See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/2102595

What is Stochastic Independence?

Article · July 2002

DOI: 10.1142/9789812705242_0008 · Source: arXiv

CITATIONS READS
13 2,160

1 author:

Uwe Franz
Université de Franche-Comté
111 PUBLICATIONS 1,130 CITATIONS

SEE PROFILE

All content following this page was uploaded by Uwe Franz on 20 December 2013.

The user has requested enhancement of the downloaded file.

 What is Stochastic Independence?
arXiv:math/0206017v1 [math.QA] 3 Jun 2002

Uwe Franz
Institut für Mathematik und Informatik
Ernst-Moritz-Arndt-Universität Greifswald
Friedrich-Ludwig-Jahn-Str. 15 a
D-17487 Greifswald, Germany
Email: [email protected]
http://hyperwave.math-inf.uni-greifswald.de/algebra/franz

Abstract
The notion of a tensor product with projections or with inclusions is de-
fined. It is shown that the definition of stochastic independence relies on such
a structure and that independence can be defined in an arbitrary category
with a tensor product with inclusions or projections. In this context, the clas-
sifications of quantum stochastic independence by Muraki, Ben Ghorbal, and
Schürmann become classifications of the tensor products with inclusions for
the categories of algebraic probability spaces and non-unital algebraic prob-
ability spaces. The notion of a reduction of one independence to another
is also introduced. As examples the reductions of Fermi independence and
boolean, monotone, and anti-monotone independence to tensor independence
are presented.

1 Introduction
In this paper we will deal with the question, what stochastic independence
is. Since the work of Speicher[13] and Schürmann[12, 2, 1] we know that a
‘universal’ notion of independence should come with a product that allows
to construct the joint distribution of two independent random variables from
their marginal distributions. It turned out that in classical probability there
exists only one such product satisfying a natural set of axioms. But there
are several different good notions of independence in non-commutative prob-
ability. The most important ones were classified in the work of Speicher[13],
Ben Ghorbal and Schürmann[1, 2], and Muraki[9, 10], they are tensor inde-
pendence, free independence, boolean independence, monotone independence
and anti-monotone independence.
We present and motivate here the axiomatic framework used in these arti-
cles. We show that the classical notion of stochastic independence is based on
a kind of product in the category of probability spaces, which is intermediate
to the notion of a (universal) product in category theory - which does not
exist in this category - and the notion of a tensor product. Furthermore, we
show that the classification of stochastic independence by Ben Ghorbal and
Schürmann[12, 1, 2] and by Muraki[9, 10] is also based on such a product,
which we call a tensor product with projections or inclusions, cf. Definition
3.3. This notion allows to define independence for arbitrary categories, see
Definition 3.4. If independence is something that depends on a tensor prod-
uct and projections or inclusions between the original objects and their tensor
product, then it is clear that a map between categories that preserves indepen-
dence should be tensor functor with an additional structure that takes care of

1
the projections or inclusions. This is formalized in Definition 4.2. We show in
several examples that these notions are really the correct one, see Subsection
3.1, 3.2, 3.3, and 4.1, and Section 7.
But let us first look at the notion of independence in classical probability.

2 Independence for Classical Random Vari-

ables
Two random variables X1 : (Ω, F, P ) → (E1 , E1 ) and X2 : (Ω, F, P ) →
(E2 , E2 ), defined on the same probability space (Ω, F, P ) and with values
in two possibly distinct measurable spaces (E1 , E1 ) and (E2 , E2 ), are called
stochastically independent (or simply independent) w.r.t. P , if the σ-algebras
X1−1 (E1 ) and X2−1 (E2 ) are independent w.r.t. P , i.e. if




P (X1−1 (M1 ) ∩ X2−1 (M2 ) = P (X1−1 (M1 ) P X2−1 (M2 )

holds for all M1 ∈ E1 , M2 ∈ E2 . If there is no danger of confusion, then the

reference to the measure P is often omitted.
This definition can easily be extended to arbitrary families of random
variables. A family Xj : (Ω, F, P ) → (Ej , Ej ))j∈J , indexed by some set J, is
called independent, if
n
! n
\ −1
Y

P (Xjk (Mjk ) = P Xj−1k
(Mjk )
k=1 k=1

holds for all n ∈ N and all choices of indices k1 , . . . , kn ∈ J with jk 6= jℓ for

j 6= ℓ, and all choices of measurable sets Mjk ∈ Ejk .
There are many equivalent formulations for independence, consider, e.g.,
the following proposition.
Proposition 2.1 Let X1 and X2 be two real-valued random variables. The
following are equivalent.
(i) X1 and X2 are independent.
(ii) For all bounded measurable functions f1 , f2 on R we have




E f1 (X1 )f2 (X2 ) = E f1 (X1 ) E f2 (X2 ) .

(iii) The probability space (R2 , B(R2 ), P(X1 ,X2 ) ) is the product of the proba-
bility spaces (R, B(R), PX1 ) and (R, B(R), PX2 ), i.e.

P(X1 ,X2 ) = PX1 ⊗ PX2 .

We see that stochastic independence can be reinterpreted as a rule to

compute the joint distribution of two random variables from their marginal
distribution. More precisely, their joint distribution can be computed as a
product of their marginal distributions. This product is associative and can
also be iterated to compute the joint distribution of more than two indepen-
dent random variables.
The classifications of independence for non-commutative probability [13,
1, 2, 9, 10] that we are interested in are based on redefining independence as
a product satisfying certain natural axioms.

3 Tensor Categories and Independence

We will now define the notion of independence in the language of category
theory. The usual notion of independence for classical probability theory and
the independences classified in [13, 1, 2, 9, 10] will then be instances of this

2
general notion obtained by considering the category of classical probability
spaces or the category of algebraic probability spaces.
First we recall the definitions of a product, coproduct and a tensor product,
see also MacLane[7] for a more detailed introduction. Then we introduce
tensor categories with inclusions or projections. This notion is weaker than
that of a product or coproduct, but stronger than that of a tensor category.
It is exactly what we need to get an interesting notion of independence.
Definition 3.1 (See, e.g., Maclane[7]) A tuple (B1 Π B2 , π1 , π2 ) is called a
product or universal product of the objects B1 and B2 in the category C, if
for any object A ∈ Ob C and any morphisms f1 : A → B1 and f2 : A → B2
there exists a unique morphism h such that the following diagram commutes,

A II
vv II f
f1
vvv h II2
vvv II
I$
zvv 
B1 o π1 B1 Π B2 π2 / B2 .

