Professional Documents
Culture Documents
2018년 2월
서울대학교 대학원
수리과학부
황규호
A Survey on Quantum Error Correcting
Codes
by
Gyuho Hwang
A DISSERTATION
Abstract
Quantum error-correcting codes is used to protect quantum information
against noise. In this thesis we explain the structure and error correction
process of the Calderbank-Shor-Stean codes(CSS) codes and the stabilizer
codes. Also we describe CSS codes in the context of stabilizer formalism.
Furthermore we will discuss the logical basis states and the encoding circuit
of stabilizer codes.
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
1 Introduction 1
2 Preliminaries 3
2.1 Quantum mechanics . . . . . . . . . . . . . . . . . . . . . . 3
2.1.1 The postulates of quantum Mechanics . . . . . . . . 3
2.1.2 Density operators . . . . . . . . . . . . . . . . . . . . 5
2.1.3 Quantum channel . . . . . . . . . . . . . . . . . . . 6
4 Calderbank-Shor-Stean codes 13
4.1 Classical linear codes . . . . . . . . . . . . . . . . . . . . . . 13
4.2 CSS codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Error correction process of CSS codes . . . . . . . . . . . . 17
5 Stabilizer codes 21
5.1 Stabilizer codes . . . . . . . . . . . . . . . . . . . . . . . . . 21
ii
CONTENTS iii
The bibliography 41
국문초록 43
Chapter 1
Introduction
1
CHAPTER 1. INTRODUCTION
correcting code.
Chapter 2 provides a overview of basic materials in quantum mechan-
ics. Chapter 3 introduces the concept of quantum error correcting codes,
and describes the quantum error correction condition. Chapter 4 is about
the Calderbank-Shor-Stean codes(or CSS codes, after the initials of the in-
ventors). Calderbank and Shor [1], and Steane [9] developed CSS codes by
using classical error correcting codes. We will review the theory of classi-
cal linear codes, and explain the structure and error correction process of
CSS codes. Finally, Chapter 5 discusses the stabilizer codes. Gottesman [2]
introduced the stabilizer formalism. It is a large class of quantum codes in-
cluding CSS codes. After explaining the structure and the error correction
properties of the stabilizer code, we will describe CSS codes in terms of
stabilizer codes. Also, we will discuss the logical basis states and encoding
circuit for stabilizer codes.
This thesis is primarily based on [4] and [2].
2
Chapter 2
Preliminaries
3
CHAPTER 2. PRELIMINARIES
holds for some unitary operator U (t1 , t2 ) which depends only on the system
and the times t1 and t2 . This is derived from the Schrödinger equation.
See the subsection 2.2.2 of [4] for details.
A quantum measurements is described by a family {Mm } of operators
acting on the state space and satisfying the completeness equation:
X
∗
Mm Mm = I.
m
The operators Mm are called the measurement operators and the index m
refers to the measurement outcomes. If we have a state vector |ψi then the
result m occur with the probability
∗
p(m) = hψ|Mm Mm |ψi.
The completeness equation shows the fact that the sum of probabilities is
one:
X X X
∗ ∗
p(m) = hψ|Mm Mm |ψi = hψ| Mm Mm |ψi = hψ|ψi = 1.
m m m
Given that outcome is m occurred, the state after the measurement is
Mm |ψi Mm |ψi
p =p .
∗
hψ|Mm Mm |ψi p(m)
There is a special class of general measurements. A projective measure-
ment is described by a Hermition operator, called an observable. Suppose
that
X
M= λm Pm
m
4
CHAPTER 2. PRELIMINARIES
5
CHAPTER 2. PRELIMINARIES
ρ(t2 ) = U ρ(t1 )U ∗
∗
p(m) = tr(Mm Mm ρ),
U (ρ ⊗ ρenv )U ∗
6
CHAPTER 2. PRELIMINARIES
where U is the unitary operator associated with the evolution. If the princi-
pal system does not interact with the environment after the transformation
U , then we can obtain the state of the principal system
Φ : B(H1 ) → B(H2 )
Theorem 2.1.1 (Corollary 2.27 in [10], section 8.2 of [4]). For a linear map
Φ on a finite dimensional Hilbert space H, the followings are equivalent.
