Quantum Computation with Topological Codes

— from qubit to topological fault-tolerance —

Keisuke Fujii1,2,3
arXiv:1504.01444v1 [quant-ph] 7 Apr 2015

1
The Hakubi Center for Advanced Research, Kyoto University
Yoshida-Ushinomiya-cho, Sakyo-ku, Kyoto 606-8501, Japan
2
Graduate School of Informatics, Kyoto University
Yoshida Honmachi, Sakyo-ku, Kyoto 606-8501, Japan
3
Department of Physics, Graduate School of Science, Kyoto University
Kitashirakawa Oiwake-cho, Sakyo-ku, Kyoto 606-8502, Japan

April 8, 2015

Abstract
This is a comprehensive review on fault-tolerant topological quantum computation
with the surface codes. The basic concepts and useful tools underlying fault-tolerant
quantum computation, such as universal quantum computation, stabilizer formalism,
and measurement-based quantum computation, are also provided in a pedagogical
way. Topological quantum computation by brading the defects on the surface code
is explained in both circuit-based and measurement-based models in such a way that
their relation is clear. The interdisciplinary connections between quantum error cor-
rection codes and subjects in other fields such as topological order in condensed matter
physics and spin glass models in statistical physics are also discussed. This manuscript
will be appeared in SpringerBriefs.

1
Contents

1 Introduction to quantum computation 6

1.1 Quantum bit and elementary operations . . . . . . . . . . . . . . . . . . . . 6
1.2 The Solovay-Kitaev algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Multi-qubit gates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Universal quantum computation . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 Quantum algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5.1 Indirect measurement and the Hadamard test . . . . . . . . . . . . . 17
1.5.2 Phase estimation, quantum Fourier transformation, and factorization 19
1.5.3 A quantum algorithm to approximate Jones polynomial . . . . . . . 21
1.6 Quantum noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2 Stabilizer formalism and its applications 28

2.1 Stabilizer formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2 Clifford operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Pauli basis measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4 Gottesman-Knill theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.5 Graph states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.6 Measurement-based quantum computation . . . . . . . . . . . . . . . . . . . 40
2.7 Quantum error correction codes . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.8 Magic state distillation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.8.1 Knill-Laflamme-Zurek protocol . . . . . . . . . . . . . . . . . . . . . 54
2.8.2 Bravyi-Kitaev protocol . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3 Topological stabilizer codes 61

3.1 Z2 chain complex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2 A bit-flip code: exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.3 Definition of surface codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.3.1 Surface code on a torus: toric code . . . . . . . . . . . . . . . . . . . 66
3.3.2 Planar surface code . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4 Topological quantum error correction . . . . . . . . . . . . . . . . . . . . . . 70
3.5 Error correction and spin glass model . . . . . . . . . . . . . . . . . . . . . 76
3.6 Other topological codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

2
 3.7 Connection to topological order in condensed matter physics . . . . . . . . 84

4 Topological quantum computation 91

4.1 Defect pair qubits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.2 Defect creation, annihilation, and movement . . . . . . . . . . . . . . . . . . 93
4.3 Logical CNOT gate by braiding . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4 Magic state injections and distillation . . . . . . . . . . . . . . . . . . . . . 99
4.5 Topological calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.6 Faulty syndrome measurements and noise thresholds . . . . . . . . . . . . . 108
4.7 Experimental progress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

5 Topologically protected MBQC 114

5.1 Topological cluster state in 3D . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.2 Vacuum, defect, and singular qubit regions . . . . . . . . . . . . . . . . . . 116
5.3 Elementary operations in topological MBQC . . . . . . . . . . . . . . . . . 117
5.4 Topological quantum error correction in 3D . . . . . . . . . . . . . . . . . . 123
5.5 Applications for MBQC on thermal states . . . . . . . . . . . . . . . . . . . 124

A Fault-tolerant quantum computation 129

A.1 Fault-tolerant syndrome measurements . . . . . . . . . . . . . . . . . . . . . 129
A.2 Fault-tolerant gate operations . . . . . . . . . . . . . . . . . . . . . . . . . . 132
A.3 Concatenated quantum computation . . . . . . . . . . . . . . . . . . . . . . 132

B Decoding stabilizer codes 134

3
Preface

In 1982, Richard Feynman pointed out that a simulation of quantum systems on classi-
cal computers is generally inefficient because the dimension of the state space increases
exponentially with the number of particles [Fey82]. Instead, quantum systems could be
simulated efficiently by other quantum systems. David Deutsch put this idea forward by
formulating a quantum version of a Turing machine [Deu85]. Quantum computation en-
ables us to solve certain kinds of problems that are thought to be intractable with classical
computers such as the prime factoring problem and an approximation of the Jones poly-
nomial. It has a great possibility to disprove the extended (strong) Church-Turing thesis,
i.e., that any computational process on realistic devices can be simulated efficiently on a
probabilistic Turing machine.
However, for this statement to make sense, we need to determine whether or not
quantum computation is a realistic model of computation. Rolf Landauer criticized it
(encouragingly) by suggesting to put a footnote: “This proposal, like all proposals for
quantum computation, relies on speculative technology, does not in its current form take
into account all possible sources of noise, unreliability and manufacturing error, and prob-
ably will not work.” [Llo99]. Actually, quantum coherence, which is essential for quantum
computation is quite fragile against noise. If we cannot handle the effect of noise, quantum
computation is of a limiting interest, like classical analog computers, as a realistic model
of computation. To solve this, many researchers have investigated the fault-tolerance of
quantum computation with developing quantum error correction techniques. One of the
greatest achievements of this approach is topological fault-tolerant quantum computation
using the surface code proposed by R. Raussendorf et al. [RHG06, RHG07, RH07]. It says
that nearest-neighbor two-qubit gates and single-qubit operations on a two-dimensional
array of qubits can perform universal quantum computation fault-tolerantly as long as the
error rate per operation is less than ∼ 1%.
In this book, I present a self-consistent review of topological fault-tolerant quantum
computation using the surface code. The book covers everything required to understand
topological fault-tolerant quantum computation, ranging from the definition of the sur-
face code to topological quantum error correction and topological operations on the surface
code. The basic concepts and powerful tools for understanding topological fault-tolerant
quantum computation, such as universal quantum computation, quantum algorithms, sta-
bilizer formalism, and measurement-based quantum computation, are also introduced in

4
the first part (Chapter 1 and Chapter 2) of the book. In particular, in Chapter 1, I also
mention a quantum algorithm for approximating the Jones polynomials, which is also re-
lated to topological quantum computation with braiding non-Abelian anyons. In Chapter.
3, the definition of the surface code and topological quantum error correction on it is ex-
plained. In Chapter 4, topological quantum computation on the surface code is described
in the circuit-based model, where topological diagrams are introduced to understand the
logical operations on the surface code diagrammatically. In Chapter. 5, I explain the
same thing in the measurement-based model, as done in the original proposal [RHG06].
Hopefully, it would be easy to see how these two viewpoints are related.
Throughout the book, I have tried to explain the quantum operations using circuit
and topological diagrams so that the readers can get a graphical understanding of the
operations. The graphical understanding should be helpful to study the subjects more
efficiently. Topological quantum error correction codes are a nice play ground for studying
the interdisciplinary connections between quantum information and other fields of physics,
such as condensed matter physics and statistical physics. Actually, there is a nice corre-
spondence between topological quantum error correction codes and topologically ordered
systems in condensed matter physics. Furthermore, if we consider a decoding problem of
a quantum error correction code, a partition function of a random statistical mechanical
model is naturally appeared as a posterior probability for the decoding. These interdisci-
plinary topics are also included in Chapter 3.
Almost all topics, except for the basic concepts in the first part, are based on the results
achieved after the appearance of the standard textbook of quantum information science
entitled “Quantum Computation and Quantum Information” (Cambridge University Press
2000) by M. A. Nielsen and I. L. Chuang. In this sense, the present comprehensive review
on these topics would be helpful to learn and update the recent progress efficiently. In
this book, I concentrated on the quantum information aspect of topological quantum
computation. Unfortunately, I cannot cover the more physical and condensed matter
aspects of topological quantum computation, such as non-Abelian anyons and topological
quantum field theory. In this sense, this book is complemented by the book “Introduction
to topological quantum computation” (Cambridge University Press 2012) written by J. K.
Pachos. Readers who are interested in the more physical aspects of topological quantum
computation are recommended to read it.
Hopefully, this review will encourage both theoretical and experimental researchers to
find a more feasible way of quantum computation. It will also bring me great pleasure
if this review provides an opportunity to reunify and refine various subdivided fields of
modern physics in terms of quantum information.

Kyoto, Japan Keisuke Fujii,

April, 2015

5
Chapter 1

Introduction to quantum
computation

In this chapter, we introduce the basic concepts of quantum computation. We first describe
the minimum unit of quantum information, the qubit, and define several gate operations for
it. Then, we explain the Solovay-Kitaev algorithm, which provides a way to decompose an
arbitrary single-qubit unitary operation into an elementary set of single-qubit gates. Using
multi-qubit gates, we construct an arbitrary n-qubit unitary operation from an elementary
universal set of gates, which we call universal quantum computation. Quantum algorithms,
which run on a universal quantum computer, are also presented. One example of this is
Shor’s prime factoring algorithm based on the phase estimation algorithm. Another is an
approximation of the Jones polynomial. Finally, we will introduce quantum noise and see
how we can describe a quantum system coupled with an environment.

1.1 Quantum bit and elementary operations

In classical information science, the minimum unit of information is described by a binary
digit or bit, which takes the value 0 or 1. Its quantum counterpart is a quantum bit, the
so-called qubit.
 The qubit is defined
  as a linear superposition of two orthogonal quantum
1 0
states |0i = and |1i = ,
0 1

|ψi = α|0i + β|1i, (1.1)

where α and β are arbitrary complex values satisfying |α|2 + |β|2 = 1. The complex
amplitudes can be expressed as
θ θ
α = cos , β = eiφ sin , (1.2)
2 2
up to an unimportant global phase. By using the angles θ and φ, the qubit can be mapped
onto a point on a three-dimensional (3D) sphere, the so-called Bloch sphere, as shown in
Fig. 1.1.

6
 Figure 1.1: The Bloch sphere.

The time evolution, which maps a quantum state into another, is given as a unitary
operator in quantum mechanics. The most important operators are the Pauli operators
       
1 0 0 1 0 −i 1 0
I= , X= , Y = , Z= . (1.3)
0 1 1 0 i 0 0 −1

The computational basis states {|0i, |1i} are eigenstates of the Pauli Z operator. The
Pauli X operator flips the computational basis state:

|1i = X|0i, |0i = X|1i. (1.4)

We define the eigenstates of the Pauli X operator as

|0i + |1i |0i − |1i

|+i = √ , |−i = √ , (1.5)
2 2
which we call the X-basis states. Similarly, the eigenstates of the Pauli Y operator are
defined as
|0i + i|1i |0i − i|1i
| + ii = √ , | − ii = √ , (1.6)
2 2
which we call the Y -basis state.
The second most important operators are the Hadamard H and phase S operators
   
1 1 1 1 0
H=√ and S = .
2 1 −1 0 i

7
These gates transform between the different Pauli-basis states, i.e., H : {|0i, |1i} ↔
{|+i, |−i} and S : {|+i, |−i} ↔ {| + ii, | − ii}. Equivalently, we may say that these
operators transform a Pauli operator into another Pauli operator under their conjuga-
tions:

X = HZH, Y = SXS † . (1.7)

From this property, the H and S gates are called Clifford gates. The Clifford gates are
depicted as circuit diagrams as follows:

If we measure the qubit |ψi in the Z-basis {|0i, |1i}, we obtain the measurement
outcomes 0 and 1 with probabilities

p0 = |h0|ψi|2 = Tr [|0ih0|ψihψ|] , (1.8)

p1 = |h1|ψi|2 = Tr [|1ih1|ψihψ|] , (1.9)

respectively. More generally, we may use the measurement operators {Mi }, satisfying
P
that Ei ≡ Mi† Mi is positive semidefinite and that i Ei = I. (If an operator A satisfies
∀ |ui, hu|A|ui ≥ 0, it is said to be positive semidefinite.) The probability of obtaining the

measurement outcome i is given by

h i
pi = hψ|Mi† Mi |ψi = Tr Mi† Mi |ψihψ| = Tr [Ei |ψihψ|] . (1.10)

Such a measurement and the set of positive operators {Ei = Mi† Mi }, are called a positive-
operator-valued measure (POVM) measurement and POVM elements, respectively. The
post-measurement state conditioned on the measurement outcome i is
Mi |ψi
q . (1.11)
Tr[Mi† Mi |ψihψ|]

Suppose we have quantum states |ψi and |φi with probability pψ and pφ , respectively.
We perform a measurement with the measurement operator {Mi }. If we assume that the
measurement outcome i is obtained with probability
h i h i
pi = pψ Tr Mi† Mi |ψihψ| + pφ Tr Mi† Mi |φihφ| , (1.12)

we can express the classical mixture of |ψi and |φi as an operator

ρ = pψ |ψihψ| + pφ |φihφ|, (1.13)

8
which is called the density matrix or density operator. More generally, if we have pure quan-
P
tum states {|ψk i} with probability {pk }, the density matrix is given by ρ = k pk |ψihψ|.
A density matrix, which has information of a quantum state including its statistical prop-
erty, has to be a positive hermitian (self-adjoint) operator ρ satisfying Tr [ρ] = 1. Note that
for a given density matrix ρ, the pure state decomposition of it is not uniquely determined,
and any decomposition has the same statistical property.
A mixed state of a qubit ρ can be represented as a point inside the Bloch sphere
through the three coordinates calculated from the density matrix

(rx , ry , rz ) = (Tr[Xρ], Tr[Y ρ], Tr[Zρ]). (1.14)

If the state is a pure state |ψi = cos 2θ |0i + eiφ sin θ2 |1i, the coordinates can be calculated
to be

(rx , ry , rz ) = (sin θ cos φ, sin θ sin φ, cos θ), (1.15)

which is consistent with the previous definition. The completely mixed state I/2 corre-
sponds to the origin of the coordinate system.

1.2 The Solovay-Kitaev algorithm

The Pauli operators {I, X, Y, Z} and single-qubit Clifford operators {H, S} form a finite
group, and hence cannot cover all unitary operations for a qubit. If a non-Clifford oper-
ation exists, e.g., e−i(π/8)Z , we can generate an arbitrary single-qubit unitary operation
using the Solovay-Kitaev algorithm [DN05]. The underlying Solovay-Kitaev theorem states
that if a set of single-qubit operations generates a dense subset of SU (2), then that set is
guaranteed to fill SU (2) quickly.
Suppose we have a basic (0th order) approximation U0 of an arbitrary unitary operator
U , and that it approximates U with a certain constant error ǫ0 :

kU0 − U k ≤ ǫ0 , (1.16)

where k...k indicates an operator norm. The Solovay-Kitaev algorithm takes the (n − 1)th
order approximation Un−1 with an error ǫn−1 and returns the nth order approximation
Un with an error ǫn as follows.
†
First, U Un−1 is decomposed in terms of the unitary operators V and W as a group
commutator:
†
U Un−1 = V W V †W †, (1.17)

where V and W are chosen such that

√ √
kV − Ik < c ǫn−1 , kW − Ik < c ǫn−1 , (1.18)

9
with a constant c. The above decomposition always exists from the following argument.
Let V and W be rotations with an angle φ about the x- and y-axes, respectively. Then
V W V † W † is a rotation with an angle θ about some axis, where
q
2
sin(θ/2) = 2 sin (φ/2) 1 − sin4 (φ/2). (1.19)

We assume that φ, θ ≪ 1, and hence θ ≃ φ2 . By inverting the above argument, if

kI − V W V † W † k ≃ θ/2 + O(θ 3 ) (1.20)

we can find V and W such that

kI − V k, kI − W k ≃ φ/2 + O(φ3 ). (1.21)

Because
†
kI − U Un−1 k = kI − V W V † W † k < ǫn−1 , (1.22)
√
we can find V and W such that kI − V k, kI − W k < c ǫn−1 .
Second, we calculate the (n − 1)th order approximations Vn−1 and Wn−1 of V and W ,
respectively. Then the Solovay-Kitaev algorithm returns the nth order approximation
† †
Un = Vn−1 Wn−1 Vn−1 Wn−1 Un−1 , (1.23)

which satisfies

kUn − U k ≤ ǫn . (1.24)

Next, we will calculate ǫn as a function of ǫn−1 . By using the property of the operator
norm, we obtain
† † †
kUn − U k ≤ kVn−1 Wn−1 Vn−1 Wn−1 − U Un−1 kkUn−1 k (1.25)
† † † †
≤ kVn−1 Wn−1 Vn−1 Wn−1 − V W V W kkUn−1 k (1.26)

By denoting Vn−1 = V + ∆V , Wn−1 = W + ∆W , we obtain

† †
kUn − U k ≤ kVn−1 Wn−1 Vn−1 Wn−1 − V W V †W †k (1.27)
≤ k∆V W V W + V† †
W ∆†V W † k † †
+ kV ∆W V W + V W V †
∆†W k + O(∆2 ).
(1.28)

Moreover, in terms of δV and δW defined by V = I + δV and W = W + δW respectively,

we obtain

kUn − U k ≤ k∆V V † + V ∆†V k + k∆W W † + W ∆†W k (1.29)

+k∆V δW k + k∆†V δW
†
k + kδV ∆W k + kδV† ∆†W k + O(∆2 ). (1.30)
3/2 2 2
≤ 4cǫn−1 + O(∆ , δ ∆). (1.31)

10
Here we have used that
∆V V † + V ∆†V = ∆V ∆†V , (1.32)
which can be derived from the unitarity of V and Vn−1 = V + ∆V . To leading order, the
3/2
error is given by ǫn ≤ c′ ǫn−1 with a certain constant c′ and is calculated to be
n
ǫn ≤ (2c′ ǫ0 )(3/2) /(2c′ ). (1.33)
If ǫ0 < 1/(2c′ ), the error decreases super-exponentially in n. On the other hand, the nth
order approximation calls the (n − 1)th order approximation three times, i.e., through
Un−1 , Vn−1 , and Wn−1 . Including the resource RD required for the decomposition (1.17),
the overhead Rn for the nth order approximation is given by
Rn = 3Rn−1 + RD (1.34)
⇔ Rn = O(3n ). (1.35)
Similarly, the number of unitary operations employed in the nth order approximation is
calculated to be Mn = O(5n ). By using (1.33) and (1.35), the overhead Rǫ and the number
Mǫ of gates required to obtain an approximation of U with an error ǫ can be estimated:
Rǫ = O(lnln 3/ ln(3/2) (1/ǫ)), (1.36)
ln 5/ ln(3/2)
Mǫ = O(ln (1/ǫ)). (1.37)
Thus, both Rǫ and Mǫ scale as polylogarithmic functions of 1/ǫ.

1.3 Multi-qubit gates

An n-qubit state is given by a superposition of tensor product states
X
|Ψi = Ci1 i2 ...in |i1 i2 ...in i,
i1 ,i2 ,...,in

where ik = 0, 1 and |i1 i2 ...in i ≡ |i1 i ⊗ |i2 i ⊗ ... ⊗ |in i. A single qubit gate A acting on the
kth qubit is denoted by
k−1 n−k−1
z }| { z }| {
Ak = I ⊗ · · · ⊗ I ⊗A ⊗ I ⊗ · · · I . (1.38)
An important two-qubit gate is the controlled-NOT (CNOT) gate,
Λc,t (X) = |0ih0|c It + |1ih1|c Xt .
For a computational basis input state |iic |jit , the CNOT gate acts as Λc,t(X)|iic |jit =
|i ⊕ ji. In this sense, the CNOT gate is a quantum generalization of the XOR operation
in classical computation. If the input state is |+ic |0it , the output of the CNOT gate is a
maximally entangled state:
√
Λc,t (X)|+ic |0it = (|00i + |11i)/ 2. (1.39)
The CNOT gate is depicted by a circuit diagram as follows:

11
 We may also define the controlled-Z (CZ) gate,
Λc,t (Z) = |0ih0|c It + |1ih1|c Zt .
The CZ operation is symmetric; Λc,t(Z) = Λt,c (Z), and is represented by the circuit
diagram

Because the Pauli Z operator is transformed into the Pauli X operator by the Hadamard
gate, we have the following relation between the CZ and CNOT gates:

These CNOT and CZ gates are both Clifford gates, i.e., Λ(A) (A = X, Z) transforms
the two-qubit Pauli group onto itself under the conjugation Λ(A)[· · ·]Λ(A)† . For example,
Λ(X)c,t Xc ⊗It Λ(X)†c,t = Xc ⊗Xt , Λ(X)c,t Ic ⊗Zt Λ(X)†c,t = Zc ⊗Zt , Λ(Z)c,t Xc ⊗It Λ(Z)†c,t =
Xc ⊗ Zt , etc.
For an arbitrary unitary operator U , the controlled-U gate is denoted by
Λc,t (U ) = |0ih0|c It + |1ih1|c Ut , (1.40)
where the qubits c and t are called the control and target qubits, respectively. By de-
composing a single-qubit unitary gate into U = eiα AXBXC with the unitary operators
A, B, C satisfying ABC = I, the controlled-U operation Λ(U ) can be implemented as
follows:

12
(Note that we always have such a decomposition for an arbitrary single-qubit gate [BBC+ 95,
NC00].)
For example, the controlled-Hadamard gate Λ(H) can be represented by

Next, we will discuss one of the most important multi-qubit gates, the Toffoli gate:

Λ2c1 ,c2 ,t (X) = (Ic1 Ic2 − |1ih1|c1 |1ih1|c2 )It + |1ih1|c1 |1ih1|c2 Xt , (1.41)

which is depicted as a circuit diagram:

For a computational basis input state |i1 ic1 |i2 ic2 |jit , the Toffoli gate acts as

Λ2c1 ,c2 ,t (X)|i1 ic1 |i2 ic2 |jit = |i1 ic1 |i2 ic2 |j ⊕ (i1 · i2 )it . (1.42)

The state of the third qubit is equivalent to the output of the NAND operation in classical
computation. In this sense, the Toffoli operation can be regarded as a quantum extension
of the NAND operation. The NAND operations are known to be universal in classical
computation in the sense that any logic gate (boolean function) can be constructed from
them. This implies that quantum computation trivially includes classical computation.
More importantly, because unitary operations are reversible, quantum computation can
simulate classical computation in a reversible way. Suppose we want to calculate a boolean
function f (x) for an input state x. Then, we can construct a quantum circuit Uf consisting
of Toffoli and Pauli X gates:

Uf |xiinput |0...0iancilla |0ianswer = |xiinput |g(x)iancilla |f (x)ianswer , (1.43)

where the qubits |·iinput , |·iancilla , and |·ianswer indicate the registers for the input state,
the ancillae for the Toffoli operations, and the answer of the calculation, respectively. The
output state of the ancilla register |g(x)iancilla is the garbage of the computation. However,
the garbage can be uncomputed as follows (see also the circuit diagram below):

Uf† Λanswer,out (X)Uf |xiinput |0...0iancilla |0ianswer |0iout

= Uf† |xiinput |g(x)iancilla |f (x)ianswer |f (x)iout
= |xiinput |0...0iancilla |0ianswer |f (x)iout . (1.44)

13
In this way, we can calculate an arbitrary boolean function in a reversible way.
Finally, we introduce the multi-controlled unitary gate

Λk (U ) = (I ⊗k − |1ih1|⊗k ) ⊗ I + |1ih1|⊗k ⊗ U, (1.45)

where U is applied to the target qubit if all k control qubits are |1i. The multi-controlled
gate can be implemented by using the Toffoli gates and k ancilla qubits as follow:

The Λ2 (U ) gate can be decomposed into CNOT and single-qubit gates by using an idea
similar to the decomposition of the Λ(U ) gate:

For example, the Toffoli gate can be constructed from the CNOT, Hadamard, and π/8
operations as follows:

14
 √
where we have used the fact that that the controlled- X gate is decomposed into the
CNOT, Hadamard, and π/8 (T = e−i(π/8)Z ) gates as follows:

It should be noted that a multi-controlled gate Λk (U ) can also be constructed without any
ancilla qubit from multi-controlled gates Λ(k−1) (U ) with lower controlled qubits by using
the trick employed for decomposing Λ2 (U ) into Λ(U ) [BBC+ 95].

1.4 Universal quantum computation

Here we will show that an arbitrary unitary operation can be decomposed into single-
qubit and CNOT gates. To this end, we will first show how to decompose an arbitrary
unitary operation into a product of two-level unitary gates. Second, the two-level gates are
decomposed into the multi-controlled gates, which can be constructed from single-qubit
and CNOT gates as seen in the previous section.
Let U be an arbitrary n-qubit unitary operator, represented by an m×m unitary matrix
with m ≡ 2n . Let Tij be a unitary operator such that (Tij )kl = δkl if k, l 6= i, j, which we
call a two-level unitary gate. (The kl element of a matrix A is denoted by (A)kl .) The ii,
ij, ji, and jj elements define a unitary operation on the two-level subsystem spanned by
|ji and |ii. By choosing Tm m−1 appropriately, we have
 
u11 ··· u1 m−1 u1 m
 .. .. .. .. 
U Tm m−1 = 
 . . . . 
, (1.46)
 um−1 1 · · · u′m−1 m−1 u′m−1 m 
um 1 · · · 0 u′m m

where (U )kl = ukl . By repeating this procedure, we obtain

 
u′′11 ··· u′′1 m−1 u′′1 m
 .. .. .. .. 

U Tm m−1 Tm m−2 · · · Tm 1 =  . . . . 
. (1.47)
 u′′m−1 1 · · · u′′m−1 m−1 u′′m−1 m 
0 ··· 0 u′′m m

15
Due to unitarity, u′′1 m = · · · = u′′m−1 m = 0 and |u′′mm | = 1. Defining Rm ≡ Tm m−1 Tm m−2 · · · Tm 1 ,
we can decompose U into a product of Rk and a diagonal unitary operator D:
U = D(Rm · · · R1 )† . (1.48)
It is obvious that D can be decomposed into two-level unitary gates. Thus, an arbitrary
unitary operator U can be decomposed into two-level unitary gates.
Next, we show that any two-level unitary operator Tij can be implemented by using
CNOT and single-qubit gates. Let us rewritte i and j (i, j = 0, ..., m−1) by using the n-bit
strings s = s1 s2 ...sn and t = t1 t2 ...tn , respectively. It is easy to find a sequence of n-bit
strings {gk }dk=1 such that s = g1 , t = gd , and gk and gk+1 differ by only one bit. By using
the Pauli X gate and the multi-controlled-NOT gate controlled by the same n − 1 bits
and targeting the one different bit, we can transform the basis |gk i to |gk+1 i. In this way,
the basis |ii = |si = |g1 i is transformed into |gd−1 i. After this basis transformation, we
now want to apply a two-level unitary gate between |gd−1 i and |ji = |ti = |gd i. Because
gd−1 and gd differ by only one bit, we can perform such a two-level unitary gate by the
multi-conditional gate, using the Pauli X gate for bit flips. Finally, the basis |gd−1 i is
returned to |si by applying the inverse of the basis transformation.
For example, a two-level unitary operator acting on a subspace spanned by {|000i, |111i}
can be implemented as follows:

As seen previously, the multi-conditional gate can be decomposed into CNOT and
single-qubit gates. Moreover, an arbitrary single-qubit unitary operation can be approx-
imated by using the Hadamard and π/8 operations by virtue of the Solovay-Kitaev algo-
rithm. Thus, the CNOT, Hadamard, and π/8 operations form a universal set of operations
for quantum computations. The Toffoli and Hadamard operations also form a universal
set as shown in [RZBB94, BBC+ 95, DiV95, BV93, NC00].

1.5 Quantum algorithms

In this section, we explain two representative quantum algorithms; Shor’s prime factoriza-
tion algorithm [Sho94, Sho97] and the Aharonov-Jones-Landau algorithm for an additive
approximation of the Jones polynomial [AJL09, AA06]. We do not go deep into the math-
ematically rigorous details, but we aim to understand how they work. The readers who are
interested in more details should read Refs. [Sho94, Sho97, Kit95a, NC00, AJL09, AA06].

16
1.5.1 Indirect measurement and the Hadamard test
An indirect measurement of an observable (hermitian) A with eigenvalues ±1 can be
performed by using Λ(A):

By denoting the input state |ψi, the post-measurement state is given by

I + (−1)s A p
|ψi/ Tr[(I + (−1)s A)/2|ψihψ|], (1.49)
2
where s = 0, 1 is the measurement outcome. For example, a circuit measuring the eigen-
value of the operator X1 X2 X3 is given by

According to the measurement outcome s = 0, 1, the post-measurement state is projected

s
by I+(−1) 2X1 X2 X3 . The indirect measurement will be employed frequently in quantum
error correction to measure the eigenvalues of the stabilizer operators.
The Hadamard test of an arbitrary unitary operator U is defined by the following
circuit:

The probabilities of the measurement outcomes 0, 1 of the X-basis measurement are cal-
culated to be
1

p0 = 1 + Reh0|⊗n U |0i⊗n , (1.50)
2
1

p1 = 1 − Reh0|⊗n U |0i⊗n . (1.51)
2

17
Similarly, the Hadamard test for the imaginary part is defined:

Suppose we perform the Hadamard test N times and obtain the measurement outcome 0,
N0 times. By using the Chernoff-Hoeffding bound,
 
N0 2
Prob − p0 > ǫ < 2e−2ǫ N , (1.52)
N

we can estimate the matrix element h0|⊗n U |0i⊗n with an error ǫ by repeating the Hadamard
test N = poly(1/ǫ) times. The Hadamard test is employed in various quantum algorithms
such as approximations of the Jones and Tutte polynomials [AJL09, AA06, AAEL07]
and the partition functions of statistical mechanical models [DlCDVdNMD11, ICBB14,
MFI14].
Specifically, if we choose the input state to be a completely mixed n-qubit state,

then, the Hadamard test provides the trace Tr[U ]/2n of the unitary operator U . Such a
restricted type of quantum computation is called a deterministic quantum computation
with one clean qubit (DQC1) [KL98]. DQC1 seems to be less powerful than universal
quantum computation, because only one qubit is a pure state. However, DQC1 can
evaluate functions, which would be intractable on a classical computer, such as the Jones
and Homefly polynomials [SJ08, JW09], spectral density function [KL98], and fidelity
decay [PBKLO04]. Recently, classical sampling of the output of DQC1 with a few qubits
measurements has been shown to be intractable unless the polynomial hierarchy collapses
to the third level [MFF14, MNFT14].

18
1.5.2 Phase estimation, quantum Fourier transformation, and factoriza-
tion
If we have an eigenstate |Ei i of a unitary operator U , for which Λ(U ) is described by a
polynomial number of gates, we can estimate the eigenvalue λi of U with a polynomial
accuracy using the Hadamard test as follows:

When we do not have the eigenstate, the input state is projected onto one of the eigenstates
by the Hadamard test. Thus, by repeating the Hadamard test, we can obtain one of the
eigenvalues with polynomial accuracy.
k k
Moreover, if a controlled-U 2 gate Λ(U 2 ) can be described by a polynomial number
of gates, we can estimate the eigenvalue with exponential accuracy [Kit95b]. Suppose the
eigenvalue is given by λi = eiφ = e2πi0.j1 j2 ...jn , where we employ a decimal of a binary
number,
n
X
0.j1 j2 ...jn = jk (1/2)k . (1.53)
k=1
Then, the Kitaev’s phase estimation algorithm is given by

k
where QFT stands for the quantum Fourier transform. The controlled gate Λ(U 2 ) kicks
k
back the phase e2 φ = e2πijk .jk+1 ...jn to the ancilla state |+i. The phase information
is transformed into a computational basis state by using the inverse quantum Fourier
transformation:
QF T † (|0i + e2πi0.j1 j2 ...jn |1i)(|0i + e2πi0.j2 ...jn |1i)...(|0i + e2πi0.jn |1i) = |j1 j2 ...jn i.
(1.54)

19
Here QFT is given by

Then, we can obtain the phase φ = (2πi)0.j1 j2 ...jn through the Z-basis measurements.
Note that the accuracy of the estimate is improved exponentially compared to the Hadamard
k
test (provided that we have a polynomial size description of Λ(U 2 )).
Let N and x be co-prime integers. The order-finding problem is about finding an order
r such that xr = 1 mod N . This problem can be solved by using the phase estimation
against a unitary operator
X
Ux = |xy mod N ihy|, (1.55)
y

because we have

Ux |us i = e2πi(s/r)) |us i, (1.56)

for a state
r−1
1 X −2πi(s/r)k k
|us i = √ e |x mod N i. (1.57)
r
k=0

Note that we can prepare the initial state |us i randomly by using the fact that |1i =
Pr−1
s=0 |us i. After estimating the phase 2πi(s/r), the continued fraction provides r with
k
high probability. The modular exponentiation x2 mod N can be calculated by using the
k k
square k times. Thus we can implement controlled-Ux2 , Λ(Ux2 ), by a polynomial number
of gates.
If we randomly choose s, we obtain an even order r with high probability. Thus, we
have (xr/2 − 1)(xr/2 + 1) = 0 mod N . Finally, the euclidean algorithm gives us the greatest
common divisor of xr/2 − 1 and N or xr/2 + 1 and N , and it is a factor of N . This is the
so-called Shor’s prime factoring algorithm. Kitaev, who formulated the phase estimation
algorithm, generalized this idea for the more general Abelian stabilizer problem [Kit95a].

20
 Figure 1.2: (a) A link diagram. (b) A braid diagram. (c) A tangle diagram.

1.5.3 A quantum algorithm to approximate Jones polynomial

Next we will explain Aharonov-Jones-Landau algorithm for approximating the Jones poly-
nomial [AJL09, AA06], which is an algorithmic version of the original proposal based on
topological quantum field theory [FKLW03, Wit89]. We will show that the approximation
of the Jones polynomial is BQP-complete; the approximation of the Jones polynomial can
be done efficiently by using a universal quantum computer, and inversely it is as hard
as any problems (BQP) solvable by a universal quantum computation. Below we will
introduce the braid diagram and the braid group, which are related to link diagrams with
an appropriate closure. A unitary representation of the braid group is constructed by
embedding the braid group to the Temperley-Lieb (TL) algebra. Using the constructed
representation, the Jones polynomial and quantum algorithm are connected.
Let us first define the Jones polynomial. The Jones polynomial is an invariant of a link,
that is, closed loops in a 3D space, which are tangled in general [Lic97]. It is convenient
to describe the link on a 2D projected space as shown in Fig. 1.2 (a), which we call a link
diagram. Specifically two links are equivalent if two link diagrams can be transformed into
each others under the Reidemeister moves:

which appropriately reflect the continuous deformations in the 3D space. The Jones poly-
nomial is calculated from a directed link diagram L as follows: (i) Smooth each crossing
∧
∧
) (

տ
ր in two ways { ∧
) (, }. Let s be the resultant diagram consisting of closed loops with
∧


no crossing, which we call a state. (ii) For each state s, we assign a weight
+ −s−
W (s) = As d|s|−1 , (1.58)

where d = −(A2 + A−2 ) is a complex constant called a loop value, |s| is the number of the
∧
∧
) (

crossing, and s+ and s− are the number of smoothings by ∧ ) ( and , respectively. By

∧

21
Figure 1.3: (a) A multiplication for two braid diagrams b1 and b2 . (b) The generator σi
of the braid group. (c) The generator Ei of the TL algebra.

taking the summation over all states, the Kauffman bracket hLi of the link L is defined:
X
hLi = W (s). (1.59)
s

The Jones polynomial is defined as a function of t = A−4 by multiplying a factor to the

Kauffman bracket:

VL (t) = (−A)3ω(L) hLi, (1.60)

where the writhe ω(L) of the link L is defined as the number of տր 

type crossings minus
ր
that of տ type. It is easy and a good exercise to confirm that the Jones polynomial is
invariant under the Reidemeister moves based on the above definition [Lic97].
Let Bn be a braid group consisting of the braid diagrams of n strands, where two
endpoints of each strand are tied at the top and bottom, respectively as shown in Fig. 1.2
(b). The multiplication of two braids b1 and b2 are defined by connecting the bottom and
top endpoints of two braid diagrams as shown in Fig. 1.3 (a). The braid group Bn is
generated by n − 1 generators {σi } subject to

σi σj = σj σi for |i − j| ≥ 2 (1.61)
σi σi+1 σi = σi+1 σi σi+1 . (1.62)

The generator σi corresponds to the braid diagram with only one crossing of ith and i+1th
strands as shown in Fig. 1.3 (b). Then, the latter equality is nothing but the Reidemeister
move III. By introducing an appropriate closure, which closes the endpoints of the strands,
a braid diagram is related to a link diagram as we will see later.
In order to construct a unitary representation of Bn , we embed it into the Temperley-
Lieb (TL) algebra T Ln (d) [TL71] on tangle diagrams with no crossing as shown in Fig. 1.2
(c). Similarly to the previous case, T Ln (d) is generated by a set of generators {E1 , ..., En−1 }
subject to

Ei Ej = Ej Ei for |i − j| ≥ 2, (1.63)

22
Figure 1.4: (a) The Markov trace closure for the braid b. (b) The plat closure for the
braid b. (c) The plat closure is deformed into the Markov trace closure.

