0% found this document useful (0 votes)
12 views17 pages

1223654528

This paper discusses the additivity conjecture in quantum information theory, which posits that certain functions of quantum states and channels are additive under tensor products. The conjecture has been proven to unify several previously separate conjectures, and despite substantial efforts, it remains unresolved, with recent progress indicating counterexamples to stronger conjectures. The author provides an exposition of the conjecture, its background, and methods used to approach it, while calling for further assistance in its resolution.

Uploaded by

Jeffry Immanuel
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
12 views17 pages

1223654528

This paper discusses the additivity conjecture in quantum information theory, which posits that certain functions of quantum states and channels are additive under tensor products. The conjecture has been proven to unify several previously separate conjectures, and despite substantial efforts, it remains unresolved, with recent progress indicating counterexamples to stronger conjectures. The author provides an exposition of the conjecture, its background, and methods used to approach it, while calling for further assistance in its resolution.

Uploaded by

Jeffry Immanuel
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 17

IP Current Developments in Mathematics, 2005

The Additivity Conjecture in Quantum


Information Theory
Peter W. Shor

Abstract. In this paper, I discuss the additivity conjecture in


quantum information theory. The additivity conjecture was orig-
inally a set of at least four conjectures. These conjectures said
that certain functions of quantum states and channels were addi-
tive under tensor products. While some of these conjectures were
previously known to be stronger than others, they have recently
all been proved equivalent. This conjecture is a very intriguing
mathematical question which the best efforts of a large number of
quantum information theorists have not been able to resolve for
nearly a decade. It is a mathematically elegant question that is
one of the most important open questions in the field of quantum
information and computation. This paper then is intended to be
both an exposition of the conjecture, its background, and some of
the methods that have been used to yield partial results for it, as
well as a plea for help in resolving this conjecture. Very recently
(summer 2007), substantial progress has been made on this conjec-
ture, in that counterexamples to a set of stronger conjectures have
been found. These will be described briefly.

1. Introduction
Earlier in this decade, the title of this paper would have been am-
biguous. Until 2003, there was not any one problem which could have
been titled “The additivity conjecture.” Rather, there were a host of
different additivity questions, which we will describe in more detail later
in this manuscript. Two major open additivity questions were the ad-
ditivity of classical channel capacity and the additivity of entanglement
of formation, but researchers had also conjectured the additivity of
the minimum entropy output of a quantum channel, and conjectured
a stronger property than additivity of entanglement of formation called
the strong superadditivity of entanglement of formation. In 2003, it was
shown that these four open conjectures were all equivalent [29, 33].
c 2007 International Press

173
174 PETER W. SHOR

This unified four seemingly separate conjectures into one unified con-
jecture. Holevo [17, 18] has written recent, thorough surveys of the
state of affairs of the additivity question. There, however, has been
significant progress since then [37, 13, 10] as will be discussed later
in this paper. Several other quantities had earlier been conjectured to
be additive, including the relative entropy of entanglement, the distill-
able entanglement. These now both appear to be non-additive; the first
has been rigorously proved so [35], and for the second there is strong
evidence of non-additivity [34]
The question of the additivity of a quantity f , which is a function
defined over quantum channels (or quantum states) can be expressed as
follows. Suppose Φ and Ψ are two quantum channels (or states):

Is f (Φ ⊗ Ψ) = f (Φ) + f (Ψ)?

where Φ ⊗ Ψ is the tensor product of Φ and Ψ. In general, and certainly


for all these quantities mentioned above, inequality in one direction is
trivial, while the other may be quite difficult.

2. Background
Before we can describe this in more detail, we need to give some
definitions. We will be dealing solely with the quantum mechanics of
finite dimensional systems in this paper. This is because it is easier
to deal with finite dimensions than with infinite dimensions and also
because , at least for the channel capacity question, the additivity con-
jecture for infinite dimensional systems is implied by the conjecture for
finite-dimensional systems [31].
In quantum mechanics, a d-dimensional system can be represented
by a complex vector space Cd . A pure quantum state of the system is
a unit complex vector in this vector space. Quantum states are often
represented as kets, | vi, although in this paper we often omit the kets,
yielding a notation which is more familiar to mathematicians. We still
use | 0i, | 1i, . . ., for basis states of a quantum system. We will be using
the generally accepted convention in physics of representing quantum
states by column vectors.
Two pure quantum states are equivalent if they differ only by a
phase, so mathematically, a quantum state lies in a projective complex
vector space. In many calculations, however, it is much more convenient
to represent a state as a unit complex vector, and we will be doing this.
A mixed quantum state, or density matrix is a trace 1 positive1 d × d
matrix. A pure state v is equivalent to the rank one density matrix vv † .
If we have a finite ensemble of quantum states, where the probability of

