Simplified circuit-level decoding using Knill error correction

Ewan Murphy,1, 2, 3, 4, ∗ Subhayan Sahu,4 and Michael Vasmer1, 3, 4

1
Inria Paris, 48 rue Barrault, 75013 Paris, France
2
Quandela, 7 Rue Léonard de Vinci, 91300 Massy, France
3
Institute for Quantum Computing, University of Waterloo,
200 University Ave W, Waterloo, ON N2L 3G1, Canada
4
Perimeter Institute for Theoretical Physics,
31 Caroline St N, Waterloo, ON N2L 2Y5, Canada
Quantum error correction will likely be essential for building a large-scale quantum com-
puter, but it comes with significant requirements at the level of classical control software. In
arXiv:2603.05320v2 [quant-ph] 23 Apr 2026

particular, a quantum error-correcting code must be supplemented with a fast and accurate
classical decoding algorithm. Standard techniques for measuring the parity-check operators
of a quantum error-correcting code involve repeated measurements, which both increases
the amount of data that needs to be processed by the decoder, and changes the nature of
the decoding problem. Knill error correction is a technique that replaces repeated syndrome
measurements with a single round of measurements, but requires an auxiliary logical Bell
state. Here, we provide a theoretical and numerical investigation into Knill error correction
from the perspective of decoding. We give a self-contained description of the protocol, prove
its fault tolerance under locally decaying (circuit-level) noise, and numerically benchmark its
performance for quantum low-density parity-check codes. We show analytically and numer-
ically that the time-constrained decoding problem for Knill error correction can be solved
using the same decoder used for the simpler code-capacity noise model, illustrating that
Knill error correction may alleviate the stringent requirements on classical control required
for building a large-scale quantum computer.

I. INTRODUCTION whose input is the parity-check measurement out-

comes (the syndrome) and whose output is a recov-
Quantum error correction (QEC) is widely- ery operator. Error correction is successful if the
believed to be essential for building a large-scale combined effect of the noise and recovery operator
quantum computer capable of implementing quan- acts trivially on the encoded information. In the
tum algorithms with many layers of gates [1]. Nu- context of fault-tolerant quantum computing, it is
merous techniques from classical error correction crucial that the decoder is fast enough to keep up
have been ported over to the quantum setting, but with the rate of syndrome extraction. Otherwise a
a key difference is that measuring the parity-checks backlog of syndrome data will build up, leading to
of a quantum code is itself a noisy process [2]. This an exponential slowdown of the computation [16].
motivates the study of error models incorporating
measurement errors such circuit-level noise, which Repeating parity-check measurements changes
generalise error models that only consider errors on the nature of the decoding problem, and therefore
the data qubits, such as code-capacity noise. A va- a decoder that worked well for code-capacity noise
riety of techniques have been proposed to deal with might not work at all for more realistic error mod-
noisy measurements, including Shor [3], Steane [4], els such as circuit-level noise. This can be particu-
and Knill [5] error correction, as well as schemes us- larly problematic when a code has been optimised
ing flag qubits [6–8], repeated parity-check measure- to have good performance for code-capacity noise,
ments [9–12], and combinations of the above [13–15]. as this performance might not carry over to circuit-
A quantum error-correcting code must be sup- level noise models [17]. However, certain syndrome
plemented by a classical algorithm called a decoder, extraction techniques such as Knill and Steane error
correction have the useful property that the code-
capacity decoder can be used for phenomenological
∗
[Link]@[Link] and circuit-level noise. This property stems from
 2

the fact that these techniques use auxiliary logical results and directions for future research.
states, which are measured destructively.
In this work, we provide a theoretical and numer-
ical investigation into Knill error correction (Knill II. PRELIMINARIES
EC), which we argue is particularly well-suited to
the constraints of fast decoding. For Knill EC there A stabilizer code is the +1 eigenspace of an
are two relevant decoding problems, which we call abelian subgroup S of the n-qubit Pauli group that
online and offline decoding. Online decoding refers does not contain −I [21]. We summarise a code us-
to decoding the measured error syndrome, and must ing the notation [[n, k, d]], where k = n−rk S denotes
be fast in order to avoid the backlog problem. Of- the number of encoded logical qubits, and d denotes
fline decoding refers to preparing the auxiliary log- the minimum weight of a non-trivial logical Pauli
ical states, and is less constrained as state prepara- operator. A Calderbank-Shor-Steane (CSS) code is
tion can be parallelised. a stabilizer code whose stabilizer generators can be
We prove that the code-capacity decoder is suf- chosen to be either X-type or Z-type, where an X-
ficient for fault-tolerant error correction subject to type (Z-type) generator is a tensor product of X and
locally decaying noise (assuming a supply of auxil- I (Z and I) operators [22, 23]. Of particular inter-
iary logical Bell states), complimenting the previous est are the class of quantum low-density parity-check
analyses of Knill [5, 18, 19] and Gottesman [2]. We (LDPC) codes [24], which are defined by the prop-
perform numerical simulations of Knill EC, includ- erty that their stabilizer generators have low weight,
ing state preparation, for surface codes and high-rate and each qubit participates in a small number of gen-
codes. In the second case, we use the lifted product erators. Examples of quantum LDPC codes include
codes from Ref. [20], which were optimised to have the surface code [9, 25], as well as lifted product
good performance for belief propagation decoding codes [26].
against code-capacity noise (without post process- In a stabilizer code, decoding proceeds by mea-
ing). Our results show that an identical decoder can suring the eigenvalues of the stabilizer generators
be used for online decoding in Knill EC. Given that using a syndrome extraction circuit, which yields
belief propagation is well-suited to fast implemen- the error syndrome. A (classical) decoding algo-
tation on specialised hardware, this shows that the rithm takes the error syndrome as input and out-
online decoding can be extremely fast, making Knill puts a recovery operator; error correction is success-
EC a compelling option for architectures optimised ful if the product of the error and recovery oper-
for fast error correction and logical gates. ator is a stabilizer. The minimum-weight perfect
To perform our simulations, we develop a modu- matching (MWPM) algorithm is a well-known effi-
lar framework for benchmarking quantum error cor- cient decoder for the surface code [9], while belief
rection circuits, that may be of independent interest. propagation (BP) is a commonly used decoder for
Our framework enables the end-to-end simulation of general quantum LDPC codes (often supplemented
composed fault-tolerant protocols, where each one with post-processing such as ordered statistics de-
performs its own decoding. The tool automatically coding [26, 27]).
handles the interaction between the protocols and We consider two error models in this work: code-
the propagation of the Pauli frame. capacity noise and circuit-level noise. In code-
The remainder of this article is structured as capacity noise, errors occur only on the data qubits,
follows. In Sec. II we provide the necessary back- i.e., the syndrome extraction is noiseless. In circuit-
ground on quantum error correction and Knill EC. level noise, errors occur on every component of
In Sec. III, we prove that Knill EC is fault tolerant the error correction circuit: data qubits, auxiliary
under locally decaying noise, and that the online de- qubits, gates, state preparations, and measurements.
coding can be done using the code-capacity decoder. For details on the specific circuit-level error model
Next, in Sec. IV, we present our numerical results for we use, see Appendix C.
Knill EC for surface codes and lifted product codes. When decoding circuit-level noise for quantum
Finally, in Sec. V we discuss the implications of our LDPC codes, it is customary to repeat the syndrome
 3

H Calc. Calc. Calc. Calc.

|ψ⟩ D H Z1 X1 Zk Xk
H
···
D R ⊗ ⊗ ⊗ ⊗
⊕

A ⊕
Z1 X1 Zk Xk

1  ⊗k
√ |00⟩ + |11⟩ ⊕

2
B X1 Z1 ···
Xk Zk |ψ⟩

FIG. 1: Knill error correction. For a code block D encoded in a general [[n, k, d]] stabilizer code, two
auxiliary blocks A and B are initialized in k logical Bell pairs. The shaded region shows the circuit for
transversal Bell measurement on D ⊗ A. The measurement outcomes are passed to a decoder D, which
deduces a classical recovery R. The corrected logical operator measurements are then computed and used
to determine the logical Pauli correction applied to B, completing a logical teleportation from D to B.

