Sunteți pe pagina 1din 15

A Domain Theoretic Model of Qubit Channels

Keye Martin
Naval Research Laboratory
Center for High Assurance Computer Systems
Washington, DC 20375
keye.martin@nrl.navy.mil

Abstract. We prove that the spectral order provides a domain theoretic


model of qubit channels. Specically, we show that the spectral order
is the unique partial order on quantum states whose least element is
the completely mixed state, which satises the mixing law and has the
property that a qubit channel is unital i it is Scott continuous and has a
Scott closed set of xed points. This result is used to show that the Holevo
capacity of a unital qubit channel is determined by the largest value of
its informatic derivative. In particular, these channels always have an
informatic derivative that is necessarily not a classical derivative.

Introduction

The study of measurement was initiated within the context of computation [3].
In [5], it is shown that measurement can be used to prove xed point theorems for
mappings that are not monotone and unique xed point theorems for mappings
that are monotone. Results like these can be used to provide a unied view
of numerical algorithms, for instance. In such applications, we are primarily
concerned with operators f whose iterates f n (x) converge to a xed point p.
The informatic derivative df (p) then measures the rate at which f converges
to p.
The view of computation taken in the study of measurement, that a computation is a process that evolves on a space of informatic objects, and that as it
evolves we can measure the amount of information lost or gained, in retrospect
lends itself very naturally to considerations in other areas, such as physics or the
study of communication. In [2], it was discovered that natural domain theoretic
structure existed in quantum mechanics. And developments such as [7] and [6]
establish the importance of domains and measurements in classical information
theory.
In this paper, we establish the signicance of domain theory and measurement
in quantum information theory. We rst show that a classical binary channel is
Scott continuous and has a Scott closed set of xed points i it is a binary symmetric channel, while a qubit channel is Scott continuous and has a Scott closed
set of xed points i it is unital. The binary symmetric channels are exactly the
entropy increasing binary channels; the unital qubit channels are exactly the entropy increasing qubit channels. One reason such channels are important is that
L. Aceto et al. (Eds.): ICALP 2008, Part II, LNCS 5126, pp. 283297, 2008.
c Springer-Verlag Berlin Heidelberg 2008


284

K. Martin

they provide eective ways of interrupting communication. For instance, assuming


all inputs are equally likely, the best way to interrupt communication for a xed
probability of error is to use a binary symmetric channel. The class of unital qubit
channels includes most of the models used to describe noise: bit ipping, phase
ipping, bit-phase ipping, phase damping (decoherence), depolarization, unitary channels and projective measurements.
In fact, the connection between entropy increasing channels and Scott continuous channels with Scott closed sets of xed points also turns out to uniquely
determine the spectral order on quantum states. We have known since [2] that
the unique partial order on 2 = {(x, y) [0, 1] : x + y = 1} that satises the
mixing law and has a least element of = (1/2, 1/2) is the Bayesian order. What
we prove in this paper is the quantum analogue of this result: the spectral order
is the unique partial order on two dimensional mixed quantum states 2 that
satises the mixing law, has least element = I/2 and the additional property
that every unital channel is Scott continuous and has a Scott closed set of xed
points. This additional property is trivially satised in the classical case 2 .
Finally, we use these results to give a method for calculating the Holevo capacity of a unital qubit channel. Surprisingly, each unital qubit channel has
an informatic derivative dened everywhere except . The largest value of a
channels informatic derivative determines its Holevo capacity. This informatic
derivative is not a classical derivative. This demonstrates a completely new use
for informatic rates of change.

The Domains of Classical and Quantum States

We review the basic ideas in the study of domains and measurements, and then
the two examples of domains that are of interest in this paper.
2.1

Domain Theory and Measurement

A domain is a partially ordered set with intrinsic notions of completeness and


approximation dened by the order. A measurement is a function that to
each informative object x assigns a number x which measures the information
content of the object x. We now dene each of these terms precisely before
discussing them further.
The intrinsic notion of completeness that a domain has is that it forms a dcpo:
Denition 1. Let (P, ) be a partially ordered set or poset. A nonempty
subset

S P is directed if (x, y S)(z S) x, y  z. The supremum S of S P
is the least of its upper bounds when it exists. A dcpo is a poset in which every
directed set has a supremum.
The intrinsic notion of approximation possessed by a domain is formalized by
continuity:

A Domain Theoretic Model of Qubit Channels

285

Denition 2. Let (D, ) be a dcpo.For elements x, y D, we write x  y i


for every directed subset S with y  S, we have x  s, for some s S. We set
x := {y D : y  x} and x := {y D : x  y}
x := {y D : y  x} and x := {y D : x  y}
and say D is continuous if x is directed with supremum x for each x D.
Denition 3. A domain is a continuous dcpo. A Scott domain is a continuous
dcpo in which any pair of elements with an upper bound has a supremum.
Denition 4. The Scott topology on a continuous dcpo D has as a basis all sets
of the form x for x D. A set S D is Scott closed if it is a lower set that is
closed under directed suprema.
A function f : D E between domains is Scott continuous if the inverse image
of a Scott open set in E is Scott open in D. This is equivalent [1] to saying that
f is monotone,
(x, y D) x  y f (x)  f (y),
and that it preserves directed suprema:


f ( S) =
f (S),
for all directed S D. In particular, for the domain [0, ) of nonnegative reals
in their opposite order, a Scott continuous function : D [0, ) will satisfy
1. For all x, y D, x  y x y, and
2. If (xn ) is an increasing sequence in D, then

xn = lim xn .
n1

This is the case of Scott continuity that pertains to measurements:


Denition 5. A Scott continuous : D [0, ) is said to measure the content of x D if for all Scott open sets U D,
x U ( > 0) x (x) U
where
(x) := {y D : y  x & |x y| < }
are called the -approximations of x.
We often refer to as measuring x D or as measuring X D when it
measures each element of X.
Denition 6. A measurement : D [0, ) is a Scott continuous map that
measures the content of ker() := {x D : x = 0}.

286

K. Martin

In this paper, all measurements we work with measure all of D. This implies [5]
that they are strictly monotone:
x  y & x = y x = y
This property enables denition of the informatic derivative:
Denition 7. Let (D, ) be a domain with a measurement that measures all
of D. If f : D D is a function and p D is not compact, then
df (p) = lim

xp

f (x) f (p)
x p

is called the informatic derivative of f , provided that this limit exists.


2.2

The Bayesian Order on Classical States

The set of classical states


2 := {(x, y) [0, 1]2 : x + y = 1}
has a natural domain theoretic structure introduced in [2]:
Denition 8. For x, y 2 ,
x  y (y1 x1 1/2) or (1/2 x1 y1 ) .
The relation  on 2 is called the Bayesian order.
This order is derived from the graph of entropy H(x) = x log2 (x) (1
x) log2 (1 x) as follows:
(1, 0)

(0, 1)

6
x
-

f lip

= ( 12 ,

1
2)

Theorem 1 ([2]). (2 , ) is a Scott domain with maximal elements


max(2 ) = {(0, 1), (1, 0)}
and least element = (1/2, 1/2). The Shannon entropy H : 2 [0, ) , given
by
H(x) = x1 log2 (x1 ) x2 log2 (x2 )
is a measurement.

A Domain Theoretic Model of Qubit Channels

2.3

287

The Spectral Order on Quantum States

Let H2 denote an two dimensional complex Hilbert space with specied inner
product |.
Denition 9. A quantum state is a density operator : H2 H2 , i.e., a selfadjoint, positive, linear operator with tr() = 1. The quantum states on H2 are
denoted 2 .
Quantum states are also sometimes call density operators or mixed states. The
set of eigenvalues of an operator , called the spectrum of , is denoted spec().
Denition 10. A quantum state on H2 is pure if
spec() {0, 1}.
The set of pure states is denoted 2 . They are in bijective correspondence with
the one dimensional subspaces of H2 .
Classical states are distributions on the set of pure states max(2 ). An analogous
result holds for quantum states: density operators encode distributions on the
set of pure states 2 .
Denition 11. A quantum observable is a self-adjoint linear operator e : H2
H2 .
Now, if we have the operator e representing the energy observable of a system
(for instance), then its spectrum spec(e) consists of the actual energy values a
system may assume. If our knowledge about the state of the system is represented
by density operator , then quantum mechanics predicts the probability that a
measurement of observable e yields the value spec(e). It is
pr( e ) := tr(pe ),
where pe is the projection corresponding to eigenvalue and e is its associated
eigenspace in the spectral representation of e.
Denition 12. Let e be an observable on H2 with spec(e) = {1, 2}. For a
quantum state on 2 ,
spec(|e) := (pr( e1 ), pr( e2 )) 2 .
We assume that all observables e have |spec(e)| = 2. Intuitively, then, e is an
experiment on a system which yields one of 2 dierent outcomes; if our a priori
knowledge about the state of the system is , then our knowledge about what
the result of experiment e will be is spec(|e). Thus, spec(|e) determines our
ability to predict the result of the experiment e.
Let [a, b] = ab ba denote the commutator of operators.

