Journal of Modern Optics

ISSN: 0950-0340 (Print) 1362-3044 (Online) Journal homepage: https://www.tandfonline.com/loi/tmop20

A comparison of entanglement measures

Jens Eisert & Martin B. Plenio

To cite this article: Jens Eisert & Martin B. Plenio (1999) A comparison of entanglement
measures, Journal of Modern Optics, 46:1, 145-154, DOI: 10.1080/09500349908231260

To link to this article: https://doi.org/10.1080/09500349908231260

Published online: 03 Jul 2009.

Submit your article to this journal

Article views: 469

View related articles

Citing articles: 48 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tmop20
JOURNAL OF MODERN OPTICS, 1999, VOL.46, NO. 1, 145-154

A comparison of entanglement measures

JENS EISERT
Institut fur Physik, Universitat Potsdam, Germany

M A R T I N B. P L E N I O t
Blackett Laboratory, Imperial College, Prince Consort Road,
London SW72B2, U K

(Received 2 July 1998)

Abstract. We compare the entanglement of formation with a measure

defined as the modulus of the negative eigenvalue of the partial transpose. In
particular we investigate whether both measures give the same ordering of
density operators with respect to the amount of entanglement.

1. Introduction
Entanglement is a key property that makes quantum information theory
different from its classical counterpart [l]. Maximally entangled states, for ex-
ample, are the basis of quantum state teleportation which has recently been
demonstrated [2-4]. Under realistic conditions, however, one will only be able
to generate partially entangled mixed states. It is then of interest to be able
quantify the amount of entanglement in such states. Some measures of entangle-
ment for mixed states have been suggested recently [5-81. They are useful, for
example, as upper bounds to entanglement purification protocols [9-131 and to the
quantum channel capacity of certain quantum communication channels [5, 8, 141.
Unfortunately, entanglement measures for mixed states (which are relevant in
the presence of noise) are usually quite hard to calculate analytically, although an
analytic expression for the entanglement of formation of two spin-1 12 particles is
now known [15]. T h e general case remains unsolved. For some problems,
however, it is not so important to know the exact amount of entanglement (a
quantity that is not unique anyway). It would be completely sufficient if one knew
which state of a family of states has the most entanglement. T o answer this
question it would be sufficient to find a (hopefully as simple as possible) quantity
that preserves the ordering of density operators with respect to entanglement, i.e.
that for two measures El and E2 and any two density operators p1 and p2 we have
that El(p1) > (<)El(pz) is equivalent to Ez(p1) > (<)E2(p2).
In this paper we will compare the entanglement of formation for two spin-1/2
particles, for which a closed analytical form is known, with a quantity which was
proposed in [16] as a way to quantify the degree of entanglement of a mixed state.
The basis of this ‘measure of entanglement’ is the Peres-Horodecki criterion for
the separability of bipartite systems [16, 171. Given a state of, for example, two

t Author for correspondence; email: m.plenio@ic.ac.uk.

0950-0340/99 812.00 0 1999 Taylor & Francis Ltd.
146 J . Eisert and M . B. Plenio

spin-1/2 systems, one calculates the partial transpose of the density operator. The
state is separable exactly if the partial transpose is again a positive operator. If,
however, one of the eigenvalues of the partial transpose is negative then the state is
entangled. One can now imagine that the amount of entanglement is quantified by
the modulus of this negative eigenvalue, i.e. the larger it is, the larger the
entanglement of the state.
It is important to check whether this measure indeed preserves the ordering of
density operators as it has been used for this purpose in some publications.
In Section 2 we will explain the entanglement of formation and subsequently
summarize some properties of the negative eigenvalue measure of entanglement.
Finally, we present in Section 3 both a numerical and analytical comparison of the
two measures of entanglement with respect to the ordering of density operators
induced by them. In Section 4 we sum up the results of this paper.

2. Entanglement measures
In this section we briefly describe some entanglement measures, in particular
the entanglement of formation and negative eigenvalue measure of entanglement.
There are not very many good measures of entanglement. One example is the
relative entropy of entanglement [6-81. In fact, it gives rise to the most restrictive
upper bound on the channel capacity of the depolarizing channel [8, 141.
Unfortunately, even for two spin-1/2 particles no general analytical expression
has been found for it so far, although many special cases can be solved analytically.
A second measure of entanglement and actually the first measure of entangle-
ment that has been proposed for mixed states is the entanglement of formation [S].
It basically describes the amount of entanglement that needs to be shared
previously in order to be able to create a particular ensemble in a given state p
by local operations. Mathematically, this means that we find that pure state
ensemble that realizes the state p and which has the smallest amount of entangle-
ment, i.e.