1. Φ is a quantum channel.
and
X
Φ(ρ) = Ei ρEi∗ (2.1)
i
7
CHAPTER 2. PRELIMINARIES
3. There exists the system(called environment) with the state space Henv
and the pure state |e0 i in Henv such that
In the theorem 2.1.1 the operators {Ei } and the equation 2.1 are called
operation elements and operator sum representation of the quantum chan-
nel Φ, respectively.
8
Chapter 3
In this chapter we will discuss the basic theory of the quantum error
correcting codes.
9
CHAPTER 3. QUANTUM ERROR CORRECTING CODES
n qubit space
n n
k qubit Unitary [n, k] Φ : B(C2 ) → B(C2 )
space encoding code C
10
CHAPTER 3. QUANTUM ERROR CORRECTING CODES
ρ 7→ (1 − p)ρ + pXρX ∗
!
0 1
where p ∈ [0, 1] is fixed and X is the Pauli sigma x operator .
1 0
√ √
Observe that the errors for the bit filp channel are 1 − pI and pX. Since
this channel takes the state |ψi to the state X|ψi with the probability p
and leaves the state unchanged with the probability (1 − p), the definition
tells us the type of error, including the probability of occurrence.
11
CHAPTER 3. QUANTUM ERROR CORRECTING CODES
Therefore if {Ei }ni=1 or {Fi }ni=1 is a correctable set of errors for some quan-
tum code C, then so is the other one automatically.
More generally, the following theorem holds.
12
Chapter 4
Calderbank-Shor-Stean
codes
Definition 4.1.1 (Linear code). Suppose that G ∈ M(n,k) (F2 ) has full rank.
Then the subspace
C := {Gx|x ∈ Fk2 }
13
CHAPTER 4. CALDERBANK-SHOR-STEAN CODES
of Fn2 is called the [n, k] linear code with the generator matrix G.
C = ker H.
We say that H is the parity check matrix of the linear code C. A linear code
is completly determined by the parity check matrix as well as its generator
matrix.
A [n, k] linear code C is able to correct an error e ∈ Fn2 if and only if we
can detect the error when the error e occur on any codeword in C. Suppose
that we have an corrupted codeword y 0 , that is, y 0 = y + e for some y ∈ C.
If we know what is e then we can recover original codeword y by adding
the error e on y 0 . Conversely if we get the original codeword y by some way,
then the error can be calculated easily : e = y 0 + y. The error correction
capability of a linear code is closely related to the ’distance’ of the code.
From now on we will introduce the definition of Hamming distance and a
useful lemma to calculate Hamming distance for a linear code.
d(x, y) := |{i : xi 6= yi }|
wt(x) := d(x, 0)
14
CHAPTER 4. CALDERBANK-SHOR-STEAN CODES
Lemma 4.1.1. Suppose that H is the parity check matrix for an [n, k]
linear code C. Then the followings are equivalent.
But the combination of the equations 4.1 and 4.2 is equivalent to the con-
dition (ii) clearly.
The next theorem is about the error correction condition for linear
codes.
Proof. (⇒) Suppose that d ≤ 2t. Then there are y1 , y2 ∈ C such that
d(y1 , y2 ) ≤ 2t, i.e, wt(y1 + y2 ) ≤ 2t. So we can choose e1 , e2 ∈ Fn2 such that
y1 +y2 = e1 +e2 and wt(ej ) ≤ t for j = 1, 2. Thus if the corrupted codeword
was y 0 := y1 + e1 = y2 + e2 we cannot detect that what error occur.
