Fault-tolerant quantum computing with color codes

Andrew J. Landahl,1, 2, ∗ Jonas T. Anderson,2, † and Patrick R. Rice2, 3, ‡

1
Advanced Device Technologies, Sandia National Laboratories, Albuquerque, NM, 87185, USA
2
Center for Quantum Information and Control, University of New Mexico, Albuquerque, NM, 87131, USA
3
Quantum Institute, Los Alamos National Laboratories, Los Alamos, NM, 87545, USA
We present and analyze protocols for fault-tolerant quantum computing using color codes. To
process these codes, no qubit movement is necessary; nearest-neighbor gates in two spatial dimen-
sions suffices. Our focus is on the color codes defined by the 4.8.8 semiregular lattice, as they
provide the best error protection per physical qubit among color codes. We present circuit-level
schemes for extracting the error syndrome of these codes fault-tolerantly. We further present an
integer-program-based decoding algorithm for identifying the most likely error given the (possibly
faulty) syndrome. We simulated our syndrome extraction and decoding algorithms against three
physically-motivated noise models using Monte Carlo methods, and used the simulations to estimate
arXiv:1108.5738v1 [quant-ph] 29 Aug 2011

the corresponding accuracy thresholds for fault-tolerant quantum error correction. We also used a
self-avoiding walk analysis to lower-bound the accuracy threshold for two of these noise models. We
present two methods for fault-tolerantly computing with these codes. In the first, many of the op-
erations are transversal and therefore spatially local if two-dimensional arrays of qubits are stacked
atop each other. In the second, code deformation techniques are used so that all quantum process-
ing is spatially local in just two dimensions. In both cases, the accuracy threshold for computation
is comparable to that for error correction. Our analysis demonstrates that color codes perform
slightly better than Kitaev’s surface codes when circuit details are ignored. When these details are
considered, we estimate that color codes achieve a threshold of 0.082(3)%, which is higher than
the threshold of 1.3 × 10−5 achieved by concatenated coding schemes restricted to nearest-neighbor
gates in two dimensions [Spedalieri and Roychowdhury, Quant. Inf. Comp. 9, 666 (2009)] but lower
than the threshold of 0.75% to 1.1% reported for the Kitaev codes subject to the same restrictions
[Raussendorf and Harrington, Phys. Rev. Lett. 98, 190504 (2007); Wang et al., Phys. Rev. A 83,
020302(R) (2011)]. Finally, because the behavior of our decoder’s performance for two of the noise
models we consider maps onto an order-disorder phase transition in the three-body random-bond
Ising model in 2D and the corresponding random-plaquette gauge model in 3D, our results also
answer the Nishimori conjecture for these models in the negative: the statistical-mechanical classi-
cal spin systems associated to the 4.8.8 color codes are counterintuitively more ordered at positive
temperature than at zero temperature.

I. INTRODUCTION developed, with threshold estimates ranging from as low

as 10−6 [9] to as high as 3% [10–12], depending on the
protocol and the noise and control assumptions.
The promise of fault-tolerant quantum computing is An important control constraint relevant for several
a crowning achievement of quantum information science quantum computing technologies is that the only multi-
[1–8]. Under a specific set of noise and control assump- qubit gates that are possible are those between nearest-
tions, the promise is that any ideal quantum circuit of size neighbor qubits, where the qubits are laid out in some
L can be simulated to any desired precision ε by a faulty 2D geometry in which each qubit neighbors a constant
quantum circuit whose size is at most O(ε−1 L loga L) for number of other qubits. Fault-tolerant quantum com-
some (small) constant a. Fault-tolerant quantum com- puting protocols based on concatenated quantum error-
puting protocols are judged by the resources they employ correcting codes have a fractal structure that is not
in the course of a simulation. Examples of such resources commensurate with such a geometry. Indeed, forcing
include the constant a, the hidden constant in the big- such codes into a semiregular 2D geometry requires that
O notation, and the requirements imposed by the noise one introduce a substantial number of additional qubit-
and control assumptions. Often protocols are compared movement operations that expose the protocol to more
by a requirement encapsulated in a single number, the errors, thereby diminishing its accuracy threshold. The
accuracy threshold, which is an upper bound on the error largest accuracy threshold of which we are aware for a
probability per elementary operation that a faulty cir- concatenated-coding protocol in a semiregular 2D geom-
cuit must satisfy for the protocol to work. A variety of etry is 1.3 × 10−5 [13]; that protocol is based on the
fault-tolerant quantum computing protocols have been concatenated nine-qubit Bacon-Shor code [14] embedded
in the 2D square lattice.
Cognizant of the constraints imposed by 2D geom-
∗ [email protected] etry, Kitaev introduced a family of quantum error-
† [email protected] correcting codes called surface codes that require only
‡ [email protected] local quantum processing, where locality is defined by
 2

a graph embedded in a surface [15]. Several fault- for noise that afflicts the individual quantum circuit ele-
tolerant quantum computing protocols have been devel- ments used in a fault-tolerant color-code-based quantum
oped around surface codes [16–18], and these protocols computing protocol. The accuracy threshold for noise af-
have significantly higher accuracy thresholds than their flicting the circuit model is perhaps the most instructive
concatenated-coding counterparts. Numerical threshold of all table entries. This is because this threshold es-
estimates for surface-code protocols range from 0.75% to tablishes the target error rate per elementary operation
1.1% [17–19]; an analytic proof in Ref. [16] guarantees that a quantum technology must meet to admit fault-
that it is no less than 1.7 × 10−4 . tolerant quantum computation using these codes. It also
Recently Bombin and Martin-Delgado proposed a new allows for a fair “apples-to-apples” comparison to the
family of quantum error-correcting codes they call color high thresholds estimated for Kitaev’s surface codes in
codes which are also defined to be local relative to a graph the circuit model.
embedded in a surface [20]. Specifically, they are defined In this Article, we analyze the accuracy threshold of
by face-three-colorable trivalent graphs in the following the 4.8.8 color codes for fault-tolerant quantum compu-
way: on each vertex of the graph lies a qubit, and for each tation under several noise and control models. We have
face f of the graph, one defines two “stabilizer genera- restricted our analysis to protocols which use the decoder
tors” or “checks,” Xf and Zf . Xf is the tensor product that identifies the most likely error (MLE) given the er-
of Pauli X operators on each qubit incident on face f , ror syndrome. We formulate the MLE decoder as an in-
while Zf is the tensor product of Pauli Z operators on teger program (IP), which in general is NP-hard to solve
each qubit incident on face f . The color code’s codespace [33]. Although the decoder is inefficient, it establishes
is defined as the simultaneous +1 eigenspace of each of a threshold that we expect is close the the maximum
the check operators. threshold possible for these codes, namely the one ob-
A fault-tolerant quantum computing protocol based tainable by an optimal decoder, which identifies the most
on color codes requires an infinite family of color codes likely logical operation given the error syndrome. For
of increasing size in order to be able to simulate arbi- small codes, the MLE IP can be solved “offline” ahead of
trarily large ideal quantum circuits to increasing preci- time to generate a lookup table that can be used during
sion. A natural source for an infinite color-code family the course of a “live” fault-tolerant quantum comput-
is a uniform tiling of the plane by a trivalent face-three- ing protocol. Our results comprise both numerical esti-
colorable lattice. Such a lattice can be embedded in any mates of the accuracy threshold achieved via Monte Carlo
orientable surface, although later we will restrict atten- simulations and a rigorous lower bound on the accuracy
tion to embeddings in planar discs. These “semiregular” threshold that we prove using combinatorial counting ar-
or “Archimedean” lattices are described in vertex nota- guments.
tion as r.s.t, where each vertex is locally surrounded by The remainder of this Article is organized as follows. In
an r-gon, an s-gon, and a t-gon. The only possible triva- Sec. II, we lay out the control model and the three noise
lent face-three-colorable tilings of the plane are the 4.8.8 models we consider. In Sec. III, we summarize the prop-
lattice, the 6.6.6 (hex) lattice, and the 4.6.12 lattice, de- erties of the 4.8.8 triangular color codes we study, present
picted in Fig. 1 [21]. two circuit schedules for extracting the error syndrome in
these codes, and formulate MLE decoders for these codes
as integer programs for each of the noise models that we
consider. In Sec. IV, we report our numerical estimates
for the accuracy threshold for fault-tolerant quantum er-
ror correction of these codes for each of the noise models
that we consider. In Sec. V, we use a self-avoiding-walk
analysis to prove rigorous lower bounds for the accuracy
(a) 4.8.8 (b) 6.6.6 (c) 4.6.12 thresholds of fault-tolerant quantum error correction of
these codes against two of the noise models that we con-
FIG. 1: The three possible face-three-colorable trivalent uni- sider. In Sec. VI, we relate the quantum error correction
form tilings of the plane. accuracy threshold to the quantum computation accu-
racy threshold for two scenarios: one in which logical
Accuracy thresholds for fault-tolerant quantum com- qubits are associated with 2D planes that are stacked
puting have been estimated for color codes in several atop one another like pancakes and the other in which
highly idealized noise models numerically. The values of logical qubits are associated with “defects” in a single
these thresholds are summarized in Table I, along with 2D substrate. In Sec. VII we conclude, summarizing and
analogous estimates for a well-studied surface code and interpreting our results both in terms of the accuracy
two recently-proposed topological subsystem codes. This thresholds we report and in terms of their consequences
table contains numerous gaps, some of which we fill in for “re-entrant behavior” of an order-disorder phase tran-
with the results of this Article—the entries containing our sition in two associated classical statistical-mechanical
results are highlighted in bold. The most significant gap, models. We cap off our conclusions with some parting
which we fill, is an estimate of the accuracy threshold thoughts about future directions that we believe are wor-
 3

Code Capacity Phenomenological Circuit-based

Code Other MLE Optimal MLE Optimal Other MLE
8.87 %a,b [22] 10.56(1) % 10.9(2) % [23] 3.05(4) % “∼ 0.1 %”a,b,c [22] 0.082(3) %
4.8.8
8.7 %d [24] (Our result) 10.925(5) % [25] (Our result) (Our result)
10.9(2) % [23] 4.5(2) % [26]
6.6.6
10.97(1) % [25]
4.6.12
10.31(1) % [27] 10.9187 % [28] 2.93(2) % [27] 3.3 % [29] 0.75 %a [17]
4.4.4.4 Kitaev
10.939(6) % [30] 1.1 %a [19]
3.4.6.4 TSCC 1.3 %a,c [24]
“SBT” [31] 1.3 %a,c [31]
a Reference computes threshold against DP channel, not BP chan-

nel. For non-circuit-based noise models, the decoder used does not
account for correlations between bit flips and phase flips in DP
channel. In these models, we reported the result for the equivalent
effective BP channel of strength 23 p.
b Decoder based on hypergraph matching heuristic.
c Limited numerics only weakly suggest this value.
d Decoder based on mapping to two Kitaev codes.

TABLE I: Numerically-estimated accuracy thresholds for several topological quantum error-correcting codes, noise models,
and decoding algorithms. The first three codes (4.8.8, 6.6.6, 4.6.12) are the color codes described in Fig. 1 and its preceding
text. The last three codes are the Kitaev surface code on the square lattice [15], a topological subsystem color code on the
3.4.6.4 lattice [32], and a hypergraph-based topological subsystem code proposed by Suchara, Bravyi, and Terhal [31]. The
details of the noise models (code capacity, phenomenological, and circuit-based) and decoders (MLE, optimal, and other) are
discussed in the text; when possible, results from other references have been translated into one of these models. The notation
“x.y1 · · · yk (z)%” means x.y1 · · · yk % ± (z × 10−k )%. When such notation is not used, it means that the no error analysis was
reported in the reference from which the value was drawn.

thy of study. Some FTQC protocols also make the following archi-
tectural assumptions, which generally lead to higher ac-
curacy thresholds; we make these assumptions here:
II. NOISE AND CONTROL MODEL
4. Reliable classical computation. Classical com-
The performance of a fault-tolerant quantum comput- putations always return the correct result.
ing (FTQC) protocol is strongly influenced by underlying
5. Fast classical computation. Classical computa-
architectural assumptions, so it is important to clearly
tions are instantaneous.
list what they are. Indeed, when those assumptions are
not borne out in real quantum information technologies, 6. No qubit leakage. Qubits never “leak” out of the
an FTQC protocol may fail entirely [34, 35]. computational Hilbert space.
Every existing FTQC protocol makes the following ar-
chitectural assumptions—assumptions which appear to 7. Uncorrelated noise. Each qubit and gate is af-
be necessary: flicted by an independent noise source.
1. Nonincreasing error rate. The asymptotic scal-
Some additional architectural assumptions, which have
ing of the error rate as a function of the circuit’s
a less clear impact on the accuracy threshold, are fre-
size is nonincreasing. This allows the performance
quently made as well; we also make these assumptions:
of fault-tolerant circuits to increase asymptotically.
2. Parallel operation. The asymptotic parallel- 8. Standard gate basis. The set of (faulty) quan-
processing rate is larger than a constant times the tum gates (including preparation and measure-
asymptotic error rate. This allows error correction ment) available consists of |0i, |+i, I, X, Z, T ,
to keep ahead of the errors. S, CNOT , MZ , and MX . The definition of what
these gates are can be found in standard textbooks,
3. Reusable memory. The asymptotic rate at e.g., in Refs. [36, 37].
which one can erase or replace qubits is larger than
a constant times the asymptotic error rate. This al- 9. Equal-time gates. Each gate, including prepara-
lows entropy to be flushed from the computer faster tions and measurements, takes the same amount of
than it is generated by errors. time to complete.
 4

10. Uniformly faulty gates. Each k-qubit gate, circuits other than those used for syndrome ex-
including preparations and measurements, is as traction (e.g., for encoded computation) are still
equally as faulty as every other k-qubit gate. subject to the BP and DP channels, respectively,
as before.
Inspired by the limitations of 2D geometry for some
quantum computing technologies, we also make the fol- 13(c). Code capacity noise. This model is the same
lowing assumptions: as the phenomenological noise model, except that
11. 2D layout. Qubits are laid out on a struc- the syndrome-bit error rate is assumed to be zero.
ture describable by a graph embedded in a two- Because there is no need to repeat syndrome mea-
dimensional surface. surements in this model, and because the accuracy
threshold for “defect-braided” quantum computa-
12. Local quantum processing. Gates can only cou- tion is the same as that for quantum memory (as
ple nearest-neighbor qubits in the graph describing argued later), the accuracy threshold for this noise
their layout. model is the same as what in quantum informa-
Finally, we make the following three variants of a thir- tion theory is called the single-shot, single-letter
teenth assumption about the noise model afflicting each quantum capacity for color codes subject to the
gate. Of all the assumptions we make, we have found BP channel.
that this one is most likely to vary in the literature.
Commonly-studied alternatives for this assumption in-
clude stochastic adversarial noise [38–41], purely depo- III. FAULT-TOLERANT ERROR CORRECTION
larizing noise [42], and noise that has a strong bias, such OF COLOR CODES
as having phase flips significantly more probable than bit
flips [43]. A. Code family

