PHYSICAL REVIEW A VOLUME 54, NUMBER 2 AUGUST 1996

Good quantum error-correcting codes exist

A. R. Calderbank and Peter W. Shor
AT&T Research, 600 Mountain Avenue, Murray Hill, New Jersey 07974
~Received 12 September 1995!
A quantum error-correcting code is defined to be a unitary mapping ~encoding! of k qubits ~two-state
quantum systems! into a subspace of the quantum state space of n qubits such that if any t of the qubits
undergo arbitrary decoherence, not necessarily independently, the resulting n qubits can be used to faithfully
reconstruct the original quantum state of the k encoded qubits. Quantum error-correcting codes are shown to
exist with asymptotic rate k/n5122H 2 (2t/n) where H 2 ( p) is the binary entropy function
2plog2p2(12p)log2(12p). Upper bounds on this asymptotic rate are given. @S1050-2947~96!00708-1#

PACS number~s!: 03.65.Bz, 89.70.1c

I. INTRODUCTION be perfectly recovered from the remaining n2t qubits.

Suppose that we have a coherent quantum state of k qu-
With the realization that computers that use the interfer- bits that we wish to store using a physical quantum system
ence and superposition principles of quantum mechanics which is subject to some decoherence process. For example,
might be able to solve certain problems, including prime during computation on the quantum computer proposed by
factorization, exponentially faster than classical computers Cirac and Zoller @3#, we would need to store quantum infor-
@1#, interest has been growing in the feasibility of these quan- mation in entangled electronic states of ions held in an ion
tum computers, and several methods for building quantum trap. The decoherence time of the quantum state of k en-
gates and quantum computers have been proposed @2,3#. One tangled qubits is in general 1/k of the decoherence time of
of the most cogent arguments against the feasibility of quan- one qubit ~this makes the optimistic assumption that coher-
tum computation appears to be the difficulty of eliminating ence between different qubits is as stable as coherence of a
error caused by inaccuracy and decoherence @4#. Whereas the single qubit!. Thus one might expect that the best way to
best experimental implementations of quantum gates accom-
store the state of k entangled qubits is to store them in k
plished so far have less than 90% accuracy @5#, the accuracy
physical qubits. Our results show that if we use quantum
required for factorization of numbers large enough to be dif-
error-correcting codes, it is possible to store the k qubits in
ficult on conventional computers appears to be closer to one
part in billions. We hope that the techniques investigated in n.k qubits so that the decoherence time for the encoded
this paper can eventually be extended so as to reduce this quantum state is a small constant fraction of the decoherence
quantity by several orders of magnitude. time of one qubit. These results thus show that some mea-
In the storage and transmission of digital data, errors can surable nonlocal properties of entangled systems are much
be corrected by using error-correcting codes @6#. In digital more stable under decoherence than is the entire entangled
computation, errors can be corrected by using redundancy; in system.
fact, it has been shown that fairly unreliable gates could be Physical quantum channels will be unlikely to leave
assembled to form a reliable computer @7#. It has widely been n2t qubits perfectly untouched and subject the remaining t
assumed that the quantum no-cloning theorem @8# makes er- qubits to decoherence. To analyze the behavior of our error-
ror correction impossible in quantum communication and correcting code for physical quantum channels, we must
computation because redundancy cannot be obtained by du- make some assumptions about the decoherence process. In
plicating quantum bits. This argument was shown to be in Sec. VI, we will show that our error-correction method per-
error for quantum communication in Ref. @9#, where a code forms well if the decoherence of different qubits occurs in-
was given that mapped one qubit ~two-state quantum system! dependently, i.e., if each of the qubits is coupled to a sepa-
into nine qubits so that the original qubit could be recovered rate environment. Our error-correction method will actually
perfectly even after arbitrary decoherence of any one of these work for more general channels, as it can tolerate coupled
nine qubits. This gives a quantum code on nine qubits with a decoherence behavior among small groups of qubits.
rate 91 that protects against one error. Here we show the ex- The lower bound of 122H 2 (2t/n) shown in our paper
istence of better quantum error-correcting codes, having a should be compared with the theoretical upper bounds of
higher information transmission rate and better error-
correction capacity. Specifically, we show the existence of
quantum error-correcting codes encoding k qubits into n qu- min@ 12H 2 ~ 2t/3n ! ,H 2 „@ 21 1 A~ 12t/n ! t/n…#
bits that correct t errors and have an asymptotic rate
122H 2 (2t/n) as n→`. These codes work not by duplicat-
ing the quantum state of the encoded k qubits, but by spread- for t/n, 21 , and 0 for t/n> 21 . These are obtained from
ing it out over all n qubits so that if t or fewer of these qubits bounds on the quantum information capacity of a quantum
are measured, no information about the quantum state of the channel, which we derive in Sec. VI from results of Refs.
encoded bits is revealed and, in fact, the quantum state can @10,11#. These bounds are plotted in Fig. 1.

1050-2947/96/54~2!/1098~8!/$10.00 54 1098 © 1996 The American Physical Society

54 GOOD QUANTUM ERROR-CORRECTING CODES EXIST 1099

0000000, 0001011, 0010110, 0011101,

0100111, 0101100, 0110001, 0111010,
1000101, 1001110, 1010011, 1011000,
1100010, 1101001, 1110100, 1111111. ~1!