(⇐) Suppose that d ≥ 2t + 1. First we claim that if e1 and e2 are distinct
15
CHAPTER 4. CALDERBANK-SHOR-STEAN CODES
elements of Fn2 whose weight are both less than t, then He1 6= He2 . If
e1 6= e2 and wt(ej ) ≤ t for j = 1, 2, then 0 < wt(e1 − e2 ) ≤ 2t < d.
So H(e1 − e2 ) is a sum of less than d columns of H. By the lemma 4.1.1
H(e1 − e2 ) 6= 0 and so He1 6= He2 . Now assume that the original codeword
y is corrupted by an error e ∈ Fn2 such that wt(e) ≤ t. Then we cen detect
the error e by computing as
Hy 0 = H(y + e) = Hy + He = 0 + He = He.
y = y0 + e
Definition 4.1.3. Let C be a [n, k] linear code with the generator matrix
G and the parity check matrix H. Then the linear code with the generator
matrix H T and the parity check matrix GT is called the dual of C, and
denoted by C ⊥ .
16
CHAPTER 4. CALDERBANK-SHOR-STEAN CODES
where
1 X
|x + C2 i = |x + yi (4.5)
|C2 |
y∈C2
17
CHAPTER 4. CALDERBANK-SHOR-STEAN CODES
18
CHAPTER 4. CALDERBANK-SHOR-STEAN CODES
to the corrupted state where hij is the (i, j) entry of H1 , the parity check
matrix of C1 . Here CNOT(i,j) is the controlled-NOT gate which perform a
bit flip on j th qubit if the i th qubit is set to |1i in terms of the computa-
tional basis. Then we get the state
1 X
(−1)(x+y)·e2 |x + y + e1 i|H1 (x + y + e1 )i. (4.8)
|C2 |
y∈C2
19
CHAPTER 4. CALDERBANK-SHOR-STEAN CODES
Putting z 0 := z + e2 , we have
1 X X
p (−1)(x+y)·(e2 +z) |zi (4.13)
|C2 |2n z 0 ∈Fn y∈C2
2
X X
1 (x+y)·(e2 +z)
X X
(x+y)·(e2 +z)
=p (−1) |zi + (−1) |zi
|C2 |2n
z 0 ∈C2⊥ y∈C2 / 2⊥
z 0 ∈C y∈C2
X X (4.14)
1 x·z 0 y·z 0
X X
x·z 0 y·z 0
=p (−1) (−1) |zi + (−1) (−1) |zi
|C2 |2n
z 0 ∈C2⊥ y∈C2 / 2⊥
z 0 ∈C y∈C2
(4.15)
By the lemma 4.3.1, this state written as
1 X 0
p (−1)x·z |z 0 + e2 i. (4.16)
2n /|C2 |
z 0 ∈C2⊥
Repeat the same process from the equation 4.8 to the equation 4.11 using
the parity check matrix H2 for C2 to correct e2 . Then we obtain the state
1 X 0
p (−1)x·z |z 0 i. (4.17)
n
2 /|C2 | 0 ⊥
z ∈C2
20
Chapter 5
Stabilizer codes
A|ψi = |ψi.
Gn = {ασ1 ⊗ · · · ⊗ σn : α = ±1, or ± i, σj = I, X, Y, or Z}
Then we say that S is the stabilizer of the space VS . Now let’s see the
necessary conditions for VS to be non trivial.
21
CHAPTER 5. STABILIZER CODES
Proof. Suppose that VS contains an non-zero vecror |ψi. Recall that any
pair of elements in Gn is either commute or anti-commute. If there are
elements M , N of S such that M N = −M N , then
Also, if −I ∈ S then
|ψi = −I|ψi = −|ψi.
But the eqaution |ψi = −|ψi contradicts that |ψi =
6 0.
In fact, two conditions stated in the proposition 5.1.1 are also sufficient
conditions of S for VS to be non trivial. We will show that later.
