VOLUME 81, NUMBER 26 PHYSICAL REVIEW LETTERS 28 DECEMBER 1998

Quantum Repeaters: The Role of Imperfect Local Operations in Quantum Communication

H.-J. Briegel,1,2, * W. Dür,1 J. I. Cirac,1,2 and P. Zoller1
1
Institut f ür Theoretische Physik, Universität Innsbruck, Technikerstrasse 25, A-6020 Innsbruck, Austria
2
Departamento de Fisica, Universidad de Castilla-La Mancha, 13071 Ciudad Real, Spain
(Received 20 March 1998)
In quantum communication via noisy channels, the error probability scales exponentially with the
length of the channel. We present a scheme of a quantum repeater that overcomes this limitation. The
central idea is to connect a string of (imperfect) entangled pairs of particles by using a novel nested
purification protocol, thereby creating a single distant pair of high fidelity. Our scheme tolerates general
errors on the percent level, it works with a polynomial overhead in time and a logarithmic overhead in
the number of particles that need to be controlled locally. [S0031-9007(98)08063-6]

PACS numbers: 03.67.Hk, 03.65.Bz, 42.50. – p

Quantum communication deals with the transmission the distance between the nodes is presently limited by (a
and exchange of quantum information between distant few times) the absorption length of the fiber [11].
nodes of a network. Remarkable experimental progress The theory of fault-tolerant quantum computing [12]
has been reported recently, for example, on secret key implies that any computation can be performed with
distribution for quantum cryptography [1,2], teleportation polylogarithmic cost in time and space [13], if the
of the polarization state of a single photon [3,4], and the error probability for each gate operation can be made
creation of entanglement between different atoms [5]. On sufficiently small. A special case of a computation is the
the other hand, first steps towards the implementation of transmission of information, for which these fault-tolerant
quantum logical operations, which are the building blocks methods must therefore have the same (or a better)
of quantum computing, have been demonstrated [6]. In asymptotic complexity. An explicit scheme for quantum
view of this progress, it is not farfetched to expect the transmission has been discussed by Knill and Laflamme,
creation of small quantum networks in the near future. using concatenated quantum codes [14]. Their method
Such networks will involve nodes, where qubits are stored requires one to encode a single qubit into an entangled
and locally manipulated, and which are connected by state of a polynomially large number of qubits, and to
quantum channels over which communication takes place operate on this code repeatedly during the transmission
by sending qubits. This will open the possibility for more process. The tolerable error probabilities for transmission
complex activities such as multiparty communication and are less than 1022 , whereas for local operations they are
distributed quantum computing [7]. less than 5 3 1025 . This seems to be outside the range
The bottleneck for communication between distant of any practical implementation in the near future. A
nodes is the scaling of the error probability with the length crucial figure for any experiment will be the number of
of the channel connecting the nodes. For channels such particles that can be manipulated locally in a coherent
as an optical fiber, the probability for both absorption fashion, together with the precision with which such local
and depolarization of a photon (i.e., the qubit) grows manipulations can be realized.
exponentially with the length l of the fiber. This has In this Letter, we present a model of a quantum repeater
two effects: (i) to transmit a photon without absorption, that allows the creation of an entangled (EPR) pair of par-
the number of trials scales exponentially with l; (ii) even ticles over arbitrary large distances with a tolerability of
when a photon arrives, the fidelity of the transmitted state errors in the percent region. Once an EPR pair is cre-
decreases exponentially with l. One may think that this ated, it can be employed to teleport any quantum infor-
last problem can be circumvented by standard purification mation [15,16]. Our solution of this problem comprises
schemes [8–10]. However, purification schemes require three novel elements: (i) entanglement purification with
a certain minimum fidelity Fmin to operate, which cannot imperfect means, including analytic results for the range
be achieved as l increases. Furthermore, in any realistic and the working conditions of standard protocols; (ii) a
situation, the operations that are part of the purification method for creation of entanglement between particles at
protocol are themselves imperfect, and this defines a distant nodes that uses auxiliary particles at intermediate
maximum attainable fidelity Fmax and limits the efficiency “connection points” and a nested purification protocol;
of the scheme. For this reason, it is not obvious, first, (iii) a scheme for which the time needed for entanglement
what the allowed error tolerances of local operations are creation scales polynomially whereas the required material
for entanglement purification to be applicable at all and, resources per connection point grow only logarithmically
second, how the resources that are needed for purification with the distance. Since our model is based on two-way
grow with the length of the channel. In the experiments, classical communication, it is qualitatively different from