The minimum distance is the minimum Hamming weight of

a nonzero codeword, which is 3, so this is a one-error cor-
FIG. 1. The solid line shows the asymptotic rate R of our quan- recting code. It is easily verified that the dual code consists
tum codes versus the error rate of the channel t/n. Two upper of all vectors in the Hamming code with an even weight.
bounds for this quantity are also plotted: the Levitin-Holevo upper The quantum Hilbert space Hn2 over n qubits is the com-
bound with a dashed line and the entanglement upper bound with a plex space generated by basis vectors u b 0 & , u b 1 & , . . . ,
dotted line. u b 2 n 21 & where b i is the representation of the number i in
binary. This Hilbert space has a natural representation as a
II. DEFINITIONS tensor product of n copies of H2 , with the ith copy corre-
sponding to the ith bit of the basis vectors. We refer to each
Our constructions of quantum error-correcting codes rely of these copies of H2 as a qubit.
heavily on the properties of classical error-correcting codes. We define a quantum error-correcting code Q with rate
We will thus first briefly review certain definitions and prop- k/n to be a unitary mapping of Hk2 into Hn2 . Strictly speak-
erties related to binary linear error-correcting codes. We only ing, this is actually a unitary mapping of Hk2 into a
consider vectors and codes over F2 , the field of two ele- 2 k -dimensional subspace of Hn2 ; it can alternatively be
ments, so we have 11150. A binary vector v PF2 with d viewed as a unitary mapping of Hk2 ^ Hn2k 2 into Hn2 , where
1’s is said to have Hamming weight d, denoted by the quantum state in H2 is taken to be that where all the
n2k

wt( v )5d. The Hamming distance d H ( v ,w) between two bi- qubits have quantum state u 0 & . In our model of error ana-
nary vectors v and w is wt( v 1w). The support of a vector lyzed in Sec. IV, we will assume that the decoherence pro-
v , denoted by supp( v ), is the set of coordinates of v where cess affects only t bits; that is, the decoherence is modeled
the corresponding entry is not 0, that is, supp( v )5 $ i: v i by first applying an arbitrary unitary transformation D to the
Þ0 % . Suppose that S is a set of coordinates. Then v u S de- space consisting of the tensor product Ht2 ^ Henv of any t of
notes the projection of v onto S, i.e., the vector that agrees the qubits and some arbitrary Hilbert space Henv designating
with v on the coordinates in S and is 0 on the remaining the environment, and then tracing over the environment
coordinates. For a binary vector E we use v u E to mean Henv to obtain the output of the channel, which will thus in
v u supp(E) . We also use e<E to mean that general be an ensemble of states in Hk2 . We say that a quan-
supp(e)#supp(E). tum code can correct t errors if the original state u x & PHk2
A code C of length n is a set of binary vectors of length can be recovered from the decohered encoded state DQu x &
n, called codewords. In a linear code the codewords are by applying a unitary transformation R ~independent of D)
those vectors in a subspace of Fn2 ~the n-dimensional vector to Hn2 ^ Hanc , where Hanc is a Hilbert space representing the
space over the field F2 on two elements!. The minimum dis- state of an ancilla ~i.e., a supplementary quantum system!. It
tance d5d(C) of a binary code C is the minimum distance turns out that if our quantum code will correct an arbitrary
between two distinct codewords. If C is linear then this mini- decoherence of t or fewer qubits, it will also be able to trans-
mum distance is just the minimum Hamming weight of a mit information with high fidelity for a large class of chan-
nonzero codeword. nels with physically plausible decoherence processes; this is
A linear code with length n, dimension k, and minimum discussed in Sec. VI.
distance d is called an @ n,k,d # code. For a code C with Since the error correction must work for any encoded
minimum distance d, any binary vector in Fn2 is within Ham- state Qu x & , the property of being a quantum error-correcting
ming distance t5 b (d21/2) c of at most one codeword; thus, code depends only on the subspace QHk2 of Hn2 , and not on
a code with a minimum distance d can correct t errors made the actual mapping Q. However, for ease of explanation, we
in the bits of a codeword; such a code is thus said to be a t will nonetheless define an orthogonal basis of this subspace
error-correcting code. The rate R of a linear code of length of Hn2 , which can be used to obtain an explicit mapping
n is dim(C)/n; this is the ratio of the information content of Q, and call the elements of this basis codewords.
a codeword to the information content of an arbitrary string
of length n. The dual code C' of a code C is the set of
vectors of perpendicular to all codewords, that is, C' 5 $ v III. QUANTUM CODES
PFn2 : v •c50;cPC% . From linear algebra, dim(C)
1dim(C' )5n. We will now define our quantum code. Suppose that we
In this paper, we will use the @ 7,4,3# Hamming code as an have a linear code C1 ,Fn2 . We let HC1 be the subspace of
example to illustrate our construction of quantum error- Hn2 generated by vectors u c & with cPC1 . Let M be a genera-
correcting codes. This code contains the following 16 binary tor matrix for C1 ; this means that C1 is the row space of
vectors of length 7: M , so that v M ranges over all the codewords in C1 as v
1100 A. R. CALDERBANK AND PETER W. SHOR 54