1In this paper, by positive, I will mean positive semidefinite Hermitian matrix.
THE ADDITIVITY CONJECTURE IN QUANTUM INFORMATION THEORY 175

vi is pi , then the density matrix associated with this ensemble is


pi vi vi†
X

One example of quantum states with which many people are familiar
is the polarization states of light. This is a two-dimensional quantum
system (which quantum information scientists have started calling a
qubit) and thus has two basis vectors. For example, we can take the
vertically and horizontally polarized photons a basis. We call these
states
| li and | ↔i
Any other pure state of polarization is a linear combination of these
states. For example,
1
(1) |ր
ւi = √ (| li + | ↔i)
2
1
(2) |ց
տi = √ (| li + | ↔i)
2
and right and left circular polarization states are the linear combinations
with imaginary coefficients √12 (| li ± i | ↔i).
When you consider two quantum systems, their joint state space is
represented by the tensor product of the two individual state spaces.
This joint state space includes states which are not tensor products.
These states are said to be entangled. For example, the state
1
√ (| li ⊗ | ↔i − | ↔i ⊗ | li)
2
is such a state, this one called an EPR pair (after Einstein, Podolsky
and Rosen’s famous thought experiment [7], which involved a similar
physical state of two entangled particles).
Density matrices arise from pure quantum states in two ways. The
first, as was described above, is if we have a quantum system about
which we have incomplete information. If this quantum system is in
state vi with probability pi , then the density matrix is
pi vi vi†
X
ρ=
i

The second way is if we have an entangled system, and we discard or


ignore one part. To describe this mathematically, we need to introduce
the partial trace operator. Suppose we have a density matrix ρ over a
tensor product vector A ⊗ B. If we discard system B, this corresponds
to taking the partial trace over B of the state ρ, written as TrB ρ. The
partial trace for tensor product states ρA ⊗ ρB is defined as
TrB (ρA ⊗ ρB ) = ρA TrρB
176 PETER W. SHOR

and it is defined for arbitrary states ρ by linearly extending this defini-


tion.
Any two orthogonal quantum states are distinguishable, meaning
that there is some theoretical measurement which distinguishes per-
fectly between these two states. If two quantum states are not orthog-
onal, then no such measurements exist, although many measurements
may give some probability of distinguishing between these two states.
The most general quantum measurement is called a POVM (positive
operator valued measurement). In this paper, we will restrict consider-
ation to POVM’s with a finite number of outcomes. Such POVM’s can
be represented by a set of positive Hermitian matrices, E1 , E2 , . . ., Ek ,
where each matrix is associated with a different outcome of the mea-
surement. The Pone constraint on these matrices is that they sum to the
identity, i.e., i Ei = I. Given a pure quantum state v, the probability
of outcome i of the measurement is
Prob(i) = v † Ei v.
For a mixed quantum state ρ, the probability of outcome i is
Prob(i) = TrρEi .
The constraint that their sum is the identity is equivalent to the con-
straint that the sum of the probabilities of the outcomes of a measure-
ment is 1.
A valid map on quantum states allowed by physics is a trace pre-
serving completely positive (TPC) map Φ, taking density matrices to
density matrices. Here, trace preserving means
TrΦ(ρ) = Trρ,
positive means that Φ maps positive matrices to positive matrices, and
completely positive means that Φ ⊗ Id is positive for the identity map Id
on a d-dimensional quantum system, for all d. There are two alternative
characterizations of trace preserving completely positive maps. The
first is the Krauss operator sum representation. Any trace-preserving
completely positive operator can be written as
Ak ρA†k , where
X X †
Φ(ρ) = Ak Ak = I.
k k
The second is the Stinespring representation. This says that any d-
dimensional TPC map can be represented as the composition of the fol-
lowing three operations. First, taking the tensor product of this system
with a fixed state v0 in a Hilbert space of dimension at most din · dout .
Second, making a unitary transform U on the resulting state. Finally,
taking the partial trace to leave a dout dimensional subsystem. That
is, Φ can be represented as the composition of operations:
ρ → ρ ⊗ v0 v0† → U (ρ ⊗ v0 v0† )U † → Tr2 U (ρ ⊗ v0 v0† )U † .
THE ADDITIVITY CONJECTURE IN QUANTUM INFORMATION THEORY 177

3. The maximum purity conjectures


The minimun entropy output of a quantum channel Φ is defined as
Hmin (Φ) = min H(Φ(ρ)).
ρ

Because of the convexity of the entropy function, and the linearity of Φ,


it is immediate that this minimum is achieved on the boundary of the
set of density matrices, i.e., on a rank-one density matrix ρ = vv † .
The additivity question in this case is whether
Hmin (Φ ⊗ Ψ) = Hmin (Φ) + Hmin (Ψ)
The ≤ is easy to prove. Let the states ρ0 and ρ1 be those where Φ and
Ψ achieve their minimum entropy. Then
H( (Φ ⊗ Ψ)(ρ0 ⊗ ρ1 )) = H(Φ(ρ0 )) + H(Ψ(ρ1 )).
The minimum entropy output seems like it may be the simplest
of the various equivalent additivity conjectures to prove. Indeed, the
proofs that the other conjectures imply additivity of Hmin are all quite
easy, while the proofs in the other direction are substantially harder.
Amosov, Holevo and Werner [1] have come up with a very natural con-
jecture on p-norms of channels that would imply the minimum entropy
output conjecture. This is called the output purity conjecture. This is
a means by which substantial progress has been made on the minimum
entropy output conjecture for special classes of channels. Very recently,
this conjecture has been shown to be false [37, 13]. However, the coun-
terexample does not appear to extend to give a counterexample to the
additivity conjecture for entropy.
We now describe the output purity conjecture. We first define a
p-norm on a quantum channel Φ.
νp (Φ) = max ||Φ(ρ)||p
ρ