288

K. Martin

Denition 13. For quantum states , 2 , we have  i there is an


observable e : H2 H2 such that [, e] = [, e] = 0 and spec(|e)  spec(|e)
in 2 .
This is called the spectral order on quantum states.
Theorem 2 ([2]). ( 2 , ) is a Scott domain with maximal elements
max( 2 ) = 2
and least element = I/2, where I is the identity matrix. The von Neumann
entropy S : 2 [0, ) given by S() = tr( log2 ()) is a measurement.
The Hilbert space formalism makes things seem much more complicated than
they really are in this case: the spectral order on 2 has a much simpler description which we now consider.
There is a 1-1 correspondence between density operators on a two dimensional
state space and points on the unit ball B3 = {x R3 : |x| 1}: each density
operator : H2 H2 can be written uniquely as


1 1 + rz rx iry
=
2 rx + iry 1 rz

where r = (rx , ry , rz ) R3 satises |r| = rx2 + ry2 + rz2 1. The vector r


B3 is called the Bloch vector associated to . Bloch vectors have a number of
aesthetically pleasing properties.
If and are density operators with respective Bloch vectors r and s, then
(i) the eigenvalues of are (1 |r|)/2, (ii) the von Neumann entropy of is
S = H((1 + |r|)/2) = H((1 |r|)/2), where H : [0, 1] [0, 1] is the base two
Shannon entropy, (iii) if and are pure states and r + s = 0, then and
are orthogonal, and thus form a basis for the state space; conversely, the Bloch
vectors associated to a pair of orthogonal pure states form antipodal points on
the sphere, (iv) the Bloch vector for a convex sum of mixed states is the convex
sum of the Bloch vectors, (v) the Bloch vector for the completely mixed state
I/2 is 0 = (0, 0, 0).
Because of the correspondence between 2 and B3 , we regard the two as equal
for the rest of the paper.
Example 1. From [2], using the Bloch representation of density operators, the
spectral order on 2 is given by x  y i the line from the origin to y passes
through x. That is,
x  y (p [0, 1]) x = py
for all x, y 2 .

Classical and Quantum Channels

We review classical binary channels, qubit channels and then a special subclass
of each of them: the entropy increasing channels.

A Domain Theoretic Model of Qubit Channels

3.1

289

Classical Channels

A binary channel has two inputs (0 and 1) and two outputs (0 and 1).
An input is sent through the channel to a receiver. Because of noise in the
channel, what arrives may not necessarily be what the sender intended. The eect
of noise on input data is modelled by a noise matrix u. If data is sent through
the channel according to the distribution x, then the output is distributed as
y = x u. The noise matrix u is given by

aa

u=
b b
where a = P (0|0) is the probability of receiving 0 when 0 is sent and b = P (0|1)
is the probability of receiving 0 when 1 is sent and x := 1 x for x [0, 1]. Thus,
the noise matrix of a binary channel can be represented by a point (a, b) in the
unit square [0, 1]2 and all points in the unit square represent the noise matrix of
some binary channel.
The noise matrix u of a binary channel denes a function f : 2 2 ,
given by f (x) = x u, which maps an input distribution x 2 to an output
distribution f (x) 2 .
3.2

Quantum Channels

A classical binary channel f : 2 2 takes an input distribution to an


output distribution. In a similar way, a qubit channel is a function of the form
: 2 2 . Specically,
Denition 14. A qubit channel is a function : 2 2 that is convex linear
and completely positive.
To say that is convex linear means that preserves convex sums i.e. sums of
the form x + (1 x) . Complete positivity, dened in [8], is a condition
which ensures that the denition of a qubit channel is compatible with natural
intuitions about joint systems. For our purposes, there is no need to get lost
in too many details of the Hilbert space formulation: thankfully, qubit channels
also have a Bloch representation.
Denition 15. For a qubit channel : 2 2 , the mapping it induces on
the Bloch sphere f : B3 B3 is called the Bloch representation of .
The set of qubit channels is closed under convex sum and composition. If is a
qubit channel and f is its Bloch representation, then (i) the function f is convex
linear, (ii) composition of quantum channels corresponds to composition of Bloch
representations: for channels 1 , 2 , we have f1 2 = f1 f2 , (iii) convex sum
of quantum channels corresponds to convex sum of Bloch representations: for
channels 1 , 2 and x [0, 1], we have fx1 +x2 = xf1 + x
f2 .
To illustrate how these properties make it simple to calculate the Bloch representation of a qubit channel, consider the bit ipping channel,
() = (1 p)I () + p x ()