We use the useful notations for characterization of the stabilizer.
22
CHAPTER 5. STABILIZER CODES
Proof. Write r(g1 ) = [a1 · · · an |an+1 · · · a2n ] and r(g2 ) = [b1 · · · bn |bn+1 · · · b2n ].
Then j-th components of g1 and g2 commute if and only if (aj bn+j +
an+j bj ) = 0. Thus we have
Pn
g1 g2 = (−1) j=1 (aj bn+j +an+j bj ) g2 g1 (5.2)
Since r(g1 )Λr(g2 ) = nj=1 (aj bn+j + an+j bj ), the equation 5.2 complete the
P
proof.
Proof. Note that gi2 = I for i = 1, · · · , l since S does not contain −I. Also
r(g) + r(g 0 ) = r(gg 0 ). Suppose that r(g1 ), · · · , r(gl ) are linearly indepen-
dent. Then there are not all zero a1 , · · · al ∈ F2 , say aj = 1, such that
Pl Q ai
i=1 ai r(gi ) = 0. So i gi ∈ S is the identity up to multiplicative factor.
But by the hypothesis that −I ∈ S implies the multiplicative factor is 1.
Thus i giai ∈ S = I and we have
Q
Y a
gj = gj−1 = gi i (5.4)
i6=j
23
CHAPTER 5. STABILIZER CODES
GΛx = ei
(ii) −I ∈
/S
24
CHAPTER 5. STABILIZER CODES
Firstly we will show that P is the projector onto VS . Recall that every
elements of Gn is unitary. Also note that every gj has multiplicative factor
either +1 or −1 since S does not contain −I. So gj ’s are Hermition. Direct
I+g I+g
calculation shows that ( 2 j )2 = 2 j . Now using commutativity of {gj } we
get P ∗ = P and P 2 = P . Let P = i |vi ihvi | be the spectral decomposition.
P
Now we claim that PSx ’s are orthogonal for distinct x’s. If x = (x1 , · · · , xn−k )
and x0 = (x01 , · · · , x0n−k ) are distinct, xj 6= x0j for some j. Without loss of
0
generality, assume xj = 0 and x0j = 1. Then PSx (resp. PSx ) contains the pro-
I+g I−g
jector 2 j (resp. 2 j ) onto (+1)-eigenspace of gj (resp. (−1)-eigenspace
25
CHAPTER 5. STABILIZER CODES
Combining the equation 5.6 and the fact that distinct PSx ’s are orthogonal,
we have
X
2n = dim I = dim PSx = 2n−k dim VS .
x∈Fn−k
2
Ei∗ Ej ∈
/ N (S) − S for all i, j (5.7)
26
CHAPTER 5. STABILIZER CODES
P Ei∗ Ej P = 0 (5.9)
Now put
1 if E ∗ E ∈ S
i j
αij = (5.10)
0 if E ∗ E ∈ G − S
i j n
Error correction
Let C(S) be a stabilizer code as in the definition 5.1.3. Suppose that
{Ej } ⊂ Gn satisfy
Ei∗ Ej ∈
/ N (S) − S for all i, j (5.11)
Suppose that an error Ej occur to a state |ψi ∈ C(S). Then we have two
cases.
27
CHAPTER 5. STABILIZER CODES
2. If distinct errors Ej and Ej 0 have the same error syndrome (β1 , · · · , βn−k ),
Hence
(Ej∗0 Ej )gl (Ej∗0 Ej )∗ = gl for l = 1, · · · , n − k (5.14)
and so
Ej∗0 Ej ∈ Z(S) = N (S). (5.15)
Ej∗0 Ej ∈ S
28
CHAPTER 5. STABILIZER CODES
|0i H • H
E|ψi gl
the measurement is
I+gl
E|ψi
p2 = E|ψi.
p(+1)
If βl = −1, then
I + gl I − gl
E|ψi = E |ψi = 0 (5.18)
2 2
I − gl I + gl
E|ψi = E |ψi = E|ψi (5.19)
2 2
So we have p(−1) = 1 and post state is E|ψi. Thus we can conclude that
+1, if the measurement outcome is + 1
βl = (5.20)
−1, if the measurement outcome is − 1
29
CHAPTER 5. STABILIZER CODES
min {weight of E}
E∈N (S)−S
(iii) If C(S) is [n, k] stabilizer code with distance d, then we say that C(S)
is an [n, k, d] stabilizer code.