An object K is called terminal, if for all objects A ∈ Ob C there exists exactly

one morphism from A to K.
The product of two objects is unique up to isomorphism, if it exists. Further-
more, the operation of taking products is commutative and associative up to
isomorphism and therefore, if a category has a terminal object and a product
for any two objects, then one can also define a product for any finite set of
objects.
The notion of coproduct is dual to that of a product, i.e., its defining
property can be obtained from that of the product by ‘reverting the arrows’.
The notion dual to terminal object is an initial object, i.e. an object K such
that for any object A of C there exists a unique morphism from K to A.
Let us now recall the definition of a tensor category.
Definition 3.2 A category (C, ) equipped with a bifunctor  : C × C → C,
called tensor product, that is associative up to a natural isomorphism
∼
=
αA,B,C : A(BC) → (AB)C, for all A, B, C ∈ Ob C,

and an element E that is, up to isomorphisms

∼
= ∼
=
λA : EA → A, and ρA : AE → A, for all A ∈ Ob C,

a unit for , is called a tensor category or monoidal category, if the pentagon

axiom

(AB)(CD)
5 SSS
αA,B,CD kkk SSSαAB,C,C
kkkkk SSS
kkk SSS
SS)
kkk



A B(CD) (AB)C D
O
idA αA,B,C αA,B,C idD


A (BC)D / A(BC)
D
αA,BC.D

and the triangle axiom

αA,E,C
A(EC) / (AE)C
LLL t
LLL tt
L ttt
idA λC LLL
ty t ρA idC
t
%
AC
are satisfied for all objects A, B, C, D of C.

3
 If a category has products or coproducts for all finite sets of objects, then
the universal property guarantees the existence of the isomorphisms α, λ, and
ρ that turn it into a tensor category.
In order to define a notion of independence we need less than a (co-)
product, but a little bit more than a tensor product. What we need are
inclusions or projections that allow us to view the objects A, B as subsystems
of their product AB.
Definition 3.3 A tensor category with projections (C, , π) is a tensor cat-
egory (C, ) equipped with two natural transformations π1 :  → P1 and
π2 :  → P2 , where the bifunctors P1 , P2 : C × C → C are defined by
P1 (B1 , B2 ) = B1 , P2 (B1 , B2 ) = B2 , on pairs of objects B1 , B2 of C, and sim-
ilarly on pairs of morphisms. In other words, for any pair of objects B1 , B2
there exist two morphisms πB1 : B1 B2 → B1 , πB2 : B1 B2 → B2 , such that
for any pair of morphisms f1 : A1 → B1 , f2 : A2 → B2 , the following diagram
commutes,
πA1 πA2
A1 o A1 A2 / A2

f1 f1 f2 f2
  
B1 o πB1 B1 B2 πB2
/ B2 .

Similarly, a tensor product with inclusions (C, , i) is a tensor category

(C, ) equipped with two natural transformations i1 : P1 →  and i2 : P2 → ,
i.e. for any pair of objects B1 , B2 there exist two morphisms iB1 : B1 →
B1 B2 , iB2 : B2 → B1 B2 , such that for any pair of morphisms f1 : A1 →
B1 , f2 : A2 → B2 , the following diagram commutes,

A1 / A1 A2 o A2
iA1 i A2

f1 f1 f2 f2
 iB1  iB2 
B1 / B1 B2 o B2 .

In a tensor category with projections or with inclusions we can define a

notion of independence for morphisms.
Definition 3.4 Let (C, , π) be a tensor category with projections. Two mor-
phism f1 : A → B1 and f2 : A → B2 with the same source A are called inde-
pendent (with respect to ), if there exists a morphism h : A → B1 B2 such
that the diagram

(1)
v A HHH
vvv
f1 HHf2
v h HH
vvv HH
{v  #
B1 o πB B1 B2 πB / B2
1 2

commutes.
In a tensor category with inclusions (C, , i), two morphisms f1 : A1 → B
and f2 : A2 → B with the same target B are called independent, if there exists
a morphism h : A1 A2 → B such that the diagram

; A cH (2)
v vv O HHH f2
f1
vv h
HH
vv HH
H
vv
B1 / B1 B2 o B2
i B1 i B2

commutes.

4
This definition can be extended in the obvious way to arbitrary sets of mor-
phisms.
If  is actually a product (or coproduct, resp.), then the universal property
in Definition 3.1 implies that for all pairs of morphisms with the same source
(or target, resp.) there exists even a unique morphism that makes diagram
(1) (or (2), resp.) commuting. Therefore in that case all pairs of morphism
with the same source (or target, resp.) are independent.
We will now consider several examples. We will show that for the category
of classical probability spaces we recover usual stochastic independence, if we
take the product of probability spaces, cf. Proposition 3.5.

3.1 Example: Independence in the Category of Clas-

sical Probability Spaces
The category Meas of measurable spaces consists of pairs (Ω, F), where Ω is
a set and F ⊆ P(Ω) a σ-algebra. The morphisms are the measurable maps.
This category has a product,

(Ω1 , F1 ) Π (Ω2 , F2 ) = (Ω1 × Ω2 , F1 ⊗ F2 )

where Ω1 × Ω2 is the Cartesian product of Ω1 and Ω2 , and F1 ⊗ F2 is the

smallest σ-algebra on Ω1 × Ω2 such that the canonical projections p1 : Ω1 ×
Ω2 → Ω1 and p2 : Ω1 × Ω2 → Ω2 are measurable.
The category of probability spaces Prob has as objects triples (Ω, F, P )
where (Ω, F) is a measurable space and P a probability measure on (Ω, F). A
morphism X : (Ω1 , F1 , P1 ) → (Ω1 , F2 , P2 ) is a measurable map X : (Ω1 , F1 ) →
(Ω1 , F2 ) such that

P1 ◦ X −1 = P2 .

