a

r
X
i
v
:
q
u
a
n
t
-
p
h
/
0
4
0
2
1
3
0
v
5


5

M
a
r

2
0
0
7
A categorical semantics of quantum protocols
Samson Abramsky and Bob Coecke
Oxford University Computing Laboratory,
Wolfson Building, Parks Road, Oxford OX1 3QD, UK.
samson.abramsky bob.coecke@comlab.ox.ac.uk
Abstract
We study quantum information and computation from a
novel point of view. Our approach is based on recasting
the standard axiomatic presentation of quantum mechan-
ics, due to von Neumann [28], at a more abstract level, of
compact closed categories with biproducts. We show how
the essential structures found in key quantum information
protocols such as teleportation [5], logic-gate teleportation
[12], and entanglement swapping [29] can be captured at
this abstract level. Moreover, from the combination of the
apparently purely qualitative structures of compact
closure and biproducts there emerge scalars and a Born
rule. This abstract and structural point of view opens up
new possibilities for describing and reasoning about quan-
tum systems. It also shows the degrees of axiomatic free-
dom: we can show what requirements are placed on the
(semi)ring of scalars C(I, I), where C is the category and
I is the tensor unit, in order to perform various proto-
cols such as teleportation. Our formalism captures both
the information-ow aspect of the protocols [8, 9], and the
branching due to quantum indeterminism. This contrasts
with the standard accounts, in which the classical informa-
tion ows are outside the usual quantum-mechanical for-
malism. We give detailed formal descriptions and proofs of
correctness of the example protocols.
1. Introduction
Quantum information and computation is concerned
with the use of quantum-mechanical systems to carry out
computational and information-processingtasks [20]. In the
few years that this approach has been studied, a number of
remarkable concepts and results have emerged. Our par-
ticular focus in this paper is on quantum information proto-
cols, which exploit quantum-mechanical effects in an essen-
tial way. The particular examples we shall use to illustrate
our approach will be teleportation [5], logic-gate telepor-
tation [12], and entanglement swapping [29]. The ideas
illustrated in these protocols form the basis for novel and
potentially very important applications to secure and fault-
tolerant communication and computation [7, 12, 20].
We now give a thumbnail sketch of teleportation to mo-
tivate our introductory discussion. (A more formal stan-
dard presentation is given in Section 2. The radically
different presentation in our new approach appears in
Section 9.) Teleportation involves using an entangled pair
of qubits (q
A
, q
B
) as a kind of communication channel to
transmit an unknown qubit q from a source A (Alice) to a
remote target B (Bob). A has q and q
A
, while B has q
B
.
We rstly entangle q
A
and q at A (by performing a suitable
unitary operation on them), and then perform a measure-
ment on q
A
and q.
1
This forces a collapse in q
B
because
of its entanglement with q
A
. We then send two classical
bits of information from A to B, which encode the four
possible results of the measurement we performed on q and
q
A
. Based on this classical communication, B then per-
forms a correction by applying one of four possible oper-
ations (unitary transformations) to q
B
, after which q
B
has
the same state that q had originally. (Because of the mea-
surement, q no longer has this state the information in
the source has been destroyed in transferring it to the tar-
get). It should be born in mind that the information required
to specify q is an arbitrary pair of complex numbers (, )
satisfying [[
2
+ [[
2
= 1, so achieving this information
transfer with just two classical bits is no mean feat!
Teleportation is simply the most basic of a family of
quantum protocols, and already illustrates the basic ideas,
in particular the use of preparations of entangled states
as carriers for information ow, performing measurements
to propagate information, using classical information to
control branching behaviour to ensure the required be-
haviour despite quantum indeterminacy, and performing lo-
cal data transformations using unitary operations. (Local
1
This measurement can be performed in the standard computational
basis. The combination of unitary and measurement is equivalent to mea-
surement in the Bell basis.
here means that we apply these operations only at A or at
B, which are assumed to be spatially separated, and not si-
multaneously at both).
Our approach is based on recasting the standard ax-
iomatic presentation of Quantum Mechanics, due to von
Neumann [28], at a more abstract level, of compact closed
categories with biproducts. Remarkably enough, all the
essential features of quantum protocols mentioned above
nd natural counterparts at this abstract level of which
the standard von Neumann presentation in terms of Hilbert
spaces is but one example. More specically:
The basic structure of a symmetric monoidal cate-
gory allows compound systems to be described in a
resource-sensitive fashion (cf. the no cloning and no
deleting theorems of quantum mechanics [20]).
The compact closed structure allows preparations and
measurements of entangled states to be described, and
their key properties to be proved.
Biproducts allow indeterministic branching, classical
communication and superpositions to be captured.
We are then able to use this abstract setting to give precise
formulations of teleportation, logic gate teleportation, and
entanglement swapping, and to prove correctness of these
protocols for example, proving correctness of telepor-
tation means showing that the nal value of q
B
equals the
initial value of q. Moreover, from the combination of the
apparently purely qualitativestructures of compact clo-
sure and biproducts there emerge scalars and a Born rule.
One of our main concerns is to replace ad hoc calcula-
tions with bras and kets, normalizing constants, unitary ma-
trices etc. by conceptual denitions and proofs. This allows
general underlying structures to be identied, and general
lemmas to be proved which encapsulate key formal proper-
ties. The compact-closed level of our axiomatization allows
the key information-ow properties of entangled systems
to be expressed. Here we are directly abstracting from the
more concrete analysis carried out by one of the authors in
[8, 9]. The advantage of our abstraction is shown by the fact
that the extensive linear-algebraic calculations in [8] are re-
placed by a few simple conceptual lemmas, valid in an arbi-
trary compact closed category. We are also able to reuse the
template of denition and proof of correctness for the basic
teleportation protocol in deriving and verifying logic-gate
teleportation and entanglement swapping.
The compact-closed level of the axiomatization allows
information ow along any branch of a quantum proto-
col execution to be described, but it does not capture
the branching due to measurements and quantum inde-
terminism. The biproduct structure allows this branch-
ing behaviour to be captured. Since biproducts induce
a (semi)additive structure, the superpositions characteris-
tic of quantum phenomena can be captured at this abstract
level. Moreover, the biproduct structure interacts with the
compact-closed structure in a non-trivial fashion. In par-
ticular, the distributivity of tensor product over biproduct
allows classical communication, and the dependence of ac-
tions on the results of previous measurements (exemplied
in teleportation by the dependence of the unitary correction
on the result of the measurement of q and q
A
), to be cap-
tured within the formalism. In this respect, our formalism
is more comprehensive than the standard von Neumann ax-
iomatization. In the standard approach, the use of measure-
ment results to determine subsequent actions is left informal
and implicit, and hence not subject to rigorous analysis and
proof. As quantum protocols and computations grow more
elaborate and complex, this point is likely to prove of in-
creasing importance.
Another important point concerns the generality of our
axiomatic approach. The standard von Neumann axiomati-
zation ts Quantum Mechanics perfectly, with no room to
spare. Our basic setting of compact closed categories with
biproducts is general enough to allow very different models
such as Rel, the category of sets and relations. When we
consider specic protocols such as teleportation, a kind of
Reverse Arithmetic (by analogy with Reverse Mathemat-
ics [26]) arises. That is, we can characterize what require-
ments are placed on the semiring of scalars C(I, I) (where
I is the tensor unit) in order for the protocol to be realized.
This is often much less than requiring that this be the eld
of complex numbers (but in the specic cases we shall con-
sider, the requirements are sufcient to exclude Rel). Other
degrees of axiomatic freedom also arise, although we shall
not pursue that topic in detail in the present paper.
The remainder of the paper is structured as follows. Sec-
tion 2 contains a rapid review of the standard axiomatic
presentation of Quantum Mechanics, and of the standard
presentations of our example protocols. Section 3 intro-
duces compact closed categories, and presents the key lem-
mas on which our analysis of the information-ow proper-
ties of these protocols will be based. Section 4 relates this
general analysis to the more concrete and specic presen-
tation in [8]. Section 5 introduces biproducts. Sections 6
and 7 present our abstract treatments of scalars and adjoints.
Section 8 presents our abstract formulation of quantum me-
chanics. Section 9 contains our formal descriptions and ver-
ications of the example protocols. Section 10 concludes.
2. Quantum mechanics and teleportation
In this paper, we shall only consider nitary quan-
tum mechanics, in which all Hilbert spaces are nite-
dimensional. This is standard in most current discussions
of quantum computation and information [20], and corre-
sponds physically to considering only observables with -
nite spectra, such as spin. (We refer briey to the extension
of our approach to the innite-dimensional case in the Con-
clusions.)
Finitary quantum theory has the following basic ingredi-
ents (for more details, consult standard texts such as [13]).
1. The state space of the system is represented as a nite-
dimensional Hilbert space H, i.e. a nite-dimensional
complex vector space with an inner product written
[ ), which is conjugate-linear in the rst argu-
ment and linear in the second. A state of a quantum
system corresponds to a one-dimensional subspace /
of H, and is standardly represented by a vector /
of unit norm.
2. For informatic purposes, the basic type is that of
qubits, namely 2-dimensional Hilbert space, equipped
with a computational basis [0), [1).
3. Compound systems are described by tensor products
of the component systems. It is here that the key
phenomenon of entanglement arises, since the general
form of a vector in H
1
H
2
is
n

