Frederic Dambreville1

1Delegation

Generale pour lArmement, DGA/CEP/GIP/SRO, 16 Bis, Avenue Prieur

de la Cote dOr, Arcueil, F 94114, France

Pre-Boolean algebra, ordered DSmT

and DSm continuous models

Published in:
Florentin Smarandache & Jean Dezert (Editors)
Advances and Applications of DSmT for Information Fusion
(Collected works), Vol. II
American Research Press, Rehoboth, 2006
ISBN: 1-59973-000-6
Chapter V, pp. 131 - 153

Abstract: When implementing the DSmT, a difficulty may arise from the possible
huge dimension of hyper-power sets, which are indeed free structures. However, it is
possible to reduce the dimension of these structures by involving logical constraints.
In this chapter, the logical constraints will be related to a predefined order over the
logical propositions. The use of such orders and their resulting logical constraints
will ensure a great reduction of the model complexity. Such results will be applied to
the definition of continuous DSm models. In particular, a simplified description of
the continuous impreciseness is considered, based on impreciseness intervals of the
sensors. From this viewpoint, it is possible to manage the contradictions between
continuous sensors in a DSmT manner, while the complexity of the model stays
handleable.

5.1

Introduction

Recent advances [6] in the Dezert Smarandache Theory have shown that this theory was able
to handle the contradiction between propositions in a quite flexible way. This new theory has
been already applied in different domains; e.g.:
Data association in target tracking [9] ,
Environmental prediction [2] .
Although free DSm models are defined over hyper-power sets, which sizes evolve exponentially
with the number of atomic propositions, it appears that the manipulation of the fusion rule
is still manageable for practical problems reasonably well shaped. Moreover, the hybrid DSm
models are of lesser complexity.
If DSmT works well for discrete spaces, the manipulation of continuous DSm models is still an
unknown. Nevertheless, the management of continuous data is an issue of main interest. It is
necessary for implementing a true fusion engine for localization informations; and associated
with a principle of conditioning, it will be a main ingredient for implementing filters for the
localization. But a question first arises: what could be an hyper-power set for a continuous
DSm model? Such first issue does not arises so dramatically in Dempster Shafer Theory or for
Transfer Belief Models [7]. In DST, a continuous proposition could just be a measurable subset.
On the other hand, a free DSm model, defined over an hyper-power set, will imply that any
pair of propositions will have a non empty intersection. This is disappointing, since the notion
of point (a minimal non empty proposition) does not exist anymore in an hyper-power set.

131

132

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

But even if it is possible to define a continuous propositional model in DST/TBM, the manipulation of continuous basic belief assignment is still an issue [4, 8]. In [4] , Ristic and Smets
proposed a restriction of the bba to intervals of IR . It was then possible to derive a mathematical relation between a continuous bba density and its Bel function.
In this chapter, the construction of continuous DSm models is proposed. This construction is
based on a constrained model, where the logical constraints are implied by the definition of an
order relation over the propositions.
A one-dimension DSm model will be implemented, where the definition of the basic belief assignment relies on a generalized notion of intervals. Although this construction has been fulfilled
on a different ground, it shares some surprising similarities with Ristic and Smets viewpoint.
As in [4], the bba will be seen as density defined over a 2-dimension measurable space. We will
be able to derive the Belief function from the basic belief assignment, by applying an integral
computation. At last, the conjunctive fusion operator, , is derived by a rather simple integral
computation.
Section 5.2 makes a quick introduction of the Dezert Smarandache Theory. Section 5.3 is about
ordered DSm models. In section 5.4, a continuous DSm model is defined. This method is restricted to only one dimension. The related computation methods are detailed. In section 5.5,
our algorithmic implementation is described and an example of computation is given. The paper
is then concluded.

5.2

A short introduction to the DSmT

The theory and its meaning are widely explained in [6]. However, we will particularly focus on
the notion of hyper-power sets, since this notion is fundamental subsequently.
The Dezert Smarandache Theory belongs to the family of Evidence Theories. As the Dempster Shafer Theory [3] [5] or the Transferable Belief Models [7], the DSmT is a framework for
fusing belief informations, originating from independent sensors. However, free DSm models
are defined over Hyper-power sets, which are fully open-world extensions of sets. It is possible
to restrict this full open-world hypothesis by adding propositional constraints, resulting in the
definition of an hybrid Dezert Smarandache model.
The notion of hyper-power set is thus a fundamental ingredient of the DSmT. Hyper-power
sets could be considered as a free pre-Boolean algebra. As these structures will be of main
importance subsequently, the next sections are devoted to introduce them in details. As a
prerequisite, the notion of Boolean algebra is quickly introduced now.