extraction circuit O(d) times. Instead of using the measurement between the code blocks D and A is
stabilizer generators as parity-checks, one instead performed, using transversal CNOT gates between
uses the difference between the syndromes measured D and A, i.e. n CNOT gates controlled on di qubits
in consecutive rounds, which are often called de- with targets on ai qubits, followed by transversal X
tectors. A detector error model is then a Tanner and Z measurements on di and ai qubits, respec-
graph [28] with detectors as check nodes and error tively.
locations as variable nodes, where there an edge be- The transversal CNOT and Z- and X-basis mea-
tween a variable node and a check node if the error surements is equivalent to measuring Xdi ⊗ Xai and
at that location would flip the detector [29]. Zdi ⊗ Zai for all i ∈ {1, . . . , n}, where Xai is an X
Pauli operator on the ith physical qubit in block A
and similarly for the other operators. These 2n op-
A. Knill error correction erators form a basis for all Pauli operators of the
form P ⊗ P supported on D ⊗ A. We can there-
Knill error correction [5] (Knill EC) is a fault tol- fore find the eigenvalue for any of these Pauli oper-
erant error correction protocol that uses logical tele- ators from the 2n bits of measurement data of the
portation and a bifurcated decoding strategy. We transversal Bell measurement. For a stabilizer, the
first describe the protocol, followed by a proof of its eigenvalue of P ⊗ P will be the XOR of the syn-
fault tolerance. dromes in the code blocks D and A for that stabi-
Suppose that the logical information of k encoded lizer. We relay this syndrome information to a de-
qubits at some stage of a quantum computation is coder D, which outputs a recovery operator R. To
encoded in n physical qubits in a [[n, k, d]] stabilizer find the correct measurement outcomes, we flip the
code block D ≡ {di }ni=1 , where di refer to the ith value of the measurement outcomes of Xdi ⊗ Xai
physical qubit. At this stage, an error correction and Zdi ⊗ Zai if R anticommutes with Xdi and Zdi ,
gadget is applied to this state, which usually involves respectively. We can then deduce the measurement
repeated (faulty) syndrome measurement. In Knill result of the logical operators Xα ⊗ Xα and Zα ⊗ Zα
EC, we instead consider two auxiliary [[n, k, d]] code on D ⊗ A, where α = 1, · · · , k denote the encoded
blocks: A ≡ {ai }ni=1 , and B ≡ {bi }ni=1 , initialized in logical qubits. Based on the measurement result, we
k logical Bell pair states [30]. Next, a logical Bell perform logical corrections on block B, to teleport
 4

the initially encoded information from D to B. The We postpone the proof of this lemma to the ap-
protocol is depicted in Fig. 1. pendix Sec. A 2. Using this lemma, we find that the
Importantly, the decoding necessary for this step final state (before destructive measurement) is effec-
of the Knill EC has the same complexity as code- tively erroneous with LD rate peff (including mea-
capacity decoding, and thus can be achieved at surement errors modelled by bit or phase flips prior
faster rate than spacetime decoding necessary for re- to measuring). Hereafter, we can assume perfect syn-
peated syndrome measurements. This constitutes drome extraction, since the qubits are destructively
the online decoder. In the next section, we will measured, followed by a faultless classical computa-
prove that Knill EC is fault tolerant for low enough tion to extract the syndromes.
error in the auxiliary state preparation. However, We now posit that there exist [[n, k, d]] code fami-
the preparation of these logical states on A ⊗ B with lies and (code-capacity) decoders based on ideal syn-
low enough error requires a separate offline error cor- drome measurements, such that for a locally decay-
rection protocol, which may involve higher decoding ing noise channel on the input with error rate τ be-
complexity than the online decoder. low some threshold τth > 0, the logical failure rate
satisfies:
 d
cτ
III. FAULT TOLERANCE WITH KNILL Pr(logical failure) ≤ f (τ ) ≡ g(τ, τth ) , (2)
τth
ERROR CORRECTION
for some constant c < 1 and a n-independent func-
We will now analyze logical error propagation un- tion g(τ, τth ).
der the transversal Bell measurement, which is an es- We can now prove the fault tolerance of Knill EC
sential ingredient of Knill EC. We will assume each under locally decaying noise. To simplify the nota-
component of the circuit is subjected to locally de- tion, we assume that all the associated LD noise
caying noise, and study the propagation of logical rates in the faulty Knill gadget are p. Following
error through the protocol. Locally decaying (LD) Gottesman [2], we define c(P, Q) be a function that
noise means that large, spread-out error events are takes as input two Pauli operators P and Q, and
exponentially unlikely: the chance that errors simul- outputs 1 if they anti-commute and 0 if they com-
taneously hit any specified set of t qubits (or t fault mute.
locations) drops like τ t , with the noise rate being Lemma 2. For any Pauli operators P, Q, R, we
τ . Equivalently, most of the probability mass is con- have c(P, QR) = c(P, Q) + c(P, R).
centrated on small supports, while long correlated
Proof. We have
patterns are strongly suppressed. This model still al-
lows correlations, but it rules out “macroscopic” cor- P QR = (−1)
c(P,QR)
QRP, (3)
related failures except with very small probability. P QR = (−1)c(P,Q) QP R = (−1)c(P,Q)+c(P,R) QRP.
The formal definitions are provided in the appendix (4)
Sec. A 1. We first present the following lemma,
Comparing the exponents, we have the result.
Lemma 1. Error propagation with transversal Suppose that the error before the logical Bell
Bell measurement circuit. Assume that the measurement is ED ⊗ EA . Let FD denote the Z
blocks D and A are subjected to LD noise with rate and Y part of the CNOT error on block D along
p1 and p2 respectively. The transversal CNOT gate with the measurement error, and let FA denote the
F ⊆ {1, . . . , n} has LD error with rate p3 , and the X and Y part of the CNOT error on block A along
single-qubit measurements is LD with rate p4 . Then, with the measurement error. Then, for any Pauli
the effective LD noise rate on the two blocks D and operator P ⊗ P , the measured value is
A propagated through the faulty Bell measurement
m̃P ⊗P = mP ⊗P + c(P, ED ) + c(P, EA )
circuit is,
+ c(P, FD ) + c(P, FA ), (5)
√
peff = p1 + p 2 + p 3 + p 4 . (1) = mP ⊗P + c(P, ED EA FD FA ),
 5

f4

|ψ⟩ H H
f1 f3
|0⟩ ⊕ ⊕ → |0⟩ P1 ⊕ ⊕

f2 P3
|+⟩ X Z |ψ⟩ |+⟩ P2 P4 = X , Z

FIG. 2: Clifford fault tolerant protocols can be defined by: a physical circuit, a (classical) function that
processes the measurement data sampled from this circuit, and a set of Pauli corrections that are applied
based on the function.