Ei Ei±1 Ei = Ei , (1.64)
Ei2 = dEi . (1.65)

The generator Ei corresponds to ∪ and ∩ for the ith and (i + 1)th strands as shown in
Fig. 1.3 (c). The braid group is embedded into the TL algebra by

σi = AEi + A−1 1. (1.66)

In order to relate the braid diagram and the Jones polynomial, we employ a Markov
trace closure on the tangle diagram as shown in Fig. 1.4 (a). The Markov trance closure
denoted by mtr has the following property:

mtr(1) = 1, (1.67)
mtr(AB) = mtr(BA) for A, B ∈ T Ln (d), (1.68)
mtr(A) = dmtr(AEn−1 ) for A ∈ T Ln−1 (d). (1.69)

Now the representation of the braid group and the Jones polynomial are related. Eq.
∧
∧
) (

(1.66) corresponds to superposition of two smoothings { ∧ ) (, } of a crossingտր . More

∧


precisely, let ρ and ρ̃ be a representation of the TL algebra and an induced representation
of the braid group, respectively. Moreover, the matrix trace has the same property as the
Markov trace closure. Thus, for a given braid diagram b, we have

Vbmtr (A−4 ) = ∆Tr[ρ̃(b)], (1.70)

23
where bmtr is a link diagram generated from the braid diagram b with the Markov trace
mtr
closure, and ∆mtr = (−A)3ω(b ) dn−1 .
Let us construct a representation of the TL algebra, the so-called path-model represen-
tation. Suppose G is a one-dimensional graph with k vertices labeled by 1, ..., k from left
to right and k − 1 edges. We consider an n step walk on G starting from the left endpoint.
Let p = (v1 , ..., vn ) be a path of such an n step walk, where v1 = 1 and vj = 0 and = 1
mean moving to the left and right neighboring vertices, respectively. Then we define a
Hilbert space Hn,k spanned by all possible paths as a basis {|pi}. For each generator Ei ,
we define its representation Φi = ρ(Ei ) on Hn,k as follows:

Φi |...vi−1 00vi+1 ...i = 0, (1.71)

p
λzi −1 λzi +1 λzi −1
Φi |...vi−1 01vi+1 ...i = |...vi−1 01vi+1 ...i + |...vi−1 10vi+1 ...i,
λzi λz i
(1.72)
p
λzi +1 λzi +1 λzi −1
Φi |...vi−1 10vi+1 ...i = |...vi−1 10vi+1 ...i + |...vi−1 01vi+1 ...i,
λzi λz i
(1.73)
Φi |...vi−1 11vi+1 ...i = 0, (1.74)

where zi ∈ {1, ..., k} is the label of the vertex at the ith step and λj = sin(jθ) with θ = π/k.
The representation ρ̃(σi ) = AΦi + A−1 1 for the braid group Bn is induced by ρ. It is easy
to confirm that, with A = ie−iθ/2 and d = 2 cos θ, the representation ρ is hermitian for all
Ei , and hence the induced representation ρ̃ is unitary. Since ρ(σi ) is a unitary operator
acting on the neighboring two-qubit, a multiplication of such unitary operators can be
implemented by using universal quantum computer. Since, by using the Hadamard test in
Sec. 1.5.1, we can evaluate the Jones polynomial Vbmtr (A−4 ) with an additive error ∆2n ǫ
taking poly(n, m, k, 1/ǫ) overhead, where m is the number of the crossing.
The approximation scale is further improved by considering another closure, the so-
called plat closure as shown in Fig. 1.4 (b). We deform the link diagram obtained by the
plat closure to another link diagram obtained by the Markov trace closure as shown in
Fig. 1.4 (c), Then, we have ∪ and ∩ for all strands. In this case, the support of ρ(b) is
only on |10101...i, which allows us to replace the matrix trace as follows:

Vbplt (A−4 ) = ∆plt h10101...|ρ(b)|10101...i, (1.75)

where bplt is a link generated from the braiding b with the plat closure and ∆plt =
plt
(−A)3ω(b ) dn/2−1 . (Note that for each ρ(Ei ) of n/2 ∪s and ∩s, a factor λ2 /λ1 = d is
substituted.) Again using the Hadamard test, we can estimate Vbplt (A−4 ) with an addi-
tive error ∆plt ǫ taking poly(n, m, k, 1/ǫ) overhead.
Finally, we show that the approximation of the Jones polynomial with an additive
error ∆plt ǫ is BQP-hard. Unfortunately, the computational basis defined by the path
p = (v1 , ..., vn ) is inappropriate for this purpose, since the n-step walks cannot span the

24
whole 2n -dimensional Hilbert space. Instead, we employ the four-step encoding [WY06],
where the computational basis states are defined by the following two four-steps:

|0̄i ≡ |1100i, |1̄i ≡ |1010i. (1.76)

The space H4n,k spanned by the 4n-step walks can support the whole 2n -dimensional space
for {|0̄i, |1̄i}⊗n . The dimension of the space spanned by the 8-step walks starting from
vertex 1 ending at vertex 1 is 14. Thus the unitary representation ρ of the braid group B8
results in unitary operators in SU (14). If they are dense in SU (14), we can approximate
an arbitrary unitary operator in SU (4) spanned by two four-step encoded qubits by using
the Solovay-Kitaev algorithm [AA06], which are enough to implement universal quantum
computation. The density is achieved by k > 4 and k 6= 6. For such a parameter, the
approximation of the Jones polynomial with an additive error ∆plt ǫ is enough to solve a
BQP-complete problem.
The approximation of the Jones polynomial can be done by universal quantum com-
puter and also is enough to solve the problems solvable by universal quantum computer.
Thus the approximation of the Jones polynomial is a BQP-complete problem [AJL09,
AA06]. The AJL algorithm for approximation of the Jone polynomial was extended for
the Tutte polynomial in Ref. [AAEL07], where the Solovay-Kitaev algorithm for non-
unitary linear operators was developed.

1.6 Quantum noise

Quantum coherence, one of the essential properties of quantum systems, is quite fragile
against noise, due to interactions between the system and the environment. Suppose that
the system S of interest interacts via a unitary operation U with the environment E,
where the system and environment are initially uncorrelated. The reduced density matrix
ρ′S of the system after the interaction is calculated to be
h i
ρ′S = TrE U (ρS ⊗ ρE )U † . (1.77)
P
Using a spectral decomposition of the initial state in the environment, ρE = k pk |ek ihek |,
we obtain a map of the system S:
X †
ρ′S = K(k′ ,k)ρS K(k ′ ,k) , (1.78)
(k ′ ,k)

where
√
K(k,k′ ) = pk he′k |U |ek i. (1.79)

This map satisfies

X †
K(k ′ ,k) K(k ′ ,k) = pk hek |U † |e′k ihe′k |U † |ek i
k
= IS , (1.80)

25
where IS is the identity operator in the system S.
In general, a map of a quantum state is given as a completely-positive-trace-preserving
(CPTP) map E, subject to

• Tr[Eρ] = 1 for any density matrix ρ (preservation of probability)

P P
• E( i qi ρi ) = i qi Eρi (convex linear map)

• IA ⊗ ES (ρAS ) ≥ 0 (complete positivity) for an arbitrary ancilla system A and density

matrix ρAS on the composite system AS.

Such a CPTP map can always be written by Kraus operators {Kk } [NC00]:
M
X
Eρ = Kk ρKk† . (1.81)
k=1

Under the Born-Markov (with rotating-wave) approximation, the time evolution of

a two-level system coupled with an environment is given by a master equation of the
Lindblad form [PB02]:
γ+
ρ̇(t) = − [σ− σ+ ρ(t) + ρ(t)σ− σ+ − 2σ+ ρ(t)σ− ]
2
γ−
− [σ+ σ− ρ(t) + ρ(t)σ+ σ− − 2σ− ρ(t)σ+ ]
2
γ0
− [σz ρ(t) + ρ(t)σz − 2σz ρ(t)σz ] ≡ Lρ(t), (1.82)
2
where σ+ = |0ih1|, σ− = |1ih0| and γα (α = 0, +, −) are the decay rates of the decay
channels. One can easily find the eigenoperators of the Lindblad super-operator. These
eigenoperators form the damping basis [BE93]:
γ+ + γ− + 2γ0
Lσ1 = σ1 ≡ λ1 σ1 , (1.83)
2
γ+ + γ− + 2γ0
Lσ2 = σ2 ≡ λ2 σ2 , (1.84)
2
Lσ3 = (γ+ + γ− )σ3 ≡ λ3 σ3 , (1.85)
Lρeq = 0, (1.86)

where ρeq = (γ+ |0ih0|+γ− |1ih1|)/(γ+ +γ− ) ≡ (σ0 +aσ3 )/2 with a = (γ+ −γ− )/(2γ+ +2γ− ).
The solution of this master equation is given by the CPTP map E(t):
X
E(t)ρ = p0 (t)ρ + pi (t)σi ρσi + f (t)(σ3 ρ + ρσ3 − iσ1 ρσ2 + iσ2 ρσ1 ),
i=1,2,3
(1.87)

where
1
p0 (t) = (1 + e−λ1 t + e−λ2 t + e−λ3 t ), (1.88)
4

26
 1
p1 (t) = (1 + e−λ1 t − e−λ2 t − e−λ3 t ), (1.89)
4
1
p2 (t) = (1 − e−λ1 t + e−λ2 t − e−λ3 t ), (1.90)
4
1
p3 (t) = (1 − e−λ1 t − e−λ2 t + e−λ3 t ), (1.91)
4
a
f (t) = (1 − e−λ3 t ). (1.92)
4
If we consider a high temperature case (i.e., a → 0), Eq. (1.87) can be rewritten as
3
X
E(t)ρ = [1 − p1 (t) − p2 (t) − p3 (t)] ρ + pi (t)σi ρσi . (1.93)
i=1

Hence, the CPTP map can be viewed as a stochastic Pauli error with probabilities pi (t).
In general, noise cannot be written by a stochastic Pauli error. However, by performing
an appropriate operation, one can depolarize the CPTP map into a stochastic Pauli error
as a standard form, where the noiseless part of the evolution is not altered [DHCB05].
Otherwise, the Pauli basis measurements can collapse the CPTP map into a stochastic
Pauli error, as we will see later. In the rest of this book, therefore, we will consider
Markovian stochastic Pauli errors only.
The fault-tolerance against more general noise has been discussed in Refs. [ABO97,
ABO08, AGP06, AGP08, AP09]. If the decoherence is non-Markovian, the dynamical de-
coupling or quantum Zeno effect can be used to suppress the decoherence [Zur84, VGW96,
DG98, VL98, VKL99a, FTP+ 05]. Besides, if the decoherence is spatially correlated,
one can utilize a passive error-prevention scheme, the so-called decoherence free subspace
(DFS), which is immune to collective noise [ZR97, DG97, LCW98b].

27
Chapter 2

Stabilizer formalism and its

applications

In general, the description of quantum states is a difficult task because it requires expo-
nentially many parameters in the number of qubits as shown in Eq. (1.38). To understand
these complex quantum systems, it is essential to have efficient tools. The stabilizer for-
malism is one such powerful tool to describe an important class of entangled states. It also
provides a diagrammatic understanding of quantum states and operations. The stabilizer
states, described by the stabilizer formalism, play important roles in quantum computa-
tion, such as for quantum error correction codes and resource states in MBQC. In this
chapter, we introduce the stabilizer formalism, especially focusing on its diagrammatic
understanding. Based on the stabilizer formalism, we explain quantum error correction,
magic state distillation, and MBQC.

2.1 Stabilizer formalism

We first define an n-qubit Pauli group Pn :

Pn := {±1, ±i} × {I, X, Y, Z}⊗n . (2.1)

An element of the Pauli group is called a Pauli product. For example, the two-qubit Pauli
group is given by

P2 := {±1, ±i}
×{II, IX, IY, IZ, XI, XX, XY, XZ, Y I, Y X, Y Y, Y Z, ZI, ZX, ZY, ZZ},(2.2)

where A ⊗ B is denoted by AB for simplicity. (We will frequently use this notation when
there is no possibility for confusion.) Next, we define an n-qubit stabilizer group S as an
Abelian (commutative) subgroup of the n-qubit Pauli group:

/ S and ∀ Si , Sj ∈ S, [Si , Sj ] = 0.
S := {Si } s.t. − I ∈ (2.3)

28
Because −I is not included in the stabilizer group, all elements are hermitian Si = Si† ,
which guarantees that the eigenvalues = ±1. An element of the stabilizer group is called
a stabilizer operator. The maximum independent subset Sg of the stabilizer group is
called stabilizer generators. Here, independence means that any element of Sg cannot be
expressed as a product of other elements in Sg . Any element of the stabilizer group can
be generated as a product of the stabilizer generators. The stabilizer group S generated
by the generators Sg is denoted by S = hSg i.
Let us, for example, consider a two-qubit stabilizer group:

SBell = {II, XX, ZZ, −Y Y }. (2.4)

Because they contain two anticommuting Pauli operators, XX and ZZ commutes. The
stabilizer group SBell is generated by {XX, ZZ}, because −Y Y can be expressed as a
product of XX and ZZ. Thus, we can write SBell = h{XX, ZZ}i.
For a given stabilizer group S, the stabilizer state is defined as a simultaneous eigenstate
of all stabilizer elements Si ∈ S with the eigenvalue +1:
∀
Si ∈ S, Si |ψi = |ψi. (2.5)

It is sufficient that the state is an eigenstate of all stabilizer generators:

∀
Si ∈ Sg , Si |ψi = |ψi. (2.6)

Let k be the number of elements in the stabilizer generator Sg . Each stabilizer generator
divides an n-qubit system (Hilbert space) into two orthogonal subspaces associated with
the eigenvalues ±1. Because all stabilizer operators commute with each other, the k
stabilizer generators divide the n-qubit system into 2k orthogonal subspaces. Thus, the
dimension of the space spanned by the stabilizer states, which we call a stabilizer subspace,
is 2d = 2n−k . When n = k, we can define the quantum state uniquely. The number of
stabilizer generators is at most n for an n-qubit stabilizer group. In the case of k < n, the
degrees of freedom in the stabilizer subspace can be addressed by using logical operators,
which commute with all stabilizer generators and also are independent of them.
Let us consider the stabilizer group SBell again. The stabilizer state is the eigen-
state of XX and ZZ with eigenvalue +1, and hence given by the Bell state (|00i +
√
|11i)/ 2 [EPR35]. If XX is removed from the generators, the two-dimensional subspace
spanned by |00i and |11i is stabilized. By choosing logical operators LX = XX and
LZ = ZI, we can specify the state in the subspace. For example, the eigenstate of LX
with the eigenvalue +1 is the Bell state. The eigenstate of LZ with the eigenvalue +1 is
|00i. Another representative example of the stabilizer states is an n-qubit cat state,
1
|cati = √ (|00...0i + |11...1i), (2.7)
n
whose stabilizer group is given by
* n
+
Y
Z1 Z2 , ..., Zn−1 Zn , Xi . (2.8)
i=1

29
Figure 2.1: The stabilizer formalism: a Heisenberg picture of quantum computation.
A Clifford operation is represented as a transformation of the stabilizer group by the
conjugation of U .

The cat state is a representative example of a macroscopically entangled state. If it is

determined whether a particle is |0i or |1i, the superposition is completely destroyed. If an
Q
element ni=1 Xi is removed from the stabilizer generator, it defines a stabilizer subspace
Q
spanned by |00..0i and |11..1i. We can choose LX = ni=1 Xi and LZ = Zi as logical
operators, which anti-commute with each other and behave as logical Pauli operators.

2.2 Clifford operations

In the stabilizer formalism, we can describe a restricted class of unitary operations, the
so-called Clifford operations, acting on the stabilizer states quite efficiently. The Clifford
operation is defined as an operation U that transforms a Pauli product into another Pauli
product under its conjugation, [...] → U [...]U † . Let us consider the action of a Clifford
operation U on the stabilizer state |ψi defined by a stabilizer group S = h{Si }i:

U |ψi = U Si |ψi = U Si U † U |ψi = Si′ U |ψi, (2.9)

where we define Si′ ≡ U Si U † . The above equality indicates that the state U |ψi is an eigen-
state of the operator Si′ with an eigenvalue +1 for all Si′ . Because U is a Clifford (unitary)
operation, the group {Si′ } is also an Abelian subgroup of the Pauli group. Accordingly, the
state U |ψi is a stabilizer state with respect to the stabilizer group {Si′ }. In this way, the
action of U on the stabilizer state can be represented as a transformation of the stabilizer
groups under the conjugation of U as shown in Fig. 2.1. For example, the stabilizer state
stabilized by hX1 I2 , I1 Z2 i is |+i1 |0i2 . The stabilizer group is transformed by Λ(X)1,2 into
√
hX1 X2 , Z1 Z2 i, whose stabilizer state is (|00i + |11i)/ 2.
The stabilizer formalism corresponds to the Heisenberg picture of quantum computa-
tion, where a minimum number of operators are employed to describe a restricted type of
quantum states and operations [Got97, Got98a]. This representation is powerful because
it requires us to keep a time evolution of at most n operators, while a straightforward
state-based approach needs exponentially many states. For example, let us consider the
following quantum circuit:

30
A straightforward calculation yields the output state |ψi,

|ψi = (|0000000i + |1010101i + |0110011i + |1100110i

+|0001111i + |1011010i + |0111100i + |1101001i
+|1111111i + |0101010i + |1001100i + |0011001i
+|1110000i + |0100101i + |1000011i + |0010110i)/4. (2.10)

It is rather cumbersome to write down the above state. Instead, we can understand the
output state as a stabilizer state whose stabilizer generators are

{ZIZIZIZ, IZZIIZZ, IIIZZZZ, XXXIIII, XXIXXII, IXIXIXI, XIIXIIX}.

(2.11)

Equivalently, we may also choose the following stabilizer generators because they generate
the same stabilizer group:

{ZIZIZIZ, IZZIIZZ, IIIZZZZ, XXXXXXX, IIIXXXX, XIXIXIX, IXXIXXI}.

(2.12)

Actually, these stabilizer generators are enough to understand the properties of the quan-
tum state |ψi. If an explicit description of the state is required, we can systematically
write it down as follows:
I + S4 I + S3 I + S2 I + S1
|ψi = 4 |0000000i, (2.13)
2 2 2 2
where S1 = XIXIXIX, S2 = IXXIIXX, S3 = IIIXXXX, and S4 = XXXXXX.
The above equation means that |0000000i is an eigenstate for all Z’s stabilizer operators.
By projecting it into the +1 eigenstate of the stabilizer generator Si by the projection
I+Si
2 , we obtain the stabilizer state |ψi.
In order for the above calculation to work, we have to obtain the stabilizer generators
of the output state. This can easily be done graphically. We introduce commutation rules
between the Pauli operators and Clifford operations below. In the case of the Hadamard
operation, HX = ZH and ZH = HX, and hence we have

31
meaning that the Pauli X operator acting before the Hadamard operation is equivalent
to the Pauli Z operator acting after the Hadamard operation and so on. Similarly, for the
phase operation X, we have

The CNOT operation transforms the Pauli operators under its conjugation as follows:

Λc,t (X)Xc Λc,t (X) = Xc Xt , (2.14)

Λc,t (X)Xt Λc,t (X) = Xt , (2.15)
Λc,t (X)Zc Λc,t (X) = Zc , (2.16)
Λc,t (X)Zt Λc,t (X) = Zc Zt . (2.17)

The commutation relation between the CNOT operation and the Pauli operators is un-
derstood as follows:

In the above circuit diagram, the solid circle commutes with the Pauli Z operator, while
the Pauli X operator is propagated as the Pauli X operator on the target qubit, making
a correlation. Similarly, the open circle commutes with the Pauli X operator, while the
Pauli Z operator is propagated as the Pauli Z operator on the control qubit, making a
correlation. By recalling that the CNOT operation is transformed into the CZ operation
by the Hadamard operations on the target qubit, the commutation relation between the
CZ operation and the Pauli operators are obtained straightforwardly. This is described
graphically as follows:

32
In this case, note that the Pauli X operation is propagated as the Pauli Z operation.
This graphical understanding allows us to calculate the stabilizer generators of the
output of the Clifford circuits. For example, in the following circuit diagram, the first
qubit is stabilized by X before the Clifford operation. The Pauli X operator is propagated
toward the right, and we obtain the stabilizer operator ZIZIZIZ for the output:

The reader should use this graphical technique to calculate the other stabilizer generators
and verify Eq. (2.11).

2.3 Pauli basis measurements

Next, we will see how the Pauli-basis measurements on the stabilizer states are described
in the stabilizer formalism. Suppose the A-basis (A = X, Y, Z) measurement is performed
on a stabilizer state |ψi, whose stabilizer group is given by hSi i. (We assume that the
number of stabilizer generators is equal to the number of qubits, and hence that the
stabilizer state is uniquely defined.) Depending on the stabilizer group h{Si }i and A,
there are two possibilities:

33
 (i) The Pauli operator A commutes with all stabilizer generators. In that case, either A
or −A is an element of the stabilizer group. If A (−A) is an element, the eigenvalue
+1 (−1) is obtained with probability 1. The post-measurement state is the same as
the stabilizer state before measurement.

(ii) At least one stabilizer operator does not commute with A. In this case, we can
choose another set of generators {Si′ } such that S1′ anti-commutes with A but all
other generators commute with A. The measurement outcomes +1 and −1 are ob-
tained with an equal probability of 1/2. The post-measurement state is given by
h(−1)m A, S2′ , ..., Sk′ i depending on the measurement outcomes m = 0, 1 correspond-
ing to the eigenvalues (−1)m .

For example, suppose we perform the Y -basis measurement on the first qubit of the
Bell state stabilized by SBell = hXX, ZZi. We can redefine the stabilizer generators by
{XX, −Y Y }. Then the stabilizer group after the measurement is given by hY I, −Y Y i =
hY I, −IY i. Thus, we obtain | − ii as the post-measurement state on the second qubit.

2.4 Gottesman-Knill theorem

Because the stabilizer states and Clifford operations are described efficiently in the sta-
bilizer formalism, it implies that such a restricted type of quantum computation can be
simulated efficiently on a classical computer. This is stated by the Gottesman-Knill theo-
rem [Got97, Got98a, NC00].

Theorem 1 Any Clifford operations, applied to the input state |0i⊗n followed by the Z
measurements, can be simulated efficiently in the strong sense.

Here, an efficient strong classical simulation of a quantum circuit C is a classical polynomial-

time computation that calculates the probability PC (x) for a given output x of the circuit
P
C, including an arbitrary marginal distribution x′ PC (x). (See, for example, Ref. [BJS11]
for the definition of a strong simulation.) Note that this theorem holds true even when the
initial state is generalized to an arbitrary stabilizer state, and also any Pauli products are
measured, because they are done in the above setup by modifying the Clifford operations
appropriately.
Proof: The stabilizer group of the input state is h{Zi }i (i = 0, 1, ..., n − 1). By apply-
ing the Clifford operations as mentioned, we obtain the stabilizer generators h{Si }i of the
quantum output before the measurements. Suppose the measurement outcome, the clas-
sical output, is given by {mi = 0, 1}. Then the probability of obtaining the measurement
outcome {mi } can be calculated as follows:

i) Set the stabilizer generators S (0) = h{Si }i and the initial probability p(0) = 1.

ii) For k = 0, 1, ..., n − 1, repeat the following procedures.

34
 1) If (−1)mk Zk ∈ S (k) , update the probability p(k+1) = p(k) , because the measure-
ment outcome mk is obtained with probability 1. The stabilizer group after the
measurement is also updated to S (k+1) = S (k) .
2) Else, if (−1)mk ⊕1 Zk ∈ S (k) , update the probability p(k+1) = 0, because such a
measurement outcome does not appear. (You may stop the calculation at this
stage, and return the probability 0.)
3) Else, S (k) is updated into S (k+1) by removing an anticommuting generator and
adding (−1)mk Zk as a new generator. Because the measurement outcome is ob-
tained randomly with probability 1/2, the probability is taken as p(k+1) = p(k) /2.

iii) Return p(n) as the probability of obtaining the measurement outcome {mi }.

Note that, in step (ii), we can efficiently decide which of the three is the case for any k by
checking the commutability of Zk with the stabilizer generators of S (k) . 
The statement of Theorem 1 can be extended by weakening the notion of the classical
simulation.

Theorem 2 Any Clifford operations, applied to any product states of convex mixtures
of the Pauli basis states, followed by Z measurements can be efficiently simulated in the
weak sense.

Here, an efficient weak classical simulation of a quantum circuit C is a classical polynomial-

time randomized computation that samples the output x according to the probability
distribution PC (x) of the output of the circuit C. (See, for example, Ref. [BJS11] for the
definition of weak simulation.) Apparently, a strong simulation includes a weak simulation,
because we sample the output by using the marginal distributions [TD04].
Proof: Suppose that the ith input qubit is given by
(i) (i) (i) (i)
ρi = px,+ |+ih+| + px,− |−ih−| + py,+ | + iih+i| + py,− | − iih−i|
(i) (i)
+pz,+ |0ih0| + pz,− |1ih1|, (2.18)
P P (i) (i)
where α=x,y,z ν=+,− pα,ν = 1. By using the probability distribution {pα,ν }, the input
state of each qubit is randomly sampled. Conditioned by the sampling result, the input
state is a product of the Pauli basis states, and hence the output probability distribution
can be calculated as shown in Theorem 1. Combined with the random sampling of the
input state, this provides an efficient weak simulation of the Clifford circuit with noisy
input states (convex mixture of the Pauli basis states). 
The input state can be generalized into a classical mixture of stabilizer states, when
its polynomial size description of the probability distribution is provided. Similarly, the
Clifford operations can be extended to stochastic Clifford operations such as the stochastic
Pauli error.
The convex mixture of the Pauli basis state lies inside the octahedron of the Bloch
sphere as shown in Fig. 2.2. It is natural to ask whether or not the Clifford circuit allows

35
Figure 2.2: A convex mixture of the Pauli basis states lies inside the octahedron of the
Bloch sphere.

universal quantum computation if the input state lies outside the octahedron. If the input
state is a pure non-stabilizer state such as e−i(π/8)Z |+i, we can implement a non-Clifford
gate e−i(π/8)Z by using gate teleportation, explained in Sec. 2.6. Even some mixed states
can be converted into a pure non-stabilizer state, the so-called magic state, by using only
Clifford operations. Such a protocol is called magic state distillation [BK05] and will be
explained in Sec. 2.8.

2.5 Graph states

In this section, we introduce an important class of stabilizer states, the so-called graph
states [HDE+ 06], whose stabilizer generators are defined on graphs. The graph states are
employed as resource states for MBQC as explained in the next section.
A graph state is defined by a graph G = (V, E). Here, V and E are the sets of the
vertices and edges, respectively. A qubit is located on each vertex of the graph. The
stabilizer generator of the graph state |Gi is defined as
Y
Ki = Xi Zj for all i ∈ V, (2.19)
j∈Vi

where we define a set of vertices Vi := {j|(i, j) ∈ E}, which are connected to the vertex i
by an edge on the graph G (see Fig. 2.2). The graph state |Gi is generated from a product

36
Figure 2.3: The graph state |Gi associated with a graph G = (V, E). A stabilizer generator
Ki is also shown.

state |+i⊗|V | by applying the CZ gate on each of the graphs:

Y
|Gi = Λ(Z)i,j |+i⊗|V | , (2.20)
(i,j)∈E

where |V | indicates the number of vertices of the graph G = (V, E). This can be un-
derstood that the stabilizer generator Xi for the state |+i is transformed into Ki by the
Q
CZ operations U ≡ (i,j)∈E Λ(Z)i,j . Especially, when the graphs are regular lattices such
as one-dimensional (1D), square, hexagonal, and cubic lattices, the corresponding graph
states tend to be referred to as cluster states [BR01]. Any stabilizer state is equivalent
to a certain graph state up to local Clifford operations [VdNDDM04, HDE+ 06]. For ex-
ample, the cat state is equivalent to the following graph state by applying the Hadamard
operation on the kth qubit:

Unfortunately, the graph associated with a stabilizer state is not uniquely defined, because
there are local Clifford operations that change the underlying graph. This property is
called the local complementarity of the graph states [VdNDDM04, HDE+ 06].
Next, we will see how the Pauli basis measurements transform the graph states. For
simplicity, we assume that the state is projected into an eigenstate with eigenvalue +1.
Let us consider a 1D graph state as follows:

37
whose stabilizer generator is given by

Ki = Zi−1 Xi Zi+1 . (2.21)

We first consider the Z basis measurement (projective measurement of the observable Z)

on the ith qubit. Following the procedure seen in Sec. 2.3, Ki is removed from the stabilizer
generator. By adding Zi instead, we obtain the stabilizer group for the post-measurement
state

h..., Ki−1 , Zi , Ki+1 , ...i. (2.22)

After the projection, the ith qubit is |0i and hence decoupled from the other qubits. By
rewriting the stabilizer generators, we obtain three decoupled stabilizer groups

h..., Zi−2 Xi−1 i, hZi i, hXi+1 Zi+2 , ...i. (2.23)

This means that the graph is divided into two parts as follows:

For any graph, this property of the Z-basis measurement holds; the post-measurement
state is defined by a modified graph, where the vertex corresponding to the measured
qubit and the edges incident to it are removed from the original graph.
Next, we consider the X-basis measurement. The observable Xi does not commute
with Ki−1 and Ki+1 , but does commute with Ki−1 Ki+1 = Zi−2 Xi−1 Xi+1 Zi+2 . Following
the procedure in Sec. 2.3, the stabilizer group for the post-measurement state is calculated
to be

h..., Zi−2 Xi−1 Xi+1 Zi+2 , Zi−1 Zi+1 , ...i, hXi i. (2.24)

By performing the Hadamard operation H on the (i−1)th qubit, we obtain a new stabilizer
group

h..., Zi−2 Zi−1 Xi+1 Zi+2 , Xi−1 Zi+1 , ...i, hXi i, (2.25)

which indicates that the graph is transformed into the following graph with the Hadamard
operation:

38
Instead of the (i − 1)th qubit, we can obtain a similar result by performing the Hadamard
operation on the (i + 1)th qubit as shown above.
Suppose the ith and (i + 1)th qubits are measured in the X-basis on a 1D graph state.
This is equivalent to measuring the (i + 1)th qubit of the above post-measurement graph
state in the Z basis, because the Hadamard operation is applied on it as a byproduct.
From the previous argument, the Z-basis measurement remove the measured qubit from
the graph. Thus, two neighboring X-basis measurements remove the measured qubits and
connect the left and right hand sides directly:

which we call a contraction.

Finally, we consider the Y -basis measurement. The observable Yi does not commute
with either Ki−1 , Ki , or Ki+1 , but does commute with Ki−1 Ki = Zi−2 Yi−1 Yi Zi+1 and
Ki Ki+1 = Zi−1 Yi Yi+1 Zi+2 . The stabilizer group for the post-measurement state is calcu-
lated to be
h..., Zi−2 Yi−1 Zi+1 , Zi−1 Yi+1 Zi+2 , ...i, hYi i. (2.26)
By performing the phase gates S on the (i − 1)th and (i + 1)th qubits, we obtain a new
stabilizer group
h..., Zi−2 Xi−1 Zi+1 , Zi−1 Xi+1 Zi+2 , ...i, hYi i. (2.27)
This indicates that the graph is directly connected up to the phase operation S as a
byproduct:

Suppose three neighboring qubits (i − 1), i, and (i + 1) are measured in the Y -basis.
This is equivalent to measuring the ith qubit in the Y -basis, and then measuring the
(i − 1)th and (i + 1)th qubits of the post-measurement graph state in the X-basis, because
there is a phase operation S acting on them as a product. As seen previously, the X-basis
measurements on two neighboring qubits result in a contraction of the two qubits on the
graph. Thus, the Y -basis measurements on three neighboring qubits contract them from
the 1D graph state.

This property is useful to change even and odd of the length of the 1D graph state.
While we have considered the Pauli-basis measurements only on the 1D graph state,
we can generalize these arguments into graph states of general structures. A graph state
is still mapped into another graph state up to some single-qubit Clifford operations as
byproducts.