290

K. Martin

where I is the identity


channel and x () = x x , with x being the spin

01
operator x =
.
10
The Bloch representation of I is fI (r) = r. Using the correspondence between density operators and Bloch vectors, we calculate directly that the Bloch
representation of x is fx (rx , ry , rz ) = (rx , ry , rz ). Thus, by property (iii) of
Bloch representations,
f (rx , ry , rz ) = (1 p)(rx , ry , rz ) + p(rx , ry , rz ) = (rx , (1 2p)ry , (1 2p)rz )
Notice that states of the form (rx , 0, 0) are unchanged by this form of noise, they
are all xed points of f .
3.3

Entropy Increasing Channels

The classical channels f : 2 2 which increase entropy (H(f (x)) H(x))


are exactly those f with f () = . They are the strict mappings of domain theory, which are also known as binary symmetric channels in information theory.
Similarly, the entropy increasing qubit channels are exactly those for which
() = . These are called unital in quantum information theory.
Denition 16. A qubit channel : 2 2 is unital if () = .
A qubit channel is unital i its Bloch representation f satises f (0) = 0. Let
us consider a few important examples of unital channels.
Example 2. Unitary channels. If U is a unitary operator on H2 , then () =
U U is unital since U U = I. The Bloch representation f is given by f (r) =
M r where M is a 3 3 orthogonal matrix with positive determinant, a rotation.
Example 3. Projective measurements. If {P0 , P1 } are projections with P0 + P1 =
I, then
() = P0 P0 + P1 P1
is a unital channel since P02 = P0 and P12 = P1 . In this case, the Bloch representation f satises f2 = f .
Just as with qubit channels, unital channels are also closed under convex sum
and composition: if 1 and 2 are unital channels, then 1 2 and p1 +(1p)2
are unital for p [0, 1].
Example 4. Let x , y and z denote the spin operators





01
0 i
1 0
x =
y =
z =
10
i 0
0 1
Each is unitary and self-adjoint.
(i) Each spin operator i denes a unital channel i () = i i . For a Bloch
vector r = (rx , ry , rz ), the respective Bloch representations sx , sy , sz are
sx (r) = (rx , ry , rz ), sy (r) = (rx , ry , rz ) and sz (r) = (rx , ry , rz ).

A Domain Theoretic Model of Qubit Channels

(ii)
(iii)
(iv)
(v)

The
The
The
The

291

Bit ipping channel = (1 p)I + p x is unital.


phase ipping channel = (1 p)I + p z is unital.
bit-phase ip channel = (1 p)I + p y is unital.
depolarization channel
d(x) = p + (1 p)x

is unital, for a xed p [0, 1].


Not all qubit channels are unital of course. Amplitude damping provides a wellknown example of a qubit channel that is not.

Scott Continuity of Unital Channels

Our rst result establishes that from the domain theoretic perspective, unital
qubit channels are the quantum analogue of binary symmetric channels in the
classical case.
Theorem 3
A classical channel f : 2 2 is binary symmetric i it is Scott continuous and its set of xed points is Scott closed.
A quantum channel f : 2 2 is unital if and only if it is Scott continuous
and its set of xed points is Scott closed.
Proof. First consider the classical case. If a classical channel f is Scott continuous, then it has a least xed point, and since the set of xed points is Scott closed,
= (1/2, 1/2) must be a xed point. This implies that f is binary symmetric
since

aa

(1/2, 1/2)
= ((a + b)/2, (
a + b)/2) = (1/2, 1/2)
b b
Conversely, suppose that f is binary symmetric. Then it can be written as
f (a, b) = (1 p) (a, b) + p (b, a)
for some p [0, 1]. First we show that f is Scott continuous. For the monotonicity
of f , let x, y 2 with x  y. Then we want to show f (x)  f (y). Writing
x = (x1 , x2 ) and y = (y1 , y2 ), we have
(y1 x1 1/2)

or

(1/2 x1 y1 )

by the denition of  on 2 ; we seek to establish


(f1 (y) f1 (x) 1/2)

or

(1/2 f1 (x) f1 (y))

where we have written f (x) = (f1 (x), f2 (x)) and f (y) = (f1 (y), f2 (y)). Notice
that
f1 (x) = (1 2p)x1 + p and f1 (y) = (1 2p)y1 + p.
We consider the cases y1 x1 1/2 and 1/2 x1 y1 separately.