13(a). Circuit-level noise. Each faulty single-qubit

preparation and faulty coherent single-qubit gate We confine our analysis of color codes to the 4.8.8 color
(|0i, |+i, I, X, Z, H, T , S) acts ideally, followed codes; our choice is motivated by two factors. First, of
by the bit-flip channel of strength p, which applies the three color codes on semiregular 2D lattices, the 4.8.8
bit flips (Pauli X operators) with probability p, code uses the fewest qubits per code distance. Second,
followed by the phase-flip channel of strength p, the 4.8.8 code is the only one of the three which can
which applies phase flips (Pauli Z operators) with realize encoded versions of the entire “Clifford group”
probability p. We call this channel the BP chan- [36] of quantum gates, namely the gates which conjugate
nel. Each faulty single-qubit measurement (MX , Pauli operators to Pauli operators in the Heisenberg pic-
MZ ) acts as the BP channel of probability p fol- ture, in a transversal fashion [20], i.e., by applying the
lowed by a measurement that returns the incorrect same operation to every qubit in a code block or be-
result with probability p. Importantly, this noise tween corresponding qubits in two code blocks. In par-
model assumes that the state after a measurement ticular, the gates X, Z, H, S, and CNOT have transver-
is in an eigenstate of the observable measured, just sal encoded implementations for these codes. When en-
perhaps not the eigenstate that the measurement coded gates are implemented transversally, fault-tolerant
indicates. Each CNOT gate acts ideally followed quantum computing protocols for simulating these gates
by a channel in which each of the 16 two-factor are generally simpler, leading to more favorable accuracy
Pauli products (II, IX, XI, XY , etc.) is ap- thresholds. The Clifford group of gates is an important
plied with probability p/16. We call this chan- group of gates for stabilizer codes such as the color codes,
nel the DP channel. This model differs slightly since error correction can be carried out solely using those
from a frequently-studied variant in the literature gates [44].
in which each of the 15 nontrivial two-factor Pauli We further restrict our analysis to planar color codes,
products is applied with probability p/15 and the namely those which are embedded in the disc (a sphere
identity is applied with probability 1 − p. with one puncture). We do this because, for all quantum-
computing technologies of which we are aware, arranging
13(b). Phenomenological noise. This noise model is qubits on a flat disc is more plausible than arranging
the same as the circuit-level noise model (13(a)), them on a more general surface like a torus. The graph
except that the circuit for syndrome extrac- constraints defining color codes require that planar color
tion (described later) is modeled “phenomenolog- codes have a boundary shaped like a polygon having 3m
ically,” having a probability p for returning the sides for some positive integer m. A 3m-sided planar
wrong syndrome bit value. In this model, the color code encodes m logical qubits; we restrict attention
propagation of errors between data qubits and to the simplest case in which m = 1. In other words, our
between data and ancilla qubits induced by the focus on this paper is on triangular color codes. Exam-
syndrome extraction circuit are ignored. Single- ples of three different triangular color codes are depicted
qubit and two-qubit gates on the data qubits in in Fig. 2.
 5

The code distance of a triangular color code is equal to side colors are indicated explicitly in Fig. 4.
its side length, namely the number of qubits along a side
of the defining triangle. To see this, notice that the logi-
cal X and Z operators for the logical qubit are transver-
sal because they are encoded Clifford gates. Thus, when
one multiplies a logical X or Z operator by all checks
of the same Pauli type, except the checks incident on a
specified side, one obtains an equivalent logical opera-
tor whose Pauli-weight is equal to the that side’s length.
The family of 4.8.8 triangular codes we study is generated
according to the pattern depicted in Fig. 3. Note that
the smallest triangular code (for any of three triangular
code families depicted in Fig. 2) is equivalent to the well-
known Steane [[7, 1, 3]] code [45]; triangular codes offer a FIG. 4: A green-colored chain in a triangular code. The chain
connects a green-colored side of the 4.8.8 triangular code to a
way to generate an infinite code family from the Steane
green octagonal face. If qubits are flipped (are in error) along
code by a means other than concatenation [72]. this chain, it will only be detected by this terminal octagonal
check operator.

B. Syndrome extraction

To record each error-syndrome bit, the relevant data

(a) 4.8.8 code (b) 6.6.6 code (c) 4.6.12 code qubits interact with one or more ancilla qubits and the
ancilla qubits are then measured. Shor [1], Steane [46],
FIG. 2: Three distance d = 11 triangular codes encoding and Knill [10] have devised elaborate methods for ex-
one qubit, drawn from the 4.8.8, 6.6.6, and 4.6.12 lattices tracting an error syndrome to minimize the impact of
respectively. For general d, these codes have length n equal ancilla-qubit errors spreading to the data qubits. For
to 12 d2 + d − 21 , 34 d2 + 14 and 23 d2 − 3d + 52 respectively. The
topological codes, however, such elaborate schemes are
asymptotic ratio of d2 to n is highest for the 4.8.8 codes.
not necessary; a single ancilla qubit per syndrome bit
suffices. This is because, by choosing an appropriate or-
der in which data qubits interact with the ancilla qubit,
the locality properties of the code will limit propagation
of errors to a constant-distance spread. Using more elab-
orate ancillas is possible, and in general there is a trade-
off in the resulting accuracy threshold one must examine
between the reduction in error propagation complexity
(a) d = 3 (b) d = 5 (c) d = 7 offered versus the additional verification procedures re-
quired. Here, we examine the simplest case, with one
FIG. 3: 4.8.8 color codes of sizes 3, 5, and 7. ancilla qubit per syndrome bit. By placing two syn-
drome qubits at the center of each face f (one for the
Although the colors of the faces in a color code have Xf measurement and one for the Zf measurement), the
no intrinsic meaning for the algebraic structure of the syndrome extraction process can be made spatially local,
code other than constraining the class of graphs on which in keeping with the spirit of the semiregular 2D geometry
color codes are defined, it is useful to use the colors as constraints we are imposing.
placeholders in discussions from time to time. To that Because color codes are Calderbank-Shor-Steane
end, we will refer to the colors of the faces as “red,” (CSS) codes [47, 48], syndrome bits can be separated
“green,” and “blue.” We will further assign a color to into those which identify Z errors (phase flips) and those
each edge so that an edge’s color is complementary to which identify X errors (bit flips). These correspond to
the colors of the two faces upon which it is incident. We the bits coming from measuring the Xf and Zf opera-
will call a set of vertices lying on a collection of edges of tors respectively. The circuit for measuring an operator
the same color connected by faces also having that color a Xf is identical to the one for measuring the operator Zf ,
“colored chain;” an example of a colored chain is depicted except with the basis conjugated by a Hadamard gate;
in Fig. 4. We will assign colors to each side of a triangular examples of bit-flip and phase-flip extraction circuits for
code so that the color of the side is complementary to the the square faces in the 4.8.8 color code are depicted in
colors of the faces terminating on that side; for example, Fig. 5.
in Figs. 2 and 3, the left sides of the triangles are blue, In a full round of syndrome extraction, both Xf and Zf
the right sides are green, and the bottoms are red. These must be measured for each face f . One way of schedul-
 6

MX %"#$ |0i     MZ %"#$


|+i • • • • 1,5 2,6 1,5 2,6


•


• 3,7 4,8 1,5 2,6 3,7 4,8


, •
• 5,1 3,7 4,8 5,1
6,2 6,2
1,5 2,6
FIG. 5: Six-step circuits for measuring X ⊗4 and Z ⊗4 . 7,3 8,4 7,3 8,4
1,5 2,6 3,7 4,8 1,5 2,6

ing this is to perform all Xf measurements in parallel 3,7 4,8 5,1 3,7 4,8
6,2
followed by all Zf measurements in parallel. The mini- 1,5 2,6 1,5 2,6
mal number of steps (ignoring preparation and measure- 7,3 8,4
ment) for parallel Xf measurements is eight; an example 3,7 4,8 1,5 2,6 3,7 4,8
of such a schedule is depicted in Fig. 6. The Zf mea-
surements can be carried out by the same schedule, but 5,1 3,7 4,8
6,2 5,1 6,2
in the Hadamard-conjugated basis as depicted in Fig. 5.
A complete syndrome extraction round using this sched- 7,3 8,4 7,3 8,4
ule then takes 20 steps: 10 for the Xf measurement and
10 for the Zf measurement. For this schedule, one only FIG. 7: Schedule with X and Z syndromes measured concur-
needs to have one, not two, syndrome qubits at the center rently, in “interleaved” fashion. This schedule takes 8 steps,
of each face. plus an extra step for syndrome qubit preparations, plus an
extra step for syndrome qubit measurements. The label m, n
1 2 at a vertex indicates that at time step m the qubit at that
vertex interacts with the X-syndrome qubit via a CNOT gate
and at time step n the qubit at that vertex interacts with the
3 4 Z-syndrome qubit via a CNOT gate.

5 6 code syndrome-extraction circuit [39]. For color codes,

a priori, it is not clear that using a simpler syndrome-
1 2 extraction circuit will yield an analogous improvement.
7 8 This is because these circuits are not constructed using
3 4 1 any fault-tolerant design principles—catastrophic error
2
propagation is halted by the codes’ structure, not by
circuit-design principles. It may be the case, in fact, that
5 6 3 4 a simpler circuit will allow errors to propagate to a larger
set of qubits than a less simple one. The set of errors to
which individual errors are propagated by a syndrome-
7 8 extraction circuit are called “hooks” in Ref. [16]. An ex-
ample of how an error can propagate to a “hook” using
FIG. 6: Simple syndrome extraction circuit schedule. A round the schedule of Fig. 7 is depicted in Fig. 8.
of X checks is followed by a round of Z checks. The number Neither the 20-step nor the 10-step schedule is neces-
at each vertex corresponds to the discrete time step in which sarily optimal in the sense of yielding the highest thresh-
the physical qubit at that vertex interacts with the syndrome old for a fixed number of time steps; we leave that op-
qubit at the face’s center via a CNOT gate. The same sched- timization to others. Indeed any schedule that satisfies
ule is used for both X and Z checks, but with the direction two constraints is valid: (1) no qubit can be acted upon
of the CNOT gates reversed. by two gates at the same time and (2) any stabilizer
generator for an error-free input state (including ancilla
The circuit for a full syndrome extraction round can be syndrome qubits) must propagate to an element of the
optimized to use fewer time steps when both syndrome stabilizer group for an error-free output state. Satisfy-
qubits in a face can be processed in parallel. An exam- ing this second criterion is not trivial; for example, an
ple of an “interleaved” schedule that uses ten steps is “obvious” schedule that acts on each face in a clockwise
depicted in Fig. 7. fashion in a manner obeying constraint (1) will not sat-
We calculate estimates for the accuracy threshold for isfy constraint (2).
both schedules, to assess the impact of compressing The number of steps in the syndrome extraction round
the schedule. Some authors who have reported im- can be reduced further to eight steps if we prepare the
proved thresholds for concatenated-coding schemes us- ancillas for the octagon measurements not √ in single-qubit
ing Bacon-Shor codes attribute the improvement in large states but in cat-states (|0i⊗8 +|1i⊗8 )/ 2 and use Shor’s
part to the simplicity of the fault-tolerant Bacon-Shor- method of syndrome extraction [1]. (One can also use
 7

9 10 9 10 to be faulty, such as in the circuit-level and phenomeno-

logical noise models that we study. This ensures that
11 12 5 12
6 11 errors in the syndrome bit values can be suppressed as
well as errors in the data qubits can be suppressed.
13 14 7 8 13 14
19 2 10
15 16 15 16 C. Decoding algorithm
5 6 3 11 4 12 5 6

The process of decoding refers to a classical algorithm

7 8 5 13 7 8
9 10 6 14 9 10
for identifying a recovery operation given an error syn-
drome, regardless of whether the code from which the
7 15 8 16
12 5 12 syndrome was derived is classical or quantum. Impor-
11 6 11
tantly, decoding does not refer to “unencoding,” or per-
forming the inverse of encoding. For classical linear
13 14 7 8 13 14
codes, the optimal decoding algorithm is the Most Likely
Error (MLE) algorithm, which identifies the recovery op-
15 16 15 16 eration to be the most likely pattern of bit-flip errors
given the syndrome. In general, this algorithm is NP-
FIG. 8: A single X error that occurs between time steps five
hard [33], but there are many families of codes for which
and six on the syndrome qubit for measuring X ⊗8 , indicated
by the small red circle, will propagate to other X errors ac-
the algorithm is known to be efficient.
cording to the arrows. Note that an even number of X flips For quantum stabilizer codes, MLE decoding identi-
is equivalent to no flip at all. Errors that propagate to other fies the recovery operation to be the most likely n-qubit
syndrome qubits will not propagate further because the syn- Pauli-group error given the syndrome. (The process of
drome qubits are refreshed before each syndrome extraction extracting the syndrome forces every error to “collapse”
round. This particular error causes three data qubits to flip. onto a definite n-qubit Pauli-group operator, which is
These flips are correctly detected by the yellow-colored syn- why it is sufficient to restrict to this family of opera-
drome bits. tors.) MLE decoding is not necessarily optimal for quan-
tum stabilizer codes. This is because quantum error-
correcting codes can be degenerate, meaning that two
four-qubit cat states and create an eight-step schedule, distinct correctable errors can map to the same error syn-
as demonstrated in Ref. [49].) Eight steps is the absolute drome. Color codes are examples of highly degenerate
minimum possible for syndrome extraction, since each codes. The optimal decoding algorithm for quantum sta-
qubit must be checked by six different syndrome bits, bilizer codes instead identifies the recovery operation to
which must also be prepared and measured. While using be one that causes the most likely logical operator to be
cat states reduces the circuit depth, the cat states need applied after recovery. This is akin to a doctor prescrib-
to be verified. We opted not to study this schedule be- ing medicine that is most likely to cure the ailment rather
cause the verification is stochastic, which would lead to a than prescribing medicine that cures the most likely ail-
difficult synchronization problem for a large-sized code. ment.
That said, such a schedule has the potential to offer a Once a decoding algorithm has identified a recovery
larger accuracy threshold. operation, which is some n-qubit Pauli-group operator,
Because there is an inherent asymmetry in the order in it need not necessarily be applied. Because the process
which we choose to perform Xf and Zf measurements, of applying the recovery operation is subject to faults,
we will report two threshold results, one for the Xf mea- it is wiser to wait until the end of the computation and
surements and one for the Zf measurements. When we apply the net recovery operation rather than apply it
only report one value, we are reporting the lower of the after each decoding step. One can even propagate the
two threshold values. For the phenomenological noise correction past the final qubit measurements at the end
model, we choose to model the Xf and Zf syndrome ex- of the quantum computation, where the recovery oper-
traction processes as occurring synchronously rather than ation becomes completely classical and fault-free. The
one followed by the other, since so many of the details of catch is that one must (classically) adaptively update
the circuit are washed away in the model anyway. This one’s “Pauli frame” after each decoding iteration by per-
has the advantage of enabling the accuracy threshold in muting the interpretation of the Pauli operators X, Y ,
the phenomenological model to be identified with a phase and Z on each qubit as suggested by the recovery oper-
transition in an associated random-bond Ising model, as ation. (The Pauli operators get conjugated by the Pauli
described in Ref. [23]. We will discuss this connection in error identified by the decoder.)
more detail in Secs. IV A and VII B. For fault-tolerant quantum error correction and a num-
Finally, it is worth reminding that the entire syndrome ber of interesting encoded quantum circuits, only Clif-
extraction round is repeated a number of times equal to ford gates are required. Since Clifford gates propagate
the distance of the code when measurements are allowed Pauli operators to Pauli operators in the Heisenberg pic-
 8