39
2.6 Measurement-based quantum computation
Measurement-based quantum computation (MBQC) is a model of quantum computation,
where quantum gates are implemented by adoptive measurements on a highly entangled
resource state [RB01, RBB03, Rau03]. Specifically, certain graph states, the so-called
cluster states, are employed as resource states in MBQC. Below we will first demonstrate
quantum teleportation, a building block of MBQC. Then, we explain how adoptive mea-
surements on a graph state enable us to emulate universal quantum computation via
quantum teleportation.
Quantum teleportation is a quantum communication protocol, in which Alice sends
a quantum state to Bob by using a shared entangled state and classical communica-
tion [BBC+ 93]. Suppose Alice and Bob share a maximally entangled state, the Bell state,
|0ia |0ib + |1ia |1ib
√ . (2.28)
2
For an unknown input state |ψii and the half of the Bell state, Alice performs a Bell
measurement, which is a projection onto the Bell basis states
|0ii |0ia + |1ii |1ia
|Ψ(m1 , m2 )ii,a = Xim1 Zim2 √ , (2.29)
2
where m1 , m2 = 0, 1 correspond to the measurement outcomes. A straightforward calcu-
lation provides
 
|0ia |0ib + |1ia |1ib
hΨ(m1 , m2 )|i,a |ψii √ = Zbm2 Xbm1 |ψib /2. (2.30)
2
Hence, the unknown input state is teleported to Bob with a byproduct operator X m1 Z m2 .
If Bob does not know the measurement outcomes (m1 , m2 ), the teleported state is a
completely mixed state for Bob. However, if Alice sends the measurement outcome as a
classical message, Bob can undo the byproduct and obtain the unknown quantum state
at Bob’s side. The circuit diagram of quantum teleportation is:

By using the following circuit equivalence, we can decompose the teleportation circuit into
two elementary teleportations, the so-called one-bit teleportations:

40
One-bit teleportation is useful as a building block of the teleportation-based gates em-
ployed in MBQC. A single-qubit Z rotation eiθZ can be implemented in a teleportation-
based way. Its action can be understood from the following circuit equivalence:

where we utilized the fact that eiθZ and Λ(Z) commute. The controlled-Z operation is
also implemented in a teleportation-based way as follows:

41
That is, instead of performing the Λ(Z) gate after one-bit teleportations, we can pre-
pare a special resource state, on which the Λ(Z) gate is pre-implemented, and the Λ(Z)
gate is then performed via teleportation. These quantum operations based on quantum
teleportation are called gate teleportation [GC99].
Now we are ready to formulate MBQC. An arbitrary single-qubit unitary operation U
can be decomposed, up to an unimportant global phase, into

U = HeiφZ eiθX eiξZ (2.31)

= HeiφZ HeiθZ HeiξZ . (2.32)

This indicates that we can perform an arbitrary single-qubit unitary operation by a se-
quence of one-bit teleportations. The resource state for the sequential one-bit teleporta-
tions is a 1D cluster state:

where the stabilizer generator for the left-most qubit is removed and the input state
is encoded. We have to take care of the byproduct Pauli operators depending on the
measurement outcomes. Fortunately, we can propagate the Pauli byproduct operators
forward as follows:
′ ′
U = X mi+2 Heiφ Z X mi+1 Heiθ Z X mi HeiξZ . (2.33)
i(−1)mi+1 φ′ Z i(−1)mi θ ′ Z
= X mi+2 ⊕mi Z mi+1 He He HeiξZ . (2.34)

42
By choosing θ ′ = (−1)mi θ and φ′ = (−1)mi+1 φ adaptively, depending on the previous mea-
surement outcomes, the random nature of the measurements can be managed. This proce-
dure is called feedforward. The Pauli byproduct is propagated and updated throughout the
computation. Note that the classical processing required to determine the measurement
angle has only XOR (addition modulo two) operations [RBB03].
Next, we will investigate the measurement-based two-qubit gate operation. The re-
source state for the gate teleportation is the following cluster state:

To adjust the timing of the two-qubit operation, we can insert identity operations depend-
ing on the even and odd lengths as follows:

Without loss of generality, we can assume that all input states of the quantum computation
are given by |+i, which are automatically encoded by preparing the graph state. At the
end of the computation (on the right-most qubits), measurements are performed to read
out the result as follows:

43
In this way, universal quantum computation is simulated solely by measurements on a
brickwork-like cluster state. This state can be generated from a cluster state on a square
lattice by using the Pauli basis measurements as shown above. Accordingly, the square
lattice cluster states are universal resources for MBQC.
The above circuit-based explanation of MBQC [Nie06] is very intuitive and straight-
forward. However, for a complicated resource state, as will be shown, an operator-based
understanding of MBQC [RBB03] is quite useful. Let us reformulate MBQC from an
operator viewpoint. Suppose again an MBQC on a 1D cluster state. The measurements
are executed up to the (i − 1)th qubits, and hence the operator Kl (l ≤ i) is removed from
the stabilizer generators. The logical degree of freedom on the remaining resource state
can be specified by the ith logical operators
(i)
LX = Xi Zi+1 , (2.35)
(i)
LY = Yi Zi+1 , (2.36)
(i)
LZ = Zi . (2.37)

These logical operators commute with all remaining stabilizer generators Kl (l ≥ i + 1).
Moreover, they anticommute with each other, satisfying the commutation relations for the
Pauli operators. Thus, they specify the state encoded in the graph state. As seen above,

44
 (i)
a Z-rotation e−i(θ/2)Zi is applied before the X-basis measurement. Because Zi = LZ , this
rotation induces a unitary transformation U of the logical operator
(i) (i) (i)
LX → cos θLX + sin θLY , (2.38)
(i) (i) (i)
LY → cos θLY − sin θLX . (2.39)
(i) (i+1)
Because LX = Xi LZ , the logical X operator after the X-basis measurements is given
(i+1)
by (−1)mi LZ depending on the measurement outcome mi = 0, 1. On the other hand,
(i)
the logical operators LY,Z do not commute with the X-basis measurement; they are not
relevant logical operators after the measurement. If two operators are equivalent up to
multiplications of the stabilizer operators, their action on the stabilizer state is also the
same. By using this fact, we can replace the logical operators in (2.39) with
(i) (i) (i+1)
LZ ∼ LZ Ki+1 = Xi+1 Zi+2 ≡ LX , (2.40)
(i) (i+1)
LY ∼ Ki+1 = Xi Yi+1 Zi+2 ≡ Xi LY , (2.41)

where ∼ indicates that two operators are equivalent up to stabilizer operators. After the
X-basis measurement, Xi can be replaced by its eigenvalue (−1)mi . Then the logical
operator of the post-measurement state is given by
(i) (i+1) (i+1) (i+1)
LX → (−1)mi (cos θLZ + sin θLY ) = U LX U †, (2.42)
(i) (i+1) (i+1) (i+1)
LY → (−1)mi (cos θLY + sin θLZ ) = U LY U †, (2.43)
(i) (i+1) (i+1)
LZ → LX = U LZ U †. (2.44)

We now realize that the logical operators for the ith step are transformed into those for
the (i + 1)th step rotated by U ≡ X mi He−i(θ/2)Z .
Similarly, a two-qubit gate in MBQC can also be regarded as a propagation of a
correlation by a projection on the stabilizer state. Consider the following graph state.

(i) (i) (i) (i)

The logical operators for the ith step are given by {LX1 , LZ1 } and {LX2 , LZ2 }. By mul-
tiplying the stabilizer operator, we obtain
(i) (i+1) (i+1)
LZ1 ∼ X1,i+1 Z1,i+1 Z2,i+1 = LX1 LZ2 , (2.45)
(i) (i+1) (i+1)
LZ2 ∼ X2,i+1 Z2,i+1 Z1,i+1 = LX2 LZ1 . (2.46)

45
The logical operators for the (i + 1)th step after the projections are calculated to be
(i) (i) (i+1) (i+1) (i+1) (i+1) (i+1)
{LX1 , LZ1 } → {(−1)m1 LZ1 , LX1 LZ2 } = {V LX1 V † , V LZ1 V † }, (2.47)
(i) (i) m2 (i+1) (i+1) (i+1) (i+1) (i+1)
{LX2 , LZ2 } → {(−1) LZ2 , LX2 LZ1 } = {V LX2 V † , V LZ2 V † }. (2.48)

Again, we realize that the logical operators for the ith step are transformed into those for
the (i + 1)th step with a two-qubit unitary operation

V ≡ (X1 Z2 )m1 (X2 Z1 )m2 Λ1,2 (Z)H1 H2 . (2.49)

By combining single-qubit rotations X m Hei(θ/2)Z and the two-qubit operation (X1 Z2 )m1 (X2 Z1 )m2 Λ1,2 (Z)H1 H
as seen above, we can perform a universal quantum computation. In this way, MBQC can
be understood in the Heisenberg picture.
Suppose the logical Pauli operators of the kth input and output qubits are related by
the measurements as follows:
(In) (In) (Out) (Out)
{LX,k , LZ,k } → {U LX,k U † , U LZ,k U † }. (2.50)

The unitary operator U is performed on the input qubits. Here a Pauli byproduct, de-
pending on the measurement outcomes, is also included in U . Moreover, if two graph
states, which perform U and V , are concatenated with the appropriate feedforwarding of
the Pauli byproducts, then V U is performed:

Let us consider the example shown in Fig. 2.4 (a). The logical operators on the inputs
are replaced by multiplying stabilizer generators so that they commute with the X-basis
measurements. Then the X operators on the measured qubits are replaced by ±1. The
measurements transform the input logical operators as follows:
(In) (In) (In) (In)
{LX,1 , LZ,1 , LX,2 , LZ,2 } →
(Out) (Out) (Out) (Out) (Out) (Out)
{(−1)m1 ⊕m5 LX,1 LZ,2 , (−1)m3 ⊕m6 LZ,1 , (−1)m2 ⊕m6 LX,2 LZ,1 , (−1)m4 ⊕m5 LZ,2 }.
(2.51)

Thus, the Λ1,2 (Z) gate is implemented up to a Pauli byproduct.

Using this fact and concatenation of the input-output relations, we can construct a
measurement-based CNOT gate between the separated two-qubit as follows [RBB03]:

46
The input-output relation of the above graph state is equivalent to that for the following
circuit:

In this way, a CNOT gate between two arbitrary separated qubits can be implemented
using a constant depth (constant width) resource state.
Topologically protected quantum computation was first proposed as a MBQC on a 3D
cluster state [RHG06, RHG07] as we will see in Chapter 5. It was further applied to the
circuit-based model in two dimensions (2D) [RH07, FSG09], which will be explained in
detail in Chapter 4. Specifically, in the formulations, the operator-based understanding is
quite useful for tracking how correlation is propagated in a topologically protected way. In
Chapter 5, we will formulate a topologically protected MBQC from the operator viewpoint
and see how it is related to the circuit-based understanding explained in Chapter 4.
To conclude, we summarize the properties of MBQC and recent progress in this area.
A unique feature of MBQC is that the resource state for universal quantum computa-
tion is prepared offline. Entangling operations, which would be one of the most difficult
tasks in experiments, are employed only in this stage. Quantum computation is executed
solely by adaptive measurements. This property is useful for experimental realization in
certain physical systems. For example, a deterministic entangling operation is difficult to
achieve in an optical quantum computation. In such a case, we can utilize linear optics
and measurement-induced nonlinearity to generate a cluster state [YR03, Nie04, BR05].
Importantly, the entangling operation can be nondeterministic, as long as the successful
or non-successful outcome is heralded. By using such a probabilistic entangling operation,
we can gradually expand the cluster state. After successful cluster state generation, we

47
can start the measurements for quantum computation. Note that the probability of suc-
cessful cluster state generation is not exponentially small by using a divide and conquer
approach [Nie04, BR05, DR05, MBF10, LBSB10, FT10].
The clear separation between the quantum stage requiring entangling operations and
the measurement stage is useful, not only for the physical implementation, but also in
a quantum cryptographic scenario. Suppose that Bob (server) possesses a fully fledged
quantum computer and that Alice (client), who has a less advanced quantum device, such
as a single-qubit state generator, wants to delegate quantum computation to Bob. By using
the idea of MBQC, such a delegated quantum computation can be made unconditionally
secure. This is called a blind quantum computation and was proposed by Broadbent,
Fitzsimons, and Kashefi (BFK) [BFK09] (see also the related earlier works [Chi05, AS06]).
In the BFK protocol, Alice sends randomly rotated qubits {eiθj Z |+i} to Bob, where the
angle is chosen to be θj = kj π/4 (kj = 0, 1, ..., 7). Bob generates a cluster state by
using the randomly rotated qubits. In the computation phase, Alice sends a classical
message δj = φj + θj + rj π. Here, φj is the measurement angle with which Alice want to
perform a measurement. The angle θj is for the randomly rotated state (which is secret
to Bob). The random bit r ∈ {0, 1} makes the measurement angle completely random
for Bob. Then Bob performs the measurement in the {eiδj |±i} basis. Because the initial
state is pre-rotated by θj (from Alice’s viewpoint), Bob performs the measurement in the
{ei(φj +rj π) |±i} basis, which is what Alice wants to do. However, from Bob’s viewpoint,
the state is a completely mixed state with no information about {φj }. Thus, Bob is
blind to any information about the input, the algorithm, and the output. Instead of the
state generation, Alice, who has a measurement device, can also perform a blind quantum
computation, whose security is guaranteed by the no-signaling principle [MF13]. A fault-
tolerant blind quantum computation has been proposed, based on topologically protected
MBQC [MF12].

2.7 Quantum error correction codes

In this section, we introduce stabilizer codes, which are a class of quantum error correction
(QEC) codes.

Three-qubit bit-flip code

The QEC codes can be described elegantly in the stabilizer formalism. Let us first consider
the simplest one, the three-qubit bit flip code, whose stabilizer generators are given by
S1 = Z1 Z2 , S2 = Z2 Z3 . (2.52)
The stabilizer subspace is spanned by the following two logical states:
|0L i = |000i, |1L i = |111i. (2.53)
The logical Pauli-X operator is given by LX ≡ X1 X2 X3 . The logical Pauli-Z operator
is defined as LZ ≡ Z1 . We may, equivalently, choose the logical Pauli Z operator to be

48
Z2 or Z3 , because their actions on the code space are equivalent. The present code is a
quantum analogue of the classical three-bit repetition code. Consider a bit flip error with
an error probability p:

Ei ρ = (1 − p)ρ + pXi ρXi . (2.54)

If the initial state |ψL i = α|0L i + β|1L i undergoes the bit flip error independently, the
output state is transformed in leading order as
X
E1 ◦ E2 ◦ E3 |ψL ihψL | = (1 − p)3 |ψL ihψL | + p(1 − p)2 Xi |ψL ihψL |Xi + O(p2 ).
i
(2.55)

The error Xi maps the code space to an orthogonal space. We perform a projective
measurement onto the orthogonal subspaces, Pk± = (I ± Sk )/2, which we call a syndrome
measurement, to know in which orthogonal space the state lies. Note that the encoded
quantum information is not destroyed by the syndrome measurement, because it commutes
with the logical operators. According to the measurement outcomes, the logical state can
recover from the error as follows:

R ◦ E1 ◦ E2 ◦ E3 |ψL ihψL | = [(1 − p)3 + 3p(1 − p)2 ]|ψL ihψL | + O(p2 ), (2.56)

where the recovery operator is given by

Rρ = P1+ P2+ ρP2+ P1+ + X1 P1− P2+ ρP2+ P1− X1 + X2 P1− P2− ρP2− P1− X2
+X3 P1+ P2− ρP2− P1+ X3 . (2.57)

The four terms in Rρ correspond to the measurement outcomes (eigenvalues) (+1, +1),
(−1, +1), (−1, −1), and (+1, −1) of the stabilizer generators, respectively. By comparing
Eqs. (2.55) and (2.56), one can understand that if p is sufficiently small, the fidelity of the
logical state is improved.
Similarly, we can construct a three-qubit phase flip code, which can correct a phase
flip error, by changing the basis with the Hadamard transformation:

hZ1 Z2 , Z2 Z3 i → hX1 X2 , X2 X3 i. (2.58)

9-qubit Shor code

The three-qubit bit-flip code cannot correct Z errors, which commute with the stabilizer
generators. A QEC code that can correct all X, Y , and Z errors was developed by Shor
based on a concatenation of three-qubit bit-flip and phase-flip codes [Sho95]. The stabilizer

49
generators of the 9-qubit Shor code are given as follows:

X X X X X X I I I
I I I X X X X X X
Z Z I I I I I I I
I Z Z I I I I I I
(2.59)
I I I Z Z I I I I
I I I I Z Z I I I
I I I I I I Z Z I
I I I I I I I Z Z

The code space is spanned by

|0L i + |1L i (|000i + |111i) (|000i + |111i) (|000i + |111i)

√ = √ , (2.60)
2 2 2
|0L i − |1L i (|000i − |111i) (|000i − |111i) (|000i − |111i)
√ = √ . (2.61)
2 2 2

The logical Pauli operators are given by XL = X ⊗9 and ZL = Z ⊗9 , which are bitwise
tensor products of physical Pauli operators. If the logical A operator is given by a bitwise
tensor product of the physical A operators on the QEC code, we say that the operation A
has transversality. The 9-qubit code is capable of correcting all X, Y , and Z errors for each
qubit, which can be understood because the three-qubit phase flip code {|+++i, |−−−i} is
constructed by using the three logical qubits of the three-qubit bit flip codes {|000i, |111i}.
Note that any single-qubit noise E can be described by using the Kraus operators {Kj }:
X
Eρ = Kj ρKj† . (2.62)
j

Any operator Kj can be decomposed into the Pauli operators σ0 = I, σ1 = X, σ2 = Y ,

and σ3 = Z:
X
Kj = cjl σl . (2.63)
l

Thus, if the X and Z errors on a single qubit are both corrected appropriately, we can
correct any single-qubit noise automatically. Specifically, because noise contains a super-
position of the Pauli errors, it can be collapsed by the syndrome measurements.

Stabilizer codes
To summarize the above examples, let us formalize the stabilizer quantum error correc-
tion codes and their properties. The code space of a stabilizer QEC code is defined by
a stabilizer group h{Si }i. The encoded degree of freedom is specified by the mutually
independent logical operators {LZj }, which commute with all stabilizer generators and are
independent of the stabilizer generators. The computational basis state of the code state

50
is completely determined by the stabilizer group h{Si }, {(−1)mj LZ
j }i. We can always find
X
another set of the logical operators {Lj } being subject to
LX Z δij Z X
j Li = (−1) Li Lj , (2.64)
where δij is the Kronecker delta. Hence, the pair of logical operators LZ X
i and Li represents
the ith logical qubit. In terms of the numbers n and (n − k) of qubits and stabilizer
generators, respectively, the number of pairs of logical operators is k.
Let us define the weight wt(Si ) of a Pauli product Si as the number of qubits on which
a Pauli operator (except for the identity I) is acting. The minimum weight of the logical
operator over all possible logical operators is called the code distance d. This implies that
all Pauli products whose weights are smaller than d are elements of the stabilizer group
or anticommute with the stabilizer generators. Thus, they act trivially on the code state
or map the code state into an orthogonal subspace. If the weight of a Pauli product as
an error is less than (d − 1)/2, we can find a unique recovery operator that returns the
erroneous state into the code space. Thus, we can correct weight-⌊(d − 1)/2⌋ errors. Such
a stabilizer QEC code is called a [[n, k, d]] stabilizer code. For example, the code distance
of the 9-qubit code is a [[9, 1, 3]] stabilizer code correcting weight-one errors.
The nine-qubit code is not the smallest QEC code that can correct all weight-one X, Y ,
and Z errors. The smallest code is the five-qubit code, found independently by Laflamme
et al. [LMPZ96] and Bennett et al. [BDSW96]. The stabilizer generators and the logical
Pauli operators are given as follows:
S1 = X Z Z X I
S2 = I X Z Z X
S3 = X I X Z Z
(2.65)
S4 = Z X I X Z
XL = X X X X X
ZL = Z Z Z Z Z
We see that the code distance is three, and hence an arbitrary single-qubit error can be
corrected.

Calderbank-Shor-Steane codes
The readers who are familiar with classical coding theory might already be aware of the
correspondence between stabilizer codes and classical linear codes. Let us recall the 9-
qubit code. The X and Z errors are detected independently through the Z-type and
X-type stabilizer generators, respectively. This implies that X and Z error corrections
are described by classical coding theory, where two classical error corrections are subject
to a certain constraint to appropriately form a stabilizer group.
To formulate this, we briefly review classical linear codes. A [[n, k]] classical linear
code C is defined as a k-dimensional space VC over GF (2n ) by using an n × k generator
matrix
G = (b1 , ..., bk ), (2.66)

51
where the column vectors {bi } are the basis vectors of VC . A k-bit classical information
y is encoded into the code c as

a = Gc. (2.67)

To detect and analyze the errors, we define an (n − k) × n parity check matrix H such
that Hbk = 0 for all basis vectors {bk }. Suppose an error e occurs on the code state,
a′ = a ⊕ e, where ⊕ indicates a bitwise addition modulo two. By using the parity check
matrix H, we can detect the error

Ha′ = H(a ⊕ e) = He ≡ s, (2.68)

where s is called an error syndrome.

For example, the three-bit repetition code is defined by the generator
 
1
G=  1 . (2.69)
1

A classical bit 0 and 1 is encoded into (0, 0, 0)T and (1, 1, 1)T , respectively. The parity
check matrix is defined to be
 
1 1 0
H= . (2.70)
0 1 1
Now, we realize that the positions of the 1s of the parity check matrix are exactly the
same as those of the Zs in the stabilizer generators of the three-qubit bit flip code. This
suggests to use the parity check matrices Hx and Hz of the two classical linear codes Cx
and Cz , respectively, in the definition of the X-type and Z-type stabilizer generators:
(i)
Y (Hx ) (i)
Y (Hz )
SX = Xj ij , SZ = Zj ij . (2.71)
j j

For these operators to commute with each other, the two parity check matrices have to
satisfy

Hx HzT = 0, (2.72)

where 0 indicates a matrix with all elements = 0. To define the logical Z operators, we
define a quotient space Ker(Hx )/VHz := {c|Hx c = 0, c ∈ VCz }/VHz , where VHz is the
space spanned by the column vectors of HzT . Denoting the basis vectors of the quotient
space Ker(Hx )/VHz by {[bzk ]}, we define the logical Z operators
(k)
Y (bz )i
LZ = Zi k . (2.73)
i

Similarly, we can define the logical X operators

(k)
Y (bx )i
LX = Xi k , (2.74)
i

52
using the basis vectors {[bxk ]} of a quotient space Ker(Hz )/VHx , where bxk is chosen such
(i) (j) (j) (i)
that LZ LX = (−1)δij LX LZ . Note that dimensions of these kernel subspaces are the
same, and we can easily find such pairs of anticommuting logical operators. The above
stabilizer code constructed from two classical linear codes is called a Calderbank-Shor-
Steane (CSS) code.
Let us see an important example of CSS codes, the 7-qubit code introduced by Steane [Ste96].
Specifically, we utilize a classical linear code, the [[7, 4, 3]] Hamming code, whose generator
and parity check matrices are given by
 
1 0 0 0
 0 1 0 0 
   
 0 0 1 0  1 0 1 0 1 0 1
 
G=  0 0 0 1 , H =
  0 1 1 0 0 1 1 . (2.75)
 0 1 1 1  0 0 0 1 1 1 1
 
 1 0 1 1 
1 1 0 1

Because HH T = 0, we can employ the Hamming code to define both X- and Z-type
stabilizer generators:
S1 = I I I X X X X
S2 = I X X I I X X
S3 = X I X I X I X
(2.76)
S4 = I I I Z Z Z Z
S5 = I Z Z I I Z Z
S6 = Z I Z I Z I Z

There is an element (1, 1, 1, 1, 1, 1, 1)T in the quotient space Ker(H)/VH . The logical
operators are given by

LX = X X X X X X X,
(2.77)
LZ = Z Z Z Z Z Z Z.

The 7-qubit code is quite useful for fault-tolerant quantum computation. Both the X-
and Z-type stabilizer generators are defined from the Hamming code, and the stabilizer
group is invariant under the transversal Hadamard operation H̄ ≡ H ⊗7 . Moreover, the
logical X operator is mapped into the logical Z operator, H̄LX H̄ = LZ . Thus, the
transversal Hadamard operation acts as the logical Hadamard operation for the encoded
degree of freedom. Similarly, a transversal phase operation S̄ ≡ (ZS)⊗7 acts as a logical
phase operation, S̄LX S̄ † = LX LY . Furthermore, a transversal CNOT operation Λ̄(X) =
Λ(X)⊗7 keeps the stabilizer group of two logical qubits invariant:
(i) (i) (i) (i)
h{SX ⊗ I ⊗7 }, {SZ ⊗ I ⊗7 }, {I ⊗7 ⊗ SX }, {I ⊗7 ⊗ SZ }i (2.78)
(i) (i) (i) (i) (i) (i)
= h{SX ⊗ SX }, {SZ ⊗ I ⊗7 }, {I ⊗7 ⊗ SX }, {SZ ⊗ SZ }i. (2.79)

The logical Pauli operators are subject to the transformation rule of the CNOT gate. Ac-
cordingly, the transversal CNOT operation Λ̄(X) acts as a logical CNOT operation for the

53
encoded degree of freedom. Because the Hadamard, phase, and CNOT operations are im-
plemented transversally, whole Clifford group elements can be implemented by transversal
operations.
The transversal implementation is fault-tolerant because the operations apparently do
not increase the number of errors on a code block and there is no internal interaction
between the qubits in the same code block. Combined with a fault-tolerant gadget for
measuring the error syndrome as explained in Appendix A.1, we can implement Clifford
operations fault-tolerantly.
Unfortunately, the non-Clifford operation does not transform a Pauli operator into
another Pauli operator. For example, the π/8 operation e−i(π/8)Z transforms the Pauli X
operator into a Clifford operator:
√
e−i(π/8)Z Xei(π/8)Z = (X + Y )/ 2. (2.80)

This implies that a transversal non-Clifford operation hardly results in a logical non-
Clifford operation. Thus, a fault-tolerant non-Clifford gate operation is not so straightfor-
ward. To settle this, we can utilize magic state distillation, consisting of noisy non-Clifford
resource states and ideal (fault-tolerant) Clifford operations, as explained in the next sec-
tion.

2.8 Magic state distillation

2.8.1 Knill-Laflamme-Zurek protocol
A fault-tolerant implementation of a non-Clifford gate was first proposed in an earlier
paper by Knill, Laflamme, and Zurek [KLZ98b, KLZ98a]. Instead of implementing the
non-Clifford gate directly, we consider a fault-tolerant preparation of the non-stabilizer
state, the so-called magic state,

|Ai ≡ e−i(π/8)Y |+i. (2.81)

The magic state can be utilized to implement a non-Clifford gate A ≡ e−i(π/8)Y by using
one-bit teleportation consisting of Clifford gates and the Pauli basis state preparations and
measurements. Thus, if we can prepare a clean magic state, we can create a fault-tolerant
non-Clifford gate by using fault-tolerant Clifford gates.
The Knill-Laflamme-Zurek construction of the fault-tolerant preparation of the magic
state was based on the fact that |Ai is an eigenstate of H. The Hadamard operation has
transversality, e.g., on the 7-qubit code. Hence, if we perform a projective measurement

54
of H ⊗7 , we obtain a clean magic state. The circuit is given as follow

(2.82)

where A = e−i(π/8)Y and we used the fact that At Λc,t (X)A†t = Λc,t(H). The above
circuit consists of three parts, encoding of the logical magic state into the 7-qubit code,
indirect measurement of H ⊗7 , and decoding by one-bit teleportation. Note that all Clifford
operations are assumed to be ideal, because they are easily made fault-tolerant by using
a stabilizer code. With an appropriate randomization operation made up by Clifford
operations, a noisy magic state can be transformed into a mixture of |Ai and |A⊥ i = Y |Ai:

ρA = (1 − p)|AihA| + p|A⊥ ihA| (2.83)

Hence, the noise on the magic state can be expressed as a Y error. The Y error can
be detected by transversal Z measurements for the decoding (see also Knill’s gadget for
a fault-tolerant syndrome measurement in Appendix A.1). Assuming an ideal Clifford
operation, the error probability p decreases as O(p3 ), when we employ the 7-qubit code
with distance 3. Including the A gate for the initial encoding, we need 15 noisy magic
states |Ai to obtain a clean magic state. This distillation protocol works for all self-dual
CSS codes, which are symmetric between the X- and Z-type stabilizer generators and has
transversality for the Hadamard gate. Recently, improved protocols have been proposed
based on this approach [MEK12, Jon13b].

2.8.2 Bravyi-Kitaev protocol

Bravyi and Kitaev proposed another magic state distillation protocol based on a 15-qubit
code [BK05]. While their and the Knill-Laflamme-Zurek protocols seem quite different,
they are, interestingly, known to be equivalent [Rei05]. In the Bravyi-Kitaev protocol, a
[[15,1,3]] quantum code is defined by the [[15,7,3]] classical Reed-Muller code, whose parity

55
check matrix is given by
 
1 0 0 0 0 1 1 0 0 1 1 1 1 0 1
 0 1 0 0 1 0 1 0 1 0 1 1 0 1 1 
Hx =   0
 (2.84)
0 1 1 0 0 1 0 1 1 0 0 1 1 1 
0 0 1 0 1 1 0 1 0 0 1 0 1 1 1
(i) Q (H)ij
The X-type stabilizer generators are defined by SX = j Xj . Then, we choose a
parity check matrix of another classical code
 
0 0 1 1 0 0 0 1 0 0 0 0 0 0 1
 0 1 0 0 1 0 0 1 0 0 0 0 0 0 1 
 
 1 0 0 0 0 1 0 1 0 0 0 0 0 0 1 
 
 1 1 0 1 0 0 1 0 0 0 0 0 0 0 0 
 
 0 1 0 1 0 0 0 0 1 0 0 0 0 0 1 
Hz =  1 0 0 1 0 0 0 0
 (2.85)
 0 1 0 0 0 0 1 

 1 1 0 0 0 0 0 1 0 0 1 0 0 0 0 
 
 1 1 0 0 0 0 0 0 0 0 0 1 0 0 1 
 
 1 0 0 1 0 0 0 1 0 0 0 0 1 0 0 
0 1 0 1 0 0 0 1 0 0 0 0 0 1 0

so that Hx HzT = 0 and the Z-type stabilizer generators are defined similarly. The logical
operators are given by LX = X ⊗15 and LZ = Z ⊗15 . The logical states are written down
explicitly as
4
Y (i)
|0L i = (I + SX )|00...0i, (2.86)
i=1
Y4
(i)
|1L i = (I + SX )|11...1i. (2.87)
i=1

The number of 1s in each term of |0L i and |1L i is 8 and 7, respectively. By applying the
T = e−i(π/8)Z gate transversally, we obtain

T ⊗15 |0L i = eiπ/8 |0L i, (2.88)

⊗15 −iπ/8
T |1L i = e |0L i. (2.89)

Thus, the transversal T gate acts as a logical T † gate. Note that this transversality
does not hold in the orthogonal (erroneous) subspace, e.g., spanned by {Xk |0L i, Xk |1L i}.
However, we can show that this is enough to perform a fault-tolerant logical T gate.
Instead of applying the T gate directly, we implement it using a one-bit teleportation:

(2.90)

56
The Z-basis measurement and CNOT operation are both implemented transversally on
a CSS code. Thus, if the preparation of the non-Clifford ancilla state |T i = (|0L i +
√
e−iπ/4 |1L i)/ 2, called a magic state, is done fault-tolerantly, one can ensure fault-tolerance
of the logical T gate.
By using an appropriate randomization process, we can prepare a noisy magic state as
follows:

ρT = (1 − p)|T ihT | + pZ|T ihT |Z. (2.91)

Thus, a phase error Z is located on the ideal magic state with probability p. This phase
error causes a Z error after the T gate by one-bit teleportation. Because the code space
is invariant under the transversal T gate, we can detect such a Z error by the following
circuit.

The first part of the above circuit consisting of CNOT operations is an encoding circuit
for the quantum Reed-Muller code. The transversal T gate is applied by using one-bit
teleportation. The logical |T i state is measured in the X-basis transversally, which detects
Z errors on the code state, projecting the code state on the local X-basis. The input state
in the second lowest wire is entangled with the ancilla qubit in the lowest wire, where the
distilled magic state is teleported.
Q (c)
Let c be a 15-bit string specifying the location of the Z errors, E(c) ≡ i Zi i .
If E(c) commutes with the X-type stabilizer generators, the state passes through the
distillation circuit. To calculate this probability, we define a weight enumerator of a

57
subspace V ∈ GF (2n ),
X
WV (x, y) = xn−wt(c) y wt(c) . (2.92)
c∈V

The probability of passing the distillation circuit is calculated to be

1 1 + 15(1 − 2p)8
ppass = WV ⊥ (1 − p, p) = WVHx (1, 1 − 2p) = , (2.93)
Hx |Vx | 16

where the orthogonal subspace VH⊥x is equivalent to the kernel of Hx , Ker(Hx ). We also
used the MacWilliams identity [MS77]:
1
WV (x, y) = W ⊥ (x + y, x − y). (2.94)
|V | V
Similarly, the error probability of the output can be calculated to be
1
WVHz (p, 1 − p) = W ⊥ (1, 2p − 1) (2.95)
|VH⊥z | VHz
1 + 15(2p − 1)8 + 15(2p − 1)7 + (2p − 1)15
= . (2.96)
32
Accordingly, the error probability, under the condition of passing the distillation circuit,
is given by

1 + 15(2p − 1)8 + 15(2p − 1)7 + (2p − 1)15

p′ = = 35p3 + O(p4 ). (2.97)
2[1 + 15(1 − 2p)8 ]

If p′ > 0.141, we can reduce the error probability on the magic state via the distillation
√ l √
circuit. After l rounds of distillation, the error probability decreases to ( 35p)3 / 35. At
each round, we need 15 noisy magic states. Because the probability of successfully passing
the distillation circuit converges rapidly to 1, the average number of noisy magic states
consumed after l rounds becomes 15l . Accordingly, the average number of noisy magic
states required to achieve an error probability ǫ of the magic state scales like
√ √
[log( 35ǫ)/ log( 35p)]log(15)/ log(3) = O(log2.5 ǫ). (2.98)

In Sec. 2.4, we saw that if an input state is a convex mixture of the Pauli basis
states followed by Clifford operations and Pauli basis measurements, the measurement
outcomes can be simulated classically in the weak sense. The noisy magic state ρT =
(1 − p)|T ihT | + p|T ihT | lies on a line in the x–y plane, as shown in Fig. 2.5. If p >
√
(1− 2/2)/2 = 0.146, ρπ/8 lies inside the octahedron, and the Gottesman-Knill theorem is
applicable. On the other hand, if p < 0.141, magic state distillation allows us to implement
universal quantum computation with an arbitrary accuracy as seen above. Unfortunately,
magic state distillation based on the Reed-Muller code does not provide a tight distillation
threshold against the classically simulatable region. In Ref. [Rei05], a distillation protocol
√
using the 7-qubit code was proposed and achieved a tight threshold p = (1 − 2/2)/2. In

58
this sense, the classically simulatable and quantum universal regions are divided sharply
on the x-y plane.
By combining the magic state distillation and fault-tolerant Clifford operations on
the CSS code, we can perform universal quantum computation fault-tolerantly. In order
to make the error probability arbitrarily small, we can employ concatenated quantum
computation, in which logical qubits of a lower concatenation level are utilized as the
physical qubits at a higher level. At the higher level, all operations, including logical qubit
preparations and syndrome measurements for QEC, have to be done fault-tolerantly. If
the error probability is smaller than a certain constant value, which we call the noise
threshold, the logical error probability at the highest concatenation level decreases super-
exponentially. On the other hand, the overhead increases exponentially. Thus, we can
make the logical error probability small enough to maintain a quantum computation of
size N with a polylogarithmic overhead polylog(N ). This implies that we can obtain a
quantum benefit even for a quantum algorithm with a quadratic speedup, such as the
Grover algorithm [Gro96]. In Appendix A.1, we briefly review fault-tolerant syndrome
measurements, concatenated quantum computation, and the threshold theorem.
While the resource increment for protecting quantum computation scales polylogarith-
mically, its constant factor is quite huge. Almost all overheads for the fault-tolerant quan-
tum computation are employed for the magic state distillation [VMLFY10, JVMF+ 12].
Thus, much effort has been spent recently on developing resource-efficient magic state
distillation [MEK12, BH12, Eas13, Jon13a].

59
Figure 2.4: (a) A graph state and a measurement pattern. (b) The logical X operator of
input 1 is multiplied by the stabilizer generators and we obtain a correlated operator on
outputs 1 and 2 (left). The logical Z operator of input 1 is multiplied by the stabilizer
generators and we obtain the logical Z operator of output 1. The gray colored X operators
are replaced by ±1 depending on the measurement outcomes.

Figure 2.5: The noisy magic state in the Bloch sphere.

60
Chapter 3

Topological stabilizer codes

Protecting quantum information from decoherence is of prime importance to realize quan-

tum information processing. Several approaches have been proposed toward reliable
quantum information processing, ranging from passive to active protections, such as
decoherence-free subspaces [LCW98a], dynamical decoupling [VKL99b], and quantum er-
ror correction [Sho95]. Among these, the most comprehensive approach is fault-tolerant
quantum computation based on quantum error correction [DS96]. Quantum fault-tolerance
theory ensures scalable quantum computation with noisy quantum devices as long as the
error probability of such devices is smaller than a threshold value (see Appendix A.1).
The first threshold values were obtained, at almost the same time (1996), independently
by Aharonov et al., who used a polynomial code of distance five [ABO97, ABO08]; Knill
et al., who used the Steane seven-qubit code [KLZ98b, KLZ98a]; and Kitaev, who used
toric (surface) codes [Kit97]. All of these studies were based on a concatenated quan-
tum computation and achieved similar threshold values ∼ 10−6 . Since then, fault-tolerant
quantum computing has been studied as one of the most important issues in quantum
information science. In 2005, Knill proposed a novel scheme based on the C4 /C6 error-
detecting code, namely the Fibonacci scheme [Kni05], and achieved a considerably higher
threshold, a few %. Recently, Fujii et al. have further improved this approach by making
use of measurement-based quantum computation (MBQC) [RHG06, RHG07] on logical
cluster states, which gives the highest threshold value so far, ∼ 5% [FY10a, FY10b].
All these approaches rely on the availability of two-qubit gates between arbitrarily
separated qubits. However, all physically available interactions are local in space. More-
over, if we consider manufacturing convenience and the selective addressability of in-
dividual qubits, 2D nearest-neighbor architectures are necessary. It was thought that,
if we restrict the two-qubit gates to nearest-neighbor ones in 2D, then the threshold
value would decrease significantly [STD07]. The situation changed completely by the
proposal made by Raussendorf et al., i.e., topologically protected quantum computa-
tion [RH07, RHG07, RHG06]. They utilized a surface code, a kind of topological quan-
tum code, originally proposed by Kitaev as a toric code on a torus [Kit03] and inves-
tigated by Dennis et al. in detail [DKLP02]. All the stabilizer generators of the sur-

61
face code are spatially local, as will be seen later, and hence syndrome measurements
are done by using nearest-neighbor two-qubit gates. More importantly, Raussendorf
et al. developed a novel way to implement a fault-tolerant logical gate operation by
braiding defects on the surface, which can be done by using only single-qubit measure-
ments and nearest-neighbor two-qubit gates. Nevertheless, its threshold value is very high
∼ 1% [RH07, RHG07, RHG06, WFH11]. Based on these analyses, physical implementa-
tions and quantum architecture designs have recently been suggested [VMLFY10, LBSB10,
FT10, JVMF+ 12, FYKI12, LB12, NLB13, GFG12, FMMC12], which clarifies the experi-
mental requirements for building a fault-tolerant quantum computer. On the other side,
extensive experimental resources have been used to achieve these requirements, and very
encouraging results have already been obtained in several experiments in superconducting
systems [CGC+ 12, CGM+ 14, BKM+ 14a, CMS+ 14].
Another important motivation for studying topological quantum codes are their con-
nection with topologically ordered systems in condensed matter physics [Kit03]. Topo-
logical stabilizer codes, whose stabilizer generators are local and translation invariant,
provide toy models for topologically ordered systems. Using quantum coding theory with
a geometrical constraint such as locality and translation invariance, we can understand
the nature of topologically ordered condensed matter systems. One of the most important
issues in this direction is to answer whether or not thermally stable topological order exists
in three or lower dimensions [BT09, Yos11b].
When we consider the decoding problems of topological quantum codes, we encounter
classical statistical physics [DKLP02]. Specifically, the posterior probabilities for the error
correction correspond to partition functions of classical spin glass models [DKLP02], such
as the random-bond Ising model. If the corresponding random-bond Ising model lies inside
a ferromagnetic phase, the error correction on a surface code succeeds, and the logical error
probability decreases exponentially.
In this chapter, we introduce a representative example of the topological stabilizer
codes, the surface codes. Before that, we also briefly review the Z2 chain complex, which
is a useful tool to define the surface codes. We also apply it to a classical repetition
code as an exercise. After introducing the surface codes, we explain how topological error
correction is performed. We also mention the relation between topological error correction
and the classical spin glass model. Another example of topological stabilizer codes, the
topological color codes, is also presented. Finally, we explain the connection between
topological stabilizer codes and topologically ordered systems studied in condensed matter
physics.

