell-pair error Logical Bell-pair error

logical state, we observe that there are strong correlations between

the stabilizers of the two blocks (Extended Data Figs. 4 and 5) owing to
0.3
propagation of physical errors between the codes during the transver-
sal CNOT (ref. 36) (Fig. 2b). We use these correlations to improve per-
d=5
0.2 Conventional decoding formance by decoding the logical qubits jointly, realized by a joint
decoding graph that includes edges and hyperedges connecting the
0.11 stabilizers of the two logical qubits (Fig. 2b, Methods). Using this cor-
Correlated decoding
related decoding procedure, we measure roughly 0.95 populations in
ZL d=7
0.09
the XLXL and ZLZL bases (Fig. 2c), showing entanglement between the
d = 7 logical qubits.
Studying the performance as a function of code size (Fig. 2d) reveals
0.07 that the logical Bell pair improves with larger code distance, dem-
XL
3 5 7 onstrating improvement of the entangling operation. By contrast,
Data X stabilizer
Surface-code distance d Ancilla Z stabilizer we note that, when conventional decoding, that is, independent
minimum-weight perfect matching within both codes4, is used, the
Fig. 2 | Transversal entangling gates between two surface codes. a, Illustration
fidelity decreases with code distance. This is in part because of the nFT
of transversal CNOT between two d = 7 surface codes based on parallel atom
state preparation, whose effect is partially mitigated by the correlated
transport. b, The concept of correlated decoding. Physical errors propagate
decoding (Methods).
between physical qubit pairs during transversal CNOT gates, creating
correlations that can be used for improved decoding. We account for these
We emphasize that, although these results demonstrate surpass-
correlations, arising from deterministic error propagation (as opposed to ing an effective threshold for the entire circuit (implying that we
correlated error events), by adding edges and hyperedges that connect the surpass the threshold of the transversal CNOT), such a threshold
decoding graphs of the two logical qubits. c, Populations of entangled d = 7 is higher owing to projective readout after the transversal CNOT.
surface codes measured in the XX and ZZ basis. d, Measured Bell-pair error as a In practice, the transversal CNOT should be used in combination
function of code distance, for both conventional (top) and correlated (bottom) with many repeated rounds of noisy syndrome extraction6, which is
decoding. We estimate Bell error with the average of the ZZ populations and expected to have a lower threshold and is an important goal for future
the XX parities (Methods). To reduce code distance, we simply remove selected research.
atoms from the grid, as shown on the right, ensuring unchanged experimental
conditions (for d = 3, four logical Bell pairs are generated in parallel). Error bars
represent the standard error of the mean. See Extended Data Figs. 4 and 5 for Fault-tolerant logical algorithms
further surface-code data.
All logical algorithms we perform in this work are built from trans-
versal gates, which are intrinsically fault-tolerant2. We now also use
fault-tolerant state preparation to explore programmable logical
zone and then pulse our single global Rydberg excitation laser to real- algorithms. We use 2D d = 3 colour codes3,37, which are topological
ize a physical entangling gate on each twin pair of the blocks (Figs. 1a codes akin to the surface code, but with the useful capability of trans-
and 2a). This process realizes a high-fidelity, fault-tolerant transversal versal operations of the full Clifford group: Hadamard (H), π/2 phase
CNOT in a single parallel step. (S) gate and CNOT (ref. 37). This transversal gate set can realize any
Clifford circuit fault-tolerantly. As a test case, here we create a logical
GHZ state. Figure 3a shows the implementation of a ten-logical-qubit
Improving entangling gates with code distance algorithm, in which all ten qubits are first encoded by a nFT encoding
A key property of QEC codes is that, for error rates below some thresh- circuit (Methods). Then, five of the codes are used as ancilla logicals,
old, the performance should improve with system size, associated with performing parallel transversal CNOTs to fault-tolerantly detect errors
a so-called code distance4,24. Recently, this property has been experi- on the computation logicals38, and are then moved into the storage
mentally verified by reducing idling errors of a code6. Neutral-atom zone, in which they are safely kept. Subsequently, four computation
qubits can be idly stored for long times with low errors, and the central logicals are used to prepare the GHZ state and logical Clifford rotations

60 | Nature | Vol 626 | 1 February 2024

a Encoding circuit Fault-tolerant GHZ circuit b
|0L〉 C
2
L /Z
L X/Z |0L〉 H C
4
stabilizers
L /Y

|0L〉 C
0.010
3 5
X

SPAM infidelity
|0L〉 C
1 7 6
Physical |0L〉
+ |0L〉

Move to storage
+
+ |0L〉 0.001
H
+ H |0L〉
+ H
|0L〉
+
+ H |0L〉
Logical Physical Logical FT
nFT atom (with CNOT)
c d e
1.0
1.0 Real(Ulogical) Imag(Ulogical)
Increasing
0.8 error
Logical GHZ state fidelity

0.9 detection 0.5 0.5

GHZ fidelity

0.6 0.4 0.4

0.8
0.3 0.3
0.4 FT
0.7 1 0.2 0.2
2 Max. 0.1
0.2 flag
3 errors 0000 0.1 0000
0.6
4 0 0
0 0000 1111 0000 1111
nFT FT EDFT 0.2 0.4 0.6 0.8 1.0 1111 1111
Accepted fraction

Fig. 3 | Fault-tolerant logical algorithms. a, Circuit for preparation of logical qubit SPAM. c, Logical GHZ fidelity without postselecting on flags (nFT),
GHZ state. Ten colour codes are encoded non-fault-tolerantly and then parallel postselecting on flags (FT) and postselecting on flags and stabilizers of the
transversal CNOTs between computation and ancilla logical qubits perform computation logical qubits, corresponding to error detection (EDFT). d, GHZ
fault-tolerant initialization. The ancilla logical qubits are moved to storage fidelity as a function of sliding-scale error-detection threshold (converted into
and a four-logical-qubit GHZ state is created between the computation qubits. the probability of accepted repetitions) and of the number of successful flags
Logical Clifford operations are applied before readout to examine the GHZ in the circuit. e, Density matrix of the four-logical-qubit GHZ state (with at most
state. b, SPAM infidelity of the logical qubits without (nFT) and with (FT) three flag errors) measured by means of full-state tomography involving all
the transversal-CNOT-based flagged preparation, compared with physical 256 logical Pauli strings.

are used at the end of the circuit for direct fidelity estimation39 and full The use of the zoned architecture directly allows scaling circuits
logical state tomography. to larger numbers, without increasing the number of controls, by
We first benchmark our state initialization5,40,41 (Fig. 3b). Averaged encoding and operating on logical qubits, moving them to storage
over the five computation logicals, we find that, by using the and then accessing storage as appropriate. This process is illustrated
fault-tolerant initialization (postselecting on the ancilla logical flag in Fig. 4a,b, in which ten colour codes are made and operated on with
not detecting errors) our |0L⟩ initialization fidelity is 99.91+0.04 −0.09% , parallel transversal CNOTs, moved to storage and then more qubits are
exceeding both our physical qubit |0⟩ initialization fidelity (99.32(4)% accessed from storage. Repeating this process four times, we create
(ref. 8)) and physical two-qubit gate fidelity (99.5% (ref. 8)). Then, Fig. 3c 40 colour codes with 280 physical qubits, at the cost of slow idling
shows that the resulting GHZ state fidelity obtained using the errors of roughly 1% logical decoherence per additional encoding step
fault-tolerant algorithm is 72(2)% (again using correlated decoding), (Fig. 4c). These storage idling errors primarily originate from global
demonstrating genuine multipartite entanglement. Furthermore, we Raman π pulses applied for dynamical decoupling of atoms in the
can postselect on all stabilizers of our computation logicals being entangling zone, which could be greatly reduced with zone-specific
correct; using this error-detection approach, the GHZ fidelity increases Raman controls.
to 99.85+0.1
−1.0 %, at the cost of postselection overhead. Because mid-circuit readout10–15 is an important component of logi-
Because not all nontrivial syndromes are equally likely to cause cal algorithms, we next demonstrate a fault-tolerant entanglement
algorithmic failure, we can perform a partial postselection, in which teleportation circuit. We first create a three-logical-qubit GHZ state
syndrome events most likely to have caused algorithmic failure are |0L0L0L⟩ + |1L1L1L⟩ (Fig. 4d,e) from fault-tolerantly prepared colour
discarded, given by the weight of the correlated matching in the whole codes. Mid-circuit X-basis measurement of the middle logical creates
algorithm. Figure 3d shows the measured GHZ fidelity as a function |0L0L⟩ + |1L1L⟩ if measured as |+L⟩ and |0L0L⟩ − |1L1L⟩ if measured as |−L⟩. We
of this sliding threshold converted into a fraction of accepted experi- recover |0L0L⟩ + |1L1L⟩ by applying a logical S gate to the first and third
mental repetitions, continuously tuning the trade-off between the logicals conditioned in real time on the state of the middle logical, akin
success probability of the algorithm and its fidelity; for example, to the magic-state-teleportation circuit24. Measurements in Fig. 4e
discarding just 50% of the data improves GHZ fidelity to approximately indicate that, although ⟨XLXL⟩ and ⟨YLYL⟩ indeed vanish without the
90%. (As discussed below, for certain applications, purifying samples feedforward step, by applying the feedforward correction, we recover
can be advantageous in improving algorithmic performance.) Finally, a Bell-state fidelity of 77(2)%, limited by imperfections in the original
fault-tolerantly measuring all 256 logical Pauli strings, we perform full underlying GHZ state. By repeating this experiment without mid-circuit
GHZ state tomography (Fig. 3e). readout and instead postselecting on the middle logical being in |+L⟩,

Nature | Vol 626 | 1 February 2024 | 61

Article
a b scrambling circuits using small 3D codes, which are used for native
|0L〉
|0L〉 non-Clifford operations (CCZ).

Move to storage
Access storage
|0L〉
|0L〉
|0L〉 We focus on small 3D [[8,3,2]] codes16,17,26,27 (Fig. 5a), which have
|0L〉
Entangling zone |0L〉
|0L〉
various appealing features. They encode three logicals per block,
|0L〉
|0L〉 feature d = 2 (d = 4) in the Z basis (X basis), implying error-detection
|0L〉
|0L〉
|0L〉
(error-correction) capabilities for Z (X) errors and can realize a trans-
|0L〉
|0L〉
|0L〉
versal CNOT between blocks. Most importantly, by using physical {T, S}
|0L〉
|0L〉 rotations (T is π/4 phase gate), we can realize transversal {CCZ, CZ, Z}
|0L〉
Storage zone |0L〉
|0L〉
gates on the logical qubits encoded within each block, as well as intra-
|0L〉
|0L〉 block CNOTs by physical permutation26,27 (Methods). This gate set
c Total number of colour codes
|0L〉
|0L〉 allows us to transversally realize the circuits illustrated in Fig. 5a,c,
|0L〉
10 20 30 40 |0L〉
|0L〉 alternating between layers of {CCZ, CZ, Z} within blocks and layers of
Logical decoherence

|0L〉
0.02
|0L〉
|0L〉
CNOTs between blocks. Although transversal H is forbidden, initializa-
Physical
Detection
|0L〉
|0L〉 tion and measurement in either the X or the Z basis effectively allows
Correction |0L〉
0
|0L〉
|0L〉
H at the beginning and end of the circuit.
|0L〉
|0L〉 We use these transversal operations to realize logical algorithms
|0L〉
0 1 2 3
Additional encoded groups
|0L〉
that are difficult to simulate classically45,46. More specifically, these
e |0L〉 S circuits can be mapped to instantaneous quantum polynomial (IQP)
d
|0L〉 H circuits20,45,46. Sampling from the output distribution of such circuits
|0L〉 S
is known to be classically hard in certain instances20, implying that a
quantum device can be exponentially faster than a classical computer
Entangling zone
for this task.
Measured logical parity

S S 0.5

Figure 5b shows an example implementation of a 12-logical-qubit

0 sampling circuit. Here we prepare all logical blocks in |+L⟩, implement a
Feedforward

No feedforward

No mid-circuit

scrambling circuit with 28 logical entangling gates and then measure all
Storage zone
Feedforward –0.5 logicals in the X basis. Figure 5b shows the probability of observing each
of the 212 = 4,096 possible logical bitstring outcomes, showing that, as
Readout zone we progressively apply more error detection (that is, postselection) in
XX
YY
XX
YY
XX
YY

post-processing, the distribution more closely reproduces the ideal

Fig. 4 | Zoned logical processor: scaling and mid-circuit feedforward. theoretical distribution. To characterize the distribution overlap, we
a, Atoms in storage and entangling zones and approach for creating and
use the cross-entropy benchmark (XEB)18,47, which is a weighted sum
entangling 40 colour codes with 280 physical qubits. b,c, 40 colour codes
between the measured probability distribution and the ideal calculated
are prepared with a nFT circuit and then 20 transversal CNOTs are used to
distribution, normalized such that XEB = 1 corresponds to perfectly
fault-tolerantly prepare 20 of the 40 codes, whose fidelity is plotted. Logical
decoherence is smaller than the physical idling decoherence experienced
reproducing the ideal distribution and XEB = 0 corresponds to the
during the encoding steps. d, Mid-circuit measurement and feedforward for uniform distribution, which occurs when circuits are overwhelmed by
logical entanglement teleportation. The middle of three logical qubits is noise. Consistent with Fig. 5b, the 12-logical-qubit circuit XEB increases
measured in the X basis and, by applying a mid-circuit conditional, locally from 0.156(2) to 0.616(7) when applying error detection (Fig. 5e). We
pulsed logical S rotation on the other two logical qubits, the state |0L0L⟩ + |1L1L⟩ note that the XEB should be a good fidelity benchmark for IQP circuits
is prepared. e, Measured logical qubit parity with and without feedforward, (Methods).
showing that feedforward recovers the intended state with Bell fidelity of We next explore scaling to larger systems and circuit depths. To
77(2)% (ZZ parities of 83(4)% not plotted; Methods). No mid-circuit refers to ensure high complexity of our logical circuits, we use nonlocal con-
turning off the mid-circuit readout and postselecting on the middle logical nections to entangle the logical triplets on up to 4D hypercube graphs
being in state |+L⟩ in the final readout. By postselecting on perfect stabilizers of (Extended Data Fig. 6 and Supplementary Video 3), which results in
only the two computation logicals (error detection in the final measurement),
fast scrambling19. Exploring entangled systems of 3, 6, 12, 24 and 48
the feedforward Bell fidelity is 92(2)% (not plotted). In d, three of the extra
logical qubits, in all cases, we find a finite XEB score, which improves
blocks are flag qubits and the other four are prepared but unused for this circuit.
with increased error detection (Fig. 5e,f). The finite XEB indicates suc-
cessful sampling and the improvement with error detection shows the
benefit of using logical qubits. Although this improvement comes at
we find a similar Bell fidelity of 75(2)%, indicating high-fidelity perfor- the cost of measurement time owing to error detection, improving the
mance of the readout and feedforward operations. sample quality cannot be replaced by simply generating more samples.
Thus, improving the XEB score yields substantial practical gains. We
obtain an XEB of approximately 0.1 for 48 logical qubits and hundreds of
Complex logical circuits using 3D codes nonlocal logical entangling gates, up to roughly an order of magnitude
One important challenge in realizing complex algorithms with logical higher than previous physical qubit implementations of digital circuits
qubits is that universal computation cannot be implemented trans- of similar complexity18,48, showing the benefits of a logical encoding
versally42. For instance, when using 2D codes such as the surface code, for this application.
non-Clifford operations cannot be easily performed37, and relatively Assuming our best measured physical fidelities, the estimated upper
expensive techniques are required for nontrivial computation24,43, as bound for an optimized physical qubit implementation in our sys-
Clifford circuits can be easily simulated44. By contrast, 3D codes can tem is also greatly below the measured logical XEB (blue line in Fig. 5f;
transversally realize non-Clifford operations, but lose the transversal Methods). In attempting to run these complex physical circuits, in prac-
H (ref. 37). However, these constraints do not imply that classically tice, we find that realizing non-vanishing XEB is much more challeng-
hard or useful quantum circuits cannot be realized transversally or ing; we confirm with small physical instances that we measure values
efficiently. Motivated by these considerations, we explore efficient well below this upper bound (Methods). As well as the error-detecting
realization of classically hard algorithms that are co-designed with benefits, it seems that the logical circuit is substantially more toler-
a particular error-correcting code. Specifically, we implement fast ant to coherent errors, exhibiting operation that is inherently digital,