ture, one can efficiently track the changing Pauli frame as follows:
through these gates, as guaranteed by the Gottesman-
Knill theorem [50]. One can safely defer applying recov- min 1T x (4)
ery operations until after final measurement in each of sto Hx = s mod 2 (5)
these circuits. However, for universal quantum compu- x ∈ Bn , (6)
tation, at least one non-Clifford gate is required. In our
protocols, the only such gate we use is the classically- where 1 denotes the all-ones vector and H is the parity
controlled S † gate, depicted later in the circuit of Fig. 24. check matrix associated with the Zf -checks. (For color
Because this gate propagates a Pauli error to a sum of codes, this is the face-vertex incidence matrix.)
Pauli errors, it is necessary to actually apply the recovery To take advantage of well-developed numerical opti-
operation before all but a constant number of these gates mization software, it is helpful to replace the linear alge-
in order to prevent the number of terms required to track bra over GF (2) in this mathematical program with linear
one’s “Heisenberg frame” from growing exponentially. algebra over R. One way to do this is to introduce “slack
We develop MLE decoders for triangular 4.8.8 color variables” into the optimization problem. Because each
codes for the three noise model settings we study: code check operator in the code has Pauli weight four or Pauli
capacity, phenomenological, and circuit-based. For the weight eight, each row of H has Hamming weight four or
code-capacity and phenomenological settings, the only Hamming weight eight. This means that the f th compo-
operations are single-qubit measurements and identity nent of the vector on the left hand side of constraint (5) is
gates. This means that they involve no circuitry that a sum of four or eight binary xv variables that must equal
could map X errors to Z errors or vice-versa. Because of sf modulo 2. The modulo 2 restriction can be dropped by
this, and because our noise model is one in which single- replacing s by s + 2z1 + 4z2 + 8z3 in the constraint, where
qubit operations are subject to BP channel noise (which the zi are binary “slack variable” vectors that allow the
applies X errors and Z errors independently), decoding LHS to sum to any integer from 0 . . . 15. While there
can factor into bit-flip decoding and phase-flip decoding can be many degenerate solutions to this revised opti-
separately. Because color codes are also “strong” CSS mization problem having different zi values, any solution
codes [37], the MLE decoders for bit-flip and phase-flip generates the same optimal x as before. By combining
errors are in fact identical; for concreteness, we formulate the zi variables and the x variables into a single vector
the decoder for Zf syndrome bits here. y = (xT , zT1 , zT2 , zT3 )T , the slack-variable version of the
program becomes the following linear binary integer pro-
gram in which the variables are restricted to be binary
but in which the linear algebra is over R:
1. Code capacity MLE decoder
min cT y (7)
In the code-capacity setting, we have a single error- sto Ay = s (8)
free m-bit syndrome s = (s1 , . . . , sm )T where sf = 0 y ∈ Bn , (9)
when Zf is measured to have eigenvalue +1 and sf = 1
when Zf is measured to have eigenvalue −1. (The value where c is a vector containing n ones followed by 3m zeros
of m is a function of the code size; for the triangular and A is the matrix generated by adjoining matrices to
n-qubit distance-d 4.8.8 color code, m = (d + 1)2 /4 − 1 H as
and n = (d + 1)2 /2 − 1.) We assign a binary variable xv  
to each vertex v indicating whether or not the recovery A := H | −2I | −4I | −8I , (10)
operation calls for the qubit at vertex v to be bit-flipped
(have Pauli X applied). The objective of MLE decod- in which each I denotes the m × m identity matrix.
ing is to minimize the number of xv variables that are There are a number of symmetries that color codes
assigned the value 1 subject to the constraint that the possess which allow one to significantly reduce the com-
parity of the xv variables on each face is consistent with plexity of this binary IP. For example, if y satisfies the
the observed syndrome. This can be expressed as the constraints of the IP, then so does y with any number
following mathematical optimization problem: of faces complemented. Since complementing the face of
X any optimal solution will not reduce its weight, we know
min xv (1) that each face’s sum will never be more than half the
v weight of that face. This means that for any particular
M instance of the IP specified by the syndrome vector s, the
sto xv = sf ∀f (2)
sums for the octagon and square faces can only take the
v∈f
syndrome-dependent values listed in Table II, thereby re-
xv ∈ B := {0, 1}. (3) ducing the number of slack variables required. We take
advantage of these kind of symmetries in the software we
This optimization problem can be expressed as a linear developed code for estimating the code capacity of 4.8.8
binary integer program (IP) over the finite field GF (2) triangular color codes. For example, we never need to
 9

Octagon Square
s = 0 0, 2, 4 0, 2
s=1 1, 3 1

TABLE II: Possible values octagonal and square face check

sums can take for an optimal IP solution if the face check
sum’s parity s is fixed.

use three slack variables and some times we need none at

all. (a) Measurement Error. (b) Data Error.
Maximum likelihood decoding is generally an NP-hard
problem, and the color codes do not appear to fall into FIG. 9: If syndrome qubits are also allowed to be in error, we
an “easy” subset of instances. This is unfortunate be- repeat syndrome measurements. Time advances from bottom
cause their close cousins, the surface codes, do have ef- to top. Yellow circles indicate syndrome bits with the value
ficient MLE decoders that can be solved as a minimum- 1. Solid yellow circles indicate bit-flip errors.
weight perfect matching problem [16]. Nevertheless, we
can solve the associated IP for reasonably small instance
sizes. variables to make the problem a linear binary IP over
the reals. Because the left-hand side of the constraints
in Eq. (13) can sum to up to ten for octagon constraints
2. Phenomenological noise MLE decoder and up to six for square constraints, three slack variables
again suffice, allowing us to formulate the optimization
In the phenomenological noise model, the syndrome problem as
values themselves can be faulty so we repeat the syn-
drome extraction process a number of times equal to the min cT y (15)
distance of the code. In this setting, it is the difference
in syndrome bit values from one time step to the next sto Ay = ∆s (16)
rather than the absolute values at particular times step y ∈ Bn , (17)
that indicate data errors. This is because a single data
error at one time step will lead to flipped syndrome bits where c is a vector containing (n + m)d ones followed by
for all future time steps (assuming that the syndrome ex- 3md zeros, ∆s is the vector (∆sT1 , . . . , ∆sTd )T , y is the
traction is not faulty), and such a syndrome-bit history vector (xT1 , . . . , xTd , rT1 , . . . , rTd , zT1 , zT2 , zT3 )T and A is the
should not imply that data errors occurred at each time matrix
step—it should imply that a data error occurred only at
the time step when the syndrome bit first changed its  
value. The difference in persistence between data and H I
 H I I
 
syndrome errors is depicted in Fig. 9. The input to a 
A= .. .. −2I −4I −8I
.
MLE decoder is therefore the collection of syndrome dif- 
 . . 
ference vectors for all time steps, namely
H I I
∆st = st − st−1 = (st + st−1 ) mod 2 ∀t, (11) (18)
where s0 := 0.
For a distance d color code, the optimization problem Finally, as we did for the code capacity setting, we can
to solve is again to minimize the number of errors given use symmetries to reduce the complexity of solving this
the observed syndrome, except we now have d time steps’ IP; Table III summarizes what the possible values are for
worth of data-error vectors, x1 , . . . , xd , and d time steps’ the square-faced and octagonal-faced constraints.
worth of syndrome-error vectors, r1 , . . . , rd , as variables
in the optimization problem. Mathematically, we can
write the optimization problem as Octagon Square
X s = 0 0, 2, 4, 6 0, 2, 4
min 1T x t (12)
s = 1 1, 3, 5 1, 3
t
sto (Hxt + rt + rt−1 ) mod 2 = ∆st mod 2 ∀t (13)
TABLE III: Possible values octagonal and square face check
x ∈ Bn . (14) sums can take for an optimal IP solution if the face check
sum’s parity s is fixed.
As we did for the code-capacity scenario, we can collect
these constraints into a single constraint and add slack
 10

3. Circuit-level decoder operator, indicating failure of the decoding algorithm.

Since error-correction is assumed to be error-free in this
In the circuit-level noise model, each component of the noise model, the corrected state is guaranteed to be in
syndrome extraction circuit can fail with a probability the codespace. Because (a) the logical bit-flip operator
that is a function of a parameter p, so that the over- is transversal, (b) all stabilizer group elements have even
all probability of a syndrome bit being in error, ps , is weight, and (c) there are an odd number of qubits in
a complicated function of p. Even more dauntingly, the every triangular code, it follows that one can identify a
circuits can induce correlated errors between syndrome decoding failure quickly by computing whether the parity
bits and between syndrome bits and data qubits. The of the error pattern equals the parity of its IP-inferred
phenomenological-noise model does not capture these correction; this means that is suffices to just store the
noise correlations. parity of the inferred correction for each pre-computed
We developed an MLE decoder for the circuit-level IP instance. The probability of failure, pfail is therefore
noise model that accounts for both these induced error X
correlations and the fact that in this noise model, single- pfail = p|E| (1 − p)n−|E| , (19)
qubit operations are subject to BP-channel noise while failing patterns E
CNOT gates are subject to DP-channel noise. However,
this decoder uses exponentially many more constraints where |E| denotes the Hamming weight of the bit-flip
than the phenomenological decoder as a function of code error pattern E.
size. Because the IP decoder is already NP-hard, we We carried out this tabulation for the smallest trian-
opted not to study this truly MLE decoder but rather gular 4.8.8 color codes of distances 1, 3, 5, and 7 (corre-
use the phenomenological-noise MLE decoder, which ig- sponding to 1, 7, 17, and 31 qubits respectively) and com-
nores these subtleties. Taking correlations into account puted the corresponding exact polynomials. To speed up
will likely boost the accuracy threshold, but probably not the computation, we used several symmetries. For ex-
by large factors [51]. By way of comparison, the thresh- ample, it suffices to examine only half of the error pat-
old for the square-lattice surface code in the circuit-level terns because if the decoding algorithm succeeds on an
noise model is 0.68% when the phenomenological decoder error pattern, it fails on its complement and vice versa.
is used [52] (0.75% [17] when using a non-MLE decoder Also, up to overall complementation, every error pattern
that takes into account some entropic effects), a thresh- can be uniquely expressed as the modulo-2 sum of an
old value that has recently been boosted to 1.1% [19] by IP-inferred minimal-weight error pattern and a pattern
accounting for some of the correlations in the noise. We where a bit-flip stabilizer group element has support. Fi-
leave the refinement of true MLE decoding of this noise nally, the decoding algorithm is guaranteed to work on
model to others. all errors whose weight is less than the code’s distance,
so those error patterns do not need to be examined.
The formulas we obtained for the smallest codes of
IV. NUMERICAL ESTIMATE OF THE distance 1, 3, and 5 (code sizes 1, 7, and 17) are:
ACCURACY THRESHOLD FOR
FAULT-TOLERANT QUANTUM ERROR (1)
pfail = p (20)
CORRECTION
(3)
pfail = p7 + 7p6 (1 − p) + 28p4 (1 − p)3 (21)
A. Code capacity noise model 3
+ 7p (1 − p) + 21p (1 − p)4 2 5
(22)
(5) 17 16 15 2
Because the [[n, 1, d]] triangular 4.8.8 color codes are pfail =p + 17p (1 − p) + 136p (1 − p) (23)
14 3 13 4
CSS codes, when they are subject to BP-channel noise + 348p (1 − p) + 725p (1 − p) (24)
of strength p, their code capacity is the same as their 12
+ 3861p (1 − p) + 4764p (1 − p) 5 11 6
(25)
bit-flip or phase-flip capacity; we focus on the bit-flip
10 7 9 8
capacity here for definiteness. The number of distinct + 12136p (1 − p) + 9747p (1 − p) (26)
bit-flip syndromes is 2(n−1)/2 and the number of distinct 8
+ 14563p (1 − p) + 7312p (1 − p) 9 7 10
(27)
bit-flip errors is 2n . For small n, one can pre-solve the 6 11 5 12
MLE decoding IP for each of the 2(n−1)/2 distinct bit- + 7612p (1 − p) + 2327p (1 − p) (28)
4 13 3 14
flip syndromes. One can then iterate through each of the + 1655p (1 − p) + 332p (1 − p) . (29)
2n distinct error patterns, compute its syndrome, and
determine whether the combination of the error pattern The formula we obtained for the distance-7 triangular
plus the inferred correction by the IP leads to a logical 4.8.8 color code (31 qubits) is a bit more hefty:
 11

(7)
pfail = p31 + 31p30 (1 − p) + 465p29 (1 − p)2 + 4495p28 (1 − p)3 + 25658p27 (1 − p)4 + 96790p26 (1 − p)5 (30)
25 6 24 7 23 8 22 9
+ 344858p (1 − p) + 1288630p (1 − p) + 3742943p (1 − p) + 10488241p (1 − p)
+ 21436239p21 (1 − p)10 + 44259329p20 (1 − p)11 + 67781868p19 (1 − p)12 + 106951476p18 (1 − p)13
+ 127137964p17 (1 − p)14 + 155845748p16 (1 − p)15 + 144694447p15 (1 − p)16 + 138044561p14 (1 − p)17
+ 99301599p13 (1 − p)18 + 73338657p12 (1 − p)19 + 40412986p11 (1 − p)20 + 22915926p10 (1 − p)21
+ 9671834p9 (1 − p)22 + 4145782p8 (1 − p)23 + 1340945p7 (1 − p)24 + 391423p6 (1 − p)25 + 73121p5 (1 − p)26
+ 5807p4 (1 − p)27 .

Our computing resources did not allow us to compute

the exact polynomial for the next-sized code (distance 0.45
d=3
9 code on 49 qubits), so we resorted to a Monte Carlo 0.4 0.165
(a) d=5
d=7
d=9
estimate for pfail (p). We did this by first selecting three 0.16
0.35
values of p near where we believed the threshold to be. 0.155
0.15
For each p, we generated N trial error patterns drawn 0.3
0.145
from the Bernoulli distribution, namely in which we ap-

PFailure
0.25 0.14

plied a bit-flip on each of the n qubits with probability 0.135

0.2 0.13
p. We then inferred the syndrome for each error pattern 0.125
(b)

and checked whether or not it led to a decoding failure 0.15 0.101 0.103 0.105 0.107 0.109 0.111 0.1422
0.142
for the MLE decoder. The optimal unbiased estimator 0.1 0.1418
0.1416
for pfail that we used is 0.05
0.1414
0.1412
0.141
0.1054 0.1055 0.1056 0.1057
(est) Nfail 0
pfail = (31) 0 0.05 0.1
PError
0.15 0.2

N
with a variance of
 
(est) (est)
pfail 1 − pfail FIG. 10: Code capacity for the 4.8.8 triangular color codes.
2
(σfail )(est) = . (32) pth = 10.56(1)%. Error bars on Monte Carlo data reflect 105
N instances studied at each of the three corresponding values of
To get reasonably small error bars in these estimates, p. The inset figures are zoom-ins near the crossing point to
show greater resolution there.
given where we believed the threshold to be, we chose
N = 105 . The polynomials for pfail (p) are plotted in
Fig. 10, including our three points of Monte Carlo data.
From these plots, we estimate the accuracy threshold for phase transition in a random-bond Ising model (RBIM)
this noise model to be 10.56(1)%. The error we report of classical spins [16, 23]. For the color codes, the Ising
in this value comes from the error analysis method we model features 3-body interactions, whereas for the sur-
describe in detail in the next section. face codes, the Ising model features 2-body interactions.
To put our result in context, we reference Table I. The MLE decoder in both settings corresponds to the
The threshold value of 10.56(1)% we find is is slightly order-disorder transition in the spin model at zero tem-
higher than the corresponding MLE threshold for the perature, whereas the optimal decoder corresponds to
code capacity 10.31(1)% of 4.4.4.4 surface codes. In- the order-disorder transition at the temperature along
tuitively this makes sense, as the 4.8.8 color code has the so-called “Nishimori line,” where the randomness in
both weight-8 and weight-4 stabilizer generators, both the bond couplings equals the randomness in the state
of which are modeled as being measured instantaneously arising from finite temperature fluctuations. In both the
and ideally. Being able to measure high-weight genera- surface-code and color-code settings, the small decrease
tors quickly should improve the performance of a code, in accuracy threshold when going from optimal to MLE
which is the effect we observe. decoding reflects that the phase-boundary in these mod-
Our threshold is also less than the threshold value of els is re-entrant, but only by a small amount. Our results
10.925(5)% for optimal decoding, which is also not sur- therefore imply a violation of the so-called Nishimori con-
prising. As with the 4.4.4.4 surface codes, the reduction jecture [53, 54], which conjectures that the spin model
in threshold is not very significant. For both the surface shouldn’t become more ordered as the temperature in-
codes and the 4.8.8 color codes, the accuracy thresh- creases. The violation that our results imply is depicted
old in the code capacity noise model corresponds to a in cartoon fashion in Fig. 11. To our knowledge, the vio-
 12

lation of the Nishimori conjecture for the 3-body RBIM Algorithm 1 : pfail (p) by Monte Carlo
is unknown before our work. We expand more on this
connection in Sec. VII B. 1: nfaces ← 14 (d + 1)2 − 1.
2: for i = 1 to N do

3: // Generate data and syndrome errors for d time slices.