5.2. A SHORT INTRODUCTION TO THE DSMT

5.2.1

133

Boolean algebra

Definition.

A Boolean algebra is a sextuple (, , , , , >) such that:

is a set, called set of propositions,

, > are specific propositions of , respectively called false and true,

: is a unary operator,
: and : are binary operators,

and verifying the following properties:

A1. and are commutative:

, , = and = ,
A2. and are associative:
, , , ( ) = ( ) and ( ) = ( ) ,
A3. > is neutral for and is neutral for :
, > = and = ,
A4. and are distributive for each other:
, , , ( ) = ( ) ( ) and ( ) = ( ) ( ) ,
A5. defines the complement of any proposition:
, = and = > .
The Boolean algebra (, , , , , >) will be also referred to as the Boolean algebra , the
structure being thus implied. An order relation is defined over by:
, ,
Fundamental examples.
gebra.

= .

The following examples are two main conceptions of Boolean al-

Example 1. Let be a set and P() be the set of
its subsets. For any A , denote
A = \ A its complement. Then P(), , , , , is a Boolean algebra.
The proof is immediate by verifying the properties A1 to A5.

ni . Let = { , . . . , }
Example 2. For any i {1, . . . , n}, let i = {0, 1}i1
1
n

{0} {0, 1}
n
n
and denote = , > = {0, 1} and B() = P {0, 1} . Define the operators , and by

= , = and = > \ for any , B() . Then B(), , , , , >
is a Boolean algebra.

The second example seems just like a rewriting of the first one, but it is of the most importance. It is called the free Boolean algebra generated by the set of atomic propositions .
Figure 5.1 shows the structure of such algebra, when n = 2. The free Boolean algebra B() is
deeply related to the classical propositional logic: it gives the (logical) equivalence classes of the
propositions generated from the atomic propositions of .

Although we give here an explicit
n
definition of B() by means of its binary coding P {0, 1} , the truly rigorous definition of B()
is made by means of the logical equivalence (which is out of the scope of this presentation).
Thus, the binary coding of the atomic propositions i is only implied.

134

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

ab

a b

a b

b
a

>
a b

a b

a b

Figure 5.1: Boolean algebra B {a, b} ; (partial)
Fundamental proposition.
Proposition 3. Any

Boolean algebra is isomorph to a Boolean algebra derived from a set, i.e.
P(), , , , , .
Proofs should be found in any good reference; see also [1].

5.2.2

Hyper-power sets

Definition of hyper-power

set. Lets consider a finite set of atomic propositions, and
denote B(), , , , , > the free Boolean
generated by . For
 any P(),

W algebra
V
define (), element of B(), by1 () = . The set < >= () P()
is called hyper-power set generated by .


It is noticed that both = () and > = P() are elements of < >. Figure 5.2 shows
the structure of the hyper-power set, when n = 2. Typically, it appears that the elements of
the hyper-power set are built only from -free components.
_ _ _ _ _






_ _ _ _ _

_ _ _






_ _ _ _ _ 

ab


_ _ _ _ _

_ _ _



_ _ _




_ _ _

ab

>

Figure 5.2: Hyper-power set < a, b >= {, a b, a, b, a b, >}

Example 3. Hyper-power set generated by = {a, b, c}.


< a, b, c >= , a, b, c, a b c, a b, b c, c a, a b c, >
a b, b c, c a, (a b) c, (b c) a, (c a) b,

(a b) c, (b c) a, (c a) b, (a b) (b c) (c a)

The following table associates some P() to their related hyper-power element ().
1

It is assumed

= and

= >.

5.2. A SHORT INTRODUCTION TO THE DSMT

135

This table is partial; there is indeed 256 possible choices for . It appears that is not
one-to-one:

()

{}
>


a

bc
{a};
{b};
{c}


(a

b)

(b

c) (c a)
{a,
b};
{b,
c};
{c,
a}


{a,c}; {b, c}; {a,
b, c}
(a c) (b c) (a b c)
{a, c}; {b, c}
(a c) (b c)

reduced form in < >

>
abc
(a b) (b c) (c a)
(a b) c
(a b) c

Remark. In the DSmT book 1 [6], the hyper-power sets have been defined by means of the
Smarandache encoding. Our definition is quite related to this encoding. In fact this encoding
is just implied in the definition of .
Hyper-power set as a free pre-Boolean algebra. It is easy to verify on example 3 that
< > is left unchanged by any application of the operators and . For example:


(a b) (b c) a = (a b b c) (a b a) = a b .

