PHYSICA L R EVIEW LET T ERS week ending
VOLUME 90, N UMBER 5 7 FEBRUARY 2003
Classical Correlations and Entanglement in Quantum Measurements
V. Vedral
Optics Section, Blackett Laboratory, Imperial College, Prince Consort Road, London SW7 2BZ
(Received 1 August 2002; published 3 February 2003)
We analyze a quantum measurement where the apparatus is initially in a mixed state. We show that
the amount of information gained in a measurement is not equal to the amount of entanglement between
the system and the apparatus, but is instead equal to the degree of classical correlations between the two.
As a consequence, we derive an uncertainty-like expression relating the information gain in the
measurement and the initial mixedness of the apparatus. Final entanglement between the environment
and the apparatus is also shown to be relevant for the efficiency of the measurement.
DOI: 10.1103/PhysRevLett.90.050401 PACS numbers: 03.65.Ta, 03.67.–a
Any measurement can be modeled as an establishment In this Letter, we analyze a quantum measurement
of correlations between two random variables: one ran- when the apparatus is ‘‘fuzzy,’’ i.e., it is initially in a
dom variable represents the values of the quantity per- mixed state. Our approach is entropic in character and
taining to the system to be measured, while the other is therefore closest in spirit to that of Lindblad [4]. We
random variable represents the states of the apparatus show that the amount of information gained via the
used to measure the system [1]. It is by looking at the apparatus is proportional to the classical correlations
states of the apparatus, and discriminating them, that we between the systems and the apparatus, rather than the
infer the states of the system. Looking at the apparatus, of amount of entanglement between them. We then derive an
course, is another measurement process itself, which uncertainty-like expression which says that the sum of
correlates our mental states (presumably another random the information gained in the measurement and the mix-
variable) with those of the apparatus, so that indirectly edness of the apparatus (quantified by the Von Neumann
we become correlated with the system as well. It is at this entropy [3]) is bounded from the above by logN, where N
point that we can say that we have gained a certain is the dimension of the apparatus. Our analysis builds on
amount of information about the system. This description recent results in quantum information theory concerning
of the measurement process is true both in classical and quantification of entanglement in bi- [5,6] and tripartite
quantum physics. (Note that in this way there is no more systems [7] and separating classical and quantum corre-
mystery in the ‘‘quantum state collapse’’ than there is in lations [8]. Quantum information theory has mainly been
the corresponding classical measurement.) The differ- developed to understand computation and communication
ence between the two lies in the way we represent states supported by quantum systems, but this knowledge can
of systems and the way we represent their mutual inter- now be applied back to quantum mechanics to study its
action and evolution. Classically, physical states of an foundations from a new perspective.
n-dimensional system are vectors in a real n dimensional We first review the existing measures of entangled and
vector space whose elements are various occupational total correlations [9]. In classical information
P theory, the
probabilities for the states. The evolution of a classical Shannon entropy, HX Hp i pi logpi , is used
system is in general some stochastic map acting on this to quantify the information in a random variable, X, that
vector space. Quantum mechanically, on the other contains states xi with probabilities pi [10]. In the quan-
hand, states are in general represented using density tum context, the results of a projective measurement fEy g
matrices, while the evolution is a completely positive, on a state represented by a density matrix,
, comprise a
trace preserving transformation acting on these matrices. probability distribution py TrEy
. Von Neumann
Using this representation, classical physics becomes a showed that the lowest entropy of any of these probability
limiting case of quantum mechanics when the density distributions generated from the state
was achieved by
matrices are strictly diagonal in one and the same fixed the probability distribution composed of the eigenvalues
basis and the completely positive map then becomes the of the state, f i g [3]. This probability distribution
stochastic map. Because of this fact, it is enough to would arise from a projective measurement onto the
analyze properties of quantum systems and quantum state’s eigenvectors. The Von Neumann entropy is then
evolutions and all the results are automatically applicable given by S
Tr
log
H . The quantum
to classical physics when we restrict ourselves to the relative entropy of a state
with respect to another
diagonal density operators only. A comprehensive survey state
is defined as S
jj
S
Tr
log
.
