Concentrating Partial Entanglement

by Local Operations

Charles H. Bennett

IBM Research Division, T.J. Watson Center, Yorktown Heights, NY 10598, USA.

Herbert J. Bernstein
arXiv:quant-ph/9511030v1 21 Nov 1995

Hampshire College, Institute for Science and Interdisciplinary Studies, Amherst, MA, 01002 USA.

Sandu Popescu

Physics Dept. Tel Aviv University, Israel.

Benjamin Schumacher
Physics Department, Kenyon College, Gambier, OH 43022, USA.

(February 1, 2008)

Abstract

If two separated observers are supplied with entanglement, in the form of n pairs of
particles in identical partly-entangled pure states, one member of each pair being given
to each observer; they can, by local actions of each observer, concentrate this entangle-

ment into a smaller number of maximally-entangled pairs of particles, for example Einstein-
Podolsky-Rosen singlets, similarly shared between the two observers. The concentration
process asymptotically conserves entropy of entanglement—the von Neumann entropy of
the partial density matrix seen by either observer—with the yield of singlets approaching,

for large n, the base-2 entropy of entanglement of the initial partly-entangled pure state.
Conversely, any pure or mixed entangled state of two systems can be produced by two
classically-communicating separated observers, drawing on a supply of singlets as their sole
source of entanglement.

PACS numbers: 03.65.Bz, 42.50.Dv, 89.70.+c

 I. INTRODUCTION

Recent results in quantum information theory have shed light on the channel resources

needed for faithful transmission of quantum states, and the extent to which these resources
can be substituted for one another. The fundamental unit of quantum information trans-
mission is the quantum bit or qubit [1]. A qubit is any two-state quantum system, such as a
spin-1/2 particle or an arbitrary superposition of two Fock states. If two orthogonal states

of the system are used to represent the classical Boolean values 0 and 1, then a qubit differs
from a bit in that it can also exist in arbitrary complex superpositions of 0 and 1, and it
can be entangled with other qubits. Schumacher’s quantum data compression theorem [1,2]
characterizes the number of qubits, sent through the channel from sender to receiver, that
are asymptotically necessary and sufficient for faithfully transmitting unknown pure states

drawn from an arbitrary known source ensemble.

Quantum superdense coding [3] and quantum teleportation [4] consume a different quan-
tum resource—namely entanglement, in the form of maximally entangled pairs of particles
initially shared between sender and receiver—and use it to assist, respectively, in the perfor-

mance of faithful classical and quantum communication. Following Schumacher’s terminol-

ogy, we define an ebit as the amount of entanglement between a maximally entangled pair of
two-state systems, such as two spin- 21 particles in the singlet state, and we inquire how many
ebits are needed for various tasks. In [4], for example, it is shown that the consumption

of one shared ebit, together with the transmission of a two-bit classical message, can be
substituted for the transmission of one qubit.
An important concept in quantum data transmission is fidelity, the probability that a
channel output would pass a test for being the same as the input conducted by someone who
knows what the input was. If a pure state ψ sent into a quantum channel emerges as the (in

general) mixed state represented by density matrix W , the fidelity of transmission is defined
as F = hψ | W | ψi. A quantum channel will be considered faithful if in an appropriate limit
the expected fidelity of transmission tends to unity. This means that the outputs are almost

2
always either identical to the inputs, or else so close that the chance of distinguishing them
from the inputs by any quantum meausrement tends to zero.
Note that qubits are a directed channel resource, sent in a particular direction from

sender to the receiver; by contrast, ebits are an undirected resource shared between sender
and receiver. For example, if you prepare two particles in a singlet state and give me one of
them, the result is the same as if I had prepared the particles and given you one of them.
Ebits are a weaker resource than qubits, in the sense that transmission of one qubit can, as

just described, be used to create one ebit of entanglement; but the sharing of an ebit, or
many ebits, does not by itself suffice to transmit an arbitrary state of a 2-state quantum
system, or qubit, in either direction. To do that, the ebits must be supplemented by directed
classical bits, as in teleportation.