This result is formalized by the following proposition.

Proposition 4. Let , < >. Then < > and < >.
Proof. Let , < >.
There are P() and P() such that = () and = () .
By applying the definition of , it comes immediately:
() () =

It is also deduced:
() () =

_ ^

_ ^

By applying the distributivity, it comes:

!
_ _ ^  ^ 

() () =

222

(,)




Then () () = () , with = (, ) .

Corollary and definition. Proposition 4 implies

that and infer inner operations within
< > . As a consequence, < >, , , , > is an algebraic structure by itself. Since it does
not contains the negation , this structure is called the free pre-Boolean algebra generated by
.

136

5.2.3

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

Pre-Boolean algebra

Generality. Typically, a free algebra is an algebra where the only constraints are the intrinsic constraints which characterize its fundamental structures. For example in a free Boolean
algebra, the only constraints are A1 to A5, and there are no other constraints put on the
propositions. But conversely, it is indeed possible to derive any algebra by constraining its
free counterpart. This will be our approach for defining pre-Boolean algebra in general: a preBoolean algebra will be a constrained free pre-Boolean algebra. Constraining a structure is a
quite intuitive notion. However, a precise mathematical definition needs the abstract notion of
equivalence relations and classes. Let us start with the intuition by introducing an example.
Example 4. Pre-Boolean algebra generated by = {a, b, c} and constrained by a b = a c
and a c = b c.
For coherence with forthcoming
notations, these constraints
will be designated by using the set


of propositional pairs = (a b, a c), (a c, b c) .
The idea is to start from the free pre-Boolean algebra < a, b, c >, propagate the constraints, and
then reduce the propositions identified by the constraints.
It is first deduced a b = a c = b c = a b c.
It follows (a b) c = c, (b c) a = a and (c a) b = b.
Also holds (a b) c = (b c) a = (c a) b = (a b) (b c) (c a) = a b c .
By discarding these cases from the free structure < a, b, c >, it comes the following constrained
pre-Boolean algebra:


< a, b, c > = , a b c, a, b, c, a b, b c, c a, a b c, >

Of course, it is necessary to show that there is actually no further reduction in < a, b, c > .
This is done by explicating a model; for example the structure of figure 5.3.
2_ _ _ _ _
2

1 _ _211
_
_ _ _

1
1
11

1
2
1  2
1  2

22
22

 _ _
1 _ _211
_
_


11
1

1 2
1
1  2
1  2

22

22

1 _ _211
_

11
1

12
1

2


1
2
1  22

_2 _ _ _ _
2

2
11

11

11

2

11  22
11  22



abc

ab

bc

2_ _ _ _ _
2

1_ _ _211

1

1
11

1
2

1  22
1  22

22
22
2


 _ _
11
_

11

11

2

11  2
11  2



ca

2_ _ _ _ _
2

2
11
_ _ _
11

11

11  2 2
11  2


22
22
2



11

11

11

11 222
11  22



abc

Figure 5.3: Pre-Boolean algebra < a, b, c > ; ( and > are omitted)
For the reader not familiar with the notion of equivalence classes, the following construction
is just a mathematical formalization of the constraint propagation which has been described in
example 4. Now, it is first introduced the notion of morphism between structures.

5.2. A SHORT INTRODUCTION TO THE DSMT

137

Magma. A (, , , >)-magma, also called magma for short, is a quintuple (, , , , >)

where is a set of propositions, and are binary operators on , and and > are two
elements of .
The magma (, , , , >) may also be referred to as the magma , the structure being thus
implied. Notice that an hyper-power set is a magma.
Morphism. Let (, , , , >) and (, , , , >) be two magma. A morphism from the
magma to the magma is a mapping from to such that:
( ) = () () and ( ) = () () ,
() = and (>) = > .
A morphism is an isomorphism if it is a bijective mapping. In such case, the magma and the
magma are said to be isomorph, which means that they share the same structure.
The notions of (, )-magma and of (, )-morphism are defined similarly by discarding and
>.
Propagation relation. Let < > be a free pre-Boolean algebra. Let < > < >
be a set of propositional pairs; for any pair (, ) is defined the constraint = . The
propagation relation associated to the constraints, and also denoted , is defined recursively by:
, for any < >,
If (, ) , then and ,
If and , then ,
If and , then ( )( ) and ( )( )
The relation is thus obtained by propagating the constraint over < >. It is obviously
reflexive, symmetric and transitive; it is an equivalence relation. An equivalence class for
contains propositions which are identical in regards to the constraints.
It is now time to define the pre-Boolean algebra.
Pre-Boolean algebra.
Proposition 5. Let be given a free pre-Boolean algebra < > and a set of propositional pairs
< > < > . Then, there is a magma < > and a morphism :< >< >
such that:
(


< > =< > ,
, < >, () = () .