292

K. Martin

In the rst case, f1 (y) f1 (x) 1/2 holds when p 1/2, and 1/2 f1 (x)
f1 (y) holds when p 1/2. In the second case, 1/2 f1 (x) f1 (y) holds when
p 1/2, and f1 (y) f1 (x) 1/2 holds for p 1/2. Thus, f (x)  f (y),
which proves f is monotone. Its Scott continuity now follows from its Euclidean
continuity and the fact that suprema in the Bayesian order coincide with limits
in the Euclidean topology.
Now we show that the xed points of f form a Scott closed set. If p = 0, then
f is the identity mapping, in which case its set of xed points is Scott closed.
Otherwise, its only xed point is , since for p > 0,
(a, b) = f (a, b) = (a, b) = (b, a) = (a, b) = (1/2, 1/2) = .
Either way, the xed points of f form a Scott closed subset of 2 .
In the quantum case, any channel f that is Scott continuous and has a Scott
closed set of xed points must have as a xed point, and so must be unital. For
the converse, we rst show that any unital f is Scott continuous. Recall that f
can be written in Bloch form as f (r) = M r for some 3 3 real matrix M . Then
f is Euclidean continuous, and since suprema in the spectral order are limits in
the Euclidean topology, f is Scott continuous in the spectral order provided it
is monotone.
For the monotonicity of f , let r  s in the spectral order on 2 . Then the
straight line segment s : [0, 1] 2 from to s, given by s (t) = t s
for t [0, 1], must pass through r. To show that f (r)  f (s), we must show
that the line from to f (s) passes through f (r). But this much is clear since
f (s (t)) = M (t s) = t f (s) = f (s) (t). Thus, all unital channels are Scott
continuous.
To see that the set of xed points x(f ) is Scott closed, we rst show that it is
a lower set. If s x(f ) and r  s, then r lies on the line segment that joins to
s. But any point on this line is a xed point of f since f (s (t)) = f (s) (t) =
s (t). In particular, r x(f ). The set x(f ) is closed under directed suprema
by the Scott continuity of f . Thus, x(f ) is Scott closed.
2
Selfmaps on Hausdor spaces have closed sets of xed points. But the Scott
topology is not Hausdor, so the result above is meaningful. The fact that the set
of xed points is Scott closed also has experimental signicance: in attempting
to prepare |0 during QKD, Alice actually prepares (1 )|00| + |11| for
some small > 0. Then this too is a xed point of the noise operator, provided
|0 is, so the only reduction in capacity is due solely to error in preparation
Alice does not suer more noise simply because she cannot prepare a qubit
exactly.

Uniqueness of the Spectral Order

The order on 2 is canonical as follows:

A Domain Theoretic Model of Qubit Channels

293

Theorem 4 ([2]). There is a unique partial order on 2 that satises the mixing
law
x  y and p [0, 1] x  (1 p)x + py  y
and has := (1/2, 1/2) as a least element. It is the Bayesian order on classical
two states.
Because of the simplicity of 2 , it then follows that the binary symmetric channels are exactly the classical channels that are Scott continuous and have a Scott
closed set of xed points. In this section, we prove the analogous result for the
spectral order.
The special orthogonal group SO(3) is the set of 33 orthogonal real matrices
M with a positive determinant i.e. those matrices M such that M 1 = M t and
det(M ) = +1. Such matrices are called rotations.
Lemma 1
(i) Every rotation f SO(3) is the Bloch representation of a unital channel,
(ii) For any x max( 2 ), there is a rotation f SO(3) such that f (x) =
(0, 0, 1).
Proof. (i) This is a folklore result, see [8] for instance.
(ii) This follows from the fact that SO(3) is a transitive group action on S 2 .
However, we want to write a self-contained paper, so let us give a simpler proof.
Every unit vector x appears as the third column M (0, 0, 1) of some orthogonal
matrix M since by the Gram-Schmidt process we can always nd an orthonormal
basis {v1 , v2 , v3 } whose rst vector is v1 = x. Given such an orthonormal basis,
we construct an orthogonal matrix f whose column vectors are the vectors in
the orthonormal basis with the third column being x.
So let us take an orthogonal matrix M such that M (0, 0, 1) = x. Then M 1
is an orthgonal matrix with M 1 (x) = (0, 0, 1). If det(M 1 ) = 1, we set