One would naturally like to know whether, in order to be useful for purposes such as
teleportation, entanglement must be supplied in the form of maximally entangled pairs. In
particular, could partly-entangled pure states, such as pairs of particles in the state

cos θ | ↑A i ⊗ | ↓B i − sin θ | ↓A i ⊗ | ↑B i (1)

be used instead, and, if so, how many such pairs would be needed to substitute for one
maximally entangled pair? Note that by using the Schmidt decomposition, and absorbing

phases into the definitions of the basis states, any entangled state can be represented by a
bi-orthogonal expression of this form, with positive real coefficients [5]
d
X
Ψ(A, B) = ci | αi i ⊗ | βi i , (2)
i=1

where | α1 i , | α2 i ... | αd i and | β1 i , | β2 i ... | βd i are orthonormal states of subsystems A and

B respectively, and the coefficients ci are real and positive. From the viewpoint of either
observer, an entangled state appears as a mixed state, described by a density matrix obtained

by tracing over the degrees of freedom of the other observer. These density matrices are
diagonal in the Schmidt basis:

c2i | αi ihαi | ,
X
ρA = TrB | Ψ(A, B)ihΨ(A, B) | = (3)
i

3
and similarly for ρB .
The entanglement of a partly-entangled pure state can be naturally parameterized by
its entropy of entanglement, defined as the von Neumann entropy of either ρA or ρB , or

equivalently as the Shannon entropy of the squares of the Schmidt coefficients.

c2i log2 c2i .

X
E = −TrρA log2 ρA = −TrρB log2 ρB = − (4)
i

Without losss of generality we choose the α and β bases such that the sequence of Schmidt

coefficients c1 , c2 ... is nonincreasing.

The quantity E, which we shall henceforth often call simply “entanglement,” ranges from
zero for a product state (eg θ = 0) to 1 ebit for a maximally entangled pair of two-state
particles (eg θ = π/4). (More generally, a maximally entangled state of two subsystems has

d equally weighted terms in its Schmidt decomposition, giving log2 d ebits of entanglement,
where d is the Hilbert space dimension of the smaller subsystem.)
If a partly entangled pair, with E < 1, is used directly for teleportation, unfaithful
transmission will result. If it is used for superdense coding, the resulting classical channel

will be noisy. In this paper we show how, by local operations on a large number n of identical
partly entangled pairs, one can concentrate their entanglement into a smaller number of
maximally entangled pairs such as singlets. This process of “entanglement concentration”
is asymptotically efficient in the sense that, for large n, the yield of singlets approaches
nE−O(log n). Conversely, local operations can be used to prepare arbitrary partly-entangled

states Ψ(AB) of two subsystems from a starting material consisting of standard singlets,
again in a manner which asymptotically conserves entropy of entanglement.
We should clarify what we mean by local operations. Initially the n partly-entangled
pairs are shared between two parties (call them Alice and Bob) with Alice receiving one

member of each pair, and Bob receiving the other. This non-local sharing establishes an
initial entanglement nE between Alice and Bob. After that Alice and Bob operate locally on
their particles, with Alice for example performing unitary operations and von Neumann or
generalized measurements in the Hilbert space of her particles, and Bob performing similar

4
operations in that of his particles. We allow Alice and Bob to coordinate their actions
through exchange of classical messages, but not to exchange any quantum systems nor
to perform any nonlocal operation after the initial sharing. This restriction is of course

necessary to force Alice and Bob to use the partly entangled pairs they already have, rather
than generating perfectly entangled pairs de novo.

II. ENTANGLEMENT CONCENTRATION

In this section we describe a method whereby the entanglement present in a supply of

identical partially-entangled pairs of 2-state particles can be concentrated into a smaller
number of perfect singlets. The generalization to d > 2 state particles is straightforward.
We call the method Schmidt projection because its essential step is a projection of the joint
state of n pairs of particles onto a subspace spanned by states having a common Schmidt

coefficient.
Let n partly-entangled pairs of 2-state particles be shared between Alice and Bob, so
that the initial state is
n
Y
Ψ(A, B) = (cos θ | α1 (i)β1 (i)i + sin θ | α2 (i)β2 (i)i). (5)
i=1

When binomially expanded, this state has 2n terms, with only n + 1 distinct coefficients,
cosn θ, cosn−1 θ sin θ... sinn θ. Let one of the parties (say Alice) perform an incomplete von
Neumann measurement projecting the initial state into one of n + 1 orthogonal subspaces

corresponding to the power k = 0...n to which sin θ appears in the coefficient. Either party
can perform this measurement locally, Alice by measuring the particles she has, or Bob by
measuring the ones he has. Let Alice perform the measurement, obtaining some outcome k.
She then tells Bob which outcome she obtained. Alternatively, if Bob and Alice wish not
to communicate, Bob can perform his version of the measurement locally, and, by virtue

of the original entanglement, he will always obtain the same value of k as Alice has. The
probability of outcomes is binomially distributed, with outcome k having probability