The magma < > is called the pre-Boolean algebra generated by and constrained by the
constraints = where (, ) .




Proof. For any < >, define = < >
; this set is called the class of
for .
It is a well known fact, and the proof is immediate, that = or = for any

138

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

, < > ; in particular, = .

Now, assume = and = , that is and .
It comes ( )( ) and ( )( ).
As a consequence, ( ) = ( ) and ( ) = ( ) .
At last:




= and = ( ) = ( ) and ( ) = ( )

222

The proof is then concluded easily, by setting:

 

< > = < > ,

, < >, = ( ) and = ( ) ,

< >, () = .

From now on, the element (), where < >, will be denoted as if were an element of
< > . In particular, () = () will imply = in < > (but not in < >).
Proposition 6. Let be given a free pre-Boolean algebra < > and a set of propositional pairs
< > < > . Let < > and < >0 be pre-Boolean algebras generated by and
constrained by the family . Then < > and < >0 are isomorph.
Proof. Let :< >< > and 0
:< >< >0 be as defined in proposition 5.
For any <
>, define
0 () .

() =
0
Then, () = () implies () = 0 () .
By definition of 0 , it is derived and then () = () .
Thus, is one-to-one.
By definition, it is also implied that is onto.
222
This property thus says that there is a structural uniqueness of < > .
Example 5. Let us consider again the pre-Boolean algebra generated by = {a, b, c} and constrained by a b = a c and a c = b c. In this case, the mapping :< >< > is
defined by:






{} = {}, a, (b c) a = {a}, b, (c a) b = {b},






c, (a b) c = {c}, {a b c} = {a b c}, {>} = {>} ,






{a b} = {a b}, {b c} = {b c}, {c a} = {c a} ,

a b c, a b, b c, c a, (a b) c, (b c) a, (c a) b,

(a b) (b c) (c a) = {a b c} .

5.2. A SHORT INTRODUCTION TO THE DSMT

139

Between sets and hyper-power sets.

Proposition 7. The Boolean algebra P(), , , , , , considered as a (, , , >)-magma,
is isomorph to the pre-Boolean algebra < > , where is defined by:
!)
(
_



, >
.
= ( , ) , and 6=

W
V
Proof. Recall the notation () = for

any
S
TP() .
Define :< > P() by setting2 () = {} for any P() .
It is immediate that is a morphism.




Now, by definition of , () = () is equivalent to ()() .
The proof is then concluded by proposition 6.

222
Thus, sets, considered as Boolean algebra, and hyper-power sets are both extremal cases of the
notion of pre-Boolean algebra. But while hyper-power sets extend the structure of sets, hyperpower sets are more complex in structure and size than sets. A practical use of hyper-power
sets becomes quickly impossible. Pre-Boolean algebra however allows intermediate structures
between sets and hyper-power sets.
A specific kind of pre-Boolean algebra will be particularly interesting when defining the DSmT.
Such pre-Boolean algebra will forbid any interaction between the trivial propositions , > and
the other propositions. These algebra, called insulated pre-Boolean algebra, are characterized
now.
Insulated pre-Boolean algebra.
A pre-Boolean
algebra < > verifies the insulation


property if < > \{, >}) < > \{, >}) .

Proposition 8. Let < > a pre-Boolean algebra verifying the insulation property. Then
holds for any , < > :
(
= ( = or = ) ,
= > ( = > or = >) .

In other words, all propositions are independent with each other in a pre-Boolean algebra with
insulation property.
The proof is immediate, since it is impossible to obtain or > without involving or > in the constraints of . Examples 3 and example 4 verify the insulation property.
On the contrary, a non empty set does not.
Corollary and definition. Let < > be a pre-Boolean algebra, verifying the insulation property. Define 
 =< > \{, >} . The operators and restrict to   , and
  , , is an algebraic structure by itself, called insulated pre-Boolean algebra. This
structure is also referred to as the insulated pre-Boolean algebra   .
2

It is defined

= .

140

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

Proposition 9. Let < > and < >0 be pre-Boolean algebras with insulation properties.
Assume that the insulated pre-Boolean algebra   and  0 are (, )-isomorph.
Then < > and < >0 are isomorph.
Deduced from the insulation property.
All ingredients are now gathered for the definition of Dezert Smarandache models.