This means that a random variable X : (Ω, F) → (E, E ) automatically be-

comes a morphism, if we equip (E, E ) with the measure

PX = P ◦ X −1

induced by X.
This category does not have universal products. But one can check that
the product of measures turns Prob into a tensor category,

(Ω1 , F1 , P1 ) ⊗ (Ω2 , F2 , P2 ) = (Ω1 × Ω2 , F1 ⊗ F2 , P1 ⊗ P2 ),

where P1 ⊗ P2 is determined by

(P1 ⊗ P2 )(M1 × M2 ) = P1 (M1 )P2 (M2 ),

for all M1 ∈ F1 , M2 ∈ F2 . It is even a tensor category with projections in

the sense of Definition 3.3 with the canonical projections p1 : (Ω1 × Ω2 , F1 ⊗
F2 , P1 ⊗ P2 ) → (Ω1 ,
F1 , P1 ), p2 : (Ω1 ×
Ω2 , F1 ⊗ F2 , P1 ⊗ P2 ) → (Ω2 , F2 , P2 )
given by p1 (ω1 , ω2 ) = ω1 , p2 (ω1 , ω2 ) = ω2 for ω1 ∈ Ω1 , ω2 ∈ Ω2 .
The notion of independence associated to this tensor product with projec-
tions is exactly the one used in probability.
Proposition 3.5 Two random variables X1 : (Ω, F, P ) → (E1 , E1 ) and X2 :
(Ω, F, P ) → (E2 , E2 ), defined on the same probability space (Ω, F, P ) and
with values in measurable spaces (E1 , E1 ) and (E2 , E2 ), are stochastically in-
dependent, if and only if they are independent in the sense of Definition 3.4 as
morphisms X1 : (Ω, F, P ) → (E1 , E1 , PX1 ) and X2 : (Ω, F, P ) → (E2 , E2 , PX2 )
of the tensor category with projections (Prob, ⊗, p).

5
 Proof. Assume that X1 and X2 are stochastically independent. We
have to find a morphism h : (Ω, F, P ) → (E1 × E2 , E1 ⊗ E2 , PX1 ⊗ PX2 ) such
that the diagram

(Ω, F, P )
jj TTTT
X1
jjjjjj TTTTX2
TTTT
jjj h
jjj TTTT
tjj  *
(E1 , E1 , PX1 ) o (E1 × E2 , E1 ⊗ E2 , PX1 ⊗ PX2 ) / (E2 , E2 , PX2 )
pE 1 pE 2

commutes. The only possible candidate is h(ω) = X1 (ω), X2 (ω) for all
ω ∈ Ω, the unique map that completes this diagram in the category of mea-
surable spaces and that exists due to the universal property of the product of
measurable spaces. This is a morphism in Prob, because we have





P h−1 (M1 × M2 ) = P X1−1 (M1 ) ∩ X2−1 (M2 ) = P X1−1 (M1 ) P X2−1 (M2 )
= PX1 (M1 )PX2 (M2 ) = (PX1 ⊗ PX2 )(M1 × M2 )

for all M1 ∈ E1 , M2 ∈ E2 , and therefore

P ◦ h−1 = PX1 ⊗ PX2 .

Conversely, if X1 and X2 are independent in the sense of Definition 3.4, then

the morphism
that makes the diagram commuting has to be again h : ω 7→
X1 (ω), X2 (ω) . This implies

P(X1 ,X2 ) = P ◦ h−1 = PX1 ⊗ PX2

and therefore




P X1−1 (M1 ) ∩ X2−1 (M2 ) = P X1−1 (M1 ) P X2−1 (M2 )

for all M1 ∈ E1 , M2 ∈ E2 .

3.2 Example: Tensor Independence in the Category

of Algebraic Probability Spaces
By the category of algebraic probability AlgProb spaces we denote the cate-
gory of associative unital algebras over C equipped with a unital linear func-
tional. A morphism j : (A1 , ϕ1 ) → (A2 , ϕ2 ) is a quantum random variable,
i.e. an algebra homomorphism j : A1 → A2 that preserves the unit and the
functional, i.e. j(1A1 ) = 1A2 and ϕ2 ◦ j = ϕ1 .
The tensor product we will consider on this category is just the usual
tensor product (A1 ⊗ A2 , ϕ1 ⊗ ϕ2 ), i.e. the algebra structure of A1 ⊗ A2 is
defined by

1A1 ⊗A2 = 1A1 ⊗ 1A2 ,

(a1 ⊗ a2 )(b1 ⊗ b1 ) = a 1 b2 ⊗ a 2 b2 ,

and the new functional is defined by

(ϕ1 ⊗ ϕ2 )(a1 ⊗ a2 ) = ϕ1 (a1 )ϕ2 (a2 ),

for all a1 , b1 ∈ A1 , a2 , b2 ∈ A2 .
This becomes a tensor category with inclusions with the inclusions defined
by

iA1 (a1 ) = a1 ⊗ 1A2 ,

iA2 (a2 ) = 1A1 ⊗ a2 ,

for a1 ∈ A1 , a2 ∈ A2 .

6
 One gets the category of ∗-algebraic probability spaces, if one assumes that
the underlying algebras have an involution and the functional are states, i.e.
are also positive. Then an involution is defined on A1 ⊗ A2 by (a1 ⊗ a2 )∗ =
a∗1 ⊗ a∗2 and ϕ1 ⊗ ϕ2 is again a state.
The notion of independence associated to this tensor product with inclu-
sions by Definition 3.4 is the usual notion of Bose or tensor independence used
in quantum probability, e.g., by Hudson and Parthasarathy.
Proposition 3.6 Two quantum random variables j1 : (B1 , ψ1 ) → (A, ϕ) and
j2 : (B2 , ψ2 ) → (A, ϕ), defined on algebraic probability spaces (B1 , ψ1 ), (B2 , ψ2 )
and with values in the same algebraic probability space (A, ϕ) are independent
if and only if the following two conditions are satisfied.
(i) The images of j1 and j2 commute, i.e.
 
j1 (a1 ), j2 (a2 ) = 0,

for all a1 ∈ A1 , a2 ∈ A2 .
(ii) ϕ satisfies the factorization property




ϕ j1 (a1 )j2 (a2 ) = ϕ j1 (a1 ) ϕ j2 (a2 ) ,

for all a1 ∈ A1 , a2 ∈ A2 .
We will not prove this Proposition since it can be obtained as a special case of
Proposition 3.7, if we equip the algebras with the trivial Z2 -grading A(0) = A,
A(1) = {0}.