5
 !
n
pk = (cos2 θ)n−k (sin2 θ)k . (6)
k

After some outcome k has been obtained, Alice and Bob will be left with a residual
 
n
state Ψk of their spins which is a maximally entangled state in a known 2 k
-dimensional
subspace of the original 22n dimensional space. Such states can be used without further
 
n
ado for faithful teleportation in an k
-dimensional or smaller Hilbert space; or they can be
transformed, as described below, into a standard form such as singlets.
Before describing this optional standardization process, we note that the measurement of

k occasionally yields a residual state Ψk with more entropy of entanglement than the original
state Ψ. However, neither the measurement of k nor any other local processing by one or
both parties can increase the expected entropy of entanglement between Alice’s and Bob’s
subsystems. Consider a measurement or other local treatment applied by Alice, resulting

in a classical outcome j and a residual pure state Ψj of the joint system. This treatement
cannot influence the partial density matrix ρB seen by Bob, since if it did one would have a
superluminal communications channel based on Alice’s applying or not applying the treat-
ment and Bob measuring ρB . Therefore, depending on the extent of correlation between the

residual state and the classical outcomes, the expected entanglement of the residual states
lies between E(Ψ) − H and E(Ψ), where E(Ψ) is the original pure state’s entanglement
and H = − pj log2 pj is the Shannon entropy of the measurement outcomes. All local
P
j

treatments (eg generalized or positive operator valued measurements [6]) that Alice might
apply can be cast in this form, if necessary by considering her operations to be performed

in an appropriately enlarged Hilbert space. In particular, unitary transformations by Alice

correspond to one-outcome measurements, which cannot change Bob’s partial density matrix
ρB at all, and can only change the eigenvectors, but not the eigenvalues, of Alice’s. By the
same argument, local actions by Bob cannot increase the expected entanglement between

his and Alice’s subsystems.

Thus, though Alice and Bob cannot by local actions increase their expected entanglement,
they can gamble with it, spending their initial amount on a chance of obtaining a greater

6
amount.
We now show how the entanglement in the above perfectly-entangled residual states
Ψk can be efficiently transformed into a standard form such as singlets. Fix some small

positive ǫ, with ǫ = 0 corresponding to perfect efficiency of transformation. Let the above

measurement of k be performed independently on a sequence of batches of n pairs each. Each
performance yields another k value; let the resulting sequence of k values be k1 , k2 , ...km ,
and let
! ! !
n n n
Dm = ... (7)
k1 k2 km
 
n
be the product of the k
values for the first m batches. The sequence is continued until the

accumulated product Dm slightly lies between 2ℓ and 2ℓ (1 + ǫ) for some power ℓ. For any
single-pair entanglement E and any positive ǫ, the probability of failing to come this close
to a power of two tends to zero with increasing m. Once a suitable Dm is found, a local
measurement is performed by Alice or Bob or both to project the joint system into one of

two orthogonal subspaces, a large space of dimension 2 · 2ℓ and a relatively small space of
dimension 2 · (Dm − 2ℓ ) < ǫ · 2 · 2ℓ . In the latter case, occurring with probability less than
ǫ, a failure has occurred, and all or most of the entanglement will have been lost. In the
former case, occurring with probability greater than 1 − ǫ, the residual state is a maximally

entangled state of two 2l dimensional subsystems, one held by Alice and one held by Bob.
Using the Schmidt decomposition, this can be converted by local unitary operations into a
product of ℓ standard singlets.

III. EFFICIENCY

We deal first with the efficiency of the initial concentration stage, which yields a bino-
mially distributed measurement result k and collapses the initial state of n partly-entangled
 
n
pairs into a maximally-entangled state between two systems of dimensionality k
. We
show below that its efficiency approaches unity for large n. Then we show that the sec-