where the optimization is over density matrices, i.e., trace 1 positive


matrices, over the input space of the channel. Here || · ||p is the p-norm
1
||σ||p = (Tr σ p ) p . Again, because of convexity, the ρ yielding maximum
νp will always be a pure state for any channel Φ. This is related to the
minimum output entropy because of the following theorem, shown in
[1]
lim νp (Φ) = Hmin (Φ)
p↓1
which can be proved using the fact that
x − xp
lim = −x log x.
p↓1 p − 1

This theorem implies that if the following conjecture is true for 1 < p <
1 + ǫ for some positive ǫ, the minimum entropy output is additive
(3) νp (Φ ⊗ Ψ) = νp (Φ)νp (Ψ).
178 PETER W. SHOR

This was realized by Amosov, Holevo and Werner, who then conjec-
tured that this minimum output purity was multiplicative for all p [1].
This is equivalent to the additivity of minimum output Rényi entropy
of order p. Unfortunately, their original conjecture was not true; two of
the proposers of this conjecture found a counterexample not long after
their original paper [36]. The counterexample shows that the conjec-
ture is false for p ≥ 4.8. There did not seem to be any obstacle to it
holding for 1 < p ≤ 2, which was the revised version of the conjecture.
However, this has recently also shown to be false. Winter [37] gave a
counterexample which showed that the conjecture was false for p > 2.
The structure of this counterexample led Hayden [13] to find a coun-
terexample for all p, 1 < p < 2. Shortly afterwards, Harrow [10] found
a counterexample for the additivity of Rényi entropy at p sufficiently
close to 0, which numerically seems to work for 0 ≤ p < 0.12. (This is
equivalent to multiplicativity of minimum output p-norm.) Thus, if the
conjecture holds, it most likely holds only for the von Neumann entropy
case of p = 1.
We now describe the structure of Winter’s counterexample [37],
which holds for 1 < p < 2. We take m random unitary transformations
over an n-dimensional Hilbert space, U1 , . . ., Um . Now, let
m
1 X †
Φ:ρ→ Um ρUm
m
j=1

and
m
1 X †
Φ∗ : ρ → Um ρUm .
m
j=1

It can be shown that if m grows faster than O(n log n), then Φ(vv † ) and
Φ∗ (vv † ) are close to random for all v, so for Φ and Φ∗ , all the eigenvalues
of any output density matrix are of order 1/n. However, if we consider
the EPR state
1 X
w=√ | ii | ii ,
n
i
then Φ ⊗ Φ∗ (ww† ) will have one large eigenvalue. This is because for
any for any unitary U ,
1 X X
w† (U ⊗ U † )w = ( hj | ⊗ hj |) U | ii ⊗ U † | ii)
n
j i
1X
= hj | U | ii ⊗ hj | U † | ii
n
i,j
1X
= |hj | U | ii|2
n
i,j
= 1,
THE ADDITIVITY CONJECTURE IN QUANTUM INFORMATION THEORY 179

where the last inequality holds because U | ii and | ji are two different
orthonomal bases for the Hilbert space.
The above equality shows that the terms where i = j in the expres-
sion
 
 
1
Ui ⊗ Uj† ww† Ui† ⊗ Uj  w
X
w† Φ ⊗ Φ∗ (ww† ) w = w†  2
m
i,j

contribute a total of 1/m. Since the other terms are positive, this shows
that (Φ ⊗ Φ∗ )(ww† ) has an eigenvalue larger than 1/m. If we choose
m = n1+ǫ for ǫ small, this will give a counterexample to the maximum
purity conjecture for any p > 2, since the maximum p-norms of each
of Φ and Φ∗ are approximately n(1−p)/p , and the maximum p-norm of
Φ ⊗ Φ∗ is approximately 1/m.
Inspired by this result, Hayden [13] found a counterexample to the
conjecture for any p > 1. This is not quite so easy to understand intu-
itively, and the calculations are somewhat more difficult, but the basic
idea behind the counterexample is the same. Here, the two channels are
Φ(ρ) = TrB U (ρ ⊗ v0 v0† )U †
and
Φ∗ (ρ) = TrB U † (ρ ⊗ v0 v0† )U,
where U is a random unitary transform.