3.1 Z2 chain complex

Before defining the surface codes, we introduce a mathematical tool, the Z2 chain complex,
which is useful for describing the surface codes.
Consider a surface G = (V, E, F ) that consists of the vertices V = {vk }, edges E = {el },
and faces F = {fm }. (Here we consider a graph G = (V, E) embedded on a 2D manifold,

62
the surface, on which the faces F are defined.) We define a vector space C0 over Z2 , using
each vertex vk ∈ V as a basis B(C0 ) = {vk }. A vector in C0
X
c0 = zk vk , (3.1)
k

where zk = {0, 1}, is referred to as a 0-chain. The vector space C0 is an Abelian group
under component-wise addition modulo 2. Similarly, we can define the Abelian groups C1
and C2 using the edges E = {el } and faces F = {fm } as base vectors B(C1 ) = {el } and
B(C2 ) = {fm }, respectively. The elements
X
c1 = zl el , (3.2)
l
X
c2 = zm fm , (3.3)
m

where zl , zm = {0, 1}, are called the 1-chain and 2-chain, respectively. With a mild abuse
of the notation, a set of i-dimensional elements (i.e., vertices, edges, and faces) is also
specified by a i-chain ci ∈ Ci as a set of elements having zj = 1.
We can define a homomorphism ∂i : Ci → Ci−1 such that

∂i ◦ ∂i−1 = 0. (3.4)

Specifically, ∂i ci is defined as an (i − 1)-chain that is the boundary of ci . For example,

∂el is a 1-chain for which zk = 1 if and only if (iff) vk is an endpoint of el . Thus, the
homomorphism ∂i is called a boundary operator. When there is no risk for confusion, we
will denote ∂i simply by ∂. A chain ci is called a cycle, if it is in a kernel of the boundary
operator ∂i , i.e., ∂ci = 0. For example, the boundary ∂2 fm of a face fm is a cycle, because
∂2 ◦∂1 = 0. Such a sequence of Abelian groups Ci , connected by the boundary operators ∂i
with ∂i ◦ ∂i−1 = 0, is called a Z2 chain complex. We define a trivial i-cycle ci if there exists
an (i + 1)-chain such that ci = ∂ci+1 , i.e., ci ∈ Img(∂i+1 ). By regarding the trivial cycles
as generators of the continuous deformation, we may define an equivalent class having the
same topology. More precisely, a homology group hi is defined as a quotient group formed
by the cycles and the trivial cycles:

hi = ker(∂i )/Img(∂i+1 ). (3.5)

An element of the quotient group hi is called a homology class. If two i-chains ci and
c′i belong to the same homology class, there exists an (i + 1)-chain ci+1 such that ci =
c′i + ∂ci+1 . Two such i-chains ci and c′i are said to be homologically equivalent.
We also define the dual surface Ḡ = (V̄ , Ē, F̄ ), where we identify the elements of the
dual and original (primal) lattices such that V̄ = F , Ē = E and F̄ = V . Specifically, the
dual lattice is constructed such that the two vertices v̄, v̄ ′ ∈ V̄ are connected by an edge
ē, if the corresponding two faces f and f ′ share the same edge e. We can also define a Z2
chain complex on Ḡ using the dual bases B̄(Ci ), a dual i-chain c̄i ∈ C̄i and a boundary
operator ∂¯i : C̄i → C̄i−1 . When there is no risk of confusion, ∂¯i will be denoted by ∂.

63
 In the construction of the surface codes, a qubit is defined on each edge el ∈ E of the
surface G (or equivalently each dual edge ēl of the dual surface Ḡ). The Pauli product is
P
defined using a 1-chain c1 = l zl el (or dual 1-chain c̄1 ) such that
Y z
W (c1 ) = Wl l , (3.6)
l

where Wl ∈ {Xl , Yl , Zl } is a Pauli operator acting on the qubit on edge el . Specifically, we

have

W (c1 )W (c′1 ) = W (c1 + c′1 ). (3.7)

Consider the two operators X(c1 ) and Z(c′1 ), defined by the two 1-chains c1 and c′1 ,
respectively. The commutability of these two operators is determined by the inner product
P
of the two chains (vectors) c1 · c′1 ≡ l zl zl′ , where the addition is taken modulo 2:

c1 · c′1 = 0, iff X(c1 )Z(c′1 ) = Z(c′1 )X(c1 ) (commute), (3.8)

c1 · c′1 = 1, iff X(c1 )Z(c′1 ) = −Z(c′1 )X(c1 ) (anticommute). (3.9)

Let M (∂i ) be a matrix representation of ∂i with respect to the basis vectors B(Ci ) and
B(Ci−1 ). We have

(M (∂i )ci ) · ci−1 = ci · (M (∂i )T ci−1 ). (3.10)

By identifying the primal and dual bases B(C0 ) = B̄(C2 ), B(C1 ) = B̄(C1 ), and B(C2 ) =
B̄(C0 ), the duality relation between primal and dual lattices can be expressed by

M (∂1 ) = M (∂¯2 )T , M (∂2 ) = M (∂¯1 )T , (3.11)

where [·]T indicates the matrix transpose.

Using Eq. (3.4) ∂1 ◦ ∂2 = 0, we obtain M (∂¯2 )T M (∂2 ) = 0. Thus, for any primal and
dual 2-chains, c2 and c̄2 , we have

¯ 2 · ∂c2 = 0,
∂c̄ (3.12)

¯ 2 ) always commute. This property is useful to construct a stabilizer

i.e., X(∂c2 ) and Z(∂c̄
group, because stabilizer generators have to commute with each other. Using Eqs. (3.10)
and (3.11), we also have

c1 · ∂¯2 c̄2 = ∂c1 · c̄2 . (3.13)

3.2 A bit-flip code: exercise

As an exercise, we define a bit-flip code (classical repetition code) using the Z2 chain
complex. (Readers who are familiar with this topic may skip this section.) Consider
a regular polygon G(V, E, F ) on a sphere consisting of n = |E| edges and two faces

64
Figure 3.1: (a) A regular
Q polygon embedded on a sphere. (b) A stabilizer operator defined
as Ak = Z(∂ f¯k ) = δvk Zk . The logical X operator is given by LX = X(∂f1 ). Because
primal and dual 1-chains share an even number of qubits, LX and Ak commute with each
other.

F = {f1 , f2 } (corresponding to the top and bottom hemispheres) as shown in Fig. 3.1.
The number of qubits is n.
We define a stabilizer generator for each dual face f¯k = vk as follows:
Y
Ak = Z(∂ f¯k ) = Zl , (3.14)
el ∈δvk

where, for convenience, δvk ≡ ∂ f¯ is defined as a set of edges that are incident to the vertex
vk . (Note that both the dual and primal objects are identified.)
Q
Because vk ∈V Ak = I, there are n − 1 independent stabilizer generators. The di-
mension of the stabilizer subspace is two. The code subspace is described by a one-cycle
c1 = ∂f1 = ∂f2 surrounding the sphere. We define the logical X operator

LX = X(∂f1 ) = X(c1 ). (3.15)

Note that LX = X(∂f1 ) and Ak = Z(∂ f¯k ) commute, due to Eq. (3.12). The logical Pauli
Z operator, which commutes with {Ak } and anticommutes with LX , is defined as LZ = Zl .
(We may choose any of the Zl s on the edge el , because they act equivalently on the code
space.) The code is a bit-flip code, which protects quantum information against bit flip
errors.
A string of bit errors X(c1 ) is defined using a 1-chain c1 , which we call an error chain.
The error is detected by measuring the eigenvalues of the stabilizer generators Ak =
Z(δvk ), i.e., through syndrome measurements. From Eq. (3.13), we have c1 · δvk = ∂c1 · vk .
Thus, X(c1 ) anticommutes with the stabilizer generator Ak if ∂c1 · vk = 1. Hence, X(c1 )
anticommutes with Ak on the vertex vk that is a boundary vk ∈ ∂c1 of the error chain
c1 . The eigenvalue of the stabilizer generator becomes -1. Error correction is the task of
finding a recovery chain c′1 such that ∂(c1 + c′1 ) = 0. In the case of the bit-flip code, there

65
Figure 3.2: An X error chain X(c1 ) and the corresponding error syndrome ∂c1 . The
eigenvalue of the stabilizer generator becomes −1 on vertices in ∂c1 .

are only two possibilities c1 + c′1 = 0 or c1 + c′1 = ∂f1 . In the former case, error correction
succeeds, while the latter case results in a logical error LX .
The bit-flip code cannot protect quantum information against phase errors. A natural
extension of the classical repetition code to handle both bit and phase errors is the surface
codes introduced in the next section.

3.3 Definition of surface codes

3.3.1 Surface code on a torus: toric code
Now we are ready to define the surface code. Let us consider a square lattice G = (V, E, F )
on a torus with a periodic boundary condition as shown in Fig. 3.3. A dual square lattice
Ḡ = (V̄ , Ē, F̄ ) is also defined on the torus. We define stabilizer generators of the Z- and
X-types for each face fm and vertex vk as follows:

Am = Z(∂fm ), Bk = X(δvk ) = X(∂ f¯k ). (3.16)

Because ∂fm · ∂ f¯k = 0 from Eq. (3.12), Am and Bk commute. The stabilizer generators
Am and Bk are called plaquette and star operators , respectively. By the definition of the
stabilizer code, the code state |Ψi satisfies

Am |Ψi = |Ψi, Bk |Ψi = |Ψi, (3.17)

for all fm ∈ F and vk ∈ V .

If two 1-chains, c1 and c′1 , are homologically equivalent, the actions of Z(c1 ) and Z(c′1 )
on the code state are the same, because there exists a 2-chain c2 such that
 
Y
Z(c′1 ) = Z(c1 )Z(∂c2 ) = Z(c1 )  Am  . (3.18)
fm ∈c2

66
Figure 3.3: A square lattice on a torus with a periodic boundary condition. The Z and
X stabilizer generators are defined by being associated with each face and vertex (dual
face). The logical Z and X operators are defined by non-trivial cycles.

We will denote this simply by Z(c′1 ) ∼ Z(c1 ). This is also the case for homologically
equivalent dual 1-chains, c̄1 and c̄′1 , and the actions of X(c̄1 ) and X(c̄′1 ) on the code state.
Hence, the homology classes, which are equivalent classes over trivial cycles, of the primal
and dual 1-chains correspond to the actions of Z- and X-type operators on the code state,
because their actions are equal up to stabilizer operators.
Let us define the logical operators Z(c1 ) and X(c̄1 ). The logical operators have to com-
mute with all stabilizer generators and also be independent of them. The former condition
implies that ∂c1 = 0 and ∂c̄1 = 0 from Eq. (3.13). This is because the commutability im-
poses that c1 · δvk = ∂c1 · vk = 0 for all vertices k and c̄1 · ∂fm = ∂c̄1 · v̄m = 0 for all
faces (dual vertices) m. The latter condition implies that c1 and c̄1 are nontrivial cycles,
because for two homologically equivalent cycles c1 and c′1 , the actions of Z(c1 ) and Z(c′1 )
on the code state are the same as seen in Eq.(3.18). We can find two non-trivial cycles,
which belong to different homology classes, for each primal and dual 1-chain, as shown in
Fig. 3.3. (This is a natural consequence, because the homology group h1 on the torus is
Z2 × Z2 .) Then, we define two pairs of logical Pauli operators
(1) (1) (1) (1) (2) (2) (2) (2)
{LZ = Z(c1 ), LX = X(c̄1 )} and {LZ = Z(c1 ), LX = X(c̄1 )}. (3.19)
(i) (j)
Note that LZ and LX anticommute with each other if i = j. Otherwise, they commute:
(i) (j) (j) (i)
LZ LX = (−1)δij LX LZ . (3.20)

That is, they satisfy a commutation relation equivalent to that of the Pauli operators for
two qubits. The logical Pauli basis states are defined as follows:
(i)
LZ |ΨZ (s1 , s2 )i = (−1)si |ΨZ (s1 , s2 )i, (3.21)
(i) si
LX |ΨX (s1 , s2 )i = (−1) |ΨX (s1 , s2 )i. (3.22)

67
The number of stabilizer generators of the n × n square lattice on the torus is given by

|F | + |F̄ | − 2 = |F | + |V | − 2 = 2n2 − 2, (3.23)

Q Q
where −2 comes from the fact that fm ∈F Am = I and vk ∈V Bk = I, and hence there
are two non-independent operators. On the other hand, the number of qubits is given by

|E| = 2n2 . (3.24)

Thus, we have a 2|E|−(|F |+|V |−2) = 22 -dimensional stabilizer subspace. The above two
pairs of logical operators appropriately describe the degrees of freedom in the code space.
The code distance, the minimum weight of the logical operators, is determined by the
linear length of the square lattice. Thus, the surface code on the torus is an [[n2 , 2, n]]
stabilizer code.
In short, an operator that commutes with the stabilizer generators corresponds to a
cycle, i.e., ker(∂1 ). The stabilizer operators correspond to the boundaries of the 2-chains
c2 ∈ C2 . The logical operators, which commute with and are independent of the stabilizer
generators, correspond to the homology class h1 = ker(∂1 )/Img(∂2 ). We summarize the
correspondence between the surface code and the Z2 chain complex in Tab. 3.1.

Table 3.1: The correspondence between Z2 chain complex and stabilizer codes.

stabilizer code chain complex

Z-type stabilizer generator face fm
X-type stabilizer generator vertex vk (dual face f¯k )
Z- and X-type stabilizer operators boundaries of 2-chains,
Img(∂2 ) and Img(∂¯2 )
commutability cycle conditions: ∂c1 = 0 and ∂c̄1 = 0
the operators commuting with stabilizer generators ker(∂1 ) and ker(∂¯1 )
the logical Z and X operators nontrivial cycles c1 and c̄1
(homology classes [c1 ] = ker(∂1 )/Img(∂2 )
and [c̄1 ] = ker(∂¯1 )/Img(∂¯2 ))
Z and X errors 1-chains c1 and c̄1
Z and X error syndromes ∂c1 and ∂c̄1

All properties so far hold for a general tilling G = (V, E, F ), and hence we can define
a surface code on general, e.g., triangular and hexagonal, lattices [FT12]. Moreover, the
numbers of edges, faces, and vertices are subject to the Euler characteristic formula

|F | + |V | − |E| = 2 − 2g, (3.25)

where g is the genus, the number of “handles” of the surface. Thus, the dimension of
the stabilizer subspace is calculated to be 2|E|−(|F |+|V |−2) = 22g . These are the degrees of
freedom equivalent to 2g qubits in the stabilizer subspace.

68
Figure 3.4: A planar surface code. Top and bottom are smooth boundaries consisting of
three-qubit star operators. Left and right are rough boundaries consisting of three-qubit
plaquette operators. At the boundary, we can define sets of 1-chains Γ1 = {γ1 } ∈ C1
and Γ̄1 = {γ̄1 } ∈ C̄1 for the primal and dual lattices, respectively, by which the relative
homology is defined.

Because almost all arguments are made from the operator viewpoint (Heisenberg pic-
ture), there is no need to write down |ΨA (s1 , s2 )i explicitly. However, for the readers who
worry about such things, we will give an explicit description from the state viewpoint:
 
(1) (2)
Y I + Bk  ⊗n2
|ΨZ (s1 , s2 )i = N Z(c1 )s1 Z(c1 )s2  |0i (3.26)
2
vk ∈V
 
(1) (2)
Y I + Am  2
|ΨX (s1 , s2 )i = N X(c̄1 )s1 X(c̄1 )s2  |+i⊗n (3.27)
2
fm ∈F

where N is a normalization factor. The code state can be viewed as an equal weight
superposition of all cycles belonging to the same homology class.

3.3.2 Planar surface code

The periodic boundary condition might be hard to implement in experiments. We can
also define a surface code on a planar n × (n − 1) square lattice with an appropriate
boundary condition as shown in Fig. 3.4. The top and bottom boundaries consist of
three-qubit star operators, which we call smooth boundaries, because they are complete
plaquette operators. On the other hand, the left and right boundaries consist of three-
qubit plaquette operators, which we call rough boundaries. The X operators on a dual
1-chain can terminate at the smooth boundary, while the Z operators on a 1-chain can
terminate at the rough boundary. Thus, we should use a relative homology to define logical
operators, where two chains ci and c′i are said to be (relative) homologically equivalent iff

69
Figure 3.5: A planar surface code with a defect. The stabilizer generators are not defined
inside a defect, which provides a nontrivial cycle wrapping around it.

c′i = ci +∂ci+1 +γi with γi ∈ Γi ⊂ Ci . In this case, Γ1 is chosen, specifically, to be the vector
space spanned by the set of 1-chains each of which consists of three edges corresponding to
the three-qubit plaquette operator at the top and bottom smooth boundaries. Similarly,
we also define Γ̄1 at the left and right rough boundaries of the dual lattice. Because
Z(γ1 ) and X(γ̄1 ) are both stabilizer operators, if the shapes of two logical operators are
homologically equivalent, their actions on the code state are the same. The number of
edges, faces, and vertices are now |E| = 2n2 −2n+1, |V | = n2 −n, and |F | = n2 −n. Thus,
the stabilizer subspace is a 2D subspace. We can define the logical operators LZ = Z(c1 )
and LX = X(c̄1 ) using horizontal and vertical 1-chains c1 and c̄1 , as shown in Fig. 3.4.
If one chooses all boundaries smooth, the stabilizer state is uniquely defined, and
hence there is no logical degree of freedom. By punching a hole, we can introduce a
defect on the surface, whereby we can define a nontrivial cycle for a logical Z operator, as
shown in Fig. 3.5. The logical X operator can be chosen as a dual 1-chain that connects
the defect and the smooth boundary, as shown in Fig. 3.5. As mentioned previously,
the properties of the logical operators are the same if the corresponding 1-chains are
homologically equivalent.

3.4 Topological quantum error correction

Let us return to the surface code on the torus to explain how errors are corrected. There
2
are 22n −2 orthogonal subspaces, the so-called syndrome subspaces, each of which is an
eigenspace of the stabilizer generators and has the same structure as the code space. These
orthogonal subspaces are utilized to identify the location of errors and to infer a recovery
operation. Suppose the X and Z errors X(c̄e1 ) and Z(ce1 ), defined by error 1-chains c̄e1
and ce1 , respectively, occur on the surface code, as shown in Fig. 3.6. The code state is
now mapped into one of the orthogonal subspaces. From Eq. (3.13), X(c̄e1 ) and Z(ce1 )

70
Figure 3.6: (a) A Z error chain Z(c1 ) is detected at the boundary ∂c1 . The eigenvalue of
the stabilizer generator Bk on the face f¯k ∈ c̄2 = ∂c1 becomes -1. (b) An X error chain
X(c̄1 ) is detected at the boundary ∂c̄1 . The eigenvalue of the stabilizer generator Am on
the face fm ∈ c2 = ∂c̄1 becomes -1.

anticommute with Am and Bk on the face fm ∈ ∂c̄e1 and vertex vk ∈ ∂ce1 . Thus, the
orthogonal subspace is specified by error syndromes

∂c̄e1 ≡ cs2 , ∂ce1 ≡ cs0 . (3.28)

More precisely, the eigenvalues with respect to the stabilizer generators Am and Bk are
s s P s P s
given by (−1)zm and (−1)zk , where cs2 = m zm fm and c0 =
s
k zk vk (see Fig. 3.6),
respectively. Error correction is the task of finding recovery 1-chains c̄r1 and cr1 such that
∂(c̄e1 + c̄r1 ) = 0 and ∂(ce1 + cr1 ) = 0, meaning that the state is returned into the code space
by applying X(c̄r1 ) and Z(cr1 ), respectively. Below we will, for simplicity, explain only how
to correct the Z errors, but the extension to the X errors is straightforward.
Suppose each Z error occurs with an independent and identical probability p. Con-
ditioned on the error syndrome cs0 = ∂ce1 , the posterior probability of an error Z(c1 )
P
occurring with c1 = l zl el is written as
Y  p z l
s
P (c1 |c0 ) = N |∂c1 =cs (3.29)
1−p 0
l

where N is a normalization factor. One way to find a recovery chain is by maximizing the
posterior probability
!
X
r s
c1 ≡ arg max P (c1 |c0 ) = arg min zl |∂c1 =cs0 , (3.30)
c1 c1
l

71
Figure 3.7: From left to right, error chain Z(ce1 ), recovery chain Z(cr1 ) with cr1 = cs0 , and
net actions Z(cr1 + ce1 ). If cr1 + ce1 is a trivial cycle (top), error correction succeeds. If cr1 + ce1
is a non-trivial cycle (bottom), the error correction results in a logical error.

P
where c1 = l zl el . As seen in the l.h.s., this corresponds to minimizing the number of
P r
errors l zl such that ∂cr1 = cs0 . Hence, this is called minimum distance decoding. If
cr1 + ce1 is a trivial cycle, the error correction succeeds, because the net action Z(cr1 + ce1 )
after the recovery operation is identity on the code state as shown in Fig 3.7. If cr1 + ce1 is
a nontrivial cycle, the recovery operation results in the logical operation Z(cr1 + ce1 ) ∼ LZ ,
i.e., a logical error.
Minimum distance decoding is hard in general, because it can be mapped into an inte-
ger programing problem, which is NP-hard [HLG11, KL12, IP13]. However, in the present
case, there is a nice geometrical property that makes the decoding problem feasible. The
condition ∂c1 = cs0 and minc1 read that it is sufficient to find a 1-chain that connects
pairs of two vertices in cs0 with a minimum Manhattan length. There is a classical poly-
nomial time algorithm to do such a task, the so-called minimum-weight perfect matching
(MWPM) algorithm of Edmonds [Edm65, Bar82, CR99]. The algorithm scales like O(n6 ),
with n being the linear length of the lattice and with the fixed error probability p. Typ-
ical examples of the error chain ce1 and the error plus recovery chain ce1 + er1 are shown
in Fig. 3.8 for each p = 0.05 (b), p = 0.10 (c), and p = 0.15 (d). Here, we employ an
implementation of MWPM, blossom V [Kol09]. The higher the physical error probability
is, the longer the error plus recovery chain becomes. For a high physical error probability
p = 0.15, the error plus recovery chain becomes a large cycle, and unfortunately results
in a logical error. Such a logical error probability is plotted as a function of the physical
error probability p in Fig. 3.8 (a) for each n = 10 (solid line), n = 20 (dashed line), and
n = 30 (dotted line). If the error probability is sufficiently smaller than a threshold value,
the logical error probability decreases for increasing n. The threshold value for decoding
by the MWPM algorithm has been estimated to be 10.3% (MWPM) [DKLP02, WHP03].
The minimum distance decoding with MWPM is not optimal for our purpose, i.e.,

72
Figure 3.8: (a) The logical error probability is plotted as a function of the physical error
probability p. Error chains ce1 and error plus recovery chains ce1 + cr1 for p = 0.05 (b),
p = 0.10 (c), and p = 0.15 (d).

making the logical error probability as small as possible. An error correction with another
′ ′
recovery chain cr1 , for which ce1 +cr1 belongs to the same homology class as ce1 +cr1 , provides
′
exactly the same result. We may use such a recovery chain cr1 to correct the error. Thus,
we should maximize, not each posterior probability p(cr1 |cs0 ), but a summation of it over
the same homology class. This problem originated from the degeneracy of the surface
code, where each syndrome is assigned not uniquely, but for many error instances. A
prototypical example of an error syndrome, for which we should consider not olny the
weight of errors, but also combinatorics (an entropic effect) of the error configurations, is
shown in Fig. 3.9.
Denoting the homology class by hi , the posterior probability for a homology class hi
is given by
X ′
pi = P (cr1 |cs0 ), (3.31)
′ ′
cr1 |cr1 +cr1 ∈hi

where cr1 is a recovery chain satisfying ∂cr1 = cs0 and chosen arbitrarily as a reference frame,
′ ′
and the summation is taken over all 1-chains cr1 such that cr1 + cr1 belongs to the homology
class hi .
The posterior probability may be rephrased by using the stabilizer language as follows
(see also Appendix B). Let G and L be the stabilizer and logical operator groups, respec-
tively. For a given error syndrome cs0 , we define the recovery operator Z(cr1 ) a priori, such

73
Figure 3.9: Suppose the top and bottom boundaries are connected by a periodic boundary
condition. The minimum-weight error is shown in the left panel. The error shown in the
right panel is not minimum-weight but has a fourfold degeneracy. The recovery chain
should be chosen by comparing the total error probabilities p3 and 4 × p4 .

that the erroneous state is returned into the code space. We can decompose an arbitrary
error operator Z(ce1 ), providing the syndrome cs0 , into

Z(ce1 ) = Z(cr1 )GLi , (3.32)

where G ∈ G and Li ∈ L are a stabilizer and logical operator, respectively. The posterior
probability of the logical operator Li is calculated by summing over all stabilizer operators
G∈G
1 X
pi = P (Li |cs0 ) = P [Z(cr1 )GLi ]. (3.33)
N
G∈G

We choose the most likely homology class hi or, equivalently, the most likely logical oper-
ator Li that maximizes the probability pi = P (Li ):

Lī ≡ arg max P (Li |cs0 ). (3.34)

Li

The error correction is completed by applying Lī Z(cr1 ), and the logical error probability
is given by 1 − pī .
In the next section, we relate the posterior probability summed over the same homology
class to a partition function of the random-bond Ising model. The partition function of the
Ising model on a planar graph with general coupling strengths (without magnetic fields)
can be calculated in polynomial time using the Kasteleyn-Barahona algorithm with the
Pfaffian method [Kas61, Bar82, BR07]. Thus, the optimal decoding is also implemented
by a polynomial time classical processing, though it takes more overhead than MWPM.
The threshold value for optimal decoding is also discussed in the next section.
Efficient decoding has been one of the most important issues for the realization of fault-
tolerant quantum computing, because the coherence time of quantum information would

74
Figure 3.10: (a) A unit cell of the level-1 logical qubit. Two pairs of logical Pauli operators
(top). The 6 stabilizer operators and the pure error operators, each of which anticommutes
with an stabilizer operator (middle). The 4 pairs of edge operators. These 12 pairs of
mutually anti-commuting operators generate the Pauli group of the 12 qubits. (b) Using
the level-1 logical operators, the level-2 stabilizer generators are defined. (c) Using 12
level-1 logical qubits, a unit cell of the level-2 logical qubit is defined.

be very short; a fast classical processing is essential. Fowler et al. proposed an efficient
decoding method based on MWPM [FWH12]. Because a long error chain is exponentially
suppressed, we can assign weights between each pairs of vertices having −1 eigenvalues
according to their Manhattan length. This allows us to reduce the number of edges from
O(n4 ) to O(n2 ). Moreover, the exponential suppression of longer error chains allows us to
search the pairs within a local small region; matching a long-distance pair is exponentially
rare. Because the matching process employs almost exclusively local information, this
algorithm can be parallelized to an O(1) average time per round.
Another decoding method is using a renormalization technique [DCP10a, DCP10b]. As
explained in Appendix B, we can efficiently execute an optimal decoding on a concatenated
quantum code by using a brief-propagation on a tree factor graph [Pou06]. Although the
surface code itself does not have such a hierarchal structure, a renormalization technique
is employed to introduce a hierarchal structure on the surface code [DCP10a, DCP10b].
Using 12 qubits as a unit cell, we define a couple of level-1 logical qubits, as shown
in Fig. 3.10 (a), which include 2 pairs of logical operators L(1) , 6 stabilizer generators
G (1) . We define 6 pure error operators Ḡ (1) each of which anticommutes with a stabilizer
generator as shown in Fig. 3.10 (a) (middle). Moreover, we define 4 pairs of anticommuting
operators as shown in Fig. 3.10 (a) (bottom), which we call edge operators E (1) , to generate
the Pauli group of the 12 qubits. Any Pauli operator A on the unit cell can be decomposed
in terms of these operator,

A = L(1) G(1) Ḡ(1) E (1) , (3.35)

75
where B (1) ∈ B (1) for B = L, G, Ḡ, E. The pure error operator Ḡ(1) is chosen uniquely
according to the error syndrome S (1) to return the state into the code space. (If a stabilizer
generator G1 has an eigenvalue −1, then we employ Ḡ1 , which anticommutes with G1 .)
From the error distribution P (A), the posterior probability of the level-1 logical operator
is calculated by taking a marginal over G (1) and E (1) ,
X
P (L(1) |S (1) ) = P (A = L(1) G(1) Ḡ(1) E (1) |S (1) ), (3.36)
G(1) ∈G (1) ,E (1) ∈E (1)

which are utilized to model the error distribution at the level 2.

Similarly, a level-k unit cell is defined by using 8 level-(k − 1) unit cells and the 12 pairs
of the level-(k − 1) logical operators on them. Similarly to the level-1 case, we define the
level-k logical, stabilizer, pure error, edge operators. At the highest level k = l, we obtain
the logical operators of the surface code. For example, the level-2 stabilizer generators are
shown in Fig. 3.10 (b). A unit cell of the level-2 logical qubit is shown in Fig. 3.10 (c).
The posterior probability of the level-k logical operator P (L(k) |S (k) ) is calculated us-
ing the posterior probabilities at the level (k − 1) by assuming that they are independent
for each level-(k − 1) unit cell. Under this assumption, we can calculate the posterior
probability P (L(l) |S (l) ) conditioned on all error syndrome S (l) by using the belief propa-
gation [Pou06, DCP10a, DCP10b] (see Appendix B for decoding the concatenated codes
by using the belief propagation). Then the maximization procedure over all logical oper-
ators provide us a most likely logical operator,

L∗ = arg max P (L(l) |S (l) ). (3.37)

L(l)

The level-(k −1) unit cells share physical qubits with each other, and hence the conditional
logical error probability is not independent for level-(k − 1) unit cells. If we employ a
message passing in order to reweigh the level-(k − 1) logical error probability as the error
model for the level-k unit cell, we can further improve the approximation of P (L(l) |S (l) ).
If the decoding process is done in parallel, the belief propagation takes O(log2 (n)) time,
similar to the case of the concatenated code as explained in Appendix B. Because the
posterior probability is approximated, the decoding based on the renormalization is not
optimal, but results in a reasonable threshold value, ∼ 9% for the independent X and
Z error and ∼ 15.2% for the depolarizing error (against 10.3% and 15.5% obtained by
MWPM, respectively) [DCP10a, DCP10b]. The renormalization decoder can also be
applied for an arbitrary topological code [BDCP12].

3.5 Error correction and spin glass model

The behavior of the logical error probability in Fig. 3.8 suggests the existence of a critical
phenomenon behind the error correction problem. Indeed, there is a beautiful correspon-
dence between quantum error correction on the surface code and a spin glass model, the

76
so-called random-bond Ising model (RBIM) [DKLP02]. More precisely, the posterior prob-
ability Eq. (3.31) of the logical operator is mapped into a partition function of the RBIM
as seen below.
To solve the condition ∂c1 = cs0 in Eq. (3.29), we rewrite the recovery chain as c1 =
(k)
c1 +ct1 , where ct1 ∈ Img(∂2 ) is a trivial cycle. cr1k is further decomposed into cr1k = ce1 +c1 ,
rk
(k)
where ce1 determines the actual location of errors, and c1 is a (nontrivial) cycle belonging
to the homology class hk . Note that this decomposition corresponds to Eq. (3.32). The
posterior probability is rewritten as
r
Y  p zlt ⊕zl k
P (c1 |cs0 ) = N , (3.38)
1−p ct1 ∈Img(∂2 )
l

where cα1 = zlα el , with zlα ∈ {0, 1}. In order to take the condition ct1 ∈ Img(∂2 ) auto-
g
matically, we introduce a gauge valuable zm ∈ {0, 1} on each dual vertex v̄m and a dual
P
0-chain c̄g0 = m zm g g
v̄m . Any trivial cycle ct1 is replaced by the gauge valuables using the
following relation
M
zlt = g
zm . (3.39)
v̄m ∈∂ ēl

For example, if we choose c̄g0 = v̄m , we obtain ct1 = ∂fm . (This corresponds to a multi-
plication of the face stabilizer generator defined on the face fm .) There is a one-to-one
correspondence between a trivial cycle and a gauge dual 0-chain. Using cg0 we can formally
solve the condition ct1 ∈ Img(∂2 ) in Eq. (3.29),
r L g
Y  p z l k v̄m ∈∂ ēl zm
s
P (c1 |c0 ) = N . (3.40)
1−p
l

The binary valuables zi ∈ {0, 1} are transformed into spin variables σi ∈ {+1, −1} by
p
σi = (−1)zi . Moreover, we define a coupling constant e−J = p/(1 − p). Then Eq. (3.40)
is rewritten as
P r g g
J σl k σm(l) σm
P (c1 |cs0 ) = N ′ e l ′ (l)
, (3.41)

where N ′ is a normalization factor and m(l) and m′ (l) are end points of the edge el .
By changing the notation l → ij and m(l), m′ (l) → i, j, the posterior probability is
reformulated as a Boltzmann factor of the ±J RBIM:
P (k)
Jij σig σjg
P (c1 |cs0 ) = N ′ e ij , (3.42)
(k) rk
where Jij = Jσij . On the edges where the errors are located, anti-ferromagnetic in-
teractions are assigned. The posterior probability of the logical operator Eq. (3.31) is
calculated by taking summation over all gauge spin configurations (this corresponds to

77
the summation over all stabilizer operators in Eq. (3.31)):
X ′
pk = P (cr1 |cs0 ) (3.43)
′ r ′
cr1 |c1k +cr1 ∈hk
X
= P (cr1k + ct1 |cs0 ) (3.44)
ct1 |ct1 ∈Img(∂2 )
X P (k) g g
= N e′ ij Jij σi σj . (3.45)
{σig }
(k)
= N ′ Z({Jij }), (3.46)
(k)
where Z({Jij }) is the partition function of the ±J RBIM.
Let us consider the performance under this decoding by taking an average of the
logarithm of pk over the error distribution, which corresponds to a sample average with
respect to quenched randomness:
X
p̄ln
k ≡ [ln(pk )] = P (ce1 )pk (3.47)
ce1
X (k)
= P ({Jije }) ln Z({Jij }) + ln N ′ (3.48)
{Jij }

≡ −Fk + ln N ′ , (3.49)

where [·] indicates the sample average and p̄ln

k is the sample average of the logarithm of
the logical error probability pk . The quenched randomness is determined by the error
distribution

P (Jije ) = (1 − p′ )δ(Jije − 1) + p′ δ(Jije + 1). (3.50)

Note that p is a parameter in the posterior probability and could be different from the
actual error probability p′ . Of course, optimal decoding, in the sense of Bayesian inference,
is achieved by maximizing the posterior probability with a true error probability p = p′ .
(0) (k)
F0 = −[ln Z({Jij })] is the free-energy of the RBIM, while Fk = −[ln Z({Jij })] (k =
(k)
1, 2, 3) is the free-energy with respect to the interactions {Jij }, where a domain wall
(k)
of anti-ferromagnetic interactions corresponding to a cycle c1 of a homology class hk is
inserted.
From Eqs. (3.46) and (3.49), the relation between quantum error correction and a
spin glass model becomes apparent; the posterior probability of the logical operator is
proportional to the partition function of RBIM, whose Hamiltonian is given by H =
P (k) (k)
− ij Jij σig σjg . The location of the Z errors, represented by Jij = −1, corresponds to
the anti-ferromagnetic interaction due to disorder as shown in Fig. 3.11 (see also Tab. 3.2).
The error probability and the coupling strength (inverse temperature) are related by e−J =
p
p/(1 − p). The error syndrome cs0 = ∂ce1 corresponds to the distribution of frustrations
of the Ising interactions and also the end points (Ising vortex) of the excited domain walls.
The probability of the homology class is expressed by the domain-wall free energy:

78
Figure 3.11: A gauge spin σfg located at the face-center of a plaquette. The distribution
of anti-ferromagnetic interactions corresponds to the error chain ce1 . The error syndrome
cs0 = ∂ce1 corresponds to the distribution of frustrations. The ground state configuration
is determined by a domain-wall consisting of excited domain-wall and anti-ferromagnetic
bonds, both of which have frustrations at their end points.