i=1

i

i

i
Such a vector may encode correlations between the
rst and second components of the system, and cannot
simply be resolved into a pair of vectors in the compo-
nent spaces.
The adjoint to a linear map f : H
1
H
2
is the linear map
f

: H
2
H
1
such that, for all H
2
and H
1
,
[ f())
H2
= f

() [ )
H1
.
Unitary transformations are linear isomorphisms
U : H
1
H
2
such that
U
1
= U

: H
2
H
1
.
Note that all such transformations preserve the inner prod-
uct since, for all , H
1
,
U() [ U())
H2
= (U

U)() [ )
H1
= [ )
H1
.
Self-adjoint operators are linear transformations
M : H H
such that M = M

.
4. The basic data transformations are represented by uni-
tary transformations. Note that all such data transfor-
mations are necessarily reversible.
5. The measurements which can be performed on the sys-
tem are represented by self-adjoint operators.
The act of measurement itself consists of two parts:
5a. The observer is informed about the measurement out-
come, which is a value x
i
in the spectrum (M) of
the corresponding self-adjoint operator M. For con-
venience we assume (M) to be non-degenerate (lin-
early independent eigenvectors have distinct eigenval-
ues).
5b. The state of the system undergoes a change, repre-
sented by the action of the projector P
i
arising from
the spectral decomposition
M = x
1
P
1
+ . . . + x
n
P
n
In this spectral decomposition the projectors P
i
: H H
are idempotent and self-adjoint,
P
i
P
i
= P
i
and P
i
= P

i
,
and mutually orthogonal:
P
i
P
j
= 0, i ,= j.
This spectral decomposition always exists and is unique
by the spectral theorem for self-adjoint operators. By our
assumption that (M) was non-degenerate each projector
P
i
has a one-dimensional subspace of H as its xpoint set
(which equals its image).
The probability of x
i
(M) being the actual outcome
is given by the Born rule which does not depend on the
value of x
i
but on P
i
and the system state , explicitly
Prob(P
i
, ) = [ P
i
()) .
The status of the Born rule within our abstract setting will
emerge in Section 8. The derivable notions of mixed states
and non-projective measurements will not play a signicant
r ole in this paper.
The values x
1
, . . . , x
n
are in effect merely labels distin-
guishing the projectors P
1
, . . . , P
n
in the above sum. Hence
we can abstract over them and think of a measurement as a
list of n mutually orthogonal projectors (P
1
, . . . , P
n
) where
n is the dimension of the Hilbert space.
Although real-life experiments in many cases destroy the
system (e.g. any measurement of a photons location de-
stroys it) measurements always have the same shape in the
quantum formalism. When distinguishing between mea-
surements which preserve the system and measurements
which destroy the system it would make sense to decom-
pose a measurement explicitly in two components:
Observation consists of receiving the information on
the outcome of the measurement, to be thought of as
specication of the index i of the outcome-projector
P
i
in the above list. Measurements which destroy the
system can be seen as observation only.
Preparation consists of producing the state P
i
().
In our abstract setting these arise naturally as the two build-
ing blocks which are used to construct projectors and mea-
surements.
We now discuss some important quantum protocols
which we chose because of the key r ole entanglement plays
in them they involve both initially entangled states, and
measurements against a basis of entangled states.
2.1 Quantum teleportation
The quantum teleportation protocol [5] (see also [8] 2.3
and 3.3) involves three qubits a, b and c (corresponding to
q, q
A
and q
B
respectively in our preliminary sketch in the
Introduction). Qubit a is in a state [) and qubits b and c
form an EPR-pair, that is, their joint state is [00) + [11).
After spatial relocation (so that a and b are positioned at
the source A, while c is positioned at the target B), one
performs a Bell-base measurement on a and b, that is, a
measurement such that each P
i
projects on one of the one-
dimensional subspaces spanned by a vector in the Bell ba-
sis:
b
1
:=
1

2
([00)+[11)) b
2
:=
1

2
([01)+[10))
b
3
:=
1

2
([00)[11)) b
4
:=
1

2
([01)[10)) .
This measurement can be of the type observation only. We
observe the outcome of the measurement and depending on
it perform one of the unitary transformations

1
:=
_
1 0
0 1
_

2
:=
_
0 1
1 0
_

3
:=
_
1 0
0 1
_

4
:=
_
0 1
1 0
_
on c
1
,
2
,
3
are all self-inverse while
1
4
=
4
.
Physically, this requires transmission of two classical bits,
recording the outcome of the measurement, from the loca-
tion of a and b to the location of c.
|00+|11
M
Bell
Ux
x B
2
|
|
6
time
The nal state of c proves to be [) as well. We will be able
to derive this fact in our abstract setting.
Since a continuous variable has been transmitted while
the actual classical communication involved only two bits,
besides this classical information ow there has to exist a
quantum information ow. The nature of this quantum ow
has been analyzed by one of the authors in [8, 9], building
on the joint work in [2]. We recover those results in our ab-
stract setting (see Section 4), which also reveals additional
ne structure. To identify it we have to separate it from
the classical information ow. Therefore we decompose the
protocol into:
1. a tree with the operations as nodes, and with branch-
ing caused by the indeterminism of measurements;
2. a network of the operations in terms of the order they
are applied and the subsystem to which they apply.
|00+|11
M
Bell
U
00
U
01
U
10
U
11
00 01 10 11
...
...
...
a b c
The nodes in the tree are connected to the boxes in the net-
work by their temporal coincidence. Classical communica-
tion is encoded in the tree as the dependency of operations
on the branch they are in. For each path from the root of the
tree to a leaf, by lling in the operations on the included
nodes in the corresponding boxes of the network, we ob-
tain an entanglement network, that is, a network
|00+|11
Px
Ux
a b c
6
time
for each of the four values x takes. A component P
x
of an
observation will be referred to as an observational branch.
It will be these networks, from which we have removed the
classical information ow, that we will study in Section 4.
(There is a clear analogy with the idea of unfolding a Petri
net into its set of processes [21]). The classical informa-
tion ow will be reintroduced in Section 9.
2.2 Logic gate teleportation
Logic gate teleportation [12] (see also [8] 3.3) gener-
alizes the above protocol in that b and c are initially not
necessarily an EPR-pair but may be in some other (not ar-
bitrary) entangled state [). Due to this modication the
nal state of c is not [) but [f

()) where f

is a linear
map which depends on . As shown in [12], when this
construction is applied to the situation where a, b and c are
each a pair of qubits rather than a single qubit, it provides a
universal quantum computational primitive which is more-
over fault-tolerant [25] and enables the construction of a
quantumcomputer based on single qubit unitary operations,
Bell-base measurements and only one kind of prepared state
(so-called GHZ states). The connection between , f

and
the unitary corrections U
,x
will emerge straightforwardly
in our abstract setting.
2.3 Entanglement swapping
Entanglement swapping [29] (see also [8] 6.2) is an-
other modication of the teleportation protocol where a is
not in a state [) but is a qubit in an EPR-pair together with
an ancillary qubit d. The result is that after the protocol c
forms an EPR-pair with d. If the measurement on a and b is
non-destructive, we can also performa unitary operation on
a, resulting in a and b also constituting an EPR-pair. Hence
we have swapped entanglement:
|00+|11
|00+|11
|00+|11 |00+|11
;
b
a d
c b
a d
c
In this case the entanglement networks have the shape:
|00+|11 |00+|11
Px
Ux Ux
d a b c
6
time
Why this protocol works will again emerge straightfor-
wardly from our abstract setting, as will generalizations of
this protocol which have a much more sophisticated com-
positional content (see Section 4).
3. Compact closed categories
Recall that a symmetric monoidal category consists of a
category C, a bifunctorial tensor
: CC C,
a unit object I and natural isomorphisms