4: for t = 1 to d do
5: for j = 1 to n do
Disordered
T

6: E[t, j] ← 1 with probability p. // Data errors.

Phase 7: end for
8: for j = n + 1 to n + 1 + nfaces do
Ordered 9: E[t, j] ← 1 with probability p. // Synd. errors.
Phase 10: end for
11: end for
12: Emin ← LDecode(Syndrome(E)). // 3D error volume.
13: E 0 ← t E[t] ⊕ Emin [t]. // 2D error plane.
0
14: Emin ← Decode(Syndrome(E 0 )). // Ideal decoding.

P if ( i E 0 [i] ⊕ Emin
L 0
15: [i] = 1) then
Pc,0 Pc 16: Nfail ← Nfail + 1.
17: end if
FIG. 11: Phase diagram for 3-body random-bond Ising 18: end for
model. The dark circle is called the Nishimori point. The (est)
19: return pfail = Nfail /N .
dotted line is the expected phase boundary given by the Nishi-
mori conjecture. Our value of code capacity (10.56(1)%) es-
tablishes that the T = 0 intercept is Pc,0 , while results of
Ohzeki [25] (10.925(5)%) establish that the Nishimori point are depicted in Fig. 12. Just as for surface codes, the
occurs at Pc . Because P c 6= Pc,0 , the Nishimori conjecture phenomenological noise MLE decoder can be mapped
for this model is false. to a random-plaquette gauge model (RPGM) on clas-
sical spins such that the zero-temperature order-disorder
phase transition in the spin model corresponds to the ac-
curacy threshold of the color codes. Because of this, as
B. Phenomenological noise model argued in Ref. [27], the mutual intersection of the curves
in Fig. 12 at the threshold pc corresponds to critical be-
havior in the spin model such that the spin correlation
In the phenomenological noise model, our fault-
length ξ scales as
tolerant quantum error correction protocol repeats syn-
drome extraction multiple times to increase the reliability ξ ∼ |p − pc |−ν0 , (33)
of the syndrome bits. This causes the number of possi-
ble error patterns for a given code size to grow so rapidly where ν0 is a critical exponent set by the universality
that obtaining exact curves for pfail (p) even for small code class of the spin model.
sizes is intractable. We therefore resorted to Monte Carlo
estimates for these curves for even the smallest code sizes.
The specific Monte Carlo algorithm we used for comput- 0.2
ing pfail at a fixed value of p is listed in Algorithm 1. d=5
d=7
d=9
In words, Algorithm 1 creates an estimator for pfail 0.18

by assessing the performance of many simulated trials of 0.16

faulty quantum error correction. In each trial, errors are
laid down, giving rise to an observed syndrome history. 0.14
PFailure

From the syndrome history, a correction is inferred. The 0.12

actual error history and the inferred error history are

0.1
XORed onto a single effective time slice, but the state in
this effective time slice is not necessarily in the codespace. 0.08

To achieve this, a fictional ideal (error-free) round of error

0.06
correction is simulated. If this succeeds (i.e., if it does
not generate a logical bit-flip operation), then the trial 0.04
0.024 0.026 0.028 0.03 0.032 0.034 0.036
is deemed a success; otherwise it is deemed a failure. By PError
repeating many trials, one obtains an optimal unbiased
estimator for the failure probability pfail , with mean and
variance given by Eqs. (31–32), identical to the formulas
relevant in the code capacity noise model setting. FIG. 12: Monte Carlo data used to estimate the accuracy
Our plots of pfail versus p for small-distance color codes threshold in the phenomenological noise model.
 13

For a sufficiently large code distance d, then, the failure C. Circuit-level noise model
probability should scale as
pfail = (p − pc )d1/ν0 . (34) As with the phenomenological noise model, computing
pfail (p) exactly even for small code sizes is intractable, so
We use our Monte Carlo data to fit to this form, but we again appeal to Monte Carlo estimation. Our Monte
as in Ref. [27], we allow for systematic corrections com- Carlo simulation algorithm is similar to Algorithm 1, ex-
ing from finite-size effects that create a constant offset. cept the manner in which the error pattern E is generated
Specifically, we use the method of differential corrections is different. To generate E, we simulate BP and DP chan-
[55] to fit the curves to the form nel noise as described by the noise model on the explicit
pfail = A + B(p − pc )d1/ν0 . (35) circuit given for syndrome extraction. This results in a
correlated error model for syndrome and data qubits. We
The linear fits to our data are plotted in Fig. 13. Using then use the phenomenological noise MLE decoder and
the software of Ref. [55], we found the following values assess success or failure as we did for that noise model.
for pc and ν0 : We estimated the pfail (p) curves for several small 4.8.8
pc = 0.030 534 ± 0.000 385 (36) triangular color codes for both the X-then-Z schedule of
Fig. 6 and the interleaved X-Z schedule of Fig. 7. Our
ν0 = 1.486 681 ± 0.166 837. (37)
results are plotted in Figs. 14 and 15.

0.2 0.6
d=5 d=5
d=7 d=7
d=9 d=9

0.5

0.15

0.4

PFailure
PFailure

0.14

0.1 0.3
0.12

0.10

0.2 0.08

0.06
0.05
0.04
0.1
0.02
0.0006 0.0007 0.0008 0.0009 0.0010

0 0
0.02 0.025 0.03 0.035 0.04 0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01

PError PError

FIG. 13: Linear fit near curve crossings of phenomenological- FIG. 14: Monte Carlo data used to estimate accuracy thresh-
noise-model Monte Carlo data. Estimated accuracy threshold old in the circuit-based noise model in which the noninter-
is pth = 3.05(4)%. leaved syndrome extraction circuit is used.

To put our results in context, as we did in the code ca- To compute the accuracy thresholds from our data,
pacity setting, we reference Table I. For the same reasons we again fit our data near the crossings to an equations
as in the code capacity noise model setting, the threshold whose form is similar to that of by Eq. (35). However, the
we compute is larger than the MLE decoder’s threshold motivation for such a fit is a bit more tenuous in this case
for the 4.4.4.4 surface codes. We conjecture that is it because while the MLE decoder we are using maps to a
also measurably less than the threshold for the optimal RPGM, the noise model which generates it is correlated.
color-code decoder, as is the case for optimal vs. MLE For this reason, as also found in Ref. [27], we found it
decoding for surface codes. So far, the threshold for opti- necessary to include a quadratic term, unlike the case for
mal decoding of 4.8.8 color codes has not been estimated, the pure phenomenological noise model. In other words,
but the analysis for optimal decoding of 6.6.6 color codes we fit our data to an equation of the form
suggests that the threshold will be near 4.5%. If true, our pfail = A + B(p − pc )d1/ν0 + C(p − pc )2 d2/ν0 . (38)
data would signal a violation of the Nishimori conjecture
for the RPGM associated with the 4.8.8 color code, some- The quadratic fits to our data for the X-then-Z sched-
thing we are not aware of being reported elsewhere. ule are plotted in Fig. 16. Again using the software of
Finally, we note that while the value of ν0 is consistent Ref. [55], we found the following values for pc and ν0 for
with value of ν0 = 1.463(6) obtained for the 4.4.4.4 sur- the X-then-Z schedule:
face code [27] and the 6.6.6 color code, the uncertainty in
pc = 0.000 820 ± 0.000 022 (39)
the value we obtained is too high to draw any meaningful
conclusions. ν0 = 1.350 954 ± 0.079 188. (40)
 14

circuit-level noise model is about a factor of ten less than

0.18
d=5
the the corresponding 0.68% accuracy threshold for MLE
0.16
d=7
d=9 decoding of 4.4.4.4 surface codes in the circuit-level noise
model. We believe that the difference comes from the fact
0.14
that the 4.8.8 codes have some weight-8 stabilizer genera-
0.12 tors while the 4.4.4.4 codes only have weight-4 stabilizer
PFailure

0.1
generators. This causes the circuits for extracting the
0.18
syndrome for the weight-8 generators in the 4.8.8 codes
0.08 0.16
0.14 to be larger, inviting more avenues for failure. Indeed, we
0.12
0.06
0.1 have investigated the finite-sized error-propagation pat-
0.08
0.04 0.06 terns for the 4.8.8 codes such as the one depicted in Fig. 8,
0.04

0.02
0.02
0
and they are significantly larger and more complex than
0.0006 0.0008 0.001 0.0012 0.0014
the corresponding patterns for the 4.4.4.4 surface codes.
0
0.0006 0.0008 0.001 0.0012 0.0014 0.0016 0.0018 0.002 Expanding this line of reasoning, we predict that the 6.6.6
PError color codes will have an MLE-decoded accuracy thresh-
old in the circuit-based noise model that is somewhere
between the 4.8.8 and 4.4.4.4 accuracy thresholds in this
noise model.
FIG. 15: Monte Carlo data used to estimate the accuracy
threshold in the circuit-based noise model in which the inter-
leaved syndrome extraction circuit is used.
V. ANALYTIC BOUND ON THE ACCURACY
THRESHOLD FOR FAULT-TOLERANT
To be clear, there is both a Z-error and an X-error QUANTUM ERROR CORRECTION
accuracy threshold; we report the smaller of the two here.
While numerical estimates of the accuracy threshold
are valuable, equally valuable are analytic proofs that the
0.1
d=5 accuracy threshold is no smaller than a given value. One
d=7
0.09 d=9 method of obtaining such a lower bound is to use the self-
0.08
avoiding walk (SAW) method, first proposed in Ref. [16].
The idea behind this method begins with the observation
0.07
that our goal is to lower-bound the failure probability of
PFailure

0.06 decoding, which is the probability that the actual errors

0.05
plus the inferred correction (modulo 2) lead to an error
chain that corresponds to a logical operator. For color
0.04
codes, logical operators can be not only string-like but
0.03
also string-net like, as described in the original paper on
0.02 color codes [20]. They must also have a Pauli-weight at
0.01
least as large as the distance of the code. The probability
0.0005 0.0006 0.0007
PError
0.0008 0.0009 0.001
that a logical operator is present in the post-corrected
state is therefore at least as large as the probability that
an error-chain string of Pauli-weight equal to the code
distance is present. Certainly this is a very pessimistic
FIG. 16: Quadratic fit near curve crossings of noninterleaved- bound; there are many error chain strings and string-nets
circuit circuit-based-noise-model Monte Carlo data. Esti- of this Pauli weight that do not result in failure!
mated accuracy threshold is pth = 0.082(3)%. The SAW lower-bound method can be applied rela-
tively straightforwardly to the code-capacity and phe-
Similarly, for the XZ-interleaved schedule we found nomenological noise models with MLE decoding. The
method begins to break down when applied to the circuit-
pc = 0.000 800 ± 0.000 037 (41) level noise model with phenomenological MLE decoding.
ν0 = 1.509 871 ± 0.151 690. (42) One reason for this is that the circuit introduces corre-
lated errors, called “hooks” in Ref. [16], which suggest
To remind, our results are for the smaller of the X- that the SAW bounding the failure probability should be
error and Z-error thresholds. allowed to sometimes take more than one step in a sin-
Our results show that despite our efforts to shorten the gle iteration. With some finesse, this can be accounted
schedule of the syndrome extraction circuit, the impact for and bounded as in Ref. [16]. However, for the color
on the resulting accuracy threshold is essentially indis- codes, the steps need not be path-connected either. For
tinguishable. The value of 0.082(3)% for the accuracy example, the circuit may create three separated errors
threshold for MLE decoding of the 4.8.8 color codes in the on a single octagon plaquette. Calling such a process
 15

a “walk” or attempting to bound the behavior of the Fig. 17. To our knowledge, the connective constant for
process by a true SAW method is dubious at best. For this lattice is not known, but it could be computed in
this reason, we have chosen to omit bounding the accu- principle using standard methods, e.g., those outlined in
racy threshold in the circuit-level noise model and instead Refs. [56–58]. We opted to bypass this analysis and in-
have bounded the accuracy threshold only for the other stead compute a coarser bound on the failure probability.
two noise models, as described below.

A. Code capacity noise model

As argued by Dennis et al. in Ref. [16], the probability

that an [[n, k, d]] topological code decoded by an error-free
MLE decoder fails is upper-bounded by the probability
that a self-avoiding walk creates a closed path (i.e., a
self-avoiding polygon or SAP) of length d or greater:
X
pfail ≤ ProbSAP (d) (43)
L≥d
X
≤n nSAP (L) (4p(1 − p))L/2 . (44)
L≥d

Self-avoiding walks on the 4.8.8 lattice have been stud-

ied, and it is known that the number of self-avoiding poly-
gons of length L on the lattice scales asymptotically as
[56]

nSAP (L) ≤ P (L)µL

4.8.8 , µ4.8.8 ≈ 1.808 830 01(6), (45)

where P is a polynomial and µ4.8.8 is the so-called con-

nective constant for the 4.8.8 lattice. (The value µ4.8.8 FIG. 17: Prismatic lattice on which a self-avoiding walk oc-
has been rigorously bounded to be 1.804 596 ≤ µ4.8.8 ≤ curs in the analysis of the accuracy threshold for fault-tolerant
1.829 254 [57, 58].) For small p, each summand in quantum error correction using color codes in the phenomeno-
Eq. (43) is upper-bounded by the term with L = d, and logical noise model.
the number of summands is at most a polynomial in d,
so that pfail → 0 as d → ∞ as long as Because the lattice in Fig. 17 has vertices of degree ∆
equal to 6, 8, and 10, we can bound the number of SAPs
1 of length L by
p(1 − p) ≤ . (46)
4µ24.8.8
nSAP (L) ≤ 2∆max (2∆max − 1)L−1 . (48)
Solving this equation for p, we find that the code ca-
pacity threshold is at least Using ∆max = 10, we obtain a formula similar to that of
Eq. (46), namely
pc ≥ 8.335 745(1)%. (47) 1 1
p(1 − p) ≤ = . (49)
4(9)2 324
Despite the crudeness of the SAW bound, it comes
surprisingly close to the numerical value of 10.56(1) that Solving this equation for p, we find that the phe-
we estimate in Sec. IV A. nomenological noise threshold is at least
√
9−4 5
pc ≥ ≈ 0.3096%. (50)
B. Phenomenological noise model 18
This bound is nearly a factor of ten less than the value
The SAW bound method is essentially the same as for of pc = 3.05(4)% that we estimate in Sec. IV B. With
the code capacity noise model, except now errors can hap- further computational effort in determining the connec-
pen on syndrome qubits as well as data qubits and the tive constant of the governing lattice, we suspect that the
set of all relevant qubits forms a three-dimensional vol- SAW bound will still be below our numerical estimate,
ume. The relevant SAW traverses a 3D lattice that con- but significantly closer, in analogy with the relationship
nects syndrome qubits and data qubits both with them- between our SAW bound for the code capacity and the
selves and each other as dictated by the color code; the value we estimate numerically. We leave this analysis to
corresponding nonregular prismatic lattice is depicted in others wishing to tighten this bound.
 16