7
ond, or standardization, stage, which distills standard singlets from these high-dimensional
maximally-entangled states, also approaches unit efficiency for large m.
We adopt the local viewpoint of one of the observers, say Alice. From her viewpoint,

the initial state is a mixed state of entropy nE. Performing the measurement of k splits off
some of this entropy in the form of the entropy of the distribution of outcomes k, and leaves
the rest of it as concentrated entropy of entanglement between the two residual maximally-
 
n
entangled k
-dimensional subsystems. The expected amount of concentrated entropy of

entanglement is given by
n−1
! !
n n
(cos2 θ)n−k (sin2 θ)k
X
log2 . (8)
k=1 k k

Because the entropy of the binomial distribution of k values increases only logarithmically

with n, the fraction of the original entanglement nE captured as concentrated entanglement

approaches 1 in the limit of large n.
We now the show that the efficiency of distilling standard singlets from large-dimensional
fully-entangled states produced by the measurements of k also tends to unity. In the light of

the previous discussion, it suffices to show that for any fixed batch size n > 1, the sequence

Zm = mod(log2 Dm , 1) = log2 Dm − ℓ (9)

of mantissas of base-2 logarithms of Dm , where Dm is given in equation 7, has an infimum

   
n n
of 0. This in turn follows from the fact that the binomial coefficients k1
, k2
... are inde-

pendently drawn from a fixed distribution (eq 6). The evolution of Zm with increasing m
may therefore be viewed as a random walk on the unit interval (with wraparound), starting
 
n
at the origin and taking steps of sizes mod(log2 k
, 1) and probabilities given by equation
6. It is elementary to show that for any distribution of step sizes, and any positive ǫ, such

a walk visits the interval [0, log2 (1 + ǫ)] with probability 1, from which the theorem follows.
The Schmidt projection method of entanglement concentration requires at least n = 2
partly-entangled pairs, and only becomes efficient for large n. We now describe another
method that works, albeit inefficiently, even with a single partly entangled pair, as in eq. 1.

8
We call this procedure the Procrustean method of entanglement concentration, because its
goal is to cut off and discard the extra probability of the larger term in equation 1, leaving
a perfectly entangled state. Assume for the moment that θ < π/4 so that if Alice measures

particle 1 in the up/down basis, the up outcome is more likely. Instead of performing this von
Neumann measurement, she passes her particle through a polarization-dependent absorber,
or a polarization-dependent-reflector (eg, for light, a Brewster window), which has no effect
on down spins but absorbs, or deflects into a different beam, a fraction tan2 θ of the up spins.

If the particle is absorbed or deflected, it is rejected; otherwise it is kept. This treatment does
not correspond to any von Neumann measurement in the original 2-dimensional spin space,
but rather to a two-outcome generalized measurement or POVM (positive-operator-valued
measurement) [6,7]. If the particle is not absorbed or deflected, its residual state after this

treatment will be a maximally mixed state of spin up and spin down. Now suppose Alice
tells Bob the result of her generalized measurement, and suppose that he does not measure
his particle at all, but simply discards it if Alice has discarded hers. The result will be a
perfectly entangled state of two particles. The Procrustean method is especially suitable for

the type of gambling mentioned earlier: when it works, it always yields more entanglement
than the parties started out with.
Both the Schmidt projection and the Procrustean method can be generalized to work
on larger Hilbert spaces. Like von Neumann’s method for obtaining unbiased random bits
from a coin of unknown but time-independent head/tail ratio (setting HT=1, TH=0, and

TT=HH=do over), Schmidt projection works even when Alice and Bob do not know how
entangled their partly-entangled pairs are, provided all n pairs have equal biases θ. The
Procrustean method, on the other hand, requires the bias to be known in advance.
Figure 1 plots the yield of perfectly entangled pairs as a function of cos2 θ obtained by

the Schmidt projection method with n = 2, 4, 8, and 32 (lower 4 curves), in comparison with
ideal asymptotic yield nE = nH2 (cos2 θ), (top curve) and the yield from the Procrustean
method (inverted-V shaped curve). Note that for n < 5 Schmidt projection is absolutely
less efficient than the Procrustean method.

9
 IV. RELATION OF ENTANGLEMENT CONCENTRATION TO QUANTUM

DATA COMPRESSION

Like the Schmidt projection method of entanglement concentration described above,

the technique of quantum data compression [2,1] involves projecting the state of a high-
dimensional system onto a set of orthogonal subspaces depending on eigenvectors and eigen-
values of an associated density matrix. However, the goals and means of the two techniques
are sufficientlly different that neither can be substituted for the other. Indeed, certain quan-
tum data transmission tasks can only be accomplished efficiently by using the two techniques

together.
Quantum data compression (QDC) has the goal of encoding an unknown sequence of
signals from a known quantum source—ie an ensemble of pure states {ψj } emitted with
specified probabilities {pj }—into a smaller Hilbert space than it originally occupies, while