A
: A I A
A
: A A I

A,B,C
: A(B C) (A B) C

A,B
: A B B A
which satisfy certain coherence conditions [17]. A cate-
gory C is -autonomous [4] if it is symmetric monoidal,
and comes equipped with a full and faithful functor
( )

: C
op
C
such that a bijection
C(A B, C

) C(A, (B C)

)
exists which is natural in all variables. Hence a -
autonomous category is closed, with
A B := (A B

.
These -autonomous categories provide a categorical se-
mantics for the multiplicative fragment of linear logic [23].
A compact closed category [15] is a -autonomous cate-
gory with a self-dual tensor, i.e. with natural isomorphisms
u
A,B
: (A B)

u
I
: I

I .
It follows that
A B A

B .
3.1 Denitions and examples
A very different denition arises when one considers a
symmetric monoidal category as a one-object bicategory.
In this context, compact closure simply means that every
object A, qua 1-cell of the bicategory, has an adjoint [16].
Denition 3.1 (Kelly-Laplaza) A compact closed cate-
gory is a symmetric monoidal category in which to each
object A a dual object A

, a unit

A
: I A

A
and a counit

A
: A A

I
are assigned in such a way that the diagram
A

A
-
A I
1
A

A
-
A (A

A)
A
1
A
?

1
A
I A

A
1
A
(A A

) A

A,A

,A
?
and the dual one for A

both commute.
The monoidal categories (Rel, ) of sets, relations and
cartesian product and (FdVec
K
, ) of nite-dimensional
vector spaces over a eld K, linear maps and tensor product
are both compact closed. In (Rel, ), taking a one-point
set as the unit for , and writing R

for the converse

of a relation R:

X
=

X
= (, (x, x)) [ x X .
The unit and counit in (FdVec
K
, ) are

V
: K V

V :: 1
i=n

i=1
e
i
e
i

V
: V V

K :: e
i
e
j
e
j
(e
i
)
where n is the dimension of V , e
i

i=n
i=1
is a basis of V and
e
i
is the linear functional in V

determined by e
j
(e
i
) =
ij
.
Denition 3.2 The name f and the coname f of a
morphism f : A B in a compact closed category are
A

A
1
A
f
-
A

B I
I

A
6

-
AB

f 1
B

-
BB

B
6
For R Rel(X, Y ) we have
R = (, (x, y)) [ xRy, x X, y Y
R = ((x, y), ) [ xRy, x X, y Y
and for f FdVec
K
(V, W) with (m
ij
) the matrix of f in
bases e
V
i

i=n
i=1
and e
W
j

j=m
j=1
of V and W respectively:
f : K V

W :: 1
i,j=n,m

i,j=1
m
ij
e
V
i
e
W
j
f : V W

K :: e
V
i
e
W
j
m
ij
.
3.2 Some constructions
Given f : A B in any compact closed category C we
can dene f

: B

as follows [16]:
B

-
I B

A
1
B

-
A

AB

1
A

1
A

B
A

B B

1
A
f 1
B

?
This operation ( )

is functorial and makes Denition 3.1

coincide with the one given at the beginning of this section.
It then follows by
C(A B

, I) C(A, B) C(I, A

B)
that every morphism of type I A

B is the name of
some morphism of type A B and every morphism of
type A B

I is the coname of some morphism of type

A B. In the case of the unit and the counit we have

A
= 1
A
and
A
= 1
A
.
For R Rel(X, Y ) the dual is the converse, R

= R

Rel(Y, X), and for f FdVec
K
(V, W), the dual is
f

: W

V

:: f .
In any compact closed category, there is a natural iso-
morphism d
A
: A

A.
The following holds by general properties of adjoints and
the fact that the tensor is symmetric [16].
Proposition 3.3 In a compact closed category C we have
I

A

-
A

,A
-
A A

A
?

,A
-
A A

d
A
1
A

?
A

1
A
d
1
A
?

-
I

A
?
for all objects A of C.
3.3 Key lemmas
The following Lemmas constitute the core of our inter-
pretation of entanglement in compact closed categories.
Lemma 3.4 (absorption) For
A
f
-
B
g
-
C
we have that
(1
A
g) f = g f.
Proof: Straightforward by Denition 3.2. 2
Lemma 3.5 (compositionality) For
A
f
-
B
g
-
C
we have that

1
C
(f 1
C
) (1
A
g)
A
= g f .
Proof: See the diagram in the appendix to this paper which
uses bifunctoriality and naturality of and . 2
Lemma 3.6 (compositional CUT) For
A
f
-
B
g
-
C
h
-
D
we have that
(
1
A
1
D
)(1
A
g1
D
)(fh)
I
= hgf.
Proof: See the diagram in the appendix to this paper which
uses Lemma 3.5 and naturality of and . 2
On the right hand side of Lemma 3.5 we have g f, that
is, we rst apply f and then g, while on the left hand side
we rst apply the coname of g, and then the coname of f. In
Lemma 3.6 there is a similar, seemingly acausal inversion
of the order of application, as g gets inserted between h and
f.
For completeness we add the following backward ab-
sorption lemma, which again involves a reversal of the com-
position order.
Lemma 3.7 (backward absorption) For
C
g
-
A
f
-
B
we have that
(g

1
A
) f = f g.
Proof: This follows by unfolding the denition of g

, then
using naturality of
A
,
I
=
I
, and nally Lemma 3.6. 2
The obvious analogues of Lemma 3.4 and 3.7 for conames
also hold.
4. Abstract entanglement networks
We claim that Lemmas 3.4, 3.5 and 3.6 capture the quan-
tum information ow in the (logic-gate) teleportation and
entanglement swapping protocols. We shall provide a full
interpretation of nitary quantum mechanics in Section 8
but for now the following rule sufces:
We interpret preparation of an entangled state as a
name and an observational branch as a coname.
For an entanglement network of teleportation-type shape
(see the picture below) applying Lemma 3.5 yields
U
_

1
C
(f 1)
_
((1 g)
A
) = U g f .
Note that the quantuminformation seems to ow following
the line while being acted on by the functions whose name
or coname labels the boxes (and this fact remains valid for
much more complex networks [8]).
g
f
U
6
time
Teleporting the input requires U g f = 1
A
we assume
all functions have type A A. Logic-gate teleportation of
h : A B requires U g f = h.
We calculate this explicitly in Rel. For initial state x
X after preparing
S (Y Z)
we obtain
x (y, z) [ S(y, z)
as the state of the system. For observational branch
R (X Y )
we have that z Z is the output iff R 1
Z
receives an
input (x, y, z) X Y Z such that (x, y)R . Since
S(y, z) ySz and (x, y)R xRy
we indeed obtain x(R; S)z. This illustrates that the com-
positionality is due to a mechanism of imposing constraints
between the components of the tuples.
In FdVec
C
the vector space of all linear maps of type
V W is V W and hence by
V

W V W
we have a bijective correspondence between linear maps
f : V W and vectors V

W (see also [8, 9]):

2
f
= f(1) f =

2
f
[) .
In particular we have for the Bell base:

2 b
i
=
i
(1)
i
=

2 b
i
[) .
Setting
g :=
1
= 1
V
, f :=
i
, U :=
1
i
indeed yields

1
i
1
A

i
= 1
A
,
which expresses the correctness of the teleportation proto-
col along each branch.
Setting g := h and f :=
i
for logic-gate teleportation
requires U
i
to satisfy U
i
h
i
= h that is
h
i
= U

h
(since U has to be unitary). Hence we have derived the
laws of logic-gate teleportation one should compare this
calculation to the size of the calculation in Hilbert space.
Deriving the swapping protocol using Lemma 3.4 and
Lemma 3.6 proceeds analogously to the derivation of the
teleportation protocol. We obtain two distinct ows due to
the fact that a non-destructive measurement is involved.
1 1