where { I + i ) } is a set of not necessarily orthogonal pure states and E v ~ ( l + ) ( + Ithe

)
entropy of entanglement. The entanglement of formation is known to be larger
than the relative entropy of entanglement which proves that, in general, quantum
state purification methods cannot recover all the entanglement that has been
invested in the creation of the quantum state [8].
A nice feature of the entanglement of formation is the fact that it can be solved
analytically for a system of two spin-1/2 particles [15]. This allows for fast
numerical studies as the cumbersome minimization equation (1) can be avoided.
The entanglement of formation can be expressed in terms of the function

E(C)=h(
1 + (1 - c2)”2
2 )
where
h ( x ) = -xlog2 x - (1 - x) log2 (1 - x). (3)
For a density operator p one defines the spin-flipped state
 Comparison of entanglement measures 147

i j = (PY @ PY>P*(PY @ P Y ) (4)

where the * denotes the complex conjugate in the standard computational basis
loo), [ O l ) , [lo), 111). One then finds the entanglement of formation to be
EF(p) = E ( C ( p ) ) (5)
where the so-called concurrence is defined by
C ( p )= max (0, A1 - A2 - A3 - A4}. (6)
Here, X I ,. . . ,A4 are the eigenvalues, in decreasing order, of the Hermitian matrix
R =( fi$fi)”2. For properties of this measure of entanglement the reader should
consult the literature [5, 151. It is interesting to note that since € as a function of
the concurrence C is a strictly monotonous function and maps the interval [0,1] on
[0,1] C can in fact also be regarded as a measure for entanglement.
We will now consider the negative eigenvalue of the partial transpose of a
density operator as a measure of entanglement. In the next section we will then
compare it to the entanglement of formation.
For two spin-1/2 particles (which form the two systems A and B) any
disentangled state p can be written as the convex sum of product states

States which permit a representation of the form equation (7) are also called
separable. For two spin-1/2 particles there is a simple criterion to decide whether a
given state is separable or not [16, 171. One calculates the partial transpose pTU of
the density operator p in the computational basis. This means that we transpose
only one subsystem, either subsystem A or B. If the resulting matrix is positive
semidefinite then the density operator p is separable; otherwise it is not. Therefore
the partial transpose of an entangled state has a negative eigenvalue and the idea of
the negative eigenvalue measure is to use the modulus of the negative eigenvalue to
quantify the entanglement of the state p. In a mathematical form this reads as

where the are the eigenvalues of the partial transpose p T B . However, we do not
know whether this way of quantifying the entanglement constitutes a proper
measure of entanglement. Therefore we will investigate in the next section
numerically whether the entanglement of formation and the negative eigenvalue
measure are compatible from a different point of view.

3. Ordering induced by entanglement measures

We would expect that any two ‘good’ entanglement measures should generate
the same ordering of the density operators. This means that for two entanglement
measures El and E2,and any two density operators p1 and p2 we have that

Why do we expect this relation to be true? If in one measure of entanglement p1

contains more entanglement than p2 then we would expect that a quantum state
purification method would generate more singlets from an ensemble in state p1
than ensemble p2. If E2 is also a measure of entanglement then we would expect
148 J . Eisert and M . B. Plenio

that p2 would yield more singlets than p i . While this reasoning is not strict, it
nevertheless indicates that equation (9) should be true for two ‘good’ measures of
entanglement.
In the next two subsections we will now check whether equation (9) is satisfied
for the entanglement of formation EF and the negative eigenvalue measure of
entanglement E N .

3 .l.A n analytical comparison

For some classes of density operators we can easily check analytically whether
relation equation (9) is true. For pure states it is sufficient to consider states of the
+
form I$) = (~100) pI11) as it follows from the Schmidt decomposition [18]. T h e
entanglement of formation EF then reduces to the von Neumann entropy of
entanglement while the negative eigenvalue measure yields

E ~ ( l $ ) ( $ l )= ( ~ =p~ r ( -l a2)ll2. (10)