5932 0031-9007y98y81(26)y5932(4)$15.00 © 1998 The American Physical Society

VOLUME 81, NUMBER 26 PHYSICAL REVIEW LETTERS 28 DECEMBER 1998

quantum error correction. By exploiting this property we on the initial fidelity, the degradation of the fidelity un-
will obtain a higher efficiency and significantly more fa- der connections, and the efficiency of the purification pro-
vorable error tolerances. tocol. The total number of elementary pairs involved in
In classical communication, the problem of exponential constructing one of the more distant pairs of length L is
attenuation can be overcome by using repeaters at certain LM. On the second level, we connect L of these more
points in the channel, which amplify the signal and distant pairs at every checkpoint CkL (k ­ 1, 2, . . .) except
restore it to its original shape. Guided by these ideas, at CL2 , C2L2 , . . . , CN2L2 . As a result, we have NyL2 pairs
for quantum communication, we divide the channel into of length L2 between A and CL2 , CL2 and C2L2 , and so
N segments with connection points (i.e., auxiliary nodes) on, of fidelity $FL . Again, we need M parallel copies of
in between. We then create N elementary EPR pairs these long pairs to repurify up to the fidelity $F1 . The
of fidelity F1 between the nodes A and C1 , C1 and total number of elementary pairs involved in constructing
C2 , . . . , CN21 and B, as in Fig. 1(a). The number N one pair of length L2 is thus sLMd2 . We iterate the pro-
is chosen such that Fmin , F1 & Fmax . Subsequently, cedure to higher and higher levels, until we reach the nth
we connect these pairs by making Bell measurements at level. As a result, we have obtained a final pair between
the nodes Ci and classically communicating the results A and B of length N and fidelity $F1 . In this way, the
between the nodes as in the schemes for teleportation total number R of elementary pairs will be sLMdn . We
[15] and entanglement swapping [15,17]. Unfortunately, can reexpress this result in the form
with every connection, the fidelity F 0 of the resulting pair R ­ N logL M11 , (1)
will decrease: on the one hand, the connection process
which shows that the resources grow polynomially with
involves imperfect operations which introduce noise; on
the distance N. A similar formula was obtained in [14]
the other hand, even for perfect connections, the fidelity
for the overhead required in propagating the concatenated
decreases. Both effects lead to an exponential decrease
quantum code. Note that R depends only on L and M. In
of the fidelity FN with N of the final pair shared between
order to evaluate M, we need to know the specific form
A and B. Eventually, the value of FN drops below Fmin ,
of the error mechanisms involved in the purification and
and therefore it will not be possible to increase the fidelity
connections, which in turn depend on the specific physical
by purification. The number of pairs L ø N that may be
implementation of the quantum network. In general, we
connected by this method seems therefore to be restricted
have only limited knowledge of these details. In order
by the condition FL . Fmin .
to estimate M, we will choose a generic error model for
Our proposal, the nested purification protocol, combines
imperfect operations and measurements.
the methods of entanglement swapping and purification
We define imperfect operations on states of one or more
into a single (meta) protocol that circumvents this restric-
qubits by the following maps:
tion. For simplicity, assume that N ­ Ln for some inte-
1 2 p1
ger n. On the first level, we simultaneously connect the r ! O1 r ­ p1 O1ideal r 1 tr1 hrj ≠ I1 , (2)
pairs (initial fidelity F1 ) at all of the checkpoints except 2
at CL , C2L , . . . , CN2L . As a result, we have NyL pairs of 1 2 p2
length L and fidelity FL between A and CL , CL and C2L , r ! O12 r ­ p2 O12 ideal
r1 tr12 hrj ≠ I12 , (3)
4
and so on. To purify these pairs, we need a certain num- the first of which describes an imperfect one-qubit op-
ber M of copies that we construct in parallel fashion. We eration on particle 1, and the second an imperfect two-
then use these copies on the segments A and CL , CL and qubit operation on particles 1 and 2. In these expressions,
C2L , etc., to purify and obtain one pair of fidelity $F1 on O ideal is the ideal operation, and I1 and I12 denote unit
each segment. This last condition determines the (aver- operators on the subspace where the ideal operation acts.
age) number of copies M that we need, which will depend The quantities p1 and p2 measure the reliability of the
operations. The expressions (2) and (3) describe a situ-
1
ation where we have no knowledge about the result of an
(a) (c) error occurring during some operation (“depolarization”),
0.9 F
A C1 C2 B 0.8
max
F´ except that it happens with a certain probability s1 2 pj d.
0.7 Any sequence of two one-qubit operations on the same
(b) 0.6 Fmin qubit is equivalent to a single one-qubit operation, and is
0.5 FL therefore described by a single parameter p1 . Similarly,
0.4 a sequence of a two- and a one-qubit operation counts
0.3 as a single two-qubit operation and is thus described by
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 p2 . An imperfect measurement on a single qubit in the
F computational basis is described by a POVM (positive-
FIG. 1. (a) Connection of a sequence of N EPR pairs; (b) operator-valued measure) corresponding to
h
nested purification with repeated creation of auxiliary pairs; P0 ­ hj0l k0j 1 s1 2 hd j1l k1j ,
(c) “purification loop” for connecting and purifying EPR pairs. (4)
h
Parameters are L ­ 3, h ­ p1 ­ 1, and p2 ­ 0.97. P1 ­ hj1l k1j 1 s1 2 hd j0l k0j .
5933
VOLUME 81, NUMBER 26 PHYSICAL REVIEW LETTERS 28 DECEMBER 1998