5.2.4

The free Dezert Smarandache Theory

Dezert Smarandache Model. Assume that is a finite set. A Dezert Smarandache model
(DSmm) is a pair (, m), where is a set of propositions and the basic belief assignment m is
a non negatively valued function defined over < > such that:
X
X
m() =
m() = 1 .
<>

The property

 m()

Belief Function.
is defined by:



= 1 implies that the propositions of are exhaustive.

Assume that is a finite set. The belief function Bel related to a bba m
< >, Bel() =

m() .

(5.1)

<>:

The equation (6.1) is invertible:

< >, m() = Bel()

m() .

<>:(

Fusion rule. Assume that is a finite set. For a given universe , and two basic belief
assignments m1 and m2 , associated to independent sensors, the fused basic belief assignment is
m1 m2 , defined by:
X
m1 m2 () =
m1 (1 )m2 (2 ) .
(5.2)
1 ,2 <>:1 2 =

Remarks. It appears obviously that the previous definitions could be equivalently restricted
to  , owing to the insulation properties.

From the insulation property ( 6= and 6= ) ( ) 6= and the definition of the

fusion rule, it appears also that these definitions could be generalized to any algebra < >
with the insulation property.

5.2.5

Extensions to any insulated pre-Boolean algebra

Let   be an insulated pre-Boolean algebra. The definition of bba m, belief Bel and
fusion is thus kept unchanged.
A basic belief assignment m is a non negatively valued function defined over  
such that:
X
m() = 1 .


5.3. ORDERED DSM MODEL

141

The belief function Bel related to a bba m is defined by:

  , Bel() =

m() .

 :

Being given two basic belief assignments m1 and m2 , the fused basic belief assignment
m1 m2 is defined by:
m1 m2 () =

m1 (1 )m2 (2 ) .

1 ,2  :1 2 =

These extended definitions will be applied subsequently.

5.3

Ordered DSm model

From now on, we are working only with insulated pre-Boolean structures.
In order to reduce the complexity of the free DSm model, it is necessary to introduce logical
constraints which will lower the size of the pre-Boolean algebra. Such constraints may appear
clearly in the hypotheses of the problem. In this case, constraints come naturally and approximations may not be required. However, when the model is too complex and there are no
explicit constraints for reducing this complexity, it is necessary to approximate the model by
introducing some new constraints. Two rules should be applied then:
Only weaken informations3 ; do not produce information from nothing,
minimize the information weakening.
First point guarantees that the approximation does not introduce false information. But some
significant informations (e.g. contradictions) are possibly missed. This drawback should be
avoided by second point.
In order to build a good approximation policy, some external knowledge, like distance or order
relation among the propositions could be used. Behind these relations will be assumed some
kind of distance between the informations: more are the informations distant, more are their
conjunctive combination valuable.

5.3.1

Ordered atomic propositions

Let (, ) be an ordered set of atomic propositions. This order relation is assumed to describe
the relative distance between the information. For example, the relation implies
that and are closer informations than and . Thus, the information contained in is
stronger than the information contained in . Of course, this comparison does not matter
when all the information is kept, but when approximations are necessary, it will be useful to be
able to choose the best information.
3

Typically, a constraint like = will weaken the information, by erasing from .

142

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

Sketchy example. Assume that 3 independent sensors are giving 3 measures about a continuous parameter, that is x, y and z. The parameters x, y, z are assumed to be real values,
not of the set IR but of its pre-Boolean extension (theoretical issues will be clarified later4 ).
The fused information could be formalized by the proposition x y z (in a DSmT viewpoint).
What happen if we want to reduce the information by removing a proposition. Do we keep
x y , y z or x z ? This is of course an information weakening. But it is possible that one
information is better than an other. At this stage, the order between the values x, y, z will be
involved. Assume for example that x y < z . It is clear that the proposition x z indicates a
greater contradiction than x y or y z . Thus, the proposition x z is the one which should
be kept! The discarding constraint x y z x y z = x z is implied then.

5.3.2

Associated pre-Boolean algebra and complexity.

In regard to the previous example, the insulated pre-Boolean algebra associated to the ordered
propositions (, ) is   , where is defined by:



= ( , ) , , and .

The following property give an approximative bound of the size of   in the case of a
total order.
Proposition 10. Assume that (, ) is totally ordered. Then,   is a substructure of
the set 2 .
proof. Since the order is total, first notice that the added constraints are:
, , , = min{, , } max{, , } .
Now, for any , define by5 :

It is noteworthy that:

and




, (1 , 2 ) 2 1 2



= (1 , 2 ) 2 1 min{, } and max{, } 2



= (1 , 2 ) 2 1 min{, , } and max{, , } 2 .