62 | Nature | Vol 626 | 1 February 2024

 RY
a c π/2 H d
7
Logical qubits
XL3 Stabilizers q Z 1
Z Z Z Z ZZ 100 Measured
5

Runtime per
SZ1 = Z1Z2Z3Z4

bitstring (s)
3 1
ZL1 SZ2 = Z1Z3Z5Z7 10–2
ZL2 XL2
XL1 ZL3 SZ3 = Z1Z2Z5Z6
8 6 SZ4 = Z1Z2Z3Z4Z5Z6Z7Z8 10–4
4 2 SX1 = X1X2X3X4X5X6X7X8

6 12 24 48
Logical operations Number of logicals
Permutation 106
CCZL1,L2,L3 CZL1,L2 Projected
CNOTL1,L2 Transversal CNOT
48 logicals

Runtime per
S S†

bitstring (s)
T T† 104
ap

T S† S
Sw

T† 1 3 5 7
1 3 5 7
102
T 2 4 6 8
2 4 6 8
T† T† 100
T 0 1 2 3
Extra CNOT layers
q48

b e f

12 logicals 48 logicals
Hypercube dimension
3 4 5 6 7
1.00 1.00
Logical bitstring probability

Logical XEB

Logical XEB
0.10 0.10
Theory
Postselection 3
6
g n 12 Raw
0.01 sin tio 0.01 0.01
ea c 24 Error detection
cr te
Raw In r de 48 Physical upper bound
0 ro
er
0 4,096 0.001 0.010 0.100 1.000 3 6 12 24 48
Logical bitstring index Accepted fraction Number of logical qubits

Fig. 5 | Complex logical circuits using 3D codes. a, [[8,3,2]] block codes individual bitstring probability; bottom plot is estimated on the basis of
can transversally realize {CCZ, CZ, Z, CNOT} gates within each block and matrix multiplication complexity. e, Measured normalized XEB as a function
transversal CNOTs between blocks. By preparing logical qubits in |+L⟩, of sliding-scale error detection for 3, 6, 12, 24 and 48 logical qubits. For all sizes,
performing layers of {CCZ, CZ, Z} alternated with inter-block CNOTs and we observe a finite XEB score that improves with increased error detection.
measuring in the X basis, we realize classically hard sampling circuits with Diagram shows 48-logical connectivity, with logical triplets entangled on a 4D
logical qubits. b, Measured sampling outcomes for a circuit with 12 logical hypercube. f, Scaling of raw (red) and fully error-detected (black) XEB from e.
qubits, eight logical CZs, 12 logical CNOTs and eight logical CCZs. By increasing Physical upper-bound fidelity (blue) is calculated using best measured physical
error detection, the measured distribution converges towards the ideal gate fidelities (see Methods and Extended Data Fig. 7 for scaling discussion).
distribution. c, Circuit involving 48 logical qubits with 228 logical CZ/CNOT Diagrams show physical connectivity. [[8,3,2]] cubes are entangled on 4D
gates and 48 logical CCZs. d, Classical simulation runtime for calculating an hypercubes, realizing physical connectivity of 7D hypercubes.

just with imperfect fidelity (see, for example, Extended Data Fig. 7a), In particular, we use a Bell-basis measurement made on two copies of
consistent with theoretical predictions49. We also note that, for the the quantum state (Fig. 6a), which is a powerful tool that can efficiently
logical algorithms, we optimize performance by optimizing the stabi- extract many properties of an unknown state21,22,52. With this two-copy
lizer expectation values (rather than the complex sampling output), technique, in Fig. 6b, we plot the measured entanglement entropy in
providing further advantage for logical implementations. the scrambled system. We observe a characteristic Page curve51 associ-
Our 48-logical circuit, corresponding to a physical qubit connectiv- ated with a maximally entangled, highly scrambled, but globally pure
ity of a 7D hypercube, contains up to 228 logical two-qubit gates and state. These measurements also reveal a final state purity of 0.74(3),
48 logical CCZ gates. Simulation of such logical circuits is challenging compared with the measured XEB of 0.616(7) in Fig. 5f, consistent with
because of the high connectivity (rendering tensor networks ineffi- the XEB being a good proxy for the final state fidelity. Despite postselec-
cient) and large numbers of non-Cliffords50. To benchmark our circuits, tion overhead, we find that error detection greatly improves signal to
we structure them such that we can use an efficient simulation method noise here, as near-zero entropies are exponentially faster to measure
(Methods), which takes about 2 s to calculate the probability of each (Extended Data Fig. 9).
bitstring (Fig. 5d and Extended Data Fig. 8). Modelling noise in our Two-copy measurements can also be used to simultaneously extract
logical circuits is even more complicated, as they are composed of 128 information about all 4N Pauli strings22. Using this property and an analy-
physical qubits and 384 T gates, thereby making experimentation with sis technique known as Bell difference sampling53, we experimentally
logical algorithms necessary to understand and optimize performance. evaluate and directly verify the amount of additive Bell magic53 in our
circuits as a function of the number of applied logical CCZs (Fig. 6c).
This measurement of magic, associated with non-Clifford operations,
Quantum simulations with logical qubits quantifies the number of T gates (assuming decomposition into T)
Finally, we explore the use of logical qubits as a tool in quantum simu- required to realize the quantum state by observing the probability
lation, probing entanglement properties of our fast scrambling cir- that sampled Pauli strings commute with each other (see Methods).
cuits, potentially related to complex systems such as black holes19,51. Moreover, combining encoded qubits and two-copy measurement

Nature | Vol 626 | 1 February 2024 | 63

Article
a Bell-basis measurement b Our observations open the door for exploration of large-scale logical
10
Increasing qubit devices. A key future milestone would be to perform repetitive
error detection

Logical entanglement entropy

Scrambling error correction6 during a logical quantum algorithm to greatly extend
circuit 8
X its accessible depth. This repetitive correction can be directly realized
6 using the tools demonstrated here by repeating the stabilizer measure-
ment (Fig. 2) in combination with mid-circuit readout (Fig. 4). The use
4
of the zoned architecture and logical-level control should allow our
Scrambling

techniques to be readily scaled to more than 10,000 physical qubits by

circuit

2
Z
increasing laser power and optimizing control methods, whereas QEC
0
efficiency can be improved by reducing two-qubit gate errors to 0.1%
0 2 4 6 8 10 12
Logical subsystem size (ref. 8). Deep computation will further require continuous reloading
c d of atoms from a reservoir source11,15. Continued scaling will benefit
Experiment 0.25 Increasing error detection
8 Theory from improving encoding efficiency, for example, by using quantum
low-density-parity-check codes55,56, using erasure conversion13,33,57 or
Additive Bell magic

0.20
Logical Pauli string
expectation value

noise bias35 and optimizing the choice of (possibly several) atomic

0.15
4
species11,14,47, as well as advanced optical controls34. Further advances
0.10 could be enabled by connecting processors together in a modular
0.05
fashion using photonic links or transport10,58 or more power-efficient
trapping schemes such as optical lattices59. Although we do not expect
0 0 clock speed to limit medium-scale logical systems, approaches to speed
0 1 2 3 0 0.2 0.4 0.6 0.8 1.0
Number of logical CCZ gates Logical purity
up processing in hardware60 or with nonlocal connectivity61 should also
be explored. We expect that such experiments with early-generation
Fig. 6 | Logical two-copy measurement. a, Identical scrambling circuits are logical devices will enable experimental and theoretical advances that
performed on two copies of 12 logical qubits and then measured in the Bell greatly reduce anticipated costs of large-scale error-corrected systems,
basis to extract information about the state. Z-basis measurements are corrected
accelerating the development of practical applications of quantum
with an [[8,3,2]] decoder (when full error detection is not applied). b, Measured
computers.
entanglement entropy as a function of subsystem size, showing expected
Page-curve behaviour19 for the highly scrambled state, improving with increased
error detection. c, Measured and simulated magic (associated with non-Clifford
Online content
operations) as a function of the number of CCZ gates applied, performed on
two copies of scrambled six-logical-qubit systems. d, Pauli string measurement Any methods, additional references, Nature Portfolio reporting summa-
and zero-noise extrapolation using logical qubits. Plot shows the absolute ries, source data, extended data, supplementary information, acknowl-
values of all 412 Pauli string expectation values, which only have five discrete edgements, peer review information; details of author contributions
values for our digital circuit; Pauli strings with the same theory value are and competing interests; and statements of data and code availability
grouped. By analysing with sliding-scale error detection, we improve towards are available at https://doi.org/10.1038/s41586-023-06927-3.
the theoretical expectation values (squares) while also improving towards a
purity of 1. By extrapolating to perfect purity, we extrapolate the expectation
values and better approximate the ideal values (shaded regions are statistical 1. Preskill, J. Quantum computing in the NISQ era and beyond. Quantum 2, 79 (2018).
fit uncertainty). 2. Shor, P. W. in Proc. 37th Conference on Foundations of Computer Science 56–65
(IEEE, 1996).
3. Steane, A. Multiple-particle interference and quantum error correction. Proc. R. Soc.
Lond. A Math. Phys. Eng. Sci. 452, 2551–2577 (1996).
4. Dennis, E., Kitaev, A., Landahl, A. & Preskill, J. Topological quantum memory. J. Math.
allows for further error-mitigation techniques. As an example, Fig. 6d Phys. 43, 4452–4505 (2002).
shows the measured absolute expectation values of all 412 logical Pauli 5. Ryan-Anderson, C. et al. Implementing fault-tolerant entangling gates on the five-qubit
code and the color code. Preprint at https://arxiv.org/abs/2208.01863 (2022).
strings with sliding-scale error detection. Because in the two-copy 6. Google Quantum AI. Suppressing quantum errors by scaling a surface code logical qubit.
measurements for each error-detection threshold we also measure the Nature 614, 676–681 (2023).
overall system purity, we can extrapolate our expectation values to the 7. Bluvstein, D. et al. A quantum processor based on coherent transport of entangled atom
arrays. Nature 604, 451–456 (2022).
case of unit purity (zero noise)54. This procedure evaluates the averaged 8. Evered, S. J. et al. High-fidelity parallel entangling gates on a neutral-atom quantum
Pauli expectation values to about 10% relative precision of the ideal computer. Nature 622, 268–272 (2023).
theoretical values spanning several orders of magnitude (Methods). 9. Beugnon, J. et al. Two-dimensional transport and transfer of a single atomic qubit in
optical tweezers. Nat. Phys. 3, 696–699 (2007).
10. Deist, E. et al. Mid-circuit cavity measurement in a neutral atom array. Phys. Rev. Lett. 129,
203602 (2022).
Outlook 11. Singh, K. et al. Mid-circuit correction of correlated phase errors using an array of spectator
qubits. Science 380, 1265–1269 (2023).
These experiments demonstrate key ingredients of scalable error cor- 12. Graham, T. M. et al. Midcircuit measurements on a single-species neutral alkali atom
rection and quantum information processing with logical qubits. As well quantum processor. Phys. Rev. X 13, 041051 (2023).
as implementing the key elements of logical processing, our approach 13. Ma, S. et al. High-fidelity gates and mid-circuit erasure conversion in an atomic qubit.
Nature 622, 279–284 (2023).
demonstrates practical utility of encoding methods for improving 14. Lis, J. W. et al. Midcircuit operations using the omg architecture in neutral atom arrays.
sampling and quantum simulations of complex scrambling circuits. Phys. Rev. X 13, 041035 (2023).
Future work can explore whether these methods can be generalized, 15. Norcia, M. A. et al. Midcircuit qubit measurement and rearrangement in a 171Yb atomic
array. Phys. Rev. X 13, 041034 (2023).
for example, to more robust, higher-distance codes and if such highly 16. Campbell, E. T. The smallest interesting colour code. Earl T. Campbell https://
entangled, non-Clifford states could be used in practical algorithms. earltcampbell.com/2016/09/26/the-smallest-interesting-colour-code/ (2016).
We note that the demonstrated logical circuits are approaching the 17. Vasmer, M. & Kubica, A. Morphing quantum codes. Phys. Rev. Appl. 10, 030319 (2022).
18. Arute, F. et al. Quantum supremacy using a programmable superconducting processor.
edge of exact simulation methods (Fig. 5d) and can readily be used Nature 574, 505–510 (2019).
for exploring error-corrected quantum advantage. These examples 19. Kuriyattil, S., Hashizume, T., Bentsen, G. & Daley, A. J. Onset of scrambling as a dynamical
demonstrate that the use of new encoding schemes, co-designed with transition in tunable-range quantum circuits. PRX Quantum 4, 030325 (2023).
20. Bremner, M. J., Montanaro, A. & Shepherd, D. J. Average-case complexity versus
efficient implementations, can allow the implementation of particular approximate simulation of commuting quantum computations. Phys. Rev. Lett. 117,
logical algorithms at reduced cost. 080501 (2016).

64 | Nature | Vol 626 | 1 February 2024

21. Daley, A. J., Pichler, H., Schachenmayer, J. & Zoller, P. Measuring entanglement growth in 46. Paletta, L., Leverrier, A., Sarlette, A., Mirrahimi, M. & Vuillot, C. Robust sparse IQP sampling
quench dynamics of bosons in an optical lattice. Phys. Rev. Lett. 109, 020505 (2012). in constant depth. Preprint at https://arxiv.org/abs/2307.10729 (2023).
22. Huang, H. Y. et al. Quantum advantage in learning from experiments. Science 376, 1182–1186 47. Shaw, A. L. et al. Benchmarking highly entangled states on a 60-atom analog quantum
(2022). simulator. Preprint at https://arxiv.org/abs/2308.07914 (2023).
23. Gidney, C. & Ekerå, M. How to factor 2048 bit RSA integers in 8 hours using 20 million noisy 48. Wu, Y. et al. Strong quantum computational advantage using a superconducting quantum
qubits. Quantum 5, 433 (2021). processor. Phys. Rev. Lett. 127, 180501 (2021).
24. Fowler, A. G., Mariantoni, M., Martinis, J. M. & Cleland, A. N. Surface codes: towards 49. Bravyi, S., Englbrecht, M., König, R. & Peard, N. Correcting coherent errors with surface
practical large-scale quantum computation. Phys. Rev. A 86, 032324 (2012). codes. npj Quantum Inf. 4, 55 (2018).
25. Self, C. N., Benedetti, M. & Amaro, D. Protecting expressive circuits with a quantum error 50. Bravyi, S. et al. Simulation of quantum circuits by low-rank stabilizer decompositions.
detection code. Nat. Phys. https://doi.org/10.1038/s41567-023-02282-2 (2024). Quantum 3, 181 (2019).
26. Honciuc Menendez, D., Ray, A. & Vasmer, M. Implementing fault-tolerant non-Clifford 51. Sekino, Y. & Susskind, L. Fast scramblers. J. High Energy Phys. 2008, 065 (2008).
gates using the [[8,3,2]] color code. Preprint at https://arxiv.org/abs/2309.08663 (2023). 52. Hangleiter, D. & Gullans, M. J. Bell sampling from quantum circuits. Preprint at https://
27. Wang, Y. et al. Fault-tolerant one-bit addition with the smallest interesting colour code. arxiv.org/abs/2306.00083 (2023).
Preprint at https://arxiv.org/abs/2309.09893 (2023). 53. Haug, T. & Kim, M. S. Scalable measures of magic resource for quantum computers. PRX
28. Andersen, T. I. et al. Non-Abelian braiding of graph vertices in a superconducting Quantum 4, 010301 (2023).
processor. Nature 618, 264–269 (2023). 54. Kim, Y. et al. Evidence for the utility of quantum computing before fault tolerance. Nature
29. Patterson, D. A. & Hennessy, J. L. Computer Organization and Design: The Hardware/ 618, 500–505 (2023).
Software Interface. RISC-V Edition (Morgan Kaufmann, 2018). 55. Bravyi, S. et al. High-threshold and low-overhead fault-tolerant quantum memory.
30. Ebadi, S. et al. Quantum phases of matter on a 256-atom programmable quantum Preprint at https://arxiv.org/abs/2308.07915 (2023).
simulator. Nature 595, 227–232 (2021). 56. Xu, Q. et al. Constant-overhead fault-tolerant quantum computation with reconfigurable
31. Scholl, P. et al. Quantum simulation of 2D antiferromagnets with hundreds of Rydberg atom arrays. Preprint at https://arxiv.org/abs/2308.08648 (2023).
atoms. Nature 595, 233–238 (2021). 57. Wu, Y., Kolkowitz, S., Puri, S. & Thompson, J. D. Erasure conversion for fault-tolerant quantum
32. Jaksch, D. et al. Fast quantum gates for neutral atoms. Phys. Rev. Lett. 85, 2208–2211 (2000). computing in alkaline earth Rydberg atom arrays. Nat. Commun. 13, 4657 (2022).
33. Scholl, P. et al. Erasure conversion in a high-fidelity Rydberg quantum simulator. Nature 58. Dordević, T. et al. Entanglement transport and a nanophotonic interface for atoms in
622, 273–278 (2023). optical tweezers. Science 373, 1511–1514 (2021).
34. Graham, T. M. et al. Multi-qubit entanglement and algorithms on a neutral-atom quantum 59. Tao, R., Ammenwerth, M., Gyger, F., Bloch, I. & Zeiher, J. High-fidelity detection of large-
computer. Nature 604, 457–462 (2022). scale atom arrays in an optical lattice. Preprint at https://arxiv.org/abs/2309.04717
35. Cong, I. et al. Hardware-efficient, fault-tolerant quantum computation with Rydberg (2023).
atoms. Phys. Rev. X 12, 021049 (2022). 60. Xu, W. et al. Fast preparation and detection of a Rydberg qubit using atomic ensembles.
36. Beverland, M. E., Kubica, A. & Svore, K. M. Cost of universality: a comparative study of Phys. Rev. Lett. 127, 050501 (2021).
the overhead of state distillation and code switching with color codes. PRX Quantum 2, 61. Litinski, D. & Nickerson, N. Active volume: an architecture for efficient fault-tolerant
020341 (2021). quantum computers with limited non-local connections. Preprint at https://arxiv.org/
37. Bombín, H. Gauge color codes: optimal transversal gates and gauge fixing in topological abs/2211.15465 (2022).
stabilizer codes. New J. Phys. 17, 083002 (2015).
38. Goto, H. Minimizing resource overheads for fault-tolerant preparation of encoded states Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
of the Steane code. Sci. Rep. 6, 19578 (2016). published maps and institutional affiliations.
39. Flammia, S. T. & Liu, Y. K. Direct fidelity estimation from few Pauli measurements. Phys.
Rev. Lett. 106, 230501 (2011). Open Access This article is licensed under a Creative Commons Attribution
40. Egan, L. et al. Fault-tolerant control of an error-corrected qubit. Nature 598, 281–286 (2021). 4.0 International License, which permits use, sharing, adaptation, distribution
41. Postler, L. et al. Demonstration of fault-tolerant universal quantum gate operations. and reproduction in any medium or format, as long as you give appropriate
Nature 605, 675–680 (2022). credit to the original author(s) and the source, provide a link to the Creative Commons licence,
42. Eastin, B. & Knill, E. Restrictions on transversal encoded quantum gate sets. Phys. Rev. and indicate if changes were made. The images or other third party material in this article are
Lett. 102, 110502 (2009). included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
43. Brown, B. J. A fault-tolerant non-Clifford gate for the surface code in two dimensions. to the material. If material is not included in the article’s Creative Commons licence and your
Sci. Adv. 6, eaay4929 (2020). intended use is not permitted by statutory regulation or exceeds the permitted use, you will
44. Aaronson, S. & Gottesman, D. Improved simulation of stabilizer circuits. Phys. Rev. A 70, need to obtain permission directly from the copyright holder. To view a copy of this licence,
052328 (2004). visit http://creativecommons.org/licenses/by/4.0/.
45. Mezher, R., Ghalbouni, J., Dgheim, J. & Markham, D. Fault-tolerant quantum speedup from
constant depth quantum circuits. Phys. Rev. Res. 2, 033444 (2020). © The Author(s) 2023