ranges over all vectors in F2

dim(C1 )
. For wPFn2 , we define a why this should be true; while in the next section, we will
quantum state u c w & by work out this calculation in detail.
To show why our codes are error correcting, we must first
give another representation of our codewords. If we perform
u c w & 52 2dim~ C1 ! /2 ( dim~ C1 !
~ 21 ! v M w u v M & . ~2! the following change of basis,
v PF
2
1
Note that if w 1 1w 2 PC'1 , then u c w 1 & 5 u c w 2 & , since v M w 1 u0&→ ~u0&1u1&) ~5!
A2
dim(C )
5 v M w 2 for all v PF2 1 . Further note that ^ c w 1 u c w 2 &
50 if w 1 1w 2 ¹C'1 . This follows since ( v (21) v M w 50 un- 1
u1&→ ~ u 0 & 2 u 1 & ),
less v M w50 for all v PF 2 1 . Thus for wPFn2 /C'1 A2
dim(C )

the vectors u c w & form a basis for the space HC1 . ~Here
F n2 /C'1 stands for the cosets of C'1 in Fn2 , which are the sets to each of the bits of our codeword u c w & we obtain the state
C'1 1w where wPFn2 ; there are 2 dim(C1 ) of these cosets and
they form the natural index set for the quantum states u s w & 52 ~ dim~ C1 ! 2n ! /2 ( '
u u1w & . ~6!
u c w & .) uPC1
Suppose now that we have another linear code C2 with
$ 0 % ,C2 ,C1 ,Fn2 . Our quantum code will be constructed us- We can see this since if u x & is any basis state in the rotated
ing codes C1 and C2 . We define the codewords of our quan- basis given by Eq. ~5!, then
tum code QC1 ,C2 as the set of u c w & for all wPC'2 , Recall that
two codewords u c w & and u c w 8 & are equal if w1w 8 PC' . The ^ x u c v & 52 2 ~ n1dimC1 ! /2 (
dim~ C1 !
~ 21 ! v M ~ w1x ! , ~7!
natural index set for the codewords is thus over C'2 /C'1 , the v PF
2
cosets of C'1 in C'2 . This code thus contains 2 dim(C1 )2dim(C2 )
orthogonal vectors. Since its length is n qubits, it has a rate and this sum is 0 unless w1xPC'1 . Letting u5w1x, we get
(dim(C1 )2dim(C2 ))/n. To construct a quantum error- Eq. ~6!. For our example quantum code,
correcting code from the Hamming code given in Eq. ~1!, we
will take C1 to be this code and C2 to be C'1 . Thus, 1
u s 0& 5 ~ u 0000000& 1 u 0011101& 1 u 0100111&
dim(C1 )54 and dim(C2 )53, so our quantum error- 2 A2
correcting code will map 42351 qubit into 7 qubits. There
are thus two codewords. The first is 1 u 0111010& 1 u 1001110& 1 u 1010011&
1 u 1101001& 1 u 1110100& ) ~8!
u c 0 & 5 41 ~ u 0000000& 1 u 0011101& 1 u 0100111& 1 u 0111010&
and
1 u 1001110& 1 u 1010011& 1 u 1101001& 1 u 1110100&
1 u 0001011& 1 u 0010110& 1 u 0101100& 1 u 0110001& 1
u s 1& 5 ~ u 0001011& 1 u 0010110& 1 u 0101100&
1 u 1000101& 1 u 1011000& 1 u 1100010& 1 u 1111111& ), 2 A2

~3! 1 u 0110001& 1 u 1000101& 1 u 1011000&

and the second is 1 u 1100010& 1 u 1111111& ). ~9!

We can now see how these codes are able to correct er-
u c 1 & 5 41 ~ u 0000000& 1 u 0011101& 1 u 0100111& 1 u 0111010&
rors. In the u c w & representation, all the codewords are super-
1 u 1001110& 1 u 1010011& 1 u 1101001& 1 u 1110100& positions of basis vectors u v & with v PC1 . Thus any t bit
errors ~those errors taking u 0 & → u 1 & and u 1 & → u 0 & ) can be
2 u 0001011& 2 u 0010110& 2 u 0101100& 2 u 0110001& corrected by performing a classical error-correction process
for the code C1 in the original basis. In the u s w & representa-
2 u 1000101& 2 u 1011000& 2 u 1100010& 2 u 1111111& ).
tion, all the codewords are superpositions of basis vectors
~4! u v & with v PC'2 . Thus any t bit errors in the rotated basis can
be corrected by performing a classical error-correction pro-
Note that in u c 1 & all the codewords of the Hamming code cess for the code C'2 in the rotated basis. However phase
with an odd weight have a negative amplitude, and all the errors in the original basis ~errors taking u 0 & → u 0 & and
codewords with an even weight have a positive amplitude. u 1 & →2 u 1 & ) are bit errors in the rotated basis and vice versa.
This is the effect of the (21) v M w term in Eq. ~2!. Thus our quantum code can correct t bit errors and t phase
We will show that if C1 and C'2 have a minimum distance errors in the original basis.
d, then the quantum code QC1 ,C2 can correct t5 b (d21/2) c The correction process we use for our quantum error-
errors. ~For our example code, C1 5C'2 has a minimum dis- correcting codes is indeed to first correct bit errors in the
tance 3, so our quantum code will correct one error.! In the u c v & basis classically and then to correct bit errors in the
remainder of this section, we will give some intuition as to u s v & basis classically. It remains to be shown that the correc-
54 GOOD QUANTUM ERROR-CORRECTING CODES EXIST 1101