1
i

1
i 6
time
How
i
has to relate to
i
such that they make up a true
projector will be discussed in Section 8.
For a general entanglement network of the swapping-
type (without unitary correction and observational branch-
ing) by Lemma 3.6 we obtain the following reduction:
;
f h
g
hgf
This picture, and the underlying algebraic property ex-
pressed by Lemma 3.5, is in fact directly related to Cut-
Elimination in the logic corresponding to compact-closed
categories. If one turns the above picture upside-down, and
interprets names as Axiom-links and conames as Cut-links,
then one has a normalization rule for proof-nets. This per-
spective is developed in [11].
5. Biproducts
Biproducts have been studied as part of the structure of
Abelian categories. For further details, and proofs of the
general results we shall cite in this Section, see e.g. [19].
A zero object in a category is one which is both initial
and terminal. If 0 is a zero object, there is an arrow
0
A,B
: A
-
0
-
B
between any pair of objects A and B. Let C be a category
with a zero object and binary products and coproducts. Any
arrow
A
1

A
2
A
1

A
2
can be written uniquely as a matrix (f
ij
), where f
ij
: A
i

A
j
. If the arrow
_
1 0
0 1
_
is an isomorphism for all A
1
, A
2
, then we say that C has
biproducts, and write AB for the biproduct of A and B.
Proposition 5.1 (Semi-additivity) If C has biproducts,
then we can dene an operation of addition on each hom-set
C(A, B) by
A
f + g
-
B
A A

?
f g
-
B B

6
for f, g : A B, where
= 1
A
, 1
A
) and = [1
B
, 1
B
]
are respectively the diagonal and codiagonal. This op-
eration is associative and commutative, with 0
AB
as an
identity. Moreover, composition is bilinear with respect to
this additive structure. Thus C is enriched over abelian
monoids.
Proposition 5.2 If C has biproducts, we can choose pro-
jections p
1
, . . ., p
n
and injections q
1
, . . ., q
n
for each

k=n
k=1
A
k
satisfying
p
j
q
i
=
ij
and
k=n

k=1
q
k
p
k
= 1
L
k
A
k
where
ii
= 1
Ai
, and
ij
= 0
Ai,Aj
, i ,= j.
Proposition 5.3 (Distributivity of over ) In monoidal
closed categories there are natural isomorphisms

A,B,C
: A (B C) (A B) (A C) ,
explicitly,

A,,
= 1
A
p
1
, 1
A
p
2
)
1
A,,
= [1
A
q
1
, 1
A
q
2
] .
A left distributivity isomorphism

A,B,C
: (A B) C (A C) (A C)
can be dened similarly.
Proposition 5.4 (Self-duality of for ( )

) In any com-
pact closed category there are natural isomorphisms

A,B
: (A B)

I
: 0

0.
Writing n X for

i=n
i=1
X it follows by self-duality of the
tensor unit I that

1
I,...,I
(n u
I
) : n I (n I)

.
Matrix representation. We can write any arrow of the
form f : AB C D as a matrix
M
f
:=
_
p
C,D
1
f q
A,B
1
p
C,D
1
f q
A,B
2
p
C,D
2
f q
A,B
1
p
C,D
2
f q
A,B
2
_
.
The sumf +g of such morphisms corresponds to the matrix
sum M
f
+M
g
and composition g f corresponds to matrix
multiplication M
g
M
f
. Hence categories with biproducts
admit a matrix calculus.
Examples. The categories (Rel, , +) where the biprod-
uct is the disjoint union and (FdVec
K
, , ) where the
biproduct is the direct sum are examples of compact closed
categories with biproducts. More generally, the category of
relations for a regular category with stable disjoint coprod-
ucts; the category of nitely generated projective modules
over a commutative ring; the category of nitely generated
free semimodules over a commutative semiring; and the cat-
egory of free semimodules over a complete commutative
semiring are all compact closed with biproducts. Compact
closed categories with biproducts, with additional assump-
tions (e.g. that the category is abelian) have been studied in
the mathematical literature on Tannakian categories [10].
They have also arisen in a Computer Science context in the
rst authors work on Interaction Categories [3].
6. Scalars
In any compact closed category we shall call endomor-
phisms s : I I scalars. As observed in [16], in any
monoidal category C, the endomorphismmonoid C(I, I) is
commutative. Any scalar s induces a natural transformation
s
A
: A A by
A

-
I A
s 1
A
-
I A

1
-
A.
Here naturality means that all morphisms preserve scalar
multiplication. We write s f for f s
A
, where s is a
scalar and f : A B. If C moreover has biproducts, the
scalars C(I, I) form a commutative semiring.
Examples. In FdVec
K
, linear maps s : K K are
uniquely determined by the image of 1, and hence corre-
spond biuniquely to elements of K; composition and addi-
tion of these maps corresponds to multiplication and addi-
tion of scalars. Hence in FdVec
K
the commutative semir-
ing of scalars is the eld K. In Rel, there are just two
scalars, corresponding to the classical truth values. Hence
in Rel the commutative semiring of scalars is the Boolean
semiring 0, 1.
7. Strong compact closure
In any compact closed category C, there is a natural iso-
morphism A A

. It will be notationally convenient to

assume that ( )

is strictly involutive, so that this natural iso-

morphism is the identity. The following denition allows
the key example of (complex) Hilbert spaces to be accomo-
dated in our setting.
Denition 7.1 A compact closed category C is strongly
compact closed if the assignment on objects A A

ex-
tends to a covariant functor, with action on morphisms
f

: A

for f : A B, such that

f

= f (f

= (f

: B A.
Examples. Any compact closed category such as Rel, in
which ( )

is the identity on objects, is trivially strongly

compact closed (we just take f

:= f). The category of

nite-dimensional real inner product spaces and linear maps
offers another example of this situation, where we take A =
A

, and dene

A
: [ ) .
Our main intended example, FdHilb, the category of
nite-dimensional Hilbert spaces and linear maps, exhibits
this structure less trivially, since the conjugate-linearity in
the rst argument of the inner product prevents us frompro-
ceeding as for real spaces. Instead, we dene H

as follows.
The additive abelian group of vectors in H

is the same as
in H. Scalar multiplication and the inner product are

H
:=
H
[ )
H
:= [ )
H
where is the complex conjugate of . The covariant ac-
tion is then just f

= f. Note that the identity map from

H to H

is a conjugate-linear isomorphism, but not linear

and hence does not live in the category FdHilb! Im-
portantly, however, H and H

have the same orthonormal

bases. Hence we can dene

H
: 1
i=n

i=1
e
i
e
i

H
: [ )
H
where e
i

i=n
i=1
is an orthonormal basis of H.
7.1 Adjoints, unitarity and inner products
Each morphism in a strongly compact closed category
admits an adjoint in the following sense.
Denition 7.2 We set
f

:= (f

= (f

,
and call this the adjoint of f.
Proposition 7.3 The assignments A A on objects, and
f f

on morphisms, dene a contravariant involutive

functor:
(f g)

= g

= 1 f

= f.
In FdHilb and real inner product spaces, f

is the usual
adjoint of a linear map. In Rel, it is relational converse.
Denition 7.4 An isomorphism U is called unitary if its
adjoint is its inverse (U

= U
1
).
Denition 7.5 Given , : I A we dene their abstract
inner product [ ) as
I

I
-
II
1
I
u
I
-
II

-
AA

A
-
I.
In FdHilb, this denition coincides with the usual inner
product. In Rel we have for x, y X:
x [ y) = 1
I
, x y ,= x [ y) = 0
I
, x y = .
Lemma 7.6 If : I A, then

is given by
A

A
-
AI
1
A
u
I
-
AI

1
A

-
AA

A
-
I.
Proof: See the diagramin the appendix to this paper, which
uses Proposition 3.3 twice (with d
A
= 1
A
). 2
Proposition 7.7 For [ ) as dened above we have
[ ) =

.
Proof: Using bifunctoriality of tensor and naturality of , it
is easy to see that [ ) can be written as
I

-
A
A
-
AI
1AuI
-
AI

1A
-
AA

A
-
I .
We now apply Lemma 7.6 to conclude. 2
Proposition 7.8 For
: I A : I B f : B A
we have
f

[ )
B
= [ f )
A
.
Proof: By Proposition 7.7,
f

[ ) = (f

f = [ f ). 2
Proposition 7.9 Unitary morphisms U : A B preserve
the inner product, that is for all , : I A we have
U [ U )
B
= [ )
A
.
Proof: By Proposition 7.8,
U [ U )
B
= U