Both measures decrease monotonically with a so that for pure states equation (9) is
satisfied. It should be mentioned that for pure states I$)($[ with
+
I$) = (~100) PI11) the concurrence C is given by
C(l$)($l) = 2 4 1 - ( Y 2 p 2 (11)
hence, also for arbitrary pure states p = /$)($I the negative eigenvalue measure and
the concurrence are connected by the simple equation
c(p)= 2EN ( P ) . (12)
Werner states are defined as

where I$-) = (101) - [lo))/& is the singlet state and F E [ 1 / 4 , 1 ] .T h e concur-

rence of Werner states is given by C ( ~ F=) 2F - 1 for F > 1 / 2 (otherwise p~ is a
separable state) while the entanglement of formation is found to be
) -~1og2 P - ( 1 - P ) log2 ( 1 - P )
E F ( P F= (14)
with p = 1 / 2 + [F(1- F)]’’2.T h e negative eigenvalue measure simply gives
EN(PF) = F - i, (15)
which means that equation (12) is also valid for Werner states. Again, both
measures decrease monotonically with F and therefore equation (9) is satisfied.
Moreover, no counterexample to equation (9) can be constructed from pure
states and Werner states, i.e. no two density operators pi and p2 can be found
which violate equation (9) when pi is assumed to be a pure state and p2
corresponds to a Werner state. Because of the monotony of & in equation ( 2 )
this can already be seen when comparing the concurrences of pi and p2 with the
negative eigenvalue measure of the respective states: from equation ( 1 2 ) it then
immediately follows that
c ( p l > > c(p2)A EN(P1) < EN(p2) (16)
cannot be fulfilled for any value of F .
 Comparison of entanglement measures 149

One may therefore suspect that equation (9) also holds for arbitrary mixed
states. T o consider the question whether equation (9) is generally satisfied and
hence the two entanglement measures induce the same ordering, we have em-
ployed Monte Carlo simulations. If equation (9) is violated at all, it is-roughly
speaking-also of interest to investigate how ‘badly’ it is violated after all. In the
following subsection we present the results of the numerical test that we have
performed.

3 . 2 . A numerical comparison
In order to test the ordering induced by the two entanglement measures under
consideration we have numerically generated a million pairs of random density
matrices with respect to a certain well-defined distribution. In the case of two spin-
112 particles, which is the case of interest in this paper, one has to generate random
density operators, i.e. linear, self adjoint, positive semidefinite operators p of finite
trace ( = l ) , acting in the four-dimensional Hilbert space isomorphic to C2 8 C2.
One can represent p by the decomposition
4

with pairwise orthogonal projections, where p ; 2 0 and cip; 1 . This in turn

=
corresponds to a 4 x 4 matrix of the form

p = UDUtl (18)
where the diagonal matrix D is defined by DQ = 6 ~ p i .
A plausible choice for the ensemble of unitary 4 x 4 matrices which can be used
to construct density matrices according to equation (18) is the one with the
normalized Haar measure on the group of unitary matrices U(4) which is called
the circular unitary ensemble [19]. It has been shown in [20] that random unitary
matrices representative for the circular unitary ensemble can be obtained as
follows: Let us set

where the complex 4 x 4 matrices U ( i i j )i, j = 1 . . . , 4 , with three real parameters

4, x, and 1c, are given by
k=l,k#i,j
sin 4e’X, k = i , l =j ,

I i s i n 4e-’X1 k =j , l = i,
otherwise.
150 J . Eisert and M . B. Plenio

Figure 1. AE& p2) versus AE,(pl, p2) for lo4 pairs of entangled states (p1, p2). Pairs
of states with Al?~(p,pz)AEN(pi, pz) < 0 are represented by dots in the second and
fourth quadrant.

We now take $+j and xi to be independent random variables with a uniform

distribution in the interval [0,27c), and for a given random number J distributed
uniformly in the interval [0, 1) we choose 46 to be arcsin (J1’(2i)) for i = 1,. .. 3.
Matrices generated in this way are then random unitary matrices of the wanted
type (except for a random phase which is of no significance for our purposes).
Finally, a set of random density matrices can be obtained by appropriately
choosing the diagonal matrices D . More precisely, the random vector with entries
p i , . . . ,$4 should be uniformly distributed on the manifold defined by xipi = 1.
According to [ 2 1 ] this can be achieved, e.g. by setting pl = 1 - Ji’31p2 =
(1 - J i ’ 2 ) ( 1 -PI), P3 = (1 -&)(I -pi - h ) , and P4 = 1 - P I - p 2 -p3, where
again, (1, J 2 , and J3 are random numbers drawn with respect to a uniform
distribution in the interval [0, 1 ) .
To investigate whether equation (9) is satisfied we now draw random density
matrices from the previously described ensemble and calculate the eigenvalues of
the partial transposes and the eigenvalues of the products of the respective density
operators with the spin flipped states. After discarding those density matrices
which correspond to separable states we then determine values for EN and EF as
described in Section 2 . Finally, we check for pairs ( p i , p2) whether the sign of

and

is identical. Figure 1 shows a diagram in which A E ~ ( p t , p 2 )is plotted versus