− (p̄ln ln
k − p̄0 ) = Fk − F0 . (3.51)

If the physical error probability p is smaller than the threshold value, −p̄lnk (k = 1, 2, 3)
diverges in the large n limit. On the other hand, if p is higher than the threshold value,
−p̄ln
k (k = 0, 1, 2, 3) converges to 2 ln 2 meaning the stored information becomes completely
destroyed. This non-analytical behavior also appeared in the r.h.s. of Eq. (3.49), i.e., the
free-energy of RBIM. Indeed, the difference between the free energies ∆ = Fk − F0 is
an order parameter, the so-called domain-wall free energy [Huk99], for a ferromagnetic
ordered phase. The phase diagram of RBIM with respect to J = (1/2) ln[p/(1 − p)]) and
p′ is shown in Fig. 3.13. As mentioned before, optimal decoding is achieved with p = p′ ,
p
i.e., e−J = p′ /(1 − p′ ), which is called the Nishimori line [Nis81]. The critical point on
the Nishimori line, which is called a multi-critical point, has been numerically calculated
to be 10.94 ± 0.02% by Honecker et al. and Merz et al. [RDQDS99, HPP01, MC02]. The
optimal threshold value of the surface code is good in the following sense: The existence
of CSS codes with asymptotic rate R ≡ k/n is guaranteed if

R = 1 − 2H2 (p), (3.52)

by the quantum Gilbert-Varshamov bound under independent X and Z errors with prob-
ability p [CS96]. The rate becomes zero with p = 11.00%. The optimal threshold of
the surface code, which consists only of local stabilizer generators, achieves a value very

79
 (k)
Figure 3.12: Examples of ground states in a ferromagnetic phase are shown for {Jij }
with k = 0, 1. In the ferromagnetic case, the insertion of the anti-ferromagnetic bonds
results in an excited domain-wall and hence increases the free energy (ground state energy
at zero temperature).

close to this. Indeed, Nishimori conjectured that the multi-critical point of the RBIM is
determined by

H2 (p) = 1/2, (3.53)

arguing from the self-duality in RBIM with a replica method [NN02, Nis79, KW41] (if
the reader is interested this derivation, please see a review in Ref. [Ohz13]). Ohzeki has
evaluated the multi-critical point precisely to be 10.92% by using a real-space renormal-
ization [Ohz09b], which is in a good agreement with the numerical result [HTPV08].
The minimum distance decoding with MWPM is achieved in the limit p → 0, which is
the low temperature limit J → ∞ where the entropic effect is suppressed. The threshold
of MWPM corresponds to the critical point at zero temperature [Bar82], which has been
investigated numerically and determined to be 10.4 ± 0.1 by Kawashima et al. [KR97,
WHP03, DKLP02].

3.6 Other topological codes

The surface codes have been also studied on general lattice tillings, such as, triangle,
hexagonal, and random lattices [Ohz09b, RWH+ 12, FT12, OF12, ASWP13, Kay14]. Sup-
pose a surface code is defined on a lattice G = (V, E, F ). As seen in the previous section,
the X error correction is mapped into a RBIM on a lattice G = (V, E). On the other
hand, the Z error correction is mapped into a RBIM on its dual lattice Ḡ = (V̄ , Ē). The
mutual duality relation [NN02] between G and Ḡ allows us to predict the relation between

80
Figure 3.13: A phase diagram of RBIM with respect to the coupling strength J =
(1/2) ln[p/(1 − p)] and the probability of anti-ferromagnetic interaction p′ . The proba-
bility p in j corresponds to the parameter in the posterior probability. When p = p′ ,
called the Nishimori line, an optimal decoding is achieved.

the optimal threshold values for X and Z errors:

H(px ) + H(pz ) = 1, (3.54)

where px and pz are the X and Z error probabilities, respectively. This equality is the same
as the quantum Gilbert-Varshamov bound evaluated for the independent X and Z errors
with probabilities px and pz , respectively. The precise locations of the optimal thresholds
for regular lattices have been investigated by Ohzeki using a real-space renormalization
technique [Ohz09b]. The thresholds have been investigated using the MWPM algorithm
by Fujii et al. [FT12]. The thresholds with MWPM (i.e., the critical points of RBIMs at
zero temperature) even approaches Eq. (3.54), as shown in Fig. 3.14. These codes with
an asymmetry between the X and Z error tolerances would be useful to correct a biased
error [ABD+ 09]. In Refs [RWH+ 12, OF12], the asymmetry is continuously controlled by
changing a lattice parameter.
A leakage process or qubit loss is an important source of noise. Unlike the X and Z
errors discussed so far, the qubit loss is detectable (heralded), and hence we can tolerate
much more loss rate than for the (undetectable) error rate. Stace et al. proposed to cope
with the qubit loss on the surface code [SBD09, SB10]. Suppose a qubit is lost on the
surface as shown in Fig. 3.15 (a). The plaquette operators containing the lost qubit are
undetermined. However, we can construct a super-plaquette multiplying two neighboring
plaquette so that the super-plaquette does not have the lost qubit. By using the super-
plaquette as a stabilizer generator, we can perform MWPM. If two super-plaquettes are
neighbors, the weight between these super-plaquettes has to be modified appropriately
in MWPM, because the plaquettes share two physical qubits and the error probability is
effectively doubled. The logical operator is defined by avoiding the lost qubits. Unless

81
Table 3.2: The correspondence among random-bond Ising model, Z2 chain complex and
stabilizer codes.

RBIM stabilizer code chain complex

dual lattice Z error correction primal lattice
primal lattice X error correction dual lattice
Ising interactions qubits edge el
gauge spin Z-type stabilizer generators face fm
gauge spin X-type stabilizer generators vertex vk (dual face f¯k )
domain wall logical X and Z operators nontrivial cycles c1 and c̄1
anti-ferromagnetic interactions X and Z errors 1-chains c1 and c̄1
frustration X and Z error syndromes ∂c1 and ∂c̄1

the lost qubits, which are located on edges, are percolated throughout the lattice, we can
find such a logical operator. Thus, the threshold for qubit loss in the large lattice limit
without any (undetected) error is determined by the bond percolation threshold. The
trade-off curves between qubit loss and (unheralded) error rates with MWPM for various
lattice tillings are shown in Fig. 3.14 (b).
On the other hand, when the error correction problem is mapped into the RBIM, the
qubit loss corresponds to a bond-dilution. The duality relation in the presence of the
bond-dilution on the square lattice is given by [Ohz12]

(1 − q)h(p) + q = 1/2. (3.55)

From the numerical data in Refs. [FT12], we could expect a more general equality,

(1 − qx )h(px ) + (1 − qz )h(pz ) + qx + qz = 1, (3.56)

where qx and qz are the probabilities of the heralded X and Z errors, respectively. With
px = pz = 0, this is reduced to qx + qz = 1, which corresponds to Kesten’s duality relation
of the bond percolation thresholds between mutually dual lattices. With qx = qz = 0, Eq.
(3.56) is reduced to Eq. (3.54). With qx = qz = q and pz = px = p, Eq. (3.56) is reduced
to Eq. (3.55).
Another important class of local stabilizer codes is the topological color code proposed
by Bombin and Martin-Delgado [BMD06, BMD07]. The topological color codes are defined
on trivalent graphs with faces that can be colored with three colors, such as hexagonal
(6.6.6) and (4.8.8) lattices, as shown in Fig. 3.17 (a) and (b), respectively. A qubit is
defined on each vertex v and stabilizer generators are defined on each face f :
Y Y
Bx = Xv , B z = Zv . (3.57)
v∈f v∈f

Specifically, the color code on a (4.8.8) lattice, shown in Fig. 3.16 (c), allows all single-
qubit Clifford gates transversally. The distance-3 topological color code on the (4.8.8)

82
Figure 3.14: (a) The thresholds (px , pz ) of the surface codes on square, Kagome, hexagonal,
and triangle-hexagonal lattices, as well as their duals. The curve is the quantum Gilbert-
Varshamov bound hitting zero asymptotic rate under a bit and phase flip channel with
probabilities px and pz , respectively. (b) The trade-off curves between qubit loss rate
ploss and unheralded error rates px and pz are shown for square, Kagome, hexagonal, and
triangle-hexagonal lattices. If the loss and error rates are below these curves, the logical
information is protected.

lattice is equivalent to Steane’s 7-qubit code [Ste96]. The extension to a 3D lattice also
enables a transversal non-Clifford gate on the code space [BMD07], whose distance-3
version corresponds to the Reed-Mullar 15-qubit code [BK05].
The topological color codes are also described by a Z2 chain complex on hyper-
graphs [Del14]. Consider a trivalent graph G = (V, E), on which a topological color
code is defined. We define a hyper-graph G = (V, E, F) consisting of the sets of hyper-
vertices V, hyper-edges E, and hyper-faces F as follows (see Fig. 3.17): The hyper-vertices
V are vertices V̄ of the dual graph Ḡ. A hyper-edge ẽ ∈ E is defined as a triplet of
(hyper-)vertices v̄ ∈ V̄ on a face f¯ of the dual graph Ḡ. A hyper-face f˜ ∈ F is defined as
a set of hyper-edges ẽ that are incident to the vertex v̄ of the dual graph Ḡ. As defined
in Sec. 3.1, we can define both Z2 valued vector spaces and Abelian groups C1,2,3 with
the hyper-graph elements as their bases. Because the hyper-edge and hyper-face are de-
fined as sets of hyper-vertices and hyper-edges, respectively, we can define boundary maps
∂i : Ci → Ci−1 naturally by such sets. The plaquette and star stabilizer generators of the
surface code defined on such a hyper-graph are as follows:
Y Y
Am = Zẽ , Bk = Xẽ . (3.58)
ẽ∈∂ f˜ ẽ∈δṽ

This definition is equivalent to the previous definition, Eq. (3.57). In Ref. [Del14], the
authors defined the projections from the Z2 chain complex on the hyper-graph into chain
complexes on a dual graph Ḡ. Specifically, a certain subset of vertices is removed in each

83
Figure 3.15: (a) A super-plaquette defined as a product of two plaquette operators. The
lost qubit is not contained in the super-plaquette operator. (b) The logical operator
is chosen appropriately by avoiding the lost qubits. The lost qubits (bonds) are not
percolated, so we can find such a logical operator.

projection. This procedure corresponds to the removal of stabilizer generators colored by

one of the three colors. This allows us to utilize MWPM to decode the topological color
codes efficiently on the projected surface codes [Del14].
The decoding problem of topological color codes is mapped into random three-body
Ising models using the mapping between quantum error correction and spin glass models.
This can be understood as follows. For each face center a gauge spin valuable is located
associated with each stabilizer generator; three gauge spin valuables (stabilizer genera-
tors) that share the same qubit (vertex) interact with each other. The locations of the
optimal thresholds have also been investigated via the spin glass theory and Monte Carlo
simulations [Ohz09a, BAO+ 12].

3.7 Connection to topological order in condensed matter

physics
Topological order is an exotic quantum phase of matter, which cannot be characterized
by the Landau-Ginzburg theory of symmetry breaking, in which an ordered phase can
be characterized by local order parameters. The ground state of a topologically ordered
system is degenerate, but its degeneracy cannot be destroyed by any local perturbation.
Thus, no local order parameter can succeed to capture the topologically ordered phase.
Understanding the nature of topological order is one of the most important goals of modern
condensed matter physics. Moreover, the ground state of a topologically ordered system is,
by definition, robust against any local perturbations, and hence it is also useful for storing
quantum information. There is a beautiful correspondence between topological quantum
codes and topologically ordered systems, which provides a promising way to understand
condensed matter physics via quantum information.
Let us first consider the bit-flip code defined in Sec. 3.2. The Hamiltonian, which we

84
Figure 3.16: (a) and (b) Trivalent lattices that can be colored with three distinct colors.
(c) Color codes defined on (4.8.8) lattices with a triangle open boundary condition. The
logical operators are shown by solid lines. Any string at the boundaries can be employed
as a logical operator.

call a stabilizer Hamiltonian, is defined as follows:

X n−1
X
HIsing = −J Ak = −J Zi Zi+1 . (3.59)
k i=1

Here, the summation is taken over all stabilizer generators, but one stabilizer operator,
which is not independent, is removed. The Hamiltonian Eq. (3.59) corresponds to the
Ising model in 1D with an open boundary condition. By its construction, the Hamiltonian
is diagonalizable, and the stabilizer state becomes the ground state. The bit-flip code has
a 2D stabilizer subspace spanned by {|00...0i, |11...1i}. This means that the ground state
is degenerate. The bit-flip errors occurring on the code space excite the ground state to an
excited state. Thus, the states in the orthogonal subspaces of the bit-flip code correspond
to excited states.
To address topological order, let us consider the robustness of the ground state against
P
perturbations from transversal fields hx i Xi . By using standard perturbation the-
ory [Kat95], we can show that the degeneracy of the ground states is not lifted up to
the (n − 1)th order of perturbation. This happens because the code distance of the n-
qubit bit flip code against X errors is n, and any X errors of weight up to n − 1 map the
code state into an orthogonal space. Accordingly, the ground state degeneracy is robust
against the transversal fields.

85
Figure 3.17: (left) A hexagonal lattice with three-coloring. (right) A triangular lattice, the
dual of the hexagonal lattice. A hyper-graph is defined on a triangular lattice. The color
code on the hexagonal lattice corresponds to a surface code defined on the hyper-graph.

P
On the other hand, longitudinal fields hz i Zi spoil the ground state degeneracy in
the large n limit, even if hz is small. More precisely, the energy between |00..0i and
|11..1i is shifted by nhz . Thus, a superposition α|00..0i + β|11..1i in the ground subspace
is easily destroyed by the longitudinal fields. In this sense, the stabilizer Hamiltonian
constructed by the bit-flip code is not topologically ordered. However, if any perturbation
with respect to the longitudinal fields is prohibited due to some symmetry of nature, the
ground state degeneracy is robust under that symmetry. This type of robustness of ground-
state degeneracies is called symmetry protected topological order [GW09, PTBO10]. (Note
that in this case the ground-state degeneracy is not related to the geometrical property of
the underlying manifold in contrast to the genuine topological order in 2D.)
The symmetry prohibiting the longitudinal fields seems to be somewhat artificial. We
can, however, impose the symmetry by transforming the Ising Hamiltonian to a free-
fermion model using the following Jordan-Wigner transformation [JW28]:
i−1
Y
a2i−1 = Xk Zi , (3.60)
k=1
i−1
Y
a2i = Xk Yi . (3.61)
k=1

The operators, called Majorana fermion operators, are hermitian ak = a†k and satisfy
the fermion commutation relation {ak , ak′ } = δk,k′ I. The Ising stabilizer Hamiltonian is
reformulated in terms of ak :
n−1
X
HIsng = −J (−i)a2i a2i+1 . (3.62)
i=1

86
Figure 3.18: (left top) (1) A pair of X-type excitations are created, moved around the
torus, and annihilated. (2) A pair of Z-type excitations are created, moved around the
torus, and annihilated. (3) Do the process (1) again. (4) Do the process (2) again.
(middle) The creation, movement, and annihilation process is continuously deformed.
(right bottom) A braiding operation of an X-type excitation around a Z-type excitation.

The logical operators acting on the degenerated ground states are given by
n−1
!
Y
LZ = a1 = Z1 , LY = a1 a2n = Y1 Xk Yn . (3.63)
k=2

The degree of freedom in the degenerated ground states is called the Majorana zero mode
or the unpaired Majorana fermion [Kit01]. If the parity of the number of fermions is
preserved, the fermion operators would appear with a quadratic form ak a′k . Under such
a symmetry, there is no perturbation that lifts the ground state degeneracy. Thus, the
ground state of the unpaired Majorana fermion is symmetry protected.
Next, we will provide a genuine topologically ordered system based on the surface
code (Kitaev’s toric code). The stabilizer Hamiltonian, the so-called Kitaev’s toric code
Hamiltonian [Kit97], is given as a summation of all plaquette and star operators:
X X
HKitaev = −J Am − J Bk . (3.64)
m k

The ground state has a fourfold degeneracy corresponding to the code space. Errors on
the code state correspond to excitations. Specifically, there are two types of excitations,
corresponding to the Z error Z(c1 ) and X error X(c1 ). Excitations appear at the bound-
aries of the error chains ∂c1 and ∂c̄1 , because the local energy changed from −J to +J
there. Such excitations, i.e., at the end points of the error chains, are always created as
pairs, can be viewed as a pair creation process on the ground state.
Suppose these two types of the excitations were created on the system and moved as
shown in Fig. 3.18 (left top). This process can be described by
(2) (2) (1) (1)
Z(c1 )X(c̄1 )Z(c1 )X(c̄1 )|Ψi = −|Ψi. (3.65)

87
On the other hand, by continuously changing the trajectory of the particles, as shown in
Fig. 3.18, this process can also be regarded as a braiding process of an X-type excitation
around a Z-type excitation. After the braiding operations, a phase factor is applied
to the state as shown in the r.h.s. of Eq. (3.65). Thus, the excitations are neither
bosonic nor fermionic, which are invariant under the braiding operation, i.e., the swapping
operation twice. In this sense, the excitations on the surface code are referred to as anyons.
Specifically, since the phase factors is Z2 , they are called Z2 Abelian anyons. By using
the generalized Pauli operators on for the qudit, we can also define Zd Kitaev’s toric
code [Kit03], on which excitations are Zd Abelian anyons. More generally, using a finite
group G, we can define a quantum state |gi (g ∈ G) in a |G|-dimensional Hilbert space.
Then we define four types of operators for each g ∈ G:
X X
Lg+ = |ghihh|, Lg− = |hg −1 ihh|, T+h = |hihh|, T−h = |h−1 ihh−1 |. (3.66)
h∈G h∈G

The non-Abelian Kitaev’s toric code model, which is called the quantum double model [Kit03],
is defined as
X X
H = −J A(fm ) − J B(vk ), (3.67)
m k

in terms of the following plaquette and star operators

X
A(fm ) = T−g1 (em g2 m g3 m g4 m
l1 )T− (el2 )T+ (el3 )T+ (el4 ) (3.68)
g1 g2 g3 g4 =I
1 X g k g k g k g k
B(vk ) = L+ (ēl1 )L+ (ēl2 )L− (ēl3 )L− (ēl4 ) (3.69)
|G|
g∈G

where the four edges em k ¯

l1,2,3,4 ∈ ∂fm and ēl1,2,3,4 ∈ δvk = ∂ fk are labeled clock wise. The
quantum double model supports non-Abelian anyonic excitations [Kit03, Pac12], which
allows us to implement universal quantum computation solely by braiding them.
Let us return to the Z2 Kitaev’s toric code Hamiltonian. The code distance of the
surface code on the n × n torus is n. Thus, neither X-type (transverse) nor Z-type
(longitudinal) fields can lift the ground state degeneracy up to the (n − 1)th order of
perturbation. More generally, no local perturbation can lift the ground state degeneracy
in the large n limit. Thus, the ground state of the Kitaev’s toric code Hamiltonian is
topologically ordered. For any properly defined stabilizer code, we can define a stabilizer
Hamiltonian, whose ground state is topologically ordered. However, in condensed matter
physics, the local interactions are of central importance. Thus, it is natural to restrict the
stabilizer generators to be spatially local, i.e., topological stabilizer codes.
Only the nearest-neighbor two-body interactions are attained in physically natural
systems. It has been known that the Kitaev’s toric code Hamiltonian is obtained as an
effective low energy model of another model consisting only of two-body nearest-neighbor
interactions [Kit06]. Let us consider the following two-body nearest-neighbor model, called

88
Figure 3.19: A hexagonal lattice on which Kitaev’s compass model is defined. In the large
Jz limit, the two spins on each vertical edge are confined into a 2D subspace forming a
dimmer. The dimmer is located on each edge of a square lattice shown by the dotted lines.

Kitaev’s compass model:

X X X
Hcomp = −Jx Xi Xj − Jy Yi Yj − Jz Zi Zj , (3.70)
(i,j)∈Ex (i,j)∈Ey (i,j)∈EZ

where Ex , Ey , and Ez are sets of right-up, left-up, and vertical bonds, respectively, of a
hexagonal lattice (see Fig. 3.19 (left)). If we take the large Jz limit, each vertical bond
favors the two-dimensional subspace spanned by {|00i, |11i}, because it is stabilized by
Zi Zj . Thus, in the large Jz limit, we can derive an effective low energy Hamiltonian, which
commutes with the Zi Zj interactions, by using perturbation theory:

Jx2 Jy2 X X
Heff = − X̃ f X̃ f Ỹ f Ỹ f − JZ
e1 e2 e3 e4
Z̃e , (3.71)
16|Jz |3
f e∈EZ

where Ãe = Ai Aj with e = (i, j) and A = X, Y, Z, depending on e ∈ Ex , Ey , Ez . The

P
edges {efi } are left-up and right-up edges on a hexagonal face f and the summation f
is taken over all faces.
Let us define a qubit {|0̄i = |00i, |1̄i = |11i} for each dimerized edge and Pauli opera-
tors X̄ = XX or = Ȳ Y and Z̄ = Z ⊗ I or = I ⊗ Z. Now a qubit is assigned to each edge
of a square lattice, which is an vertical edge in the hexagonal lattice as shown in Fig. 3.19
(right). Using this definition, the effective Hamiltonian can be reformulated as

Jx2 Jy2 X
H̄eff = − X̄ef X̄ef Z̄ef Z̄ef , (3.72)
16|Jz |3 l r t b
f

where {efl , efr , eft , efb } are the left, right, top, and bottom edges, respectively, on a square
face f . If we apply the Hadamard transformation on all horizontal edges, we obtain the
Kitaev’s toric code Hamiltonian. In this way, the stabilizer Hamiltonian can be obtained

89
as a low-energy effective model of a two-body nearest-neighbor system [Kit06]. This
shows the validity of employing topological stabilizer codes and quantum coding theory
to understand the quantum phase of matter in condensed matter physics.
A complete classification of the topological stabilizer codes in 2D has been obtained
by Yoshida [Yos11a] (see also the classification by Bombin et al.[BDCP12]). Specifically,
the quantum phases of the 2D stabilizer Hamiltonians are classified by the geometric
shapes of the logical operators. The thermal stability of the topological order in stabilizer
Hamiltonian systems at finite temperatures has also been studied via quantum coding
theory by Bravyi and Terhal [BT09] for 2D, and Yoshida for 3D [Yos11b].
If there is a thermally stable topological order, we can store quantum information re-
liably even at a finite temperature without any active error correction, i.e., we have a self-
correcting quantum memory. Of course, if a fault-tolerant quantum computer were real-
ized, we could store quantum information reliably with a repetitively performing quantum
error correction, which, however, requires selective addressing of each individual qubits.
There are also several intermediate approaches for a reliable quantum storage without
selective addressing using global dissipative dynamics [FNIK14], an interaction with an
engineered environment [HCC09, PHWL13, KCS14], and decoding by cellular automata
with local update rules [HCEK14, Har04]. However, a genuine topologically ordered self-
correcting quantum memory seems to be hard to achieve in 2D even in the presence of
effective long-range interactions [LCYPP15].

90
Chapter 4

Topological quantum computation

In this chapter, we explain how to perform topological fault-tolerant quantum computation

on the surface code. All operations employed are nearest-neighbor (at most) two-qubit
gates and signle-qubit measurements on a 2D array of qubits. This property is quite favor-
able for the fabrication of qubits and their control lines on a chip. Fault-tolerance ensures
that no local noise during any sort of operations ever spoil the quantum computation, if
the noise strength is smaller than a certain threshold value, as seen below.
We first introduce the defect pair qubit, a logical qubit using a pair of defects on
the surface, which allows us to encode many logical qubits on the surface. Then, we
explain the elementary operations of defects by local quantum information processing. A
braiding operation, based on the elementary operations, of the defects is further utilized
to implement the logical CNOT gate between the defect pair qubits. Combining with the
topologically protected operations, the magic states are injected and distilled for fault-
tolerant universal quantum computation. Based on these understandings, we introduce a
topological diagram and the topological calculus on it, as a set of the transformations that
preserve the logical actions on the code space. These provide us with a deep understanding
of topological quantum computation on the surface. Finally, we return to the microscopic
viewpoint to explain how topological quantum error correction is executed and analyzed.

4.1 Defect pair qubits

In order to perform quantum computation of N qubits, we have to arrange N logical qubits
in the code subspace while keeping their code distance long enough. A way of doing this
is to use a surface of higher genus. Specifically, we use a large planar surface and punch
holes on it, which we call defects on the surface.
In Sec. 3.3.2, we have seen that we can define a logical qubit by introducing a defect on
a planar surface, where the stabilizer generators inside the defect region are removed from
the stabilizer group. The logical operators are defined by a nontrivial cycle surrounding
the defect and a chain connecting the defect to the boundary. If we introduced many
degrees of freedom, i.e., many defects based on this strategy, the logical operators defined

91
Figure 4.1: Primal (left) and dual (right) defect pair qubits. The primal defect pair qubit
is defined by removing the plaquette operators inside the defect regions D and D ′ . Then
the cycle ∂D around the defect becomes a nontrivial cycle, by which the logical Z operator
Z(∂D) is defined. The logical X operator X(c̄1 ) is defined as a dual 1-chain c̄1 connecting
two defects. Any homologically equivalent logical operators acts the same way on the code
space. The dual defect pair is defined similarly, but the basis is changed by the Hadamard
transformation.

by chains connecting each defect to the boundary would become complicated. To avoid
this, we use a pair of defects to define a logical qubit, which we call a defect pair qubit.
We define two defect regions D ∈ C2 and D ′ ∈ C2 , as shown in Fig. 4.1. We remove all
Z-type stabilizer generators inside these regions, i.e., {Am }fm ∈D∪D′ . Thus, the operators
Z(∂D) and Z(∂ D̄ ′ ) are not stabilizer operators. Instead, we append a Z-type stabilizer
operator Z(∂D + ∂ D̄) as a stabilizer generator. Moreover, we also append the Pauli X
operator on all edge qubits inside (not including the boundary) the regions D ∪ D ′ , i.e.,
{Xl }el ∈(D∪D′ )\(∂D∪∂D′ ) . By this definition, the star (X-type) stabilizer generators in the
defect region are still in the stabilizer group. We choose Z(∂D) as a logical operator,
because it commutes with all stabilizer operators, but does not belong to the stabilizer
group. Because Z(∂D + ∂ D̄) is a stabilizer operator, Z(∂D ′ ) also acts the same way, i.e.,
Z(∂D) ∼ Z(∂D ′ ). Moreover, the actions of operators represented by the homologically
equivalent cycles are the same. The logical X operator is given by X(c̄1 ), with a dual
1-chain c̄1 connecting two defects, as shown in Fig. 4.1. The actions of the operators
represented by any homologically equivalent chains in the sense of relative homology are
the same. That is, we may choose any dual 1-chain c̄1 which connects two defect regions
D and D ′ . The code distance is given by the circumference of the defect or the distance
between two defects. Because the plaquette operators on the primal lattice are removed,
we call these defects and the logical qubit primal defects and primal defect pair qubit,
respectively.

92
Figure 4.2: (a) A defect creation. The qubits inside the defect regions (excluding the
qubits on the boundary), denoted by circles, are measured in the X-basis. (b) The defect
region D is expanded into D ′ by measuring the qubits inside the region D ′ in the X-basis,
where the encoded quantum information is preserved. (c) A defect region D ′ is contracted
into D by measuring the plaquette operators in D ′′ = D ′ \D. The logical operator Z after
the contraction is defined depending on the measurement outcome m.

Similarly, we can also define a logical qubit by removing the star (X-type) stabilizer
operators on the dual defect regions D̄ and D̄ ′ defined on the dual lattice as shown in
Fig. 4.2. We call such defects and the logical qubit dual defects and dual defect pair qubit,
respectively. Hereafter, the planar surface code state, on which the defects are introduced,
is referred to as vacuum. Below we will explain how defect qubits are created, deformed,
and moved in the vacuum.

4.2 Defect creation, annihilation, and movement

The defect creation is accomplished by measuring the qubits inside the defect region D, not
including the qubits on its boundary, in the X-basis [see Fig. 4.2 (a)]. These measurements
remove the plaquette operators inside D from the stabilizer group, because these X-
basis measurements do not commute with them. On the other hand, the measurements
do commute with the star operators, and hence a parity of four measurement outcomes
corresponding to a star operator is always even (if there is no measurement error). While
the measurement outcomes are random, we can prepare all measured qubits to be in the
|+i state by applying Z operators according to the measurement outcomes. (In practice,

93
there is no need to apply the Z operations. It is enough to keep the information of the
byproduct Pauli operator dependent on the measurement outcomes.)
As the qubits on the boundary are not measured, the post-measurement state is sta-
bilized by Z(∂D). In order to create a pair of defects, we do the same thing for another
defect D ′ . Apparently, the resultant state is stabilized by Z(∂D + ∂D ′ ) and satisfies the
definition of the defect pair qubit. Moreover, the logical qubit is stabilized by Z(∂D) and
hence a logical Z-basis state is prepared.
The defect region can be expanded by creating a larger defect D ′ , which includes the
original defect region D, i.e., D ⊂ D ′ , as shown in Fig. 4.2 (b). We can choose Z(∂D ′ )
as a logical Z operator of the defect pair qubit, so that the information with respect to
Z(∂D ′ ) is preserved during this operation. (Recall that logical operators act the same if
the corresponding cycles are homologically equivalent. Thus we can choose a large enough
cycle surrounding the defect in advance.) We can expand the defect region step-by-step
such that, at each step, a logical X operator is untouched as shown in Fig. 4.2 (b). (Or
equivalently, we can compensate for the difference in the logical X operator according
the X-basis measurement outcomes inside the defect region.) Thus, the information with
respect to the logical X operator is also preserved. Accordingly, the logical information
of the defect pair qubit is stored, but now the defect region D is expanded into D + D ′ .
The defect annihilation is executed by measuring the plaquette operators inside the
region D to restore them to the stabilizer group. The surface with a defect D is rewritten
by
Y  I + Xl 
|Di ∝ |vi, (4.1)
2
el ∈D

where |vi indicates the surface code state without defect, i.e., vacuum. Thus, |Di is a
superposition of all possible applications of X operators on the vacuum |vi. The mea-
surements of the plaquette operators collapse the superposition. By applying a recovery
operation inside the region D such that all eigenvalues becomes +1, the defect is annihi-
lated. (Note that there is no need to actually apply the recovery operation. It is enough
to keep the record of the eigenvalues.) The parity of all measurement outcomes of the
plaquette operators inside D corresponds to the eigenvalue of Z(∂D), because we have
Y Y
Z(∂D) = Z(∂fm ) = Am . (4.2)
fm ∈D fm ∈D

This indicates that a logical Z-basis measurement can be done by annihilating the defect
completely.
Suppose a defect region D inside the defect region D ′ (i.e., D ∈ D ′ ) is annihilated by
measuring the qubits inside D, as shown in Fig. 4.2 (c). We can then obtain the eigenvalue
(−1)m (m = 0, 1) of Z(∂D). Let D ′′ be the complement of D in D ′ . Because we have

Z(∂D ′ ) = Z(∂D ′′ )Z(∂D), (4.3)

94
 Figure 4.3: A defect movement by expansion and contraction.

depending on the eigenvalue (−1)m , the operator (−1)m Z(∂D ′′ ) acts as the logical operator
of the defect D ′′ . Thus, the defect D ′ , consisting of the defect pair qubit, is contracted
into D ′′ without changing the stored quantum information (up to the logical Pauli X flip).
The defect movement on the surface is implemented by combining the defect expansion
and contraction, as shown in Fig. 4.3. At first, we expand a defect D into D ′ (D ∈
D ′ ) by the previously mentioned procedure. Second, the defect region D is annihilated,
and we obtain the eigenvalue (−1)m of Z(∂D). As we already pointed out, the stored
logical information is unchanged under these procedures up to the logical Pauli X operator
depending on the measurement outcome m.
The logical X-basis state preparation is executed by combining the defect creation and
annihilation, as shown in Fig. 4.4. We first create a defect region D0 = D1 + D2 + D3 ,
which consists of three adjacent defect regions D1 , D2 , and D3 . The two defects D1 and
D3 are utilized as a defect pair. We can define a logical X operator by employing the
qubits inside the defect region D2 . Because all qubits inside the defect region are in the
|+i state, the eigenvalue of the logical X operator at this time is +1. Then the defect
region D2 in-between D1 and D3 is annihilated by measuring the star operators. These
measurements commute with the logical X operator, and hence its eigenvalue is still +1
after the annihilation. Now we have a defect pair qubit, which is the eigenstate of the

95
Figure 4.4: A preparation of the logical X-basis state by creation of a defect region D0
and annihilation of a defect region D2 between two defect regions D1 and D3 . Note that
after the first defect creation, the state is stabilized by the logical X operator X(c̄1 ), which
is untouched during the following defect annihilation.

logical X operator. A measurement in the logical X basis can be implemented by doing

the X-basis state preparation in an inverse way. More precisely, two defects D1 and D3
are connected by making a larger defect D0 = D1 + D2 + D3 . We can choose a logical X
operator such that all constituent qubits belong to the defect region D0 . Then, we obtain
the eigenvalue of the logical X operator.
The dual defect creation, annihilation, expansion, contraction, and propagation can be
done in the same way on the dual lattice with the Hadamard transformation. In Sec. 4.5,
these elementary operations of the primal and dual defects are depicted by a topological
diagram.

4.3 Logical CNOT gate by braiding

Using the elementary operations explained in the previous section, we can perform logical
gate operations on the defect pair qubits. Indeed, the defects created from the vacuum
behave like anyonic particles, i.e., braiding a defect around another defect results in a
nontrivial operation. (It is better to say that we change the manifold continuously to
deform the ground state, in contrast to the anyons appeared as an excitation on the toric
code Hamiltonian.) All operations employed in the elementary operations are single-qubit
or local stabilizer measurements, which can be done by nearest-neighbor two-qubit gates
and single-qubit measurements. This contrasts with concatenated quantum computation,
where the logical operations are implemented as transversal operations and hence, essen-
tially, non-local two-qubit gates are employed.

96
Figure 4.5: A logical CNOT gate created by braiding. (a) A primal defect is braided around
a dual defect. (b,c) The logical X operator is deformed continuously during the braiding
operations. (d) After the braiding, the logical X operator is represented by the dual chain
c̄′1 winding around the primal defect. By applying star operators inside the rectangle
denoted by dotted lines, c̄′1 is decoupled into a chain connecting two primal defects and
a chain surrounding the dual defect. The corresponding operators are equivalent to the
products of the logical X operators on the primal and dual qubits.