where mP ⊗P is the measurement outcome in the ab- Therefore, the total logical error probability after
sence of error and we use Lemma 2. In particular, M rounds of faulty Knill EC is q[M ] ≈ M f (peff ).
for a stabilizer generator S ⊗ S, the measurement Choosing a large enough distance for the code can
outcome is m̃S⊗S = c(S, ED EA FD FA ), and so we re- suppress f (·), and thus q[M ] below a target error rate
cover the syndrome of the error ED EA FD FA . If the for an M ∼ poly(n)-depth logical computation.
√
peff = 3p+p ≤ τth , then with probability 1−f (peff )
the recovery operator returned by the code-capacity
decoder D satisfies R = ED EA FD FA S for some sta- IV. NUMERICAL RESULTS
bilizer S. Then, for any logical Pauli operator L ⊗ L,
we can compute the corrected measurement outcome A. Simulating the composition of fault tolerant
as protocols

m̃L⊗L + c(L, R), Fault tolerant quantum computing requires the

(6) replacement of logical subcircuits with fault toler-
= mL⊗L + c(L, ED EA FD FA ) + c(L, R),
ant protocols. Numerical simulations have provided
= mL⊗L + c(L, S) = mL⊗L . important information on the performance of these
protocols. However, understanding how these proto-
We can therefore recover the correct measurement cols perform when composed into larger logical cir-
outcomes for all the logical operators, and thus reli- cuits is still an under-explored area.
ably compute the logical correction to apply to block
Motivated by simulating logical Bell teleporta-
B.
tion, in the form of Knill EC, we developed a
After one round of Knill EC, the new data block
tool that allows the end-to-end simulation of com-
has physical error rate equal to the error in the aux-
posed fault-tolerant protocols, building on Stim [29].
iliary blocks in the previous step. There is also a
The tool is designed to be modular, with the fault-
rate of logical failure carried over from the previous
tolerant protocols defined independently of each
rounds. Let p′[n] be the probability that the (n+1)th
other and their interaction handled automatically.
round does not eliminate the accumulated logical er-
The fault-tolerant protocols we consider are Clifford
ror from the previous n rounds. The logical error
circuits with an associated classical function that de-
rate after n + 1 rounds is, thus,
termines what Pauli corrections to apply based on
q[n+1] = q[n] × p′[n] + (1 − q[n] ) × f (peff ) the circuits measurements. An example is shown in
Figure 2.
≤ q[n] + f (peff ) (7) Given the modular nature of our tool, each pro-
 6

tocol performs its own decoding, which is in con- d rounds

trast with other work in the literature that decodes

the whole logical Clifford circuit as one module [31– |0⟩ ⊕
34]. We note that our approach is also different from |0⟩ ⊕
.. SX SZ SX SZ SX SZ
overlapping- or sliding-window decoding [9, 35–37], . ··· ..
.
where the decoding problem for the Clifford circuit |0⟩ ⊕

separated into smaller decoding problems that have

overlapping detectors [38–42]. |+⟩
Although, the original motivation to build this |+⟩
tools was to simulate Knill EC we believe it will be of .. SX SZ SX SZ ···
SX SZ ..
. .
interest to the wider QEC community as it supports |+⟩
the simulation of many types of logical circuits. A
detailed explanation of how this tool works is given
 ⊗k
= √1 |00⟩ + |11⟩
 
in Appendix B.
2
B. Simulations FIG. 3: Circuit for the fault tolerant preparation of
logical Bell states used in the simulations. The
We performed simulations to verify that a code- syndrome data from the d rounds of X and Z
capacity decoder can be used for the online decoder stabilizer measurement are decoded separately
part of Knill EC. The Knill EC simulations were per- using the offline decoder.
formed using the modular simulation tool mentioned
in the previous section. This allowed us to have sep-
arate the online and offline decoding for the Bell two larger codes having girth 8. We performed code-
measurement and state preparation, respectively. capacity simulations using BP and verified that we
To performed these simulations we had to pick a were able to observe a threshold. Then, we per-
method for state preparation, for this work we chose formed Knill EC simulations using the same decoder
to use repeated syndrome measurement. We would for the online decoding and BP+OSD for the of-
  ⊗k fline decoding. These simulations confirmed that
like to prepare the state √12 |00⟩ + |11⟩ across we could achieve threshold behaviour under circuit
the 2k logical qubits of two [[n, k, d]] stabilizer code level noise using a code-capacity decoder for online
blocks. We first prepare all the physical qubits of decoding, provided we had access to fault tolerantly
the first code block in the |0⟩ state, and all the phys- prepared Bell states. To provide a comparison with
ical qubits of the second code block in the |+⟩ state. existing methods in the literature we performed over-
Then perform d rounds of stabilizer measurement lapping window decoding. We simulated 32d rounds
and decoding, with the X and Z syndromes for each of stabilizer measurement with the syndrome par-
state being decoded separately using the offline de- titioned into overlapping windows of size 2d, with
⊗k
coder. This will prepare the logical states |0⟩ and commit size of d. The detector error models for
|+⟩⊗k . Performing a transversal CNOT controlled these windows were decoded using BP to compare
on the second code block will create the logical Bell to the online decoding of Knill EC. We were not
state. A diagram of this circuit is shown in Figure 3. able to achieve threshold behaviour when perform-
This method of state preparation is possible for CSS ing overlapping window decoding with BP for the
codes. lifted product codes. This agrees with our under-
We performed numerical simulations for a set standing that the large girth of these lifted prod-
of high rate lifted product (LP) codes from [20]; uct codes allowed them to be decoded in the code-
the results are shown in Figure 4(a), (c), and (e). capacity setting, however, this property is not main-
The LP codes have parameters: [[175, 19, ≤ 10]], tained when moving to the detector error model for
[[225, 21, ≤ 12]], [[425, 29, ≤ 18]] and [[475, 31, ≤ 20]], repeated syndrome measurement.
with the two smaller codes having girth 6 and the In addition to the lifted product codes we per-
 7

100

10−1
pL

pL
10−1
[[175,19,≤ 10]] 10−2 [[81,1,9]]
[[225,21,≤ 12]] [[121,1,11]]
[[425,29,≤ 18]] [[225,1,15]]
[[475,31,≤ 20]] [[441,1,21]]
0.06 0.08 0.10 0.12 0.06 0.08 0.10 0.12 0.14
Physical Error Rate Physical Error Rate
(a) LP codes, code-capacity noise (b) Surface codes, code-capacity noise

100

10−1
pL

pL
10−1
[[175,19,≤ 10]] [[81,1,9]]
[[225,21,≤ 12]] [[121,1,11]]
[[425,29,≤ 18]] 10−2 [[225,1,15]]
[[475,31,≤ 20]] [[441,1,21]]
0.003 0.004 0.005 0.006 0.007 0.008 0.004 0.005 0.006 0.007 0.008 0.009 0.010
Physical Error Rate Physical Error Rate
(c) LP codes, circuit noise with Knill EC (8 rounds) (d) Surface codes, circuit noise with Knill EC (8 rounds)

10−1

10−2
pL

pL

10−1
[[175,19,≤ 10]] [[81,1,9]]
[[225,21,≤ 12]] 10−3 [[121,1,11]]
[[425,29,≤ 18]] [[225,1,15]]
[[475,31,≤ 20]] [[441,1,21]]
10−4
0.001 0.002 0.003 0.004 0.005 0.004 0.006 0.008 0.010
Physical Error Rate Physical Error Rate
(e) LP codes, circuit noise with repeated syndrome (f) Surface codes, circuit noise with repeated syndrome
measurement (32d rounds) measurement (32d rounds)

Code-capacity Knill error correction Repeated noisy syndrome

Offline Online extraction∗
Lifted product codes BP BP+OSD BP BP†
Surface codes MWPM MWPM MWPM MWPM
(g) Decoding strategies for numerics. ∗ Overlapping window decoding. † No threshold.