3.3 Example: Fermi Independence

Let us now consider the category of Z2 -graded algebraic probability spaces
Z2 -AlgProb. The objects are pairs (A, ϕ) consisting of a Z2 -graded unital
algebra A = A(0) ⊕ A(1) and an even unital functional ϕ, i.e. ϕ|A(1) = 0.
The morphisms are random variables that don’t change the degree, i.e., for
j : (A1 , ϕ1 ) → (A2 , ϕ2 ), we have
(0) (0) (1) (1)
j(A1 ) ⊆ A2 and j(A1 ) ⊆ A2 .

The tensor product (A1 ⊗Z2 A2 , ϕ1 ⊗ ϕ2 ) = (A1 , ϕ1 ) ⊗Z2 (A2 , ϕ2 ) is defined

as follows. The algebra A1 ⊗Z2 A2 is the graded tensor product of A1 and
(0) (0) (1) (1) (1)
A2 , i.e. (A1 ⊗Z2 A2 )(0) = A1 ⊗ A2 ⊕ A1 ⊗ A2 , (A1 ⊗Z2 A2 )(1) = A1 ⊗
(0) (0) (1)
A2 ⊕ A1 ⊗ A2 , with the algebra structure given by

1A1 ⊗Z2 A2 = 1A1 ⊗ 1A2 ,

(a1 ⊗ b1 ) · (a2 ⊗ b2 ) = (−1)deg b1 deg a2 a1 b1 ⊗ a2 b2 ,

for all homogeneous elements a1 , b1 ∈ A1 , a2 , b2 ∈ A2 . The functional ϕ1 ⊗ ϕ2

is simply the tensor product, i.e. (ϕ1 ⊗ ϕ2 )(a1 ⊗ a2 ) = ϕ1 (a1 ) ⊗ ϕ2 (a2 ) for
all a1 ∈ A1 , a2 ∈ A2 . It is easy to see that ϕ1 ⊗ ϕ2 is again even, if ϕ1
and ϕ2 are even. The inclusions i1 : (A1 , ϕ1 ) → (A1 ⊗Z2 A2 , ϕ1 ⊗ ϕ2 ) and
i2 : (A2 , ϕ2 ) → (A1 ⊗Z2 A2 , ϕ1 ⊗ ϕ2 ) are defined by

i1 (a1 ) = a1 ⊗ 1A2 and i2 (a2 ) = 1A1 ⊗ a2 ,

for a1 ∈ A1 , a2 ∈ A2 .
If the underlying algebras are assumed to have an involution and the func-
tionals to be states, then the involution on the Z2 -graded tensor product is
defined by (a1 ⊗ a2 )∗ = (−1)deg a1 deg a2 a∗1 ⊗ a∗2 , this gives the category of
Z2 -graded ∗-algebraic probability spaces.
The notion of independence associated to this tensor category with inclu-
sions is called Fermi independence or anti-symmetric independence.

7
Proposition 3.7 Two random variables j1 : (B1 , ψ1 ) → (A, ϕ) and j2 :
(B2 , ψ2 ) → (A, ϕ), defined on two Z2 -graded algebraic probability spaces (B1 , ψ1 ),
(B2 , ψ2 ) and with values in the same Z2 -algebraic probability space (A, ϕ) are
independent if and only if the following two conditions are satisfied.
(i) The images of j1 and j2 satisfy the commutation relations

j2 (a2 )j1 (a1 ) = (−1)deg a1 deg a2 j1 (a1 )j2 (a2 )

for all homogeneous elements a1 ∈ A1 , a2 ∈ A2 .

(ii) ϕ satisfies the factorization property




ϕ j1 (a1 )j2 (a2 ) = ϕ j1 (a1 ) ϕ j2 (a2 ) ,

for all a1 ∈ A1 , a2 ∈ A2 .

Proof. The proof is similar to that of Proposition 3.5, we will only

outline it. It is clear that the morphism h : (B1 , ψ1 ) ⊗Z2 (B2 , ψ2 ) → (A, ϕ)
that makes the diagram in Definition 3.4 commuting, has to act on elements
of B1 ⊗ 1B2 and 1B1 ⊗ B2 as

h(b1 ⊗ 1B2 ) = j1 (b1 ) and h(1B1 ⊗ b2 ) = j2 (b2 ).

This extends to a homomorphism from (B1 , ψ1 ) ⊗Z2 (B2 , ψ2 ) to (A, ϕ), if and
only if the commutation relations are satisfied. And the resulting homomor-
phism is a quantum random variable, i.e. satisfies ϕ ◦ h = ψ1 ⊗ ψ2 , if and only
if the factorization property is satisfied.

4 Reduction of Independences
In this Section we will study the relations between different notions of inde-
pendence. Let us first recall the definition of a tensor functor.
Definition 4.1 (see, e.g., Section XI.2 in MacLane[7]) Let (C, ) and (C ′ , ′ )
be two tensor categories. A cotensor functor or comonoidal functor F :
(C, ) → (C ′ , ′ ) is an ordinary functor F : C → C ′ equipped with a morphism
F0 : F (EC ) → EC′ and a natural transformation F2 : F ( ·  · ) → F ( · )′ F ( · ),
i.e. morphisms F2 (A, B) : F (AB) → F (A)′ F (B) for all A, B ∈ Ob C that
are natural in A and B, such that the diagrams

F (αA,B,C )

F A(BC) / F (AB)C (3)

F2 (A,BC) F2 (AB,C)
 
F (A)′ F (BC) F (AB)′ F (C)

idF (A)  ′ F2 (B,C) F2 (A,B) ′ idF (C)




F (A)′ F (B)′ F (C) / F (A)′ F (B) ′ F (C)
α′F (A),F (B),F (C)

F2 (B,EC )
F (BEC ) / F (B)′ F (EC ) (4)

F (ρB ) idB  ′ F0
 
F (B) o ρF (B)
F (B)′ EC′

8
 F2 (EC ,B)
F (EC B) / F (EC )′ F (B) (5)

F (λB ) F0  ′ idB
 
F (B) o EC′ ′ F (B)
λF (B)

commute for all A, B, C ∈ Ob C.