of major papers on quantum measurement can be found in The joint entropy S
AB for a composite system
[2] and the first fully quantum analysis was due to
AB with two subsystems A and B is given by
Von Neumann [3]. S
AB Tr
AB log
AB and theVon Neumann mutual
050401-1 0031-9007=03=90(5)=050401(4)$20.00 2003 The American Physical Society 050401-1
PHYSICA L R EVIEW LET T ERS week ending
VOLUME 90, N UMBER 5 7 FEBRUARY 2003
information between the two subsystems is defined as projective measurement onto the states fjiig of subsystem
I
AB S
A S
B S
AB . The mutual informa- A. Therefore the classicalP correlations are given by
tion is the relative entropy between
AB and
A
B , and CA
AB S
B i pi S
iB , which is, for this state,
is used to measure the total correlations between the two equal to the mutual information I
AB . This is to be
subsystems. The entanglement of a bipartite quantum expected since there are no entangled correlations and
state
AB may be measured by how distinguishable it is so the total correlations between A and B should be equal
from the ‘‘nearest’’ separable state. Relative entropy of to the classical correlations. This measure of classical
entanglement, ERE
AB min
AB 2D S
AB jj
AB , has correlations has other important properties such as
been shown to be a useful measure of entanglement (D C
AB 0 if and only if
AB
A
B ; it is also
is the set of all separable or disentangled states) [5]. Note invariant under local unitary transformations and non-
that ERE
AB I
AB , by definition of ERE
AB , since increasing under any general local operations [8].
the mutual information is also the relative entropy be- Let us now introduce the general framework for a
tween
AB and a completely disentangled state. There are quantum measurement (for a special P case, see [14]). We
other ways of measuring the entanglement of a bipartite have a system in the state P j i i ai jii, and an appara-
quantum state [9], but they can all be unified under the tus in the state
i ri jri ihri j in the eigenbasis. The
formalism of relative entropy [6]. An advantage of relative purpose of a measurement is to correlate the system with
entropy is that it can be generalized to any number of the apparatus so that we can extract the information about
subsystems, a property that will be useful in understand- the state jji of the system. In a perfect measurement, by
ing the measurement process when the environment is looking at the apparatus we can unambiguously identify
also present. The relative entropy will be used exclusively the state of the system. Therefore, when the system is in
throughout to quantify entanglement. We stress that the state jji we would like the apparatus to be in the state
all the measures used here are entropic in nature, which
j , such that
i
j 0, i.e., different states of the appa-
means that they are generally attainable only asymptoti- ratus lie in orthogonal subspaces and can be discrimi-
cally. The advantage of using entropic measures is that nated with a unit efficiency. If this condition is not
our results will be universally valid, although they fulfilled, which is frequently the case, then the measure-
will almost always be overestimates in the finite-case ment is imperfect and the amount of information
scenario. obtained is not maximal (this is what defines an ‘‘imper-
Recently we have suggested that correlations in a state fect measurement’’). We now compute the amount of
AB can also be split into two parts, the quantum and the information gained in general and show that it is more
classical part [8] (see also [11,12] for alternative ap- appropriately identified with the classical rather than
proaches). The classical part is seen as the amount of quantum correlations between the system and the appa-
information about one subsystem, say A, that can be ratus. Suppose that the measurement transformation
obtained by performing a measurement on the other is given by a unitary operator, U, acting on both the
subsystem, B. The resulting measure is the difference system and the apparatus, such that U
jiihjjUy
between the initial and P the residual entropy [8]:
ij
jiihjj, where we assume that the measurement trans-
CB
AB maxBy Bi S
A i pi S
iA , where Byi Bi is a formation acts such that the state jrk ijli of the apparatus
i
positive operator valued measure performed on the sub- and the system, respectively, is transformed into the state
system B and
iA trB Bi
AB Byi =trAB Bi
AB Byi is the j~r kl ijli, such that the states of the apparatus correspond-
remaining state of A after obtaining the outcome