1 0 0
f = 0 1 0 M 1
0 01
then f is a rotation with f (x) = (0, 0, 1). Otherwise, det(M 1 ) = +1, in which
case we set f = M 1 .
2
Theorem 5. There is a unique partial order on 2 with the following three
properties:
(i) It has least element = I/2,
(ii) It satises the mixing law: if r  s, then r  tr + (1 t)s  s, for all
t [0, 1],
(iii) Every unital channel f : 2 2 is monotone and has a lower set of xed
points.
It is the spectral order, and gives 2 the structure of a Scott domain on which all
unital channels are Scott continuous and have a Scott closed set of xed points.

294

K. Martin

Proof. In this proof, we work with Bloch representations. By the mixing law,
the depolarization channel dt (x) = tx + (1 t) = tx is deationary, so tx  x
for each t [0, 1]. Thus,  contains the spectral order.
Now suppose r  s. We want to show that r precedes s on the line that travels
from to s and on to a pure state. Draw the line a from to r until it hits
the boundary of the Bloch sphere at a point a. Similarly, let b denote the line
from to s and on to a pure state b. Since r  s and s  b, we have r  b by
transitivity and thus r  a, b.
Let f be a rotation such that f (b) = (0, 0, 1). Then f (b) = (0, 0, 1). Let
p be the Bloch representation of a projective measurement in the basis whose
Bloch vectors are {f (b), f (b)}. Then
Im(p) = {(0, 0, t) : t [1, 1]}
Since r  b, f (r)  f (b) and thus p(f (r))  p(f (b)). But, p(f (b)) = f (b), so
f (r) is also a xed point, since the xed points of p are Scott closed. Then
f (r), f (b) Im(p). This means f (r) and f (b) lie on a line that joins a pure state
to its antipode. Because f is a rotation, the same is true of r and b. However, by
the mixing law, the line from r to b, which increases with respect to , does not
pass through since is the least element (otherwise, r = and the proof is
nished). Then a = b, which means r and s lie on a line that joins to a pure
state a.
So let us write r = xa and s = ya for x, y [0, 1]. If x y, the proof is done.
If x > y, then s = (y/x)r  r using the depolarization operator dy/x . But since
r  s, we have r = s by antisymmetry of .
2
Notice that the discrete order on 2 \ {0} with 0 adjoined as the least element
gives a domain that makes all unital channels Scott continuous with a Scott
closed set of xed points, so requiring the mixing law is essential in uniquely
characterizing the spectral order.

Holevo Capacity from the Informatic Derivative

A standard way of measuring the capacity of a quantum channel in quantum


information is the Holevo capacity; it is sometimes called the product state
capacity since input states are not allowed to be entangled across two or more
uses of the channel.
Denition 17. For a trace preserving quantum operation f , the Holevo capacity
is given by





xi i
xi S(f (i ))

C(f ) = sup S f
{xi ,i }

where the supremum is taken over all ensembles {xi , i } of possible input states
i to the channel.

A Domain Theoretic Model of Qubit Channels

295

The possible inputstates i to the channel are in general mixed and the xi are
probabilities with i xi = 1. If f is the Bloch representation of a qubit channel,
the Holevo capacity of f is given by





1 + |f ( i xi ri ) |
1 + |f (ri )|
xi H
C(f ) = sup H

2
2
{xi ,ri }
i
where ri are Bloch vectors for density operators in an ensemble, and we recall
that eigenvalues of a density operator with Bloch vector r are (1 |r|)/2.
Theorem 6. Let (x) = 1 |x| denote the standard measurement on 2 . For
any unital channel f and any p 2 dierent from ,
df (p) =

|f (p)|
|p|

Thus, the Holevo capacity of f is determined by the largest value of its informatic
derivative. Explicitly,