tion process for the bit errors does not interfere with the IV. DECODING QUANTUM CODES
correction process for the phase errors, and that arbitrary
nonunitary errors on t or fewer quantum bits of our code will In this section we will show that errors in any t qubits of
also be corrected by this procedure. This is done through our quantum codes can be corrected by first correcting bit
calculations which are performed in Sec. IV of our paper. errors in the u c & basis, and then correcting bit errors in the
As in Ref. @9#, we correct the error by correcting the de- u s & basis. For this section and the remainder of this paper, we
coherence without disturbing the encoded information. Intu- will assume for simplicity that dim(C1 )5n2k and
itively, what we do is to measure the decoherence without dim(C2 )5k; thus, the rate of our codes will be 122k/n.
observing the encoded state; this then lets us correct the de- However, all of our results are easily extendable to quantum
coherence while leaving the encoded state unchanged. In our
codes derived from classical codes C2 ,C1 ,Fn2 of any di-
decoding procedure, we thus learn which qubits had bit er-
rors and which had phase errors, which tells us something mension.
about the decoherence process but which gives no informa- In order to prove that errors in quantum codes can be
tion about our encoded state. Linear codes are very well corrected, we first need a lemma about purely classical
suited for this application: each codeword has the same rela- codes.
tion to all the other words in the code, and this property is Lemma 1. Suppose that C is a binary linear code of length
what enables us to measure the error without learning which n. Let e, E P Fn2 , with e<E and wt(E),d(C' ). Then there
codeword it is that is in error. exists a vector v e PC such that v e u supp(E) 5e.
Recently we learned that related work has been done by Proof. The projection of C onto E has to have full rank,
Steane @12#. Steane generates his quantum code using code- because otherwise C' would contain a vector w with
words wt(w)<wt(E),d(C' ).
We now need the following lemma about the states
u s w8 & 52 2dim~ C2 ! /2 (
v PC2
u v 1w & , ~10! u c w& .
Lemma 2. Suppose that C1 has a minimum distance d. Let
e, E P Fn2 with e<E. Let P be the projection onto the
where w is chosen from C1 /C2 . This is the same as our
u s w & basis if the codes C1 and C'2 are interchanged. It should subspace of Hn2 generated by all u v & where v is in the set
also be noted that these codewords u s w8 & generate exactly the $ v PFn2 : v u E 5e % , that is, with v equal to e on supp(E). Then
same subspace of Hn2 as the codewords u c w & given by Eq.
~2!, and thus effectively give a different basis for the same
quantum code.

^ c w 1 u P u c w 2 & 52 2 ~ n2k ! (
v : v M u E 5e
~ 21 ! v M ~ w 1 1w 2 ! ~11a!

5 H ~ 21 ! e• ~ c1w 1 1w 2 ! /2wt~ E !
0 otherwise.
if 'cPC'1 such that c1w 1 1w 2 <E,
~11b!

Proof. From the definition of u c w & in Eq. ~2!, it is straight- Now, because the set $ v M : v M u E 50 % is an
forward to show Eq. ~11a!. We must now show that this is n2k2wt(E) dimensional subspace of Fk2 , the sum ~12b! is
equal to Eq. ~11b!. Since wt(e),d(C'1 ), by Lemma 1 there 0 unless v M (w 1 1w 2 )50 for all v M in this subspace. It is
is a vector v e such that v e M u E 5e. We can obtain the linear clear that if there is a cPC'1 such that w 1 1w 2 1c<E,
space $ v PFn2k 2 : v u E 5e % by taking every vector in the set then v M (w 1 1w 2 )50 if v M u E 50, and v e M (w 1 1w 2 )
$ v PFn2k : v u E 50 % and adding the vector v e . Using this sub- 5e•(c1w 1 1w 2 ). This shows the first part of Eq. ~11b!.
2
stitution in Eq. ~11a! gives We now prove the other direction. Suppose that
v M (w 1 1w 2 )50 for all v with v M u E 50. Let e j be the vec-
tor that is 1 on the jth coordinate of E and 0 on the other
coordinates. We know from Lemma 1 that there is a vector
^ c w 1 u P u c w 2 & 52 2 ~ n2k ! (
v : v M u E 50
~ 21 ! ~ v 1 v e ! M ~ w 1 1w 2 ! v j PFn2k
2 such that v j M u E 5e j . Let s j 5 v j M (w 1 1w 2 ). We
~12a! consider the vector c 8 5w 1 1w 2 1 ( wt(E)
j51 s j e j ; we will show
that this vector satisfies the conditions for the c in Eq. ~11b!.
Clearly, w 1 1w 2 1c 8 <E. We need also to show that c 8
52 2 ~ n2k ! ~ 21 ! v e M ~ w 1 1w 2 ! PC'1 . Consider any vector v PFn2k 2 . We can decompose it
i51 a i v i where v 0 M u E 50, and a i is 0 or 1.
into v 5 v 0 1 ( wt(E)
3 (
v : v M u E 50
~ 21 ! v M ~ w 1 1w 2 ! . ~12b! Note that v i M e j 5 d (i, j) where d is the Kronecker d func-
tion. Now,
1102 A. R. CALDERBANK AND PETER W. SHOR 54

S
v M c 85 v 01
wt~ E !