FIG. 4: Numerical simulations comparing the performance of the same decoder for different noise models
and error correction gadgets. (a,c,e) Lifted product (LP) codes decoded using BP. We observe results
consistent with a non-zero threshold in (a) and (c). (b,d,f) Surface codes decoded using MWPM. We
observe results consistent with a non-zero threshold in all cases. Error bars show 95% confidence intervals
calculated using the Agresti-Coull method [43]. (g) Summary of decoders used in the simulations.
 8

formed the same three numerical experiments for propagation, it is likely that the performance could
the rotated surface code; see Figure 4(b), (d), and be improved by using the circuit structure to opti-
(f) for the results. We performed the simulations on mise the error priors given to the decoder.
surface codes with parameters [[81, 1, 9]], [[121, 1, 11]], Knill EC is compatible with all stabilizer codes,
[[225, 1, 15]] and [[441, 1, 21]]. We used MWPM to per- but for CSS codes there is a compressed version that
form the online and offline decoding. All three nu- only uses auxiliary logical |0⟩ and |+⟩ states [14, 46,
merical experiments were able to achieve threshold 47]; see Appendix D. This version is also compatible
behaviour. Although MWPM is used for both the with fast decoding and may therefore be preferable
online part of Knill EC and overlapping window de- for CSS codes. We argue that compressed Knill EC
coding, the size of the graph they are solving the is likely to be superior to Steane error correction
problem on is different, O(d2 ) vertices compared to for two reasons. Firstly, it inherits the robustness
O(d3 ). This suggests that although overlapping win- to leakage, loss, and coherent errors [47–49] of Knill
dow decoding can achieve threshold behaviour with EC. Secondly, in Steane error correction the recovery
the same algorithm as online decoding in Knill EC, operator is applied to the data block, and therefore
there still may be advantages in terms of real time errors in the auxiliary state can lead to spurious op-
decoding and control system design to use Knill EC. erators being applied to the data block. This can be
For MWPM we used the PyMatching [44] im- dealt with by considering an approach reminiscent
plementation. For BP and BP+OSD [27] we used of overlapping window decoding [50] but at the cost
the implementations of these algorithms from the of making the decoding problem more complex.
python library ldpc [45]. The parameters for BP Knill EC can be modified to implement logical
were taken from the same paper as the LP codes [20]. gates through gate teleportation [5, 51]. The only
Minimum sum was used as the belief propagation real change is that different auxiliary logical states
method with a scaling factor of 0.75, a serial sched- need to be prepared; the online decoding remains
ule and maximum number of iterations set at 100. the same. Therefore, a fault-tolerant quantum com-
For BP+OSD the parameters were the same and puting architecture based on Knill EC could achieve
we used the “osd_0” ordered statistics decoding high logical clock speeds, as long as the pace of aux-
method. Table (g) in Figure 4 summarises which iliary logical state generation can keep up. This
decoder was used for each of the numerical experi- makes architectures based on Knill EC an attractive
ments. proposition for hardware platforms with slower phys-
ical operations, such as neutral atom and trapped
ion qubits [52–55]. It is worth contrasting the Knill
V. DISCUSSION approach with the paradigm of algorithmic fault tol-
erance [33, 34, 42], where syndrome measurements
In this work, we demonstrated that Knill EC is are not repeated but fault tolerance is nonetheless
compatible with fast decoding, and moreover that maintained. This approach is more qubit efficient
the online decoding during the protocol can be per- than the Knill approach, but at the cost of a sig-
formed using a code-capacity decoder. We illus- nificant increase in the complexity of the decoding
trated this point by using belief propagation (with problem. We leave a detailed comparison of these
no post-processing) as the online decoder for a fam- two approaches to future work.
ily of lifted product codes. This decoder is widely We benchmarked a naive approach to auxiliary
used in practice for classical codes, and can be effi- logical state preparation, where logical Bell states
ciently implemented using specialised hardware such are prepared using d rounds of error correction. Sup-
as FPGAs and ASICs. This fact combined with the pose that we want to do Knill EC for m copies of
small number of variables in the decoding problem some [[n, k, d]] code. Then, to keep pace with the
implies that the Knill approach has some of the low- online decoding, we would need approximately 2md
est classical control requirements of any quantum code blocks to prepare the auxiliary logical states
error correction scheme. For decoders compatible (for CSS codes we would use compressed Knill and
with soft information such as matching and belief save a factor of 2), a significant qubit overhead. Op-
 9

timised state preparation circuits exist for certain plotting_lib library for producing the plots in
codes, see e.g. [56–65], including code families de- Fig. 4. We acknowledge the use of TikZit to draw
signed specifically with this in mind [66]. In future the circuit diagrams used in this paper. The au-
work, we plan to investigate and devise more effi- thors are grateful to the CLEPS infrastructure from
cient techniques for logical state preparation, and to the Inria of Paris for providing resources and sup-
propose an optimised architecture for fault-tolerant port. This research was enabled in part by sup-
quantum computation for hardware platforms with port provided by Simon Fraser University (https:
access to transversal CNOT gates. //[Link]/) and the Digital Research Alliance
of Canada ([Link] EM grate-
fully acknowledges the financial support of NTT Re-
CODE AVAILABILITY
search while working in Waterloo. Research at IQC
and the Perimeter Institute is supported in part by
The modular simulation tool we built to perform the Government of Canada through the Department
the simulations in this paper is call Hex and is avail- of Innovation, Science and Economic Development;
able on GitHub [67]. and by the Province of Ontario through the Ministry
of Colleges, Universities, Research Excellence and
ACKNOWLEDGEMENTS Security. This work was co-funded by CIFRE grant
n°2025/0667. MV was supported in part by Plan
The authors thank Boris Bourdoncle, Anthony France 2030 through the project ANR-22-PETQ-
Leverrier, and Evan Peters for valuable discus- 0006.
sions. We acknowledge the use of Timo Hillmann’s

[1] E. T. Campbell, B. M. Terhal, and C. Vuillot, Roads [10] A. G. Fowler, A. M. Stephens, and P. Groszkowski,
towards fault-tolerant universal quantum computa- High-threshold universal quantum computation on
tion, Nature 549, 172 (2017). the surface code, Physical Review A 80, 052312
[2] D. Gottesman, Surviving as Quantum Computer (2009).
in a Classical World (2024), [Link] [11] A. A. Kovalev and L. P. Pryadko, Fault tolerance of
edu/class/spring2024/cmsc858G/. quantum low-density parity check codes with sublin-
[3] P. Shor, Fault-tolerant quantum computation, in ear distance scaling, Physical Review A 87, 020304
Proceedings of 37th Conference on Foundations (2013).
of Computer Science (IEEE Comput. Soc. Press, [12] D. Gottesman, Fault-tolerant quantum computa-
Burlington, VT, USA, 1996) pp. 56–65. tion with constant overhead, Quantum Information
[4] A. M. Steane, Active stabilization, quantum com- & Computation 14, 1338 (2014), arXiv:1310.2984.
putation, and quantum state synthesis, Phys. Rev. [13] C.-Y. Lai, Y.-C. Zheng, and T. A. Brun, Fault-
Lett. 78, 2252 (1997). tolerant preparation of stabilizer states for quantum
[5] E. Knill, Quantum computing with realistically Calderbank-Shor-Steane codes by classical error-
noisy devices, Nature 434, 39 (2005). correcting codes, Physical Review A 95, 032339
[6] R. Chao and B. W. Reichardt, Quantum Error Cor- (2017).
rection with Only Two Extra Qubits, Physical Re- [14] S. Huang and K. R. Brown, Between Shor and
view Letters 121, 050502 (2018). Steane: A Unifying Construction for Measuring
[7] C. Chamberland and M. E. Beverland, Flag fault- Error Syndromes, Physical Review Letters 127,
tolerant error correction with arbitrary distance 090505 (2021).
codes, Quantum 2, 53 (2018), arXiv:1708.02246. [15] S. Huang and K. R. Brown, Constructions for mea-
[8] R. Chao and B. W. Reichardt, Flag Fault-Tolerant suring error syndromes in Calderbank-Shor-Steane
Error Correction for any Stabilizer Code, PRX codes between Shor and Steane methods, Physical
Quantum 1, 010302 (2020). Review A 104, 022429 (2021).
[9] E. Dennis, A. Kitaev, A. Landahl, and J. Preskill, [16] B. M. Terhal, Quantum Error Correction for Quan-
Topological quantum memory, Journal of Mathe- tum Memories, Reviews of Modern Physics 87, 307
matical Physics 43, 4452 (2002). (2015), arXiv:1302.3428.
 10