The parameter h is a measure for the quality of the projec- works even in the presence of errors which we take into
tion onto the basis states. For example, for the state r ­ account. Connecting L neighboring pairs as explained ear-
j0l k0j the measuring apparatus will give the wrong result lier, one obtains a new “L pair” with fidelity
(“1”) with probability 1 2 h $ 0. A detailed discussion √ !L21 √ !L
of this and more general models for imperfect operations 1 3 p12 p2 s4h 2 2 1d 4F 2 1
will be given elsewhere [18]. With these error models, we FL ­ 1 . (5)
4 4 3 3
have a toolbox to analyze all of the processes involved in
the connection and purification procedures. For example,
the Bell measurement required in the connection can be This formula describes an exponential decrease of the
decomposed into a controlled-NOT (CNOT) operation, ef- resulting fidelity, unless both the elementary pairs and
fecting, e.g., j0l j0l 6 j1l j1l ! sj0l 6 j1ld j0l, followed all of the operations involved in the connection process
by two single-qubit measurements. are perfect. There are several possibilities to do the
The basic elements of the nested purification protocol purification, and we first analyze the scheme introduced
are (i) pair connections and (ii) purification. In the fol- by Bennett et al. [8] in the case of imperfect gate and
lowing we analyze these elements using the error models measurement operations. In short, the scheme takes two
introduced above. Assume now that all of the pairs in adjacent L pairs of fidelity F, performs local (1-bit and
Fig. 1(a) are in Werner states (see [8]). These states can 2-bit) operations on the particles at the same ends of the
be produced using depolarization (as in Ref. [8]) after each pairs, and obtains with a certain probability psucc a new
connection and purification process. This depolarization pair of fidelity
12F 2 12F 12F 2 12p22
fF 2 1 s 3 d g fh
2
1 s1 2 hd2 g 1 fFs 3 d 1s 3 d g f2hs1 2 hdg 1 s 8p22
d
F0 ­ 2 5 12F 12F 2 12p22
. (6)
fF 2 1 3 Fs1 2 Fd 1 9 s1 2 Fd2 g fh 2 1 s1 2 hd2 g 1 fFs 3 d 1s 3 d g f8hs1 2 hdg 1 4s 8p22
d