30
CHAPTER 5. STABILIZER CODES
In the above equation, H[C1 ]G[C2 ] = O follows from the fact that C2 ⊂ C1 .
By the equation 5.24 and 5.25, we have
Since the dimension of CSS(C1 , C2 ) and C(S) are same as 2k1 −k2 , it suffices
to show that CSS(C1 , C2 ) ⊂ C(S). Recall that {|x + C2 i|x ∈ C1 } forms
P
an orthonormal basis for CSS(C1 , C2 ). We will show that y∈C2 |x + yi is
stabilized by gl for l = 1, · · · n − (k1 − k2 ).
31
CHAPTER 5. STABILIZER CODES
i) 1 ≤ i ≤ k2
h i
r(gi ) = i th row of H[C2⊥ ] 0 · · · 0
h i (5.26)
= (G[C2 ]ei )T 0 · · · 0
and so
X X
gi |x + yi = |x + y + G[C2 ]ei i
y∈C2 y∈C2
X (5.27)
= |x + yi
y∈C2
ii) k2 + 1 ≤ i ≤ n − (k1 − k2 )
h i
r(gi ) = 0 · · · 0 (i − k2 )th row of H[C1 ]
h i (5.28)
= 0 · · · 0 eTi−k2 H[C1 ]
and so
eT
i−k H[C1 ](x+y)
X X
gi |x + yi = (−1) 2 |x + yi
y∈C2 y∈C2
X
= (−1)0 |x + yi (5.29)
y∈C2
X
= |x + yi
y∈C2
32
CHAPTER 5. STABILIZER CODES
denote the check matrix for C(S) where G1 and G2 are (n−k)×2n matrices.
Between actions on the check matrix G and the stabilizer code C(S), there
exist following correspondence:
swapping rows ←→ relabeling generators
swapping columns on both side ←→ relabeling qubits
adding two rows ←→ multiplying generators
Using Gaussian elimination we can obtain an equivalent stabilizer code with
the check matrix in the form
r n−k−r k r n−k−r k
z}|{ z}|{ z}|{ z}|{ z}|{ z}|{
" #
I A 1 A2 B 0 C }r
(5.30)
0 0 0 D I E } n-k-r
We will define tha logical Z operators and the logical computational basis
states of this code. Put
h i
Gz = 0 0 0 AT2 0 I
and choose operators Z¯1 , · · · , Z¯k ∈ Gn with the check matrix Gz . Then
" # I A1 A2 B 0 C
G
= 0 0 0 D I E
Gz
0 0 0 AT2 0 I
33
CHAPTER 5. STABILIZER CODES
has the full rank. So g1 , · · · , gn−k , Z¯1 , · · · , Z¯k are independent by the lemma
5.1.2. Also observe that
" #" #
I A 1 A 2 B 0 C 0 I h iT
GΛGTz = 0 0 0 AT2 0 I
0 0 0 D I E I 0
" #
A2 + A2
= = 0(n−k)×k
0
(5.32)
An such state is unique up to global phase since the dimension of the space
V<g1 ,··· ,gn−k ,(−1)x1 Z¯1 ,··· ,(−1)xk Z¯k > is one by the theorem 5.1.1. Furthemore, if
(x1 , · · · , xk ) 6= (x01 , · · · , x0k ) then the states |x1 · · · xk iL and |x01 · · · x0k iL are
orthonormal since these states are eigenvectors with different eigenvalues
for some Hermition operator Z̄i . Thus {|x1 · · · xk iL : (x1 , · · · , xk ) ∈ Fk2 }
forms a orthonormal basis for C(S).