[17] M. Pacenti, D. Chytas, and B. Vasic, Construction [34] M. Cain, D. Bluvstein, C. Zhao, S. Gu, N. Maskara,
and Decoding of Quantum Margulis Codes (2025), M. Kalinowski, A. A. Geim, A. Kubica, M. D. Lukin,
arXiv:2503.03936 [quant-ph]. and H. Zhou, Fast correlated decoding of transversal
[18] E. Knill, Fault-Tolerant Postselected Quantum logical algorithms (2025), arXiv:2505.13587 [quant-
Computation: Schemes (2004), arXiv:quant- ph].
ph/0402171. [35] L. Skoric, D. E. Browne, K. M. Barnes, N. I. Gille-
[19] E. Knill, Fault-Tolerant Postselected Quan- spie, and E. T. Campbell, Parallel window decod-
tum Computation: Threshold Analysis (2004), ing enables scalable fault tolerant quantum compu-
arXiv:quant-ph/0404104. tation, Nature Communications 14, 7040 (2023).
[20] N. Raveendran, N. Rengaswamy, F. Rozpędek, [36] L. Berent, T. Hillmann, J. Eisert, R. Wille, and
A. Raina, L. Jiang, and B. Vasić, Finite Rate J. Roffe, Analog Information Decoding of Bosonic
QLDPC-GKP Coding Scheme that Surpasses the Quantum Low-Density Parity-Check Codes, PRX
CSS Hamming Bound, Quantum 6, 767 (2022). Quantum 5, 020349 (2024).
[21] D. Gottesman, Stabilizer Codes and Quantum Error [37] T. R. Scruby, T. Hillmann, and J. Roffe,
Correction (1997), arXiv:quant-ph/9705052. High-threshold, low-overhead and single-shot de-
[22] A. R. Calderbank and P. W. Shor, Good quantum codable fault-tolerant quantum memory (2024),
error-correcting codes exist, Physical Review A 54, arXiv:2406.14445 [quant-ph].
1098 (1996). [38] H. Bombín, C. Dawson, Y.-H. Liu, N. Nickerson,
[23] A. M. Steane, Error Correcting Codes in Quantum F. Pastawski, and S. Roberts, Modular decoding:
Theory, Physical Review Letters 77, 793 (1996). Parallelizable real-time decoding for quantum com-
[24] N. P. Breuckmann and J. N. Eberhardt, Quantum puters (2023), arXiv:2303.04846 [quant-ph].
Low-Density Parity-Check Codes, PRX Quantum 2, [39] A. J. Malcolm, A. N. Glaudell, P. Fuentes, D. Chan-
040101 (2021). dra, A. Schotte, C. DeLisle, R. Haenel, A. Ebrahimi,
[25] S. B. Bravyi and A. Y. Kitaev, Quantum codes J. Roffe, A. O. Quintavalle, S. J. Beale, N. R.
on a lattice with boundary (1998), arXiv:quant- Lee-Hone, and S. Simmons, Computing Efficiently
ph/9811052. in QLDPC Codes (2025), arXiv:2502.07150 [quant-
[26] P. Panteleev and G. Kalachev, Degenerate Quan- ph].
tum LDPC Codes With Good Finite Length Perfor- [40] K. Sahay, Y. Lin, S. Huang, K. R. Brown, and
mance, Quantum 5, 585 (2021). S. Puri, Error Correction of Transversal cnot
[27] J. Roffe, D. R. White, S. Burton, and E. Campbell, Gates for Scalable Surface-Code Computation, PRX
Decoding across the quantum low-density parity- Quantum 6, 020326 (2025).
check code landscape, Phys. Rev. Res. 2, 043423 [41] J. Zhang, Z.-Y. Chen, J.-N. Li, T.-H. Wei,
(2020). H.-Y. Liu, X.-N. Zhuang, Q.-S. Li, Y.-C. Wu,
[28] D. MacKay, Information Theory, Inference, and and G.-P. Guo, Scalable Constant-Time Logical
Learning Algorithms (2003). Gates for Large-Scale Quantum Computation Us-
[29] C. Gidney, Stim: a fast stabilizer circuit simulator, ing Window-Based Correlated Decoding (2025),
Quantum 5, 497 (2021). arXiv:2410.16963 [quant-ph].
[30] In principle, the two auxiliary blocks can be in dif- [42] M. Serra-Peralta, M. H. Shaw, and B. M. Terhal,
ferent stabilizer codes, however here we specialize to Decoding across Transversal Clifford Gates in the
the same code. Surface Code, PRX Quantum 7, 010335 (2026).
[31] M. Cain, C. Zhao, H. Zhou, N. Meister, J. P. B. [43] L. D. Brown, T. T. Cai, and A. DasGupta, Inter-
Ataides, A. Jaffe, D. Bluvstein, and M. D. Lukin, val Estimation for a Binomial Proportion, Statisti-
Correlated Decoding of Logical Algorithms with cal Science 16, 10.1214/ss/1009213286 (2001).
Transversal Gates, Physical Review Letters 133, [44] O. Higgott and C. Gidney, Sparse Blossom: Cor-
240602 (2024). recting a million errors per core second with
[32] K. H. Wan, M. Webber, A. G. Fowler, and W. K. minimum-weight matching, Quantum 9, 1600
Hensinger, An iterative transversal CNOT decoder (2025), arXiv:2303.15933 [quant-ph].
(2024), arXiv:2407.20976 [quant-ph]. [45] J. Roffe, LDPC: Python tools for low density parity
[33] H. Zhou, C. Zhao, M. Cain, D. Bluvstein, check codes (2022).
N. Maskara, C. Duckering, H.-Y. Hu, S.-T. Wang, [46] A. Paetznick, M. P. da Silva, C. Ryan-Anderson,
A. Kubica, and M. D. Lukin, Low-overhead transver- J. M. Bello-Rivas, J. P. C. III, A. Chernogu-
sal fault tolerance for universal quantum computa- zov, J. M. Dreiling, C. Foltz, F. Frachon, J. P.
tion, Nature 646, 303 (2025). Gaebler, T. M. Gatterman, L. Grans-Samuelsson,
D. Gresh, D. Hayes, N. Hewitt, C. Holliman, C. V.
 11

Horst, J. Johansen, D. Lucchetti, Y. Matsuoka, wal, J. Simon, A. Smull, M. Sorensen, D. T. Stack,