VI. FAULT-TOLERANT COMPUTATION WITH |ψi BP QEC

COLOR CODES
FIG. 18: Noisy identity gate. BP indicates the action of the
BP channel.
To establish a threshold for fault-tolerant quantum
computation, it is sufficient to establish three things: 1) a
threshold for fault-tolerant quantum error correction, 2) Formally, we can express the equivalence between the
a procedure for performing a universal set of gates in en- accuracy threshold for the identity gate and the accuracy
coded form, and 3) that a failure in an encoded gate that threshold for fault-tolerant quantum error correction as
occurs with probability p leads to failures in each output
codeword with probability at most p. These three ingre- (I)
pth = pth
(QEC)
. (51)
dients establish that each gate in a quantum circuit can
be simulated fault-tolerantly by performing it in encoded
form followed by fault-tolerant quantum error correction. 2. CNOT gate
We previously established the first criterion in Sec. III.
We establish the second two criteria here for two possible
The color codes are Calderbank-Shor-Steane (CSS)
computer architectures.
codes [47, 48], and for all such codes, the encoded
In the first, which we call the “pancake architecture,”
controlled-NOT (CNOT ) gate can be implemented
each logical qubit is stored in its own triangular 4.8.8
transversally, namely by applying CNOT gates between
color code and the logical qubits are stacked atop one
corresponding pairs of physical qubits in two color codes.
another. This architecture is essentially the same as the
(For color codes, fewer CNOT gates than a fully transver-
one proposed in Ref. [16]. Almost all encoded opera-
sal set also suffice.) Schematically, Fig. 19 depicts a noisy
tions are implemented transversally in this model, acting
CNOT gate. Each physical CNOT gate propagates the
on single “logical qubit pancakes” or between two such
BP channel on its control to the BP channel on its target
“pancakes.” In the second, which we call the “defect
and vice versa, so that the effective noise model seen by
architecture,” each logical qubit is stored as a connected
the fault-tolerant quantum error correction procedure on
collection of missing check operators, which we call a “de-
each code block after the encoded CNOT gate is the BP
fect,” in a single 2D 4.8.8 substrate. This architecture is
channel followed by the projection of the two-qubit DP
essentially the same as the one proposed in Ref. [17].
channel onto a single qubit. Although the DP channel
Almost all encoded operations are performed in one of
can create correlated errors between output code blocks,
two ways: encoded single-qubit gates are performed by
it will never cause a correlated error within a code block.
disconnecting a region containing the defect, operating
Since our decoder treats the noise model phenomenolog-
transversally on the region, and reconnecting the region,
ically, it does not account for DP-channel features such
while the encoded CNOT gate is performed by a sequence
as the fact that in the DP channel a Y error is more
of local measurements that cause one defect to circulate
probable than the combination of separate X and Z er-
around another.
rors. For this reason, since half of the DP-channel errors
act as a bit-flip on a given code block and half of them
act as a phase-flip on a given code block, our decoder
A. Fault-tolerance by transversal gates interprets the post-CNOT noise model as a BP chan-
nel with an effective error rate of p + p/2 for bit flips
In this section, we compute the threshold for fault- and p + p/2 for phase flips. This means that the ac-
tolerant quantum computation with triangular 4.8.8 color curacy threshold for the CNOT gate is actually 2/3 of
codes when (almost) all encoded gates are implemented the value for the identity gate. The CNOT gates used
transversally. To remind, by calling a gate “transversal,” in an encoded CNOT gate must therefore meet a more
we mean that it acts identically on all physical qubits in stringent requirement than the identity gate to be im-
a code block. For example a two-qubit transversal gate plemented transversally fault-tolerantly. (However, the
between two triangular codes acts as the same two-qubit CNOT gates used in fault-tolerant quantum error correc-
physical gate between corresponding physical qubits in tion still only need to meet the threshold for the encoded
each code block. Some authors refer to this notion of identity gate.)
transversality as strong transversality [59].
(CNOT ) 2 (I)
pth = p . (52)
3 th
1. Identity gate

The accuracy threshold for the identity gate is exactly 3. Hadamard gate
the same as the accuracy threshold for fault-tolerant
quantum error correction, by definition. Schematically, The color codes are strong CSS codes, meaning that
Fig. 18 depicts the noisy identity gate circuit. the X-type and Z-type stabilizer generators have the
 17

BP • D QEC • BP D QEC B S B QEC = S BP B QEC (57)

 D  BP D
= (53)
P S P QEC = S PP QEC (58)
BP QEC QEC

FIG. 21: Noisy phase gate. B indicates the action of the bit-
FIG. 19: Noisy CNOT . BP indicates the action of the BP
flip channel; P indicates the action of the phase-flip channel.
channel; D indicates the action of the DP channel.

is p3 + 3p(1 − p)2 . The phase gate thus has separate

same structure. As with all strong CSS codes, the en-
thresholds for bit-flip and phase-flip noise. For bit-flip
coded Hadamard gate (H) can be implemented transver-
noise, the threshold is
sally.
Like the CNOT gate, the Hadamard gate propagates 1 1 1 (I)
q
(S,bit-flip) (I)
the BP channel to the BP channel. However, since faults pth = − 1 − 2pth ≈ pth . (59)
2 2 2
in the Hadamard gate are modeled as an ideal Hadamard
gate followed by the BP channel, the effective noise model For phase-flip noise, one must solve a cubic equation to
is not one but two actions of the BP channel, as depicted get a closed-form solution for the threshold as a function
in Fig. 20. of the threshold for the identity gate. While this is pos-
sible in principle, to save space we simply state the cubic
(S,phase-flip)
equation in the variable x = pth that must be
BP H BP QEC = H BP BP QEC (54) solved and its approximate solution, which we can es-
timate because we know that the accuracy threshold is
very close to 0:
FIG. 20: Noisy Hadamard BP indicates the action of the BP
channel; D indicates action of the DP channel. x3 + 3x(1 − x)2 = pth ,
(I)
(60)
1 (I)
It is straightforward to show that two successive ap- x ≈ pth . (61)
plications of the BP channel with probability p is equiv- 3
alent to one application of BP channel with probability
2p(1 − p). This is therefore the effective post-Hadamard 5. Single-qubit measurements
noise channel, so that the threshold for the Hadamard
gate is about half of that for fault-tolerant quantum er- To destructively apply the encoded single-qubit mea-
ror correction: surements MX and MZ , we transversally measure X or Z
1 1 1 (I) on each of the qubits in the code block. We then perform
q
(H) (I)
pth = − 1 − 2pth ≈ pth . (55) classical error correction on the measurement outcomes
2 2 2
(because they may be faulty) to infer the outcome of
the encoded measurement, as depicted schematically in
4. Phase gate Fig. 22.

MZ %"#$ MZ %"#$
The color codes have the feature that each stabilizer

MX %"#$ MX %"#$
generator for the code has a Pauli weight equal to 0 mod B CEC = B CEC (62)
4 and each pair of generators are incident on 0 mod 2 P CEC = B CEC (63)
qubits. One can show that because of this, the encoded
phase gate (S) has a transversal implementation [20, 60].
(Technically, it is the transversal S † operation that acts FIG. 22: Noisy measurements. B denotes the bit-flip chan-
as an encoded S.) nel, P denotes the phase-flip channel, and CEC denotes clas-
sical error correction of the measurement outcomes. Post-
While a faulty phase gate acts as an ideal phase gate
measured states are drawn with double lines to indicate that
followed by a BP channel, the phase gate itself does not they are “classical.”
propagate the BP channel preceding it symmetrically for
bit flips and phase flips. This follows from the conjuga-
The correctness of this procedure follows from the fact
tion actions
that X and Z operators can be expressed as Z = S 2
SXS † = Y = iXZ SZS † = Z. (56) and X = HZH, and the encoded operations H and S
have previously been demonstrated to have transversal
The phase gate therefore propagates a phase flip to a encoded implementations. Bit or phase errors (as rele-
phase flip and a bit flip to both a bit-flip and a phase vant) before a measurement then map to bit errors on
flip, as depicted in Fig. 21. the observed classical bit pattern.
The phase gate is correspondingly more sensitive to The reason the measurement is destructive is that af-
phase-flip noise because the effective phase-flip strength ter the measurement, the qubits are no longer in the
 18

codespace of the color code; the post-measured state is Fig. 22, has a rather high threshold equal to the code
not projected onto an X or Z eigenstate in the codespace. capacity even in the circuit-level noise model. However,
However, as pointed out by Steane [46], given the abil- it is lowered slightly by the fact the effective error rate
ity to prepare encoded |+i states, a circuit composed of is 23 p, as discussed in the analysis of the encoded CNOT
transversal CNOT and transversal destructive MX mea- gate. The other output enters a standard quantum error
surements can implement nondestructive MX measure- correction circuit, also subject to noise of strength 23 p.
ments transversally. A similar story holds for encoded Since the lowest threshold of these two thresholds is this
|0i states and MZ measurements. The circuits for gen- one, the overall threshold for an encoded nondestructive
erating these nondestructive measurements transversally measurement is the same as the threshold for the encoded
are depicted in Fig. 23. Because the encoded |0i and CNOT gate. Namely, we have the result that
|+i states are being used to enable gates, namely nonde-
structive encoded measurements, these states are called (MX ,nondestructive) (CNOT ) 2 (I)
pth = pth = p , (66)
“magic states” for the gates [61]. Ordinarily, quantum er- 3 th
ror correction would follow not just one, but both of the (MZ ,nondestructive) (CNOT ) 2 (I)
pth = pth = pth . (67)
outputs of the encoded CNOT gate in these circuits, but 3
because one of the encoded qubits is destructively mea-
sured immediately after the CNOT gate, that encoded
6. |0i and |+i preparation
qubit does not require quantum error correction; it will
be effectively performed by the classical error correction
process occurring after the destructive measurement. It is tempting to assert that the way to fault-tolerantly


prepare the encoded |0i state is to perform an encoded
|ψi QEC MX |ψi nondestructive MZ measurement. The flaw with this
reasoning is that the nondestructive MZ measurement
|0i • MX CEC requires the encoded |+i state as a magic state, and the
analogous way of preparing a |+i state requires a |0i
|ψi • QEC MZ |ψi state.


To get out of this chicken-and-egg cycle, one must use
|+i MZ CEC an independent process. We describe a two-step process
that works for preparation of an encoded |0i state; the
process for preparing an encoded |+i state is similar.
FIG. 23: Circuits for nondestructive encoded MX and MZ ,
using the states |0i and |+i as “magic states.” The first step is to prepare the product state |0i⊗n by
transversally measuring MZ on each physical qubit. This
state is a stabilizer state, having n check operators, with
The threshold for destructive MZ and MX measure- check operator i being Z on qubit i for i = 1, . . . n. The
ments is the same as the code capacity threshold for the second step is to fault-tolerantly measure the X checks
code, regardless of which noise model we are considering. for the color code. Because the only Z-type operators
This is because the physical measurements are made only consistent with all the X checks are the color codes’ Z
once, as repetition cannot improve their effective error checks for the color code and the logical Z operator, these
rate. The (flawless) classical error correction performed measurements will transform the state into the logical |0i
in post-processing has a threshold equal to the code ca- state.
pacity threshold. Hence, we have the result that It turns out that it is not necessary to also fault-
(MX ,destructive) (I,code capacity) tolerantly measure the Z checks for the color code. The
pth = pth , (64)
state is already in an eigenstate of these operators at
(M ,destructive) (I,code capacity)
pth Z = pth . (65) this point, so all the measurements can do is yield syn-
drome bits. Had one obtained these bits and processed
Although these measurements need only be smaller them, the post-corrected state would still have been sub-
than the code capacity threshold to implement the en- ject to X errors drawn from the same distribution as the
coded measurement, when these measurements are used X errors afflicting the initial |0i⊗n preparation—fault-
in the fault-tolerant quantum error correction protocol, tolerant error correction doesn’t suppress the final error
they must be smaller than the threshold set by the pre- rate to zero, it only keeps it at the same rate one started
vailing noise model—a threshold that may be signifi- with.
cantly lower. The threshold for preparation of encoded |0i and |+i
To compute the threshold for nondestructive MZ and states is therefore the same as the threshold for fault-
MX measurements, we examine how errors propagate tolerant quantum error correction, namely,
through the circuits in Fig. 23. As with the analysis of
Fig. 19, the effective noise channel we need to consider af- (|0i) (|+i) (I)
pth = pth = pth . (68)
ter the CNOT gate is the BP channel followed by the DP
channel on each output. One of these enters a destruc- It is worth noting that while the process for fault-
tive measurement, which, as we found in the analysis of tolerantly preparing |0i and |+i states is not strictly
 19

transversal, the only nontransversal operation is fault- then “distilled” using encoded gates until the resultant
tolerant quantum error correction, a process that is re- |π/4i states have an error below the accuracy threshold.
quired in addition to transversal operations in any event In the second, high-quality |π/4i states are first distilled
in order to achieve fault-tolerant quantum computation. and then injected into the code. The circuit depicted in


_ _ _ _ _ _ _ _ _ _ 
 |M i • MX 

•  Unencode%"#$  MZ 
7. T gate
 
0 / H •


Another gate that admits a transversal implementation _ _ _ _ _ _ _ _ _ _
with a magic state is the T gate, also called the π/8 gate, 0 / X Z M
defined as
" # " # FIG. 25: Circuit for injecting a single-qubit magic state M .
iπ/8 The circuit for multi-qubit magic states is similar.
1 0 −iπ/8 e 0
T := =e . (69)
0 e−iπ/4 0 e−iπ/8
Fig. 25 is not fault-tolerant, but faults are already sup-
If we we have an encoded version of the state pressed by the code on the encoded qubits; only opera-
tions from the latter-half of the decoding circuit onwards
|π/4i := T H|0i (70) are unprotected.
1   Unlike all of the previous encoded gates, this method
= √ |0i + eiπ/4 |1i , (71) for implementing an encoded |π/4i preparation requires
2
an operation which is neither transversal nor fault-
also called |Ai and |Aπ/4 i in the literature, we can im- tolerant quantum error correction. The “unencoding”
plement the T gate transversally using the circuit of portion of the circuit is the time-reversed coherent cir-
Fig. 24. This circuit is not a Clifford circuit, because the cuit for encoding a state in the color code, derivable via
classically-controlled S gate is not a Clifford gate. Never- standard stabilizer codes as shown in Ref. [50]. This
theless, it only uses gates that we have previously shown unencoding circuit does not appear to have a transver-
how to implement in encoded form by purely transversal sal implementation. While the Eastin-Knill theorem [62]
operations. asserts that at least one nontransversal operation is re-
quired to generate a universal set of encoded gates, it
|ψi • QEC S QEC does not guarantee that no transversal implementation


of this circuit exists. That is because the process of fault-
|π/4i MZ CEC • tolerant quantum error correction used to prepare |0i and
|+i states is not transversal. For 3D color codes [63],
FIG. 24: Magic-state circuit for the T gate. in which T is intrinsically transversal and in which en-
coded |0i and |+i states still require fault-tolerant quan-
To compute the T gate threshold, we again study error tum error correction for preparation, only transversal and
propagation through its defining circuit, viz.the circuit in FTQEC operations are needed, for example. It would be
Fig. 24. As shown previously, the CNOT gate creates an interesting to develop a variant of the circuit in Fig. 25
input to the first QEC cycle that has a threshold of 2/3 which only uses transversal operations and possibly fault-
of the standard QEC threshold. The S gate creates an tolerant quantum error correction to inject a |π/4i state
input to the second QEC cycle which splits the thresh- into 2D color codes. We leave that for others to explore.
old into bit-flip and phase-flip thresholds approximately While the portion of the circuit in Fig. 25 in which the
equal to 1/2 and 1/3 of the standard QEC threshold. physical |M i state interacts with the unencoded qubit
The threshold for the T gate is set by the smallest of via a CNOT appears to also not be transversal, it can
these, namely the S gate threshold, which is be made so with slight modification. In principle, one
could prepare n states of the form |M i and transversally
1 1 1 (I)
q
(T,bit-flip)
pth = −
(I)
1 − 2pth ≈ pth , (72) apply the CNOT gate between these and the code block,
2 2 2 but only the one qubit corresponding to the unencoded
(T,phase-flip) 1 (I) state will be used to classically control the X and Z gates
pth = x ≈ pth . (73)
3 that are used to inject the correct state. As usual, these
corrections do not need to be actually implemented, only
used to update the Pauli frame.
8. |π/4i preparation Both alternatives for preparing high-quality encoded
|π/4i states require a procedure for magic-state distil-
There are two alternatives for preparing encoded |π/4i lation. One option is to use the encoding circuit for
states fault-tolerantly described in the literature. In the the 15-qubit Reed-Muller code [6] (also the smallest 3D
first, low-fidelity |π/4i states are “injected” into the code color code [63]) run in reverse, as depicted in Fig. 26.
by teleportation, using the circuit in Fig. 25 [10], and For it to work, the initial states must have an error less
 20

than the |π/4i distillation threshold. For the circuit de- used. The first is fault-tolerant quantum error correc-
picted in Fig. 26, the distillation threshold for indepen- tion, a process that is required in addition to encoded
dent, identically
√ distributed (iid) depolarizing noise is computations in any event for the entire protocol to
(6 − 2 2)/7 ≈ 45.3% [64, 65], for dephasing iid noise be fault tolerant. The second is the time-reversed
√ √
is ( 2 −√ 1)/ 2 ≈√29.3% [64, 66], and for worst-case iid coherent encoding circuit for color codes. Such a circuit
noise is ( 2−1)/2 2 ≈ 14.6% [64, 66]. The entire circuit is useful for encoding unknown quantum states, but
must be run O(poly(ε−1 )) times to achieve an output in an actual quantum computation, the input state is
error less than ε; convergence should be quite rapid in known so it is not needed for this purpose. Whether
practice given the actual polynomial [64]. Various tricks this “unencoding circuit” can be replaced with another
can be used to boost the distillation threshold and reduce operation which uses only transversal operations and
the resources required to achieve high-fidelity states; any fault-tolerant quantum error correction is an interesting
of these can be readily adapted to this setting. open question. For 3D color codes, we know that the
E answer is “yes.”
π/4
g • MX The “pancake architecture,” described in Ref. [16] for
E the Kitaev surface-codes, realizes the encoded gate set
we described using only gates between spatially neigh-
 