U [ )
A
= [ )
A
. 2
7.2 Bras and kets
By Proposition 7.8 we can interpret the Dirac notation
(e.g. [20]) in our setting. For morphisms
: I A : I B f : B A
dene
[ f [ ) := f

[ )
B
= [ f )
A
.
By Proposition 7.7,
[ f [ ) =

f .
7.3 Strong compact closure and biproducts
Proposition 7.10 If C has biproducts, ( )

preserves them
and hence is additive:
(f + g)

= f

+ g

A,B
= 0
B,A
.
If a category is both strongly compact closed and has
biproducts, the adjoint acts as an involutive automorphism
on the semiring of scalars C(I, I). For Rel and real in-
ner product spaces it is the identity, while in the case of
FdHilb, it corresponds to complex conjugation.
We need a compatibility condition between the strong
compact closure and the biproducts.
Denition 7.11 We say that a category Cis a strongly com-
pact closed category with biproducts iff
1. It is strongly compact closed;
2. It has biproducts;
3. The coproduct injections
q
i
: A
i

k=n

k=1
A
k
satisfy
q

j
q
i
=
ij
.
Fromthis, it follows that we can require that the chosen
projections and injections in Proposition 5.2 addition-
ally satisfy (p
i
)

= q
i
.
Examples Finite-dimensional Hilbert spaces and real in-
ner product spaces, categories of relations, and categories of
free modules and semimodules are all examples of strongly
compact closed categories with biproducts.
7.4 Spectral Decompositions
We dene a spectral decomposition of an object A to be
a unitary isomorphism
U : A
i=n

i=1
A
i
.
(Here the spectrum is just the set of indices 1, . . . , n).
Given a spectral decomposition U, we dene morphisms

j
:= U

q
j
: A
j
A

j
:=

j
= p
j
U : A A
j
,
diagramatically
A
j

j
-
A
i=n

i=1
A
i
q
j
?
p
j
-

U
A
j

j
?
and nally projectors
P
j
:=
j

j
: A A.
These projectors are self-adjoint
P

j
= (
j

j
)

j
=
j

j
= P
j
idempotent and orthogonal:
P
i
P
j
=
i

i

j

j
=
i

ij

j
=
A
ij
P
i
.
Moreover, they yield a resolution of the identity:
i=n

i=1
P
i
=
i=n

i=1

i

i
=
i=n

i=1
U

q
i
p
i
U
= U

(
i=n

i=1
q
i
p
i
) U
= U
1
1
L
i
Ai
U = 1
A
.
7.5 Bases and dimension
A basis for an object A is a unitary isomorphism
base : n I A.
Given bases base
A
and base
B
for objects A and B respec-
tively we can dene the matrix (m
ij
) of any morphism
f : A B in those two bases as the matrix of
base

B
f base
A
: n
A
I n
B
I
as in Section 5.
Proposition 7.12 Given f : A B and
base
A
: n
A
I A and base
B
: n
B
I A
the matrix (m

ij
) of f

in these bases is the conjugate trans-

pose of the matrix (m
ij
) of f.
Proof: We have
m

ij
= p
i
base

A
f

base
B
q
j
= (p
j
base

B
f base
A
q
i
)

= m

ji
.
2
If in addition to the assumptions of Proposition 7.8 and
Proposition 7.9 there exist bases for A and B, we can prove
converses to both of them.
Proposition 7.13 If there exist bases for A and B then f :
A B is the adjoint to g : B A if and only if
f [ )
B
= [ g )
A
for all : I A and : I B.
Proof: Let (m
ij
) be the matrix of f

and (m

ij
) the matrix
of g in the given bases. By Proposition 7.7 we have
m
ij
= p
i
base

A
f

base
B
q
j
= f base
A
q
i
[ base
B
q
j
)
B
= f [ )
B
= [ g )
A
= base
A
q
i
[ g base
B
q
j
)
A
= p
i
base

A
g base
B
q
j
= m

ij
.
Hence the matrix elements of g and f

coincide so g and f

are equal. The converse is given by Proposition 7.8. 2

Proposition 7.14 If there exist bases for A and B then a
morphism U : A B is unitary if and only if it preserves
the inner product, that is for all , : I A we have
U [ U )
B
= [ )
A
.
Proof: We have
U
1
[ )
A
= U U
1
[ U )
B
= [ U )
B
and hence by Proposition 7.13, U

= U
1
. The converse
is given by Proposition 7.9. 2
Note also that when a basis is available we can assign to

: A I and : I A matrices
_

n
_
_
_
_

1
.
.
.

n
_
_
_
respectively, and by Proposition 7.7, the inner product be-
comes
[ ) =
_

n
_
_
_
_

1
.
.
.

n
_
_
_ =
i=n

i=1

i

i
.
Interestingly, two different notions of dimension arise in
our setting. We assign an integer dimension dim(A) N to
an object A provided there exists a base
base : dim(A) I A.
Alternatively, we introduce the scalar dimension as
dim
s
(A) :=
A

A

,A

A
C(I, I).
We also have:
dim
s
(I) = 1
I
dim
s
(A

) = dim
s
(A)
dim
s
(A B) = dim
s
(A)dim
s
(B)
In FdVec
K
these notions of dimension coincide, in
the sense that dim
s
(V ) is multiplication with the scalar
dim(V ). In Rel the integer dimension corresponds to the
cardinality of the set, and is only well-dened for nite sets,
while dim
s
(X) always exists; however, dim
s
(X) can only
take two values, 0
I
and 1
I
, and the two notions of dimension
diverge for sets of cardinality greater than 1.
8. Abstract quantum mechanics
We can identify the basic ingredients of nitary quantum
mechanics in any strongly compact closed category with
biproducts.
1. A state space is represented by an object A.
2. A basic variable (type of qubits) is a state space Q
with a given unitary isomorphism
base
Q
: I I Q
which we call the computational basis of Q. By using
the isomorphismn I (n I)

described in Section 5,
we also obtain a computational basis for Q

.
3. A compound system for which the subsystems are de-
scribed by Aand B respectively is described by AB.
If we have computational bases base
A
and base
B
, then
we dene
base
AB
:= (base
A
base
B
) d
1
nm
where
d
nm
: n I m I (nm) I
is the canonical isomorphism constructed using rst
the left distributivity isomorphism , and then the
right distributivity isomorphism , to give the usual
lexicographically-ordered computational basis for the
tensor product.
4. Basic data transformations are unitary isomorphisms.
5a. A preparation in a state space A is a morphism
: I A
for which there exists a unitary U : I B A such
that
I

-
A
I B
q
1
?
U
-
commutes.
5b. Consider a spectral decomposition
U : A
i=n

i=1
A
i
with associated projectors P
j
. This gives rise to the
non-destructive measurement
P
i
)
i=n
i=1
: A n A.
The projectors
P
i
: A A
for i = 1, . . . , n are called the measurement branches.
This measurement is non-degenerate if A
i
= I for all
i = 1, . . . , n. In this case we refer to U itself as a de-
structive measurement or observation. The morphisms

i
= p
i
U : A I
for i = 1, . . . , n are called observation branches. (We
leave discussion of degenerate destructive measure-
ments, along with other variant notions of measure-
ment, to future work).
Note that the type of a non-destructive measurement makes
it explicit that it is an operation which involves an inde-
terministic transition (by contrast with the standard Hilbert
space quantum mechanical formalism).
6a. Explicit biproducts represent the branching arising
from the indeterminacy of measurement outcomes.
Hence an operation f acting on an explicit biproduct AB
should itself be an explicit biproduct, i.e. we want
f = f
1
f
2
: A B C D,
for f
1
: A C and f
2
: B D. The dependency of f
i
on the branch it is in captures local classical communica-
tion. The full force of non-local classical communication is
enabled by Proposition 5.3.
6b. Distributivity isomorphisms represent non-local clas-
sical communication.
To see this, suppose e.g. that we have a compound system
QA, and we (non-destructively) measure the qubit in the
rst component, obtaining a new system state described by
(Q Q) A. At this point, we know locally, i.e. at the
site of the rst component, what the measurement outcome
is, but we have not propagated this information to the rest
of the system A. However, after applying the distributivity
isomorphism
(QQ) A (QA) (QA)
the information about the outcome of the measurement on
the rst qubit has been propagated globally throughout the
system, and we can perform operations on A depending on
the measurement outcome, e.g.
(1
Q
U
0
) (1
Q
U
1
)
where U
0
, U
1
are the operations we wish to perform on A
in the event that the outcome of the measurement we per-
formed on Q was 0 or 1 respectively.
The Born rule
We now show how the Born rule, which is the key quan-
titative feature of quantum mechanics, emerges automati-
cally from our abstract setting.
For a preparation : I A and spectral decomposition
U : A
i=n

i=1
A
i
,
with corresponding non-destructive measurement
P
i
)
i=n
i=1
: A n A,
we can consider the protocol
I