We have reversed the direction of F0 and F2 in our definition. In the case of
a strong tensor functor, i.e. when all the morphisms are isomorphisms, our
definition of a cotensor functor is equivalent to the usual definition of a tensor
functor as, e.g., in MacLane[7].
The conditions are exactly what we need to get morphisms Fn (A1 , . . . , An ) :
F (A1  · · · An ) → F (A1 )′ · · · F (An ) for all finite sets {A1 , . . . , An } of ob-
jects of C such that, up to these morphisms, the functor F : (C, ) → (C ′ , ′ )
is a homomorphism.
For a reduction of independences we need a little bit more than a cotensor
functor.
Definition 4.2 Let (C, , i) and (C ′ , ′ , i′ ) be two tensor categories with in-
clusions and assume that C is a subcategory of C ′ . A reduction (F, J) of
the tensor product  to the tensor product ′ is a cotensor functor F :
(C, ) → (C ′ , ′ ) and a natural transformation J : idC → F , i.e. morphisms
jA : A → F (A) in C ′ for all objects A ∈ Ob C such that the diagram

JA
A / F (A)

f F (f )
 
B / F (B)
JB

commutes for all morphisms f : A → B in C.

Such a reduction provides us with a system of inclusions Jn (A1 , . . . , An ) =
Fn (A1 , . . . , An ) ◦ JA1 ···An : A1  · · · An → F (A1 )′ · · · ′ F (An ) with
′ ′
J1 (A) = JA that satisfies, e.g.,
Jn+m (A1 , . . . , An+m ) = F2 F (A1 ) · ·
·  F (An ),
F (An+1 )′ · · · ′ F (An+m ) ◦ Jn (A1 , . . . , An )Jm (An+1 , . . . , An+m ) for all
n, m ∈ N and A1 , . . . , An+m ∈ Ob C.
A reduction between two tensor categories with projections would consist
of a tensor functor F and a natural transformation P : F → id.
We have to extend our definition slightly. In our applications C will often
not be a subcategory of C ′ , but we have, e.g., a forgetful functor U from C
to C ′ that “forgets” an additional structure that C has. An example for this
situation is the reduction of Fermi independence to tensor independence in
following subsection. Here we have to forget the Z2 -grading of the objects of
Z2 -AlgProb to get objects of AlgProb. In this situation a reduction of the
tensor product with inclusions  to the tensor product with inclusions ′
is a tensor function F from (C, ) to (C ′ , ′ ) and a natural transformation
J : U → F.

4.1 Example: Bosonization of Fermi Independence

We will now define the bosonization of Fermi independence as a reduction
from (AlgProb, ⊗, i) to (Z2 -AlgProb, ⊗Z2 , i). We will need the group algebra
CZ2 of Z2 and the linear functional ε : CZ2 → C that arises as the linear
extension of the trivial representation of Z2 , i.e.

ε(1) = ε(g) = 1,

if we denote the even element of Z2 by 1 and the odd element by g.

9
 The underlying functor F : Z2 -AlgProb → AlgProb is given by

Ob Z2 -AlgProb ∋ (A, ϕ) 7→ (A ⊗Z2 CZ2 , ϕ ⊗ ε) ∈ Ob AlgProb,

F :
Mor Z2 -AlgProb ∋ f 7→ f ⊗ idCZ2 ∈ Mor AlgProb.

The unit element in both tensor categories is the one-dimensional unital

algebra C1 with the unique unital functional on it. Therefore F0 has to be a
morphism from F (C1) ∼ = CZ2 to C1. It is defined by F0 (1) = F0 (g) = 1.
The morphism F2 (A1 , A2 ) has to go from F (A ⊗Z2 B) = (A ⊗Z2 B) ⊗ CZ2
to F (A) ⊗ F (B) = (A ⊗Z2 CZ2 ) ⊗ (B ⊗Z2 CZ2 ). It is defined by

(a ⊗ 1) ⊗ (b ⊗ 1) if b is even,
a ⊗ b ⊗ 1 7→
(a ⊗ g) ⊗ (b ⊗ 1 if b is odd,

and

(a ⊗ g) ⊗ (b ⊗ g) if b is even,
a ⊗ b ⊗ g 7→
(a ⊗ 1) ⊗ (b ⊗ g) if b is odd,

for a ∈ A and homogeneous b ∈ B.

Finally, the inclusion JA : A → A ⊗Z2 CZ2 is defined by

JA (a) = a ⊗ 1

for all a ∈ A.
In this way we get inclusions Jn = Jn (A1 , . . . , An ) = Fn (A1 , . . . , An ) ◦
JA1 ⊗Z2 ...⊗Z2 An of the graded tensor product A1 ⊗Z2 · · · ⊗Z2 An into the usual
tensor product (A1 ⊗Z2 CZ2 ) ⊗ · · · ⊗ (An ⊗Z2 CZ2 ) which respect the states
and allow to reduce all calculations involving the graded tensor product to
calculations involving the usual tensor product on the bigger algebras F (A1 ) =
A1 ⊗Z2 CZ2 , . . . , F (An ) = An ⊗Z2 CZ2 . These inclusions are determined by

Jn (1 ⊗ · · · ⊗ 1 ⊗a ⊗ 1 ⊗ · · · ⊗ 1) = g̃ ⊗ · · · ⊗ g̃ ⊗ã ⊗ 1̃ ⊗ · · · ⊗ 1̃,
| {z } | {z } | {z } | {z }
k − 1 times n − k times k − 1 times n − k times

for a ∈ Ak , 1 ≤ k ≤ n, where we used the abbreviations

g̃ = 1 ⊗ g, ã = a ⊗ 1, 1̃ = 1 ⊗ 1.

5 Forgetful Functors, Coproducts, and Semi-

universal Products
We are mainly interested in different categories of algebraic probability spaces.
There objects are pairs consisting of an algebra A and a linear functional ϕ on
A. Typically, the algebra has some additional structure, e.g., an involution,
a unit, a grading, or a topology (it can be, e.g., a von Neumann algebra or
a C ∗ -algebra), and the functional behaves nicely with respect to this addi-
tional structure, i.e., it is positive, unital, respects the grading, continuous, or
normal. The morphisms are algebra homomorphisms, which leave the linear
functional invariant, i.e., j : (A, ϕ) → (B, ψ) satisfies

ϕ =ψ◦j

and behave also nicely w.r.t. to additional structure, i.e., they can be required
to be ∗-algebra homomorphisms, map the unit of A to the unit of B, respect
the grading, etc. We have already seen one example in Subsection 3.3.
The tensor product then has to specify a new algebra with a linear func-
tional and inclusions for every pair of of algebraic probability spaces. If the
category of algebras obtained from our algebraic probability space by forget-
ting the linear functional has a coproduct, then it is sufficient to consider the
case where the new algebra is the coproduct of the two algebras.