Nature | Vol 626 | 1 February 2024 | 65

Article
Methods profiles30, to be homogeneous over a 35-μm-tall region. As the Rydberg
beams propagate longitudinally, the entangling zone is longer than it
System overview is tall. We optimize top hats to be homogeneous over roughly 250-μm
Our experimental apparatus (Extended Data Fig. 1a) is described horizontal extent. Taller regions are also achievable, with a trade-off
previously in refs. 7,8,30. To carry out these experiments, several key with reduced laser intensity and greater challenge in homogenization.
upgrades have been made enabling programmable quantum circuits on The 250-μm width of the zones used here is set by the bandwidth of
both physical and logical qubits. A cloud containing millions of cold 87Rb our AOD deflection efficiency. We position the readout zone on the
atoms is loaded in a magneto-optical trap inside a glass vacuum cell, other side of the storage zone to further minimize decoherence on
which are then loaded stochastically into programmable, static arrange- entangling-zone atoms.
ments of 852-nm traps generated with a SLM and rearranged with a set Our two-qubit gate parameters are similar to our previous work in
of 850-nm moving traps generated by a pair of crossed AODs (DTSX- ref. 8. During two-qubit Rydberg (n = 53) gates, we place atoms ≲2 μm
400, AA Opto-Electronic) to realize defect-free arrays30,31,62. Atoms are apart within a ‘gate site’, resulting in ≳450 MHz interaction strength
imaged with a 0.65-NA objective (Special Optics) onto a CMOS camera between pairs, much larger than the Rabi frequency of 4.6 MHz. Nota-
(Hamamatsu ORCA-Quest C15550-20UP), chosen for fast electronic bly, owing to the use of the Rydberg blockade32,68, the gate is largely
readout times. The qubit state is encoded in mF = 0 hyperfine clock independent of the exact distance between atoms. Hence, precise
states in the 87Rb ground-state manifold, with T2 > 1s (ref. 7), and fast, inter-atom positioning is not required. Gate sites are separated such
high-fidelity single-qubit control is executed by two-photon Raman that atoms in different gate sites are no closer than 10 μm during the
excitation7,63 (Extended Data Fig. 1b). A global Raman path illuminating gate, resulting in negligible long-range interactions. Throughout this
the entire array is used for global rotations (Rabi frequency roughly work, we use four gate sites vertically (five for the surface-code experi-
1 MHz, resulting in approximately 1-μs rotations with composite pulse ment) and 20 horizontally, performing gates on as many as 160 qubits
techniques7), as well as for dynamical decoupling throughout the entire simultaneously (see Extended Data Fig. 1d). Under various conditions,
circuit (typically one global π pulse per movement). Fully program- with proper calibration, we measure two-qubit gate fidelities in the
mable local single-qubit rotations are realized with the same Raman range F = 99.3–99.5%. We do not observe any error on storage-zone
light but redirected through a local path, which is focused onto tar- atoms when Rydberg gates are executed in the entangling zone. Even
geted atoms by an additional set of 2D AODs. Entangling gates (270-ns though the tail of the top-hat Rydberg excitation beams is only sup-
duration) between clock qubits are performed with fast two-photon pressed to about 10% intensity, the two-photon drive is far off-resonant
excitation using 420-nm and 1,013-nm Rydberg beams to n = 53 Rydberg owing to the approximately 20 MHz 1013 light shift detuning that is
states, using a time-optimal two-qubit gate pulse64,65, detailed in ref. 8. present for the entangling-zone atoms8. We natively realize physical
During the computation, atoms are rearranged with the AOD traps CZ gates; when implementing CNOTs, we add physical H gates. We find
to enable arbitrary connectivity7,66,67. Mid-circuit readout is carried minimal two-qubit cross-talk between gate sites, as examined with
out by illuminating from the side with a locally focused 780-nm imag- long benchmarking sequences in ref. 8. Although ref. 8 seems to find
ing beam, with scattered photons collected on the CMOS camera and some small cross-talk seemingly originating from decay into Rydberg
processed in real time by a FPGA (Xilinx ZCU102), with feedforward P states, this should be considerably suppressed in the practical opera-
control signal outputs. tion here owing to the approximately 200 μs duration between gates,
The quantum circuits are programmed with a control infrastructure during which time Rydberg atoms should either fly away or decay back
consisting of five arbitrary waveform generators (AWGs) (Spectrum to the ground state.
Instrumentation), as illustrated in Extended Data Fig. 1c, synchronized
to <10 ns jitter. The two-channel ‘Rearrangement AWG’ is used for Shuttling and transfers. The SLM tweezers can have arbitrary po-
rearranging into defect-free arrangements30 before the circuit, the sitions but are static. The AOD tweezers are mobile but have several
one channel of the ‘Rydberg AWG’ is used for entangling-gate pulses, constraints7,69. In particular, the AOD array creates rectangular grids
the four channels of the ‘Raman AWG’ are used for IQ (in-phase and (but not all sites need to be filled). During the atom-moving operations,
quadrature) control of a 6.8-GHz source7,63 (the global phase reference they are only used for stretches, compressions and translations of the
for all qubits) and pulse-shaping of the global and local Raman driving, AOD trap array, that is, atoms move in rows and columns, and rows
the two channels of the ‘Raman AOD AWG’ are used for generating and columns never cross7,69. Arbitrary qubit motions and permuta-
tones that create the programmable light grids for local single-qubit tion is achieved by shuttling atoms around in AOD tweezers and then
control and the two channels of the ‘Moving AOD AWG’ are used for transferring atoms between AOD and SLM tweezers as appropriate. We
controlling the positions of all atoms during the circuit. AODs are perform gates on pairs of atoms in both AOD–AOD traps and AOD–SLM
central to our methods of efficient control62, in which the two voltage traps, with no observed difference for gate performance as measured
waveforms (one for the x axis and one for the y axis) control many by randomized benchmarking8.
physical or logical qubits in parallel: each row and column of the grid We find that free-space shuttling of atoms (that is, no transfers) in
simply corresponds to a single frequency tone, and these tones are AOD tweezers comes essentially with no fidelity cost (other than time
then superimposed in the waveform delivered to the AOD (amplified overhead), consistent with our previous work7. Two further improve-
by Mini Circuits ZHL-5W-1+). The phase relationship between tones ments here are the use of a photodiode to calibrate and homogenize
is chosen to minimize interference. the 2D deflection efficiency of our 2D AODs to percent-level homo­
geneity across our used region and engineering atomic trajectories
Programming circuits and echo sequences to cancel out residual-path-dependent inhomo-
Most of the system parameters used in our approach do not have hard geneities. For example, we move an atom 100 μm away to realize a
limits but instead result from possible trade-offs. Next, we detail some distant entangling gate and then, before returning the atom, we per-
design decisions made for the circuits used in this work. form a Raman π pulse, so that differential light shifts accumulated
during the return trip cancel with the first trip. Motion is realized with
Zone parameter choices. For simplicity, we keep the entangling zone a cubic profile as in ref. 7, the characteristic free-space movement
fixed for all experiments. This conveniently allows us to switch between, time between gates is roughly 200 μs and acoustic-lensing effects
for example, surface code and [[8,3,2]] code experiments, without fur- from the AOD are estimated to be negligible. We pulse the 1013 laser
ther calibrations. We choose our entangling-zone profile, realized by off during motion to remove loss effects from the large light shifts.
420-nm and 1,013-nm Rydberg ‘top-hat’ beams generated by SLM phase Note that the 1013-induced differential light shift on the hyperfine
qubit is only on the kHz scale but we still ensure its effects are properly complex, varied circuits on hundreds of physical qubits. Animations of
echoed out. all of our programmed circuits are attached as Supplementary Videos.
Transferring atoms between tweezers9 presents further challenges.
We measure the infidelity of each transfer, encompassing both dephas- Programmable single-qubit gates
ing and loss, to be ≲0.1%. To achieve this performance, in our transfer To enable individual single-qubit gates, we use the same Raman laser
from SLM to AOD, we ramp up the intensity of the AOD tones (with system as our global rotation scheme and illuminate only chosen atoms
quadratic intensity profile when possible) corresponding to the appro- using a pair of crossed AODs. The focused beam waist in the plane of
priate sites over a time of 100–200 μs to a trap depth about two times the atoms is 1.9 μm, which is large enough to be robust to fluctuations
larger than the SLM trap depth, and then move the AOD trap 1–2 μm in atomic positions and small enough to prevent cross-talk to neigh-
away over a time of 50–100 μs. These timescales can probably be short- bouring atoms separated by ≳6 μm. For Raman excitation, polarization
ened considerably while suppressing errors using optimal control needs to be carefully considered. Unlike the global path, the local
techniques. During subsequent motion, we leave the AOD trap depth at beam-propagation direction is perpendicular to the atom-quantization
this 2× value. To transfer an atom AOD to SLM, we perform the reversed axis (set by the external magnetic field). Therefore, the fictitious mag-
→
process. During these transfer processes, the differential light shifts netic field Bfict responsible for driving the transitions, as described in
on the transferred atoms are dynamically changing and can result in ref. 63, preferentially drives σ± hyperfine transitions rather than the
large unechoed phase shifts. As such, whenever possible, we engineer desired π clock transition73. There exist two possible approaches to
circuits such that pairs of transfers will echo with appropriately chosen single-qubit gates, as illustrated in Extended Data Fig. 2a. First,
π pulses. When echoing pairs of transfers is not possible, we perform off-resonant σ± dressing generates differential light shifts between
one cycle of XY4 or XY8 dynamical decoupling during the transfer. qubit states, enabling fast local Z(θ) gates. Global π/2 rotations convert
Finally, we note that low-loss transfer is highly sensitive to alignment these to local X(θ) gates. Second, we can directly apply local X(θ) gates
of the AOD and SLM grid. We fix small optical distortions between the with direct π transitions by slightly rotating the quantization axis
AOD and SLM tweezer grids by fine adjustment of individual SLM grid towards the local beam direction; this could be achieved with an exter-
→
tweezers, which can be arbitrarily positioned, to overlap with the AOD nal field but, conveniently, Bfict has a DC component that naturally
traps as seen on an image plane reference camera. It is important to rotates the axis. Note that, if the local beam is quickly turned on, this
adjust the SLM and not the AOD, as small adjustments of individual AOD same fictitious DC field causes leakage out of the mF = 0 subspace,
tones deviating from a frequency comb causes beating and atom loss. therefore Gaussian-smoothed pulses are used throughout this work.
Although we realize both the π and σ± versions above, in these
Dynamical decoupling and local gates. In our circuit design, we engi- experiments, we use the off-resonant σ± dressing procedure because
neer our echo sequences to cancel out as many deleterious aspects as of reduced polarization sensitivity, as our polarization homogeneity
possible. We ensure that, in our dynamical decoupling, we have an odd was affected by the sharp wavelength edge of a dichroic after the AOD.
number of π pulses between CZ gates (whenever possible), as this ech- Furthermore, as for most circuits, we perform local rotations row by
oes out both systematic and spurious contributions to the single-qubit row (only one Y tone at a time); this enables arbitrary fine-tuning of X
phase7,8. We apply appropriate X(π) and Z(π) rotations between local coordinates and powers at each site for homogenizing and calibrating
addressing with the local Raman to cancel out errors induced by the rotations (Extended Data Fig. 2b). We calibrate using the procedure
global π/2 pulses, as well as between pulses of the 420-nm laser (when in Extended Data Fig. 2c and find that these calibrations are stable on
used for entangling-zone single-qubit rotations7) to echo out small month timescales.
cross-talk experienced in the storage zone by the tail of the 420-nm To quantify the fidelity, we perform randomized benchmarking using
beam. For our global decoupling pulses, we use both BB1 pulses70 and 0, 10, 20, 30, 40 and 50 local Z(π/2) rotations (per site) on 16 sites,
SCROFULOUS pulses71. To benchmark and optimize coherence dur- obtaining F = 99.912(7)%, as shown in Extended Data Fig. 2d (note that
ing our complex circuits, we perform a Ramsey fringe measurement the single-qubit gates we execute globally have fidelity closer to 99.99%
encompassing the entire movement and single-qubit gate sequence (refs. 7,8)). This approaches the Raman scattering limit for our σ±
and optimize the observed contrast7. When performing properly, our scheme (error of about 7 × 10−4 per π/2 pulse), but when not well cali-
total single-qubit error is consistent with state preparation and meas- brated is limited by inhomogeneity, in particular, associated with
urement (SPAM)8, an effective coherence time of 1–2 s and the Raman distortions of the y position of the rows. In the future, the performance
scattering error of all the Raman pulses7,63. We note that these measured can be further improved by using X(θ) gates, which enables robust
coherence times include the movement within and between zones; composite sequences such as BB1 (ref. 70), has an improved Raman
although we use fewer pulses (typically one per movement) than the scattering contribution and is faster (roughly 1 μs duration).
XY16-128 sequence used to benchmark 1.5 s coherence in ref. 7, the
coherence times here are naturally longer because of further-detuned Mid-circuit readout and feedforward
tweezers used (852 nm rather than 830 nm). To perform mid-circuit readout10–15 of selected qubits without affecting
Local single-qubit gates34,72 with the Raman AOD are realized in arbi- the others, we use a local imaging beam focused on the readout zone
trary positions in space on both AOD and SLM atoms. Targeted logical that is roughly 100 μm spatially separated from the entangling zone7,35.
qubit blocks are addressed by a grid illumination of the logical block. The local imaging beam consists of 780-nm circularly polarized light,
Arbitrary patterns of rotations on the qubit grid (for example, during with a near-resonant component from F = 2 to F′ = 3 and a small repump
colour-code preparation) are realized with row-by-row serializing, with component. This beam is sent through the side of our glass vacuum
the targeted x coordinates in each row simultaneously illuminated. The cell, co-propagating with the global Raman and 1,013-nm Rydberg
duration of each row is 5–8 μs (corresponding to several tens of μs for beams (Extended Data Fig. 1a). We use cylindrical lenses to shape the
an arbitrary pattern of rotations), which can be sped up considerably, beam, with focused beam waists of 30 μm in the plane of the atom array
as discussed in the next section. For simplicity, we carefully calibrate and 80 μm out of the plane. After moving some of the atoms to this
rotations on 80–160 specific sites across the array, but also perform readout zone, we first perform local pushout of population in the F = 2
rotations in arbitrary spots using the nearest calibrated values. ground-state manifold (by turning off the repump laser frequency),
With the local single-qubit gates and entangling-zone two-qubit followed by local imaging of the remaining F = 1 population.
gates calibrated, the entire circuit is simply defined by the appropriate As depicted in Extended Data Fig. 3a, we collect an average of about
trapping SLM phase profile and waveforms for our several AWG chan- 50 photons per imaged atom. To avoid losing the atoms too quickly
nels and TTL pulse generator. These several channels then program during mid-circuit imaging (which, unlike our global imaging scheme,
Article
does not have multi-axis cooling), we use deep (roughly 5-mK) traps to Steane error correction80, for which errors are intentionally propa-
(helping retain the atoms) and stroboscopically pulse them on and gated from a data logical qubit onto an ancilla logical qubit, which
off out of phase of the local imaging light to avoid deleterious effects is then projectively measured to extract the syndrome of the data
of the deep traps, such as inhomogeneous light shifts and fluctuating logical qubit.
dipole force heating74 (Extended Data Fig. 3b). From a double-Gaussian To perform correlated decoding, we solve the problem of finding
fit to the two distributions in Fig. 3a, we extract an imaging fidelity of the most likely error given the measured syndrome. We start by con-
more than 99.9%. Because this fit can lead to an overestimate of the structing a decoding hypergraph based on a description of the logical
imaging fidelity (for example, owing to atom loss during imaging), algorithm, which describes how each physical error mechanism (for
we compare the total SPAM error (measured by the amplitude of the example, a Pauli-error channel after a two-qubit gate) propagates onto
Ramsey fringe) with local imaging versus with global imaging for the the measured stabilizers76,81. The hypergraph vertices correspond to
same state-preparation sequence, extracting 0.14(5)% higher error with the stabilizer measurement results. Each edge or hyperedge corre-
local imaging; with these considerations, we conservatively estimate sponds to a physical error mechanism that affects the stabilizers it
a local imaging fidelity of around 99.8%. connects, with an edge weight related to the probability of that error.
Various design considerations facilitate local imaging in the readout Each hyperedge can connect stabilizers both within and between logi-
zone while preserving coherence of the data qubits in the entangling cal qubit blocks (see Fig. 2b). We then run a decoding algorithm that
zone35 (Extended Data Fig. 3e–g). The main sources of decoherence are uses this hypergraph, along with each experimental snapshot, to find
rescattering of photons from the locally imaged atoms, as well as beam the most likely physical error consistent with the measurements. This
reflections and tails of the local imaging beam hitting the data qubits. correction is then applied in software (with the exception of Fig. 4e,
As shown in Fig. 1c, for the 500-μs mid-circuit imaging used in this which is decoded in real time).
work, we are able to achieve unchanged coherence (identical within the Concretely, to construct the hypergraph for a given logical circuit,
error bars) of the data qubits with the local imaging light on as without we perform the following procedure. For each logical algorithm (in
it. To understand these effects more quantitatively, we measure the this section, considering only Clifford gates), we identify a set of N
error probability of the data qubits in the entangling zone while the detectors (vertices of the hypergraph) Di ∈ {0, 1} for i = 1,…, N, which
local imaging beam is on in the readout zone for up to 20 ms and with are sensitive to physical errors occurring during the logical circuit.
higher intensities than used for local imaging in this work. We suppress A detector is either on (1) or off (0) to indicate the presence of an error.
decoherence by light shifting the 780-nm transition of the data qubits For the general case, we let Di = 0 if the ith stabilizer measurement
to be different from that of the locally imaged qubits by several tens of matches the measurement of its backwards-propagated Pauli operator
MHz, as studied in Extended Data Fig. 3f–g. Data qubit decoherence is at a previous time and 1 otherwise (the latter indicates that an error has
further suppressed by the large spatial separation between the readout occurred). In particular, for our surface-code experiments, detectors
zone and the entangling zone, in which intensity from the Gaussian tail in the final projective measurement are computed by comparing the
of the local imaging beam should theoretically fall off rapidly. Even at final projective measurement of the stabilizers with the value of the
large separations, we find that stray beam reflections (for example, ancilla-based stabilizer measurement that occurred before the CNOT
from the glass cell window and other optical elements) can hit the data (note that, owing to our state-preparation procedure, the initial stabi-
qubit region. To mitigate this effect, we displace reflections away from lizer measurement is randomly ±1, but the detector is deterministically
the atom array by angling the local imaging beam as it hits the glass cell zero in the absence of noise). For our 2D colour-code experiments, the
window. The estimated effects of rescattered photons from the imaged initial stabilizers are deterministically +1, so each detector is equal to
atoms, especially with the added relative detuning, is negligible. With zero if the corresponding stabilizer in the final projective measurement
all these considerations, we find that we are able to suppress data qubit is +1. To construct the concrete hypergraph and hyperedge weights, we
decoherence rates to ≲0.1% per 500 μs of local imaging exposure, as then use Stim76 to identify the probability pj (j = 1,…, M) of each error
illustrated in Extended Data Fig. 3h. mechanism Ej in the circuit using a Pauli-channel noise model with
The full mid-circuit readout and feedforward cycle occurs in slightly approximate experimental error rates, along with the detectors that
less than 1 ms, including local pushout, local imaging, readout of the are affected by Ej.
camera pixels, decoding of the logical qubit state on the FPGA and a To find the most likely physical error, we encode it as the optimal solu-
local Raman pulse, which is gated on or off by a conditional trigger tion of a mixed-integer program, a canonical problem in optimization
(Extended Data Fig. 3d). In future work, this approach to mid-circuit with commercial solvers readily available82, similar to previous work in
readout and feedforward can be considerably improved to enable ref. 83. We associate each error mechanism Ej with a binary variable that
mid-circuit readout close to the 100-μs scale75. This method can be is equal to 1 if that error occurred and 0 otherwise. Our goal is then to
directly extended to perform many rounds of measurement and feed- find the error assignment {0, 1}M with maximum total error probability
forward, in which groups of ancilla atoms are consecutively brought (alternatively, the error with the minimum total weight, in which the
to the readout zone throughout a deep quantum circuit. weight of error i is wi = log[(1 − pi)/pi]), subject to the constraint that
the error is consistent with the measured detectors. To be consist-
Correlated decoding ent with the measured detectors, the parity of the error variables for
During transversal CNOT operations, physical CNOT gates are applied all the hyperedges connected to a given detector should match the
between the corresponding data qubits of two logical qubits. These parity of that detector. Concretely, let f be a map from each detec-
physical CNOT gates propagate errors between the data qubits in a tor Di to the subset of error mechanisms that flip its parity. The most
deterministic way: X errors on the control qubit are copied to the target likely error is then the optimal solution to the following mixed-integer
qubit and Z errors on the target qubit are copied to the control qubit program:
(see Extended Data Fig. 4b). As a result, the syndrome of a particu-
M
lar logical qubit can contain information about the errors that have Maximize ∑ log(pj )E j + log(1 − pj )(1 − E j)
occurred on another logical qubit, at the point in time in which the j =1
pair underwent a transversal CNOT operation. We can use the infor-
mation about these correlations and improve the circuit fidelity by
subject to ∑ E j − 2Ki = Di ∀i = 1, …, N
E j ∈ f (D i )
jointly decoding the logical qubits involved in the algorithm. We note
E j ∈ {0, 1} ∀j = 1, …, M
that this is closely related to other recent developments in decoding
entire circuits, or so-called space-time decoding76–79. It is also related Ki ∈ Z≥0 ∀i = 1, …, N
 The objective function evaluates to the logarithm of the probability between 1 and 3 stabilizer errors and also because errors deterministi-
of the assigned error configuration, and each variable Ki ensures that cally propagate between codes during the transversal CNOT gates, such
the sum of the error variables in f(Di) matches Di, modulo 2. Finally, that a single physical error on one code can lead to detected errors on all
we solve the mixed-integer program to optimality using Gurobi, a codes, but which are still all correctable errors. As such, we perform the
state-of-the-art solver82, and apply the correction string associated sliding-scale error detection using the correlated decoding technique
with the error indices j for which Ej = 1 in the optimal assignment. We and set the confidence threshold as a threshold weight of the overall
explore this correlated decoding in more detail, including its conse- correction weight on the decoding hypergraph. For example, in the
quences on error-corrected circuits and the asymptotic runtimes of colour code GHZ experiment, a stabilizer error on all four logical qubits
different decoders (M.C. et al., manuscript in preparation). See sections that is just consistent with a single physical qubit error that propagated
‘Surface code and its implementation’ and ‘Correlated decoding in the to all four logical qubits is in fact a low-weight (or high-probability)
surface code’ for further discussion on the surface code in particular. error, as it corresponds to just a single physical qubit error. If the weight
of hypergraph correction (inversely related to the log of the probabil-
Direct fidelity estimation and tomography ity that a given error mechanism would have occurred leading to the
One challenge with logical qubit circuits is that convenient probes that observed syndrome outcome) is below the cut-off threshold weight,
are accessible with physical qubits may no longer be accessible. The then the measurement is accepted; otherwise, it is rejected. For each
GHZ state studied here provides such an example, as conventional threshold, we then calculate the average algorithm result (y axis of
parity-oscillation measurements cannot be performed84. Instead, we Fig. 3d), as well as the fraction of accepted data (x axis of Fig. 3d).
use a technique known as direct fidelity estimation39, which can be In Fig. 5 with [[8,3,2]] codes, for 3, 6, 24 and 48 logical qubits, we
understood as follows. The target state ψ is the simultaneous eigenstate apply our sliding-scale detection simply as given by the total number
of the N stabilizer generators {Si} and, so, the projector onto the target of stabilizer errors detected, although—as illustrated above—this can
N
state is ∣ψ ⟩ ⟨ψ∣ = ∏i (Si + 1)/2 (which is 1 if Si = 1 ∀ i and 0 otherwise). probably be improved by considering which stabilizer error patterns
Thereby, we can directly measure fidelity by measuring the expectation are more likely to cause an algorithmic failure. For the 12 logical qubits,
values of all terms in this product, which—in other words—refers to to have a more fine-grained sliding scale, for each of the 24 = 16 possible
measuring the expectation values of all elements of the stabilizer group stabilizer outcomes, we calculate the XEB to rank the likelihood that
given by the exponentially many products of all the Si. The logical GHZ each of the observed stabilizer outcomes leads to an algorithmic failure
fidelity is defined as the average expectation value of all measured and then use this ranking when deciding whether a given measurement
elements of the stabilizer group. With our four-qubit GHZ state, with is above/below the cut-off threshold. In Fig. 6b, we set the threshold by
four stabilizer generators {XXXX, ZZII, IZZI, IIZZ}, the 16-element sta- the number of stabilizer errors and in Fig. 6d, to have more fine-grained
bilizer group is given by all possible products: {IIII, ZZII, IZZI, IIZZ, ZIIZ, sliding-scale information, we take different subsets of stabilizer out-
IZIZ, ZIZI, ZZZZ, XXXX, XYYX, YXXY, XXYY, YYXX, YXYX, XYXY, YYYY}. come events that are all below the threshold of the allowed number
We measure the expectation values of all 16 of these operators; for each of stabilizer errors and calculate the y axis (Pauli expectation value)
element, we simply rotate each logical qubit into the appropriate and x axis (purity) for all of them. Broadly, there are many ways to per-
logical basis and then calculate the average parity of the four logical form this sliding-scale error detection, and this can be useful both as
qubits in this measurement configuration. We then directly average continuous trade-offs between fidelity and acceptance probability, as
all 16 elements equally (with appropriate signs, as some of the stabilizer well as for use in techniques such as zero-noise extrapolation in data
products should have −1 values) and, in this way, compute the logical analysis (Fig. 6d).
GHZ state fidelity. This is an exact measurement of the logical state
fidelity39. Scaling to larger states can be achieved by measuring ele- Overview of QEC methods
ments of the stabilizer group at random39. To perform full tomography Here we provide a brief overview of key QEC methods used in our work.
in Fig. 3e, we measure in all 34 = 81 bases, thereby measuring the expec-