We first consider the CNOT gate between primal (control) and dual (target) defect
pair qubits. Suppose there are primal and dual qubits on the surface code as shown in
Fig. 4.5 (a), where logical operators are specified by the chains {∂D, c̄1 } and {∂ D̄, c1 } for
the primal and dual qubits, respectively. We can freely move the defect everywhere we
want by the defect expansion and contraction. Let us braid the primal defect around the
dual defect, as shown in Fig. 4.5 (a)-(d). After the braiding operation, the operator X(c̄′1 )
in Fig. 4.5 (d) has the same information as the operator X(c̄1 ) before the braiding. Using
the equivalence relation,

X(c̄′1 ) ∼ X(c̄1 )X(∂D), (4.4)

we understand that a correlation is made by the braiding operation between the logical X
operators X(c̄1 ) and X(∂D) of the primal and dual defect pair qubits. A similar observa-

97
Figure 4.6: The time evolution of the logical Z operator caused by braiding. The logical
Z operator on the dual qubit is transformed into the product of the logical Z operators
on the primal and dual qubits.

tion holds for the logical operator Z(c1 ), which is transformed into Z(c′1 ) ∼ Z(c1 )Z(∂D)
(see Fig. 4.6). On the other hand, Z(∂D) and X(∂ D̄) are invariant under this operation.
In short, the braiding operation transforms the logical Pauli operators as follows:

Z(∂D) → Z(∂D), (4.5)

Z(c1 ) → Z(c1 )Z(∂D), (4.6)
X(∂ D̄) → X(∂ D̄), (4.7)
X(c̄1 ) → X(c̄1 )X(∂ D̄). (4.8)

This transformation is equivalent to that for the CNOT gate, Eqs. (2.14-2.17). Thus,
braiding the primal defect around the dual defect results in a logical CNOT gate between
the primal and dual defect pair qubits.
Unfortunately, in the above CNOT gate, the primal (dual) defect is always a control
(target) qubit. Such CNOT gates always commute with each other, which is a natural
consequence of the fact that the defect qubits on the surface code are Abelian. In order to
realize a genuine CNOT gate between the same type of qubits (and hence noncommuting
gates), we utilize a teleportation-based gate [GC99], as shown in Fig. 4.7. Only CNOT
gates between primal (control) and dual (target) qubits are employed. The Pauli basis
measurements are also done as mentioned in the previous section. Accordingly, the CNOT
gate between primal qubits is realized by braiding the primal defects around virtual dual
defects, which are created and annihilated as ancillae.
The above braiding operations are implemented with topological quantum error cor-
rection at each elementary step, which will be explained later in detail. All operations
considered so far can be executed keeping the defect size (circumference) and defect dis-

98
Figure 4.7: The circuit diagram of the teleportation-based CNOT gate between primal
qubits using a dual qubit as ancilla (left). The control and target qubits of the CNOT gates
are always primal and dual qubits, respectively. The corresponding braiding operation
(right).

tance larger than a length d, which provides a code distance of the logical qubits. Thus,
the logical error probability for the logical CNOT gates decreases exponentially by in-
creasing the characteristic length d of the system. In other words, the CNOT gates are
topologically protected.
We should mention that the braiding operations explained above are not the only
way to perform fault-tolerant operations on the surface code. There are another ap-
proaches to perform logical operations fault-tolerantly for the encoded degrees of freedom
of the surface code. One is the lattice surgery scheme [HFDVM12], where the bound-
ary conditions of two planar surface codes are engineered to perform a logical operation.
Another is to employ twists, which are topological defects introduced by point defects
on lattices [Bom10, Bom11]. All Clifford gates can be implemented by the twist cre-
ation, braiding, and annihilation similarly to the defect pair qubits explained above. All
these different approaches can be view as the logical operations by the code deforma-
tions [DKLP02, RHG06, BMD09] and seem to be a unique feature for the topological
codes contrasting to the transversal logical gate for the concatenated codes.

4.4 Magic state injections and distillation

Unfortunately, the topologically protected CNOT gates do not allow universal quantum
computation, because Clifford circuits can efficiently be simulated classically due to the
Gottesman-Knill theorem [Got98b]. Here, we explain how to perform single-qubit rota-
tions on the defect pair qubit, while, unfortunately, they are not topologically protected.
Suppose we have a defect pair qubit, whose logical operators are given by Z(∂D) and
X(c1 ). We first consider a logical Z rotation e−i(θ/2)Z(∂D) on the defect pair qubit. To

99
Figure 4.8: (a) A magic state injection by contracting the defect into an elementary cell.
(b) A circuit diagram for the logical Z rotation on the defect of an elementary cell.

this end, the defect region D is contracted to a single face fl by using the annihilation
process, as shown in Fig. 4.8 (a). Now the logical operator Z(fl ) is a four-body operator.
(This implies that the code distance becomes 4 at this stage.) A rotation with respect to
the logical operator Z(fl ), e−i(θ/2)Z(fl ) , can be implemented indirectly by using an acilla
qubit located at the center of the face shown in Fig. 4.8 (b). First, the four CZ gates are
applied between the ancilla and the four edge qubits on the face. Second, after applying a
X-rotation Rx (θ) = e−i(θ/2)X , the ancilla qubit is measured in the Z-basis. According to
m
the measurement outcome m = 0, 1, a logical rotation e−i(−1) (θ/2)Z(fl ) is implemented.
After the above procedure, the defect, as a single face fl , is expanded again into the defect
region D to restore the code distance of the defect pair qubit.
Next we consider a logical X rotation, e−i(θ/2)X(c1 ) . In this case, the two defects are
moved and deformed near each other such that X(c′1 ) becomes a one-body operator as
′
shown in Fig. 4.9. The rotation e−i(θ/2)X(c1 ) is now easily implemented by a single-qubit
gate. After that, the distance between two defects is restored.
In this way, we can perform an arbitrary single-qubit unitary operation on the defect
pair qubit by continuously deforming the defect pair qubit. These deformations and the
operations for the logical rotations are depicted in Fig. 4.10. Unfortunately, during the
deformation, the code distance inevitably becomes relatively small, so that we can perform
logical rotations directly with nearest-neighbor operations. Thus, these processes are not
topologically protected.
These non-topological operations can be utilized to inject noisy magic states on the
surface. Specifically, the logical e−i(θ/2)Z |+i state is injected as shown in Fig. 4.10 (c).
Then, these noisy magic states are distilled by topologically protected operations to clean
magic states, which have a fidelity high enough for reliable quantum computation. Note
that we are allowed to use only the CNOT gates, and there is no topologically protected
single-qubit Clifford gate. To manage this, we distill two types of magic states. One is
the eigenstate of the Pauli-Y operator, which is used to implement the S gate via gate

100
Figure 4.9: A magic state injection by connecting two defects. If two defects regions are
adjacent, the logical X operator is a physical Pauli X operator between two defects.

teleportation [GC99, ZLC00], as shown in Fig. 4.11 (a). Another is the eigenstate of
√
the (X + Y )/ 2 operator, which is used to implement the π/8 gate, a non-Clifford gate
necessary for universal quantum computation. The magic state distillation for the Y -basis
state is executed solely by the topologically protected CNOT gates using the Steane 7-
qubit code, similarly to the method introduced in Sec. 2.8. The topologically protected
CNOT gates and the S gates with the distilled Y -basis states are further employed to
√
distill the (X + Y )/ 2-states through the Reed-Muller 15-qubit code, as explained in
Sec. 2.8.
Using these distilled magic states and the CNOT gates, we can implement the single-
qubit gates shown in Fig. 4.11 (a) and (b), which together with the CNOT gate constitute
a universal set of gates. Accordingly, universal quantum computation is performed reliably
on the surface code. Note that all operations employed are single-qubit gates, two-qubit
nearest-neighbor gates, and single-qubit measurements on the 2D array of qubits.

4.5 Topological calculus

Based on the microscopic understanding of the topological operations, we can introduce a
topological diagram and a topological calculus, which allows us a diagrammatic description
of topological quantum computation on the surface codes.
In this diagram, the 3D space-time trajectory of the defects is depicted by projecting
it onto a 2D plane like link diagrams mentioned in Sec. 1.5.3. The trajectory of the primal
and dual defects as tubes in 3D space-time are denoted by pairs of solid and gray-colored

101
Figure 4.10: (a) The logical Z rotation. (b) The logical X rotation. (c) The logical
e−i(θ/2)Z |+i state injection.

Figure 4.11: (a,b) Single-qubit rotations by one-bit teleportations [ZLC00].

lines:

The braiding operation is denoted by a double crossing as follows:

The logical X state preparation and X-basis measurement of the primal defect are denoted
by closures:

102
The logical Z state preparation and Z-basis measurement of the primal defect are denoted
by endpoints of the tubes:

The state preparations and measurements for the dual defect pair qubit are depicted in
a similar way with gray (curved) lines, except for the basis change by the Hadamard
transformation.
The trajectories of the logical operators are represented by surfaces in 3D space-time
like a Seifert surface, which we call correlation surfaces. Specifically, the logical Z operator
for the primal defect corresponds to a surface wrapping around the primal defect. The
logical X operator for the primal defect corresponds to a surface whose boundary is the
primal defect tubes. (This is also the case for the dual defect, except for the Hadamard
transformation.) For example, the time evolution of the logical X and Z operators by the
braiding operation can be represented by the following surfaces

The logical operators before and after the operation correspond to the left and right
boundaries of these surfaces, respectively. We can confirm Eqs. (4.5)-(4.8) from the left
and right boundaries the correlation surfaces of the above diagrams.
The CNOT gate between the primal defect pair qubits is depicted as follows:

103
In this diagram, we can easily confirm that the braiding operations transform the logical
operators according to the rule for the CNOT gate, i.e., X ⊗I → X ⊗X and I ⊗Z → Z ⊗Z:

As seen above, the action on the code space is defined by the left and right boundaries
of the correlation surfaces. Thus, the logical action of the topological operation is invariant
under transformations that do not change the topology of the left and right boundaries
of the correlation surfaces. Similarly to link diagrams, the boundary topology is invariant
under the Reidemeister moves:

In the present case, we have further transformations under which the logical action on
the code space is invariant due to the properties of the defect pair qubits. At first, the
following two crossings of tubes of the same type are equivalent:

(4.9)

Second, if two defect tubes are a defect pair qubit, a dual loop wrapping around a defect
pair can be removed:

(4.10)

104
This is because the primal defect pair qubit is stabilized by a Z-type loop operator sur-
rounding the defect pair, and hence the dual loop is nothing but a trivial measurement of
the stabilizer operator. Third, because the defect pair qubit is a Z2 Abelian anyon, if we
braid a primal defect around a dual defect twice, it results in an identity operation:

(4.11)

Fourth, and most importantly, two tubes can be connected or disconnected by the following
procedure, which we call a Φ-transformation:

(4.12)

We can easily confirm that the logical operators are invariant under the Φ-transformation:

where we note that the dual ring in the middle serves to stop the Z-type correlation surface
from propagating toward the bottom and also serves to mediate the X-type correlation
surface toward the right. Specifically, if two defect pair qubits are connected by the Φ-
transformation between their closures, we can remove the dual ring by using the rule
(4.11):

By using the Φ-transformation and rules (4.11) and (4.10), we can transform the defect
pair qubit as follows:

105
We can also easily confirm that the logical operators are invariant under this transforma-
tion as follows:

The Φ-transformation is very useful to simplify a complicated topological operation on

the surface by joining the tubes as follows:

where we frequently used the Φ-transformation and rule (4.10). Here, the dual triple-ring
wrapping around each of the two defect tubes serves to reflect the logical Z operator from
the upper to the lower tubes as follows:

106
 For example, if we apply this transformation to the CNOT gate between the primal
defect pair qubits, we obtain a much simpler diagram as follows:

where the dual ring in the middle serves to keep the code space of the defect pair qubit as
shown in the left above. It is straightforward to check the transformations of the logical
operators. In general, by joining the defect tubes directly, we can transform the logical
operators under multiple CNOT gates. However, if two tubes are joined, the definition
of the defect pair is broken. In order to keep the defect pair qubit encoding, we need the
dual ring wrapping around the two tubes at the joint.
The magic state injection is denoted by the following diagram:

which can be viewed as a superposition of two correlation surfaces for the two anti-
commuting logical operators. The non-Clifford gate by one-bit teleportation can be de-
scribed in the topological diagram as follows:

107
Figure 4.12: Syndrome measurements for plaquette (left) and star (right) operators, where
qubits on a face center and on a vertex are employed as ancillae. The alternative syndrome
measurements for the plaquette and star operators are done repeatedly.

The right diagram, which is topologically equivalent to the left one, indicates that the
non-Clifford gate by one-bit teleportation with the magic state injection is equivalent to
simply performing the non-Clifford gate by the method mentioned in the previous section.
The present diagrammatic description of the topological operations and correlation
surfaces on them provide us an intuitive understanding of how topological quantum com-
putation on the surface code is performed. These transformation rules will be useful to
optimize the complexity (space-time volume required) of the braiding operations.

4.6 Faulty syndrome measurements and noise thresholds

We have seen how universal quantum computation is executed on the surface code. At
each step, we have to perform topological quantum error correction to protect the quan-
tum information encoded by the defects. In Sec. 3.4, we analyzed topological quantum
error correction on the surface code. However, at that time, we assumed ideal syndrome
measurements. However, in a fault-tolerant quantum computation, we have to take into
account all reasonable sources of noise, including faulty syndrome measurements, and the
quantum computation has to tolerate them as well. Here, we explain how faulty syndrome
measurements are handled during topological quantum computation, which completes the
big picture of fault-tolerant topological quantum computation.

108
 During the topological operations, the star and plaquette operators are measured in
the vacuum region to obtain the error syndrome. These measurements are implemented
using ancillae located on each vertex and face center for the star and plaquette operators,
respectively (see Fig. 4.12). The ancilla state is prepared to be |+i and the Λ(X) or
Λ(Z) gates are performed from the ancilla qubit as a control to the four qubits on ∂fl
or δvk . By measuring the ancilla qubit in the X-basis, we obtain the eigenvalue of the
star or plaquette operator. If an error is introduced during the measurement, we obtain
an incorrect syndrome value, which has to be dealt with in the topological quantum error
correction.
Again, we assume that the error is given as a Pauli error for simplicity. We also
assume that the X and Z errors, which might be correlated in general, are corrected
independently. The former assumption could be justified as follows. Any Kraus operator
of the noise map can be decomposed into a superposition of Pauli operators. Such a
superposition is collapsed into Pauli errors by the syndrome measurements, which map
different Pauli errors (of low weight) into orthogonal subspaces. Note that, in this case,
we have to model the error per gate carefully. The latter assumption makes the analysis
very simple, but only results in an underestimation of the noise threshold. Below we only
consider correction of the Z errors, but it can be applied straightforwardly to the X errors.
P
Let ce1 (t) = l zle (t)el be a 1-chain specifying the space-time Z error location, where
if zl (t) = 1, an Z error occurs on qubit el at time step t. The Z operators on the code
P
state at time t are denoted by c1 (t) = l zl (t) (i.e., Z[c1 (t)]).The state at time t satisfies
the following equation of the motion:

c1 (t + 1) = c1 (t) + ce1 (t + 1). (4.13)

At each time step t, we measure the star operator Bk (t) and obtain the measurement
outcome
 
M
mk (t) =  zl (t) ⊕ zke (t), (4.14)
el ∈δvk

where the bit zke (t) ∈ {0, 1} indicates an error on the measured syndrome, which we call
a measurement error. To obtain an equation consisting only of errors, we calculate the
parity of the syndrome at time steps t and t − 1,
 
M
sk (t) = mk (t) ⊕ mk (t − 1) =  zle (t) ⊕ zke (t) ⊕ zke (t − 1). (4.15)
el ∈δvk

Hence, errors on the measured syndrome can be detected by the parity (difference) of the
syndrome values at time t and t − 1. We redefine {sk (t)} as an error syndrome of the
space-time errors including the measurement error.
In the case of the perfect syndrome measurement, the error syndrome is given by
the boundary of errors ∂ce1 , which allows us to use the MWPM algorithm for the error

109
Figure 4.13: A mapping from the chain complex in 2D with time evolution to another
chain complex in 3D.

correction. What about the case with the faulty syndrome measurement? Using a chain
complex in a 3D manifold, we can again reformulate the error syndrome {sk (t)} as the
boundary of a space-time error chain as shown in Fig. 4.13. We consider a chain complex
{C0 , C1 , C2 , C3 } on a cubic lattice. The basis for the 3-chain is given by a set of cubes
P
{qr }. The 3-chain c3 = r zr qr ∈ C3 (zr ∈ {0, 1}) is a linear combination over Z2 . The
dual chain complex {C̄0 , C̄1 , C̄2 , C̄3 } is also defined. Specifically, Ci and C3−i are identified
via the duality transformation.
Let us explain how to embed the space-time error location to a 1-chain in the chain
complex in 3D. The edge el (t) and its coefficient zle (t) at time t are mapped into a vertical
face fm and its coefficient zm e in the 3D chain complex, respectively. The measurement

error z e (t) is mapped to the coefficient zm e of a horizontal face f ′ . Then, the space-time
′ m
error location including the measurement errors is described by a 2-chain or a dual 1-chain
in the 3D chain complex:
X X
ce2 = zme
fm , ↔ c̄e1 = e
zm ēm . (4.16)
m m

A syndrome sk (t) is assigned on each cube qr , or equivalently each dual vertex v̄r , in the
3D chain complex, and denoted by sr . Now we realize that
M
e
sr = zm . (4.17)
fm ∈qr

By using the same argument for the 2D case, sr = 1 iff v̄r belongs to the boundary ∂c̄e1

110
of the dual 1-chain c̄e1 , specifying the space-time error location. Thus, we can apply the
MWPM algorithm on the cubic lattice to perform topological quantum error correction in
space-time.
To calculate the noise thresholds, we have to model the noise distribution P (c̄e1 ). First,
we consider the simplest case, where the errors are located on each dual edge ēl of the
dual cubic lattice with probability p. This means that Z errors occur on each qubit
independently with probability p at each time step. Moreover, the measured syndrome is
flipped with probability p. Such a noise model is called a phenomenological noise model.
Using the MWPM algorithm, the noise threshold has been estimated to be 2.93% [WHP03].
The error correction problem under the phenomenological noise model can be mapped into
a phase transition of the random-plaquette Z2 gauge model by using the same argument
as in Sec. 3.5. Specifically, the loop condition for the dual 1-chain (primal 2-chain) in the
3D model can be solved by introducing a gauge spin on each primal edge σl and defining
P Q
the dual trivial 1-cycle c̄1 = m zm ēm ∈ Img(∂¯2 ) via (−1)zm = el ∈∂fm σl , where ēm and
fm are related through the duality relation. In this way, the variable on the face center is
provided as a product of the gauge spins on the boundary of the face, which leads to the
random-plaquette Z2 gauge model. The ordered Higgs and disordered confinement phases
correspond to fault-tolerant and non-fault-tolerant regions, respectively [DKLP02]. The
threshold value 2.93% corresponds to the critical point at zero temperature. The optimal
threshold is provided on the Nishimori line and has been estimated to be 3.3% using the
Monte Carlo simulation [OAIM04].
In fault-tolerant quantum computation, we have to consider any source of noise, in-
cluding the gate operations, during the syndrome measurement. As a standard way to
model a realistic situation, suppose that each elementary gate is followed by a depolarizing
channel. This is called a circuit-based noise model. More precisely, an ideal single-qubit
gate is followed by single-qubit depolarizing noise,
X p1
(1 − p1 )ρ + AρA. (4.18)
3
A∈{X,Y,Z}

An ideal two-qubit gate is followed by two-qubit depolarizing noise,

X p2
(1 − p2 )ρ + (A ⊗ B)ρ(A ⊗ B). (4.19)
15
A,B∈{I,X,Y,Z}\I⊗I

A faulty Pauli-basis state preparation is modeled by the state

I +A I −A
PA+ (pp ) = (1 − pp )+ pp , (4.20)
2 2
where A = X, Y, Z. A faulty Pauli-basis measurement is modeled by a POVM measure-
ment with POVM elements

{PA+ (pm ), PA− (pm )}, (4.21)

where
I −A I +A
PA− (pm ) = (1 − pm ) + pm . (4.22)
2 2

111
Thus, the measurement outcome is flipped with probability pm .
In Refs. [RHG06, RHG07, RH07], the error probabilities were parameterized as p1 =
p2 = (3/2)pp = (3/2)pm . The noise threshold with the MWPM algorithm was obtained
by numerical simulations to be p2 = 0.75%. In Ref. [WFH11], the threshold value was
further improved to ∼ 1% by assigning a weight for each edge in MWPM appropriately,
according to the amount of possible errors. Roughly speaking, the threshold is located
where the expectation value of each syndrome becomes 0.7:

h(−1)sr i = 0.7. (4.23)

In the phenomenological noise model, the expectation value is provided by

h(−1)sr i = (1 − 2p)6 , (4.24)

which hits 0.7 for p = 2.89%; thus being in good agreement with 2.93%. Moreover, in
the case of the circuit-based noise model, p is roughly given by pp + pm + 4p2 for the
measurement error, which means that one state preparation, one measurement, and four
two-qubit gates are employed. The error probability p for the code state is given by 4p2 ,
meaning that four two-qubit gates are employed in the syndrome measurements of the
star and plaquette operators. This yields

h(−1)sr i = (1 − 2pp − 2pm − 8p2 )2 (1 − 8p2 )4 , (4.25)

which hits 0.70 for p2 = 0.63% with p1 = p2 = (3/2)pp = (3/2)pm . This is again in a
good agreement with the numerical result 0.75%. This simple calculation provides rough
estimates of the threshold values, but there is no validity. If we need a more accurate
threshold value, we should perform a numerical simulation by taking noise propagation
and correlation into account [RHG07, WFH11].

4.7 Experimental progress

Topologically protected quantum computation in 2D has been utilized as a platform to
design fault-tolerant architectures for quantum computation. One promising approach is
on-chip monolithic architectures, such as quantum dots [VMLFY10, JVMF+ 12], silicon-
based nuclear spins [ONR+ 14], and superconducting qubits [GFG12, FMMC12, CGM+ 14,
CMS+ 14], where huge number of qubits are integrated on a single chip and each individ-
ual qubit and the interactions between the qubits are manipulated by multiple lasers or
electrodes. Among them, the superconducting qubits system is one of the most promis-
ing candidates for implementing topological quantum computation on the surface code,
because they can be fabricated on a 2D chip, and all elementary operations have al-
ready been demonstrated experimentally [NPT99, YPA+ 03, CNHM03, WSB+ 04]. The
coherence time and gate fidelity of the superconducting systems have improved rapidly.
An important breakthrough was made by Martinis’s group at University of California
Santa Barbara in 2014, where single-qubit gates with a fidelity of 99.92% and two-qubit

112
gates with a fidelity of 99.4% were demonstrated on a 1D array of five superconduct-
ing transmon qubits [BKM+ 14b, Gib14]. In addition, repetitive quantum non-demolition
measurements were demonstrated on a 1D array of 9 qubits, which improved the fidelity of
a state preparation by a factor of 8.5, even when using 9 qubits and faulty two-qubit gates
and measurements [KBF+ 14]. This is an important building block of the fault-tolerant
quantum computation on the surface code.
Another approach is the distributed modular architecture, where few-qubit local mod-
ules are connected with quantum channels mediating interactions between separate mod-
ules [LBSB10, FT10, FYKI12, LB12, NLB13, MRR+ 14]. The few-qubit quantum module
has already been experimentally realized in various physical systems, such as nitrogen-
vacancy centers in diamond and trapped ions. Furthermore, entangling operations be-
tween separate local modules have been experimentally demonstrated. This experimental
and theoretical progress will gradually lead us to large-scale quantum computation.

113
Chapter 5

Topologically protected MBQC

In this chapter, we reformulate topological fault-tolerant quantum computation explained

in the previous chapter in terms of meausrement-based quantum computation.

5.1 Topological cluster state in 3D

Consider a (primal) cubic lattice L and Z2 chain complex on it, {C0 , C1 , C2 , C3 }, where
X X
c0 = zk vk ∈ C0 , c1 = zl el ∈ C1 , (5.1)
k l
X X
c2 = zm fm ∈ C2 , c3 = zn qn ∈ C3 , (5.2)
m n

with zk , zl , zm , zn ∈ Z2 . We also consider a dual cubic lattice L̄ through the relations

vk ↔ q̄k , el ↔ f¯l , fm ↔ ēm , and qn ↔ v̄n .
Qubits are defined on the edges and faces of the primal lattice L (or equivalently primal
and dual edges), as shown in Fig. 5.1. We define an operator A(ci ) (i = 1, 2) in terms of
P P
a 1-chain c1 = j zj ej or a 2-chain c2 = j zj fj as
Y
A(ci ) = Azj . (5.3)
j

The stabilizer generators of a 3D cluster state for topologically protected MBQC are
defined on the primal and dual elementary faces fm , f¯m′ (see Fig. 5.2 (a)):

Kfm = Xfm Z(∂fm ), (5.4)

Kf¯m′ = X ¯ Z(∂ f¯m′ ).
fm′ (5.5)

A unit cell of the 3D cluster state is shown in Fig. 5.1. This notion of stabilizer generators
of the cluster state is quite useful; it provides a connection between the operators and the
chain complex as follows. By multiplying the two stabilizer operators Kfm and Kfm′ , we
have

Kfm Kfm′ = X(fm + fm′ )Z(∂(fm + fm′ )), (5.6)

114
Figure 5.1: A unit cell of the cluster state for topological MBQC. The primal and dual
cubes, faces, and edges are also shown.

(see also Fig. 5.2 (b)). By using this property, we can define a stabilizer operator on a
2-chain c2 ,
Y
K(c2 ) ≡ Kfzm
m
= X(c2 )Z(∂c2 ), (5.7)
m

(see also Fig. 5.2 (c)). Furthermore, for the two 2-chains, c2 and c′2 , we have

K(c2 + c′2 ) = K(c2 )K(c′2 ). (5.8)

Let us see how the 3D cluster state is related to topological quantum computation on
the surface code explained in Chapter 4. Recall the circuit diagrams for the syndrome
measurements of the plaqette and star operators in Fig. 4.12. The measurement for the
plaquette operator is done by applying the CZ gates between the ancilla qubit on the face
center and the four qubits on the edges. This operation is the same as generation of the
cluster state stabilized by K(fm ) with a horizontal face fm . The syndrome measurement
for the star operator can be done by the CZ gates with the basis change by the Hadamard
gates. This corresponds to generation of the cluster state stabilized by K(f¯l ) with a
horizontal dual face f¯l . Moreover, the horizontal edge qubits, which constitute the surface
code, are connected by applying the CZ gates vertically in order to perform the Hadamard
gates for the basis change. In this way, we recover the 3D cluster state stabilized by
K(fm ) and K(f¯l ) for all primal and dual faces fm and f¯l . Two of three dimensions are
employed for the spatial degrees of freedom, constituting the surface code. One is for
the time evolution of measurement-based quantum computation. The measurements are

115
Figure 5.2: (a) A stabilizer operator
Q defined on a primal face. (b) Kfm Kfm′ = X(fm +
fm′ )Z(∂(fm + fm′ )). (c) K(c2 ) ≡ m Kfzm
m
= X(c2 )Z(∂c2 ).

done along the time-like axis, where even and odd layers, corresponding to the syndrome
measurements of the plaquette and star operators respectively, together with constitute
an elementary time step of the topologically protected MBQC. Below we will see how the
topological operations on the surface code are translated into a measurement pattern of
the MBQC on the 3D cluster state.

5.2 Vacuum, defect, and singular qubit regions

The cubic lattice is divided into three regions: the vacuum V, defect D, and singular qubits
S (the detailed definitions are provided later). In the vacuum region, the topological
quantum computation is protected through topological quantum error correction. The
defect regions are utilized to implement topological quantum computation by braiding
defects. We have two types of defects: the primal (D) and dual (D̄) defects. For simplicity,
we only consider the primal defect. The extension to the dual case is straightforward by
replacing primal by dual in the derivation. The primal defect D is defined as a set of
primal cubes. The primal face qubits inside the primal defect (except for those on the
boundary ∂D) are measured in the Z-basis to remove the corresponding bonds of the
cluster state (or, equivalently, we can prepare the cluster state without those bonds from
the beginning). On the boundary ∂D, the primal face qubits are measured in the X-

116
 (t) (t)
Figure 5.3: (a) A defect pair logical qubit. (b) The logical operators LZ and LX at time
step t.

basis. The primal edge qubits belonging to the primal defect (including its boundary) are
measured in the X-basis. The measurement pattern for the dual defect is defined similarly.
Unfortunately, only Clifford circuits such as Pauli-basis preparations, measurements,
and CNOT gates, are implemented in a topologically protected way. For universal quan-
tum computation, magic states for the non-Clifford gates are injected on the singular
qubits, which are always located in-between two defects. The injections are executed by
√
measuring the singular qubits in the Y - and (X + Y )/ 2-bases. These measurements cor-
√ √
respond to injections of (|0i+e−iπ/2 |1i)/ 2 and (|0i+e−iπ/4 |1i)/ 2 (up to global phases),
which are utilized to implement the S = e−i(π/4)Z and T = e−i(π/8)Z gates via gate tele-
portation, respectively. The singular qubits are not topologically protected, because two
defects are made close to each other resulting in shortening the code distance. However,
we can obtain clean magic states with topologically protected Clifford gates through the
magic state distillation protocols [BK05]. In this way, an arbitrary quantum computation
is executed fault-tolerantly. Below, we will define these three regions more precisely and
see how topological quantum computation is executed in a measurement-based way.

5.3 Elementary operations in topological MBQC

Definition of a logical qubit
The logical information is encoded by using a pair of two defects as shown in Fig. 5.3
(a) and (b), where the measurements are done from left to right. The logical degree of
information at time step t is described a primal 1-chain c1 surrounding the defect and a
dual 1-chain c̄1 connecting the two defects, as shown in Fig. 5.3 (a). After measuring qubits
up to the (t − 1)th even and odd layers, according to the measurement patterns presented
before, the following two operators may become logical operators, which commute with

117
Figure 5.4: A logical identity gate. The logical operators at time steps t and t′ are related
by the correlation surfaces K(c2 ) and K(c̄2 ) via the measurements.

the stabilizer group of the remaining cluster state and are independent of it:

LZ = Z(c1 ), LX = X(←
c̄−
(t) (t)
1 )Z(c̄1 ), (5.9)

where ← c̄−
1 indicates the dual face qubits on the even layer at time step t that are the left
neighbor of c̄1 , as shown in Fig. 5.3 (b). These two operators anticommute with each
other and represent a logical qubit. (If the cluster state ends at the even layer at the time
step t, then the two logical operators are equivalent to the logical operators of the surface
code. Because there are the time-like CZ gates for the Hadamard gates, the logical X
operator in Eq. (5.9) accompanied by the Z operators.)

Identity gate
Next, we will see how these logical operators evolve with the measurements. We consider
a correlation surface defined by a primal 2-chain c2 and a dual 2-chain c̄2 as shown in
Fig. 5.4. A stabilizer operator K(c2 ) on the correlation surface c2 surrounding the defect
is obtained by multiplying the stabilizer generators Kfm on the primal 2-chain c2 :
Y
K(c2 ) ≡ Kfzm
m
= Z(∂c2 )X(c2 ). (5.10)
m

Similarly, a stabilizer operator on the dual correlation surface c̄2 is defined by multiplying
Kf¯m on the dual 2-chain c̄2 :
Y
K(c̄2 ) ≡ Kfz¯m = Z(∂c̄2 )X(c̄2 ). (5.11)
m
m

118
 (t)
Figure 5.5: (a) A logical Z-basis state preparation. (b) K(c′2 ) = X(c′2 )LZ L′Z (t)

Suppose measurements are done from the left to the right, except for those qubits on the
final even layer. Using the correlation surface, we obtain equivalence relations between
the logical operators at time t and t′ :
(t) (t) (t′ )
LZ ∼ Z(c1 )K(c2 ) = Z(c1 )X(c2 ). (5.12)
← −
∼ X(c̄1 )Z(c̄1 )K(c̄2 ) = X(←
c̄−
(t) (t) (t) (t′ )
LX 1 )X(c̄2 )Z(c̄1 ) (5.13)

Here, A ∼ B means that A and B are equivalent up to a multiplication of the stabilizer

operator of the cluster state, meaning that both A and B act the same on the cluster
state. When the qubits on c2 have been measured in the X-basis, we obtain
(t) (t+1)
LZ ∼ LZ , (5.14)
(t) (t+1)
LX ∼ LX , (5.15)

where we assumed that all measurement outcomes are +1 for simplicity. This relation
indicates that the logical information at time step t is propagated to time step t′ without
any operation, i.e., a logical identity operation.

State preparation and measurement

Next, we consider how the logical qubit is prepared from the vacuum. To prepare the
(t)
eigenstate of LZ , we utilize the defect shown in Fig. 5.5 (a). By considering the correlation
surface c2 , we obtain
(t)
K(c2 ) = X(c2 )LZ . (5.16)

Because X(c2 ) commutes with the measurements, the state at time step t is stabilized by
(t)
LZ , and hence a logical Z-basis state is prepared. Considering another surface c′2 , shown
(t)
in Fig. 5.5 (b), the state at time step t is also stabilized by Z(∂c′2 ) = LZ L′Z (t) . Thus the
(t)
pair of the defects is appropriately encoded into the code space. Both LZ and L′Z (t) act
equivalently as logical Z operators.

119
 Figure 5.6: A logical X-basis state preparation.

Next, we consider the defect shown in Fig. 5.6. Considering a correlation surface c̄2 ,
we obtain

K(c̄2 ) = Z(∂c̄2 )X(c̄2 ). (5.17)

After the measurements, the state at time step t is stabilized by LX = X(← c̄−
(t)
1 )Z(c1 ), where
c1 is a 1-chain on the tth even layer connecting the two defects. Thus, a logical X-basis
state is prepared. Again, the logical state is stabilized by Z(c1 )Z(c′1 ), with c1 and c′1 being
(t) (t′ )
a cycle surrounding each defect. Hence, we can choose either LZ = Z(c1 ) or LZ = Z(c′1 )
to serve as the logical operator. The logical measurements of the defect pair qubits can
be done with the same defects as the state preparations, but by reversing the time-like
direction.

CNOT gate by braiding

Let us consider primal defects braiding around a dual defect as shown in Fig. 5.7 (a).
Similarly to the previous case, we calculate the time evolution of logical operators by the
(t) (t)
measurements. The state at time step t is described by {LZ , LX } and {L′Z (t) , L′X (t) }
(t) (t)
corresponding to {c1 , c̄1 } and {c′1 (t) , c̄′1 (t) }, respectively. We first consider a correlation
(t) (t)
surface c2 with respect to c1 , as depicted in Fig. 5.7 (b). Similarly to the identity gate, LZ
(t)
is transformed into L′ Z . An interesting thing happens when we consider the correlation
surface c′2 with respect to c′1 (t) , as shown in Fig. 5.7 (c). The stabilizer operator on c′2 is

120
Figure 5.7: (a) A diagram for defect braiding for a logical CNOT gate. (b-e) Time
evolutions of the logical operators and the corresponding correlation surfaces.

given by

K(c′2 ) = Z(∂c′2 )X(c′2 )

(t) ′ (t′ )
= Z(c′1 )Z(c1 (t ) )Z(c′1 )Z(c′′1 )X(c′2 ), (5.18)

where c′′1 is a cycle in the dual defect of a loop as shown in Fig. 5.7 (c). Then after the
measurements, we obtain an equivalence relation,
(t) (t′ ) (t′ )
LZ ∼ LZ L′ Z . (5.19)

Note that inside the dual defect region, the dual face (primal edge) qubits are measured
in the Z-basis, and hence we can obtain the eigenvalue of Z(c′′1 ). After a similar argument
using a defect surface c̄2 and c̄′2 with respect to the dual 1-chain c̄1 and c̄′1 , shown in
Fig. 5.7 (d) and (e), respectively, we obtain
(t) (t)
LX ∼ LX , (5.20)
(t′ ) (t′ )
(t′ )
LX ∼ LX L′ X . (5.21)

These relations between the logical operators at time steps t and t′ are equivalent to those
for the CNOT gate. Thus, the defect braiding in Fig. 5.7 results in a logical CNOT gate.
Now we realize that the correlation surface introduced in Chapter 4 as a trajectory of the
logical operator corresponds to the correlation surface defined by the stabilizer operator
of the cluster state.