introducing negligible distortion. QDC is useful when the source has less than maximal
entropy (permitting it to be compressed at all) and consists of non-orthogonal states ψj ,
necessitating the use of quantum operations to do the compression.
As has previously been noted [1,8,12], a quantum source is not fully specified by its

density matrix ρ = pj | ψj ihψj |. By the same token, it is also not fully specified by
P
j

giving an entangled state ΨAB of which its density matrix is the partial trace, eg ρ = ρA =
TrB | ΨAB ihΨAB |. A quantum source can, however, be fully specified by giving both such
an entangled state ΨAB and a von Neumann or generalized measurement to be performed
by Bob, who holds subsystem B. This is done in such a way that each of Bob’s possible

measurement outcomes projects Alice’s subsystem into one of the states ψj , and the outcomes
occur with the required probabilities {pj }. Then each of Bob’s measurements tells him
which state Alice received from the source at that instant. For example, depending on the
measurement performed by Bob, the entangled state of equation 1 can be used to generate

either of the following two sources for Alice, one classical, the other distinctively quantum.

• Source Q, consisting of orthogonal states | ↑i and | ↓i emitted with unequal probabili-

10
 ties cos2 θ and sin2 θ, respectively; and

• Source Q′ consisting of non-orthogonal states ψ 0 = (cos θ | ↑i + sin θ | ↓i) and ψ 1 =

(cos θ | ↑i − sin θ | ↓i) emitted with equal probabilities.

The first source, Q, is purely classical in the sense that it could be faithfully compressed
by making a complete von Neumann measurement in the up/down basis, and applying
conventional data compression (eg Huffman coding) to the resulting classical data.
Although the other source, Q′ , would yield statistically similar data when measured in

the up/down basis, the resulting data would be useless for reliably encoding a sequence of
n states from the source, because the data would be utterly uncorrelated with which of the
2n equiprobable non-orthogonal spin sequences the source had emitted.
At this point some notation is helpful. Let x denote an arbitrary n-bit sequence, where

n > 2, and let Ψx denote the n-spin product state resulting when source Q′ emits a sequence
of states indexed by the bits of x. For example, taking x = 011, we have Ψ011 = ψ 0 ⊗ψ 1 ⊗ψ 1 .
In order to transmit such sequences faithfully and economically, one uses quantum data
compression [2,1]. This consists of performing a very gentle, incomplete measurement on
the the joint state Ψx of the spins, which projects the state into one of two complementary

subspaces :

• a “likely” subspace of dimensionality 2n(H(ρ)+δ) spanned by the eigenvectors of the

largest eigenvalues of the joint density matrix ρ(n) , defined as the tensor product of n
copies of ρ.

• the “unlikely” subspace spanned by the remaining eigenvectors.

If the joint state projects into the likely subspace, one transmits the resulting projected state;

if the projection fails, one transmits an arbitrary state. Using W x to denote the (slightly
mixed) state resulting from applying quantum data compression to a source sequence Ψx
from the source Q′ , it can be shown that for any positive δ and ǫ there exists an n0 such
that for all n > n0 , the fidelity of the quantum data compression, ie the probability

11
 p(x) hΨx | W x | Ψx i
X
F = (10)
x

that its output would pass a test for being the same as the input sequence Ψx , conducted
by someone who knew what the input sequence was, is greater than 1 − ǫ. Infidelity can

be thought of as resulting from two causes: failure to project into the likely subspace, and
failure of even a successful projection into that subspace to agree with the original state
when subsequently tested. Both kinds of infidelity become negligible in the limit of large
n 1.

Because a less-than-maximally entangled pure state appears as a less-than-maximally

random mixed state to its two separate observers, one could imagine using quantum data
compression as an approximate means of entanglement concentration. In other words, by
separately compressing their respective subsystems, Alice and Bob could squeeze the original

entanglement into a smaller number of shared pairs of qubits. Applying this two-sided
compression to n shared pairs of entanglement E, Alice’s and Bob’s projections into the likely
subspace would be isomorphic (this can be seen by considering the Schmidt representation
of the original entangled state), and the result would be to leave each observer with slightly
more than nE qubits having slightly less than nE bits entropy of entanglement with an

equal number of qubits held by the other observer. However, the entangled sates produced
by such two-sided quantum data compression are never maximally entangled, and in an
important sense they become poorer approximations of maximally-entangled states as n
increases. A perfectly-entangled state must have all its Schmidt coefficients equal, but

the compressed bipartite state Ψc (A, B), resulting from applying two-sided quantum data

1 Although one is usually more interested in data compression, it is also possible to perform classical

or quantum data expansion, in other words to encode the output of a nonredundant source into a

larger number of redundant bits or qubits. For example, all states of two qubits can be encoded

into the majority-up subspace of three qubits, each of which has a single-particle density matrix

with eigenvalues 3/4 and 1/4 and entropy of approximately 0.811 bits.