tation values of all 256 logical Pauli strings, and reconstruct the density Code distance, decoding and thresholds. [[n,k,d]] notation describes
matrix by solving the system of equations with optimization methods. a code with several physical qubits n, several logical qubits k and a code
distance d. The code distance d sets how many errors a code can detect
Sliding-scale error detection or correct. The code distance is the minimum Hamming distance be-
Here we provide more information about the sliding-scale error- tween valid codewords (logical states), that is, the weight of the smallest
detection protocol applied for Figs. 3, 5 and 6. Typically, error detec- logical operator85. In the case of the 2D surface and colour codes studied
tion refers to discarding (or postselecting) measurements in which here, d is equivalent to the linear dimension of the system24.
any stabilizer errors occurred. In the context of an algorithm, however, Following this definition, quantum codes of distance d can detect
discarding the result of an entire algorithm if just one physical qubit any arbitrary error of weight up to d − 1. Such errors cause stabilizer
error occurred may be too wasteful and we may want to only discard violations, indicating that errors occurred. Postselecting on the results
measurements in which many physical qubits fail and the probability with no such stabilizer violations corresponds to performing error
of algorithm success is greatly reduced. For this reason, for the algo- detection, which protects the quantum information up to d − 1 errors
rithms here, we explore error detection on a sliding scale, for which we at the cost of postselection overhead. Conversely, codes can correct
can set a desired ‘confidence threshold’ such that, on the basis of the fewer errors than they detect (but without any postselection overhead).
syndrome outcomes, we determine whether to accept a given measure- The correction procedure brings the system back to the closest logical
ment. Sliding this confidence threshold enables a continuous trade-off state (codeword); thus, if more than d/2 errors occur, the resulting state
(in data analysis) between the fidelity of the algorithm and the accept- may be closer to a codeword different from the initial one, resulting
ance probability. When sliding-scale error detection is applied, in in a logical error85. For this reason, codes of distance d can correct any
all applicable cases, we also apply error correction to return to the arbitrary error of weight up to (d − 1)/2 (rounded down if d is even24). The
codespace. process of decoding refers to analysing the observed pattern of errors
We apply such a sliding-scale error detection for the colour-code and determining what correction to apply to return to the original
logical GHZ fidelity measurements in Fig. 3d. One possible method code state and undo the physical errors created. In many cases, such
would be to discard measurements based on the number of detected as with the 2D surface and colour codes, one does not need to apply
stabilizer errors. However, this is suboptimal, both because on the the correction in hardware (physically flipping the qubits); instead, it
colour code a single physical qubit error can result from anywhere is sufficient to undo an unintended XL/ZL operator that was applied by
Article
hardware errors by simply applying a ‘software’ XL/ZL operator24, also colour code (Figs. 3 and 4) and d = 3 surface code (part of Fig. 2), but
described as Pauli frame tracking86. is non-fault-tolerant for the state preparation of our d = 5, 7 surface
As the size of an error correcting code and the corresponding code codes and [[8,3,2]] codes. Thus, all of our experiments with the d = 3
distance is increased, so are the opportunities for errors to occur as the colour codes are fault-tolerant from beginning to end, and so the entire
number of physical qubits increases. This leads to a threshold behaviour algorithm is fault-tolerant and theoretically has a failure probability
in QEC: if the density of errors p is above a (possibly circuit-dependent) that scales as p2. However, we note that having a fault-tolerant algorithm
characteristic error rate pth, then increasing code distance will worsen also does not imply that errors do not build up during execution of the
performance. However, if p < pth, then increasing code distance will circuit. For this reason, deep circuits require repetitive error correc-
improve performance24. Theoretically, because we require (d + 1)/2 tion6,88 to constantly remove errors and continuously benefit from, for
errors to create a logical error, the logical error rate will be exponen- example, the p2 suppression.
tially suppressed as ∝(p/pth)(d+1)/2 at sufficiently low error rates24. The Our logical GHZ state theoretically has a failure probability scaling
performance improvement with increasing code distance, observed as p2. Nevertheless, the error build-up (increasing p) during the opera-
for the preparation and entangling operation in Fig. 2, implies that tions of the circuit and the spreading of errors through transversal gates
we surpass the threshold of this circuit. We note that, in this regime, limits our logical GHZ fidelity to 72%. This is consistent with numeri-
improving fidelities by, for example, a factor of 2× can then lead to an cal modelling. Similar to the surface-code modelling (Extended Data
error reduction of 24 = 16× for the distance-7 code studied and further Fig. 4), we use empirical error rates consistent with 99.4% two-qubit gate
exponential suppression with increasing code distance. This rapid fidelity, as well as roughly 4% data qubit decoherence error (including
suppression of errors with reduced error rate and increased code dis- SPAM) over the entire circuit. We simulate the experimental circuit
tance is the theoretical basis for realizing large-scale computation. We (including the fault-tolerant state preparation with the ancilla logical
emphasize that thresholds can be circuit-dependent, as discussed in flag) and measurements of all 16 elements of the stabilizer group (see
detail in the surface-code section below. the ‘Direct fidelity estimation and tomography’ section), and extract
a simulated logical GHZ fidelity of 79%. This is slightly higher than our
Fault tolerance and transversal gates. A common definition of fault measured 72% logical GHZ fidelity, possibly originating from imperfect
tolerance in quantum circuits85 (which we use in this work) is that a experimental calibration. This modelling indicates that our logical GHZ
weight-1 error (that is, an error affecting one physical qubit) cannot fidelity is limited by residual physical errors, which will be reduced
propagate into a weight-2 error (now affecting two physical qubits) quadratically as p2 with reduction in physical error rate p, in particular
within a logical block. This property implies that errors cannot spread by reducing residual single-qubit errors, which were larger during this
within a logical block and thereby prevents a single error from growing measurement and are dominating the error budget here.
uncontrollably and causing a logical error.
Distance-3 codes, which are of notable historical importance3,87, can Surface code and its implementation. In 2D planar architectures,
correct any weight-1 error. Fault tolerance is particularly important for such as those associated with superconducting qubits6,88, stabilizer
these codes because otherwise a weight-1 error can lead to a weight-2 measurement is the most important building block of error-corrected
error and thereby cause a logical fault. An important characteristic circuits24. In such systems, stabilizers need to be constantly measured
of a fault-tolerant circuit that uses distance-3 codes85 is that (in the to correct qubit dephasing and increase coherence time, as demon-
low-error-rate regime) physical errors of probability p lead to logical strated recently6. Logic operations are implemented by changing
errors with probability ∝p2. We emphasize that the notion of fault toler- stabilizer measurement patterns, enabling realization of techniques
ance refers to circuit structuring to control propagation of errors, but a such as braiding24 and lattice surgery89. Similar techniques can be used
circuit can be fault-tolerant with low fidelity or non-fault-tolerant with to move logical degrees of freedom to implement nonlocal logical
high fidelity. For example, even if a weight-1 error can lead to a weight-2 gates23. Owing to this gate-execution strategy, d rounds of stabilizer
error but the code has high distance, or if this error-propagation measurement are required for each entangling gate for ensuring fault
sequence is possible but highly unlikely, then this property may not tolerance24.
be of practical importance (for this reason, definitions of fault tol- Neutral-atom quantum computers feature different challenges and
erance may vary). In practice, the goal of QEC is to execute specific opportunities. Specifically, they feature long qubit coherence times
algorithms with high fidelity, and fault-tolerant structuring of a circuit (T2 > 1s), which can be further increased to the scale of tens to hun-
is one of many tools in the design and execution of high-fidelity logical dreds of seconds with well-established techniques72. By using the stor-
algorithms. age zone, qubits can be idly and safely stored for long periods without
Transversal gates, defined here as being composed of independent repeated stabilizer measurements. Hence, from a practical perspective,
gates on the qubits within the code block (that is, entangling gates are increasing qubit coherence by using a logical encoding does not pro-
not performed between qubits within the same code block)42, consti- vide immediate gains in improving quantum algorithms and the gains
tute a direct approach to ensure fault-tolerant structuring of a logical will be from improving the fidelity of entangling operations. Moreover,
algorithm. Because transversal gates imply performing independent logic gates and qubit movement do not have to be performed with
operations on the physical constituents of a code block, errors cannot stabilizer measurements. Instead, they can be executed with nonlocal
spread within the block and fault tolerance is guaranteed. In this work, atom transport and transversal gates. Because such transversal gates
all logical circuits we realize (following the logical state preparation) are intrinsically fault-tolerant, they do not necessarily require d rounds
are fault-tolerant, as all logical operations we perform are transversal. of correction after each operation. Even syndrome measurement may
Note, in particular, that even though the transversal CNOT allows errors be better executed in certain cases by techniques such as Steane error
to propagate between code blocks, this is still fault-tolerant, as it does correction80 (similar to our ancilla logical flag with colour codes as used
not lead to a higher-weight error within the block and, thereby, a single in Fig. 3), as opposed to repeated stabilizer measurement. For these
physical error can neither lead to a logical failure nor an algorithmic reasons, the transversal CNOT is among the most important building
failure. Notably, the large family of codes referred to as Calderbank– blocks in error-corrected circuits. Hence, we focus here on improving
Shor–Steane (CSS) codes all have a transversal CNOT (ref. 2), all of the transversal CNOT by scaling code distance.
which can be implemented with the single-step, parallel-transport Specifically, we use the so-called rotated surface code6, which has
approach here. code parameters [[d2,1,d]]. Our distance-7 surface codes (as drawn in
Although all the logical circuits we implement are fault-tolerant, Fig. 2d) are composed of 49 physical data qubits, with 24 X stabilizers
the logical qubit state preparation is fault-tolerant for our d = 3 (light-blue squares) and 24 Z stabilizers (dark-blue squares), and one
encoded logical qubit described by anticommuting weight-7 operators, is only 2). Our correlated decoding technique is thus essential to our
the horizontally oriented XL and the vertically oriented ZL. The X and observation of improved Bell performance with code distance.
Z stabilizers commute with the XL and ZL logical operators, allowing Finally, we elaborate on our evaluation of Bell-pair error. Bell-state
the measurement of the stabilizers without disturbing the underlying fidelity is given by the average of the populations and the coherences,
logical degrees of freedom. In our experiments, we prepare one surface which—for physical qubits—can be measured as the ZZ populations
code in |+L⟩ and one surface code in |0L⟩. In the first code, this is realized and the amplitude of parity oscillations. In the language of stabiliz-
by preparing all physical data qubits in |+⟩, thereby preparing an eigen- ers, the parity oscillation amplitude is given by the average of ⟨XX⟩
state of XL and the 24 X stabilizers, and then projectively measuring and −⟨YY⟩ (ref. 90). With the surface code, we cannot conveniently
the 24 Z stabilizers with 24 ancilla qubits (Fig. 2d red dots) using four measure the YL operators fault-tolerantly (and that is why we use colour
entangling-gate pulses24. The second code is prepared similarly but codes for programmable Clifford algorithms and full tomography; see
with all physical qubits initialized in |0⟩, thus preparing an eigenstate next section). For this reason, we estimate the logical coherences as
of ZL and the 24 Z stabilizers, and then projectively measuring the 24 ⟨XLXL⟩, which we then average with the populations for calculating the
X stabilizers with 24 ancillas. The CNOT is directly transversal because Bell-pair error. To support the validity of this analysis, we can instead
these two surface-code blocks have the same orientation and does calculate a lower bound on the Bell-state fidelity90, which also shows
not require rotation of the lattice to implement a H. The projective the same improvement in performance as we increase code distance
measurement of the ancillas defines the values of the stabilizers. Dur- (Extended Data Fig. 4d).
ing the transversal CNOT, the values of the stabilizers are copied onto
the other code as well and is tracked in software. Correlated decoding in the surface code. Following the above dis-
Because we only perform a single round of stabilizer measurement, cussion, we provide more insights related to the correlated decoding
our state-preparation scheme is nFT for the d = 5, 7 codes. Consider, for in the case of the surface-code transversal CNOT. Consider a circuit in
instance, the case when all stabilizers are defined as +1 and no errors which perfect (noiseless) surface codes are initialized, a transversal
are present in the system, but an ancilla measurement error in the CNOT is executed and then projective readout is performed. If errors
middle of the surface-code lattice yields a stabilizer measurement occur before the transversal CNOT, then these errors can propagate; for
of −1. Correction then causes a large-weight pairing of this apparent example, an X physical error on the control logical qubit will propagate
stabilizer violation to the boundary4. Hence, this single ancilla meas- onto the target logical qubit and thereby double the density of errors
urement error can lead to several data qubit errors, resulting in nFT on the target logical qubit. By multiplying the projectively measured
operation. The d = 3 code initialization is a special case that does not Z stabilizers of the target logical qubit with those of the control logical
suffer from this issue38. Higher-order considerations about fault toler- qubit, the propagation is undone. Now the target logical qubit only has
ance given by gate ordering during stabilizer measurement can also be to decode its original density of X errors. The same considerations can
considered6. be made for Z errors originating on the target logical qubit that propa-
The effect of these nFT errors from noisy syndrome extraction is to gate onto the control logical qubit. However, if there are errors after the
cause X physical errors on the |+L⟩ state and Z physical errors on the transversal CNOT, then multiplying the stabilizers instead doubles the
|0L⟩ state. Thus, in performing just a SPAM measurement, the presence density of such errors. Thus, the optimal decoding strategy if errors are
of these errors would not be directly apparent, as these errors commute only after the transversal CNOT is to perform independent matching
with measuring the |+L⟩ in the X basis and |0L⟩ in the Z basis. As such, within both codes. The general case in which errors are present both
this circuit would not be a good benchmark of surface-code state before and after the transversal CNOT corresponds to neither case and
preparation. Conversely, the transversal CNOT experiment is sensitive is modelled by our decoding hypergraph that has edges and hyperedges
to the various aspects of the circuit and a good indication of perfor- connecting the two logical qubits, with edge weights informed by our
mance. Because we measure the Bell state in both the XL1XL2 and ZL1ZL2 experimental error model. Extended Data Fig. 5 explores decoding
bases, the nFT errors in both bases will propagate through the logical performance with different values of the scaled weights of the edges
CNOT and cause errors on both logical qubits in both the X and Z bases. and hyperedges that connect the stabilizers of the two logical qubits.
For these reasons, unlike a surface-code SPAM measurement, this These results illustrate that the correlated decoding is robust (but not
experiment is a good indication of logical performance. In fact, the completely insensitive) to the nFT errors associated with ancilla meas-
effect of these nFT errors is such that, if we just apply conventional urement errors. This feature would also be recovered by the simpler
decoding within each logical block, then we find that the Bell state multiplication decoder, which would be entirely insensitive to errors
degrades substantially with increased code distance (Fig. 2d). from ancilla measurement, but is—however—more sensitive to errors
The effects of this nFT preparation are suppressed (but not entirely after the CNOT. Specifically, Extended Data Fig. 5c shows that our op-
removed) by using the correlated decoding technique. For example, timized decoder is not simply a ‘multiplication decoder’, as the ancilla
consider a nFT-induced apparent stabilizer violation to the left of the measurement values indeed contribute to the correction procedure and
middle line in the lattice of the d = 7 |+L⟩ state, corresponding to a chain make the correlated decoding more robust to decoder parameters. For
of three physical X errors to the boundary. These errors will propagate a given logical circuit, our correlated decoding procedure generates
through the logical CNOT onto the second logical qubit and affect the a decoding hypergraph, which we then solve using most likely error
independent measurement of both logical qubits in the Z basis when methods, which is done here for both surface-code and colour-code
investigating the ZL1ZL2 stabilizer. When decoded independently, if experiments, and can generically be applied to any stabilizer codes and
another single X error occurs on the first block after the CNOT moving Clifford circuits79. More theoretical details and discussion of correlated
the stabilizer violation to the right of the middle line, becoming a chain decoding will be presented in M.C. et al, manuscript in preparation.
of four X physical errors, this will cause an incorrect pairing and lead
to an independent XL1 error on this code only and thereby corrupt the 2D colour codes. 2D colour codes are topological codes that are simi-
ZL1ZL2 stabilizer and would correspond to a total weight-6 correction lar to surface codes91. Often portrayed in a triangular geometry, the
between the two codes. However, when decoded jointly with correlated colour codes used here are a tiling of three colours of weight-4 and
decoding, these errors can be effectively decoded, as they will appear weight-6 stabilizers, with XL and ZL operator strings running along the
on the stabilizers of both logical qubits. In this example, the boundary of the code91. In this work, we study 2D d = 3 colour codes,
lowest-weight pairing would remove this chain of three X errors from as portrayed in Fig. 3a, which only contain weight-4 stabilizers given
both codes and leave only the one remaining X error on the first block, by the products of X and Z on the qubits of each coloured plaquette.
which can also be decoded successfully (the total pairing weight here This d = 3 colour code is identical to the seven-qubit Steane code.
Article
However, we emphasize that the techniques used here directly apply superposition of Pauli strings, that is, X ⊗ I ⊗ I → 1/2(X ⊗ I ⊗ I + X ⊗
to larger-distance colour codes92. Z ⊗ I + X ⊗ I ⊗ Z − X ⊗ Z ⊗ Z), as an X flip now changes whether a CZ
Although the colour codes are similar to surface codes, an important operator will be applied on the other qubits, resulting in four times
difference is that, in the colour code, the X and Z stabilizers lie directly more operators to track after the single CCZ. (The CZ operator matrix
on top of the same qubits (as opposed to being on dual lattices with is simply equal to 1/2[I ⊗ I + Z ⊗ I + I ⊗ Z − Z ⊗ Z]). This causes not only
respect to each other) and, similarly, the XL and ZL operators lie on top operator spreading but also so-called operator entanglement94. As
of each other (as opposed to propagating in the orthogonal directions we apply further non-Clifford gates, the number of operators to track
on the surface code). In other words, the operators here are symmetric will grow exponentially and eventually will become computationally
and related by a global basis transformation. This has important conse- intractable. For example, state-of-the-art Clifford + T simulators can
quences for the allowed transversal gate set41,93. In particular, although handle roughly 16 CCZ gates50. This is the basis behind our complex
the surface code technically has a transversal H that transforms XL ↔ ZL, sampling circuits, in which the 48 CCZs on the 48 logical qubits cre-
it requires a physical 90° rotation of the code block. Although such ate a high degree of scrambling and magic (defined below), rendering
lattice rotations are possible using atom-motion techniques, for many Clifford + T simulation impractical.
circuits, it is inconvenient. Conversely, in the colour code, H is transver-
sal: it directly exchanges XL ↔ ZL as well as the X and Z stabilizers. This [[8,3,2]] circuit implementation
difference is even more important for the transversal S gate, which is Here we provide more detail about our [[8,3,2]] circuit implementa-
possible for the colour code. Here transversal S exchanges XL ↔ YL (for tions. The [[8,3,2]] code blocks are initialized in the |−L,+L,−L⟩ state with
which YL is given by the product of XL and ZL, which lie on top of each the circuit in Extended Data Fig. 6, which can be understood as prepar-
other) as intended, and the X stabilizer of a given plaquette returns ing two four-qubit GHZ states (corresponding to [[4,2,2]] codes95), that
to itself by multiplying the Z stabilizer of that same plaquette. (This is, GHZ1,3,5,7
Z ⊗ GHZX2,4,6,8, and subsequently entangling them as illus-
is in contrast to the surface code, which does not have a transversal trated in Extended Data Fig. 6a (as well as applying Z gates). In our
S, for which the YL operator is a product of horizontally propagating circuit implementations, for system sizes of 3–24 logical qubits for
XL and vertically propagating ZL (ref. 24)). Because the colour code both sampling and two-copy measurements, we prepare eight blocks
has the entire transversal gate set of {H, S, CNOT} and also does not encoded over 64 physical qubits. For the 48-logical-qubit circuit (128
require tracking any lattice rotations, it is well suited to exploration of physical qubits in total), we encode eight blocks and entangle them,
programmable logical Clifford algorithms. and then drop them into storage; then, we pick up 64 new physical
For fault-tolerant preparation of the d = 3 colour code, we use a modi- qubits from storage, encode them into eight blocks in the entangling
fied version of the scheme summarized in ref. 38, in which, instead of the zone and entangle them. Finally, we bring the original eight blocks
eight-gate encoding circuit, we use a nine-gate encoding circuit that is from storage and entangle them with the second group of eight blocks
more conveniently mapped to specific atom movements in our system in the entangling zone (Extended Data Fig. 6) (see Supplementary
(corresponding to graph-state preparation similar to ref. 7), followed by Video).
a transversal CNOT with an ancilla logical flag. The logical SPAM fidelity The transversal gate set of the [[8,3,2]] code is enabled as follows
is then calculated as the probability of observing |0L⟩ after decoding. (see also refs. 16,17,26,27). The transversal CNOT between blocks imme-
We note that, in Fig. 3, we could also have made a five-qubit GHZ state diately follows from the fact that the [[8,3,2]] code is a CSS code.
but instead made a four-qubit GHZ state for simplicity of performing In-block CZ gates between two logical qubits Li and Lj (CZLi,Lj) can be
full tomography. In Fig. 4, when Bell-state fidelities with feedforward realized by S, S† gates on the face corresponding to logical qubit Lk. For
are reported, we estimate the logical coherences as the average of example, consider applying the pattern of S, S† gates to the top face in
⟨XLXL⟩ and −⟨YLYL⟩, which we then average with the ZLZL populations Fig. 5, that is, S 1S3†S †5S 7, which transforms XL1 = X1X2X3X4 to X′L1 = − Y1X 2Y3X 4 ,
(not plotted) for calculating the Bell-pair fidelity. Finally, we note that which is equal to X′L1 = XL1ZL2 , and the same applies to give X′L2 = XL2ZL1 ,
the feedforward Bell state in Fig. 4e could also be performed with a that is, a CZ is realized between logical qubits 1 and 2. This procedure
software ZL rotation on either of the two qubits, allowing correction can also be used to understand why the pattern of T, T† realizes a CCZ
to the appropriate Bell state, but here we perform the feedforward S between the three encoded qubits. CCZ gates should map XL3 to
on both qubits to test our feedforward capabilities; this technique XL3 ⊗ CZL1,L2. By applying the pattern of T, T† in Fig. 5a, each X face maps
is directly compatible with performing magic-state teleportation24. to itself multiplied by a pattern of S, S†, for example, XL3 = X1X3X5X7
maps to X′L3 = XL3S 1S3†S †5S 7, or then X′L3 = XL3 ⊗ CZL2,L3. This happens for
Clifford and non-Clifford gates and universality. 2D topological all three XL faces, thereby realizing a CCZ gate. Finally, we detail the
codes such as the surface and colour codes have transversal imple- permutation CNOT, which was also developed in ref. 27. Physically
mentation of Clifford gates (for example, {H, S, CNOT}). This gate set permuting atoms to swap qubits 4 ↔ 8 and 3 ↔ 7 takes XL1 = X1X2X3X4
is not universal, that is, it cannot alone be used to realize an arbitrary to X′L1 = X1X 2X7X 8 or instead X′L1 = XL1XL2 (also by multiplying the global
quantum computation and requires a non-Clifford gate such as {T, CCZ} X stabilizer) and, similarly, it can be seen by tracking the qubit permu-
for achieving universal computation. Moreover, circuits composed tations that Z′L2 = ZL2ZL1, that is, realizing a CNOT. Finally, although these
solely of stabilizer states and Clifford gates can be simulated in poly- 3D codes do not have a transversal H, because they are CSS codes, they
nomial time because of the Gottesman–Knill theorem44. This can be can be initialized and measured in either the X or the Z basis, effectively
understood as stabilizer tracking; for example, consider a three-qubit allowing H gates at the beginning or end of the circuit.
system in which a stabilizer of the state is X ⊗ I ⊗ I, such that X stabilizes In-block logical entangling gates are applied block by block and any
the |+⟩ state and I is the identity. Applying two CZ entangling gates in-block gate combination can be realized. For conceptual simplicity,
CZ1,2 ⊗ CZ1,3 transforms this stabilizer to X ⊗ Z ⊗ Z because an X flip we apply only two particular local Raman patterns in layers. The first
before the CZ simply changes whether a Z flip will be applied to the is the gate combination CCZL1,L2,L3 ⋅ CZL1,L2 ⋅ CZL1,L3 ⋅ CZL2,L3 ⋅ ZL1 ⋅ ZL2 ⋅ ZL3,
other qubits. Even though Clifford circuits create superposition and given by applying T† on the entire physical qubit block, and the second
entanglement between qubits, the N initial stabilizers of the state can gate combination we apply is CCZL1,L2,L3 ⋅ CZL2,L3 ⋅ CZL1,L3 ⋅ ZL3, given by
simply be tracked as they propagate through the circuit (so-called applying T on the top row and T† on the bottom row. In our circuits,
operator spreading94) and thereby simulation of the circuit can be we alternate layers of in-block transversal entangling gates and
easily accomplished. out-block transversal CNOTs, entangling logical blocks on up to 4D
The effect of non-Clifford gates, however, is far more complex. For hypercubes19,96,97 (see Extended Data Fig. 6). We keep the control and
example, passing the stabilizer X ⊗ I ⊗ I through a CCZ maps into a target qubits the same throughout the circuit for conceptual
simplicity, allowing the local physical H gates on the target qubits to Moreover, the XEB turns out to be a better benchmark for IQP circuits
be compiled with the in-block gate layers, but the control-target direc- than for generic random circuit sampling settings (such as Haar-random
tion can also be chosen arbitrarily. We ensure that in-block logical circuits)102–104. For IQP, the XEB is close to the many-body fidelity and
entangling gates are applied such that they do not trivially commute the difference can be theoretically bounded under reasonable noise
through and cancel with earlier entangling-gate applications. As an assumptions (M.K. et al., manuscript in preparation). Intuitively, this
experimental note, we remark that, for the Clifford states realized in fact is related to the diagonal structure of the IQP circuits, which allows