121
 Figure 5.8: A state injection on the singular qubit.

A singular qubit injection for magic state distillation

So far, we have shown that the Clifford circuits, Pauli-basis preparation, measurements,
and CNOT gate, can all be implemented in a topological way. Unfortunately, these opera-
tions are not enough to generate universal quantum computation. To implement universal
√
quantum computation, we inject Y - and (X + Y )/ 2-basis states by measuring singular
√
qubits in the Y - and (X + Y )/ 2-bases, respectively, as shown in Fig. 5.8. Let us see how
this measurement works. Similarly to the previous case, we have two correlation surfaces
c2 and c̄2 :

K(c2 ) = Z(∂c2 )X(c2 ) = Z(∂c2 )X(c2 \s)Xs , (5.22)

K(c̄2 ) = Z(∂c̄2 )X(c̄2 ) = Zs Z(∂c̄2 \s)X(c̄2 ), (5.23)

where As is a Pauli operator on the singular qubit and [·]\s indicates a chain with a
removal of an element corresponding to the singular qubit. Suppose the singular qubit is
measured in the Y -basis. After the measurements, the state at time step t is stabilized by
(t) (t) (t)
K(c2 )K(c̄2 ) ≃ LX LZ ≡ LY . (5.24)

Thus, a logical Y -basis state is prepared. When the singular qubit is measured in the
√
(X + Y )/ 2-basis, the state at time step t is stabilized by
√ (t) (t) √
[K(c̄2 ) + K(c2 )K(c̄2 )]/ 2 ≃ (LX + LY )/ 2, (5.25)
√
which means that a logical (X + Y )/ 2-basis state is prepared.
These states are utilized to implement S, T , HSH, and HT H gates using gate tele-
portation with the CNOT gate. These gates form a universal set of gates.

122
5.4 Topological quantum error correction in 3D
Next, we will see how topological quantum error correction is done in 3D. Indeed, in the
3D case, the argument for the noisy syndrome measurements made in Chapter 4 becomes
more simple as follows. All measurements in the vacuum region are done in the X-basis.
We consider a stabilizer operator on a unit primal cube q,
Y
K(∂qn ) = Xfm , (5.26)
fm ∈∂qn

where there is no Z operator due to ∂ ◦ ∂qn = 0. This implies that, if there is no

error, the parity of each six X-basis measurement outcomes on the primal cube is always
even. The errors are described by using a dual 1-chain E = Z(c̄1 ). At a unit cube qn
belonging to ∂c̄1 , we have |qn ∩ ∂c̄1 | = odd. (Recall that the primal 3-chain and the
dual 0-chain are identified.) From a set of odd parity cubes ∂c̄1 , we estimate the actual
location of errors E ′ = Z(c̄′1 ) such that ∂c̄1 = ∂c̄′1 . If the total of the actual and estimated
error chains c̄1 + c̄′1 results in a trivial cycle, meaning that there is no defect inside the
cycle, it can be contracted and removed by a continuous deformation. If the total of the
actual and estimated error chains c̄1 + c̄′1 results in a nontrivial cycle, meaning a cycle
winding around a defect, EE ′ = Z(c̄1 + c̄′1 ) may result in a logical operator. In such a
case, the topological error correction has failed. This property is completely the same
as the topological quantum error correction under faulty syndrome measurements argued
in Sec. 4.6. If the error probability is smaller than a constant value (the threshold), the
failure probability of the topological quantum error correction decreases exponentially in
the characteristic size and distance of the defects.
Inside the defect region, the face qubits are measured in the Z-basis. Especially,
the Z-basis measurement outcomes near the defect boundary are employed to evaluate
the correlation surface. Note that these Z-basis measurements and the removal of the
corresponding bonds of the cluster state can instead be done by generating a cluster
without connecting the corresponding bonds in advance. In such a case, the errors on the
Z-basis measurements do not appear. We can obtain an additional parity Kfm and Kf¯m′
at the primal and dual faces on the boundary of the defects, respectively. If the errors
on the face qubits fm and f¯m′ are suppressed, the errors on the boundary are reduced
into errors on a toric code on a 2D surface, ∂D or ∂ D̄. Again, if the error probability is
sufficiently smaller than a constant threshold value, we can correct it faithfully.
The X- and Z-basis state preparations and measurements, and the CNOT gate ob-
tained by braiding, are topologically protected because we can execute these topological
operations by keeping the defect size and distance larger than an arbitrarily large con-
stant length. Unfortunately, through the state injection, we shrink the defect size into
an elementary unit cell, where the defect size and distance become very small. Thus, the
topological protection is broken down around the singular qubit (see Fig. 5.8 (a)). There
will also be lower weight errors, which effectively increase the logical error probability on
the injected logical states. However, noisy injected states can be purified by using the

123
topologically protected Clifford gates, the so-called magic state distillation. The Y - and
√
(X + Y )/ 2-basis states are distilled by using the 7-qubit Steane and 15-qubit Reed-
Muller codes, respectively as explained. The distilled states are utilized to implement
non-Clifford gates via gate teleportation, as seen before. In this way, universal quantum
computation is executed with arbitrary accuracy.

5.5 Applications for MBQC on thermal states

Topologically protected MBQC in 3D is useful to study the quantum computational ca-
pacity of quantum many-body states at finite temperature. Consider the stabilizer Hamil-
tonian of the 3D cluster state for topological MBQC [RBB03, RBH05, BBD+ 09]:
 
X X
Hfc = −J  K(f ) + K(f¯) . (5.27)
f f¯

The thermal state at temperature T = 1/(βJ) is given by

ρfc = e−βHfc /Tr[e−βHfc ]. (5.28)
Q
Using a unitary operator UCZ = (fm ,f¯l ) Λfm ,f¯l (Z), consisting of CZ gates on all nearest-
neighbor two qubits, the thermal state can be mapped into the thermal state of an
interaction-free spin model:
†
UCZ ρfc UCZ = e−βHf /Tr[e−βHf ], (5.29)
where
X †
Hf ≡ −J Xi = UCZ Hfc UCZ . (5.30)
i
Because Hf is an interaction-free Hamiltonian, the stabilizer Hamiltonian, which we will
call the free cluster Hamiltonian, does not undergo any thermodynamic phase transition.
The thermal state of the free cluster Hamiltonian is given as a product state of the
single-spin density matrix:
Y
ρf = e−βHf /Tr[e−βHf ] = eβJXi /Tr[eβJXi ] (5.31)
i
Y
⊗n
= Ei (pβJ )(|+ih+|) , (5.32)
i

where
Ei (p) = (1 − p)ρ + pZi ρZi , (5.33)
and pβJ = e−2βJ /(1 + e−2βJ ). Because Ei and UCZ commute, the thermal state of Hfc is
rewritten as
" # " #
†
Y †
Y
ρfc = UCZ ρf UCZ = Ei (pβJ ) UCZ (|+ih+|)⊗n UCZ = Ei (pβJ ) |Ψ3D ihΨ3D |,
i i
(5.34)

124
 Figure 5.9: A PEPS picture of the cluster state.

where |Ψ3D i is the ground state of Hfc , i.e., the 3D cluster state. This means that the
thermal state is given as an ideal 3D cluster state, followed by an independent dephasing
for each qubit with probability pβJ = e−2βJ /(1 + e−2βJ ). From the argument made in
the previous section, if p ≤ 2.9 − 3.3% and hence T = 1/(βJ) ≤ 0.57 − 0.59, we can
then perform universal quantum computation reliably on the thermal state at a finite
temperature, where the errors originating from the thermal excitations are corrected by
the topological quantum error correction. On the other hand, in the high temperature
limit T = 1/(βJ) → ∞, the thermal state is given by a completely mixed state, and hence
MBQC on it can be simulated classically.
A projected-entangled-pair state (PEPS) picture [VC04, RBH05, BBD+ 09] allows us to
obtain a lower bound for the possibility of a classical simulation. In the PEPS picture, the
3D cluster state is described as follows. A maximally entangled pair |ψMES i ≡ (|0i|+i +
√
|1i|−i)/ 2 is shared on each bond. On each site consisting of halves of the entangled pair,
a projection

|0ih00..0| + |1ih11..1| (5.35)

is performed with an appropriate normalization. The resultant state is the 3D cluster

state. Because the projection and the Z error Ei commute, the effect of the thermal
excitations on the shared entangled state can be determined beforehand:

ρbond = Ea (p)Eb (p)|ψMES ihψMES | (5.36)

√
If p ≥ (2 − 2)/2, the decohered entangled pair ρbond becomes a separable state. If two
bonds per site are made separable, the 3D cluster state becomes a separable state. A
sampling on such a resource state can be simulated efficiently classically [BBD+ 09, FT14].
√
In this case, the Z error probability per site has to be pβJ ≥ 2−1, i.e., T = 1/(β) = 5.77J.
The true critical temperature Tc between the classically simulatable and universal quantum
computational phases is located in the range 0.59J < Tc < 5.77J. Note that this model
exhibits a transition of the computational capability, while there is no thermodynamic
phase transition in the physical system [RBH05, BBD+ 09].
In classical information processing, a thermodynamic phase transition, or, more pre-
cisely, an ordered phase below a critical temperature is utilized for robust information

125
storage in magnetic storage devises. While there is no such long range order in the pre-
vious model, it is natural to ask whether or not a long range ordered phase is useful to
enhance the measurement-based quantum computation on many-body thermal states for
quantum information processing. To address this issue, Fujii et al. proposed an interacting
cluster Hamiltonian [FNOM13],
X
Hic = −J K(f )K(f¯).
hf,f¯i

Because interactions are introduced between the cluster stabilizers, this model is mapped
by UCZ into an Ising model on a 3D lattice:
†
X
HIsing = UCZ HicUCZ = −J Xf Xf¯.
hf,f¯i

Thus, it undergoes a thermodynamic phase transition at a finite temperature. The degen-

erate ground states

UCZ |+i⊗n and UCZ |−i⊗n (5.37)

are also the 3D cluster states, up to the simultaneous spin flipping due to the global
symmetry. Because the eigenvalues of the cluster stabilizer have a long range order (they
are likely to be aligned in the same direction) in a ferromagnetic ordered phase, the
topologically protected MBQC on the symmetry-breaking thermal state has a special
robustness against the thermal excitations. In Ref. [FNOM13], topological quantum error
correction of this model is mapped to a correlated random plaquette Z2 -gauge model in
3D, where the disorder in the signs of the plaquettes has an Ising-type correlation. By
using this property and the gauge transformation on the Nishimori line [Nis01], Fujii et al.
showed that the critical temperature of this model, and hence the threshold temperature
for topological protection, is equal to the critical temperature of the 3D Ising model,
which is the unitary equivalent model of the interacting cluster Hamiltonian. This means
that the critical temperatures for the topological protection and the thermodynamic phase
transition of the underlying physical system coincides exactly. Due to this fact, we can
improve the threshold temperature for topological protection by one order of magnitude.
While the above Hamiltonian employs multi-body interactions, the 3D cluster state
can be generated from the thermal states of a nearest-neighbor two-body Hamiltonian for
spin-3/2 and composite spin-1/2 particles via local filtering operations [LBK+ 11, FM12].
Let us consider a system consisting of a spin-3/2 particle located at site r and a composite
particle of two spin-1/2 particles located at the nearest-neighbor site r + i, with i = 1, 2, 3,
as shown in Fig 5.10 (a). The Hamiltonian is given by
X
H = ∆ ~r · (I~r+1 + I~r+2 + I~r+3 )
S
r

where ~r ≡ (S x , Sry , S z ) is the spin-3/2 operator of the center particle at the position
S r r
r and ~ ~
Ir+a = Ar+a or B ~ r+a depending on the interaction types (line or dash). Here,

126
~ r+a ≡ (Ax , Ay , Az ) and B
A ~ r+a ≡ (B x , B y , B z ) are two independent spin-
r+a r+a r+a r+a r+a r+a
1/2 operators on the composite particle at the position r + a (a = 1, 2, 3). The above
Hamiltonian H can be reformulated as
X X
H = Hr = ∆/2 (T~r2 − S
~ 2 − I~2 )
r r
r r

~r + I~r . The ground state |Gi = N |gr i is given

where I~r ≡ I~r+1 + I~r+2 + I~r+3 and T~r ≡ S r
by Tr = 0, Sr = 3/2, and Ir = 3/2, where Lr (Lr + 1) (L = T, S, I) is the eigenvalue of the
operator L~ 2 . Each center particle in the ground state |Gi is filtered by using the POVM
r
measurement:
√
{F α = (Srα 2 − 1/4)/ 6} (α = x, y, z). (5.38)

If the measurement outcome is α = z, we obtain a four-qubit GHZ (Greenberger-Horne-

Zeilinger) state [GHZ89] as the post-POVM measurement state:
1
|GHZ4r i ≡ √ (|0̃ + ++i + |1̃ − −−i),
2
where |1̃i and |0̃i are eigenstates of S z with eigenvalues +3/2 and −3/2, respectively, and
|±i are the eigenstates of Az or B z with eigenvalues ±1, respectively. Even if we obtain
other outcomes, we can transform the post-POVM measurement state to |GHZ4r i by local
operations. The four-qubit GHZ state is subsequently used to construct the 2D honeycomb
cluster state, which is a universal resource for MBQC, by measuring the operators Az ⊗B x
and Ax ⊗ B z on the bond particle as shown in Fig. 5.10 (a).
N
In the case of finite temperature, we have the thermal state r ρr with ρr ≡ e−βHr /Z
instead of the ground state, where Z indicates the partition function and β = T −1 for
a temperature T . Then, the GHZ state becomes a noisy, say thermal, GHZ state, σr ≡
F α ρr F α† /Tr[F α ρr F α† ]. In the low temperature case, the thermal GHZ state is calculated,
in the leading order, to be E4 (|GHZ4r ihGHZ4r |) with

E4 = (1 − q1 − 3q2 − 3q3 )[I] + q1 [Zr ]

X X
+q2 [Zr+a ] + q3 [Zr Zr+a ], (5.39)
a=1,2,3 a=1,2,3

where q1 , q2 , and q3 are the error probabilities as functions of the temperature T , the
Pauli Z operator Zb on the qubit at the position b, and [C]ρ ≡ CρC †, respectively. The
probability of other errors such as Zr Zr+a Zr+a′ is several orders of magnitude smaller
than q1,2,3 .
To obtain the 3D cluster state for topological MBQC, as done in Ref. [LBK+ 11], the
five-qubit GHZ state |GHZ5r i is generated in a similar way by using spin-2 particles and
composite particles of spin-1/2, as shown in Fig. 5.10 (b). Instead of the spin-2 particles,
spin-3/2 particles were employed in Ref. [FM12] to obtain the 3D cluster state shown in
Fig. 5.10 (c). After the filtering operation and local operations, the two four-qubit GHZ
states are connected to obtain the five-qubit GHZ state for building the 3D cluster state.

127
Figure 5.10: (a) A system consisting of spin-3/2 particles and composite particles of two
spin-1/2. After the filtering operation on the ground state, we obtain cluster states. (b)
A system consisting of spin-2 particles and composite particles of two spin-1/2 for the 3D
cluster state. (c) A system consisting of spin-3/2 particles and composite particles of two
spin-1/2 for the 3D cluster state.

By using the threshold for topologically protected MBQC, we can calculate the thresh-
old temperatures T = 0.21∆ and T = 0.18∆ for the cases of spin-2 and spin-3/2 center par-
ticles, respectively [LBK+ 11, FM12]. Accordingly, we can perform fault-tolerant universal
measurement-based quantum computation even on the thermal states of local two-body
Hamiltonians at finite temperature.

128
Appendix A

Fault-tolerant quantum
computation

A.1 Fault-tolerant syndrome measurements

In the case of the stabilizer code, the syndrome measurements are done by simply mea-
suring the stabilizer operators. Several QEC gadgets have been proposed to implement
the stabilizer measurement fault-tolerantly [DS96, Kni05, Ste97].
DiVincenzo-Shor’s gadget— The first QEC gadget was proposed by David DiVincenzo
and Peter Shor, who used cat states as ancillae for the syndrome measurement [DS96]. It
is based on an indirect measurement of the observable A with the eigenvalues ±1:

(A.1)

For example, the stabilizer S1 of the seven-qubit code can be measured as the transversal
X measurements of the corresponding physical qubits in the code block:

(A.2)

129
Unfortunately, this measurement is not fault tolerant, because the errors in the CNOT
gates [A = X in Eq. (A.1)] are spread by the following CNOT gates, as shown in the
√
above circuit. To make it fault-tolerant, a cat state |cati = (|000 · · · 0i + |111 · · · 1i)/ 2 is
used as an ancilla for the measurement as follows:

(A.3)

where the cat state is verified before connecting with the code state. Because the qubits
in the code block interact with different ancilla qubits, this measurement does not spread
the errors in the CNOT gates. Similarly, other stabilizers S2 , · · · , S6 are measured fault-
tolerantly to obtain the error syndrome. Instead of the verification, one can perform a
suitable recovery operation by postprocessing the ancilla state after its interaction with
the code state [DA07]:

(A.4)

The DiVincenzo-Shor’s QEC gadget and its improved version both require a lot of physical
gate operations, which results in deterioration of the performance.
Steane’s gadget— Subsequently, a relatively simple QEC gadget was proposed by An-
drew Steane [Ste97], where encoded ancilla states are used to extract the syndrome with
transversal operations. In particular, for the case of the CSS code, the logical code states

130
can be used as ancilla states. The following circuit executes the Z and X error syndrome
extractions by using the ancilla |0L i states,

(A.5)

Because the ancilla states are the logical code states, one can obtain the error syndrome
by simply measuring the ancilla states. The syndrome extraction is repeated a couple of
times to extract reliable error information. An optimized way to extract the syndrome
information was proposed in Ref. [PVK97], where the subsequent syndrome extraction
is performed conditionally according to the preceding syndrome information. For these
schemes to work fault-tolerantly, the encoded ancilla |0L i states have to be prepared with
high fidelity. This is achieved by using either verification or entanglement purification
[Kni05, Ste97, DAB03, ADB05].
Knill’s gadget— Another interesting QEC gadget was proposed by Emanuel Knill
[Kni05]. It is based on quantum teleportation as illustrated in the following circuit:

(A.6)

Here, the encoded data qubit |ψL i is teleported to the fresh encoded qubit of the ancilla
Bell state. Thus, the encoded ancilla Bell state has to be prepared with high fidelity
by using verification or entanglement purification, similarly to the Steane’s gadget. The
logical Bell measurement completes the teleportation, namely error-correcting teleporta-
tion. There is no need to identify the error syndrome, but it is sufficient to find the
logical measurement outcomes of the logical Bell measurement. Thus, it is not neces-
sary to repeat the syndrome extraction in this QEC gadget. The outcome of the Bell
measurement is properly propagated to the subsequent computation as the Pauli frame
[Kni05, DHN06b, DHN06a].

131
A.2 Fault-tolerant gate operations
Fault-tolerant computation is now executed by the logical gate operations that are followed
by the QEC gadgets. This is illustrated for the fault-tolerant CNOT gate as follows:

(A.7)

where the code block is depicted as though it is a three-qubit code. A QEC gadget
is attached to each logical output of the transversal CNOT gate. Because a single error
never will be propagated as multiple errors in a fault-tolerant gate, a logical error is caused
by two (or more) simultaneous physical errors. Denoting the number of such faulty pairs
of error locations as C, the logical error probability is given by Cp2 , with p being the
physical error probability. If the physical error probability is sufficiently small, p < 1/C,
one can improve the accuracy of the gate and achieve a fault-tolerant computation.

A.3 Concatenated quantum computation

For a reliable computation of a large size, the logical error probability should be reduced
arbitrarily. This is done by a concatenated fault-tolerant computation [KLZ98b, KLZ98a,
ABO97, ABO08]. In the concatenated computation, each physical gate operation is re-
peatedly replaced by a logical gate operation followed by QEC gadgets.
Suppose A is a quantum computation, which consists of some physical gate operations.
Then the first level concatenated computation is defined by C(A), where the operation C
indicates replacing each physical gate with a logical one, followed by QEC gadgets. For
example, C(CNOT) is described in the diagram (A.7). The second level concatenated
CNOT gate C ◦ C(CNOT) is also described as follows:

(A.8)

where each physical gate in the QEC gadgets is replaced by the logical one, followed by
the QEC gadgets. By repeating this procedure, the lth level concatenated computation of

132
A is given by C l (A). Specifically, for a physical gate operation G (e.g., Hadamard, CNOT
gate, etc.), C l (G) is called the level-l G gate. The lth level concatenated code state is
called the level-l qubit, denoted by|0(l) i, |+(l) i.
(1)
As mentioned previously, the logical error probability pg of the level-1 gate is given
by
p(1) = C(p(0) )2 , (A.9)
where p(0) = p and C denotes the number of all faulty pairs of error locations. The constant
C differs between the logical gates. It is, however, sufficient to choose the maximum value.
Due to the self-similarity of the concatenation, the logical error probability of the level-2
gate is given in terms of p(1) by
p(2) = C(p(1) )2 . (A.10)
Similarly, the logical error probability p(l) of the level-l gate is obtained recursively as

p(l) = C(p(l−1) )2
l
= (Cp(0) )2 /C. (A.11)

We conclude that if p(0) < pth ≡ 1/C, the logical error probability can be reduced super-
exponentially with the concatenation level l. This is the so-called threshold condition.
On the other hand, the resources usage, R(l) , consumed for the level-l gate is estimated
roughly as
R(l) = N l , (A.12)
where N indicates the total number of physical gates in the level-1 gate. Suppose that the
size of the computation is 10n−1 = M . Then an accuracy of p(l) < 10−n is required for
each logical gate at the highest level. The total resources to perform a reliable quantum
computation of size M amount to
 log2 N
l̄ n
Rtot = N M = M = poly(log(M ))M, (A.13)
log10 (Cp(0) )−1
 
where ¯l ≃ log2 n/ log10 (Cp(0) ) is the number of levels necessary to achieve the required
accuracy. This result clearly shows that if the physical error probability p(0) is smaller
than pth , one can execute quantum computation to an arbitrary accuracy with only poly-
logarithmic overhead. This is the celebrated threshold theorem and the critical value pth is
called the noise threshold [NC00, Kit97, KLZ98b, KLZ98a, ABO97]. The noise thresholds
have been calculated to be about 10−4 −10−2 for several fault-tolerant schemes under vary-
ing degrees of assumption and rigor [KLZ98b, KLZ98a, ABO97, ABO08, Ste03, Kni05,
AGP06, AC07, AGP08, AP09, Ste99, CDT09].

133
Appendix B

Decoding stabilizer codes

Consider an n-qubit stabilizer code, whose stabilizer group is given by G = {Gi }. The
group consisting of the logical operators is denoted by L = {Li }. Suppose a Pauli product
E ∈ {I, X, Y, Z}⊗n acts as an error on the stabilizer code state. The error syndrome, a
set of eigenvalues of the stabilizer generators, of an error E is denoted by S = S(E). For
each error syndrome S, we define a pure error operator R(S) ∈ {I, X, Y, Z}⊗n such that
S = S[R(S)]. The pure error operator R(S) is chosen arbitrarily as long as S = S[R(S)].
The error E is decomposed uniquely into logical, stabilizer, and pure error operators:

E = Li Gj R[S(E)]. (B.1)

We define a decoding map D, which computes the logical operator Li from the error E,
i.e., Li = D(E).
The decoding problem consists of finding an optimal logical operator Li that maximizes
the posterior probability P (L|S):

arg max P (Li |S). (B.2)

Li ∈L

The posterior probability is given by

1 X
P (Li |S) = P (E)δ[Li = D(E)]δ[S = S(E)] (B.3)
N
E
1 X
= P [Li Gj R(S)]. (B.4)
N
Gj ∈G

where 1/N is a normalization factor, δ(· · ·) is an indicator function, and P (E) is the
probability of the error E. In general, computing P (L|S) is hard because the summation
P
over all stabilizer operators, Gj , takes an exponential time. However, if the code has a
good structure, such as a concatenated code, we can manage it efficiently.
A concatenated code is defined recursively using the logical qubits at the lower level
(k)
as physical qubits at the higher level. Using the logical Pauli operators Li ∈ L(k) at the
(k+1)
kth level (k = 1, 2, ..), we define a stabilizer group G (k+1) = {Gj } and logical operators

134
 (k+1)
L(k+1) = {Li } at the (k + 1)th level. The whole stabilizer group of the concatenated
code is given by the union of all stabilizer groups ∪lk=1 G (k) , where we consider the lth
concatenated code. A logical operator of the lth concatenated code is given by a logical
(l)
operator at the highest level, Li ∈ L(l) . At each level, we define a level-k error E (k) . In
the case of k = 0, the level-0 error E (0) ∈ {I, X, Y, Z}⊗n is a physical error. At a higher
level, the level-k error E (k) is a level-k logical operator, which will be defined later. At
each level, we define an error syndrome S (k) = S (k) (E (k−1) ), i.e., a set of eigenvalues of
(k)
the level-k stabilizer operators {Gj } under the level-(k − 1) error E (k−1) . The level-k
pure error operator R(k) (S (k) ) is defined arbitrarily such that S (k) [R(k) (S (k) )] = S (k) .
The level-k error E (k) is now defined recursively as follows. Any physical error E (0)
can be decomposed into the stabilizer, logical, and pure error operators of level-1:
(1) (1)
E (0) = Gj Li R(1) (S (1) ). (B.5)

(1) (1)
The obtained level-1 logical operator Li = E (0) Gj R(1) (S (1) ) is further regarded as
(1)
the level-1 error E (1) ≡ Li . Recursively, the level-k error E (k) is decomposed into the
stabilizer, logical, and pure error operators of level-(k + 1):
(k+1) (k+1)
E (k) = Gj Li R(k+1) (S (k+1) ). (B.6)

(k+1) (k+1)
Then, we define the level-(k + 1) error E (k+1) = Li = E (k) Gj R(k+1) (S (k+1) ). Let
(k)
D (k) be such a level-k decoding map Li = D (k) (E (k−1) ), which computes the level-k
logical operator from the level-(k − 1) error E (k−1) . At the highest level, we have a
decomposition
l
Y
(0) (l) (k)
E = Li Gjk R(k) [S (k) (E (k−1) )], (B.7)
k=1

where E (k) is defined by Eq. (B.6).

The decoding is performed by maximizing the posterior probability of the logical op-
(l)
erator Li , conditioned on the union of the error syndrome S̄ (l) = ∪lk=1 S (k) up to the lth
concatenation level,

(l) 1 X
P (Li |S̄ (l) ) = P (E (0) ), (B.8)
N (l) (l−1) (1)
Gj ,Gj ,...,Gj
l l−1 1

where E (0) is given by Eq. (B.7). Using the hierarchal structure, Eq. (B.8), can be rewritten
as
(l)
P (Li |S̄ (l) ) (B.9)
X (l) (l−1) (l−1)
= P (Li |S̄ (l) , Lil−1 )P (Lil−1 |S̄ (l) )
(l−1)
Li ∈L(l−1)
l−1

135
 X h i P (L(l−1) , S̄ (l) )
(l) (l) (l−1) il−1
= δ L il =D (Lil−1 )
(l−1)
P (S̄ (l) )
Li ∈L(l−1)
l−1
(l−1)
X (l) (l) (l−1) (l−1)
P (Lil−1 , S̄ (l−1) )
= δ[Lil =D (Lil−1 )]δ[S (l) =S (l)
(Lil−1 )]
(l−1)
P (S̄ (l) )
Li ∈L(l−1)
l−1
(l) (l−1) (l−1)
X δ[Lil = D (l) (Lil−1 )]δ[S (l) = S (l) (Lil−1 )] (l−1)
= P (Lil−1 |S̄ (l−1) )
(l−1)
P (S̄ (l) |S̄ (l−1) )
Li ∈L(l−1)
l−1

(B.10)

If the physical errors occur independently for each qubit, we can factorize the posterior
probability of the (l − 1)th level into posterior probabilities for level-(l − 1) code blocks
(l−1)
Y (l−1,j)
P (Lil−1 |S̄ (l−1) ) = P (Lil−1 |S̄ (l−1,j) ), (B.11)
j

(l−1,j)
where Lil−1 is a level-(l − 1) logical operator acting on the jth code block, and S̄ (l−1,j) is
a level-(l − 1) syndrome with respect to the stabilizer operator on the jth code block. By
repeating this procedure, we can rewrite the posterior probability Eq. (B.8) as a summation
(k,j)
over logical operators defined at each concatenation level Lik ∈ L(k,j), where we have
defined the group of the level-k logical operators on the jth code block, L(k,j).
The summation can be reformulated as a marginalization problem on a factor graph,
which is a bipartite graph consisting of two types of nodes, circles, and boxes [Pou06].
The variable xc and a function fb (∂b) are assigned on each circle c and box b. Here, ∂b is
a set of circles neighboring b. Specifically, in the present case, the factor graph is a tree
graph. The marginal distribution on a tree factor graph can be computed efficiently by
using the brief-propagation method as follows. From circles to boxes, we pass a message,
Y
µc→b (xc ) = νb′ →c (xc ). (B.12)
b′ ∈δc\b

Then from boxes to circles, we pass another message

XY Y
νb→c (xc ) = fb (δb) µc′ →b (x′c ). (B.13)
δb\xc b c′ ∈δb\c

(l)
By repeating these procedures alternatively, we can obtain P (Li |S̄ (l) ) as a message
νb→c (xc ), where the µc→b (xc ) and νb→c (xc ) are updated from the bottom leaf nodes to
the top root node, while using marginal distributions. By replacing µc→b (xc ) and fb (δb)
(k−1,j) (k) (k−1) (k−1)
with P (Lik−1 |S̄ (k−1,j) ) and δ[Lik = D (k) (Lik−1 )]δ[S (k) = S (k) (Lik−1 )], we obtain the
(l)
posterior probability P (Li |S̄ (l) ) as νb→c (xc ) at the top root node. In this way, decoding
by maximizing the posterior probability can be executed efficiently for the concatenated
stabilizer codes [Pou06].

136
Index

Φ-transformation, 122 homology class, 73

Z2 chain complex, 72
Jones polynomial, 23, 24
7-qubit Steane code, 61
Jordan-Wigner transformation, 100
9-qubit Shor code, 57
Kauffman bracket, 25
Abelian anyon, 102
Kitaev’s compass model, 103
Aharonov-Jones-Landau algorithm, 23
Kitaev’s phase estimation algorithm, 22
Bloch sphere, 7 Kitaev’s toric code Hamiltonian, 101
braid group, 25 Kraus operators, 30
byproduct operator, 47, 49
logical operator, 34, 51, 58
Calderbank-Shor-Steane code, CSS code, 61 logical qubit, 58
Chernoff-Hoeffding bound, 20
magic state, 63
classical linear code, 59
Majorana fermion, 100
Clifford gate, 8
measurement-based quantum computation,
Clifford operation, 35
46
cluster states, 42
minimum-weight perfect matching, MWPM,
code distance, 58
84
completely-positive-trace-preserving (CPTP),
mutual duality relation, 94
30
correlation surface, 120 Nishimori line, 93

defect creation, annihilation, and movement, one-bit teleportation, 47

108
parity check matrix, 60
defect pair qubit, 107
path-model representation, 27
deterministic quantum computation with one
Pauli group, 32
clean qubit, DQC1, 21
Pauli operator, 7
feedforward, 49 Pauli product, 32
five-qubit code, 59 Pauli-basis measurement, 38
plaquette and star operators, 77
gate teleportation, 48, 49 positive-operator-valued measure measure-
Gottesman-Knill theorem, 39 ment, POVM measurement, 9
graph state, 41 projected-entangled-pair state, PEPS, 146
Hadamard test, 19 quantum error correction, QEC, 55

137
quantum Fourier transform, 22
quantum Gilbert-Varshamov bound, 93
qubit, 6
quotient group, 73

random-bond Ising model, RBIM, 89

Reed-Muller code, 64
Reidemeister move, 121

Shor’s factorization algorithm, 21

stabilizer formalism, 32
stabilizer generator, 33
stabilizer group, 33
stabilizer Hamiltonian, 99
stabilizer QEC code, 58
stabilizer state, 33
stabilizer subspace, 33
symmetry protected topological order, 100

Temperley-Lieb (TL) algebra, 26

three-qubit bit flip code, 56
three-qubit phase flip code, 57
topological calculus, 119
topological color code, 96
topological diagram, 119
topological order, topological ordered sys-
tem, 99
topological stabilizer code, 72
transversality, 57
trivial cycle, 73

weight, 58

138
References

[AA06] Dorit Aharonov and Itai Arad. The bqp-hardness of approximating the
jones polynomial. arXiv preprint quant-ph/0605181, 2006.

[AAEL07] Dorit Aharonov, Itai Arad, Elad Eban, and Zeph Landau. Polynomial
quantum algorithms for additive approximations of the potts model
and other points of the tutte plane. arXiv preprint quant-ph/0702008,
2007.

[ABD+ 09] Panos Aliferis, Frederico Brito, David P DiVincenzo, John Preskill,
Matthias Steffen, and Barbara M Terhal. Fault-tolerant computing
with biased-noise superconducting qubits: a case study. New Journal
of Physics, 11(1):013061, 2009.

[ABO97] Dorit Aharonov and Michael Ben-Or. Fault tolerant quantum com-
putation with constant error. In Proc. ACM STOC, pages 176–188–,
1997.

[ABO08] Dorit Aharonov and Michael Ben-Or. Fault-tolerant quantum com-

putation with constant error rate. SIAM Journal on Computing,
38(4):1207–1282, 2008.

[AC07] Panos Aliferis and Andrew W Cross. Subsystem fault tolerance with
the bacon-shor code. Physical review letters, 98(22):220502, 2007.

[ADB05] Hans Aschauer, Wolfgang Dür, and H-J Briegel. Multiparticle entan-
glement purification for two-colorable graph states. Physical Review A,
71(1):012319, 2005.

[AGP06] Panos Aliferis, Daniel Gottesman, and John Preskill. Quantum accu-
racy threshold for concatenated distance-3 codes. Quant. Inf. Comput.,
6:97, 2006.

[AGP08] Panos Aliferis, Daniel Gottesman, and John Preskill. Accuracy thresh-
old for postselected quantum computation. Quant. Inf. Comput., 8:181,
2008.

139
[AJL09] Dorit Aharonov, Vaughan Jones, and Zeph Landau. A polynomial
quantum algorithm for approximating the jones polynomial. Algorith-
mica, 55(3):395–421, 2009.

[AP09] Panos Aliferis and John Preskill. Fibonacci scheme for fault-tolerant
quantum computation. Physical Review A, 79(1):012332, 2009.

[AS06] Pablo Arrighi and Louis Salvail. Blind quantum computation. Inter-
national Journal of Quantum Information, 4(05):883–898, 2006.

[ASWP13] Abbas Al-Shimary, James R Wootton, and Jiannis K Pachos. Lifetime

of topological quantum memories in thermal environment. New Journal
of Physics, 15(2):025027, 2013.

[BAO+ 12] H Bombin, Ruben S Andrist, Masayuki Ohzeki, Helmut G Katzgraber,

and MA Martin-Delgado. Strong resilience of topological codes to de-
polarization. Physical Review X, 2(2):021004, 2012.

[Bar82] Francisco Barahona. On the computational complexity of ising spin

glass models. Journal of Physics A: Mathematical and General,
15(10):3241, 1982.

[BBC+ 93] Charles H Bennett, Gilles Brassard, Claude Crépeau, Richard Jozsa,
Asher Peres, and William K Wootters. Teleporting an unknown quan-
tum state via dual classical and einstein-podolsky-rosen channels. Phys.
Rev. Lett., 70(13):1895, 1993.

[BBC+ 95] Adriano Barenco, Charles H Bennett, Richard Cleve, David P DiVin-
cenzo, Norman Margolus, Peter Shor, Tycho Sleator, John A Smolin,
and Harald Weinfurter. Elementary gates for quantum computation.
Physical Review A, 52(5):3457, 1995.