P i on B. ing to different system states are orthogonal h~r ij j~r ik i
Alternatively, CA
AB maxAy Ai S
B i pi S
iB if jk . This particular interaction is chosen so that in
i
the measurement is performed on subsystem A instead of the special case of the pure apparatus we obtain
on B. Clearly CA
AB CB
AB for all states
AB such Von Neumman’s (and Everett’s) analysis. We see that the
that S
A S
B (e.g., pure states). This measure is a measurement is such that the new apparatus state de-
natural generalization of the classical mutual informa- pends on the state of the system. This is exactly how
tion, which is the difference in uncertainty about the correlations between the two are established. P Then, the
subsystem B (A) before and after a measurement on the initial state is transformed into
a a
P P f ij i j ij
correlated subsystem A (B). Note the similarity of jiihjj i jai j2
ii
jiihij ij ai aj
ij
jiihjj. The
the definition to the Holevo bound which measures the first term on the right-hand side indicates how much
capacity of quantum states for classical communication information this measurement carries. We will now
[13]. The following example provides an illustration of measure the apparatus and try to distinguish the states
this and will be the key to our discussion of the quantum
ii to the best of our ability. Once we confirm that the
measurement.PConsider a bipartite separable state of the apparatus is in the state
jj , then we can infer that the
form
AB i pi jiihijA
iB , where fjiig are orthonor- system is in the state jji. The amount of information about
mal states of subsystem A. Clearly the entanglement of the state of the apparatus (and hence the state of the
this state is zero. The best measurement that Alice can system), Im , is given by the well-known Holevo
make to gain information about Bob’s subsystem is a bound [13]:
050401-2 050401-2
PHYSICA L R EVIEW LET T ERS week ending
VOLUME 90, N UMBER 5 7 FEBRUARY 2003
X X
2 initial
P state of the system in our above scenario was
Im S jai j
ii jai j2 S
ii : (1)
i i
a jii,
Pi i 2 while the final state is a mixture of the form
i jai j jiihij. The disturbance to the state can be mea-
As we have seen, this quantity is also equal to the amount sured as a distance between the final and the initial state.
of classical correlations between
P the system and the ap- We choose the relative entropy to quantify this difference.
paratus in the state
0f i jai j2
ii
jiihij, which is, in So, while the information in the measurement P is given by
this case, the same as the Von Neumann mutual informa- Im ,Pthe disturbance is D Sj ih jjj i jai j2 jiihij
tion between the two. Note that this state is only classi- i jai j2 logjai j2 , which is the same as the maximum
cally correlated and there is no entanglement involved. amount of information possible from this measurement.
The amount of entanglement in the state
f , on the other So, the measurement described above always maximally
hand, will in general be nonzero. This may be difficult to disturbs the state, and the reason why this does not lead to
calculate. However, we can provide lower and upper the maximum information gain is because the apparatus
bounds. The lower bound on the entanglement between state is mixed. The system could be disturbed less by
the system and the apparatus is adjusting the overlap between the states of the apparatus
X
j~r ij i, so that they are not orthogonal to each other. In
E
f S jai j2
ii S
f general we can require that h~r ij j~r ik i ajk , such that
i
X jajk j < 1. We will not treat this case here: it is mathe-
S jai j2
ii S
Im : (2) matically more demanding, but does not illuminate the
i measurement issue any better. Note also that a question
may be raised as to why we consider the interaction
[Note that here S
S
ii for all i by definition of
between the apparatus and system to be unitary and not
measurement interaction.] Therefore, the entanglement
of a more general kind (a completely positive map as in,
between the system and the apparatus is larger than or
for example [9]). The reason is that any such interaction
equal to the classical correlations between the two which
can be represented by a unitary transformation [9] and
quantify the amount of information that the measurement
our analysis then also applies (although the resulting
carries. So, this shows that the information in a quantum
effective measurement would in general be less efficient
measurement is correctly identified with the classical
than the one performed unitarily).