By defining m = min{, , } and M = max{, , } , it is deduced:

.
= m
M

(5.3)

and property (5.3).

Figure 5.4 illustrates the construction of ,
4

In particular, as we are working in a pre-Boolean algebra, x y makes sense and it is possible that x y 6=
even when x 6= y .
5
wherethesymbol, means equals by definition.

5.3. ORDERED DSM MODEL

O 2


 
 

143
O


 2 =1

/ 1





m
M


 







 = M


=m


= m
M

Figure 5.4: Construction of

with and , ie.:
Let A P(2 ) be generated by |
A=

n0

n
[

k=1

k k

k , k , k

A consequence of (5.3) is that A is an insulated pre-Boolean algebra which satisfies the

constraints of . Then, the mapping:

 A


nk
nk
n \
n ^
[
_
`:
k,l , where k,l
k,l 7

k=1 l=1

k=1 l=1

is an onto morphism of pre-Boolean algebra.

Now, let us prove that ` is a one-to-one morphism.
Lemma 11. Assume:
n
[

k=1

m



[
2
1

l1 l2
k k
l=1

where jk , lj .

Then:
k , l , min{1k , 2k } min{l1 , l2 } and max{1k , 2k } max{l1 , l2 }
and
k , l , 1k 2k l1 l2 .
Proof of lemma. Let k [[1, n]] .
Define m = min{1k , 2k } and M = max{1k , 2k } .


S
1 2
Then holds (m, M ) 1k 2k , implying (m, M ) m
l=1 l l .
Let l be such that (m, M ) l1 l2 .
Then m min{l1 , l2 } and M max{l1 , l2 } .
At last, 1k 2k l1 l2 .
22

144

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

From inequalities min{1k , 2k } min{l1 , l2 } and max{1k , 2k } max{l1 , l2 } is also

deduced (1k 2k ) (l1 l2 ) = 1k 2k (definition of ) .
This property just means 1k 2k l1 l2 . It is lastly deduced:
Lemma 12. Assume:
n
[

k=1

Then:

m



[
l1 l2
1k 2k
l=1

n
_

1k

k=1

2k

m
_

l=1

where jk , lj .


l1 l2 .

From this lemma, it is deduced that ` is one to one.

At last ` is an isomorphism of pre-Boolean algebra, and   is a substructure of
2 .
222

5.3.3

General properties of the model

In the next section, the previous construction will be extended to the continuous case, ie.
(IR, ) . However, a strict logical manipulation of the propositions is not sufficient and instead
a measurable generalization of the model will be used. It has been seen that a proposition of
  could be described as a subset of 2 . In this subsection, the proposition model will
be characterized precisely. This characterization will be used and extended in the next section
to the continuous case.
Proposition 13. Let 
  .


Then ` () T , where T = (, ) 2 .
Proof. Obvious, since , T .

222
Definition 14. A subset 2 is increasing if and only if:
(, ) , , , (, ) .



Let U = T is increasing and 6= be the set of increasing non-empty subsets of
T . Notice that the intersection or the union of increasing non-empty subsets are increasing
non-empty subsets, so that (U, , ) is an insulated pre-Boolean algebra.



Proposition 15. For any
 choice
 of , ` ()   U .
When is finite, U = ` ()   .
Proof of . Obvious, since is increasing for any .

Proof of . Let  U and let (a, b) .

a and b and is increasing, it follows a
b .
Since a
b = (, ) 2 
S
W
At last, = (a,b) a
b =`
(a,b) a b .
W
Notice that (a,b) a b is actually defined, since is finite when is finite.

5.4. CONTINUOUS DSM MODEL

145

222
Figure 5.5 gives an example of increasing subsets, element of U.

When infinite -ing are allowed, notice that U may be considered as a model for   even

Figure 5.5: Example of increasing subset U

if is infinite. In the next section, the continuous pre-Boolean
algebra related to (IR, ) will


be modeled by the measurable increasing subsets of (x, y) IR2 x y .

5.4

Continuous DSm model

In this section, the case = IR is considered.

Typically, in a continuous model, it will be necessary to manipulate any measurable proposition, and for example intervals. It comes out that most intervals could not be obtained by
a finite logical combination of the atomic propositions, but rather
S by infinite combinations.
For example, considering the set formalism,
W it is obtained [a, b] = x[a,b] {x} , which suggests
the definition of the infinite disjunction x[a,b] x. It is known that infinite disjunctions are
difficult to handle in a logic. It is better to manipulate the models directly. The pre-Boolean
algebra to be constructed should verify the property x y z x y z = x z . As
discussed previously and since infinite disjunctions are allowed, a model for such algebra are
the measurable increasing subsets.