4. The other equivalent conjectures


In this section, we will describe several other conjectures equivalent
to the additivity of Hmin There are at least six other equivalent con-
jectures [33, 29, 27, 9], of which we will mention four. These deal
with two quatitites. The first of these quantities is the entanglement of
formation of a quantum state, and the second of these is the classical
information capacity of a quantum channel.
A pure quantum state in a tensor product space A ⊗ B is said to be
entangled if it cannot be written as a tensor product φ ⊗ ψ where φ ∈ A
and ψ ∈ B. Entanglement is a very important concept in quantum in-
formation theory. It is essentially the source of Einstein’s distress over
what he called “spooky action at a distance,” and is also responsible
for Bell’s inequality, and protocols for quantum information transmis-
sion such as quantum teleportation, quantum superdense coding, and
so forth. It is thus an important question as to how to quantify entan-
glement.
The canonical example of an entangled state is an EPR pair. This
is a system of two qubits in the state
1
ψEP R = √ (| 01i − | 10i),
2
180 PETER W. SHOR

or any state that can be derived from this state by local unitary trans-
formations. That is, if we have two unitary transformations on the two
Hilbert spaces UA acting on A and UB acting on B, then (UA ⊗UB )ψEP R
is also an EPR pair. By convention, we take the amount of entanglement
in a single EPR pair to be one bit.
It is fairly easy to quantify entanglement of pure quantum states
(recall these are vectors in our tensor product Hilbert space). The
amount of entanglement in a pure quantum state v consists of
EP (v) = H(TrA vv † ) = H(TrB vv † ).
It is a theorem [6] that any two pure quantum states with the same
amount of entanglement can be asymptotically interconverted with high
efficiency, using only local unitary operations and classical communica-
tion. That is, suppose we have a pure quantum state v. We denote the
quantum state of an EPR pair, ψEP R above, by ψ. Then for any fixed
small ǫ > 0, there is a large enough n so that, if two people, Alice and
Bob, hold A and B respectively, then Alice and Bob, using only local
quantum operations and classical communication, can convert v ⊗n to a
state ρ that is very close to ψ ⊗(1−ǫ)nEP , i.e., such that
⊗(1−ǫ)nEP
ψ† ρ ψ ⊗(1−ǫ)nEP > 1 − ǫ
where EP is the pure state entanglement of v. We also have the reverse
theorem, where we interchange the role of v and ψ above, so that we
start with nEP EPR pairs: ψ ⊗nEP and we end with a state ρ that is
very close to v ⊗(1−ǫ)n .
We call a transformation that can be performed using local quan-
tum operations and classical communication a LOCC operation. In this
model, we are allowed to perform local quantum operations conditioned
on classical variables, i.e., on the outcome of measurements or on clas-
sical information that Alice has received from Bob (or vice versa). This
is thus a class of transformations that is quite difficult to characterize.
When we try to prove analogous results for mixed states, that is,
we try to quantify the entanglement in a mixed state ρ, something
unfortunate happens. There is no longer a single good measure of en-
tanglement, as there was with pure states. We can look at the number
of nearly perfect EPR pairs we can obtain from many copies of a mixed
state ρ, analogous to the theorem we stated for pure state entanglement
above. This gives a quantity is known as distillable entanglement, ED .
We can also define the amount of pure entanglement asymptotically re-
quired to create a mixed state ρ using LOCC operations. This gives a
quantity called entanglement cost, known as EC . In general, these are
not the same for mixed states. There are bound entangled states ρ with
ED (ρ) = 0, so no states close to an EPR pair can be created out of
any number of copies of these states, but still with the entanglement
cost EC (ρ) > 0 [20]. We do not have any nice formulas for distillable
THE ADDITIVITY CONJECTURE IN QUANTUM INFORMATION THEORY 181

entanglement ED , but there is a formula for entanglement cost EC . Let


us define entanglement of formation by
X
EF (ρ) = min pi EP (vi )
{pi ,vi }
i
where the minimum is taken over ensembles of pure states whose density
matrix is ρ, that is
pi vi vi† = ρ.
X

i
It is then known [14] that
1
EC (ρ) = lim EF (ρ⊗n ).
n
n→∞
The question of whether entanglement of formation is equal to en-
tanglement cost, that is, whether we really need the limit in the above
expression, is equivalent to the question of whether entanglement of
formation is additive, that is, whether for two density matrices ρ1 and
ρ2 ,
EF (ρ1 ⊗ ρ2 )) = EF (ρ1 ) + EF (ρ2 ).
The structure of the states may need a little more explanation here.
We have four Hilbert spaces, A1 , A2 , B1 and B2 , and we work in the
tensor product A1 ⊗ A2 ⊗ B1 ⊗ B2 . We have ρ1 ∈ M(A1 ⊗ B1 ), ρ2 ∈
M(A2 ⊗ B2 ). The definition of the entanglement of formation assumes
that the quantum space is a tensor product of two systems. In each case,
this tensor product defining the entanglement of formation is between
the “A” and the “B” compenents of the system.
There is a stronger conjecture about entanglement of formation than
the additivity. It is called the strong superadditivity of entanglement of
formation [35, 3]. In this case, we have an arbitrary entangled state ρ
in a Hilbert space which is the tensor product of four systems, which
we again designate as A1 , B1 , A2 , B2 . The conjecture is
EF (ρ) ≤ EF (Tr2 ρ) + EF (Tr1 ρ)
where the entanglement of formation is defined over the A-B split of
subsystems, and the partial traces are over the 1-2 split of subsystems.
Strong superadditivity of EF is easily seen to imply additivity of EF
[35], whereas it is harder to prove the reverse direction.
We now turn to the definition of the classical capacity of a quantum
channel. Recall the definition of the capacity of a classical noisy channel.
A classical noisy channel can be thought of as a stochastic map from
a random variable X representing the input of a channel to a random
variable Y representing the output of the channel. Shannon’s theorem
says that a noisy channel N has the capacity
max I(X; Y ) = H(Y ) − H(Y |X)
p(x)
182 PETER W. SHOR