the other parts of this work, stabilizers take on values of either +1 or −1 the XEB to capture errors in a manner closer to fidelity, despite being
(owing to, for example, use of physical π/2 rotations instead of H), defined only in the computational basis. In other words, a Z error will
which is then simply redefined in software. Because, for our [[8,3,2]] always corrupt the X-basis measurement, and an X error (except one
circuits, we implement non-Cliffords on the physical level, it is impor- immediately before measurement) will create new Z errors that also cor-
tant to ensure that all stabilizers are initialized and maintained as +1; rupt the X-basis measurement. Thus, in the fully postselected regime,
for example, if a Z stabilizer is −1, then the logical CCZ implementation in which errors at the end of the circuit are well described by logical
sends the X stabilizer expectation value to 0. This can be understood errors, we expect the XEB to be a good measure of fidelity. We further
as a physical X on a single site transforming to a superposition (X + Y)/ 2 note that, as well as the efficient generation of complex IQP circuits
by physical T, going into an equal superposition of X stabilizer being here, the [[8,3,2]] gate set presented here can realize arbitrary IQP
+1 and −1. circuits composed of {CCZ, CZ, Z} gates105. The in-block {CCZ, CZ, Z}
operations can be applied to any groupings of qubits by noting that
Classically hard circuits with [[8,3,2]] codes combining the in-block and out-block CNOTs allow us to compose
Our implemented circuits are equivalent to IQP circuits98, which arbitrary transversal SWAP operations of targeted individual logical
gives a theoretical basis for understanding why our circuits could be qubits between different blocks.
classically hard to simulate, for which we also provide numerical evi-
dence of so-called anticoncentration99,100. IQP circuits are defined as Simulation of bitstring probabilities
initializing |+⟩⊗n on n qubits, applying a diagonal entangling unitary To calculate the logical bitstring probabilities necessary for evaluating
such as those composed by {CCZ, CZ, Z} and then measuring in the X the XEB and benchmarking our circuits, we use a hybrid simulation
basis20,98. A uniform superposition of 2n bitstrings is created, the diago- approach combining wavefunction and tensor-network106 methods.
nal gates apply −1 signs in a complicated fashion to the exponentially It works best only when performing all of the entangling gates of the
many bitstrings and then ‘undoing the superposition’ with the final hypercube a single time and relies on the fact that the final round of
H before measurement now results in an intricate ‘speckle’ interfer- CNOTs is immediately followed by a measurement, simplifying network
ence pattern18. Sampling from the output distribution of this speckle contraction. Concretely, for a D-dimensional logical hypercube, the
pattern can be done efficiently on a quantum device that implements two subsystems consisting of 2D−1 blocks are simulated independently
the circuit, but is exponentially costly on a classical device for certain and then the final layer of CNOTs and in-block operations is combined
choices of IQP circuits20. The transversal gate set of the [[8,3,2]] code, with the measurement outcomes (the bitstring of interest), which
D−1
as described above, contains diagonal gates {CCZ, CZ, Z} that apply −1 results in a contraction of two 8 2 tensors (see Extended Data Fig. 8b).
signs to the bitstrings, but is made non-diagonal by the application of This is a square-root reduction in the memory requirement compared
D
CNOTs, which permute bitstrings. Because this bitstring permutation with the full wavefunction simulation, which uses O(8 2 ) space. The
does not break the IQP framework, these circuits are equivalent to an ideal XEB value is calculated by sampling bitstrings from the ideal out-
effective IQP circuit, but which is much more complex: for example, put distribution and then averaging the corresponding probabilities.
circuits with 48 CCZs and 96 CNOTs map to effective IQP circuits with The bitstrings are sampled using a marginal sampling algorithm, which
roughly 1,000 CCZ gates. Nevertheless, because IQP circuits are a well- uses the same contraction scheme described above.
understood framework, we can discuss our circuit properties with this We next consider whether the finite XEB scores in this problem can
toolset. be easily ‘spoofed’ by foregoing exact simulation of the implemented
We experimentally explore these circuits with the XEB18, defined as circuit and using a classical algorithm with fewer resources, similar in
XEB = 2NLΣi p(xiL)q(xiL) − 1 , in which NL is the number of logical qubits, spirit to the algorithm introduced in ref. 102. For the circuits studied
q(xiL) is the measured probability distribution for our logical qubits in Fig. 5, containing only a single layer of gates on the hypercube, there
and p(xiL) is the calculated probability distribution; here we normalize is only a single round of CNOTs connecting the two 2D−1-block partitions;
the XEB by its ideal value such that the XEB for the noiseless circuit is thus, removing them from the circuit and sampling from the two inde-
1. In typical cases, if noise overwhelms the circuit, the measured distri- pendent halves might not decrease the XEB substantially while reduc-
D−2
bution will be uniform18 and the measured XEB will be 0. ing the memory requirement to 8 2 . In Extended Data Fig. 8c, we study
The IQP circuits are a good setting for quantum-advantage-type the performance of this spoofing attack and find that the obtained XEB
experiments, as the bitstring distribution of IQP circuits with ran- rapidly decreases, once further gate layers are introduced, for a par-
domly applied {CCZ, CZ, Z} gates (random degree-3 polynomials) ticular extension of our circuit.
is known to be classically hard to simulate20,101. In M.K. et al., manu- The contraction scheme above, used for both the ideal simulation
script in preparation, we show that the ensemble of random hyper- and the XEB spoofing, scales exponentially with the number of qubits.
cube IQP circuits, whose instances are experimentally explored However, the exponent is substantially reduced by using the fact that
here, converges to the uniform IQP ensemble as the depth and size the hypercube circuits can be naturally partitioned into smaller blocks,
of the hypercube is increased. In Extended Data Fig. 8a, we show that with only a single inter-partition layer of CNOTs at the end of the circuit.
hypercube IQP circuits with random in-block operations and rand- This simulation method therefore becomes less efficient if we introduce
omized control-targets on the out-block CNOT layers (realizing the extra CNOT layers (within a single partition) after the inter-partition
hypercube) anticoncentrate quickly as the dimension of the hyper- layer, as we estimate in Fig. 5d. Applying l = {0,…, D − 1} further
cube is increased, with XEB eventually reaching the uniform-IQP intra-partition CNOT layers forces the CNOT tensors in Extended Data
value of 2. We also find that the presence of non-Clifford CCZ gates, Fig. 8b to be blocked into groups of 2l, which results in the execution
l
which are critical for the computational hardness here, further time to scale roughly as O(8 2 /2l ), in which the numerator comes from
improves anticoncentration properties, as we observe that the ideal the tensor contraction complexity and the denominator accounts for
XEB of experimental circuits approach 2 as well, even without much the reduced number of contractions resulting from blocking. The
randomization. explicit times quoted in Fig. 5d as a function of extra CNOT layers are
Article
based on the above matrix-multiplication estimate and fitted such that we find that these complex circuits seem to perform much better with
the depth-1 hypercube time matches 1.44 s, which corresponds to our logical qubits than physical qubits.
implementation. In practice, the actual runtimes might differ owing
to hardware and software optimization and other factors, such as the Two-copy measurements
cost of tensor permutations; however, we expect the general trend to A powerful method to extract various quantities of interest are Bell-
hold. Finally, if the 2l-blocked tensors were to be stored directly, the basis measurements between two copies of the same state21,22,52. First,
l+1
memory requirement of this approach would grow as 8 2 , recovering we use these measurements to calculate the purity or entanglement
2D
the full 8 memory complexity for l = D − 1. entropy of the resulting state7,21,52,112,113. Measuring the occurrences
In this work, we use these circuits and XEB results for benchmarking of the singlet state 01⟩ − 10⟩ (|11⟩ outcome for our measurements
2
our logical encoding, which requires the ability to simulate these cir- after applying the final pairwise entangling operations) probes the
cuits. Future logical-algorithm experimentation can explore quantum- eigenvalue of the SWAP operator sî at a given pair of sites i. This is in
advantage18,48,107–109 tests with encoded qubits, as will be detailed in turn related to the purity of the state by observing that Tr[ ρ 2A]=
M.K. et al., manuscript in preparation. Tr[Πi ∈Asî ρA ⊗ ρA] for any subsystem A. Thus, the average purity can
be estimated by the average parity Tr[ ρ 2A] = ⟨(−1)no. of observed singlets ⟩
Physical qubit circuit implementations within A and, thus, also the second-order Rényi entanglement entropy
To compare our logical-qubit algorithms with analogous circuits on S2(A) = − log 2Tr[ ρ 2A].
physical qubits, we work out a concrete implementation of our sam- The entanglement entropy calculation only involves the singlet out-
pling/scrambling circuits on physical qubits using the same physical comes. By making use of the full outcome distribution, we can also
gate set, Clifford + T, as used in the logical circuit, which we also then evaluate the absolute value of all 4N Pauli strings, from a single dataset,
attempt to realize experimentally. We replace each [[8,3,2]] block with a in which N is the number of qubits involved in each copy of the state22.
three-physical-qubit block, decomposing the ‘in-block’ CCZ gates into More concretely, consider a given Pauli string O = ∏iPi, in which
six CNOTs and seven {T, T†} gates and implement ‘transversal’ CNOTs Pi ∈ {Xi, Yi, Zi, Ii} are individual Pauli operators on site i (and the identity),
directly between the three-qubit blocks. We note that the CZ can be → → →→
and a given observed bitstring {a , b }, in which a , b label the outcomes
compiled into the CCZ implementation, but this has a minor effect in the control and target copy. The rules of reconstructing the Pauli
on our analysis and estimates. These physical circuits are complex: strings through these Bell-basis bitstrings can be worked out through
48 qubits with 48 CCZs and 228 two-qubit gates (as realized with our considering the computational states to which the Bell states are
logical qubits) decomposes into an effective 516 two-qubit gates (384 mapped and considering which operators of XX, YY and ZZ have +1 or
if the CZ gates are compiled into the CCZs). In trying to implement −1 eigenvalue for the various Bell states. We explicitly list the analysis
these circuits in practice, the build-up of coherent errors resulted in procedure: for Pauli term Xi, we assign parity +1 if ai = 0 and −1 otherwise;
a vanishing XEB for our physical circuits. These experiments made for Pauli term Yi, we assign parity +1 if ai ≠ bi and −1 otherwise; for Pauli
it clear that the logical-circuit equivalent was greatly outperform- term Zi, we assign parity +1 if bi = 0 and −1 otherwise; for Ii, we assign
→ → 2
ing the physical circuit, thereby providing direct evidence that our parity +1 always. The contribution of the bitstring {a , b } to tr(Oρ) is
logical algorithm outperforms our physical algorithm for this specific then given by the product of the individual parities.
sampling circuit. We can perform the same analysis as a function of the amount of
More quantitatively, with a concrete physical implementation, we error detection applied. As shown in Extended Data Fig. 9a, as more
calculate an upper bound by assuming optimistic performance. We error detection is applied, the distribution of Pauli expectation values
assume our best-measured fidelities: SPAM of 99.4% (ref. 8), local that are expected to be zero and non-zero separate apart further. This
single-qubit gate fidelity of 99.91% (Extended Data Fig. 2), two-qubit also provides a natural method to perform error mitigation through
gate fidelity of 99.55% (ref. 8) and T2 = 2s. We then count the total num- zero-noise extrapolation: by performing sliding-scale error detection,
ber of entangling-gate pulses for the CZ gates, the total number of we can extract the Pauli expectation value squared for groups of Pauli
compiled local single-qubit gates and the estimated circuit duration, strings with the same expected value, as a function of the logical purity.
and use these to calculate the estimate presented in Fig. 5f. We further We perform a linear fit of the Pauli expectation value squared versus
confirm this analysis for small-scale circuit implementations. For a the logical purity and extrapolate to purity tr(ρ 2 ) = 1, corresponding
short three-qubit circuit, we benchmark the XEB for the physical cir- to the case of zero noise, to estimate the error-mitigated values. The
2
cuit as approximately 0.87, which is below the estimated three-qubit choice of a linear fit is motivated by the fact that both tr(Oρ) and
2
upper bound of roughly 0.92. We note that, in Fig. 5f, we plot estimates tr(ρ ) scale with power 2 of the density matrix. We expect that more
of physical-qubit fidelity and not the XEB, but we expect the XEB and detailed considerations of the noise model, using knowledge about
fidelity to be closely related, as discussed previously. the weight of each operator, as well as whether detected errors in each
We note several observations made in comparing physical and logical shot overlap with a given Pauli operator, can further improve the
implementations of these complex circuits. First, empirically, it seems error-mitigation results.
that the logical circuit is much more tolerant to coherent errors49,110,111, We can also compute measures of distance from stabilizer states,
and understanding the manifestations of this is a subject of continu- also known as ‘magic’, using the additive Bell magic measure in ref. 53,
ing investigation. Specifically, it seems that the logical circuit realizes which only requires O(1) number of samples and O(N) classical
inherently digital operation, for which the small coherent errors do not post-processing time. To do so, we randomly sample subsets of four
substantially shift/distort the bitstring distribution but just reduce the measured Bell-basis bitstrings r, r′, q, q′ and calculate their contribu-
overall fidelity49,110 (see, for example, the agreement in Extended Data tion to the Bell magic using the check-commute method of ref. 53:
Fig. 7a). This is in contrast to the physical implementation, in which B = ∑r, r ′, q, q′ P(r)P(r′)P(q)P(q′) ∥ [σr ⊕r ′, σq ⊕q′] ∥ , with ∥ [σr ⊕r ′, σq ⊕q′] ∥
∞ ∞
coherent errors are seen to substantially alter the shape of the bitstring ∈{0,1}2N
distribution, for example, changing relative amplitudes. Second, we being 0 when the two Pauli strings commute and 2 otherwise. r ⊕ r′
note that we optimize our [[8,3,2]] circuits only by optimizing the sta- denotes bitwise XOR between the two bitstrings. P(r) is the probability
bilizer expectation values and not by optimizing the XEB or two-copy of observing bitstring r. The Pauli string σr is of length N and the ith
result directly. When running complex circuits, the stabilizers serve as element is I, X, Z or Y when the target and control qubit at site i read
useful intermediate fidelity benchmarks, for both optimizing circuit 00, 01, 10 or 11, respectively. We convert this result to additive Bell
design and ensuring proper execution, especially in regimes in which magic through the formula Ba = − log 2(1 − B) . We use approximately
output distributions or other observables cannot be calculated. Overall, 107 samples to estimate the additive Bell magic for each dataset. The
results for the estimated additive Bell magic as a function of the num- 87. Shor, P. W. Scheme for reducing decoherence in quantum computer memory. Phys. Rev.
A 52, R2493 (1995).
ber of non-Clifford gates applied (circuits shown in Extended Data 88. Krinner, S. et al. Realizing repeated quantum error correction in a distance-three surface
Fig. 9f) are shown in Fig. 6c. These results also use the purity estimates code. Nature 605, 669–674 (2022).
in the same dataset, which are used for error mitigation as described 89. Horsman, C., Fowler, A. G., Devitt, S. & Meter, R. V. Surface code quantum computing by
lattice surgery. New J. Phys. 14, 123011 (2012).
in equations (13)–(15) of ref. 53. All additive Bell magic data shown are 90. Tóth, G. & Gühne, O. Entanglement detection in the stabilizer formalism. Phys. Rev. A 72,
with full error detection applied. 022340 (2005).
The same experiments we perform here can also be interpreted as a 91. Kubica, A., Yoshida, B. & Pastawski, F. Unfolding the color code. New J. Phys. 17, 083026
(2015).
physical Bell-basis measurement. Using this insight, in Extended Data 92. Chamberland, C., Kubica, A., Yoder, T. J. & Zhu, G. Triangular color codes on trivalent
Fig. 9c,d, we show the entanglement entropy for different subsystem graphs with flag qubits. New J. Phys. 22, 023019 (2020).
sizes, when analysing the data as physical Bell-pair measurements 93. Kubica, A. & Beverland, M. E. Universal transversal gates with color codes: a simplified
approach. Phys. Rev. A 91, 032330 (2015).
and applying different levels of stabilizer-based postselection. Nota- 94. Mi, X. et al. Information scrambling in quantum circuits. Science 374, 1479–1483
bly, the full-system parity when postselecting on all stabilizers being (2021).
correct is identical when analysing the outcomes as either a physi- 95. Linke, N. M. Fault-tolerant quantum error detection. Sci. Adv. 3, e1701074 (2017).
96. Hashizume, T., Bentsen, G. S., Weber, S. & Daley, A. J. Deterministic fast scrambling with
cal or a logical circuit. This is because, in this limit, the results of the neutral atom arrays. Phys. Rev. Lett. 126, 200603 (2021).
physical-circuit analysis can be viewed as taking the (imperfect) logical 97. Jia, Y. & Verbaarschot, J. J. Chaos on the hypercube. J. High Energy Phys. 2020, 154 (2020).
state and running a perfect encoding circuit, hence giving identical 98. Bremner, M. J., Jozsa, R. & Shepherd, D. J. Classical simulation of commuting quantum
computations implies collapse of the polynomial hierarchy. Proc. R. Soc. A Math. Phys.
results. Eng. Sci. 467, 459–472 (2011).
99. Hangleiter, D., Bermejo-Vega, J., Schwarz, M. & Eisert, J. Anticoncentration theorems for
schemes showing a quantum speedup. Quantum 2, 65 (2018).
100. Bouland, A., Fefferman, B., Nirkhe, C. & Vazirani, U. On the complexity and verification of
Data availability quantum random circuit sampling. Nat. Phys. 15, 159–163 (2019).
The data that support the findings of this study are available from the 101. Bremner, M. J., Montanaro, A. & Shepherd, D. J. Achieving quantum supremacy with
sparse and noisy commuting quantum computations. Quantum 1, 8 (2017).
corresponding author on request.
102. Gao, X. et al. Limitations of linear cross-entropy as a measure for quantum advantage.
Preprint at https://arxiv.org/abs/2112.01657 (2021).
103. Morvan, A. et al. Phase transition in random circuit sampling. Preprint at https://arxiv.org/
62. Barredo, D., De Léséleuc, S., Lienhard, V., Lahaye, T. & Browaeys, A. An atom-by-atom abs/2304.11119 (2023).
assembler of defect-free arbitrary two-dimensional atomic arrays. Science 354, 1021–1023 104. Ware, B. et al. A sharp phase transition in linear cross-entropy benchmarking. Preprint at
(2016). https://arxiv.org/abs/2305.04954 (2023).
63. Levine, H. et al. Dispersive optics for scalable Raman driving of hyperfine qubits. Phys. 105. Shepherd, D. & Bremner, M. J. Temporally unstructured quantum computation. Proc. R.
Rev. A 105, 032618 (2022). Soc. A Math. Phys. Eng. Sci. 465, 1413–1439 (2009).
64. Jandura, S. & Pupillo, G. Time-optimal two- and three-qubit gates for Rydberg atoms. 106. Pan, F. & Zhang, P. Simulation of quantum circuits using the big-batch tensor network
Quantum 6, 712 (2022). method. Phys. Rev. Lett. 128, 030501 (2022).
65. Pagano, A. et al. Error budgeting for a controlled-phase gate with strontium-88 Rydberg 107. Boixo, S. et al. Characterizing quantum supremacy in near-term devices. Nat. Phys. 14,
atoms. Phys. Rev. Res. 4, 033019 (2022). 595–600 (2018).
66. Lengwenus, A., Kruse, J., Schlosser, M., Tichelmann, S. & Birkl, G. Coherent transport 108. Zhong, H.-S. et al. Quantum computational advantage using photons. Science 370,
of atomic quantum states in a scalable shift register. Phys. Rev. Lett. 105, 170502 1460–1463 (2020).
(2010). 109. Madsen, L. S. et al. Quantum computational advantage with a programmable photonic
67. Schlosser, M., Tichelmann, S., Kruse, J. & Birkl, G. Scalable architecture for quantum processor. Nature 606, 75–81 (2022).
information processing with atoms in optical micro-structures. Quantum Inf. Process. 110. Iverson, J. K. & Preskill, J. Coherence in logical quantum channels. New J. Phys. 22,
10, 907 (2011). 073066 (2020).
68. Levine, H. et al. Parallel implementation of high-fidelity multiqubit gates with neutral 111. Iyer, P. & Poulin, D. A small quantum computer is needed to optimize fault-tolerant
atoms. Phys. Rev. Lett. 123, 170503 (2019). protocols. Quantum Sci. Technol. 3, 030504 (2017).
69. Tan, D. B., Bluvstein, D., Lukin, M. D. & Cong, J. Compiling quantum circuits for dynamically 112. Kaufman, A. M. et al. Quantum thermalization through entanglement in an isolated
field-programmable neutral atoms array processors. Preprint at https://arxiv.org/abs/ many-body system. Science 353, 794–800 (2016).
2306.03487 (2023). 113. Brydges, T. et al. Probing Rényi entanglement entropy via randomized measurements.
70. Wimperis, S. Broadband, narrowband, and passband composite pulses for use in advanced Science 364, 260–263 (2019).
NMR experiments. J. Magn. Reson. A 109, 221–231 (1994).
71. Cummins, H. K., Llewellyn, G. & Jones, J. A. Tackling systematic errors in quantum logic
gates with composite rotations. Phys. Rev. A 67, 042308 (2003). Acknowledgements We thank A. Kubica for pointing us to the connection between our
72. Barnes, K. et al. Assembly and coherent control of a register of nuclear spin qubits. Nat. transversal gate set and IQP circuits, J. Campo, S. Haney, T. Wong, T. T. Wang, P. Stroganov
Commun. 13, 2779 (2022). and especially J. Amato-Grill for contributions in the development of the FPGA technology
73. Le Kien, F., Schneeweiss, P. & Rauschenbeutel, A. Dynamical polarizability of atoms in and fast CMOS readout. We gratefully acknowledge useful discussions with B. Braverman,
arbitrary light fields: General theory and application to cesium. Eur. Phys. J. D 67, 92 H. Briegel, S. Cantu, S. Choi, J. Cong, M. Devoret, H.-Y. Huang, A. Keesling, H. Levine,
(2013). A. Lukin, K. V. Kirk, N. Meister, H. Pichler, H. Poulsen, J. Ramette, J. Sinclair, D. Tan and all
74. Hutzler, N. R., Liu, L. R., Yu, Y. & Ni, K. K. Eliminating light shifts for single atom trapping. members of the Lukin group. We acknowledge financial support from the DARPA ONISQ
New J. Phys. 19, 023007 (2017). programme (grant number W911NF2010021), the Center for Ultracold Atoms (a NSF Physics
75. Shea, M. E., Baker, P. M., Joseph, J. A., Kim, J. & Gauthier, D. J. Submillisecond, Frontier Center), the National Science Foundation, the Army Research Office MURI (grant
nondestructive, time-resolved quantum-state readout of a single, trapped neutral number W911NF-20-1-0082), IARPA and the Army Research Office, under the Entangled
atom. Phys. Rev. A 102, 053101 (2020). Logical Qubits programme (Cooperative Agreement Number W911NF-23-2-0219) and QuEra
76. Gidney, C. Stim: a fast stabilizer circuit simulator. Quantum 5, 497 (2021). Computing. D.B. acknowledges support from the NSF Graduate Research Fellowship
77. Higgott, O., Bohdanowicz, T. C., Kubica, A., Flammia, S. T. & Campbell, E. T. Improved Program (grant DGE1745303) and the Fannie and John Hertz Foundation. S.J.E. acknowledges
decoding of circuit noise and fragile boundaries of tailored surface codes. Phys. Rev. X 13, support from the National Defense Science and Engineering Graduate (NDSEG) fellowship.
031007 (2023). T.M. acknowledges support from the Harvard Quantum Initiative Postdoctoral Fellowship in
78. Gottesman, D. Opportunities and challenges in fault-tolerant quantum computation. Science and Engineering. M.C. acknowledges support from the Department of Energy
Preprint at https://arxiv.org/abs/2210.15844 (2022). Computational Science Graduate Fellowship under award number DE-SC0020347. D.H.
79. Delfosse, N. & Paetznick, A. Spacetime codes of Clifford circuits. Preprint at https://arxiv. acknowledges support from the U.S. Department of Defense through a QuICS Hartree
org/abs/2304.05943 (2023). fellowship. J.P.B.A. acknowledges support from the Generation Q G2 Fellowship and the
80. Steane, A. M. Active stabilization, quantum computation, and quantum state synthesis. Ramsay Centre for Western Civilisation. N.M. acknowledges support by the Department of
Phys. Rev. Lett. 78, 2252 (1997). Energy Computational Science Graduate Fellowship under award number DE-SC0021110.
81. McEwen, M., Bacon, D. & Gidney, C. Relaxing hardware requirements for surface code I.C. acknowledges support from the Alfred Spector and Rhonda Kost Fellowship of the Hertz
circuits using time-dynamics. Quantum 7, 1172 (2023). Foundation, the Paul and Daisy Soros Fellowship and NDSEG. M.J.G. and D.H. acknowledge
82. Gurobi Optimization. Gurobi optimizer reference manual. Gurobi Optimization https:// support from NSF QLCI (award no. OMA-2120757). The commercial equipment used in this
www.gurobi.com/documentation/current/refman/index.html (2023). work does not reflect endorsement by the NIST. The views and conclusions contained in this
83. Landahl, A. J., Anderson, J. T. & Rice, P. R. Fault-tolerant quantum computing with color document are those of the authors and should not be interpreted as representing the official
codes. Preprint at https://arxiv.org/abs/1108.5738 (2011). policies, either expressed or implied, of IARPA, the Army Research Office, or the US
84. Monz, T. et al. 14-qubit entanglement: creation and coherence. Phys. Rev. Lett. 106, Government. The US Government is authorized to reproduce and distribute reprints for
130506 (2011). Government purposes notwithstanding any copyright notation herein.
85. Gottesman, D. Stabilizer Codes and Quantum Error Correction. Thesis, California Institute
of Technology (1997). Author contributions D.B., S.J.E., A.A.G., S.H.L., H.Z., T.M., S.E. and G.S. contributed to the
86. Knill, E. Quantum computing with realistically noisy devices. Nature 434, 39–44 (2005). building of the experimental setup, performed the measurements and analysed the data.
Article
M.C., M.K., D.H., J.P.B.A., N.M., I.C. and X.G. performed theoretical analysis. P.S.R. and T.K. Additional information
developed the FPGA electronics. All work was supervised by M.J.G., M.G., V.V. and M.D.L. All Supplementary information The online version contains supplementary material available at
authors contributed to the logical processor vision, discussed the results and contributed https://doi.org/10.1038/s41586-023-06927-3.
to the manuscript. Correspondence and requests for materials should be addressed to Mikhail D. Lukin.
Peer review information Nature thanks Benjamin Brown and the other, anonymous, reviewer(s)
Competing interests M.G., V.V. and M.D.L. are co-founders and shareholders and H.Z., P.S.R. for their contribution to the peer review of this work. Peer reviewer reports are available.
and T.K. are employees of QuEra Computing. Reprints and permissions information is available at http://www.nature.com/reprints.
Extended Data Fig. 1 | Neutral-atom quantum computer architecture. AWG’) are synchronized to <10 ns jitter. During Rydberg gates, the traps are
a, Experimental layout, featuring optical tools including static SLM and 2D briefly pulsed off by a TTL. The FPGA processes images from the camera in real
moving AOD traps, global and local Raman single-qubit laser beams, 420-nm time and, in this work, sends control signals to the Raman 2D AOD for local
and 1,013-nm Rydberg beams and imaging system for both global and local single-qubit control. d, Example array layout featuring entangling, storage
imaging. b, Level structure for 87Rb atoms, with the relevant atomic transitions and readout zones. Zones can be directly reprogrammed and repositioned for
used in this work. c, Control infrastructure used for programming quantum different applications, as well as specific tweezer site locations. Tweezer beams
circuits, featuring several AWGs. In particular, the moving and Raman 2D AODs and local Raman control are projected from out of plane. The entire objective
are each controlled by two waveforms (one for the x axis and one for the y axis). field of view is 400 μm in diameter and, consequently, we do not expect or
An additional AWG is used in first-in-first-out (FIFO) mode for rearrangement observe substantial tweezer deformation near the edges of our processor.
before the circuit begins and then the moving AOD control is switched to the During two-qubit Rydberg gates, we place atoms ≲2 μm apart within a gate
‘Moving AWG’. See ref. 30 for further SLM and pre-circuit rearrangement details, site and gate sites are separated such that atoms in different gate sites are no
ref. 8 for further Rydberg AWG details and Rydberg excitation details, refs. 7,63 closer than 10 μm during the gate. At our present n = 53 and two-photon Rabi
for further Raman laser and microwave control infrastructure details and ref. 7 frequency of 4.6 MHz, the blockade radius is roughly 4.3 μm, such that adjacent
for further moving AWG details. All AWGs (other than the ‘Rearrangement atoms are well within blockade and distant atoms are well outside blockade.
Article