5.4.1

Measurable increasing subsets

A measurable subset A IR2 is a measurable increasing subset if:


(x, y) A , x y ,
(x, y) A , a x , b y , (a, b) A .
The set of measurable increasing subsets is denoted U.
Example. Let f : 
IR IR be a non decreasing
measurable mapping such that f (x) x for

any x IR. The set (x, y) IR2 f (x) y is a measurable increasing subset.
Points.

For any x IR, the measurable increasing subset x

is defined by:



x
= (a, b) IR2 a x b .

The set x
is of course a model for the point x IR within the pre-Boolean algebra (refer to
section 5.3).

146

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

Generalized intervals. A particular class of increasing subsets, the generalized intervals,

will be useful in the sequel.
For any x IR, the measurable sets x
` and x
are defined by:
(




x
` = (a, b) IR2 a b and x b ,



x
= (a, b) IR2 a b and a x .

The following properties are derived:

x
=x
`x
, x
`=

z and x
=

z[x,+[

z],x]

Moreover, for any x, y such that x y, it comes:

x
` y =

z .

z[x,y]

As a conclusion, the set x

`, x
and x
` y (with x y) are the respective models for the intervals
[x, +[ , ] , x] and [x, y] within the pre-Boolean algebra. Naturally, the quotation marks `
(opening) and(closing) are used respectively for opening and closing the intervals. Figure 5.6
illustrates various cases of interval models.
At last, the set x
` y, where x, y IR are not constrained, constitutes a generalized definition of
bO
x
`






 b=a





 
 
 
/a


x
` i.e. [x, +[

O


y 










y



x
` y  

/

x
i.e. ] , x]


 



 
 


x
` y

x
` y ; x y i.e. [x, y]










x
` y ; x > y

Figure 5.6: Interval models

the notion of interval. In the case x y, it works like classical interval, but in the case x > y,
it is obtained a new class of intervals with negative width (last case in figure 5.6). Whatever,
x
` y comes with a non empty inner, and may have a non zero measure.

The width = yx
` y could be considered as a measure of contradiction
2 of the interval x
associated with this proposition, while its center = x+y
2 should be considered as its median
value. The interpretation of the measure of contradiction is left to the human. Typically, a
possible interpretation could be:
< 0 means contradictory informations,
= 0 means exact informations,
> 0 means imprecise informations.

5.4. CONTINUOUS DSM MODEL

147

It is also noteworthy that the set of generalized intervals

I = {`
x y/x, y IR}
is left unchanged by the operator , as seen in the following proposition 16 :
Proposition 16 (Stability). Let x1 , x2 , y1 , y2 IR .
Define x = max{x

1 , x2 } and

y = min{y1 , y2 } .
Then x
`1 y1 x
`2 y2 = x
` y .
Proof is obvious.

This last property make possible the definition of basic belief assignment over generalized intervals only. This assumption is clearly necessary in order to reduce the complexity of the
evidence modeling. Behind this assumption is the idea that a continuous measure is described
by an imprecision/contradiction around the measured value. Such hypothesis has been made
by Smets and Ristic [4]. From now on, all the defined bba will be zeroed outside I. Now, since
I is invariant by , it is implied that all the bba which will be manipulated, from sensors or
after fusion, will be zeroed outside I. This makes the basic belief assignments equivalent to a
density over the 2-dimension space IR2 .

5.4.2

Definition and manipulation of the belief

The definitions of bba, belief and fusion result directly from section 5.2, but of course the bba
becomes density and the summations are replaced by integrations.
Basic Belief Assignment. As discussed previously, it is hypothesized that the measures
are characterized by a precision interval around the measured values. In addition, there is
an uncertainty about the measure which is translated into a basic belief assignment over the
precision intervals.
According to these hypotheses, a bba will be a non negatively valued function m defined over
U , zeroed outside I (set of generalized intervals), and such that:
Z
Belief function.
by:

x,yIR

m x
` y dxdy = 1 .

The function of belief, Bel, is defined for any measurable proposition U

Bel () =

x
`
y

In particular, for a generalized interval x

` y :


Bel x
` y =

m x
` y dxdy .

+ Z y

u=x

v=

m u
` v dudv .