π/4
g • MX
E boring qubits. One difference in our analysis from that
π/4
g • |π/4i performed in Ref. [16] is that we have analyzed the accu-
racy threshold not only for fault-tolerant quantum mem-
E
π/4 • MX

  

g
E
ory but also for fault-tolerant quantum computation, a
π/4 MZ feat made tractable by the strong CSS nature of the color

  

g
E codes.
π/4 MZ

  

g
E
π/4
g MZ B. Fault-tolerance by code deformation
E
π/4 • MX

  

g
The method of fault-tolerance described in Sec. VI A
E requires a three-dimensional architecture to allow the
π/4 MZ

  

g
transversal CNOT gates to remain spatially local. This
violates the spirit of using two-dimensional codes in the
E
π/4 MZ

  

g
first place. Fortunately, it is possible to use code de-
formation to achieve fault-tolerance in a strictly two-
E
π/4 MZ

  

g
E dimensional architecture. Our construction here mirrors
π/4 MZ that of Raussendorf et al.’s construction for surface codes

  

g
E [17, 67]. Fowler has independently constructed a method
π/4 MZ for using code deformation in 4.8.8 color codes that is

  

g
E similar to ours [49]. Some salient differences between our
π/4 MZ method and Fowler’s are that (i) Fowler’s logical qubits
    
g
E are always encoded in a triple of defects whereas ours are
π/4
g MZ encoded in single defects except during certain logical
gates, and (ii) Fowler’s scheme disallows different defect
FIG. 26: Distillation circuit for |π/4i states; it is the 15-qubit types from occupying the same plaquette location while
Reed-Muller code’s encoding circuit in reverse. ours does not. Each of these differences allows our scheme
to encode a higher density of information. Specifically,
our scheme allows a six-fold increase in logical qubit den-
sity over the Fowler scheme.
9. Synthesis To begin, we generate a sufficiently large 4.8.8 trian-
gular color code by performing fault-tolerant quantum
It is well-known result the gate basis error correction on a collection of qubits. We are not
{H, S, CNOT , MX , MZ , |0i, |+i, |π/4i} is universal interested in what state the triangular code encodes—
for quantum computation [36] (in fact, it is even over- all we require is that the state is in the codespace with
complete). We have presented transversal methods for arbitrarily high fidelity. We consider any logical qubits
performing color-code encoded versions of each of these associated with the entire surface to be “gauge” qubits
except for the state preparations. By the Eastin-Knill in the language of subsystem stabilizer code theory [68].
theorem [62], it is impossible to generate a complete uni- We will use this state as a substrate for generating and
versal encoded gate basis in transversal form. However, manipulating encoded qubits.
color codes offer a particularly gentle way around this Each element of the standard set of stabilizer genera-
theorem. There are only two nontransversal operations tors for a color code can be labeled by a face of a definite
 21

color (red, green, or blue) and an operator of a definite introduces a number of “gauge” qubits in the interior of
Pauli type (X or Z). Notationally, we will refer to a gen- the defect that can be ignored; the details of this are
erator as a (c, P ) generator if it is of color c and Pauli type described in the next section.
P . To prepare an encoded qubit in our color code sub-
strate, we remove a connected product of stabilizer gen-
erators of the same color and type. (Generally removal 2. Growing, shrinking, and moving defects
of any element of the stabilizer group will yield a logical
qubit; we restrict attention to this class for simplicity.) We grow a (c, P ) defect qubit on region q in the fol-
We call this removed region a defect in analogy with the lowing way. Suppose we would like to extend the defect
language used by Raussendorf et al. in Ref. [17]. This so that it includes an adjacent region q 0 of the same color
removal is entirely passive—we simply cease measuring and type. (By adjacent, we mean that the regions can
this product of stabilizer generators in future quantum be connected by a single two-qubit c-colored link.) To
error correction rounds. For this reason, it is manifestly do this, we first perform the following conditional opera-
a fault-tolerant process. tion. If P = X, then we measure ZZ on a c-colored link
In the following sections, we describe how to perform connecting the regions, while if P = Z, then we measure
a universal repertoire of encoded logic gates on defect- XX on a c-colored link connecting the regions. Exam-
based logical qubits, with arbitrarily high fidelity. This is ples of how this works for octagonal and square defects
therefore a prescription for fault-tolerant quantum com- are depicted in Figs. 27 and 28; the circuit in Fig. 29
putation using code deformation. implements this transformation. A Y Y operator can be
used to grow a X and Z-type defect at the same time.

1. Preparing a defect in |0i or |+i

(-1)a
In principle, the generator removed to form a defect
qubit can be identified with any element of the encoded XX
Pauli group for that encoded qubit. For concreteness, (-1)a
we make the choice of calling the removed generator a
logical Z when it is Z-type defect (also called a ‘primal’ or (a) Octagonal green Z defect. (b) Growth to two defects.
‘smooth’ defect in the language of Ref. [17]) and a logical
X when it is X-type defect (also called a ‘dual’ or ‘rough’
FIG. 27: Growth of an octagonal green Z defect by one site.
defect in the language of Ref. [17]). Thus removing a c-
colored X- or Z-type generator corresponds to preparing
a logical |+i(c,X) or |0i(c,Z) state respectively.
The logical Z operator for a (c, X) defect acts as Z on a a
c-colored chain of qubits connecting the defect to another 1)
(-
c-colored boundary, which may itself be another defect.
XX
If no such other boundary exists, then the defect fails a
1)
to encode a logical qubit. To avoid this complication, (-
we have chosen our substrate to be a triangular code,
having boundaries of each of the three colors. Similarly, (a) Square red Z defect. (b) Growth to two defects.
the logical X operator for a (c, Z) defect acts as X on
a c-colored chain of qubits connecting the defect to a c-
FIG. 28: Growth of a square red Z defect by one site.
colored boundary.
Preparing a |+i(c,Z) or |0i(c,X) state requires more
care. To do this, we measure MX or MZ respectively


|0i • • MZ
along a c-colored chain of qubits from the plaquette we


wish to store the logical qubit in and the nearest c-colored |ψi
boundary. This projects each qubit along the chain into
|ψi
either |+i or |−i (resp. |0i or |1i), which we can interpret
as |+i (resp. |0i) for each qubit by changing local Pauli FIG. 29: Measurement of XX to grow a Z-type defect. The
bases. We then measure the Z-checks (resp. X-checks) measurement can be performed with existing circuitry already
incident on this chain except the one at the defect loca- in place for syndrome extraction.
tion and correct any errors, which places the defect back
into the substrate in the desired state. After this measurement, the new collective defect op-
An arbitrarily large c-colored defect can be prepared erator is the product of the q and q 0 defect operators.
in a single step by ceasing to measure a collection of The ±XX or ±ZZ operator has also been added to the
c-connected defects by a similar process, enabling the list of stabilizer generators. As usual, we do not need to
preparation process to be made arbitrarily reliable. This actually correct the result to a +1 outcome: it suffices
 22

to update the Pauli frames of the stabilizer generators tolerance to errors of one Pauli type will be significantly
incident on these two interior qubits. lower, but this will not be of the type that disturbs the
Because we will no longer use the weight-two opera- measurement.
tor, we may consider it to also be a “gauge” operator To nondestructively measure a defect, one uses the cir-
in the language of subsystem stabilizer codes [68]. This cuit of Fig. 23, which uses destructive measurement of
also makes its anticommuting partner a gauge operator, MZ or MX , preparation of |0i or |+i, and the CNOT
which we may interpret to be either of the original de- gate described in the next section.
fect operators (on q or q 0 ). By introducing these two
new gauge operators, we may reinterpret the defect log-
ical operator on the collective q and q 0 region as acting 4. CNOT gate between defects
solely on its boundary. In particular, the interior of the
collective q and q 0 region need never be involved in future
syndrome extractions. It is straightforward to show that moving a (c, Z) de-
An important question is whether the defect growth fect qubit around a (c0 , X) defect qubit (or vice-versa)
process is fault-tolerant. The simplest circuit for mea- generates an encoded CNOT gate controlled by the (c, Z)
suring XX or ZZ would perform CNOT gates into or defect when c and c0 are different colors; the construc-
out of an ancilla qubit to each of the two relevant qubits, tion is essentially the same as that in Refs. [17, 49, 67].
as depicted in Fig. 29. Although a single error in this Since this process traces out a braid in spacetime, we
ancilla qubit could propagate to two errors on the two in- call this process “braiding defects.” Also drawing upon
terior qubits, because we subsequently treat these qubits Refs. [17, 67], one can generate a CNOT gate between
as encoding a gauge qubit, we do not worry about errors two Z-type defects or two X-type defects, whether they
on these. It could still be the case that the value of the are the same color or not. The circuit for doing this be-
measurement obtained is incorrect, which impacts the tween two Z-type defects is depicted in Fig. 30; the cir-
update of the Pauli frame of the two adjacent stabilizer cuit for doing this between two X-type defects is similar.
generators in a correlated way. Thus a single syndrome

  

measurement error would propagate to two syndrome-bit |controli(c,Z) • |controli(c,Z)
errors. To prevent this happening to first order in the er- |0i(c00 ,X) MZ
ror probability, we repeat the XX or ZZ measurement
twice and use the majority vote of the three outcomes to |+i(c0 ,Z) • |targeti(c0 ,Z)
update the Pauli frame. |targeti(c0 ,Z) • MX
Compared to the process of defect growth, defect con-
traction is much simpler: to shrink a defect by a single- FIG. 30: Circuit for braiding a CNOT gate between Z-type
plaquette, one simply measures that plaquette operator defects. The colors c and c0 may be the same or different,
in the next round of fault-tolerant quantum error correc- but the color c00 is a color different from these. The circuit
tion. for braiding a CNOT gate between X-type defects is simi-
lar: the CNOT gate directions are reversed, the types of the
By a combination of local growth and shrinking pro-
defects and the types of the measurements have their Pauli
cesses, one can deform the code with a (c, P ) defect at types swapped from X to Z and vice-versa, and the |0i state
one plaquette to a code with a (c, P ) defect anywhere becomes a |+i state and vice-versa.
else. In other words, the move operation for a defect can
be decomposed into a sequence of more elementary grow
and shrink operations. One can convert an X-type defect into a Z-type defect,
or vice-versa, (changing its color as a side effect) using
one of the circuits in Fig. 31. In conjunction with the
3. Measuring a defect other type of CNOT gates mentioned, this allows CNOT
gates between two defects regardless of the colors or Pauli
types they have.
To destructively measure the logical operator encir-

cling a defect, one first shrinks the defect to size of a
single plaquette. Then one measures the defect with the |ψi(c,X) MZ
existing circuitry at that plaquette as though it were a lo- |+i(c0 ,Z) • |ψi(c0 ,Z)
cal stabilizer generator. The shrunken defect will have a
significantly lower tolerance to one type of Pauli error but
that error type is in the basis being measured in and will |ψi(c,Z)

• MX
not disturb the measurement outcome. To destructively
measure the string-like logical operator connecting two |0i(c0 ,X) |ψi(c0 ,X)
defects, one brings the two operators as close together
as possible. One then measures the weight-two operator FIG. 31: Circuits for converting a Z-type defect into an X-
type defect and vice-versa.
connecting the defects using the circuitry used to grow a
defect from one site to encompass the other. Again, the
 23

5. Phase gate on a defect If we want to perform a Hadamard gate on a (c, P )

defect, we first prepare an ancilliary (c, P ) defect in the
To perform an S (phase) gate on a (c, P ) defect, we state |0i and perform a CNOT gate from the defect qubit
prepare two more qubits of (c0 , P ) and (c00 , P ) type, each to this defect ancilla using the circuit of Fig. 30. This en-
in the state |0i and use CNOT gates to put the defect codes the original defect qubit into the two-qubit bit-flip
into a three-defect repetition code. This maps our single- repetition code across the two defects. The ZZ opera-
defect logical qubits into the three-defect logical qubits tor for the two defect qubits is in the stabilizer group
Fowler uses in his construction [49]. We then grow the of this repetition code, so we can measure ZZ without
defects and connect them so that they separate an in- disturbing the encoded qubit. The logical Z operator is
terior triangular region from an exterior region, just as a c0 -colored chain of Z operators around either of the de-
described in Fowler’s construction. If P = Z, then as fects and the logical X operator is a c-colored chain of
Fowler noted, it suffices to apply S transversally (actu- X operators connecting the defects, where c0 6= c. This
ally, S † must be applied transversally) to generate a logi- encoding is the one used at all times in the Raussendorf
cal S on the triple-defect qubit. However, if P = X, then et al. scheme [17, 67], but here we only use it to perform
Fowler’s construction fails, because the “exterior trees” the Hadamard gate and go back to our original single-
in his language fail to undergo the action SXS † = Y . defect encoding once the gate is completed.
“Pruning” the exterior tree as Fowler suggests for his im- After we’ve encoded the defect qubit into two, we then
plementation of the Hadamard gate fails as well, because perform individual MZ measurements on a c0 -colored
such an operation yields only the “byproduct operator” chain of qubits surrounding both defects, where c0 6= c.
for logical X or Z on the triple-defect qubit, but not both. This separates the region of the two qubits from the sub-
To perform the S gate on X-type defects, we propose the strate, so we can then apply H transversally on the cut-
following two-step procedure. First, we arrange the de- out region without influencing the external substrate.
fects to separate a triangular interior from the exterior This operation applies a logical Hadamard gate to the
and apply S † transversally on the interior. Second, we re- two defect qubits in the interior, but also turns them
arrange the defects so that part of what was the exterior into P 0 -type defect qubits in the process, where P 0 is
becomes the new interior, and perform S † transversally conjugate to P (i.e., P 0 = Z if P = X, and P 0 = X if
on this new triangular interior. In this way, both the in- P = Z). We then stitch the cut out region back into the
terior and exterior trees experience the S gate. Once the code by measuring the P -check operators incident on the
gate is complete, we run the three-defect encoding circuit cut. The encoding circuit for the repetition code is run
in reverse and absorb the two ancilla qubit regions back in reverse, and the resulting defect can be converted back
into the substrate in subsequent quantum error correc- to its original type and color using circuits of the form
tion rounds. depicted in Fig. 31.
Of course, the S gate can also be achieved
via magic
states of the form |π/2i := √12 |0i + eiπ/2 |1i (also called
|Y i or |+ii in the literature) in a manner similar to what 7. Injecting |π/4i into a defect
is done for surface codes. But this is one of the great
benefits of 4.8.8 color codes—no magic state distillation To perform universal encoded quantum computation
and usage is required to realize this gate in encoded form. with defects, our approach requires defects encoded into
It may well be worth the lower accuracy threshold of the state |π/4i with an error below its distillation thresh-
color codes relative to surface codes in order to achieve old, as discussed in the previous “pancake” architecture.
the resource reduction for performing encoded S gates. We therefore need a method for injecting magic states
into defects such that the injection process introduces
errors at a rate below the distillation threshold. The
6. Hadamard gate on a defect single-qubit preparation threshold for a magic state is
therefore the difference between its distillation threshold
From one point of view, a logical Hadamard gate is un- and the error introduced by its injection process.
necessary because it can be implemented using the gates It is worth remarking that this kind of injection pro-
we have previously described, for example by the circuit cess is used in defect-based surface code schemes as well
of Fig. 32. However, we have developed a more resource- [17, 49, 69]. In these schemes, one must not only in-