correlations between the apparatus and the system rather
In order to show that some form of entanglement is still
than the entanglement or the mutual information between
important (albeit not the one between the system and the
the two in the final state,
f . Only in the limiting case of
apparatus) we revisit the same measurement scenario, but
the pure apparatus do we have that the amount of infor-
from the ‘‘higher Hilbert space perspective.’’ This is done
mation in the measurement is equal to the entanglement,
by adding the environment to the apparatus so that the
which becomes the same as the classical correlations,
joint state is pure, j EA i. We briefly note that our treat-
while the sum of the quantum and classical correlations
ment differs from the usual ‘‘environment induced col-
is then equal to the mutual information in the state.
lapse’’ and decoherence as in, for example, [15,16]. In our
We can recast this relationship in the form of an ‘‘un-
case, the environment is not there to cause the disappear-
certainty relation’’ between the initial mixedness of the
ance of entanglement between the system and the appa-
apparatus and the amount P of information gained. So,
2 ratus, but is there to purify the initally mixed state of the
from the fact that Im S i jai j
ii S
, we have
apparatus. The P measurement P transformation is now given
that
X by
P p j EA i
i ai jii
! i a i j i
EA ijii, where j EA i
2 i ri jei ijri i and jei i is an orthonormal basis for the
Im S
S jai j
ii logN; (3)
i
environmental states. We see that when the environment
is traced out, the state of the apparatus is equal to
. Now,
where N is the dimension of the apparatus. Thus we see the measurement implements a unitary transformation so
that the sum of the initial mixedness of the apparatus and that each of the states of the apparatus changes according
the amount of information the measurement obtains is to which state of the system it interacts with. Therefore we
always smaller than a given fixed value: the larger S
, see that the ith state of the environment and Pthe apparatus
p
the smaller Im . When
is maximally mixed [and there- after the interaction is given by j iEA i j rj jej ij~r ji i.
fore S
logN], then no information can be extracted To make a link with the first picture of the measure-
from the measurement. Note that this relation is different ment, wePtrace out the environment to obtain
A0 S0
P i j
to the usual ‘‘information versus disturbance’’ law in a ij ai aj k hek j EA ih EA jek i
jiihjj and, thus, the
quantum measurement as well as to the usual entropic quantity in brackets can be identified with
ij
P
uncertainty relations of incompatible observables. Every k rk j~
r ki ih~r kj j. Therefore, since we have no access to the
measurement that extracts information from a quantum environment, our task is to discriminate the states
ii ,
system also disturbs the state, and without this distur- and therefore identify the corresponding states jii of the
bance there would be no information gain possible. The system, and this was done in the previous analysis. If, on
050401-3 050401-3
PHYSICA L R EVIEW LET T ERS week ending
VOLUME 90, N UMBER 5 7 FEBRUARY 2003
the other hand, we had access to the environment, the fact, for the measurement to be efficient we wish the
measurement could be perfect. environment not to become entangled with the apparatus
We first apply entropic considerations to the to a large extent (after the measurement). We should
‘‘environment-apparatus-system’’ tripartite state. The mention at the end that our example is somewhat simpli-
initial and the final entropy of the environment are the fied in that the environment will not, in reality, be passive
same as its state remains unchanged, and this value is throughout the process. It would instead interact with
the same as the initial entropy of the apparatus, S
. As both the system and the apparatus making the measure-
we have seen, this is an important quantity, as it deter- ment even less effective, although all the above results
mines how much information can be extracted from a would still apply.