The value of psucc is given by p22 times the denominator from F, the fidelity FL after connecting L pairs can be
of this expression. For perfect operations, h ­ 1 and read off from the curve below the diagonal. Reflecting
p2 ­ 1, (6) reduces to the formula given in Ref. [8]. this value back to the diagonal line, as indicated by
Figure 1(c) shows the curves for connection (5) and the arrows in Fig. 1(c), sets the starting value for the
purification (6) for a certain set of parameters. The purification curve. If FL lies within the purification
purification curve has three intersection points with the interval, then iterated application of (6) leads back to the
diagonal, which are the real fixpoints of the map (6). In initial value F (staircase). Once the initial value F is
addition to the trivial point at F ­ 1y4, there are two reobtained, we have NyLk pairs and we can start with
nontrivial fixpoints. The upper point, Fmax , 1, is an the level k 1 1. In summary, each level in the protocol
attractor and gives the maximum value of the fidelity corresponds to one cycle in Fig. 1(c). Note that if, in
beyond which no pair can be purified. Note also the the loop, FL # Fmin then purification is not possible.
existence of the minimum value Fmin . 1y2. Together, Being polynomial in F, the lower curve gets steeper
they define the interval within which purification is and steeper near F ­ 1 for higher values of L. From
possible. The limiting situation Fmax ­ Fmin defines the this, one sees that for a given starting fidelity F, there
threshold for the applicability of the purification protocol. is a maximum number of pairs one can connect before
For all pairs sp2 , hd for which there is only one real purification becomes impossible. Q max smd
fixpoint (at F ­ 1y4), the imperfections of the local For the resources we obtain M ­ m m 2ypsucc ,
smd
operations introduce more noise than one gains from the where psucc is the probability of obtaining the required
purification, so the scheme breaks down. For example, outcome (00 or 11) in the measurement at the mth
for h ­ 1 the threshold is at p2 . 0.95; that is, the CNOT purification step. The total number of steps, mmax , is the
gate must work with a reliability of 95%, at least. Please same as in the staircase of Fig. 1(c).
note that the fixpoints and the threshold condition can In Fig. 2(a), M is plotted against the working fidelity F.
all be given analytically from (6). The connection curve, Because of the discrete nature of the purification process,
which looks like a simple power in Fig. 1(c), stays below the fidelity of the repurified pairs need not be exactly the
the diagonal for all values of F between 1y4 and 1. The same on each nesting level. The working fidelity is thus
offset of this curve at F ­ 1 from the ideal value F 0 ­ 1 defined as the fidelity maintained on average when going
quantifies the amount of noise that is introduced through through different nesting levels. The error parameters for
imperfect operations in the connection process. this plot are h ­ p1 ­ p2 ­ 0.995. One can see that
With the above results, we can now analyze the nested there exists an optimum working fidelity of about 0.94
purification protocol. Let us consider a given level k in which requires a minimum number of about 15 resources.
this protocol, where we have NyLk21 pairs of fidelity F A purification protocol that converges faster and
each. The two-step process connection-purification can therefore involves less parallel channels was proposed
now be visualized as follows [see Fig. 1(c)]. Starting by Deutsch et al. [9]. We have employed this protocol,
5934
VOLUME 81, NUMBER 26 PHYSICAL REVIEW LETTERS 28 DECEMBER 1998