12
compression to a state such as the Ψ(A, B) of eq. 5, describing a set of n partly-entangled
pairs, has a distribution of values of Schmidt coefficents whose variance increases with n.
In consequence, the fidelity with which Ψc (A, B) approximates any maximally entangled

state approaches zero with increasing n, as does the fidelity with which it would work in
teleporting a random state in a Hilbert space of dimensionality equal to that of one of its
two parts.
We have just shown that two-sided quantum data compression does not work as a method

of entanglement conentration. Conversely, because any mixed state can be regarded as the
partial trace of an appropriate entangled state, one can imagine attempting to use the
Schmidt projection method of entanglement concentration in a one-sided manner as a way
of performing quantum data compression. As we will show presently, this also does not

work. To use Schmidt projection in a one-sided manner would mean making a more aggres-
sive projection than that used in conventional quantum data compression, into subspaces
spanned by each distinct eigenvalue of ρ(n) , rather than into a single subspace spanned by
all the likely eigenvalues and a residual unlikely subspace. However, this projection is too

aggressive for the purposes of reliable data transmission. Because the entropy of the distri-
bution of the eigenvalues increases absolutely with n, (although it decreases as a fraction of
n), the fidelity of transmission of typical sequences from a source such as Q′ tends to zero
with increasing n.
In more detail, the proposed Schmidt projection method of quantum data-compression,

whose fidelity we seek to refute, corresponds in the case of the source Q′ to an incomplete
measurement, in which one observes the number k of down spins in an n-spin block, leaving a
residual quantum state equal to the renormalized projection of the original source sequence
into the subspace corresponding to the measurement outcome. We shall show that this

coding cannot transmit sequences from the source Q′ with asymptotically perfect fidelity,
even if the measurement outcome k is made available, as classical information, to help in the
decoding process at the receiving end of the channel. Such a dual classical/quantum channel
setup is certainly no weaker than a purely quantum channel (where the classical outcome

13
k would simply be discarded) and is analogous to the dual channel used in teleportation,
in which the quantum channel output is postprocessed in a way depending on classical
information generated at the sending end. The inability of even this strong kind of Schmidt

coding to approach perfect fidelity for source Q′ follows from the fact, to be demonstrated
presently, that with significant probability it maps distinct input sequences onto the same
output state.
Recalling notation introduced earlier, we use Ψx to denote the n-spin product state

resulting when source Q′ emits a sequence of states indexed by the bits of bit string x. For
example, taking x = 011, we have Ψ011 = ψ 0 ⊗ψ 1 ⊗ψ 1 . Let x̄ denote the Boolean complement
of x. In our example x̄ would be 100, and correspondingly Ψx̄ would be Ψ100 = ψ 1 ⊗ ψ 0 ⊗ ψ 0 .
By expanding the states Ψx and Ψx̄ in the up/down basis, and grouping terms according to

k, it can readily be seen that the states Ψx and Ψx̄ differ only by a uniform sign change in
all terms of odd k. This difference is obliterated by the Schmidt encoding process, which
by measuring k randomizes the relative phases of terms of differing k. Thus, for any x, the
encoding process, even if the classical outcome k is made available at the receiving end, maps

input states Ψx and Ψx̄ onto the same output state or distribution of output states. This
precludes the achievement of asymptotically perfect fidelity, which by definition requires
that with probability approaching unity a typical channel input be mapped onto a pure or
mixed output state arbitrarily close to the input in Hilbert space. Here, on the contrary,
we have distinct inputs Ψx and Ψx̄ , which for large n are nearly orthogonal, being mapped

onto the same output.