measurement: the more mixed the initial state of the In this Letter we have analyzed the information gained
apparatus, the less information can be extracted. If the in a quantum measurement when the apparatus used to
initial state is maximally mixed (say it is a thermal state extract this information is initially in a mixed state. We
with an arbitrarily high temperature), then there can be have shown that the amount of information is correctly
no information gain during the measurement. The initial identified with the amount of classical correlations be-
entropy of the apparatus is also equal to the entropy of tween the system and the apparatus after their correlation
the system and the apparatus after the measurement, is established and derived an entropic uncertainty relation
S
A0 S0 S
f , as well as the amount of entanglement between this amount and the mixedness of the initial
between the environment and the system and the appara- state. Further light on quantum measurement was then
tus together, EE:A0 S0 , after the measurement. The entan- shed by purifying the apparatus and including its own
glement and the mutual information between the environment in the analysis. Among open problems high-
environment and the apparatus after the measurement lighted by this work are to extend the analysis to non-
are always less than or equal to their value before the orthogonal states of the apparatus and to prove that the
measurement (since the systems becomes correlated to information gain is symmetric between the system and
the apparatus during the measurement). the apparatus.
We now use the recently derived three-party entangle- This work is funded by Engineering and Physical
ment bounds to provide further constraints on the mea- Sciences Research Council, the European grant EQUIP,
surement. For any pure tripartite states
ABC we have that and Hewlett-Packard.
[7]: maxfE
AB S
C ; E
AC S
B ; E
BC
S
A g E
ABC . Applying this to our measurement
scenario we obtain that S
EE:A0 :S0 EA0 :S0 , where
the subscripts E; A; S indicate the environment, the appa-
ratus, and the system, respectively. The primes on the [1] H. Everett, Rev. Mod. Phys. 29, 454 (1957).
subscripts indicate states after the measurement. We see [2] Quantum Theory and Measurement, edited by J. A.
that the closer the tripartite entanglement to the entan- Wheeler and W. H. Zurek, Princeton Series in Physics
glement between the system and the apparatus (with the (Princeton University Press, Princeton, New Jersey,
environment disentangled), the more efficient the mea- 1983).
[3] J. Von Neumann, Mathematical Foundations of Quantum
surement. We immediately conclude that the necessary
Mechanics (Princeton University Press, New York, 1955).
condition for the equality between the two entanglements [4] G. Lindblad, Commun. Math. Phys. 33, 305 (1973).
is that the initial entropy of the apparatus is zero. It should [5] V. Vedral and M. B. Plenio, Phys. Rev. A 57, 1619 (1998).
be remembered, however, that the measurement can still [6] L. Henderson and V. Vedral, Phys. Rev. Lett. 84, 2263
be perfect even though the apparatus is not pure and this is (2000).
because the relevant quantity is the classical correlations [7] M. B. Plenio and V. Vedral, J. Phys. A 34, 6997 (2001).
between the system and the apparatus and not their en- [8] L. Henderson and V. Vedral, J. Phys. A 34, 6899 (2001).
tanglement. In that context we can also derive from the [9] V. Vedral, Rev. Mod. Phys. 74, 297 (2002).
above inequality that EE:A0 S
S0 EE:A0 :S0 , so that [10] E. Shannon and W. Weaver, The Mathematical Theory of
Communication (University of Illinois Press, Urbana,
EE:A0 Im EE:A0 :S0 S
S
0A : Illinois, 1949).
[11] H. Ollivier and W. H. Zurek, Phys. Rev. Lett. 88, 017901
Thus, the sum of the information from the measurement (2002).
and the final entanglement between the environment and [12] B. M. Terhal et al., quant-ph/0202044.
the apparatus is limited by the final entropy of the appa- [13] A. S. Holevo, Prob. Peredachi Inf. 9, 177 (1973).
ratus and therefore by logN. Again we see that the larger [14] S. Bose and V. Vedral, quant-ph/0004016.
the information we want, the smaller the entanglement [15] W. H. Zurek, Phys. Rev. D 24, 1516 (1981).
with the environment and the apparatus will be. So, in [16] M. H. Partovi, Phys. Lett. A 137, 445 (1985).
050401-4 050401-4