M. Mills, S. A. Moses, B. Neyenhuis, A. Paz, J. Pino, M. Stone, L. Taneja, R. J. M. van de Veerdonk,
P. Siegfried, A. Sundaram, D. Tom, S. J. Wernli, Z. Vendeiro, R. T. Weverka, K. White, T.-Y. Wu,
M. Zanner, R. P. Stutz, and K. M. Svore, Demon- X. Xie, E. Zalys-Geller, X. Zhang, J. King, B. J.
stration of logical qubits and repeated error cor- Bloom, and M. A. Norcia, Repeated ancilla reuse
rection with better-than-physical error rates (2024), for logical computation on a neutral atom quantum
arXiv:2404.02280 [quant-ph]. computer (2025), arXiv:2506.09936 [quant-ph].
[47] G. Baranes, M. Cain, J. P. B. Ataides, D. Blu- [54] J.-S. Chen, E. Nielsen, M. Ebert, V. Inlek,
vstein, J. Sinclair, V. Vuletić, H. Zhou, and M. D. K. Wright, V. Chaplin, A. Maksymov, E. Páez,
Lukin, Leveraging Qubit Loss Detection in Fault- A. Poudel, P. Maunz, and J. Gamble, Benchmark-
Tolerant Quantum Algorithms, Physical Review X ing a trapped-ion quantum computer with 30 qubits,
16, 011002 (2026). Quantum 8, 1516 (2024).
[48] C. Ryan-Anderson, N. C. Brown, C. H. Bald- [55] A. Ransford, M. S. Allman, J. Arkinstall, J. P. C.
win, J. M. Dreiling, C. Foltz, J. P. Gaebler, III, S. F. Cooper, R. D. Delaney, J. M. Dreil-
T. M. Gatterman, N. Hewitt, C. Holliman, C. V. ing, B. Estey, C. Figgatt, A. Hall, A. A. Husain,
Horst, J. Johansen, D. Lucchetti, T. Mengle, A. Isanaka, C. J. Kennedy, N. Kotibhaskar, I. S.
M. Matheny, Y. Matsuoka, K. Mayer, M. Mills, Madjarov, K. Mayer, A. R. Milne, A. J. Park,
S. A. Moses, B. Neyenhuis, J. Pino, P. Siegfried, A. P. Reed, R. Ancona, M. P. Andersen, P. Andres-
R. P. Stutz, J. Walker, and D. Hayes, High- Martinez, W. Angenent, L. Argueta, B. Arkin,
fidelity and Fault-tolerant Teleportation of a Log- L. Ascarrunz, W. Baker, C. Barnes, J. Bartolotta,
ical Qubit using Transversal Gates and Lattice J. Berg, R. Besand, B. Bjork, M. Blain, P. Blan-
Surgery on a Trapped-ion Quantum Computer chard, R. Blume-Kohout, M. Bohn, A. Borgna,
(2024), arXiv:2404.16728 [quant-ph]. D. Y. Botamanenko, R. Boutelle, N. Brown,
[49] K. Chang, Q. Su, and S. Puri, Taming coherent G. T. Buckingham, N. Q. Burdick, W. C. Bur-
noise with teleportation (2025), arXiv:2508.04947 ton, V. Carey, C. J. Carron, J. Chambers, J. Chil-
[quant-ph]. dren, V. E. Colussi, S. Crepinsek, A. Cureton,
[50] A. Gong and J. M. Renes, Improved Logical Error J. Davies, D. Davis, M. DeCross, D. Deen, C. De-
Rate via List Decoding of Quantum Polar Codes, laney, D. DelVento, B. J. DeSalvo, J. Dominy,
in 2024 IEEE International Symposium on Infor- R. Duncan, V. Eccles, A. Edgington, N. Erickson,
mation Theory (ISIT) (2024) pp. 2496–2501. S. Erickson, C. T. Ertsgaard, B. Evans, T. Evans,
[51] D. Gottesman and I. L. Chuang, Demonstrating the M. I. Fabrikant, A. Fischer, C. Foltz, M. Foss-
viability of universal quantum computation using Feig, D. Francois, B. Freyberg, C. Gao, R. Garay,
teleportation and single-qubit operations, Nature J. Garvin, D. M. Gaudiosi, C. N. Gilbreth, J. Giles,
402, 390 (1999). E. Glynn, J. Graves, A. Hansen, D. Hayes, L. Heide-
[52] D. Bluvstein, S. J. Evered, A. A. Geim, S. H. mann, B. Higashi, T. Hilbun, J. Hines, A. Hlavaty,
Li, H. Zhou, T. Manovitz, S. Ebadi, M. Cain, K. Hoffman, I. M. Hoffman, C. Holliman, I. Hooper,
M. Kalinowski, D. Hangleiter, J. P. Bonilla Ataides, B. Horning, J. Hostetter, D. Hothem, J. Houlton,
N. Maskara, I. Cong, X. Gao, P. Sales Rodriguez, J. Hout, R. Hutson, R. T. Jacobs, T. Jacobs, M. Jo-
T. Karolyshyn, G. Semeghini, M. J. Gullans, hannsen, J. Johansen, L. Jones, S. Julian, R. Jung,
M. Greiner, V. Vuletić, and M. D. Lukin, Logical A. Keay, T. Klein, M. Koch, R. Kondo, C. Kong,
quantum processor based on reconfigurable atom ar- A. Kosto, A. Lawrence, D. Liefer, M. Lollie, D. Luc-
rays, Nature 626, 58 (2024). chetti, N. K. Lysne, C. Lytle, C. MacPherson,
[53] J. A. Muniz, D. Crow, H. Kim, J. M. Kindem, A. Malm, S. Mather, B. Mathewson, D. Maxwell,
W. B. Cairncross, A. Ryou, T. C. Bohdanowicz, L. McCaffrey, H. McDougall, R. Mendoza, M. Mills,
C.-A. Chen, Y. Ji, A. M. W. Jones, E. Megidish, R. Morrison, L. Narmour, N. Nguyen, L. Nugent,
C. Nishiguchi, M. Urbanek, L. Wadleigh, T. Wilka- S. Olson, D. Ouellette, J. Parks, Z. Peters, J. Pet-
son, D. Aasen, K. Barnes, J. M. Bello-Rivas, ricka, J. M. Pino, F. Polito, M. Preidl, G. Price,
I. Bloomfield, G. Booth, A. Brown, M. O. Brown, T. Proctor, M. Pugh, N. Ratcliff, D. Raymond-
K. Cassella, G. Cowan, J. Epstein, M. Feldkamp, son, P. Rhodes, C. Roman, C. Roy, C. Ryan-
C. Griger, Y. Hassan, A. Heinz, E. Halperin, Anderson, F. B. Sanchez, G. Sangiolo, T. Sawadski,
T. Hofler, F. Hummel, M. Jaffe, E. Kapit, K. Kotru, A. Schaffer, P. Schow, J. Sedlacek, H. Semenenko,
J. Lauigan, J. Marjanovic, M. Meredith, M. Mc- P. Shevchuk, S. Shore, P. Siegfried, K. Singhal,
Donald, R. Morshead, S. Narayanaswami, K. A. S. Sivarajah, T. Skripka, L. Sletten, B. Spaun, R. T.
Pawlak, K. L. Pudenz, D. R. Pérez, P. Sabhar- Sprenkle, P. Stoufer, M. Tader, S. F. Taylor, T. H.
 12