In summary, two-sided quantum data compression is too faithful to the original non-
maximally-entangled state to provide good entanglement concentration (whose goal is a
maximally entangled state rather than a good approximation to the initial state). Con-

versely, entanglement concentration by the Schmidt projection method necessarily sacrifices

fidelity to the original state in order to produce a maximally entangled output.
Despite these differences, the two techniques can sometimes be fruitfully combined. Sup-
pose Alice has a long sequence of n spins from a nonrandom quantum source such as Q′ which

14
she wishes to faithfully teleport to Bob with a minimal usage of engtanglement. Suppose
further that the entanglement Alice and Bob have at their disposal for teleportation purposes
is supplied not in the form of standard singlets but rather as pure but less-than-maximally

entangled states ψAB . An economical procedure would then be

• use quantum data compression to compress the input sequence to a bulk of slightly
greater than nH(ρ(Q′ )) spins.

• use entanglement concentration to prepare standard singlets from the supply of im-

perfectly entangled pairs.

• teleport each compressed spin using one of the standard singlets.

V. DISCUSSION

We have shown that the entanglement in any pure state of a bipartite system can be con-

centrated by local operations and classical communication into maximally entangled states
such as singlets. Here we note that, conversely, an arbitrary partly-entangled state Ψ(A, B)
of a bipartite system can be prepared by local operations and classical communication using
standard singlets as the only source of entanglement. One way to do this is for Alice first

to prepare a copy Ψ(A, C) of the entangled state she wishes to share with Bob, using two
systems in her laboratory, A and C. Here A is the system she wishes to entangle with
Bob’s system B, and C is a system similar to B but located in Alice’s laboratory instead
of Bob’s. Because A and C are in the same location, Ψ(A, C) can be prepared by purely

local operations. Next Alice uses a supply of standard singlets, in conjunction with classical
communication, to teleport the state of her local system C into Bob’s system B. This has
the effect [4] of destroying the entanglement between A and C while creating the desired en-
tangled state Ψ(A, B) shared remotely between Alice and Bob. The teleportation consumes

log2 d singlets and requires 2 log2 d classical bits of communication, where d is the dimen-
sion of the Hilbert space of system B, regardless of the entanglement E of the state being

15
teleported, which is rather inefficient if E is small. To reduce the consumption of singlets
toward the theoretical minimum, Alice need only use Schumacher coding to compress her
C systems before teleporting them to Bob, who then applies the inverse decoding operation

to re-expand the teleported B systems to their original dimensions.

Because entropy of entanglement can be thus efficiently concentrated and diluted, its
positive-definiteness and the nonincrease of its expecation under local operations represent
the only limitation on the asymptotic ability of two separated observers to prepare a final

pure state Ψ′AB given an inital one, ΨAB , by local operations and classical communication.
Thus if Ψ and Ψ′ have equal entanglement, they can be interconverted with efficiency ap-
proaching unity in the limit of large n; if they have unequal entanglement, the asymptotic
yield is E(Ψ′ )/E(Ψ). On the other hand, if the initial state is unentangled, the positive-

definiteness of entanglement and the nonincrease of its expectation under local operations
together imply that there is no way of locally preparing an entangled final state Ψ′ , even
with low yield.
We have argued that entropy of entanglement is a good entanglement measure for pure

states because local operations cannot increase its expectation, but can, with asympotically
perfect efficiency, interconvert states of equal entropy of entanglement. No similarly simple
entanglement measure is known for mixed states. For a mixed state, the subsystem entropies
HA and HB need not be equal, and neither they nor any other simple function of HA , HB ,
and the total entropy HAB can be clearly identified with the degree of entanglement. For

example, entropies of HA = HB = HAB = 2 can be realized by two very different mixed

states:

• two independent pairs of classically correlated spins, or

• one singlet and a pair of uncorrelated spins.

The former mixed state is naturally regarded as unentangled, since it can be made by mixing
unentangled pure states; the latter should be regarded as entangled, since it can be converted
into a maximally-entangled state by discarding the second pair of spins.

16
 Two possible measures of entanglement for the mixed bipartite state ρAB , each of which
reduces to entropy of entanglement when ρAB is pure, are:

• “Entanglement of formation” defined as the least number of shared singlets asymptot-

ically required to prepare ρAB by local operations and classical communication.

• “Distillable entanglement” defined as the greatest number of pure singlets that can
asymptotically be prepared from ρAB by local operations and classical communication.