1 1
+
sup df (x)
C(f ) = 1 H
2 2 x ker()
Proof. Since x  p i x = tp for some t [0, 1], x p in the topology i
t 1 , so
df (p) = lim

xp

f (x) f (p)
f (tp) f (p)
= lim
x p
(tp) p
t1
= lim

|f (p)| |f (tp)|
|p| |tp|

= lim

|f (p)|(1 |t|)
|p|(1 |t|)

t1

t1

(Linearity of f )

|f (p)|
|p|

Now we show that the Holevo capacity is determined by the largest value of its
informatic derivative. By the Euclidean continuity of |f |, there is a pure state
r 2 for which
|f (r)| = max |f (x)| = m+
|x|=1

Setting r1 = r, r2 = r and x1 = x2 = 1/2 denes an ensemble for which the


expression maximized in the denition of C(f ) reduces to 1 H((1 + m+ )/2).
Notice that in this step we explicitly make use of the fact that f is unital:
f (0) = 0. This proves 1 H((1 + m+ )/2) C(f ).
For the other inequality, any term in the supremum is bounded from above
by



1 + |f (ri )|
xi H
1
2
i

296

K. Martin

since H(x) 1. For each ri , there is a pure state pi max( 2 ) with ri  pi . By


the Scott continuity of f ,
|f (ri )| |f (pi )| sup |f (x)| = m+ ,
|x|=1

so we have
H

1 + |f (ri )|
2


H

1 + m+
2

which then gives C(f ) 1 H((1 + m+ )/2).

Thus, we see that C(f ) = 1 for any rotation f since df = 1. Notice that df 1
i f is a rotation. For each p [0, 1], the unique channel f  1 with df = p is the
depolarization channel f = dp , so that C(dp ) = 1H((1+p)/2). In fact, the map
(p, 1 p)  d12p denes an isomorphism from the nonnegative classical binary
symmetric channels onto the depolarization channels that preserves capacity.
The only unital qubit channel with capacity zero is 0 itself.
Example 5. The two Pauli channel in Bloch form is




1p
1p
(r) = p r +
sx (r) +
sy (r)
2
2
where sx and sy are the Bloch representations of the unitary channels x and
y . This simplies to
(rx , ry , rz ) = (prx , pry , (1 p)rz )
The matrix associated to is diagonal, so the diagonal element (eigenvalue) that
has largest magnitude also yields the largest value of its informatic derivative.
The capacity of the two Pauli channel is then


1 + max{p, 1 p}
1H
2
where p [0, 1].
The set of unital channels U is compact hence closed and thus forms a dcpo as
a subset of the domain [ 2 2 ].
Corollary 1. The Holevo capacity C : U [0, 1] is Scott continuous.

Closing

The set of unital qubit channels U is a convex monoid and a dcpo with respect
to which the Holevo capacity is monotone. In a similar way, the interval domain
I[0, 1], which models classical binary channels, is a convex monoid and a dcpo
with respect to which the Shannon capacity is monotone [7].

A Domain Theoretic Model of Qubit Channels

297

References
1. Abramsky, S., Jung, A.: Domain theory. In: Abramsky, S., Gabbay, D.M., Maibaum,
T.S.E. (eds.) Handbook of Logic in Computer Science, vol. III, Oxford University
Press, Oxford (1994)
2. Coecke, B., Martin, K.: A partial order on classical and quantum states. Oxford
University Computing Laboratory, Research Report PRG-RR-02-07 (August 2002),
http://web.comlab.ox.ac.uk/oucl/publications/tr/rr-02-07.html
3. Martin, K.: A foundation for computation. Ph.D. Thesis, Tulane University, Department of Mathematics (2000)
4. Martin, K.: Entropy as a xed point. In: Daz, J., Karhum
aki, J., Lepist
o, A.,
Sannella, D. (eds.) ICALP 2004. LNCS, vol. 3142. Springer, Heidelberg (2004)
5. Martin, K.: The measurement process in domain theory. In: Welzl, E., Montanari,
U., Rolim, J.D.P. (eds.) ICALP 2000. LNCS, vol. 1853. Springer, Heidelberg (2000)
6. Martin, K.: Topology in information theory in topology. Theoretical Computer Science (to appear)
7. Martin, K., Moskowitz, I.S., Allwein, G.: Algebraic information theory for binary
channels. In: Proceedings of MFPS 2006. Electronic Notes in Theoretical Computer
Science, vol. 158, pp. 289306 (2006)
8. Nielsen, M., Chuang, I.: Quantum computation and quantum information. Cambridge University Press, Cambridge (2000)
9. Shannon, C.E.: A mathematical theory of communication. Bell Systems Technical
Journal 27, 379423, 623656 (1948)

S-ar putea să vă placă și