A E ~ ( p 1 , p 2 for
) lo4 pairs ( p 1 , p z ) of random density matrices. Although there is
obviously a certain correlation between AEF and AEN there are dots in the second
and the fourth quadrant of the diagram which are associated with pairs of states
which do not satisfy equation (9). One can therefore conclude that in the case of
 Comparison of entanglement measures 151

EN

EF
Figure 2. Distribution of states: EN versus E F .

two spin-1/2 systems this relation is not satisfied for arbitrary mixed states, and
hence, the ordering induced by the entanglement of formation EF and the one
induced by the negative eigenvalue measure EN is not the same.
It is interesting to note that most of the randomly drawn density operators do
not correspond to entangled, but to separable states; so more density matrices have
to be discarded than can be kept for the investigation. In fact, from the numerical
simulations we can estimate the probability PE that a created mixed state is
entangled as PE x 0.365 f 0.001, which is in complete agreement with the
analytical and numerical findings presented in [21].
From the Monte Carlo simulation we can also give a rather accurate estimate of
how likely it is to find a pair of states which violates equation (9), given that both
states are entangled. From the relative frequency of a violation for a million pairs
of density matrices we obtain from the estimate of this probability Pv

Pv x 0.047 f 0.001. (23)

Now that we know that the ordering is different it is interesting to see to what
extent the values of EN(^) and E F ( ~differ ) for a given state p, or how these random
states are distributed in a diagram where EF is plotted versus E N . Figure 2 shows
such a plot. Since the ordering induced by EF and EN is not the same, there is no
strictly monotonous functionf : [0,1] + [ 0 , 1 / 2 ]such that EN(^) = f ( E ~ ( p )for
) all
states p. This is also obvious from figure 2 .
In figure 3 again the distribution of states is shown, but this time the negative
eigenvalue measure is plotted versus the concurrence C. We can see that most of
the dots are located close to the diagonal connecting (0,O) and ( 1 / 2 , 1 ) ; this
diagonal corresponds to states satisfying equation ( 1 2 ) . Note that the numerical
simulation also strongly suggests that for a state p with a certain value of C(p) the
upper bound for the possible values of the negative eigenvalue measure EN(^) is
given by C ( p ) / 2 ,that is, that in general C(p) 2 2 E ~ ( pholds
) for any state p.
Furthermore, since for pure states the ordering induced by the two entangle-
ment measures is the same but in general it is not, it is of interest to investigate the
dependence of the probability that equation (9) is violated for a pair of entangled
states ( P I , pz) on the 'purity of those states' with respect to a certain characteriza-
152 J . Eisert and M . B. Plenio

EN

C
Figure 3. Same as figure 2 with EN versus C.

S
Figure 4. Estimate for the probability density of finding a pair of states ( p , , p 2 ) with a
certain value of S = S +
1 &(+). The solid line is a cubic spline interpolation.
Apparently, above a certain cut-off point there are no longer any pairs of entangled
states. The histogram shows the relative number of pairs violating equation (9)
compared to the total number of pairs of entangled states for a given value of S.
Note that the scaling is different for both plots: the right axis belongs to the estimate
for the probability density (+), the left one to the relative number of violating pairs
(histogram).

tion of the purity. In figure 4 we show the relative number of pairs (p1, p2) of states
+
which do not satisfy equation (9) versus S = S1 S2 based on a million pairs of
entangled states. Here, S1 and S2 are the linear entropies

of p1 and p2, respectively, as commonly employed, e.g. in decoherence studies [22].

Obviously, pure states correspond to a vanishing linear entropy.
We observe that the fraction of pairs ( p l , p 2 ) of states with A E ~ ( p 1 , p z )
A E ~ ( p 1 , p z<
) 0 increases monotonically with the sum of the linear entropies of
the respective states. This indicates that the more mixed the two states are with
respect to the arithmetic mean of their linear entropies, the larger is the probability
that this pair violates equation (9). Above a certain cut-off no pairs of entangled
 Comparison of entanglement measures 153

states can be found at all-the numerical data shown in figure 4 give an estimate for
the probability density of finding a pair of entangled states with a certain value of S
(compare also [21]).
It should finally be mentioned that in reference [21] yet another measure of
entanglement incorporating the partial transpose is proposed: for a given state p
the quantity E = ‘&IXYl - 1 is taken as a measure, where again, the X
p denote
the eigenvalues of the partial transpose pTB of p. While the actual value of E for a
given state is of course different from that of E N , from numerical simulations we
come to the same conclusion that the ordering induced by this measure is different
from the one induced by the entanglement of formation E F .