Extended Data Fig. 2 | Single-qubit Raman addressing. a, 5S1/2 hyperfine level c, Calibration procedure used to homogenize the Rabi frequency over a
diagram illustrating the two possible implementations of local single-qubit 220 μm × 35 μm array. The position calibration is illustrated for 80 sites:
gates: resonant X(θ) (purple) and off-resonant Z(θ) (turquoise) rotations with approximate X(π/2) gates are locally performed and the horizontal/vertical
two-photon Rabi frequencies ΩRaman. In this work, we use the Z rotation scheme position of all tones is scanned in parallel such that a Gaussian fit returns the
and are blue-detuned by 2 MHz from the two-photon resonance. Owing to optimal alignment. After this, powers are iteratively calibrated until the fitted
Z
Clebsch–Gordan coefficients, Ω͠ Raman = − 3 ΩRaman
Z
. b, Schematic showing the scale factors for the individual RF tones converge to unity. d, Single-qubit
conversion of local Z(π/2) into local X(±π/2) gates, in which the pulses before randomized benchmarking of local Z(π/2) gates. The local gates are interleaved
(after) the central Y(π) have positive (negative) sign, while leaving non-addressed with random global single-qubit Clifford gates and the final operation Cf is
qubit states unchanged. The Gaussian-smoothed local pulses have duration chosen to return to the initial state. Each data point is the average of 100
2.5 μs for π/4 pulses and 5 μs for π/2 pulses and are performed on single rows random sets of Clifford gates and fitting an exponential decay to the return
at a time with a 3-μs gap between subsequent gates to allow the RF tones in the probability quantifies the fidelity F per local gate. Note that we apply all 51
AODs to be changed (including this, duration is 5–8 μs per row). In this way, global Clifford gates for each data point, such that errors from the global
arbitrary patterns of qubits, such as the example drawn, can be addressed. Clifford gates (as well as SPAM errors) do not contribute to the fitted value.
Extended Data Fig. 3 | Mid-circuit readout and feedforward. a, Single-shot feedforward cycle takes less than 1 ms and can be sped up in the future by
500-μs local image in the readout zone, in which the peak corresponds to optimizing local imaging and camera readout. e–g, Characterization of the
roughly 50 photons collected by the CMOS camera. b, Atomic transition and error probability of data qubits during local imaging. e, Data-qubit error
pulse sequence used for local imaging of ancilla qubits. The data-qubit probability (fraction of population depumped from F = 2 to F = 1) as a function
trap-light shift suppresses data qubit errors, as well as the large spatial of local imaging duration out to 20 ms to quantify the effect of the local
separation between entangling and readout zones. We avoid quickly losing the imaging beam on data-qubit coherence for very long illumination. f, Data-qubit
readout-zone atoms during local imaging by using a 5× higher trap depth and error probability after 20 ms of local imaging, as a function of detuning of the
we pulse the ancilla qubit traps and local imaging light to image directly on local imaging beam, showing suppression of error red-detuned or blue-detuned
resonance while avoiding negative effects of large trap-light shifts. c, Diagram from the data-qubit transition. g, Equivalently, increasing the trap depth of the
of components involved in mid-circuit readout and feedforward steps. Atom data qubits enables suppression of decoherence owing to the local imaging
detection and logical-state decoding occur using the FPGA, which then outputs beam. Because qubits in the readout zone are imaged while their traps are
a conditional TTL to gate local Raman pulses performed on logical qubits in pulsed off, any light shift of the data-qubit transition from the traps contributes
the entangling zone. d, Diagram of approximate timings for a mid-circuit directly to the relative detuning. h, For a long, 10.5-ms local beam illumination
feedforward cycle. First, the F = 2 population is pushed out (in 10 μs) and then with optimal local imaging parameters, we observe a 0.7(1)% increase in
the remaining F = 1 population is imaged locally for 500 μs. The 24 rows of data-qubit error during an XY8 dynamical decoupling sequence. This suggests
pixels covering the readout zone are read out to the FPGA in 200 μs, after a roughly 0.034(5)% error probability for the data qubits during the 500-μs
which processing is performed. Finally, a conditional TTL output based on the mid-circuit readout image used in this work.
decoded state gates on or off local Raman pulses. The whole readout and
Article