10
Proposition 5.1 Let (C, , i) be a tensor category with inclusions and ` F :
C → D a functor from C into another category D which has a coproduct and
an initial
` object E D . Then F is a tensor functor. The morphisms F2 (A, B) :
F (A) F (B) → F (AB) and F0 : ED → F (E) are those guaranteed by
the universal property of the coproduct and the initial object, i.e. F0 : ED →
F (E) is the unique morphism from ED to F (E) and F2 (A, B) is the unique
morphism that makes the diagram
F (iA )
F (A) / F (AB) o F (iB ) F (B)
MMM O qq
MMM qqq
M F 2 (A,B) q
iF (A) MMM& q
qq iF (B)
` xq
F (A) F (B)

commuting.

Proof. Using the universal property of the coproduct and the definition
of F2 , one shows that the triangles containing the F (A) in the center of the
diagram


` ` / F (A) ` F (B)
` F (C)
αF (A),F (B),F (C)
F (A) F (B) F (C)
hQQQ mm6
` QQ mmm `
idF (A) F2 (B,C) iF (A) iF (A) 2 F (A,B) idF (C)
mm QQQ
m QQ
` mm `
F (A) F (BC) o iF (A) F (A) Q / F (AB) F (C)
mmm QQQ
m m Q
F2 (A,BC) F (i ) F (iAQ
) F2 (AB,C)
m mm A QQQ
 vm

m Q( 

F A(BC) / F (AB)C
F (αA,B,C )

` `
commute (where the morphism from F (A) to F (AB) F (C) is F (iA ) idF (C) ),
and therefore that the morphisms

corresponding to all the different paths
form F (A) to F (AB)C coincide. Since we can get similar diagrams
with F (B) and`F (C), it`follows
from the universal property of the triple co-
product F (A)
` F (B)
` F (C)

that there exists

only a unique morphism
from F (A) F (B) F (C) to F (AB)C and therefore that the whole
diagram commutes.
The commutativity of the two diagrams involving the unit elements can
be shown similarly.
Let C now be a category of algebraic probability spaces and F the functor
that maps a pair (A, ϕ) to the algebra A, i.e., that “forgets” the linear func-
tional ϕ. Suppose that C is equipped
` with a tensor product  with inclusions
and that F (C) has a coproduct . Let (A, ϕ), (B, ψ) be two algebraic proba-
bility spaces in C, we will denote the pair (A, ϕ)(B, `ψ) also by (AB, ϕψ).
By Proposition 5.1 we have morphisms F2 (A, ` B) : A B → AB that define
a natural transformation from the bifunctor to the bifunctor . With these
morphisms we can define a new tensor product  e with inclusions by
 a 
e
(A, ϕ)(B, ψ) = A B, (ϕψ) ◦ F2 (A, B) .

The inclusions are those defined by the coproduct.

Proposition 5.2 If two random variables f1 : (A1 , ϕ1 ) → (B, ψ) and f1 :
(A1 , ϕ1 ) → (B, ψ) are independent with respect to , then they are also inde-
pendent with respect to .e

Proof. If f1 and f2 are independent with respect to , then there exists

a random variable h : (A1 A2 , ϕ1 ϕ2 ) → (B, ψ) that
`makes diagram (2) in
˜ 2 ) → (B, ψ)
Definition 3.4 commuting. Then h ◦ F2 (A1 , A2 ) : (A1 A2 , ϕ1 ϕ
makes the corresponding diagram for ˜ commuting.

11
The converse is not true. Consider the category of algebraic probability
` spaces
with the tensor product, see Subsection 3.2, and take B = A1 A2 and ψ =
(ϕ1 ⊗ ϕ2 ) ◦ F2 (A1 , A2 ). The canonical inclusions iA1 : (A1 , ϕ1 ) → (B, ψ) and
˜ but not with respect to the
iA2 : (A2 , ϕ2 ) → (B, ψ) are independent w.r.t. ⊗, `
tensor product itself, because their images do not commute in B = A1 A2 .
We will call a tensor product with inclusions in a category of quantum
probability spaces semi-universal, if it is equal to the coproduct of the corre-
sponding category of algebras on the algebras. The preceding discussion shows
that every tensor product on the category of algebraic quantum probability
spaces AlgProb has a quasi-universal version.

6 The Classification of Independences in the

Category of Algebraic Probability Spaces
We will now reformulate the classification by Muraki[10] and by Ben Ghorbal
and Schürmann[2, 1] in terms of semi-universal tensor products with inclusions
for the category of algebraic probability spaces AlgProb.
In order to define a semi-universal tensor product with inclusions on AlgProb
one needs a map that associates to a pair of unital functionals (ϕ1 , ϕ2 ) on
`two
algebras A1 and A2 a unital functional ϕ1 · ϕ2 on the free product A1 A2
(with identification of the units) of A1 and A2 in such a way that the bifunctor
a
 : (A1 , ϕ1 ) × (A2 , ϕ1 ) 7→ (A1 A2 , ϕ1 · ϕ2 )
`
satisfies all the necessary axioms. Since  is equal to the coproduct on the
algebras, we don’t have a choice for the isomorphisms α, λ, ρ implementing
the associativity and the left and right unit property, we have to take the
ones following from the universal property of the coproduct. The inclusions
and the action of  on the morphisms also have to be the ones given by the
coproduct.
The associativity gives us the condition


(ϕ1 · ϕ2 ) · ϕ3 ◦ αA1 ,A2 ,A3 = ϕ1 · (ϕ2 · ϕ3 ), (6)

for all (A1 , ϕ1 ), (A2 , ϕ2 ), (A3 , ϕ3 ) in AlgProb. Denote the unique unital func-
tional on C1 by δ, then the unit properties are equivalent to