By the definition of the logocal basis states, the operators Z̄i play a role
of the Pauli Z operators:
34
CHAPTER 5. STABILIZER CODES
Logical X operators
In order to use a certain unitary operator L ∈ Gn as a nontrivial logical
operator on C(S), the following two conditions should be satisfied.
35
CHAPTER 5. STABILIZER CODES
has full rank. So g1 , · · · , gn−k , X̄1 , · · · , X̄k are independent by the lemma
5.1.2. Also, the equation
0
# E
" # I A 1 A2 B 0 C "
G 0 I I
ΛGTx = 0 0 0 D I E
Gx T T
I 0 C
0 E I C 0 0 0
(5.35)
0
" #
C +C
= = 0n×k
E+E
implies that {g1 , · · · , gn−k , X̄1 , · · · , X̄k } is a commuting set by the lemma
5.1.1. Furthermore,
" #
h i 0 I h iT
T
Gz ΛGx = 0 0 0 AT2 0 I 0 ET I CT 0 0
I 0
= Ik×k .
(5.36)
The equation 5.37 implies that if a state |ψi is stabilized by Z̄i then the
state X̄j |ψi is stabilized by (−1)δij Z̄i . Thus the operators X̄1 , · · · , X̄k ∈ Gn
play the role of encoded Pauli X operators:
36
CHAPTER 5. STABILIZER CODES
So far we have seen that ri=1 (I + gi )|0 · · · 0i is the vector in C(S) and sta-
Q
bilized by the logical Z operators. Note that gi flips only ith qubits among
the first r qubits for i = 1, · · · , r. So {g1x1 · · · grxr |0 · · · 0i : (x1 , · · · , xr ) ∈ Fr2 }
37
CHAPTER 5. STABILIZER CODES
|0i
.. ..
n−k . .
|0i
Quantum gates |c1 · · · ck iL
|c1 i
.. ..
. .
|ck i
38
CHAPTER 5. STABILIZER CODES
|0i
r .. ··· ..
. .
|0i
|0i
.. X ··· X ..
n−k−r .
f1 fk
.
|0i
|c1 i •
.. .. ..
. . .
|ck i • ∗
Qk cj
Figure 5.3: Circuit to get j=1 X̄j |0 · · · 0i
(C T )ij
The part of X̄i acting on first r qubits, rj=1 Zj
Q
, does not change
|0 · · · 0i and so we can ignore this part for our purpose. Let X fi denote
Qn−k−r (E T )ij
j=1 Xr+j . Since the state ontained by applying X fi to a state of the
form | · · · ci · · ·i controlled on (n−k+i)th qubit equal X fi Xn−k+i | · · · 0 · · ·i,
|{z}
|{z}
n−k+i n−k+i
c
we get the state kj=1 X̄j j |0 · · · 0i by the circuit in the figure 5.3. Note that
Q
c
the first r qubits of kj=1 X̄j j |0 · · · 0i are still |0i’s.
Q
c
to the state kj=1 X̄j j |0 · · · 0i. Let gej denote the operator obtained by re-
Q
39
CHAPTER 5. STABILIZER CODES
|0i H Z b11 •
ger
.. .. .. .. ..
. . . . .
|0i H Z brr •
|0i
.. X ··· X ··· ..
.
f1 fk ge1 .
|0i
ger
|c1 i •
.. .. ··· ..
. . .
|ck i •
Figure 5.4: Circuit to get |c1 · · · ck iL
40
Bibliography
[5] J. Preskill. Lecture Notes for Physics 229: Quantum Information and
Computation. California Institute of Technology, 1998.
41
BIBLIOGRAPHY
42
국문초록