Extended Data Fig. 4 | Further surface-code data. a, Depiction of Bell-state stabilizer error probability owing to the error propagation in the transversal
circuit and d = 7 surface codes. b, Diagram showing the transversal CNOT and CNOT (reducing expectation values relative to if the transversal CNOT is not
physical error propagation rules. c, Covariance of the 48 measured stabilizers performed). f, Using the empirical error rates that correspond to data-theory
in both bases. The correlations near the diagonal corresponds to adjacent agreement for the measured stabilizers in e, our simulations for improvement
stabilizers within each block. Strong correlations are also observed with the in Bell-pair error, as a function of code distance, are in good agreement with
stabilizers of the other block owing to the error propagation in the transversal experiments. The empirical error rates used are consistent with the 99.3%
CNOT. d, Bell-pair infidelity upper bound (as opposed to estimated Bell-pair two-qubit gate fidelity, measured for this larger array, as well as the roughly 4%
error in Fig. 2d; see Methods), showing improvement with increasing code data-qubit decoherence error (integrated over the entire circuit and measured
distance. e, Probability of no detected error for each of the 96 measured by the Ramsey method). These dephasing error rates are dominated by a
stabilizers, showing agreement when compared with the theoretical values complex moving sequence as we prepare the two surface codes in a serial
from empirically chosen error rates (experiment average = 77%, theory fashion (see Supplementary Video) and would be much smaller for a repetitive
average = 82%). Note that X-basis logical 1 and Z-basis logical 2 have higher error-correction experiment.
Extended Data Fig. 5 | Surface-code preparation and decoding data. correlations between the two logical qubits, corresponding to the inter-
a, Surface-code stabilizers for the two independent d = 7 codes following state logical edges. As the decoder is optimized by tuning the inter-logical scaling
preparation. The entire movement circuit corresponding to the transversal factor, the performance for all three code distances improves, and the larger
CNOT is implemented and the transversal entangling-gate pulse is simply code distances improve faster when approaching the optimal decoding
turned off. The mean stabilizer probability of success across the 96 total configuration, as expected. These data are consistent with the decoder being
stabilizers is 83%. The high probability of stabilizer success of the two properly optimized for all three code distances, consistent with the fact that
independent codes in both the X and Z bases shows that topological surface our improvement with code size does not originate from suboptimal decoder
codes were prepared (and Extended Data Fig. 4 shows that they were preserved performance for low distance. Note that the y axis is log scale. c, Logical
during the transversal CNOT). We note that physical fidelities were slightly Bell-pair error when using (black) and not using (grey) the ancilla stabilizer
lower during this measurement because of calibration drift and, therefore, measurement values, as a function of the scaling of the inter-logical edges and
these results slightly underestimate performance relative to the data in hyperedges that connect the stabilizers of the two logical qubits. The ancilla
Fig. 2 and Extended Data Fig. 4. b, Logical Bell-pair error while optimizing measurements contribute to the correction procedure and contribute more
the decoder by (inversely) scaling the weights of the inter-logical edges and for smaller values of the inter-logical scaling, as they correspond to errors that
hyperedges that connect the stabilizers of the two logical qubits (higher values happen before the transversal CNOT. 0× inter-logical scaling corresponds to
correspond to lower pairing weights). More concretely, the probability p of conventional decoding within the two independent surface codes. For the
the error mechanism corresponding to the inter-logical edges/hyperedges is 1× inter-logical scaling plotted here, the d = 7 inter-logical scaling parameter is
scaled and the weights are calculated as log((1 − p)/p). Qualitatively, optimizing chosen slightly different from in Fig. 2d to have consistency across the three
this scaling value optimizes with respect to the probability that errors are code distances (which produces measured values within error bars).
before or after the transversal CNOT, as errors before the CNOT will lead to
Article