(
i51
DS
a i v i M w 1 1w 2 1
wt~ E !

(
j51
s je j D Note that since v M PC1 , we have now corrected our state to
some state in the Hilbert space HC1 . Recall that the vectors
u c u & with uPFn2 generated HC1 . What we do now is to con-
~13a!
sider the Hilbert space HC1 in terms of the basis elements

5 S( D S
wt~ E !

i51
a i v i M w 1 1w 2 1
wt~ E !

(
j51
s je j D u c u & for uPFn2 /C'1 instead of the basis elements u v M & . We
do this by substituting the identity
~13b!

wt~ E ! wt~ E ! u v M & 52 2 ~ n2k ! /2 ( n '

~ 21 ! v M u u c u & ~17!
5 (
i51
a i v i M ~ w 1 1w 2 ! 1 (
i51
a is i ~13c!
uPF2 /C1

in Eq. ~16!. This gives the same type of effect as the change
50, of basis in Eq. ~5! in that it produces a representation in
which it is easier to deal with phase errors. The substitution
proving the second part of Eq. ~11b!. The terms containing ~17! gives the equation
v 0 vanish in Eq. ~13a! because v 0 M (w 1 1w 2 )50 since
v 0 M u E 50, and v 0 M e i 50 since e i aE. The two terms in Eq.
~13c! cancel because of the definition of s i . R f D u c w & 52 2 ~ n2k ! (v ~ 21 ! v M w
We are now ready to prove the following theorem.
Theorem 1. If C1 and C'2 are both linear @ n,n2k,d # codes
with $ 0 % ,C2 ,C1 ,Fn2 , then the quantum code QC1 ,C2 is a 3 (u ~ 21 ! v M uu c u & e<E
( uavMu E ,e
&u A e& ,
t-error-correcting code, where t5 b (d21)/2 c .
Proof. We show how to correct any t errors. Let us start ~18!
with a codeword u c w & for wPC'2 . Now, let E be the binary
vector such that supp(E) is the set of qubits that have deco- which can be rewritten as
hered. By our hypothesis that at most t qubits decohere, we
can take wt(E)5t. We denote states of the environment by
u a i & . Since the decoherence only operates on those qubits in R f D u c w & 52 2 ~ n2k !
e<E
( u A e& (
e 8 <E
u a e 8 ,e & (u u c u &
supp(E), the most general decoherence D is a unitary pro-
cess operating on a binary vector u and the initial state of the
environment u a 0 & as follows: 3 ( ~ 21 ! v M w ~ 21 ! v M u . ~19!
v : v M u 5e
E 8

D u u,a 0 & 5 (
e<E
u u1e & u a u u E ,e & , ~14!
Now, by Lemma 2, the inner sum is 0 unless there exists
cPC'1 for which c1w1u<E. This means that u c w & can
where the states of the environment u a i & are not necessarily only decohere to u c u & if there is a cPC'1 such that
normalized. Now, we let this decoherence act on u c w & u a 0 & . wt(u1w1c)<t. We now show this means that for each
We get u c u & there is a unique u c w & with wPC'2 /C'1 which it could
have arisen from. Suppose that we have two such w’s, w 1
and w 2 with w 1 1u1c 1 5e 1 and w 2 1u1c 2 5e 2 . Then,
D u c w ,a 0 & 52 2 ~ n2k ! /2 (n2k
~ 21 ! v M w
v PF2
e 1 1e 2 5w 1 1w 2 1c 1 1c 2 PC'2 . ~20!
3 (
e<E
u v M 1e & u a v M u E ,e & . ~15!
However,

Now, we know v M PC1 , which is a code with a minimum

wt~ e 1 1e 2 ! <wt~ e 1 ! 1wt~ e 2 ! <2t. ~21!
distance d.2wt(e). Thus we can restore v M 1e to a unique
codeword v M PC1 . Intuitively, this corrects bits that have
flipped from 0 to 1 or vice versa. We can do this using a But C'2 has minimum distance d.2t; thus e 1 5e 2 , so
unitary operator R f provided we make the operation revers- w 1 1w 2 PC'1 and u c w 1 & 5 u c w 2 & .
ible; to do this we record the error e in a set of ancilla qubits This means that we can unitarily express the state in Eq.
A. After this process, the quantum state of our system is ~19! in terms of u c u & , where uPFn2 /C'1 , and then correct the
state u c u & to u c w & , since there is at most one w with
d H (w,u),t. As before, to unitarily correct u c u & to u c w & we
R f D u c w & 52 2 ~ n2k ! /2 (v ~ 21 ! v M w need to use a second ancilla A 8 to record which bits we
needed to flip to get from u to w. These flipped bits corre-
3 (
e<E
u v M & u a v M u E ,e & u A e & . ~16! spond to phase errors in the original basis. Denoting this
correction operator by Rp , we get
54 GOOD QUANTUM ERROR-CORRECTING CODES EXIST 1103