Thompson, R. Tobey, A. Tran, T. Tran, G. Vit- Appendix A: Proofs

torini, C. Volin, J. Walker, S. White, D. Wilson,
Q. Wolf, C. Wringe, K. Young, J. Zheng, K. Zuraski, We provide the detailed definitions and proofs
C. H. Baldwin, A. Chernoguzov, J. P. Gaebler, S. J.
of the lemmas for fault tolerance with Knill error
Sanders, B. Neyenhuis, R. Stutz, and J. G. Bohnet,
Helios: A 98-qubit trapped-ion quantum computer correction (Knill EC) presented earlier.
(2025), arXiv:2511.05465 [quant-ph].
[56] B. W. Reichardt, Improved ancilla preparation
1. Definitions
scheme increases fault-tolerant threshold (2004),
arXiv:quant-ph/0406025.
[57] A. Goswami, M. Mhalla, and V. Savin, Fault- Definition 1. Locally decaying (LD) proba-
tolerant preparation of quantum polar codes en- bility distribution: For a given finite set A, a
coding one logical qubit, Physical Review A 108, probability distribution p over the power set of A,
042605 (2023). p : Pow[A] → [0, 1] is locally decaying with rate τ
[58] A. Goswami, M. Mhalla, and V. Savin, Factory- iff for any subset B ⊆ A, the total probability that
based fault-tolerant preparation of quantum polar
any subset B ′ ⊆ A drawn according to it is upper
codes encoding one logical qubit, Physical Review
A 110, 012438 (2024). bounded as,
[59] A. Gong and J. M. Renes, Computation with quan- X
tum Reed-Muller codes and their mapping onto 2D p(B ′ ) ≤ τ |B| . (A1)
atom arrays (2024), arXiv:2410.23263 [quant-ph]. B ′ :B⊆B ′ ⊆A
[60] G. M. Sommers, M. Foss-Feig, D. Hayes, D. A. Huse,
and M. J. Gullans, Observation of a fault tolerance
Definition 2. Locally decaying error model on
threshold with concatenated codes, Physical Review quantum states: A locally decaying Pauli error
Research 7, 043271 (2025). model Nτ with rate τ on n qubits is a quantum chan-
[61] N. Kanomata and H. Goto, Fault-tolerant quantum nel specified by Kraus operators consisting of Pauli
computing with a high-rate symplectic double code operators E ∈ P̂n drawn with a probability pNτ (E)
(2025), arXiv:2509.15457 [quant-ph]. that satisfies the locally decaying distribution, i.e.,
[62] Y. Ibe, Y. Hirano, Y. Ozu, T. Kawakubo, and the probability distribution of its support is locally
K. Fujii, Measurement-Based Fault-Tolerant Quan-
decaying with rate τ ,
tum Computation on High-Connectivity Devices: A
Resource-Efficient Approach toward Early FTQC X
(2025), arXiv:2510.18652 [quant-ph]. pNτ (E ′ ) ≤ τ |E| , (A2)
[63] H. Yamasaki and M. Koashi, Time-Efficient E ′ : supp(E)⊆supp(E ′ )
Constant-Space-Overhead Fault-Tolerant Quantum
Computation, Nature Physics 20, 247 (2024). where supp(E) is the support of the non-trivial Pauli
[64] S. Yoshida, S. Tamiya, and H. Yamasaki, Concate- factors in E and |E| = |supp(E)| is its weight.
nate codes, save qubits, npj Quantum Information Aplocally decaying noise channel Nτ is denoted as
11, 1 (2025). { pNτ (E)E}E∈P̂n , where pNτ satisfies the locally
[65] D. Litinski, Blocklet concatenation: Low-overhead decaying condition.
fault-tolerant protocols for fusion-based quantum
computation (2025), arXiv:2506.13619 [quant-ph]. Definition 3. Locally decaying error model for
[66] M. B. Hastings, A Class of Cyclic Quantum Codes a quantum circuit: Consider a faulty quantum
(2025), arXiv:2509.06865 [quant-ph]. circuit Q, defined as a collection of fault locations
[67] E. Murphy, Hex: modular simulation toolkit for
(which could be wait locations, multi-qubit gates,
quantum error correction (2026).
[68] C. Horsman, A. G. Fowler, S. Devitt, and R. V. and single qubit measurements). By FQ we refer
Meter, Surface code quantum computing by lattice to the power set of all fault locations, over which we
surgery, New Journal of Physics 14, 123011 (2012). define a probability distribution pF : FQ → [0, 1].
[69] S. Aaronson and D. Gottesman, Improved simula- The fault model is locally decaying with rate τ , if
tion of stabilizer circuits, Physical Review A 70, the total probability that given fault chain F ∈ F is
052328 (2004). included is upper-bounded by τ |F | , where |F | is the
number of fault locations in F , and not the support
of F .
 13

2. Proof of Lemma 1 Using the locally-decaying (LD) bounds,

Proof. For E ⊆ D ∪ A, define Pr({di : i ∈ I1 } ⊆ SD ) ≤ p1 1 ,

|I |
(A6)
|I |
T (E) := {i ∈ [n] : ei ∩ E ̸= ∅}, Pr({ai : i ∈ I2 } ⊆ SA ) ≤ p2 2 , (A7)
|I |
Pr(I3 ⊆ F ) ≤ p3 3 . (A8)
where ei = (di , ai ) is an “edge” (cardinality-two
set) corresponding to the CNOT between di and ai .
Since each edge covers two qubits, Therefore,
   
|E|
|T (E)| ≥ . (A3) Pr
\
Bi  ≤
X |I | |I | |I |
p1 1 p2 2 p3 3
2
i∈T (E) I1 ,I2 ,I3
For qubits in edge ei to end up in error support partition of T (E)
on that edge, it must be because either the input = (p1 + p2 + p3 )|T (E)| . (A9)
qubits to that edge were affected by a prior error, or
the edge CNOT was faulty. Let SD and SA be the Combining (A9), and (A3), we obtain
support of the effective Pauli operators on D and
A blocks. Let F ⊆ {1, 2, · · · , n} (edge indices) be Pr(E) ≤ (p1 + p2 + p3 )⌈|E|/2⌉
the set with faulty CNOT location. Thus, for E to
be an error in the output, then for every i ∈ T (E), ≤ (p1 + p2 + p3 )|E|/2
√
|E|
either one of these options must be true: di ∈ SD or = p1 + p2 + p3 , (A10)
ai ∈ SA or i ∈ F .
Define where the first inequality holds assuming p1 + p2 +
p3 ≤ 1.
Bi := {di ∈ SD } ∪ {ai ∈ SA } ∪ {i ∈ F } .
| {z } | {z } | {z } Let Sout denote the Pauli support induced by the
1 2 3 input noise and the noisy CNOT. This is further
Introduce the indicator function σ : T (E) → affected by measurement noise. Let SM ⊆ D ∪ A
{1, 2, 3} the error type affecting the CNOT edge in denote the Pauli support induced by measurement
T (E). Thus, the probability of having error E is, noise, assumed to be locally decaying with rate p4 :

|E|
\
Pr(E) = Pr( Bi ) Pr(E ⊆ SM ) ≤ p4 ∀E ⊆ D ∪ A. (A11)
i∈T (E)
Using closure of locally-decaying distributions un-
 
der union, for any E ⊆ D ∪ A,
[ \
= Pr   {di ∈ SD }∩
σ ∈{1,2,3}T (E)
⃗ i:σ(i)=1 X
 Pr(E) ≤ Pr(E1 ⊆ Sout ) Pr(E2 ⊆ SM )
\ \
E1 ∪E2 =E
{ai ∈ SA } ∩ {i ∈ F } X √
|E1 | |E2 |
i:σ(i)=2 i:σ(i)=3 ≤ p1 + p 2 + p 3 p3
E1 ∪E2 =E
X
≤ Pr({di : i ∈ I1 } ⊆ SD ) × √
|E|
⃗
σ = p1 + p 2 + p 3 + p 4 . (A12)
Pr({ai : i ∈ I2 } ⊆ SA ) Pr(I3 ⊆ F ) , (A4)
Therefore the effective LD rate on the data and
where first auxiliary block is
Ij := {i ∈ T (E) : σ(i) = j}, √
peff = p1 + p 2 + p 3 + p 4 . (A13)
j ∈ {1, 2, 3},
|I1 | + |I2 | + |I3 | = |T (E)|. (A5)
 14