ject |π/4i states, but also inject |π/2i states as well.
|ψi S X S H |ψi However, the impact of errors introduced by errant in-
jection has not been studied to our knowledge. It is un-
|+i • S MX clear whether considering it will significantly alter the
high threshold values numerically estimated for surface
FIG. 32: Circuit for simulating H with previously-described codes—the difference between a 1% accuracy threshold
gates. and a 14% distillation threshold is not that great, so it
is reasonable to expect that it may be quite important,
efficient way to perform this gate that we describe here. especially because injection generates small-sized defects
 24

that are not arbitrarily well-protected from noise at first. codes in the code-capacity noise model is 10.56(1)%. This
We do not investigate the impact of the injection process is not significantly different from what had previously
on the threshold for color codes here either, but we ex- been estimated for optimal decoding of these and the
pect that it will be less consequential because the value 6.6.6 color codes, or most-likely-error or optimal decod-
of the color-code accuracy threshold is much lower than ing of Kitaev’s 4.4.4.4 surface codes. Indeed, the upper
that for the surface codes. bound for any CSS code is slightly more than 11%, so
To inject into a (c, Z) defect, we identify the corner of all of these codes perform close to optimally in this noise
the triangular substrate containing the c-colored plaque- model. To support our numerical estimate, we proved
tte and measure MZ on the qubit in the corner, isolating that the threshold is at least 8.335 745 (1)% using a self-
it from the code. We then apply T H to the corner qubit avoiding walk technique.
and then measure the weight-four X check in the corner, Our numerically-estimated value for the accuracy
bringing the corner qubit back into the code. We then threshold of most-likely-error fault-tolerant quantum er-
cease measuring the Z check in the standard way, creat- ror correction for 4.8.8 color codes in the phenomeno-
ing a single-plaquette Z-type defect in the corner. This logical noise model is 3.05(4)%. Again, this is not sig-
defect is not well protected from noise, so we move it nificantly different from what had previously been esti-
from the corner and grow it as fast as we can, so that the mated for optimal decoding of the 6.6.6 color codes, or
ambient noise doesn’t degrade the fidelity of the encoded most-likely-error or optimal decoding of Kitaev’s 4.4.4.4
state. surface codes. We attribute the nominal improvement we
find relative to Kitaev’s surface codes for both this and
the previous noise model to the fact that the color codes
8. T gate on a defect have higher-weight stabilizer generators, which should
be modeled as more errant, but which aren’t in these
Given |π/4i defect qubits, we can distill them and use noise models. To support our numerical estimate, we
them to perform the T gate in the same way as described proved that the threshold is at least 0.3096% using a
in Secs. VI A 7 and VI A 8 for the “pancake” architecture. self-avoiding walk technique.
Our numerically-estimated value for most-likely-error
fault-tolerant quantum error correction for 4.8.8 color
VII. CONCLUSIONS codes in the circuit-based noise model is 0.082(3)%. By
attempting to optimize the syndrome extraction circuit
A. Fault-tolerant quantum computation by hand, we ended up surprisingly decreasing our thresh-
old estimate to 0.080(3)%, suggesting that optimizing the
We studied fault-tolerant quantum computation using syndrome extraction circuit to find the highest threshold
color codes, inspired by (a) the need to minimize qubit is a nontrivial task. Unlike our findings for the previ-
transport in real technologies having 2D layouts and (b) ous two noise models, our accuracy-threshold estimate is
the high accuracy thresholds reported for similar topo- in fact significantly different from what had previously
logical codes. We framed our study with a well-defined been estimated for most-likely-error decoding of Kitaev’s
quantum control model and three physically-motivated 4.4.4.4 surface codes—it is nearly a tenth the comparable
noise models of increasing realism which we call the code- value of 0.68%. That said, it is consistent with the value
capacity noise model, the phenomenological noise model, of “about 0.1%” estimated using a different suboptimal
and the circuit-based noise model. decoder for these codes considered in Ref. [22]. However,
The strategy behind our study was to first understand the estimate in Ref. [22] lacked any error analysis, so it is
how to fault-tolerantly simulate the identity gate via hard to determine how consistent these results truly are.
fault-tolerant quantum error correction and then extend We believe that the reduction in threshold relative to the
this understanding to how to fault-tolerantly simulate a surface code threshold comes from the increased weight
universal set of quantum gates capable of general-purpose of the stabilizer generators for the 4.8.8 color code. Based
quantum computation. on this, we predict that the 6.6.6 color codes will have
In the course of studying fault-tolerant quantum er- a quantum error-correction accuracy threshold for this
ror correction, we formulated most-likely-error decoding noise model somewhere between 0.082(3)% and 0.68%
for color codes as a mathematical optimization problem without any additional optimizations. We did not prove
known as an integer program. We also developed fea- a lower bound on the threshold in this noise model, as the
sible schedules for parallelized syndrome extraction for self-avoiding walk technique breaks down for this noise
the most efficient family of color codes, the 4.8.8 color model.
codes. To better understand the performance of our To extend our results to general-purpose fault-tolerant
decoder, we elaborated a previously-established connec- quantum computing, we considered two different ap-
tion between the performance of our decoder and some proaches. In the first, the architecture consisted of 2D
statistical-mechanical classical spin models. surfaces stacked like pancakes in which each surface cor-
Our numerically-estimated value for most-likely-error responded to a logical qubit and almost all operations
fault-tolerant quantum error correction for 4.8.8 color were either global transversal operations or local syn-
 25

drome extraction operations. In the second, the archi- an “energy-minimizing” decoder in this paradigm, cor-
tecture consisted of an extended 2D surface in which log- responding to the phase boundary at zero temperature.
ical qubits were associated with “defects” and almost all Because our code-capacity value of 10.56(1)% is lower
operations were either defect braiding by local measure- than the code capacity of 10.925(5)% of a “free-energy-
ments or local syndrome extraction operations. minimizing” decoder implicitly explored by Ohzeki [25],
In the “pancake” architecture, we showed that en- this demonstrates that the phase boundary of the 3BR-
coded universal quantum computation was possible us- BIM is “re-entrant” as depicted in Fig. 11, violating the
ing only local stabilizer measurements, global transversal so-called Nishimori conjecture for this system. This re-
operations, and the time-reversed coherent encoding cir- sult is counterintuitive because it states that the 3BR-
cuit for the color code, which was used to inject magic BIM can become more ordered by increasing the tem-
states. Each gate in this architecture has its own accu- perature, depending on the system’s quenched disorder
racy threshold that is a significant fraction of the quan- parameter. It would be exciting to see experimental con-
tum error correction (memory) threshold. firmation of this effect.
In the “defect” architecture, we showed that encoded
universal quantum computation was possible using only
local stabilizer measurements, code deformation, and C. Future directions
transversal operations on isolated regions. These de-
formations came in different forms, including growing While we have been able to answer many questions
small defects into large ones, braiding defects around about fault-tolerant quantum computing using color
each other for encoded CNOT gates, and isolating defects codes, practicalities have necessarily limited the focus of
from the rest of the code. Each gate has the same accu- our analysis, leaving other related questions open. Our
racy threshold as the quantum error correction (memory) results also also raise new questions that we believe are
threshold, although errors afflicting injected magic state worthy of study.
defects before they are grown to full size may dominate One future direction we mentioned is optimizing the
the threshold for the less realistic code-capacity and phe- syndrome extraction circuit. One could also examine us-
nomenological noise models. ing more elaborate ancilla states in the circuit, such as
Because the defect architecture has a higher threshold those used in the schemes proposed by Shor [1], Steane
and is more consistent with the original motivation for [46], and Knill [10]. In any scheme one chooses, further
our study—namely that many technologies are restricted improvement may still be possible by transforming the
to a single 2D layout—we believe the defect-based ap- circuit used in an implementation.
proach to be the most practical. To that end, we ex- Another future direction we alluded to is optimizing
tended some of the defect-based approach for color codes the decoding algorithm. One could examine the per-
presented in Ref. [49] so that a significantly higher den- formance of the truly optimal decoder for the circuit
sity of defects can be stored and processed in the surface. model which accounts for the correlations in the noise
induced by the syndrome extraction circuit. This will
yield an upper bound on the accuracy threshold for the
B. Relation to statistical-mechanical phase noise model(s) studied. On the other end of the spec-
transitions trum, it would be useful to explore the performance of
faster decoders which don’t yield as high a threshold as
It has been previously established that there is a the MLE decoder but which may be more valuable in
mapping between quantum color codes and a classical practice. The renormalization group decoder [70] and
statistical-mechanical model known as the three-body minimum-weight perfect matching decoder [16] (using a
random-bond Ising model (3BRBIM). In this mapping, mapping of one color code to two Kitaev surface codes
each check maps to a classical ±1 spin and each qubit [24]) are examples of this. Another alternative is to gen-
maps to a three-body interaction, with the interaction eralize the results of Feldman et al., who developed an
being ferromagnetic if the qubit is not in error and anti- efficient linear-program decoder for binary codes based
ferromagnetic if it is. Specifically, the Hamiltonian con- on an integer-program-based decoder similar to the one
structed by this mapping is we developed here [71].
X Y The lower bound technique of self-avoiding walks that
H= Jq Sc , (74) we used is certainly not the tightest, and it may be of
qubits q checks c3q interest to establish tighter lower bounds. For tighter
bounds, it may be possible to use different techniques.
where Jq ∈ ±1 indicates a flip on qubit q and Sc ∈ ±1 In the case of the circuit-based noise model, the self-
indicates the eigenvalue of the check c. avoiding walk bound technique breaks down dramati-
A feature of the mapping is that the code capacity for cally, and it would be worth exploring other lower-bound
any particular decoding algorithm represents a point on techniques in this setting.
the boundary of the order-disorder transition of the as- While we believe the noise and control model that we
sociated 3BRBIM. Our integer-programming decoder is studied is reasonable, it is certainly not unique and can
 26

be improved upon with more experimental input. As code for some topology, having an associated classical
shown by Levy et al., [34, 35], when more realistic models statistical-mechanical model for a given quantum noise
are included, conclusions regarding fault tolerance can model. It might be interesting to use the fault-tolerant
change dramatically. Even at an abstract level, one could decoding of CSS codes generally as a tool to explore re-
modify our depolarizing noise model for CNOT gates so lated statistical-mechanical systems with quenched dis-
that it acted ideally with probability 1−p and applied one order.
of the fifteen nontrivial Pauli operators with probability
p/15 each rather than acting ideally with probability 1 −
p and applying one of the sixteen Pauli operators with
probability p/16.
Acknowledgments
While we gave a prescription for injecting magic states
into the color code for both the pancake and 2D defect-
based architectures, we did not carefully study the We would like to thank the following individuals for
threshold of the circuits used for injection. To our knowl- helpful discussions: Hector Bombin, Bob Carr, Chris
edge, this type of study has not been performed for Ki- Cesare, Guillaume Duclos-Cianci, Bryan Eastin, Austin
taev’s surface codes either. Such studies would be valu- Fowler, Anand Ganti, Peter Groszkowski, Jim Harring-
able, as it could be the case that the magic state prepara- ton, Charles Hill, Lloyd Hollenberg, Uzoma Onunkwo,
tion threshold is actually less than the accuracy threshold Cindy Phillips, David Poulin, Robert Raussendorf, and
reported for all of the other gates, even though the dis- David Wang. We would also like to thank Dave Gay
tillation threshold for the magic states is higher than the for use of AMPL mathematical programming language.
accuracy threshold for the other gates. PRR was supported by the Quantum Institute at Los
Finally, the connection between color codes and the Alamos National Laboratories. AJL and JTA were
three-body random-bond Ising model allowed us to ex- supported in part by the National Science Foundation
plore the structure of order-disorder transition in the lat- through Grant 0829944. JTA and AJL were supported
ter model by studying the former. This is one of the in part by the Laboratory Directed Research and Devel-
rare examples where a purely quantum information the- opment program at Sandia National Laboratories. San-
oretic result has led to greater understanding of a clas- dia National Laboratories is a multi-program labora-
sical system. Kitaev’s surface codes and the two-body tory managed and operated by Sandia Corporation, a
random-bond Ising model have a similar connection and wholly owned subsidiary of Lockheed Martin Corpora-
have admitted a similar study [16, 27]. It is clear that it is tion, for the U.S. Department of Energy’s National Nu-
the CSS structure of these codes that admits these stud- clear Security Administration under contract DE-AC04-
ies; one could argue that every CSS code is a topological 94AL85000.

[1] P. W. Shor, Fault-tolerant quantum computation, in arXiv:quant-ph/9611027.