In fact, explicit constructions for classical codes that attain

Rp R f D u c w & 52 2 ~ n2k ! ( u A e & e(<E u a e 8,e & the Gilbert–Varshamov bound asymptotically are not
e<E 8 known.
Consider towers of codes as shown below
3 ( ( ~ 21 ! v M w
8 9
v : v M u 5e e <E
E $ 0 % # ^^ 1 n && #C#C' #F n2 , ~27!
3 ~ 21 ! v M ~ w1e 9 ! u c w & u A 8e 9 & where dimC5k and dimC' 5n2k. Here ^^ 1 n && denotes the
subspace of Fn2 generated by the vector 1 n containing all
ones. The codes C and C' correspond to C2 and C1 , respec-
52 2 ~ n2k ! u c w & (
e<E
u A e& (
e 8 <E
u a e 8 ,e & ~22!
tively, in Sec. III; we have now added the requirement that
C'1 5C2 . We follow MacWilliams, Sloane, and Thompson
@13#. They call a code weakly self-dual if
3 ( u A e 9& v : v M(u 5e ~ 21 ! v M e 9
9
e <E E 8 ^^ 1 n && #C#C' . ~28!

52 2wt~ E ! u c w & ( u A e& ( u A 8e 9& Given a vector v with even weight we need that the number
of k-dimensional weakly self-dual codes for which v PC' is
e<E 9
e <E
independent of v . In other words, the number of
3 (
e 8 <E
~ 21 ! e 8 •e 9 u a e 8 ,e & , k-dimensional weakly self-dual codes C contained in a given
hyperplane v' is independent of v .
We apply Theorem 2.1 of Ref. @13# ~actually a stronger
which is just u c w & tensored with a state of the ancillae and the statement established in the proof!. Let s n,k,s be the number
environment that does not depend on w. We have thus uni- of k-dimensional weakly self-dual codes C@ n,k # that contain a
tarily restored the original state and corrected t decohered given s-dimensional code C@ n,s # . Then the numbers s n,k,s are
bits. independent of the code C@ n,s # that was chosen.
We separate the case v PC@ n,k # #C'@ n,k # from the case v
V. WEAKLY SELF-DUAL CODES PC'@ n,k # \ C@ n,k # . The number of k-dimensional weakly self-
To show that a family of codes contains codes that meet dual codes C@ n,k # for which v PC@ n,k # is just s n,k,2 , the num-
the Gilbert-Varshamov bound we can often employ a very ber of codes containing the two-dimensional space
simple greedy argument; this argument appears in Ref. @6#, ^^ 1 n , v && . Next we consider pairs (C@ n,k # , v ) where C@ n,k # is a
pp. 557 and 558 ~proof of Theorem 31 of Chap. 17!. k-dimensional weakly self-dual code and v PC'@ n,k # \ C@ n,k # . In
Lemma 3. Let f i be a set of @ n i ,k i # codes such that this case C@ n,k # and v generate a (k21)-dimensional weakly
~1! k i /n i .R self-dual code C@ n,k11 # containing the two-dimensional space
~2! each nonzero vector of length n i belongs to the same ^^ 1 n , v && . The number of choices for C@ n,k11 # is s n,k11,2 .
number of codes in f i . Every code C@ n,k11 # contains 2 k k-dimensional weakly self-
Then there are codes in the family that asymptotically dual codes of which 2 k21 do not contain the two-
meet the Gilbert-Varshamov bound dimensional space ^^ 1 n , v && . Hence given a vector v with
even Hamming weight, the number of k-dimensional weakly

R>12H 2 SDd
n
as n→`. ~23!
self-dual codes contained in v' is independent of v . This is
all that is needed to apply the greedy argument used to es-
tablish the Gilbert-Varshamov bound.
Proof. Let W i be the number of codes in f i that contain a The statement that there are codes meeting the Gilbert–
particular vector v . By hypothesis, Varshamov bound is that given a ratio d/n ~where d denotes
minimum distance!, we may achieve a rate
~ 2 n i 21 ! W i 5 ~ 2 k i 21 ! u f i u . ~24!

The number of vectors with weight less than d is ~ n2k ! /n>12H 2 SD d

n
. ~29!

SD
d21
ni The redundancy k/n satisfies k/n<H 2 (d/n), so that the
(
j50 j
. ~25! quantum codes achieve a rate

If R5 ~ n22k ! /n>122H 2 SDd

n
. ~30!

SD
d21
ni
Wi (
j50 j
,W i ~ 2 n i 21 ! / ~ 2 k i 21 ! 5 u f i u ~26! This function is plotted in Fig. 1.