Appendix B: Modularised circuit simulations protocols depending on the previous protocols cor-
rections. As an example consider the circuit in Fig-
Here we will provide a more detailed description ure 2. The measurements for protocol 3 will be up-
of the modular simulation tools. The goal of this dated if the corrections from protocols 1 and 2 anti-
tool is to perform end-to-end simulations of the com- commute with its measurements.
position of fault-tolerant protocols, while allowing We now give a detailed description of our simu-
each protocol to have independent decoding. lation Algorithm 1. There are a total of N fault-
Our tool is designed for Clifford fault-tolerant tolerant protocols, each with a Clifford circuit Ci ,
protocols that take the following form: a Clifford a classical function fi , and a set of possible Pauli
circuit, a classical function that processes the mea- corrections Pi where i ∈ {1, . . . , N }. There are
surements of this circuit (usually the decoder) and Ti independent Pauli corrections in Pi . Compos-
a set of Pauli corrections applied by the classical ing the circuits from all of the protocols gives our
function. We focus on these protocols because the full circuit C = C1 ◦ C2 ◦ · · · ◦ CN , each protocol
Clifford circuits allow for efficient simulation and the has Mi measurements with the total circuit having
corrections being Pauli operators allows for the clas- M = M1 + · · · + MN measurements.
sical functions to be applied during post-processing. As discussed earlier, it is straightforward to con-
Our tool supports a wide range of fault-tolerant struct a function FC : P → FM 2 that calculates which
protocols such as error correction gadgets, logical measurements in C are flipped by introducing a new
state preparation, and logical Clifford gates (e.g. lat- Pauli gate into the circuit, where P is the set of all
tice surgery [68]). But the tool does not support possible Pauli insertions into the circuit. We note
protocols that require non-Clifford gates or Clifford that P1 ∪ . . . ∪ PN ⊆ P. The map from Pauli inser-
correction. Including methods to use post-selection, tions to measurement flips is linear, FC (P ◦ P ′ ) =
such state distillation, would require a minor mod- FC (P ) ⊕ FC (P ′ ), allowing the pre-computation of
ification to the algorithm presented in this section. measurement flips for the independent Pauli correc-
A boolean value will need to be added to the output tions of each protocol. Ki = FC (Pi ) ∈ FT2 i ×M is the
of the classical function indicating if a sample needs matrix that maps each of the Pauli corrections in
to be post-selected. This is a minor change that protocol i to their measurement flips.
does not impact the rest of the algorithm, therefore, Circuit C is sampled Nshots times producing m =
shot ×M
for the remainder of this section we will not discuss [m1 |m2 | · · · |mN ] ∈ FN 2 , where mi is the sam-
post-selection. pled data from circuit Ci . The classical function
Now we consider how to simulate a sequence of takes these sampled measurements as input and de-
fault-tolerant (Clifford) protocols. A naive approach termines which combination of Pauli operators in Pi
shots ×Mi shots ×Ti
would be to implement a direct stabilizer tableau to apply fi : FN 2 → FN 2 . The measure-
simulation [69], stopping after each protocol to ap- ments that then need to be flipped to account for
ply the Pauli corrections before moving onto the the corrections of protocol i are Bi = fi (mi ) · Ki .
next protocol. However, using Stim it is much faster The measurements are updated, m ← m ⊕ Bi , be-
to sample all of the measurements of the circuit at fore the procedure moves onto the next protocol. If
once. The disadvantage is that the measurements one of these protocols includes the measurement of
of a protocol have been sampled without the cor- a logical observable then its updated measurement
rections from the previous protocols. Therefore, we values can be used to calculate logical errors.
need a way to update the measurements to what To provide clarity on what Ci , fi and Pi are for
they would be if we had applied the Pauli correc- different protocols we will give two examples. First,
tions during the simulation. the transversal Bell measurement used in Knill er-
This can be done in pre-processing by propagat- ror correction, shown in Figure 1. For this Ci will be
ing the Pauli corrections of each protocol through the circuit for transversal Bell measurement shown
the full circuit to identify which measurements are in the shaded region of the figure. The classical
flipped by these corrections. This information can function fi will then be the combination of calculat-
then be used to update the measurements of future ing the syndrome, decoding (D), applying the recov-
 15

Algorithm 1: Sampling Modular Circuit |0⟩ X ⊕ D2

Input: {Ci }, {fi }, {Pi }, Nshots |0⟩ X ⊕ D2
Output: Corrected measurements m ..
. ..
.
1 for i = 1 to N do |0⟩ X ⊕ D2
2 Compute the measurement flips caused by
the Ti independent Pauli corrections in Pi : |+⟩ Z D2 D2 D2 Z H
Ki ← FC (Pi ) ∈ FT2 i ×M ;
3 Sample the measurements of the full circuit, FIG. 5: Circuit-level error model for state
C = C1 ◦ C2 ◦ · · · ◦ CN , with Nshot shots to get preparation and stabilizer measurement.
shot ×M
m ← [m1 |m2 | · · · |mN ] ∈ FN 2 ;
4 for i = 1 to N do
5 Apply the classical function for protocol i to the bit-flip and phase-flip channels applied before
its measurement data fi (mi ) ∈ FN 2
shot ×Ti
; measuring in the Z and X basis, respectively. Ev-
6 Find which measurements need updating to ery error channel in the circuit is parametrised by a
account for this protocol’s corrections single error probability 0 < q < 1. Definitions for
shots ×M
Bi ← fi (mi ) · Ki ∈ FN 2 ; the error channels in terms of q are as follows:
7 Update the measurements m ← m ⊕ Bi ;
• Single qubit depolarising channel:
8 return m
q q q
Eq (ρ) = (1 − q)ρ + XρX + Y ρY + ZρZ,
3 3 3
ery operation (R), and calculating the logical Pauli
eigenvalues. The set of Pauli corrections Pi will be • 2-qubit depolarising channel:
the 2k logical Pauli operators applied to the third Eq (ρ) = (1 − q)ρ
code block, resulting in a Ti value of 2k. Second, we q 
+ IXρIX + IY ρIY + IZρIZ
will consider the repeated syndrome measurement 15
used in state preparation, shown in Figure 3. The +XIρXI + Y IρY I + ZIρZI
circuit Ci is the preparation of the physical qubits +XXρXX + XY ρXY + XZρXZ
and then d rounds of stabilizer measurement. The +Y XρY X + Y Y ρY Y + Y ZρY Z
classical function fi will calculate the detector values 
+ZXρZX + ZY ρZY + ZZρZZ ,
from the measurements and decode these using the
detector error model and the offline decoder. The
Pauli corrections Pi are defined as the error loca-
tions in the detector error model. • Bit-flip channel:
These examples illustrate that this tool is not Eq (ρ) = (1 − q)ρ + qXρX,
limited to decoding using detector error models, but
allows a variety of different quantum error correction
• Phase-flip channel:
protocols to be combined and simulated.
Eq (ρ) = (1 − q)ρ + qZρZ.

Appendix C: Circuit-level error model

Figure 5 shows this error model applied to state
preparation and stabilizer measurement, and Fig-
We use a circuit-level depolarising error model ure 6 shows this error model applied to transversal
for our simulations. Physical qubits are prepared in Bell measurement.
either the |0⟩ or |+⟩ state followed by a bit-flip or
phase-flip Pauli error channel, respectively. Single
qubit unitary gates are followed by the 1-qubit de- Appendix D: Compressed Knill error correction
polarising channel. Two qubit unitary gates are fol-
lowed by the 2-qubit depolarising channel. We ex- For CSS codes you can perform a compressed
clusively consider single-qubit measurements, with form of Knill EC [14, 46, 47]. Instead of using a
 16

2-qubit Depolarising Channel

Z-error channel

D2 Z H Calc. Calc. Calc. Calc.

D2 Z H Z1 X1 Zk Xk
D2 Z H
···
X-error Channel D R ⊗ ⊗ ⊗ ⊗
⊕ D2 X
⊕ D2 X Z1 X1 Zk Xk
 ⊗k ⊕ D2 X
√1 |00⟩ + |11⟩

2
X1 Z1 ···
Xk Zk

FIG. 6: Circuit-level error model for transversal Bell measurement.

H
Calc. Calc.
|ψ⟩ H DX R ···

H
X1 Xk

⊕ ⊕

⊗k DZ
Calc. Calc.

|0⟩ ⊕
Z1 ···
Zk ⊕
R ···

⊕ ⊕
Z1 Zk

|+⟩⊗k X1 ···
Xk |ψ⟩

FIG. 7: Circuit for the compressed Knill error correction protocol.

⊗k ⊗k
logical Bell state you use a logical |0⟩ and a |+⟩⊗k teleporting using the |+⟩ state we only decode us-
state and perform two teleportations. When tele- ing the Z stabilizers. A diagram of this compressed
state the online decoding Knill protocol is shown in Figure 7.
⊗k
porting using the |0⟩
is done only for the X stabilizers, similarly, when