where I(X; Y ) is the mutual information between the ramdom variables


X and Y . We have H(Y ) is the entropy of the random variable Y , and
H(Y |X) is the conditional entropy of the random variable Y . This is
defined as X
H(Y |X) = Prob(X = x)H(Y |X = x)
x
where H(Y |X = x) is the entropy of the output Y when the input is x.
One can think of this definition of the mutual information of a channel
as the entropy of the average output minus the average entropy of the
output.
One might think that the classical capacity of a quantum channel Φ
would be the maximum over all input distributions X of quantum states,
and over all measurements of the output, of the mutual information
between random variable X representing the input and the random
variable Y representing the measurement outcome. This quantity is
called the accessible information for a quantum channel Φ, which we
denote by IA (Φ). This guess is not correct; while one can clearly achieve
the accessible information, it is not an upper bound. The accessible
information can be exceeded by using joint measurements of the output
state across several uses of a quantum channel. To get the true classical
capacity of a quantum, one must take
1
lim ( IA Φ⊗n ),
n→∞ n
1
the limit of n times the accessible information of the tensor power of n
uses of the channel. This formula allows both entangled inputs to the
channel and joint measurements of the channel. To obtain this limit,
one can first take the limit for the joint measurements, and then the
limit for the entangled inputs. If we allow joint measurements, but only
unentangled inputs, one can achieve the following capacity, which was
formerly called the Holevo bound but is now often called the Holevo
capacity or Holevo formula. This was long known merely as an upper
bound for the accessible information [15]. It is now known to be the
capacity as long as the protocols are not allowed to use entangled inputs
[16, 30].
pi vi vi† )) − pi H(Φ(vi vi† ))
X X
χ(Φ) = max H(Φ(
{pi ,vi }
i i
where the maximum is taken over all probability distributions pi on pure
states vi in the input space of the channel. This is again the entropy of
the average output minus the average entropy of the output.
Recall that the classical capacity of a quantum channel is the limit
of the accessible information over protocols using both entangled inputs
and joint measurements on the output. Taking the limit just over joint
measurements of the output gives us the Holevo capacity. To get the
true classical capacity, we need to then take the limit of the Holevo
THE ADDITIVITY CONJECTURE IN QUANTUM INFORMATION THEORY 183

capacity over n uses of the channel, as n goes to ∞. We thus get the


formula
1
Capacity(Φ) = lim χ(Φ⊗n )
n→∞ n
If the Holevo capacity is additive, then this limit is the same as the
Holevo capacity. This, of course, leads us to the last conjecture, whether
χ(Φ ⊗ Ψ) = χ(Φ) + χ(Ψ)

5. Partial results on the output purity and minimum entropy


conjectures.
In this section, I will try to sketch the techniques used to prove
partial results on these conjectures. There are several cases for which
these conjectures are proved. The most imporant of these may be unital
qubit channels [21], depolarizing channels [8, 22], and entanglement
breaking channels [23, 32]. The proofs here are often fairly complicated,
and have a lot of details which may obscure the essential features of these
proofs.
An entanglement breaking channel is one which breaks the entan-
glement of any input state with a reference system. That is, Φ is an
entanglement breaking channel
I ⊗ Φ(vv † )
never has entanglement between the output space of I and that of Φ. A
unital qubit channel Φ is one which takes a 2-dimensional quantum space
to a 2-dimensional quantum space, and for which Φ(I) = I. Finally, a
depolarizing channel mixes a quantum state ρ with the maximally mixed
state I/d. That is, the depolarizing channel with parameter λ, which
we call ∆λ , takes
∆λ (ρ) = λρ + (1 − λ)I/d
where d is the dimension of the input (and output) space of the channel.
Two of the key tools for these proofs are inequalities. When dealing
with von Neumann entropy, one of the most important tools is the strong
subadditivity of entropy [24]. This says that for a quantum system in
a mixed state in a tensor product of three systems, ρ = ρABC , we have
H(ρAB ) + H(ρAC ) ≥ H(ρA ) + H(ρABC )
where the density matrices for subsystems are obtained by taking the
partial trace of ρ, that is, ρAB = TrC ρ, ρAC = TrB ρ, and ρA = TrBC ρ.
The other inequality, which has been useful for proving theorems
involving output purity, is the Lieb-Thirring inequality2 [25, 21], which
says that for two matrices,
Tr(A1/2 BA1/2 )p ≤ TrAp/2 B p Ap/2
2There are actually two inequalities, quite different, which are both known as
the Lieb-Thirring inequality. This is one of them.
184 PETER W. SHOR