These entanglement measures have the desirable feature that their expectations cannot be
increased by local operations, but the disadvantage of being hard to evaluate in particular
cases because of the implied optimizations over local procedures.
Since any mixed state can be regarded as a pure state over a larger, and partly unseen,
Hilbert space, it might seem more elegant (cf [9]) to define the entanglement of a bipartite

mixed state ρAB as the minimum entanglement of any pure state Ψ(ABCD) of a four-
part system, from which ρAB can be obtained by tracing over states of C and D, and
where the entanglement of Ψ(ABCD) is computed by partitioning it into subsystems AC
and BD. Unfortunately, this prescription is unsatisfactory because it would assign positive

entanglement to classically correlated mixed states such as ρ = 21 (| ↑↑ih↑↑ |+| ↓↓ih↓↓ |), which
can be generated from unentangled initial conditions by local actions (random choice) and
exchange of classical messages.
A nontrivial example of an entangled mixed state is provided by the Werner state [10],

whose density matrix is

1 1  ED 
W = I +  Ψ− Ψ−  (11)

8 2

where I is the 4x4 identity matrix and Ψ− is a singlet. This state may be viewed as a

50/50 mixture of totally mixed states (density matrix 41 I) and singlets; and indeed, since
totally mixed states can be manufactured by purely local means, this description consti-
tutes a local probabilistic procedure for generating n Werner states from an expected n/2
singlets. Recently it has been shown [11] that the Werner state can be constructed much

17
more economically, as a mixture of pure states each containing only about 0.1176 ebits of en-
tanglement. Moreover, a small yield ≥ 0.0001258 of arbitrarily pure singlets can be distilled
from Werner states by local actions and classical communication [11]. Thus impure Werner

states can be both created from and converted into pure singlets. This two-way conversion
is a rough mixed-state analog of the processes of pure-state entanglement concentration and
dilution which have been the principal subject of the present paper. However, it is not known
whether the interconversion is asymptotically reversible for mixed states, as it is for pure

states. In other words, it is not known whether the Werner state’s distillable entanglement
is equal to, or far less than, its entanglement of formation.

ACKNOWLEDGMENTS

We wish to thank William Wootters for advice on Schmidt decompositions and many

other subjects, and John Smolin for pointing out how to use teleportation to prepare arbi-
trary entangled states.

18
 REFERENCES

[1] B. Schumacher, “On Quantum Coding,” Phys. Rev. A (to appear 1995).

[2] R. Jozsa and B. Schumacher, J. Modern Optics 41, 2343-2349 (1994).

[3] C. H. Bennett and S. J. Wiesner, Phys. Rev. Lett. 69, 2881 (1992).

[4] C. H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W. K. Wootters, Phys.

Rev. Lett. 70, 1895 (1993).

[5] A. Peres, Quantum Theory: Concepts and Methods (Kluwer, 1993) page 131-133.

[6] C.W. Helstrom, Quantum Detection and Estimation Theory (Academic Press, New-
York, 1976), pp.74-83.

[7] N. Gisin (private communication 1993) noted that POVMs can be simply implemented
in polarization space by using a polarization-dependent absorber.

[8] C. H. Bennett, G. Brassard, R. Jozsa, D. Mayers, A. Peres, B. Schumacher, and

W. K. Wootters, J. Mod. Optics 41, 2307-2314, (1994).

[9] R. Jozsa, J. Mod. Optics 41, 2315-2323 (1994).

[10] R. F. Werner, Phys. Rev. A 40, 4227 (1989), S. Popescu, Phys. Rev. Lett. 72, 797
(1994).

[11] C.H. Bennett, G. Brassard, S. Popescu, J. Smolin and W.K. Wootters, “Purification
of Noisy Entanglement and Faithful Teleportation via Noisy Channels” submitted to
Phys. Rev. Lett. (1995).

[12] G.Jaeger, A. Shimony, and L. Vaidman, Phys. Rev. A 51, 54-67 (1995).

19
 FIGURE CAPTION

1. Yield of maximally-entangled output states from partly-entangled input states

(cos θ ↑↓ − sin θ ↓↑), as a function of cos2 θ. Highest curve is entropy of entangle-

ment of input state, equal to the asymptotic yield of the Schmidt projection method.
Successively lower smooth curves give yields of Schmidt projection applied to n = 32,
8, 4, and 2 input pairs. The Λ-shaped curve gives yield by the Procrustean method

applied to one input pair.

20
 1.0 1

0.8
0.8

0.6
0.6

E
0.4
0.4

0.2
0.2

0.0 0.2 0.4 0.6 0.8 1

0.0 0.2 0.4 0.6 0.8 1.0

cos 2 θ