Proceedings of the 37th Annual Symposium on Foun- [6] E. Knill, R. Laflamme, and W. H. Zurek, Re-
dations of Computer Science, edited by R. S. Sip- silient quantum computation: Error models and
ple, IEEE (IEEE Press, Los Alamitos, CA, 14–16 thresholds, Proc. Roy. Soc. London A 454, 365
Oct. 1996, Burlington, VT, USA, 1996), pp. 56–65, (1998), arXiv:quant-ph/9702058, URL http://www.
ISBN 0-8186-7594-2, doi:10.1137/S0097539795293172, jstor.org/stable/53171.
arXiv:quant-ph/9605011. [7] J. Preskill, Fault-tolerant quantum computation, in
[2] D. Aharonov and M. Ben-Or, Fault tolerant quan- Introduction to Quantum Computation and Informa-
tum computation with constant error, in Proceedings tion, edited by H.-K. Lo, T. Spiller, and S. Popescu
of the Twenty-Ninth Annual ACM Symposium on the (World Scientific, Singapore / River Edge, NJ,
Theory of Computing, edited by F. T. Leighton and 1998), chap. 8, pp. 213–269, ISBN 9-810-24410-X,
P. Shor (ACM Press, New York, El Paso, TX, arXiv:quant-ph/9712048.
USA, 1997), pp. 176–188, ISBN 0-89791-888-6, See [8] J. Preskill, Reliable quantum computers, Proc. Roy. Soc.
also extended version [3]., doi:10.1145/258533.258579, London A 454, 385 (1998), arXiv:quant-ph/9705031,
arXiv:quant-ph/9611025. URL http://www.jstor.org/stable/53172.
[3] D. Aharonov and M. Ben-Or, Fault tolerant quantum [9] A. M. Stephens, A. G. Fowler, and L. C. L. Hollenberg,
computation with constant error rate (1999), See also Universal fault tolerant quantum computation on bilin-
condensed version [2]., arXiv:quant-ph/9906129. ear nearest neighbor arrays, Quant. Inf. Comp. 8, 0330
[4] A. Y. Kitaev, Quantum computations: algorithms and (2008), arXiv:quant-ph/0702201, URL http://www.
error correction, Russian Math. Surveys 52, 1191 (1997), rintonpress.com/xxqic8/qic-8-34/0330-0344.pdf.
doi:10.1070/RM1997v052n06ABEH002155. [10] E. Knill, Quantum computing with realisti-
[5] A. M. Steane, Active stabilization, quantum computa- cally noisy devices, Nature 434, 39 (2005),
tion, and quantum state synthesis, Phys. Rev. Lett. doi:10.1038/nature03350, arXiv:quant-ph/0410199.
78, 2252 (1997), doi:10.1103/PhysRevLett.78.2252, [11] E. Knill, Fault-tolerant postselected quantum computa-
 27

tion: Schemes (2004), arXiv:quant-ph/0402171. accuracy threshold for quantum memory, Ann. Phys.
[12] E. Knill, Fault-tolerant postselected quan- 303, 31 (2003), doi:10.1016/S0003-4916(02)00019-2,
tum computation: Threshold analysis (2004), arXiv:quant-ph/0207088.
arXiv:quant-ph/0404104. [28] M. Ohzeki, Locations of multicritical points for
[13] F. M. Spedalieri and V. P. Roychowdhury, Latency in spin glasses on regular lattices, Phys. Rev. E 79,
local, two-dimensional, fault-tolerant quantum comput- 021129 (2009), doi:10.1103/PhysRevE.79.021129,
ing, Quant. Inf. Comp. 9, 666 (2009), arXiv:0805.4213, arXiv:0811.0464.
URL http://www.rinton.net/xxqic9/qic-9-78/ [29] T. Ohno, G. Arakawa, I. Ichinose, and T. Mat-
0666-0682.pdf. sui, Phase structure of the random-plaquette
[14] D. Bacon, Operator quantum error-correcting subsys- z2 gauge model: accuracy threshold for a toric
tems for self-correcting quantum memories, Phys. Rev. A quantum memory, Nucl. Phys. B 697, 462
73, 012340 (2006), doi:10.1103/PhysRevA.73.012340, (2004), doi:10.1016/j.nuclphysb.2004.07.003,
arXiv:quant-ph/0506023v4. arXiv:quant-ph/0401101.
[15] A. Y. Kitaev, Quantum error correction with imperfect [30] S. L. A. de Queiroz, Location and properties of
gates, in Proceedings of the Third International Confer- the multicritical point in the Gaussian and ±j Ising
ence on Quantum Communication, Computing and Mea- spin glasses, Phys. Rev. B 79, 174408 (2009),
surement, edited by O. Hirota, A. S. Holevo, and C. M. doi:10.1103/PhysRevB.79.174408, arXiv:0902.4153.
Caves (Plenum Press, New York, 1997). [31] M. Suchara, S. Bravyi, and B. Terhal, Constructions
[16] E. Dennis, A. Kitaev, A. Landahl, and J. Preskill, Topo- and noise threshold of topological subsystem codes (2010),
logical quantum memory, J. Math. Phys. 43, 4452 (2002), arXiv:1012.0425.
doi:10.1063/1.1499754, arXiv:quant-ph/0110143. [32] H. Bombin, Topological subsystem codes, Phys. Rev. A
[17] R. Raussendorf and J. Harrington, Fault-tolerant 81, 032301 (2010), doi:10.1103/PhysRevA.81.032301,
quantum computation with high threshold in arXiv:0908.4246.
two dimensions, Phys. Rev. Lett. 98, 190504 [33] E. Berlekamp, R. J. McEliece, and H. van Tilborg,
(2007), doi:10.1103/PhysRevLett.98.190504, On the inherent intractability of certain coding prob-
arXiv:quant-ph/0610082. lems, IEEE Trans. Info. Theo. 24, 384 (1978),
[18] A. G. Fowler, A. M. Stephens, and P. Groszkowski, doi:10.1109/TIT.1978.1055873.
High-threshold universal quantum computation on the [34] J. E. Levy, A. Ganti, C. A. Phillips, B. R. Hamlet, A. J.
surface code, Phys. Rev. A 80, 052312 (2009), Landahl, T. M. Gurrieri, R. D. Carr, and M. S. Car-
doi:10.1103/PhysRevA.80.052312, arXiv:0803.0272. roll, Brief announcement: the impact of classical elec-
[19] D. S. Wang, A. G. Fowler, and L. C. L. Hol- tronics constraints on a solid-state logical qubit mem-
lenberg, Surface code quantum computing with error ory, in Proceedings of the Twenty-First Annual ACM
rates over 1%, Phys. Rev. A 83, 020302(R) (2011), Symposium on Parallelism in Algorithms and Architec-
doi:10.1103/PhysRevA.83.020302, arXiv:1009.3686. tures, edited by F. Meyer auf der Heide and M. A.
[20] H. Bombin and M. A. Martin-Delgado, Topolog- Bender (ACM Press, New York, 11–13 August, 2009,
ical quantum distillation, Phys. Rev. Lett. 97, Calgary, AB, Canada, 2009), ISBN 978-1-60558-606-9,
180501 (2006), doi:10.1103/PhysRevLett.97.180501, doi:10.1145/1583991.1584039, arXiv:0904.0003.
arXiv:quant-ph/0605138. [35] J. E. Levy, M. S. Carroll, A. Ganti, C. A. Phillips, A. J.
[21] List of uniform tilings, URL http://en.wikipedia.org/ Landahl, T. M. Gurrieri, R. D. Carr, H. L. Stalford,
wiki/List_of_uniform_tilings. and E. Nielsen, Implications of electronics constraints for
[22] D. S. Wang, A. G. Fowler, C. D. Hill, and L. C. L. Hol- solid-state quantum error correction and quantum circuit
lenberg, Graphical algorithms and threshold error rates failure probability (2011), arXiv:1105.0682.
for the 2d color code, Quant. Inf. Comp. 10, 780 (2010), [36] M. A. Nielsen and I. L. Chuang, Quantum Computation
arXiv:0907.1708, URL http://www.rintonpress.com/ and Quantum Information (Cambridge University Press,
xxqic10/qic-10-910/0780-0802.pdf. Cambridge, 2000), ISBN 0-521-63235-8 (Hardback), 0-
[23] H. G. Katzgraber, H. Bombin, and M. A. Martin- 521-63503-9 (Paperback).
Delgado, Error threshold for color codes and ran- [37] J. Preskill, Lecture notes for Caltech Ph 219: Quantum
dom three-body ising models, Phys. Rev. Lett. 103, Information and Computation (1998), URL http://www.
090501 (2009), doi:10.1103/PhysRevLett.103.090501, theory.caltech.edu/~preskill/ph219/.
arXiv:0902.4845. [38] P. Aliferis, D. Gottesman, and J. Preskill, Quantum accu-
[24] G. Duclos-Cianci, H. Bombı́n, and D. Poulin, Fast de- racy threshold for concatenated distance-3 codes, Quant.
coding algorithm for subspace and subsystem color codes Inf. Comp. 6, 97 (2006), arXiv:quant-ph/0504218, URL
and local equivalence of topological phases (2011), per- http://www.rinton.net/xqic6/qic-6-2/097-165.pdf.
sonal communication. [39] P. Aliferis and A. W. Cross, Subsystem fault toler-
[25] M. Ohzeki, Accuracy thresholds of topologi- ance with the Bacon-Shor code, Phys. Rev. Lett. 98,
cal color codes on the hexagonal and square- 220502 (2007), doi:10.1103/PhysRevLett.98.220502,
octagonal lattices, Phys. Rev. E 80, 011141 (2009), arXiv:quant-ph/0610063.
doi:10.1103/PhysRevE.80.011141, arXiv:0903.2102. [40] P. Aliferis, D. Gottesman, and J. Preskill, Accuracy
[26] R. S. Andrist, H. G. Katzgraber, H. Bombin, and threshold for postselected quantum computation, Quant.
M. A. Martin-Delgado, Tricolored lattice gauge theory Inf. Comp. 8, 0181 (2008), arXiv:quant-ph/0703264,
with randomness: Fault-tolerance in topological color URL http://www.rintonpress.com/xxqic8/qic-8-34/
codes (2010), arXiv:1005.0777. 0181-0244.pdf.
[27] C. Wang, J. Harrington, and J. Preskill, Confinement- [41] P. Aliferis and J. Preskill, Fibonacci scheme for
Higgs transition in a disordered gauge theory and the fault-tolerant quantum computation, Phys. Rev. A
 28

79, 012332 (2009), doi:10.1103/PhysRevA.79.012332, doi:10.1088/0305-4470/37/48/001.

arXiv:0809.5063. [58] S. E. Alm, Upper and lower bounds for the connective
[42] A. W. Cross, D. P. DiVincenzo, and B. M. Terhal, Com- constants of self-avoiding walks on the Archimedean and
parative code study for quantum fault tolerance, Quant. Laves lattices, J. Phys. A: Math. Gen. 38, 2055 (2005),
Inf. Comp. 9, 541 (2009), arXiv:0711.1556, URL http: doi:10.1088/0305-4470/38/10/001.
//www.rinton.net/xxqic9/qic-9-78/0541-0572.pdf. [59] B. Eastin, Error channels and the threshold for fault-
[43] P. Aliferis and J. Preskill, Fault-tolerant quantum tolerant quantum computation, Ph.D. thesis, University
computation against biased noise, Phys. Rev. A of New Mexico (2007), arXiv:0710.2560.
78, 052331 (2008), doi:10.1103/PhysRevA.78.052331, [60] E. Knill, R. Laflamme, and W. H. Zurek, Thresh-
arXiv:0710.1301. old accuracy for quantum computation (1996),
[44] D. Gottesman, The Heisenberg representation of quan- arXiv:quant-ph/9610011.
tum computers, in Group22: Proceedings of the XXII [61] S. Bravyi and A. Kitaev, Universal quantum computation
International Colloquium on Group Theoretical Meth- with ideal Clifford gates and noisy ancillas, Phys. Rev. A
ods in Physics, edited by S. P. Corney, R. Delbourgo, 71, 022316 (2005), doi:10.1103/PhysRevA.71.022316,
and P. D. Jarvis (International Press, Cambridge, MA, arXiv:quant-ph/0403025.
13–17 Jul. 1998, Hobart, Australia, 1999), pp. 32–43, [62] B. Eastin and E. Knill, Restrictions on transversal
arXiv:quant-ph/9807006. encoded quantum gate sets, Phys. Rev. Lett. 102,
[45] A. M. Steane, Error correcting codes in quan- 110502 (2009), doi:10.1103/PhysRevLett.102.110502,
tum theory, Phys. Rev. Lett. 77, 793 (1996), arXiv:0811.4262.
doi:10.1103/PhysRevLett.77.793. [63] H. Bombin and M. A. Martin-Delgado, Topological
[46] A. M. Steane, Space, time, parallelism and noise re- computation without braiding, Phys. Rev. Lett. 98,
quirements for reliable quantum computing, Fortschr. 160502 (2007), doi:10.1103/PhysRevLett.98.160502,
Phys. 46, 443 (1998), arXiv:quant-ph/9708021, URL arXiv:quant-ph/0610024.
http://onlinelibrary.wiley.com/doi/10.1002/ [64] B. W. Reichardt, Quantum universality by state distilla-
%28SICI%291521-3978%28199806%2946:4/5%3C443:: tion (2006), arXiv:quant-ph/0608085.
AID-PROP443%3E3.0.CO;2-8/abstract. [65] H. Buhrman, R. Cleve, M. Laurent, N. Linden, A. Schri-
[47] A. R. Calderbank and P. W. Shor, Good quan- jver, and F. Unger, New limits on fault-tolerant quan-
tum error-correcting codes exist, Phys. Rev. A tum computation, in Proceedings of the 47th IEEE Sym-
54, 1098 (1996), doi:10.1103/PhysRevA.54.1098, posium on Foundations of Computer Science, edited
arXiv:quant-ph/9512032. by B. Werner, IEEE (IEEE Press, Los Alamitos,
[48] A. Steane, Multiple particle interference and quantum CA, Oct. 21–24, Berkeley, CA, USA, 2006), pp. 411–
error correction, Proc. Roy. Soc. London A 452, 2551 419, ISBN 0-7695-2720-5, doi:10.1109/FOCS.2006.50,
(1996), arXiv:quant-ph/9601029, URL http://www. arXiv:quant-ph/0604141.
jstor.org/stable/52827. [66] S. Virmani, S. F. Huelga, and M. B. Plenio, Clas-
[49] A. G. Fowler, Two-dimensional color-code quan- sical simulability, entanglement breaking, and quan-
tum computation, Phys. Rev. A 83, 042310 tum computation thresholds, Phys. Rev. A 71,
(2011), doi:10.1103/PhysRevA.83.042310, 042328 (2005), doi:10.1103/PhysRevA.71.042328,
arXiv:0806.4827v3. arXiv:quant-ph/0408076.
[50] D. Gottesman, Stabilizer codes and quantum er- [67] R. Raussendorf, J. Harrington, and K. Goyal,
ror correction, Ph.D. thesis, Caltech (1997), Topological fault-tolerance in cluster state
arXiv:quant-ph/9705052. quantum computation, New J. Phys. 9, 199
[51] A. G. Fowler, D. S. Wang, and L. C. L. Hollenberg, Sur- (2007), doi:10.1088/1367-2630/9/6/199,
face code quantum error correction incorporating accu- arXiv:quant-ph/0703143.
rate error propagation, Quant. Inf. Comp. 11, 8 (2011), [68] D. Poulin, Stabilizer formalism for operator quan-
arXiv:1004.0255. tum error correction, Phys. Rev. Lett 95, 230504
[52] J. Harrington, personal communication about the accu- (2005), doi:10.1103/PhysRevLett.95.230504,
racy threshold analysis done in support of Ref. [17]. arXiv:quant-ph/0508131.
[53] H. Nishimori, Internal energy, specific-heat and [69] R. Raussendorf, J. Harrington, and K. Goyal, A
correlation-function of the bond-random Ising- fault-tolerant one-way quantum computer, Ann. Phys.
model, Prog. Theor. Phys. 66, 1169 (1981), 321, 2242 (2006), doi:10.1016/j.aop.2006.01.012,
doi:10.1143/PTP.66.1169. arXiv:quant-ph/0510135.
[54] H. Nishimori, Geometry-induced phase transition in the [70] G. Duclos-Cianci and D. Poulin, Fast decoders for
±j Ising model, J. Phys. Soc. Japan 55, 3305 (1986), topological quantum codes, Phys. Rev. Lett. 104,
doi:10.1143/JPSJ.55.3305. 050504 (2010), doi:10.1103/PhysRevLett.104.050504,
[55] J. C. Pezzullo, Nonlinear least squares regression arXiv:0911.0581.
(curve fitter) (2011), (Online calculator), URL http: [71] J. Feldman, M. J. Wainwright, and D. R. Karger,
//statpages.org/nonlin.html. Using linear programming to decode binary linear
[56] I. Jensen and A. J. G. and, Self-avoiding walks, codes, IEE Trans. Info. Theo. 51, 954 (2005),
neighbour-avoiding walks and trails on semiregular lat- doi:10.1109/TIT.2004.842696.
tices, J. Phys. A: Math. Gen. 31, 8137 (1998), [72] Similarly, the three-dimensional color codes [63] offer a
doi:10.1088/0305-4470/31/40/008. way to generate an infinite code family from the fifteen-
[57] I. Jensen, Improved lower bounds on the con- qubit Reed-Muller code by a means other than concate-
nective constants for two-dimensional self-avoiding nation.
walks, J. Phys. A: Math. Gen. 37, 11521 (2004),