4. Summary
We have compared the entanglement of formation with a potential measure of
entanglement that is given by the negative eigenvalue of the partial transpose of the
density operator of the system. In particular the ordering of density operators with
respect to the amount of entanglement induced by the two measures has been
compared both numerically and analytically. We have shown that the negative
eigenvalue measure does not induce the same ordering as the entanglement of
formation. Therefore we do not expect it to be a ‘good’ measure of entanglement.
In particular it cannot, in general, be used to determine the most entangled state
for a given family of density operators as it has been used previously.

Acknowledgments
The authors would like to thank Peter Knight, Vlatko Vedral and Martin
Wilkens for discussions and useful hints. This work was supported in part by the
EPSRC, the European T M R Research Network ERBFMRXCT960066 and the
European T M R Research Network ERBFMRXCT960087.

References
[l] PLENIO, M. B., and VEDRAL, V., 1998, Contemporary Phys., to appear, also available as
Ian1 e-print quant-ph/9804075.
[2] BENNETT,C. H., BRASSARD, G., CREPEAU, C., JOZSA,R., PERES,A., and WOOTTERS,
W. K., 1993, Phys. Rev. Lett., 70, 1895.
[3] BOSCHI, D., BRANCA, S., DEMARTINI, F., HARDY, L., and POPESCU, S., 1998, Phys. Rev.
Lett., 80, 1121.
[4] BOUWMEESTER, D., PAN,J. W., MATTLE, K., EIBL,M., WEINFURTER, H., and ZEILINGER,
A., 1997, Nature, 390, 575.
[5] BENNETT, C. H., DIVINCENZO, D. P., SMOLIN, J. A., and WOO-ITERS, W. K., 1996, Phys.
Rev. A, 54, 3824.
[6] VEDRAL,V., PLENIO, M. B., RIPPIN, M. A,, and KNIGHT, P. L., 1997, Phys. Rev. Lett.,
78, 2275.
[7] VEDRAL, V., PLENIO, M. B., JACOBS, K., and KNIGHT, P. L., 1997, Phys. Rev. A, 56,
4452.
[8] VEDRAL, V., and PLENIO, M. B., 1998, Phys. Rev. A, 57, 1619.
[9] BENNETT, C. H., BERNSTEIN, H. J., POPESCU, S., and SCHUMACHER, B., 1996, Phys. Rev.
A, 53, 2046.
[lo] BENNETT, C. H., BRASSARD, G., POPESCU, S., SCHUMACHER, B., SMOLIN, J. A., and
WOOTTERS, W. K., 1996, Phys. Rev. Lett., 76, 722.
[ l l ] GISIN,N., 1996, Phys. Lett., 210, 151.
154 Comparison of entanglement measures

[12] HORODECKI, M., HORODECKI, P., and HORODECKI, R., 1997, Phys. Rev. Lett., 78, 574.
[13] MURAO, M., PLENIO, M. B., POPESCU, S., VEDRAL,V., and KNIGHT, P. L., 1998, Phys.
Rev. A, 57, R4075.
[14] RAINS,E., lanl e-print quant-ph/9707002,
[15] WOOWERS, W. K., 1998, Phys. Rev. Lett., 80, 2245.
[16] HORODECKI, M., HORODECKI, P., and HORODECKI, R., 1996, Phys. Lett A, 223, 1.
[17] PERES, A., 1996, Phys. Rev. Lett., 77, 1413.
[18] SCHMIDT, E., 1907, Math. Ann., 63, 433.
[19] For a recent analysis of measures on the set of mixed states see SLATER, P. B., Ian1 e-
print quant-ph/9806039.
[20] ZYCZKOWSKI, K., and KuS, M., 1994, J. Phys. A, 27, 4235; POZNIAK, M., ZYCZKOWSKI,
K., and KuS, M., 1998, J . Phys. A, 31, 1059.
[21] ZYCZKOWSKI, K., HORODECKI, P., SANPERA, A., and LEWENSTEIN, M., lanl e-print
quantph/9804024.
[22] ZUREK, W. H., HABIB, S., and PAZ,J. P., 1993, Phys. Rev. Lett., 70, 1187.