[BBD+ 09] Sean D. Barrett, Stephen D. Bartlett, Andrew C. Doherty, David Jen-
nings, and Terry Rudolph. Transitions in the computational power of
thermal states for measurement-based quantum computation. Phys.
Rev. A, 80:062328, Dec 2009.

[BDCP12] Héctor Bombı́n, Guillaume Duclos-Cianci, and David Poulin. Universal

topological phase of two-dimensional stabilizer codes. New Journal of
Physics, 14(7):073048, 2012.

[BDSW96] Charles H Bennett, David P DiVincenzo, John A Smolin, and

William K Wootters. Mixed-state entanglement and quantum error
correction. Physical Review A, 54(5):3824, 1996.

140
[BE93] Hans-Jürgen Briegel and Berthold-Georg Englert. Quantum optical
master equations: The use of damping bases. Physical Review A,
47(4):3311, 1993.

[BFK09] Anne Broadbent, Joseph Fitzsimons, and Elham Kashefi. Universal

blind quantum computation. In Foundations of Computer Science,
2009. FOCS’09. 50th Annual IEEE Symposium on, pages 517–526.
IEEE, 2009.

[BH12] Sergey Bravyi and Jeongwan Haah. Magic-state distillation with low
overhead. Physical Review A, 86(5):052329, 2012.

[BJS11] Michael J Bremner, Richard Jozsa, and Dan J Shepherd. Classical

simulation of commuting quantum computations implies collapse of
the polynomial hierarchy. Proceedings of the Royal Society A: Mathe-
matical, Physical and Engineering Science, 467(2126):459–472, 2011.

[BK05] Sergey Bravyi and Alexei Kitaev. Universal quantum computation with
ideal clifford gates and noisy ancillas. Phy. Rev. A, 71(2):022316, 2005.

[BKM+ 14a] R Barends, J Kelly, A Megrant, A Veitia, D Sank, E Jeffrey, TC White,

J Mutus, AG Fowler, B Campbell, et al. Superconducting quan-
tum circuits at the surface code threshold for fault tolerance. Nature,
508(7497):500–503, 2014.

[BKM+ 14b] R Barends, J Kelly, A Megrant, A Veitia, D Sank, E Jeffrey, TC White,

J Mutus, AG Fowler, B Campbell, et al. Superconducting quan-
tum circuits at the surface code threshold for fault tolerance. Nature,
508(7497):500–503, 2014.

[BMD06] Hector Bombin and Miguel Angel Martin-Delgado. Topological quan-

tum distillation. Physical review letters, 97(18):180501, 2006.

[BMD07] H Bombin and MA Martin-Delgado. Topological computation without

braiding. Physical review letters, 98(16):160502, 2007.

[BMD09] H Bombin and MA Martin-Delgado. Quantum measurements and gates

by code deformation. Journal of Physics A: Mathematical and Theo-
retical, 42(9):095302, 2009.

[Bom10] Hctor Bombin. Topological order with a twist: Ising anyons from an
abelian model. Physical review letters, 105(3):030403, 2010.

[Bom11] H Bombin. Clifford gates by code deformation. New Journal of Physics,

13(4):043005, 2011.

141
[BR01] Hans J Briegel and Robert Raussendorf. Persistent entanglement in
arrays of interacting particles. Physical Review Letters, 86(5):910, 2001.

[BR05] Daniel E Browne and Terry Rudolph. Resource-efficient linear optical

quantum computation. Physical Review Letters, 95(1):010501, 2005.

[BR07] Sergey Bravyi and Robert Raussendorf. Measurement-based quantum

computation with the toric code states. Phy. Rev. A, 76(2):022304,
2007.

[BT09] Sergey Bravyi and Barbara Terhal. A no-go theorem for a two-
dimensional self-correcting quantum memory based on stabilizer codes.
New Journal of Physics, 11(4):043029, 2009.

[BV93] Ethan Bernstein and Umesh Vazirani. Quantum complexity theory. In

Proceedings of the twenty-fifth annual ACM symposium on Theory of
computing, pages 11–20. ACM, 1993.

[CDT09] Andrew W Cross, David P DiVincenzo, and Barbara M Terhal. A com-

parative code study for quantum fault-tolerance. Quant. Inf. Comput.,
9:0541, 2009.

[CGC+ 12] Jerry M Chow, Jay M Gambetta, AD Córcoles, Seth T Merkel, John A
Smolin, Chad Rigetti, S Poletto, George A Keefe, Mary B Roth-
well, JR Rozen, et al. Universal quantum gate set approaching fault-
tolerant thresholds with superconducting qubits. Physical review let-
ters, 109(6):060501, 2012.

[CGM+ 14] Jerry M Chow, Jay M Gambetta, Easwar Magesan, Srikanth J Srini-
vasan, Andrew W Cross, David W Abraham, Nicholas A Masluk,
BR Johnson, Colm A Ryan, and M Steffen. Implementing a strand
of a scalable fault-tolerant quantum computing fabric. Nat. Commun,
5:4015, 2014.

[Chi05] Andrew M Childs. Secure assisted quantum computation. Quantum

Information & Computation, 5(6):456–466, 2005.

[CMS+ 14] AD Córcoles, Easwar Magesan, Srikanth J Srinivasan, Andrew W

Cross, M Steffen, Jay M Gambetta, and Jerry M Chow. Detecting
arbitrary quantum errors via stabilizer measurements on a sublattice
of the surface code. arXiv preprint arXiv:1410.6419, 2014.

[CNHM03] I Chiorescu, Y Nakamura, CJP Ma Harmans, and JE Mooij. Co-

herent quantum dynamics of a superconducting flux qubit. Science,
299(5614):1869–1871, 2003.

142
[CR99] William Cook and Andre Rohe. Computing minimum-weight perfect
matchings. INFORMS Journal on Computing, 11(2):138–148, 1999.

[CS96] A Robert Calderbank and Peter W Shor. Good quantum error-

correcting codes exist. Physical Review A, 54(2):1098, 1996.

[DA07] David P DiVincenzo and Panos Aliferis. Effective fault-tolerant quan-

tum computation with slow measurements. Physical review letters,
98(2):020501, 2007.

[DAB03] Wolfgang Dür, Hans Aschauer, and H-J Briegel. Multiparticle en-
tanglement purification for graph states. Physical review letters,
91(10):107903, 2003.

[DCP10a] Guillaume Duclos-Cianci and David Poulin. Fast decoders for topolog-
ical quantum codes. Physical review letters, 104(5):050504, 2010.

[DCP10b] Guillaume Duclos-Cianci and David Poulin. A renormalization group

decoding algorithm for topological quantum codes. In Information The-
ory Workshop (ITW), 2010 IEEE, pages 1–5. IEEE, 2010.

[Del14] Nicolas Delfosse. Decoding color codes by projection onto surface codes.
Physical Review A, 89(1):012317, 2014.

[Deu85] David Deutsch. Quantum theory, the church-turing principle and the
universal quantum computer. Proceedings of the Royal Society of Lon-
don. A. Mathematical and Physical Sciences, 400(1818):97–117, 1985.

[DG97] Lu-Ming Duan and Guang-Can Guo. Preserving coherence in quan-

tum computation by pairing quantum bits. Physical Review Letters,
79(10):1953, 1997.

[DG98] Lu-Ming Duan and Guang-Can Guo. Prevention of dissipation with

two particles. Physical Review A, 57(4):2399, 1998.

[DHCB05] W Dür, M Hein, JI Cirac, and H-J Briegel. Standard forms of

noisy quantum operations via depolarization. Physical Review A,
72(5):052326, 2005.

[DHN06a] Christopher M Dawson, Henry L Haselgrove, and Michael A Nielsen.

Noise thresholds for optical cluster-state quantum computation. Phys-
ical Review A, 73(5):052306, 2006.

[DHN06b] Christopher M Dawson, Henry L Haselgrove, and Michael A Nielsen.

Noise thresholds for optical quantum computers. Physical review let-
ters, 96(2):020501, 2006.

143
[DiV95] David P DiVincenzo. Two-bit gates are universal for quantum compu-
tation. Physical Review A, 51(2):1015, 1995.

[DKLP02] Eric Dennis, Alexei Kitaev, Andrew Landahl, and John Preskill.
Topological quantum memory. Journal of Mathematical Physics,
43(9):4452–4505, 2002.

[DlCDVdNMD11] G De las Cuevas, W Dür, M Van den Nest, and MA Martin-Delgado.

Quantum algorithms for classical lattice models. New Journal of
Physics, 13(9):093021, 2011.

[DN05] Christopher M Dawson and Michael A Nielsen. The solovay-kitaev

algorithm. Quantum Information and Computation, 6:081–095, 2005.

[DR05] L-M Duan and R Raussendorf. Efficient quantum computation with

probabilistic quantum gates. Physical review letters, 95(8):080503,
2005.

[DS96] David P. DiVincenzo and Peter W. Shor. Fault-tolerant error correction

with efficient quantum codes. Phys. Rev. Lett., 77:3260–3263, 1996.

[Eas13] Bryan Eastin. Distilling one-qubit magic states into toffoli states. Phys-
ical Review A, 87(3):032321, 2013.

[Edm65] Jack Edmonds. Paths, trees, and flowers. Canadian Journal of math-
ematics, 17(3):449–467, 1965.

[EPR35] Albert Einstein, Boris Podolsky, and Nathan Rosen. Can quantum-
mechanical description of physical reality be considered complete?
Physical review, 47(10):777, 1935.

[Fey82] Richard P Feynman. Simulating physics with computers. International

journal of theoretical physics, 21(6):467–488, 1982.

[FKLW03] Michael Freedman, Alexei Kitaev, Michael Larsen, and Zhenghan

Wang. Topological quantum computation. Bulletin of the American
Mathematical Society, 40(1):31–38, 2003.

[FM12] Keisuke Fujii and Tomoyuki Morimae. Topologically protected

measurement-based quantum computation on the thermal state of a
nearest-neighbor two-body hamiltonian with spin-3/2 particles. Phys-
ical Review A, 85(1):010304, 2012.

[FMMC12] Austin G Fowler, Matteo Mariantoni, John M Martinis, and Andrew N

Cleland. Surface codes: Towards practical large-scale quantum com-
putation. Physical Review A, 86(3):032324, 2012.

144
[FNIK14] Keisuke Fujii, Makoto Negoro, Nobuyuki Imoto, and Masahiro Kita-
gawa. Measurement-free topological protection using dissipative feed-
back. arXiv preprint arXiv:1401.6350, 2014.

[FNOM13] Keisuke Fujii, Yoshifumi Nakata, Masayuki Ohzeki, and Mio Murao.
Measurement-based quantum computation on symmetry breaking ther-
mal states. Phys. Rev. Lett., 110:120502, Mar 2013.

[FSG09] Austin G Fowler, Ashley M Stephens, and Peter Groszkowski. High-

threshold universal quantum computation on the surface code. Physical
Review A, 80(5):052312, 2009.

[FT10] Keisuke Fujii and Yuuki Tokunaga. Fault-tolerant topological one-

way quantum computation with probabilistic two-qubit gates. Physical
review letters, 105(25):250503, 2010.

[FT12] Keisuke Fujii and Yuuki Tokunaga. Error and loss tolerances of surface
codes with general lattice structures. Physical Review A, 86(2):020303,
2012.

[FT14] Keisuke Fujii and Shuhei Tamate. Computational quantum-

classical boundary of commuting quantum circuits. arXiv preprint
arXiv:1406.6932, 2014.

[FTP+ 05] Paolo Facchi, Shuichi Tasaki, Saverio Pascazio, Hiromichi Nakazato,
Akira Tokuse, and DA Lidar. Control of decoherence: analysis
and comparison of three different strategies. Physical Review A,
71(2):022302, 2005.

[FWH12] Austin G Fowler, Adam C Whiteside, and Lloyd CL Hollenberg. To-

wards practical classical processing for the surface code. Physical review
letters, 108(18):180501, 2012.

[FY10a] Keisuke Fujii and Katsuji Yamamoto. Cluster-based architecture for

fault-tolerant quantum computation. Phys. Rev. A, 81:042324, Apr
2010.

[FY10b] Keisuke Fujii and Katsuji Yamamoto. Topological one-way quantum

computation on verified logical cluster states. Phys. Rev. A, 82:060301,
Dec 2010.

[FYKI12] Keisuke Fujii, Takashi Yamamoto, Masato Koashi, and Nobuyuki

Imoto. A distributed architecture for scalable quantum computation
with realistically noisy devices. arXiv preprint arXiv:1202.6588, 2012.

145
[GC99] Daniel Gottesman and Isaac L Chuang. Demonstrating the viability
of universal quantum computation using teleportation and single-qubit
operations. Nature, 402(6760):390–393, 1999.

[GFG12] Joydip Ghosh, Austin G Fowler, and Michael R Geller. Surface code
with decoherence: An analysis of three superconducting architectures.
Physical Review A, 86(6):062318, 2012.

[GHZ89] Daniel M Greenberger, Michael A Horne, and Anton Zeilinger. Going

beyond bells theorem. In Bells theorem, quantum theory and concep-
tions of the universe, pages 69–72. Springer, 1989.

[Gib14] Elizabeth Gibney. Quantum computer quest. Nature, 516(7529):24–26,

2014.

[Got97] D Gottesman. Stabilizer codes and quantum error correction. PhD

thesis, California Institute of Technology, 1997.

[Got98a] Daniel Gottesman. The heisenberg representation of quantum comput-

ers. arXiv preprint quant-ph/9807006, 1998.

[Got98b] Daniel Gottesman. The heisenberg representation of quantum comput-

ers. arXiv preprint quant-ph/9807006, 1998.

[Gro96] Lov K Grover. A fast quantum mechanical algorithm for database

search. In Proceedings of the twenty-eighth annual ACM symposium on
Theory of computing, pages 212–219. ACM, 1996.

[GW09] Zheng-Cheng Gu and Xiao-Gang Wen. Tensor-entanglement-filtering

renormalization approach and symmetry-protected topological order.
Physical Review B, 80(15):155131, 2009.

[Har04] J. W. Harrington. Analysis of quantum error-correcting codes: sym-

plectic lattice codes and toric codes. PhD thesis, California Institute of
Technology, 2004.

[HCC09] Alioscia Hamma, Claudio Castelnovo, and Claudio Chamon. Toric-

boson model: Toward a topological quantum memory at finite temper-
ature. Physical Review B, 79(24):245122, 2009.

[HCEK14] M Herold, ET Campbell, J Eisert, and MJ Kastoryano. Cellular-

automaton decoders for topological quantum memories. arXiv preprint
arXiv:1406.2338, 2014.

[HDE+ 06] M Hein, W Dür, J Eisert, R Raussendorf, M Van der Nest, and
HJ Briegel. Quantum computers, algorithms and chaos. In Inter-
national School of Physics Enrico Fermi, volume 162, 2006.

146
[HFDVM12] Clare Horsman, Austin G Fowler, Simon Devitt, and Rodney Van Me-
ter. Surface code quantum computing by lattice surgery. New Journal
of Physics, 14(12):123011, 2012.

[HLG11] Min-Hsiu Hsieh and François Le Gall. Np-hardness of decoding quan-

tum error-correction codes. Physical Review A, 83(5):052331, 2011.

[HPP01] Andreas Honecker, Marco Picco, and Pierre Pujol. Universality class
of the nishimori point in the 2d±j random-bond ising model. Physical
review letters, 87(4):047201, 2001.

[HTPV08] Martin Hasenbusch, Francesco Parisen Toldin, Andrea Pelissetto, and

Ettore Vicari. Universal dependence on disorder of two-dimensional
randomly diluted and random-bond±j ising models. Physical Review
E, 78(1):011110, 2008.

[Huk99] Koji Hukushima. Domain-wall free energy of spin-glass models:

Numerical method and boundary conditions. Physical Review E,
60(4):3606, 1999.

[ICBB14] S Iblisdir, M Cirio, O Boada, and GK Brennen. Low depth quantum

circuits for ising models. Annals of Physics, 340(1):205–251, 2014.

[IP13] Pavithran Iyer and David Poulin. Hardness of decoding quantum sta-
bilizer codes. arXiv preprint arXiv:1310.3235, 2013.

[Jon13a] Cody Jones. Low-overhead constructions for the fault-tolerant toffoli

gate. Physical Review A, 87(2):022328, 2013.

[Jon13b] Cody Jones. Multilevel distillation of magic states for quantum com-
puting. Phys. Rev. A, 87:042305, Apr 2013.

[JVMF+ 12] N Cody Jones, Rodney Van Meter, Austin G Fowler, Peter L McMa-
hon, Jungsang Kim, Thaddeus D Ladd, and Yoshihisa Yamamoto.
Layered architecture for quantum computing. Physical Review X,
2(3):031007, 2012.

[JW28] P Jordan and Eugene P Wigner. About the pauli exclusion principle.
Zeitschrift fur Physik, 47:631, 1928.

[JW09] Stephen P Jordan and Pawel Wocjan. Estimating jones and homfly
polynomials with one clean qubit. Quantum Information & Computa-
tion, 9:264, 2009.

[Kas61] Pieter W Kasteleyn. The statistics of dimers on a lattice: I. the number

of dimer arrangements on a quadratic lattice. Physica, 27(12):1209–
1225, 1961.

147
[Kat95] Tosio Kato. Perturbation theory for linear operators, volume 132.
springer, 1995.

[Kay14] Alastair Kay. Implications of ignorance for quantum-error-correction

thresholds. Physical Review A, 89(3):032328, 2014.

[KBF+ 14] J. Kelly, R. Barends, A. G. Fowler, A. Megrant, E. Jeffrey, T. C. White,

et al. State preservation by repetitive error detection in a supercon-
ducting quantum circuit. arXiv preprint arXiv:1408.0959, 2014.

[KCS14] Eliot Kapit, John T Chalker, and Steven H Simon. Passive correction
of quantum logical errors in a driven, dissipative system: a blueprint
for an analog quantum code fabric. arXiv preprint arXiv:1408.0959,
2014.

[Kit95a] A Yu Kitaev. Quantum measurements and the abelian stabilizer prob-

lem. arXiv preprint quant-ph/9511026, 1995.

[Kit95b] A Yu Kitaev. Quantum measurements and the abelian stabilizer prob-

lem. arXiv preprint quant-ph/9511026, 1995.

[Kit97] A Yu Kitaev. Quantum computations: algorithms and error correction.

Russian Mathematical Surveys, 52(6):1191–1249, 1997.

[Kit01] A Yu Kitaev. Unpaired majorana fermions in quantum wires. Physics-

Uspekhi, 44(10S):131, 2001.

[Kit03] A Yu Kitaev. Fault-tolerant quantum computation by anyons. Annals

of Physics, 303(1):2–30, 2003.

[Kit06] Alexei Kitaev. Anyons in an exactly solved model and beyond. Annals
of Physics, 321(1):2–111, 2006.

[KL98] E. Knill and R. Laflamme. Power of one bit of quantum information.

Phys. Rev. Lett., 81:5672–5675, Dec 1998.

[KL12] Kao-Yueh Kuo and Chung-Chin Lu. On the hardness of decoding quan-
tum stabilizer codes under the depolarizing channel. In Information
Theory and its Applications (ISITA), 2012 International Symposium
on, pages 208–211. IEEE, 2012.

[KLZ98a] Emanuel Knill, Raymond Laflamme, and Wojciech H Zurek. Resilient

quantum computation. Science, 279(5349):342–345, 1998.

[KLZ98b] Emanuel Knill, Raymond Laflamme, and Wojciech H Zurek. Resilient

quantum computation: error models and thresholds. Proceedings of
the Royal Society of London. Series A: Mathematical, Physical and
Engineering Sciences, 454(1969):365–384, 1998.

148
[Kni05] Emanuel Knill. Quantum computing with realistically noisy devices.
Nature, 434(7029):39–44, 2005.

[Kol09] Vladimir Kolmogorov. Blossom v: a new implementation of a min-

imum cost perfect matching algorithm. Mathematical Programming
Computation, 1(1):43–67, 2009.

[KR97] Naoki Kawashima and Heiko Rieger. Finite-size scaling analysis of

exact ground states for±j spin glass models in two dimensions. EPL
(Europhysics Letters), 39(1):85, 1997.

[KW41] Hendrik A Kramers and Gregory H Wannier. Statistics of the two-

dimensional ferromagnet. part i. Physical Review, 60(3):252, 1941.

[LB12] Ying Li and Simon C Benjamin. High threshold distributed quan-

tum computing with three-qubit nodes. New Journal of Physics,
14(9):093008, 2012.

[LBK+ 11] Ying Li, Daniel E Browne, Leong Chuan Kwek, Robert Raussendorf,
and Tzu-Chieh Wei. Thermal states as universal resources for quan-
tum computation with always-on interactions. Physical review letters,
107(6):060501, 2011.

[LBSB10] Ying Li, Sean D Barrett, Thomas M Stace, and Simon C Ben-
jamin. Fault tolerant quantum computation with nondeterministic
gates. Physical review letters, 105(25):250502, 2010.

[LCW98a] D. A. Lidar, I. L. Chuang, and K. B. Whaley. Decoherence-free sub-

spaces for quantum computation. Phys. Rev. Lett., 81:2594–2597, 1998.

[LCW98b] Daniel A Lidar, Isaac L Chuang, and K Birgitta Whaley. Decoherence-

free subspaces for quantum computation. Physical Review Letters,
81(12):2594, 1998.

[LCYPP15] Olivier Landon-Cardinal, Beni Yoshida, David Poulin, and John

Preskill. Perturbative instability of quantum memory based on effective
long-range interactions. Phys. Rev. A, 91:032303, Mar 2015.

[Lic97] WB Raymond Lickorish. An introduction to knot theory, volume 175.

Springer Science & Business Media, 1997.

[Llo99] Seth Lloyd. Obituary: Rolf landauer (1927-99). Nature, 400(6746):720–

720, 1999.

[LMPZ96] Raymond Laflamme, Cesar Miquel, Juan Pablo Paz, and Wojciech Hu-
bert Zurek. Perfect quantum error correcting code. Physical Review
Letters, 77(1):198, 1996.

149
[MBF10] Yuichiro Matsuzaki, Simon C Benjamin, and Joseph Fitzsimons. Prob-
abilistic growth of large entangled states with low error accumulation.
Physical review letters, 104(5):050501, 2010.

[MC02] F Merz and JT Chalker. Two-dimensional random-bond ising model,

free fermions, and the network model. Physical Review B, 65(5):054425,
2002.

[MEK12] Adam M Meier, Bryan Eastin, and Emanuel Knill. Magic-state distil-
lation with the four-qubit code. arXiv preprint arXiv:1204.4221, 2012.

[MF12] Tomoyuki Morimae and Keisuke Fujii. Blind topological measurement-

based quantum computation. Nature communications, 3:1036, 2012.

[MF13] Tomoyuki Morimae and Keisuke Fujii. Blind quantum computation

protocol in which alice only makes measurements. Phy. Rev. A,
87(5):050301, 2013.

[MFF14] Tomoyuki Morimae, Keisuke Fujii, and Joseph F. Fitzsimons. Hardness

of classically simulating the one-clean-qubit model. Phys. Rev. Lett.,
112:130502, Apr 2014.

[MFI14] Akira Matsuo, Keisuke Fujii, and Nobuyuki Imoto. Quantum algorithm
for an additive approximation of ising partition functions. Phys. Rev.
A, 90:022304, Aug 2014.

[MNFT14] Tomoyuki Morimae, Harumichi Nishimura, Keisuke Fujii, and Shuhei

Tamate. Classical simulation of dqc1 2 or dqc2 1 implies collapse of
the polynomial hierarchy. arXiv preprint arXiv:1409.6777, 2014.

[MRR+ 14] C Monroe, R Raussendorf, A Ruthven, KR Brown, P Maunz, L-M

Duan, and J Kim. Large-scale modular quantum-computer architecture
with atomic memory and photonic interconnects. Physical Review A,
89(2):022317, 2014.

[MS77] Florence Jessie MacWilliams and Neil James Alexander Sloane. The
theory of error correcting codes, volume 16. Elsevier, 1977.

[NC00] Michael A Nielsen and Isaac L Chuang. Quantum computation and

quantum information. Cambridge university press, 2000.

[Nie04] Michael A Nielsen. Optical quantum computation using cluster states.

Physical review letters, 93(4):040503, 2004.

[Nie06] M. A. Nielsen. Cluster-state quantum computation. Rep. on Math.

Phys., 57:147, 2006.

150
[Nis79] Hidetoshi Nishimori. Conjecture on the exact transition point of the
random ising ferromagnet. Journal of Physics C: Solid State Physics,
12(23):L905, 1979.

[Nis81] Hidetoshi Nishimori. Internal energy, specific heat and correlation func-
tion of the bond-random ising model. Progress of Theoretical Physics,
66(4):1169–1181, 1981.

[Nis01] Hidetoshi Nishimori. Statistical physics of spin glasses and information

processing, volume 187. Oxford University Press Oxford, 2001.

[NLB13] Naomi H Nickerson, Ying Li, and Simon C Benjamin. Topological

quantum computing with a very noisy network and local error rates
approaching one percent. Nature communications, 4:1756, 2013.

[NN02] Hidetoshi Nishimori and Koji Nemoto. Duality and multicritical point
of two-dimensional spin glasses. Journal of the Physical Society of
Japan, 71(4):1198–1199, 2002.

[NPT99] Yu Nakamura, Yu A Pashkin, and JS Tsai. Coherent control of

macroscopic quantum states in a single-cooper-pair box. Nature,
398(6730):786–788, 1999.

[OAIM04] Takuya Ohno, Gaku Arakawa, Ikuo Ichinose, and Tetsuo Matsui. Phase
structure of the random-plaquette z2 gauge model: accuracy threshold
for a toric quantum memory. Nuclear Physics B, 697(3):462–480, 2004.

[OF12] Masayuki Ohzeki and Keisuke Fujii. Duality analysis on random planar
lattices. Physical Review E, 86(5):051121, 2012.

[Ohz09a] Masayuki Ohzeki. Accuracy thresholds of topological color codes

on the hexagonal and square-octagonal lattices. Physical Review E,
80(1):011141, 2009.

[Ohz09b] Masayuki Ohzeki. Locations of multicritical points for spin glasses on

regular lattices. Physical Review E, 79(2):021129, 2009.

[Ohz12] Masayuki Ohzeki. Error threshold estimates for surface code with loss
of qubits. Physical Review A, 85(6):060301, 2012.

[Ohz13] Masayuki Ohzeki. Spin glass: A bridge between quantum comput-

ing and statistical mechanics. In Mikio Nakahara and Shu Tanaka,
editors, Lectures on Quantum Computing, Thermodynamics and Sta-
tistical Physics, volume 8. World Scientific, 2013.

[ONR+ 14] Joe O’Gorman, Naomi H Nickerson, Philip Ross, John JL Morton, and
Simon C Benjamin. A silicon-based surface code quantum computer.
arXiv preprint arXiv:1406.5149, 2014.

151
[Pac12] Jiannis K Pachos. Introduction to topological quantum computation.
Cambridge University Press, 2012.

[PB02] Francesco Petruccione and Heinz-Peter Breuer. The theory of open

quantum systems. Oxford Univ. Press, 2002.

[PBKLO04] David Poulin, Robin Blume-Kohout, Raymond Laflamme, and Harold

Ollivier. Exponential speedup with a single bit of quantum informa-
tion: Measuring the average fidelity decay. Physical review letters,
92(17):177906, 2004.

[PHWL13] Fabio L Pedrocchi, Adrian Hutter, James R Wootton, and Daniel Loss.
Enhanced thermal stability of the toric code through coupling to a
bosonic bath. Physical Review A, 88(6):062313, 2013.

[Pou06] David Poulin. Optimal and efficient decoding of concatenated quantum

block codes. Physical Review A, 74(5):052333, 2006.

[PTBO10] Frank Pollmann, Ari M Turner, Erez Berg, and Masaki Oshikawa. En-
tanglement spectrum of a topological phase in one dimension. Physical
Review B, 81(6):064439, 2010.

[PVK97] MB Plenio, V Vedral, and PL Knight. Conditional generation of er-

ror syndromes in fault-tolerant error correction. Physical Review A,
55(6):4593, 1997.

[Rau03] R Raussendorf. Measurement-based quantum computation with cluster

states. PhD thesis, Ludwig-Maximillians Universität München, 2003.

[RB01] Robert Raussendorf and Hans J Briegel. A one-way quantum computer.

Phys. Rev. Lett., 86(22):5188–5191, 2001.

[RBB03] Robert Raussendorf, Daniel E Browne, and Hans J Briegel.

Measurement-based quantum computation on cluster states. Physical
review A, 68(2):022312, 2003.

[RBH05] Robert Raussendorf, Sergey Bravyi, and Jim Harrington. Long-range

quantum entanglement in noisy cluster states. Phys. Rev. A, 71:062313,
Jun 2005.

[RDQDS99] FDA Aarao Reis, SLA De Queiroz, and Raimundo R Dos Santos. Uni-
versality, frustration, and conformal invariance in two-dimensional ran-
dom ising magnets. Physical Review B, 60(9):6740, 1999.

[Rei05] Ben W Reichardt. Quantum universality from magic states distillation

applied to css codes. Quantum Information Processing, 4(3):251–264,
2005.

152
[RH07] Robert Raussendorf and Jim Harrington. Fault-tolerant quantum com-
putation with high threshold in two dimensions. Phys. Rev. Lett.,
98:190504, May 2007.

[RHG06] Robert Raussendorf, Jim Harrington, and Kovid Goyal. A fault-

tolerant one-way quantum computer. Annals of physics, 321(9):2242–
2270, 2006.

[RHG07] Robert Raussendorf, Jim Harrington, and Kovid Goyal. Topological

fault-tolerance in cluster state quantum computation. New Journal of
Physics, 9(6):199, 2007.

[RWH+ 12] Beat Röthlisberger, James R Wootton, Robert M Heath, Jiannis K

Pachos, and Daniel Loss. Incoherent dynamics in the toric code subject
to disorder. Physical Review A, 85(2):022313, 2012.

[RZBB94] Michael Reck, Anton Zeilinger, Herbert J Bernstein, and Philip

Bertani. Experimental realization of any discrete unitary operator.
Physical Review Letters, 73(1):58, 1994.

[SB10] Thomas M Stace and Sean D Barrett. Error correction and degeneracy
in surface codes suffering loss. Physical Review A, 81(2):022317, 2010.

[SBD09] Thomas M Stace, Sean D Barrett, and Andrew C Doherty. Thresholds

for topological codes in the presence of loss. Physical review letters,
102(20):200501, 2009.

[Sho94] P. W. Shor. Algorithms for quantum computation: discrete logarithms

and factoring. In Proceedings of the 35th Annual Symposium on Foun-
dations of Computer Science, SFCS ’94, pages 124–134, Washington,
DC, USA, 1994. IEEE Computer Society.

[Sho95] Peter W. Shor. Scheme for reducing decoherence in quantum computer

memory. Phys. Rev. A, 52:R2493–R2496, 1995.

[Sho97] Peter W Shor. Polynomial-time algorithms for prime factorization and

discrete logarithms on a quantum computer. SIAM journal on com-
puting, 26(5):1484–1509, 1997.

[SJ08] Peter W Shor and Stephen P Jordan. Estimating jones polynomials

is a complete problem for one clean qubit. Quantum Information &
Computation, 8(8):681–714, 2008.

[STD07] Krysta M. Svore, Barbara M. Terhal, and David P. DiVincenzo.

Noise threshold for a fault-tolerant two-dimensional lattice architec-
ture. Quant. Inf. Comp., 7:297–318, 2007.

153
[Ste96] Andrew M Steane. Error correcting codes in quantum theory. Physical
Review Letters, 77(5):793, 1996.

[Ste97] Andrew M Steane. Active stabilization, quantum computation, and

quantum state synthesis. Physical Review Letters, 78(11):2252, 1997.

[Ste99] Andrew M Steane. Efficient fault-tolerant quantum computing. Nature,

399(6732):124–126, 1999.

[Ste03] Andrew M Steane. Overhead and noise threshold of fault-tolerant quan-

tum error correction. Physical Review A, 68(4):042322, 2003.

[TD04] B. Terhal and D DiVincenzo. Adaptive quantum computation, constant

depth quantum circuits and arthur-merlin games. Quantum Informa-
tion & Computation, 4(2):134–145, 2004.

[TL71] Harold NV Temperley and Elliott H Lieb. Relations between

the’percolation’and’colouring’problem and other graph-theoretical
problems associated with regular planar lattices: some exact results
for the’percolation’problem. Proceedings of the Royal Society of Lon-
don. A. Mathematical and Physical Sciences, 322(1549):251–280, 1971.

[VC04] Frank Verstraete and J Ignacio Cirac. Valence-bond states for quantum
computation. Phy. Rev. A, 70(6):060302, 2004.

[VdNDDM04] Maarten Van den Nest, Jeroen Dehaene, and Bart De Moor. Graphi-
cal description of the action of local clifford transformations on graph
states. Physical Review A, 69(2):022316, 2004.

[VGW96] Lev Vaidman, Lior Goldenberg, and Stephen Wiesner. Error prevention
scheme with four particles. Physical Review A, 54(3):R1745, 1996.

[VKL99a] Lorenza Viola, Emanuel Knill, and Seth Lloyd. Dynamical decoupling
of open quantum systems. Physical Review Letters, 82(12):2417, 1999.

[VKL99b] Lorenza Viola, Emanuel Knill, and Seth Lloyd. Dynamical decoupling
of open quantum systems. Phys. Rev. Lett., 82:2417–2421, 1999.

[VL98] Lorenza Viola and Seth Lloyd. Dynamical suppression of decoherence

in two-state quantum systems. Physical Review A, 58(4):2733, 1998.

[VMLFY10] Rodney Van Meter, Thaddeus D Ladd, Austin G Fowler, and Yoshi-
hisa Yamamoto. Distributed quantum computation architecture using
semiconductor nanophotonics. International Journal of Quantum In-
formation, 8(01n02):295–323, 2010.

154
[WFH11] David S Wang, Austin G Fowler, and Lloyd CL Hollenberg. Surface
code quantum computing with error rates over 1%. Physical Review A,
83(2):020302, 2011.

[WHP03] Chenyang Wang, Jim Harrington, and John Preskill. Confinement-

higgs transition in a disordered gauge theory and the accuracy thresh-
old for quantum memory. Annals of Physics, 303(1):31–58, 2003.

[Wit89] Edward Witten. Quantum field theory and the jones polynomial. Com-
munications in Mathematical Physics, 121(3):351–399, 1989.

[WSB+ 04] Andreas Wallraff, David I Schuster, Alexandre Blais, L Frunzio, R-S
Huang, J Majer, S Kumar, Steven M Girvin, and Robert J Schoelkopf.
Strong coupling of a single photon to a superconducting qubit using
circuit quantum electrodynamics. Nature, 431(7005):162–167, 2004.

[WY06] Pawel Wocjan and Jon Yard. The jones polynomial: quantum algo-
rithms and applications in quantum complexity theory. arXiv preprint
quant-ph/0603069, 2006.

[Yos11a] Beni Yoshida. Classification of quantum phases and topology of logical

operators in an exactly solved model of quantum codes. Annals of
Physics, 326(1):15–95, 2011.

[Yos11b] Beni Yoshida. Feasibility of self-correcting quantum memory and ther-

mal stability of topological order. Annals of Physics, 326(10):2566–
2633, 2011.

[YPA+ 03] Tsuyoshi Yamamoto, Yu A Pashkin, Oleg Astafiev, Yasunobu Naka-

mura, and Jaw-Shen Tsai. Demonstration of conditional gate operation
using superconducting charge qubits. Nature, 425(6961):941–944, 2003.

[YR03] Nadav Yoran and Benni Reznik. Deterministic linear optics quan-
tum computation with single photon qubits. Physical review letters,
91(3):037903, 2003.

[ZLC00] Xinlan Zhou, Debbie W Leung, and Isaac L Chuang. Methodology for
quantum logic gate construction. Phy. Rev. A, 62(5):052316, 2000.

[ZR97] Paolo Zanardi and Mario Rasetti. Noiseless quantum codes. Physical
Review Letters, 79(17):3306, 1997.

[Zur84] Wojciech H Zurek. Reversibility and stability of information processing

systems. Physical Review Letters, 53(4):391, 1984.

155