(ϕ · δ) ◦ ρA = ϕ and (δ · ϕ) ◦ λA = ϕ,

for all (A, ϕ) in AlgProb. The inclusions are random variables, if and only if

(ϕ1 · ϕ2 ) ◦ iA1 = ϕ1 and (ϕ1 · ϕ2 ) ◦ iA2 = ϕ2 (7)

for all (A1 , ϕ1 ), (A2 , ϕ2 ) in AlgProb. Finally, from the functoriality of  we

get the condition
a
(ϕ1 · ϕ2 ) ◦ (j1 j2 ) = (ϕ1 ◦ j1 ) · (ϕ2 ◦ j2 ) (8)

for all pairs of morphisms j1 : (B1 , ψ1 ) → (A1 , ϕ1 ), j2 : (B2 , ψ2 ) → (A2 , ϕ2 ) in

AlgProb.
Our Conditions (6), (7), and (8) are exactly the axioms (P2), (P3), and
(P4) in Ben Ghorbal and Schürmann[1], or the axioms (U2), the first part of
(U4), and (U3) in Muraki[10].
Theorem 6.1 (Muraki[10], Ben Ghorbal and Schürmann[2, 1]) There exist
exactly two semi-universal tensor products with inclusions on the category of
algebraic probability spaces AlgProb, namely the semi-universal version ⊗˜ of
the tensor product defined in Section 3.2 and the one associated to the free
product ∗ of states.

12
 Voiculescu’s[14] free product ϕ1 ∗ϕ2 of two unital functionals can be defined
recursively by
→
!
X m−♯I+1
Y Y
(ϕ1 ∗ ϕ2 )(a1 a2 · · · am ) = (−1) (ϕ1 ∗ ϕ2 ) ak ϕǫk (ak )
I${1,... ,m} k∈I k6∈I

`
for a typical element a1 a2 · · · am ∈ A1 A2 , with ak ∈ Aǫk , ǫ1 6= ǫ2 6= · · · 6=
ǫm , i.e. neighboring a’s don’t Q belong to the same algebra. ♯I denotes the
number of elements of I and → k∈I ak means that the a’s are to be multiplied
in the same order in Q which they
appear on the left-hand-side. We use the
→
convention (ϕ1 ∗ ϕ2 ) k∈∅ ak = 1.
Ben Ghorbal and Schürmann[2, 1] and Muraki[10] have also considered
the category of non-unital algebraic probability nuAlgProb consisting of pairs
(A, ϕ) of a not necessarily unital algebra A and a linear functional ϕ. The
morphisms in this category are algebra homomorphisms that leave the func-
tional invariant. On this category we can define three more tensor products
with inclusions corresponding to the boolean product ⋄, the monotone product
⊲ and the anti-monotone product ⊳ of states. They can be defined by
m
Y
ϕ1 ⋄ ϕ2 (a1 a2 · · · am ) = ϕǫk (ak ),
k=1
 
→
Y Y
ϕ1 ⊲ ϕ2 (a1 a2 · · · am ) = ϕ1  ak  ϕ2 (ak ),
k:ǫk =1 k:ǫk =2
 
Y →
Y
ϕ1 ⊳ ϕ2 (a1 a2 · · · am ) = ϕ1 (ak ) ϕ2  ak  ,
k:ǫk =1 k:ǫk =2

`
for a1 a2 · · · am ∈ A1 A2 , ak ∈ Aǫk , ǫ1 6= ǫ2`
6= · · · 6= ǫm , i.e. neighboring a’s
don’t belong to the same algebra. Note that denotes here the free product
without units, the coproduct in the category of not necessarily unital algebras.
For the classification in the non-unital case, Muraki imposes the additional
condition

(ϕ1 · ϕ2 )(a1 a2 ) = ϕǫ1 (a1 )ϕǫ2 (a2 ) (9)


for all (ǫ1 , ǫ2 ) ∈ (1, 2), (2, 1) , a1 ∈ Aǫ1 , a2 ∈ Aǫ2 .
Theorem 6.2 (Muraki[10]) There exist exactly five semi-universal tensor prod-
ucts with inclusions satisfying (9) on the category of non-unital algebraic prob-
ability spaces nuAlgProb, namely the semi-universal version ⊗ ˜ of the tensor
product defined in Section 3.2 and the ones associated to the free product ∗,
the boolean product ⋄, the monotone product ⊲ and the anti-monotone product
⊳.
`
The monotone ` and the anti-monotone are not symmetric, i.e. (A1 A2 , ϕ1 ⊲
ϕ2 ) and (A2 A2 , ϕ2 ⊲ ϕ1 ) are not isomorphic in general. Actually, the anti-
monotone product is simply the mirror image of the monotone product,
a a
(A1 A2 , ϕ1 ⊲ ϕ2 ) ∼
= (A2 A1 , ϕ2 ⊳ ϕ1 )

for all (A1 , ϕ1 ), (A2 , ϕ2 ) in the category of non-unital algebraic probability

spaces. The other three products are symmetric.
At least in the symmetric setting of Ben Ghorbal and Schürmann, Con-
dition (9) is not essential. If one drops it and adds symmetry, one finds in
addition the degenerate product

ϕǫ1 (a1 ) if m = 1,
(ϕ1 · ϕ2 )(a1 a2 · · · am ) =
0 if m > 1.

13
and a whole family


ϕ1 •q ϕ2 = q (q −1 ϕ1 ) • (q −1 ϕ2 ) ,

parametrized by a complex number q ∈ C\{0}, for each of the three symmetric

˜ ∗, ⋄}.
products, • ∈ {⊗,

7 The Reduction of Boolean, Monotone, and

Anti-Monotone Independence to Tensor In-
dependence
We will now present the unification of tensor, monotone, anti-monotone, and
boolean independence of Franz[5] in our category theoretical framework. It
resembles closely the bosonization of Fermi independence in Subsection 4.1,
but the group Z2 has to be replaced by the semigroup M = {1, p} with two
elements, 1 · 1 = 1, 1 · p = p · 1 = p · p = p. We will need the linear functional
ε : CM → C with ε(1) = ε(p) = 1.
The underlying functor and the inclusions are the same for the reduction
of the boolean, the monotone and the anti-monotone ` product. They map the
algebra A of (A, ϕ) to the free product F (A) = Ã CM of the unitization Ã
of A and the group algebra CM of M . For the unital functional F (ϕ) we take
the boolean product ϕ̃ ⋄ ε of the unital extension ϕ̃ of ϕ with ε. The elements
of F (A) can be written as linear combinations of terms of the form

pα a1 p · · · pam pω

with m ∈ N, α, ω ∈ {0, 1}, a1 , . . . .am ∈ A, and F (ϕ) acts on them as

m
Y
F (ϕ)(pα a1 p · · · pam pω ) = ϕ(ak ).
k=1

The inclusion is simply

JA : A ∋ a 7→ a ∈ F (A).