for p ≥ 1. These inequalities do not appear to be related at first sight,


but there may be some connection between them, as the additivity of
entropy for entanglement breaking channels was proved using strong
subadditivity [32], while the output purity conjecture for p-norms in
entanglement breaking channels was proved by the Lieb-Thirring in-
equality [23].
I will now try to sketch the proofs of two of these theorems. First,
we show the additivity entropy for entanglement breaking channels.
We consider two channels. Let Ψ be an arbitrary channel acting on
space A, and Φ be an entanglement breaking channel acting on space B.
Now, consider a state ρAB . We know (by the definition of entanglement
breaking channel) that I ⊗ Φ(ρAB is separable, so we have
qj vj vj† ⊗ wj wj†
X
I ⊗ Φ(ρAB ) =
j

Here, the vj are in system A and the wj are in system B. We let σAB
be
q − jΨ(vj vj† ) ⊗ wj wj†
X
Ψ ⊗ Φ(ρAB ) =
j
We now add a third Hilbert space, C, to the mix, and work in the
space A ⊗ B ⊗ C. We define
qj Ψ(vj vj† ) ⊗ wj wj† ⊗ ej e†j
X
σABC =
where the ej are orthonormal vectors in system C.
We now have a tensor product of three systems, and can apply the
strong subadditivity of entropy to σABC
H(σAB ) ≥ H(σABC ) − H(σBC ) + H(σB ).
But we find
σB = Φ(ρB ) ≥ Smin (Φ)
and
qj S(Φ(vj vj† ) ≥ Smin (Ψ)
X
H(σABC ) − H(σBC ) =
j
which proves the theorem.
Note that this does not automatically prove the additivity of channel
capacity if one of the two channels is an entanglement breaking chan-
nel; the proof that additivity of minimum entropy implies additivity of
channel capacity invokes the additivity of Smin for a different channel.
However, a more complicated calculation also proves the addivity of
channel capacity if one channel is entanglement breaking.
The proof of multiplicativity of νp for depolarizing channels is more
complicated. The key tool is the Lieb-Thirring inequality [25]. Recall
that the depolarizing channel in d dimensions is
∆λ (ρ) = λρ + (1 − λ)I/d
THE ADDITIVITY CONJECTURE IN QUANTUM INFORMATION THEORY 185

The proof uses the dephasing channel. Suppose that ei , i = 1 . . . d is


an orthonomal basis for our vector space/ The dephasing channel (with
dephasing parameter λ) is
d
ei e†i ρei e†i
X
Φλ (ρ) = λρ + (1 − λ)
i=1

It is fairly easy to express the depolarizing channel ∆λ as a convex


combination of dephasing channels (each having a different set of eigen-
vectors ei ). For the proof, we need a specific combination of dephasing
channels. If our depolarizing channel is acting on a state ρ with or-
thonormal eigenvectors vj , we can express ∆λ as a convex combination
of dephasing channels, each of which has eigenvectors ei complemen-
tary to vi . That is, |e†i vi |2 = d1 . This lets us prove results on dephasing
channels, and then apply them to the depolarizing channel.
The result that is proved goes as follows.
Lemma: For a dephasing channel,
" d #1/p
(i)
X
(4) ||(Φλ ⊗ I)(ρAB )||p ≤ d1−1/p νp (∆λ ) Tr(ρB )p
i=1

where
(i)
ρB = TrA e†i ei ⊗ IρAB .
The clever step in the proof of the lemma is to factor (Φλ ⊗ I)(ρAB )
1/2 1/2
into matrices M1 M2 M1 so that we can apply the Lieb-Thirring in-
equality. We then apply this lemma to the state
ρAB = (I ⊗ Ψ)(σAB )
to get multiplicativity. The last term in equation 4 together with d−1/p
gives the bound on νp (Ψ), provided we select the proper decomposition
of the depolarizing channel into dephasing channels. The second to last
term gives the bound on νp (Φ) = νp (∆λ ), and the rest, combined with
the details of the decomposition of the depolarizing channel ∆λ into
dephasing channels, result in the desired inequality.

6. Equivalence of the additivity conjectures


The equivalence of the additivity conjectures is shown by using
several techniques, including one called channel extension. We will
sketchily describe one case of this technique; for more details, see the
papers [33, 19, 9]. What we will describe is the additivity of something
called the constrained Holevo capacity. This is the problem of maximiz-
ing the Holevo capacity, when the average input is constrained to be
a given density matrix ρ. More formally, we would like to maximize
186 PETER W. SHOR

over all probability


P distributions pi over density matrices ρi in the input
space with i pi ρi . That is, for a channel Ψ we would like to find
X
χρ (Ψ) = max H(Ψ(ρ)) − pi H(Ψ(ρi ))
i
X
subject to pi ρi = ρ.
i
Although we do not explain it in this paper, it is fairly straightforward
that the Stinespring dilation theorem shows that the additivity of con-
strained Holevo capacity is equivalent to the entanglement of formation.
Since ρ is fixed in the above problem, we can just minimize the
second term on the right-hand side, namely
pi H(Ψ(vi vi† ))
X
(5) minimize
i
subject to the same constraint on the average input. This minimization
problem is a linear program in the pi , and as such, it has a dual linear
program which has the same optimum value. This program is
maximize Tr τ ρ