-
A
P
i
)
i=n
i=1
-
n A.
We dene scalars
Prob(P
i
, ) := [ P
i
[ ) =

P
i
.
Proposition 8.1 With notation as above,
Prob(P
i
, ) = (Prob(P
i
, ))

and
i=n

i=1
Prob(P
i
, ) = 1 .
Hence we think of the scalar Prob(P
j
, ) as the prob-
ability of obtaining the jth outcome of the measurement
P
i
)
i=n
i=1
on the state .
Proof: From the denitions of preparation and the projec-
tors, there are unitaries U, V such that
Prob(P
i
, ) = (V q
1
)

q
i
p
i
U V q
1
for each i. Hence
i=n

i=1
Prob(P
i
, ) =
i=n

i=1
p
1
V

U

q
i
p
i
U V q
1
= p
1
V

U

_
n

i=1
q
i
p
i
_
U V q
1
= p
1
V
1
U
1
1
nI
U V q
1
= p
1
q
1
= 1
I
. 2
Moreover, since by denition P
j
=

j

j
, we can rewrite
the Born rule expression as
Prob(P
j
, ) =

P
j

=

j

j

= (
j
)

j

= s

j
s
j
for some scalar s
j
C(I, I). Thus s
j
can be thought of
as the probability amplitude giving rise to the probability
s

j
s
j
, which is of course self-adjoint. If we consider the
protocol
I

-
A

i
)
i=n
i=1
-
n I .
which involves an observation
i
)
i=n
i=1
, then these scalars
s
j
correspond to the branches
I

-
A

j
-
I .
9. Abstract quantum protocols
We prove correctness of the example protocols.
9.1 Quantum teleportation
Denition 9.1 A teleportation base is a scalar s together
with a morphism
prebase
T
: 4 I Q

Q
such that:
base
T
:= s prebase
T
is unitary.
the four maps
j
: Q Q, where
j
is dened by

j
:= prebase
T
q
j
,
are unitary.
2s

s = 1.
The morphisms s
j
are the base vectors of the tele-
portation base. A teleportation base is a Bell base when the
Bell base maps

1
,
2
,
3
,
4
: Q Q
satisfy
2

1
= 1
Q

2
=

Q

3
=

3

4
=

Q

3
where

Q
:= base
Q

I,I
base
1
Q
.
A teleportation base denes a teleportation observation
s

i
)
i=4
i=1
: QQ

4 I .
To emphasize the identity of the individual qubits we la-
bel the three copies of Q we shall consider as Q
a
, Q
b
, Q
c
.
We also use labelled identities, e.g. 1
bc
: Q
b
Q
c
, and la-
belled Bell bases. Finally, we introduce

4
ac
:= s

s 1
ac
)
i=4
i=1
: Q
a
4 Q
c
as the labelled, weighted diagonal. This expresses the in-
tended behaviour of teleportation, namely that the input
qubit is propagated to the output along each branch of the
protocol, with weight s

s, corresponding to the probabil-

ity amplitude for that branch. Note that the sum of the cor-
responding probabilities is
4(s

s)

s = (2s

s)(2s

s) = 1 .
2
This choice of axioms is sufcient for our purposes. One might prefer
to axiomatize a notion of Bell base such that the corresponding Bell base
maps are exactly the Pauli matrices.
Theorem 9.2 The following diagram commutes.
Q
a
=============== Q
a
import unknown state
Q
a
I

a
?
produce EPR-pair
Q
a
(Q

b
Q
c
)
1
a
(s 1
bc
)
?
spatial delocation
(Q
a
Q

b
) Q
c

a,b,c
?
teleportation observation
(4 I) Q
c
s

ab
i
)
i=4
i=1
1
c
?
classical communication
4 Q
c
_
4
1
c
_

c
?
unitary correction
4 Q
c

4
ac
?
============ 4 Q
c

i=4
i=1
(
c
i
)
1
?
The right-hand-side of the above diagram is our formal de-
scription of the teleportation protocol; the commutativity
of the diagram expresses the correctness of the protocol.
Hence any strongly compact closed category with biprod-
ucts admits quantum teleportation provided it contains a
teleportation base. If we do a Bell-base observation then
the corresponding unitary corrections are

1
i
=
i
for i 1, 2, 3 and
1
4
=
3

Q
.
Proof: For a proof of the commutativity of this diagram see
the Appendix it uses the universal property of the product,
Lemma 3.5, naturality of and the explicit form of

c
:= p
I
i
1
c
)
i=4
i=1
.
In the specic case of a Bell-base observation we use
1

Q
= 1
Q
, (

Q
)

Q
and (

Q

3
)

3
(

Q
)

Q
. 2
Although in Rel teleportation works for individual ob-
servational branches it fails to admit the full teleporta-
tion protocol since there are only two automorphisms of Q
(which is just a two-element set, i.e. the type of classical
bits), and hence there is no teleportation base.
We now consider sufcient conditions on the ambient
category C for a teleportation base to exist. We remark
rstly that if C(I, I) contains an additive inverse for 1, then
it is a ring, and moreover all additive inverses exist in each
hom-set C(A, B), so C is enriched over Abelian groups.
Suppose then that C(I, I) is a ring with 1 ,= 1. We can
dene a morphism
prebase
T
= base
Q

Q
M : 4 I Q

Q
where M is the endomorphism of 4 I determined by the
matrix
_
_
_
_
1 0 1 0
0 1 0 1
0 1 0 1
1 0 1 0
_
_
_
_
The corresponding morphisms
j
will have 2 2 matrices
determined by the columns of this 4 4 matrix, and will be
unitary. If C(I, I) furthermore contains a scalar s satisfying
2s

s = 1, then s prebase
T
is unitary, and the conditions
for a teleportation base are fullled. Suppose we start with
a ring R containing an element s satisfying 2s
2
= 1. (Ex-
amples are plentiful, e.g. any subring of C, or of Q(

2),
containing
1

2
). The category of nitely generated free R-
modules and R-linear maps is strongly compact closed with
biproducts, and admits a teleportation base (in which s will
appear as a scalar with s = s

), hence realizes teleportation.

9.2 Logic-gate teleportation
Logic gate teleportation of qubits requires only a minor
modication as compared to the teleportation protocol.
Theorem 9.3 Let unitary morphism f : Q Q be such
that for each i 1, 2, 3, 4 a morphism
i
(f) : Q Q
satisfying
f
i
=
i
(f) f
exists. The diagram of Theorem 9.2 with the modications
made below commutes.
Q
a
I
?
.
.
.
.
.
produce f-state
Q
a
(Q

b
Q
c
)
1
a
(s f)
?
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Q
c
?
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
unitary correction
4 Q
c

4
ac
f
?
========= 4 Q
c

i=4
i=1
(
i
(f))
1
?
The right-hand-side of the diagram is our formal descrip-
tion of logic-gate teleportation of f : Q Q; the commu-
tativity of the diagramunder the stated conditions expresses
the correctness of logic-gate teleportation for qubits.
Proof: See the diagram in the appendix. 2
This two-dimensional case does not yet provide a univer-
sal computational primitive, which requires teleportation of
QQ-gates [12]. We present the example of teleportation
of a CNOT gate [12] (see also [8] Section 3.3).
Given a Bell base we dene a CNOT gate as one which
acts as follows on tensors of the Bell base maps
3
:
CNOT (