The morphism F0 : F (C1) = CM → C1 is given by the trivial representation

of M , F0 (1) = F0 (p) = 1.
The only part of the reduction that is different for the three cases are the
morphisms
a a a
F2 (A1 , A2 ) : A1 A2 → F (A1 ) ⊗ F (A2 ) = (Ã CM ) ⊗ (Ã CM ).

We set

a⊗p if a ∈ A1 ,
F2B (A1 , A2 )(a) =
p⊗a if a ∈ A2 ,

for the boolean case,


a⊗p if a ∈ A1 ,
F2M (A1 , A2 )(a) =
1⊗a if a ∈ A2 ,

for the monotone case, and


a⊗1 if a ∈ A1 ,
F2AM (A1 , A2 )(a) =
p⊗a if a ∈ A2 ,

for the anti-monotone case.

14
 For the higher order inclusions Jn• = Fn• (A1 , . . . , An ) ◦ JA1 ` ··· ` An , • ∈
{B, M, AM}, one gets

JnB (a) = p⊗(k−1) ⊗ a ⊗ p⊗(n−k) ,

JnM (a) = 1⊗(k−1) ⊗ a ⊗ p⊗(n−k) ,
JnAM (a) = p⊗(k−1) ⊗ a ⊗ 1⊗(n−k) ,

if a ∈ Ak .
One can verify that this indeed defines reductions (F B , J), (F M , J), and
AM
(F , J) from the categories (nuAlgProb, ⋄, i), (nuAlgProb, ⊲, i), and (nuAlgProb, ⊳, i)
to (AlgProb, ⊗, i). The functor U : nuAlgProb → AlgProb mentioned at the
end of Section 4 is the unitization of the algebra and the unital extension of
the functional and the morphisms.
This reduces all calculations involving the boolean, monotone or anti-
monotone product to the tensor product. These constructions can also be
applied to reduce the quantum stochastic calculus on the boolean, monotone,
and anti-monotone Fock space to the boson Fock space. Furthermore, they
allow to reduce the theories of boolean, monotone, and anti-monotone Lévy
processes to Schürmann’s[11] theory of Lévy processes on involutive bialge-
bras, see Franz[5].

8 Conclusion
We have seen that the notion of independence in classical and in quantum
probability depends on a product structure which is weaker than a universal
product and stronger than a tensor product. We gave an abstract definition
of this kind of product, which we named tensor product with projections or
inclusions, and defined the notion of reduction between these products. We
showed how the bosonization of Fermi independence and the reduction of the
boolean, monotone, and anti-monotone independence to tensor independence
fit into this framework.
We also recalled the classifications of independence by Ben Ghorbal and
Schürmann[2, 1] and Muraki[10] and showed that their results classify in a
sense all tensor products with inclusions on the categories of algebraic prob-
ability spaces and non-unital algebraic probability spaces, or at least their
semi-universal versions.
There are two ways to get more than the five universal independences.
Either one can consider categories of algebraic probability spaces with addi-
tional structure, like for Fermi independence, cf. Subsection 3.3, and braided
independence, cf. Franz, Schott, and Schürmann[3], or one can weaken the
assumptions, drop, e.g., associativity, see Mlotkowski[8] and the references
therein. Romuald Lenczewski[6] has given a tensor construction for a fam-
ily of new products called m-free that are not associative, see also Franz and
Lenczewski[4]. His construction is particularly interesting, because in the limit
m → ∞ it approximates the free product. But it is not known, if a reduction
of the free product to the tensor product in the sense of Definition 4.2 exists.

References
[1] A. Ben Ghorbal and M. Schürmann. On the algebraic foundations of a
non-commutative probability theory. Prépublication 99/17, Institut E.
Cartan, Nancy, 1999, to appear in Math. Proc. Cambridge Philos. Soc.
[2] A. Ben Ghorbal. Fondements algébrique des probabilités quantiques et
calcul stochastique sur l’espace de Fock booléen. PhD thesis, Université
Henri Poincaré-Nancy 1, 2001.

15
 [3] U. Franz, R. Schott, and M. Schürmann. Braided independence and Lévy
processes on braided spaces. Prépublication Institut Elie Cartan 98/n 32,
1998.
[4] U. Franz, and R. Lenczewski. Limit theorems for the hierarchy of freeness.
Probab. Math. Stat. 19, p. 23-41, 1999.
[5] U. Franz. Unification of boolean, monotone, anti-monotone, and ten-
sor independence and Lévy processes. EMAU Greifswald Preprint-Reihe
Mathematik 4/2001, 2001, to appear in Mathematische Zeitschrift.
[6] R. Lenczewski. Unification of independence in quantum probability. Infin.
Dimens. Anal. Quantum Probab. Relat. Top., 1, no. 3, p. 383-405, 1998.
[7] S. MacLane. Categories for the Working Mathematician, Graduate texts
in mathematics, Vol. 5, Springer-Verlag, Berlin, 2nd edition, 1998.
[8] W. Mlotkowski. Free probability on algebras with infinitely many states.
Probab. Theory Related Fields 115, no. 4, p. 579–596, 1999.
[9] N. Muraki. The five independences as quasi-universal products. Infin.
Dimens. Anal., Quantum Probab. Relat. Top. 5, no. 1, p. 113-134, 2002.
[10] N. Muraki. The five independences as natural products. EMAU Greifs-
wald Preprint-Reihe Mathematik 3/2002, 2002.
[11] M. Schürmann. White Noise on Bialgebras. Lecture Notes in Math.,
Vol. 1544, Springer, Heidelberg, 1993.
[12] M. Schürmann. Direct sums of tensor products and non-commutative
independence. J. Funct. Anal. 133, no. 1, p. 1-9, 1995.
[13] R. Speicher. Universal products. In D. Voiculescu, editor, Free probability
theory. Papers from a workshop on random matrices and operator alge-
bra free products, Toronto, Canada, Mars 1995, Fields Inst. Commun.,
Vol. 12, p. 257-266. American Mathematical Society, 1997.
[14] D. Voiculescu, K. Dykema, and A. Nica. Free Random Variables. Amer-
ican Mathematical Society, Providence, RI, 1992.

View publication stats

16