(6) subject to v † τ v ≤ H(Ψ(vv † ))


where the maximization is over Hermitian matrices τ , and the con-
straint (6) holds for all for all vectors v in the input space. The linear
programming duality theorem says that for states vi which have non-
zero probability in the optimum for probability distribution for Eq. (5),
the dual constraint Eq. (6) holds with equality. We will call such states
signal states.
Now, suppose we could find a new channel Φ such that for any vector
v in the input space,
(7) H(Φ(vv † )) = H(Ψ(vv † )) + C − v † τ v.
Then, for signal states vi ,
H(Φ(vi vi† )) = H(Ψ(vi vi† )) + C − vi† τ vi = C.
holds with equality, while for other states v,
H(Φ(vv † )) ≤ C.
Thus, the signal states for this channel are exactly the minimum entropy
output states.
Now, suppose we had two channels Ψ1 and Ψ2 , and we could find a
Φ1 and Φ2 as described above. A calculation then shows the additivity
of minimum entropy output for Φ1 and Φ2 would imply the additivity
of constrained Holevo capacity for Ψ1 and Ψ2 .
Unfortunately, we cannot actually construct a channel Φ which satis-
fies Eq. (7) above. However, we can find a channel that nearly satisfies
THE ADDITIVITY CONJECTURE IN QUANTUM INFORMATION THEORY 187

this. In fact, this is very closely related to the original channel. We


choose some POVM with two elements E, I − E, which we will deter-
mine later. Now, we flip a coin which comes up heads with probability
q. If it is heads, we send the input through the original channel Ψ. If
it is tails, we measure the input with the POVM, and output either k
completely random bits or a specified pure state. We also send an indi-
cator state which tells which of the three above possibilities occured. It
is straightforward to show that for an input state v, the output of this
channel has entropy

H(Φ(vv † )) = qH(Ψ(vv † )) + C + (1 − q)kTrEρ + (1 − q)H2 (T rEρ)

where H2 (x) = −x log x−(1−x) log(1−x) is the binary entropy function,


and C is the constant H2 (q). By letting k go to ∞, and adjusting q and
E appropriately, we can approach the desired channel extension (7)
arbitrarily closely, and this construction lets us prove that additivity
of mininimum entropy output implies additivity of constrained Holevo
capacity and thus additivity of entanglement of formation.
The technique of creating the new channel Φ from Ψ is known as
channel extension, and this technique, applied somewhat differently, is
an important component in several of the other reductions contained in
the proof of equivalence.

References
[1] G.G. Amosov, A.S. Holevo, R.F. Werner, On some additivity problems in quan-
tum information theory, Problems in information transmission 36 25–34 (2000);
arXiv e-print math-ph/0003002.
[2] S. Arimoto, An algorithm for calculating the capacity of an arbitrary discrete
memoryless channel, IEEE Trans. Info. Theory 18 14–20 (1972).
[3] K.M.R. Audenaert and S.L. Braunstein, On strong superadditivity of the entan-
glement of formation, Comm. Math. Phys. 246 443–452 (2004); arXiv e-print
quant-ph/0303045.
[4] K. Audenaert, F. Verstraete, B. De Moor, Variational characterizations of sep-
arability and entanglement of formation, Phys. Rev. A 64, art. 052304, 2001;
arXiv e-print quant-ph/0006128.
[5] H. Barnum, J.A. Smolin, and B.M. Terhal, Quantum capacity is properly defined
without encodings, Phys. Rev. A 58 3496–3501 (1998).
[6] C.H. Bennett, H.J. Bernstein, S. Popescu, and B. Schumacher, Concentrating
partial entanglement by local operations, Phys. Rev. A 53 2046–2052 (1996).
[7] A. Einstein, B. Podolsky, and N. Rosen, Can quantum-mechanical description
of physical reality be considered complete? Phys. Rev. 47 777–780 (1935).
[8] Additivity of the capacity of depolarizing channels, Physics Letters A 299 469–
475 (2002).
[9] M. Fukuda, Simplification of additivity conjecture in quantum information
theory, Quantum Information Proc. 6 179–186 (2007); arXiv e-print quant-
ph/0608010.
[10] A. Harrow, Counterexamples to additivity of minimum output p-Renyi entropy
for p close to 0, talk at MIT, Sept. 10, 2007, manuscript in preparation.
188 PETER W. SHOR