VI. QUANTUM CHANNELS

then there is a code in f i with minimum distance >d.
Q.E.D. In order to carry Shannon’s theory of information to the
This proof is not constructive in that it does not produce quantum regime, it is necessary to have some reasonable
codes satisfying this bound, but merely shows that they exist. definition of a noisy quantum channel. We will define a
1104 A. R. CALDERBANK AND PETER W. SHOR 54

quantum channel W by a probability distribution P on uni-

SD
t
n n2 j
tary transformations U W mapping Hsig ^ H env . For any pure E^ y u Rp R f W n u y & > (
j50 j
F ~ 12F ! j ~32!
input state u x & the channel produces as output a mixed state
by first obtaining an ensemble of states in Hsig ^ Henv by for all u y & in our quantum code C. This quantity is close to 1
applying the transformation U W to u x & with probability dis- as long as t/n.12F. Thus, if the fidelity F for each trans-
tribution P, and secondly tracing over Henv . While the ini- mitted qubit is large enough, our quantum codes guarantee
tial state of Henv could be given by an ensemble of states, it high fidelity transmission for our encoding of k qubits. Our
may also without loss of generality be taken to be a fixed quantum codes will give good results for any channel W that
pure state, as the probability distribution given by an en- transmits states u y & PHn2 well enough that W u y & has an ex-
semble of initial states may be absorbed into the probability pected projection of length at least 12 e onto the subspace of
distribution on the unitary transformation U W . The probabil- Hn2 obtained from u x & by the decoherence at most t qubits.
ity distribution could also be concentrated entirely in the Our encoding and decoding schemes then give a channel on
inital mixed state of Hen v , and a fixed unitary transform U
the Hilbert space Hk2 which has fidelity 12 e . We will next
be used, but this leads to a slightly less intuitive description
use this observation to obtain an upper bound on the channel
of the one quantum channel that we later discuss in detail.
capacity of quantum channels.
Actual quantum channels are unlikely to produce output
An upper bound for the amount of classical information
that differs from the input exactly by the decoherence of at
carried by a quantum channel is given by the Levitin–
most t qubits, and thus are unlikely to be able to transmit
Holevo theorem @10#. If the output of the channel is a signal
quantum states perfectly using this scheme. However, if the
that has density matrix r a with probitility p a , the Levitin–
average behavior of the channel results in the decoherence of
Holevo bound on the information content of this signal is
fewer than t qubits, a channel may still be able to transmit
quantum states very well. A measure of the success of trans-
mission of quantum states that has previously been success- H~ r !2 (a p a H ~ r a ! , ~33!
ful applied in quantum information theory is fidelity @14,11#.
In this paper, we define fidelity slightly differently from the
definition in Refs. @14#; we make this change as these previ- where r 5 ( a p a r a ~the density matrix for the ensemble of
ous papers discuss channels that transmit some distribution signals!, and where H( r )52Tr( r log2r) is the von Neu-
of quantum states given a priori, whereas we want our chan- mann entropy. Since quantum information can be used to
nel to faithfully transmit any pure input state. Suppose that carry classical information, the Levitin–Holevo bound can
we have a noisy channel W that transmits quantum states in be used to obtain an upper bound for the rate of a quantum
error-correcting code.
a Hilbert space Hsig . We define the fidelity of the channel to
Consider the following quantum channel discussed in Ref.
be
@11#; this channel treats each qubit independently. With
probability 12p, a qubit is unchanged, corresponding to the
min E^ x u W u x & , ~31! identity transformation ~ 10 01!. Otherwise, with each possibility
u x & PHsig
having probability p/3, the qubit is acted on by the unitary
transformation corresponding to one of the three matrices:

S D S D S D
where the expectation is taken over the output of the channel.
In other words, we are measuring the fidelity of transmission 0 1 1 0 0 1
, , or .
of the pure state transmitted with least fidelity. We could also 1 0 0 21 21 0
measure the fidelity of transmission of a typical state in
Hsig ; this average fidelity is a quantity which is closer to the That is, each of the following possibilities has probability
previous definition, and may be more useful in some situa- p/3: the qubit is negated, or its phase is changed, or it is both
tions. negated and its phase is changed. If t/n.p1 e for e .0, the
Assume that a channel W transmits qubits with a fidelity length projection of the output of this channel onto the sub-
of F and is that the decoherence process affects each qubit space of Hn2 with at most t errors approaches 1 as n grows,
independently, i.e., each the decoherence of one qubit has no so the quantum error-correcting codes given earlier in this
correlation with the decoherence of any other qubit. This paper guarantee high fidelity. This channel can alternatively
would follow from the assumption that each qubit has a dif- be described as transmitting a qubit error-free with probabil-
ferent environment, and this situation corresponds to memo- ity 12 34 p, and producing a random quantum state with prob-
ryless channels in classical information theory. Then ability 34 p. This description shows that the entropy of the
EW ^ x u W u x & >F for every state u x & PH2 . If the output of our output of the channel is at least H 2 ( 32 p), so by the Levitin–
channel is a pure state, our error-correction procedure Holevo theorem an upper bound on the classical information
Rp R f will be successful with probability equal to the length capacity of this channel is 12H 2 ( 32 p). This bound is plotted
of the projection of the state onto the subspace of Hn2 which in Fig. 1. For this channel, the bound is achievable for clas-
results from decoherence of any t or fewer qubits. Since the sical information, but we believe it is unlikely to be tight for
decoherence process for each qubit is independent, we can quantum information.
use the binomial theorem to calculate the probability that the Another question that has been studied is: how much en-
state W n u y & is projected onto the correctable subspace of tanglement can be transmitted over a quantum channel @11#.
Hn2 , where u y & is in our quantum code C. We thus have a Since any means of transmitting quantum states with high
channel which transmits states u y & with fidelity fidelity can also be used to transmit entanglement, upper
54 GOOD QUANTUM ERROR-CORRECTING CODES EXIST 1105