Q
1
Q
) = (

Q
) CNOT
CNOT (1
Q

Q
) = (1
Q

Q
) CNOT
CNOT (
3
1
Q
) = (
3
1
Q
) CNOT
CNOT (1
Q

3
) = (
3

3
) CNOT
It follows from this that
CNOT (
4
1
Q
) = (
4

Q
) CNOT
CNOT (1
Q

4
) = (
3

4
) CNOT
fromwhich in turn it follows by bifunctoriality of the tensor
that the required unitary corrections factor into single qubit
actions, for which we introduce a notation by setting
CNOT (
i
1
Q
) =
1
(
i
) CNOT
CNOT (1
Q

i
) =
2
(
i
) CNOT
3
One could give a more explicit denition of a CNOT gate, e.g. by
specifying the matrix. However, our generalized denition sufces to pro-
vide the required corrections. Moreover, this example nicely illustrates the
attitude of focussing on the essentials by abstracting.
The reader can verify that for
4
2
(Q
c1
Q
c2
) := 4 (4 (Q
c1
Q
c2
))
and

4
2
ac
:=s

s s

s 1
ac
)
i=4
i=1
)
i=4
i=1
: Q
a1
Q
a2
4
2
(Q
c1
Q
c2
)
the following diagram commutes.
Q
a1
Q
a2
============ Q
a1
Q
a2
import unknown state
(Q
a1
Q
a2
) I

a
?
produce CNOT-state
(Q
a1
Q
a2
) ((Q
b1
Q
b2
)

(Q
c1
Q
c2
))
1
a
(s
2
CNOT)
?
spatial delocation
((Q
a1
Q

b1
) (Q
c1
Q
c2
)) (Q
a2
Q

b2
)
(, ) (1
a
(u
b
1
c
))
?
1st observation
((4 I) (Q
c1
Q
c2
)) (Q
a2
Q

b2
)
(s

a1b1
i
)
i=4
i=1
1
c
)1
2
?
1st communication
(4 (Q
c1
Q
c2
)) (Q
a2
Q

b2
)
((4
1
c
)
c
) 1
2
?
1st correction
(4 (Q
c1
Q
c2
)) (Q
a2
Q

b2
)
_

i=4
i=1
(
c
1
(
i
))
1
_
1
2
?
2nd observation
(4 (Q
c1
Q
c2
)) (4 I)
(4 1
c
)s

a2b2
i
)
i=4
i=1
?
2nd communication
(4 (4 (Q
c1
Q
c2
)))
(4
1
4c
)
4c
?
2nd correction
4
2
(Q
c1
Q
c2
)

4
2
ac
CNOT
?
======= 4
2
(Q
c1
Q
c2
)

i=4
i=1
(4
c
2
(
i
))
1
?
9.3 Entanglement swapping
Theorem 9.4 Setting

i
:= (
i
)

P
i
:= s

s (
i

i
)

ac
i
:=

i=4
i=1
_
(1

b

1
i
) (1

d

1
i
)
_

ab
:= 1

d
P
i
)
i=4
i=1
1
c

ab
:= s

s
3
(1
ba
1
dc
))
i=4
i=1
the following diagram commutes.
I I ============== I I
produce EPR-pairs
(Q

d
Q
a
) (Q

b
Q
c
)
s
2
(1
da
1
bc
)
?
spatial delocation
Q

d
(Q
a
Q

b
) Q
c

?
Bell-base measurement
Q

d
(4 (Q
a
Q

b
)) Q
c

ab
?
classical communication
4 ((Q

b
Q
a
)(Q

d
Q
c
))
(4 (, )) (, )
?
unitary correction
4 ((Q

b
Q
a
)

ab
?
(Q

d
Q
c
))

ac
i
?
The right-hand-side of the above diagram is our formal de-
scription of the entanglement swapping protocol.
Proof: See the diagram in the appendix it uses Lemma
3.4 and Lemma 3.6. 2
We use
i
= (
i
)

rather than
i
to make P
i
an endo-
morphism and hence a projector. The general denition of
a bipartite entanglement projector is
P
f
:= f f

= f f

,B
: A

B A

B
for f : A B, so in fact P
i
= P
(i)
.
10. Conclusion
Other work. Birkhoff and von Neumann [6] attempted
to capture quantum behavior abstractly in lattice-theoretic
terms see also Mackey [18] and Piron [22]. The weak
spot of this programme was the lack of a satisfactory treat-
ment of compound systems whereas in our approach the
tensor is a primitive. Different kinds of lattices do arise
naturally in our setting, but we leave a discussion of this to
future work.
Isham and Buttereld [14] have reformulated the
Kochen-Specker theorem in a topos-theoretic setting. On
the one hand, assuming that the tensor in a compact closed
category is the categorical product leads to trivialitythe
category is then necessarily equivalent to the one-object
one-arrow categoryand in this sense the compact closed
and topos axioms are not compatible. On the other hand,
each topos yields a strongly compact closed category with
biproducts as its category of relations.
The recent papers [24, 27] use categorical methods for
giving semantics to a quantum programming language, and
a quantum lambda calculus, respectively. In both cases, the
objectives, approach and results are very different to those
of the present paper. A more detailed comparison must
again be left to future work.
Further Directions. This work has many possible lines
for further development. We mention just a few.
Our setting seems a natural one for developing type
systems to control quantum behaviour.
In order to handle protocols and quantum computa-
tions more systematically, it would be desirable to have
an effective syntax, whose design should be guided by
the categorical semantics.
The information ow level of analysis using only the
compact-closed structure allows some very elegant
and convenient qualitative reasoning, while adding
biproducts allows very ne-grained modelling and
analysis. The interplay between these two levels merits
further investigation.
We have not considered mixed states and non-
projective measurements in this paper, but they can
certainly be incorporated in our framework.
In this paper, we have only studied nitary Quan-
tum Mechanics. A signicant step towards the in-
nite dimensional case is provided by the previous
work on nuclear ideals in tensored -categories [1].
The compactness axiom for nuclear ideals (see Def-
inition 5.7 in [1]) corresponds to our Compositional-
ity Lemma 3.4. One of the main intended models of
nuclear ideals is given by the category of all Hilbert
spaces and bounded linear maps.
Another class of compact closed categories with
biproducts are the Interaction Categories introduced
by one of the authors [3]. One can consider linear-
algebraic versions of Interaction Categories matri-
ces extended in time rather than relations extended
in time as in [3]. Does this lead to a useful notion of
quantum concurrent processes?
Acknowledgements
Rick Blute and Prakash Panangaden suggested some im-
provements to an earlier version of this paper.
References
[1] S. Abramsky, R. Blute, and P. Panangaden. Nuclear and
trace ideals in tensored -categories. Journal of Pure and
Applied Algebra, 143:347, 1999.
[2] S. Abramsky and B. Coecke. Physical traces: Quantum
vs. classical information processing. In Proceedings of Cat-
egory Theory and Computer Science 2002 (CTCS02), vol-
ume 69 of Electronic Notes in Theoretical Computer Sci-
ence. Elsevier Science, 2003. arXiv:cs/0207057.
[3] S. Abramsky, S. J. Gay, and R. Nagarajan. Interaction cate-
gories and foundations of typed concurrent programming. In
Deductive Program Design: Proceedings of the 1994 Mark-
toberdorf International Summer School, NATO Advanced
Science Institutes Series F, pages 35113. Springer-Verlag,
1995.
[4] M. Barr. -Autonomous Categories, volume 752 of Lecture
Notes in Mathematics. Springer-Verlag, 1979.
[5] C. H. Bennett, G. Brassard, C. Cr epeau, R. Jozsa, A. Peres,
and W. K. Wooters. Teleporting an unknown quantum state
via dual classical and Einstein-Podolsky-Rosen channels.
Physical Review Letters, 70:18951899, 1993.
[6] G. Birkhoff and J. von Neumann. The logic of quantum
mechanics. Annals of Mathematics, 37:823843, 1936.
[7] D. Bouwmeester, A. Ekert, and A. Zeilinger, editors. The
physics of quantum information. Springer-Verlag, Berlin,
2001.
[8] B. Coecke. The logic of entanglement. An invitation.
Technical Report PRG-RR-03-12, Oxford University, 2003.
web.comlab.ox.ac.uk/oucl/publications/tr/rr-03-12.html.
[9] B. Coecke. The logic of entanglement. arXiv:quant-ph/
0402014, 2004.
[10] P. Deligne. Cat egories tannakiennes. In The Grothendieck
Festschrift Volume II, volume 87 of Progress in Mathemat-
ics, pages 111196. Birkh auser, 1990.
[11] R. Duncan. Quantum entanglement and multiplicative linear
logic. Transfer Report Oxford University, November 2003.
[12] D. Gottesman and I. L. Chuang. Quantum teleportation is
a universal computational primitive. Nature, 402:390393,
1999. arXiv:quant-ph/9908010.
[13] C. Isham. Lectures on Quantum Theory. Imperial College
Press, 1995.
[14] C. J. Isham and J. Buttereld. A topos perspective on the
Kochen-Specker theorem I: Quantum states as generalized
valuations. International Journal of Theoretical Physics,
37:26692733, 1998.
[15] G. M. Kelly. An abstract approach to coherence. Lecture
Notes in Mathematics, 281:106147, 1972.
[16] G. M. Kelly and M. L. Laplaza. Coherence for compact
closed categories. Journal of Pure and Applied Algebra,
19:193213, 1980.
[17] S. Mac Lane. Categories for the Working Mathematician.
Springer-Verlag, 1971.
[18] G. M. Mackey. The mathematical foundations of quantum
mechanics. W. A. Benjamin, 1963.
[19] B. Mitchell. Theory of Categories. Academic Press, 1965.
[20] M. A. Nielsen and L. Chuang. Quantum computation and
quantum information. Cambridge University Press, 2000.
[21] C. A. Petri. Non-sequential processes. Technical Report
ISF-77-5, GMD, St-Augustin, Germany, 1977.
[22] C. Piron. Foundations of quantum physics. W. A. Benjamin,
1976.
[23] R. A. G. Seely. Linear logic, -autonomous categories
and cofree algebras. In Categories in Computer Science
and Logic, volume 92 of Contemporary Mathematics, pages
371382, June 1987, Boulder, Colorado, 1989.
[24] P. Selinger. Towards a quantum programming language.
Mathematical Structures in Computer Science, 2004. To ap-
pear.
[25] P. W. Shor. Fault-tolerant quantumcomputation. In Proceed-
ings of the 37nd Annual Symposium on Foundations of Com-
puter Science, pages 5665. IEEE Computer Society Press,
1996.
[26] S. G. Simpson. Subsystems of Second-Order Arithmetic.
Springer-Verlag, 1999.
[27] A. van Tonder. Quantum computation, categorical seman-
tics, and linear logic. arXiv:quant-ph/0312174, 2003.
[28] J. von Neumann. Mathematische grundlagen der quan-
tenmechanik. Springer-Verlag, Berlin, 1932. Translation,
Mathematical foundations of quantum mechanics, Princeton
University Press, 1955.
[29] M.