100 15 Fig. 1(b). Here, the dependence of the fixpoint on the

(a) (b) form of the auxiliary pair becomes quite important: When
75
10 we use our method together with the recurrence protocol
M 50 M of Ref. [8], which is based on Werner states, the fixpoint
5 Fsp0 d is too small and the nesting condition cannot be
25
satisfied for any L $ 2. On the other hand, the nesting
0
0.8 0.85 0.9 0.95 1 0
0.8 0.85 0.9 0.95 1
condition can be satisfied if we adopt a similar sequence
F F of local operations as in Ref. [9], which does not involve
a depolarization to Werner states.
FIG. 2. M (see text) versus working fidelity F. (a) Realiza- Using this method, the vertical axes in Fig. 2 are
tion of the repeater with the aid of the purification schemes of essentially translated into temporal resources [18]. On
Refs. [8] (upper curve) and [9] (lower curve). The error proba-
bilities of all operations are 0.5% (error parameters 0.995), and the other hand, the number of particles at each node
L ­ 2. (b) Lower curve in (a) for different error probabilities. [see Fig. 1(b)] increases by unity with every additional
From bottom to top: 0%, 0.25%, 0.5%, 0.75%, 1%. nesting level, and thus depends only logarithmically on
the distance between the initial and the final node. In
the context of a quantum optical implementation [19], for
using imperfect operations (2)–(4). As is demonstrated example, this would correspond to the number of ions that
in Fig. 2(a), M can be reduced by a factor of the order need to be controlled in a cavity at each node [20]. Note,
of 10. Since this number has to be taken to the nth however, that this method requires perfect memory during
power, this reduces the number of total resources that the process. In this particular implementation, the storage
are required at each connection point by many orders of decoherence time is orders of magnitude longer than the
magnitude. In Fig. 2(b), M is plotted versus the working estimated duration of the process [20].
fidelity for different error parameters. One can see that This work was supported in part by the Austrian
for errors in the one percent region, a working fidelity Science Foundation, and by the TMR network ERB-
can be maintained with, on average, five L pairs on each FMRX-CT96-0087.
nesting level. We note that the procedure also works
for error probabilities up to about 3%, but the number of
purification resources gets larger.
*On leave from Institut für Theoretische Physik, Uni-
In the remainder of this paper, we propose a method
versität München, Theresienstrasse 37, D-80333
for which the resources grow only logarithmically with München, Germany.
the distance, whereas the total time needed for building [1] W. Tittel et al., Phys. Rev. Lett. 81, 3563 (1998).
the pair scales polynomially. Imagine that we purify a [2] W. T. Buttler et al., Phys. Rev. Lett. 81, 3283 (1998).
pair not with the help of M copies, but instead with one [3] D. Bouwmeester et al., Nature (London) 390, 575 (1997).
auxiliary pair of constant fidelity p0 that is repeatedly [4] D. Boschi et al., Phys. Rev. Lett. 80, 1121 (1998).
created at each purification step. The purification with [5] E. Hagley et al., Phys. Rev. Lett. 79, 1 (1997).
the help of such a pair leads to a maximum achievable [6] C. Monroe et al., Phys. Rev. Lett. 75, 4714 (1995); Q. A.
fidelity Fmax sp0 d that depends on the value of p0 and, Turchette et al., ibid. 75, 4710 (1995).
more generally, on the state of the auxiliary pair. This [7] L. K. Grover, quant-ph/9704012.
purification method is different from the standard schemes [8] C. H. Bennett et al., Phys. Rev. Lett. 76, 722 (1996).
[9] D. Deutsch et al., Phys. Rev. Lett. 77, 2818 (1996).
[8,9], and the purification limit Fmax is usually smaller
[10] N. Gisin, Phys. Lett. A 210, 151 (1996).
than for the destillation method. In the context of the [11] For optical fibers, this length is typically 10 km (see [1]).
repeater protocol, it is therefore not a priori clear whether [12] P. Shor, quant-ph/9605011; A. M. Steane, Phys. Rev. Lett.
the fidelity that is lost by the connection process can be 78, 2252 (1997).
regained with this method. [13] E. Knill, R. Laflamme, and W. Zurek, Science 279, 342
When connecting L pairs of fidelity F as in Fig. 1(b), (1998); D. Aharonov and M. Ben-Or, quant-ph/9611025;
we obtain a resulting L pair of fidelity p0 ; FL . In the A. Yu. Kitaev, Russ. Math. Surv. 52, 1191 (1997).
first step, this pair is swapped to two auxiliary particles [14] E. Knill and R. Laflamme, quant/ph-9608012.
at the ends of the L pair, as indicated by the arrows [15] C. H. Bennett et al., Phys. Rev. Lett. 70, 1895 (1993).
in Fig. 1(b). In the next step, an L-pair of fidelity p0 [16] C. H. Bennett et al., Phys. Rev. A 54, 3824 (1996).
is again created by using the same string of particles [17] M. Zukowski et al., Phys. Rev. Lett. 71, 4287 (1993).
[18] W. Dür et al., quant-ph/9808065; G. Giedke et al., quant-
as before, which is now used to purify the pair stored
ph/9809043.
between the auxiliary particles. This procedure can be [19] S. J. van Enk et al., Phys. Rev. Lett. 78, 4293 (1997);
iterated and thus the stored pair be purified back to the Science 279, 205 (1998).
fidelity F given that the nesting condition Fmax sFL d . F [20] For a distance of 1280 km and a local node every 10 km
is satisfied. If this is the case, then the same procedure can [11], this amounts to, e.g., 7 ­ log2 s128d particles per
be applied at higher levels, thereby purifying correlations node and an estimated purification time of less than
between more and more distant particles as indicated in 1 s [18].

5935