bounds for entanglement transmission also apply to the ACKNOWLEDGMENTS

quantum information capacity of a quantum channel. For
the above channel, the upper bound proved in Ref. @11# is We would like to thank Peter Winkler for helpful discus-
H 2 „ 21 1 Ap(12 p)… for p, 21 and 0 if p> 21 . This bound is also sions on quantum error-correcting codes, and David DiVin-
plotted in Fig. 1. cenzo for advice on the presentation of these results.

@1# D. Deutsch, Proc. R. Soc. London Ser. A 400, 96 ~1985!; D. Correcting Codes ~North-Holland, Amsterdam, 1977!.
Simon, in Proceedings of the Thirty-Fifth Annual Symposium @7# R. L. Dobrushin and S. I. Ortyukov, Probl. Peredachi Inf. 13
on Foundations of Computer Science, edited by S. Goldwasser ~1!, 82 ~1977! @Probl. Inf. Transm. ~USSR! 13, 59 ~1977!#;
~IEEE Computer Society, Los Alamitos, CA, 1994! p. 116; P. Probl. Peredachi Inf. 13 ~3!, 56 ~1977! @Probl. Inf. Transm.
W. Shor, ibid., p. 124; A. Ekert and R. Jozsa, Rev. Mod. Phys. ~USSR! 13, 203 ~1977!#; N. Pippenger, J. Assoc. Comput.
~to be published!. Mach. 36, 531 ~1989!; N. Pippenger, G. D. Stamoulis, and J.
@2# S. Lloyd, Science 261, 1569 ~1993!; D. P. DiVincenzo, Phys. N. Tsitsiklis, IEEE Trans. Inf. Theory 37, 639 ~1991!, U.
Rev. A 51, 1015 ~1995!; A. Barenco, D. Deutsch, A. Ekert, Feige, P. Raghavan, D. Peleg, and E. Upfal, SIAM J. Comput.
and R. Jozsa, Phys. Rev. Lett. 74, 4083 ~1995!; T. Sleator and 23, 1001 ~1994!.
H. Weinfurter, ibid. 74, 4087 ~1995!; I. L. Chuang and Y. @8# W. K. Wooters and W. H. Zurek, Nature 299, 802 ~1982!; D.
Yamamoto, Phys. Rev. A 52, 3489 ~1995!. Dieks, Phys. Lett. A 92, 271 ~1982!.
@3# J. I. Cirac and P. Zoller, Phys. Rev. Lett. 74, 4091 ~1995!. @9# P. W. Shor, Phys. Rev. A 52, 2493 ~1995!.
@4# W. G. Unruh, Phys. Rev. A 51, 992 ~1995!; G. M. Palma, @10# L. B. Levitin, in Proceedings of the Fourth All-Union Confer-
K.-A. Suominen, and A. K. Ekert, Proc. R. Soc. London A ence on Information Theory, Tashkent ~1969!, p. 111, in Rus-
452, 567 ~1996!; R. Landauer, Philos. Trans. R. Soc. London, sian; A. S. Kholevo, Probl. Peredachi Inf. 9 ~3!, 3 ~1973!
Ser. A 353, 367 ~1995!; R. Landauer, in Proceedings of the @Probl. Inf. Transm. ~USSR! 9, 177 ~1973!#.
Drexel-4 Symposium on Quantum Nonintegrability — Quan- @11# C. H. Bennett, G. Brassard, S. Popescu, B. Schumacher, J. A.
tum Classical Correspondence, edited by D. H. Feng and B.-L. Smolin, and W. K. Wooters, Phys. Rev. Lett. 76, 722 ~1996!.
Hu, ~International Press, in press!; I. L. Chuang, R. Laflamme, @12# A. M. Steane, Proc. R. Soc. London Ser. A ~to be published!.
P. W. Shor, and W. H. Zurek, Science 270, 1635 ~1995!. @13# F. J. MacWilliams, N. J. A. Sloane, and J. G. Thompson, Dis-
@5# C. Monroe, D. M. Meekhof, B. E. King, W. M. Itano, and D. crete Math. 3, 153 ~1972!.
J. Wineland, Phys. Rev. Lett. 75, 4714 ~1995!. @14# B. Schumacher, Phys. Rev. A 51, 2738 ~1995!; R. Jozsa and B.
@6# F. J. MacWilliams and N. J. A. Sloane, The Theory of Error- Schumacher, J. Mod. Opt. 41, 2343 ~1994!.