[11] P. Hausladen, R. Jozsa, B. Schumacher, M. Westmoreland, and W.K. Wootters,


Classical information capacity of a quantum channel, Phys. Rev. A 54 1869–1876
(1996).
[12] M. Hayashi and H. Nagaoka, General formulas for capacity of classical-quantum
channels, IEEE Trans. Inf. Theory 49 1753–1768 (2003); arXiv e-print quant-
ph/0206186.
[13] P. Hayden, The maximal p-norm multiplicativity conjecture is false, arXiv e-print
quant-ph/0707.3291.
[14] P.M. Hayden, M. Horodecki, and B.M. Terhal, The asymptotic entanglement
cost of preparing a quantum state, J. Phys. A: Math. Gen. 34 6892–6898 (2001).
[15] A.S. Holevo, Information theoretical aspects of quantum measurements, Probl.
Info. Transm. (USSR), 9(2) 31–42 (1973) (in Russian); [translation: A.S.
Kholevo, Probl. Info. Transm., 9 177–183 (1973)].
[16] A.S. Holevo, The capacity of the quantum channel with general signal states,
IEEE Trans. Info. Theory 44 269–273 (1998).
[17] A.S. Holevo, The additivity problem in quantum information theory, in ‘Pro-
ceedings of the International Congress of Mathematicians’, Madrid, Spain, 2006,
999–1018, EMS, Zurich, 2007.
[18] A.S. Holevo, Multiplicativity of p-norms of completely positive maps and the
additivity problem in quantum information theory, Uspekhi Matematicheskikh
Nauk 61(2) 113–152 (2006); English version: Russian Math Surveys 61(2) 301–
339 (2006).
[19] A.S. Holevo and M.E. Shirokov, On Shor’s channel extension and constrained
channels, Comm. Math. Phys. 249 417–430 (2004).
[20] M. Horodecki, P. Horodecki, and R. Horodecki, Mixed-state entanglement and
distillation: Is there “bound entanglement” in nature?, Phys. Rev. Lett. 80 5239–
5242 (1998).
[21] Additivity for unital qubit channels, Journal of Math. Phys. 43 4641–4653 (2002).
[22] C. King, The capacity of the quantum depolarizing channel, IEEE Trans. Inform.
Theory 49 221–229 (2003); arXiv e-print quant-ph/0204172.
[23] C. King, Maximal p-norms of entanglement breaking channels, Quantum Inf.
Comput. 3 186–190 (2003); arXiv e-print quant-ph/0212057.
[24] E.H. Lieb and M.B. Ruskai, Proof of the strong subadditivity of quantum me-
chanical entropy, J. Math. Phys. 14 1938–1941 (1973).
[25] E.H. Lieb and W. Thirring, Inequalities for the moments of the eigenvalues
of the Schrödinger Hamiltonian and their relation to Sobolev inequalities, in
‘Studies in Mathematical Physics’, E. Lieb, B. Simon, A Wieghtman eds., 269-
303, Princeton University Press, 1976.
[26] K. Matsumoto, T. Shimono, and A. Winter, Remarks on additivity of the Holevo
channel capacity and of the entanglement of formation, Comm. Math. Phys. 246
427–442 (2004); arXiv e-print quant-ph/0206148.
[27] K. Matsumoto, Yet another additivity conjecture, arXiv e-print quant-ph/
0506052.
[28] S. Osawa and H. Nagaoka, Numerical experiments on the capacity of quantum
channel with entangled input states, in ‘IEICE Trans. Fundamentals of Electron-
ics, Communications and Computer Sciences’, E84-A, 2583-2590 (2001); arXiv
e-print quant-ph/0007115.
[29] A.A. Pomeransky, Strong superadditivity of the entanglement of formation fol-
lows from its additivity, Phys. Rev. A 68 032317 (2003).
[30] B. Schumacher and Westmoreland, Sending classical information via a noisy
quantum channel, Phys. Rev. A 56 131–138 (1997).
THE ADDITIVITY CONJECTURE IN QUANTUM INFORMATION THEORY 189

[31] M.E. Shirokov, The Holevo capacity of infinite dimensional channels and the ad-
ditivity problem, Comm. Math. Phys. 262 137–159 (2006); arXiv e-print quant-
ph/0408009.
[32] P.W. Shor, Additivity of the classical capacity of entanglement-breaking channels,
J. Math. Physics 43 4334–4340 (2002).
[33] P.W. Shor, Equivalence of additivity questions in quantum information theory,
Comm. Math. Phys. 246 453–472 (2004).
[34] P.W. Shor, J.A. Smolin, and B.M. Terhal, Nonadditivity of bipartite distillable
entanglement follows from a conjecture on bound entangled Werner states, Phys.
Rev. Lett. 86 2681-2684 (2001).
[35] K.G.H. Vollbrecht and R.F. Werner, Entanglement measures under symmetry,
Phys. Rev. A 64 062307 (2001).
[36] R.F. Werner and A.S. Holevo, Counterexample to an additivity conjecture for
output purity of quantum channels, quant-ph/0203003.
[37] A. Winter, The maximum output p-norm of quantum channels is not multiplica-
tive for any p > 2, arXiv e-print quant-ph/0707.0402 (2007).

MIT, Cambridge, MA 02139


E-mail address: [email protected]

You might also like