Zukowski, A. Zeilinger, M. A. Horne, and A. K. Ek-
ert. Event-ready-detectors Bell experiment via entangle-
ment swapping. Physical Review Letters, 71:42874290,
1993.
Appendix: Diagramatic proofs
Proof of Lemma 3.5 (compositionality). The top trapezoid is the statement of the Lemma.
A
g f
-
C
A I
1
A
g
-

A
-
A B

C
f 1
C
-
I C

1
C
-
A B

B
1
A

g
-
1
A

B
-
B B

1
C
-
f

1
B

C
-
B I
f 1
I
?
1
B

B
-
B B

B
1
B
-
1
B

g
-
f

1
B

B
-
I B
1
I
g
6
B
f
?
=======================================================================

B
-
B
g
6

1 B
-
Lemma 3.5
Compact closedness
Proof of Lemma 3.6 (compositional CUT). The top trapezoid is the statement of the Lemma.
I
h g f
-
A

D
I I
fh
-

I
-
A

BC

D
1
A
g1
D
-
A

ID

1
A

1
D

1
A

1
D

-
A

AI

A
1
I
?
1
A
f 1
I
-
A

BI
1
A
h
6

1
I
-
A

B
1
A

(
h

g
)
-
A

A
?
==================================================================

A
-
A

A
1
A
(hgf)
6

1
A

f
Lemma 3.5
Lemma 3.6
Proof of Lemma 7.6 (adjoints to points). The top trapezoid is the statement of the Lemma. The cell labelled SMC
commutes by symmetric monoidal coherence.
A

-
I
I A

A
?

I,A
-
A I
1
A
u
I
-

A
-
A I

================= A I

1
A

-
AA

A
-
I A
_
_
_
_
_
_
_
_
_
_
(I

I) A

I,A
-

1
A
-
A (I

I)
1
A

I
6
1
A

I
-
I

,A
6

1
A
-
A

,A
6

-
I
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(I I

)A

,I
1
A
?

1
I,I

,A
-

1
A -
I (I

A)
1
I
(

1
A
)
-

1
I

A
-
I A

A
1
I

A

1
A

A
-
I I

1
I
=

1
I
-
Prop. 3.3 SMC
Lemma 7.6
Prop. 3.3
Proof of Theorem 9.2 (quantum teleportation). For each j 1, 2, 3, 4 we have a diagram of the form below. The top
trapezoid is the statement of the Theorem. We ignore the scalars which cancel out against each other in this proof.
Q
a
1
ac
)
i=4
i=1
-
4 Q
c
Q
a
I
1
a
1
bc

a
-
Q
a
Q

b
Q
c

ab
i
)
i=4
i=1
1
c
-
(4 I) Q
c
p
I
i
1
c
)
i=4
i=1
-
4 (IQ
c
)
4
1
c
-
4 Q
c

i
=
4
i
=
1
(

c
i
)

1
i
-
Q
c

1
c

c
- I Q
c
p
I
j
1
c
?
1
IQc
-

a
b
j

1
c
-
I Q
c
p
IQc
j
?

1
c
-
Q
c
p
Qc
j
?
Q
a
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
1
ac
-

a
c
j
-
Q
c
p
Qc
j
?
(

c
j
)
1
-
Lemma 3.5
Quantum teleportation
Proof of Theorem 9.3 (logic-gate teleportation). The top trapezoid is the statement of the Theorem. The a, b and c-labels
are the same as above. For each j 1, 2, 3, 4 we have a diagram of the form below. We ignore the scalars which cancel
out against each other in this proof.
Q
f)
i=4
i=1
-
4 Q
QI
1
Q
f
-

Q
-
QQ

i
)
i=4
i=1
1
Q
-
(4 I)Q
p
I
i
1
Q
)
i=4
i=1
-
4 (IQ)
4
1
Q
-
4 Q

i
=
4
i
=
1

i
(
f
)

1
-
Q
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
f
j
-
Q

1
Q

Q
- I Q
p
I
j
1
Q
?
1
IQ
-

1
Q -
I Q
p
IQ
j
?

1
Q
-
Q
p
Q
j
?
Q
_
_
_
_
_
_
_
_
_
_
f
-

j
(
f
)

f
-
Q
p
Q
j
?

j
(
f
)
1
-
Lemma 3.5
Logic-gate teleportation
Proof of Theorem 9.4 (entanglement swapping). The top trapezoid is the statement of the Theorem. We have a diagram of
the form below for each j 1, 2, 3, 4. To simplify the notation of the types we set (a

, b, c

, d) for Q

a
Q
b
Q

c
Q
d
etc. We ignore the scalars which cancel out against each other in this proof.
I I
1
ba
1
dc
)
i=4
i=1
-
4 (b

, a, d

, c)
(d

, a, b

, c)

ab
-

1
d
a

1
b
c
-
(d

, 4(a, b

), c)
(, )
-
4 (d

, a, b

, c)
4
-
4 (b

, a, d

, c)

-
(d

, c)

1
d
1
c
(d

, I , c)
1

j
1
c
-
1
d

1
c
-
(d

, a, b

, c)
p
(d

,a,b

,c)
j
?

-
1

d
p
(
a
,b

) j

1
c
-
(b

, a, d

, c)
p
(b

,a,d

,c)
j
?
I

I
6

-
(I , d

, c)

d
1
c
?

1
(d

, a, b

, c)
1

d
1
ab

1
j
?
==================== (d

, a, b

, c)
1

1
j
1

1
j
?

-
(b

, a, d

, c)
p
(b

,a,d

,c)
j
?

-
I I

I
?
1
ba
1
dc

-
(b

, a, d

, c)

1
?

1
b
a

1
d

1
j
-
Entanglement swapping
Lemma 3.4
Lemma 3.4
Lemma 3.6
:=

i=4
i=1
(1

1
i
1

1
i
)
:= 1

1
j
1

1
j