Extended Data Fig. 6 | [[8,3,2]] and hypercube encoding. a, State-preparation circuit. Initially, eight [[8,3,2]] code blocks are prepared in the entangling zone
circuit for the [[8,3,2]] code, in which two four-qubit GHZ states are and atoms for later state preparation of eight additional code blocks are loaded
simultaneously prepared and subsequently entangled. This initializes an in the storage zone. The code blocks in the entangling zone are then picked up
[[8,3,2]] code with logical states |−L1,+L2,−L3⟩. b, 4D hypercube circuit performed and interlaced with adjacent blocks to perform three transversal CNOT layers.
on 48 logical qubits (128 physical qubits). The circuit is drawn on the block The two groups of eight code blocks are then swapped and the same procedure
level, in which each block consists of three logical qubits and eight physical is repeated with the second group of code blocks. The first group of code blocks
qubits. The first in-block gate layer is performed with a global T†. The local gate is then moved back into the entangling zone and interleaved with the atoms of
patterns, and the corresponding logical gates they execute within each code the first group to perform a final parallel transversal CNOT. The layers of CNOT
block, are illustrated in the inset. c, Diagram illustrating the code-block gates connect the code blocks such that a 4D hypercube on 16 blocks of [[8,3,2]]
movements and use of the processor’s zoned architecture throughout the codes is constructed. See also Supplementary Video.
Extended Data Fig. 7 | Further [[8,3,2]] circuit sampling data. a, Overlap for the raw, uncorrected data, as the circuit we apply on the physical level is
of error-detected 12-qubit sampling data with the theoretical distribution not IQP. Without applying error detection, not all errors are logical errors
(same data as fully error-detected case in Fig. 5b). Progressive zoom-ins show and, therefore, the circuit differs from IQP behaviour and can lend itself to a
the agreement between theory and experiment, down to the level of 10 −4 different scaling. For systems of 3, 6 and 12 logical qubits, several systems are
probability per bitstring. This error-detected dataset is composed of 23,545 measured in parallel and their results are averaged. We note that, although our
shots (raw dataset is 138,626 shots). Note that we simultaneously measure on preparation of [[8,3,2]] code states makes these states on a cube, it does not
two groups of 12 logical qubits; plotted here is only one of the two 12-logical have CNOTs between two pairs of qubits in the first step and, therefore, does
groups with an XEB of 0.69(1), whereas in plots Fig. 5e,f and Extended Data not have the full gate connectivity of a cube. Instead, we can interpret these
Fig. 7b, we average the two logical groups, which gives a measured XEB of CNOTs as having been included but then compiled away as they commute with
0.616(7). b, Same data as Fig. 5f but with purity (orange), as measured by the state. We neglect this in plotting our physical-qubit connectivity, which is
two-copy measurement, also plotted. The measured XEB is slightly below the derived from entangling 3D cubes on a 4D hypercube connectivity, realizing a
measured purity, providing evidence that the XEB is a faithful fidelity proxy. 7D hypercube. c, 48-qubit XEB sliding-scale error-detection data. The point
We further note that, under error detection, the logical XEB for these IQP with full postselection on all stabilizers being perfect returned only eight
circuits should be a good fidelity proxy. Notably, the behaviour can be different samples, so we omit this point from the plot in the main text for clarity.
Article

Extended Data Fig. 8 | Theoretical exploration of hypercube IQP circuits. system (controls and targets of the final CNOT layer). This contraction scheme
a, Anticoncentration property of our circuits. The circuit is said to be reduces the memory requirements to half the system size, which enables
anticoncentrated if its output distribution is spread almost uniformly among bitstring amplitude evaluation for the 48-qubit experiment. This simulation
all outcomes, without the probability being concentrated on a subset of approach can be made much more expensive by applying further out-block
bitstrings. This property is crucial for many proofs of classical hardness20,100 operations within the two subsystems, forcing the blocking of the intra-partition
and, thus, it is desired for our sampling circuits to anticoncentrate. The plot tensors, which increases the memory and runtime requirements (Fig. 5d).
shows that the output distribution of random hypercube circuits (randomized c, To understand the effects of finite XEB on required classical simulation time,
in-block operations and randomized control/target in out-block CNOT layers) we explore whether our circuit families can be ‘spoofed’ with a cheaper,
anticoncentrates as the dimension of the hypercube is increased and the XEB approximate simulation that achieves moderately high XEB scores102, studied
(which captures the output collision probability) converges to the uniform IQP here for a 24-qubit system with full state-vector simulation. The spoofing
value of 2 (here using Clifford circuits; that is, circuits comprising random CZ algorithm works by independently sampling from the two halves of the
and Z only)20. This suggests that sampling from the ideal output distribution system (two groups of 12 qubits), effectively removing the final layer of CNOTs.
can be classically hard. In general, the hypercube IQP circuit ensemble converges This further reduces the simulation complexity, as each of the halves can, in
to the uniform IQP ensemble in total variation distance as the depth and principle, be independently simulated with the efficient approach from b. The
hypercube dimension are increased (M.K. et al., manuscript in preparation). The plot shows that the spoofed XEB for the 24-qubit non-Clifford circuit can be
specific circuit instances implemented in the experiment also anticoncentrate exponentially reduced by extending the circuit with further gate layers (similar
quickly with increasing hypercube dimension. b, A single layer of the hypercube to the approach used to decrease the performance of the efficient hypercube
circuit admits an efficient tensor-network contraction scheme, which allows us contraction), for a particular extension of our circuit. This result shows that
to evaluate the ideal and experimental XEB values. The final out-block CNOT future work can consider adding extra CNOT layers into these circuits to
layer is immediately followed by the measurement, which can be incorporated demonstrate quantum advantage (in the presence of finite experimental noise).
into a non-unitary tensor that is contracted between the two halves of the
Extended Data Fig. 9 | Further Bell-basis measurement results. a, Histogram c,d, Entanglement entropy when analysing the circuit as a physical Bell-basis
2
of tr(Pρ ) for all 46 Pauli strings P in the six-logical-qubit circuit, as a function measurement as opposed to a logical Bell-basis measurement. For logical
of stabilizer postselection threshold (that is, the number of correct stabilizers entanglement entropy calculations, we average over all possible subsystems
across the 6 × 2 logical qubits). Blue (red) indicate Pauli strings that are expected of that given subsystem size, which we find behaves very similarly to, for
2
to have tr(Pρ ) = 0.0625 (0). The separation between the histograms improves example, contiguous subsystems owing to the high-dimensional hypercube
as more postselection is applied. b, Signal to noise (purity divided by statistical connectivity. In the physical qubit entanglement entropy calculations, we
uncertainty of purity) as a function of sliding-scale error detection (converted randomly choose from the possible subsystems, as there are many. c, Six logical
into accepted fraction) for the 12-logical-qubit two-copy measurements, in (16 physical) qubits per copy. d, 12 logical (32 physical) qubits per copy. The
which subsystem size 1 indicates a single logical qubit in one copy and subsystem finite sampling imposes a noise floor for very high entanglement entropy
size 12 indicates all logical qubits. For subsystem size 1, the signal-to-noise values. e, Entanglement entropy measurements, as in Fig. 6b, but as a function
ratio gets worse as data are discarded, as the signal does not change (maximally of logical subsystem size. f, Logical circuits used for benchmarking magic. For
mixed) but the number of repetitions decreases. By contrast, for the global one CCZ, we include U1 and omit U0; for two CCZs, we include U0 and omit U1; for
purity, the signal to noise increases, as near-unity purities are faster to measure113. the three CCZs, we include both U0 and U1.