148

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

Fusion rule. Being given two basic belief assignments m1 and m2 , the fused basic belief
assignment m1 m2 is defined by the curvilinear integral:
Z


m1 ()m2 () dC .
` y =
m1 m2 x
C={(,)/=`
x
y}

Now, from hypothesis it is assumed that mi is positive only for intervals of the form x
`i yi .
Proposition 16 implies:
(
x = max{x1 , x2 } ,
x
`1 y1 x
`2 y2 = x
` y where
y = min{y1 , y2 } .
It is then deduced:
Z +
Z x






m1 m2 x
m1 x
` y =
` y m2 x
`2 y2 dx2 dy2
x = y2 =y
Z x
Z +2




` y dx1 dy1
`1 y1 m2 x
m1 x
+
x = y1 =y
Z x1
Z +




m1 x
+
`1 y m2 x
` y2 dx1 dy2
x = y2 =y
Z +
Z x1




`2 y dx2 dy1 .
` y1 m2 x
m1 x
+
y1 =y

x2 =

In particular, it is now justified that a bba, from sensors or fused, will always be zeroed outside
I.

5.5

Implementation of the continuous model

Setting. In this implementation, the study has been restricted to the interval [1, 1] instead
of IR. The previous results still hold by truncating over [1, 1] . In particular, any bba m is
1 = {`
zeroed outside I1
x y/x, y [1, 1]} and its related belief function is defined by:


Bel x
` y =

1
u=x

y
v=1

m u
` v dudv ,

1 . The bba resulting of the fusion of two bbas m and m is

for any generalized interval of I1
1
2
defined by:
Z x
Z 1






m1 m2 x
` y =
m1 x
` y m2 x
`2 y2 dx2 dy2

+
+

x2 =1

x1 =1 y1 =y
Z 1
x

x1 =1 y2 =y
Z 1
x

x2 =1

y1 =y

y2 =y

m1 x
`1 y1 m2 x
` y dx1 dy1

m1 x
`1 y m2 x
` y2 dx1 dy2

m1 x
` y1 m2 x
`2 y dx2 dy1 .

5.6. CONCLUSION

149

Method. A theoretical computation of these integrals seems uneasy. An approximation of

the densities and of the integrals has been considered. More precisely, the densities have been
approximated by means of 2-dimension Chebyshev polynomials , which have several good properties:
The approximation grows quickly with the degree of the polynomial, without oscillation
phenomena,
The Chebyshev transform is quite related to the Fourier transform, which makes the
parameters of the polynomials really quickly computable by means of a Fast Fourier
Transform,
Integration is easy to compute.
In our tests, we have chosen a Chebyshev approximation of degree 128 128 , which is more
than sufficient for an almost exact computation.
Example. Two bba m1 and m2 have been constructed by normalizing the following functions
mm1 and mm2 defined over [1, 1]2 :




mm1 x
` y = exp (x + 1)2 y 2
and

` y = exp x2 (y 1)2 .
mm2 x

The fused bba m1 m2 and the respective belief function b1 , b2 , b1 b2 have been computed.
This computation has been instantaneous. All functions have been represented in the figures 5.7
to 5.14.
Interpretation. The bba m1 is a density centered around the interval [1, 0] , while m2 is a
density centered around [0, 1] . This explains why the belief b1 increases faster from the interval
[1, 1] to [1, 1] than from the interval [1, 1] to [1, 1] . And this property is of course inverted
for b2 .
A comparison of the fused bba m1 m2 with the initial bbas m1 and m2 makes apparent a
global forward move of the density. This just means that the fused bba is put on intervals
with less imprecision, and possibly on some intervals with negative width (ie. associated with
a degree of contradiction). Of course there is nothing surprising here, since information fusion
will reduce imprecision and produce some contradiction! It is also noticed that the fused bba
is centered around the interval [0, 0] . This result matches perfectly the fact that m1 and m2 ,
and their related sensors, put more belief respectively over the interval [1, 0] and the interval
[0, 1] ; and of course [1, 0] [0, 1] = [0, 0] .

5.6

Conclusion

A problem of continuous information fusion has been investigated and solved in the DSmT
paradigm. The conceived method is based on the generalization of the notion of hyper-power
set. It is versatile and is able to specify the typical various degrees of contradiction of a DSm
model. It has been implemented efficiently for a bounded continuous information. The work

150

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

pre-bba mm1
mm1

mm1(x,y)
1
0.8
0.6
0.4
0.2
0
-1

-0.5

x 0

0.5

1 -1

-0.5

0.5
y

Figure 5.7: Non normalized bba mm1

pre-bba mm2
mm2

mm2(x,y)
1
0.8
0.6
0.4
0.2
0
-1

-0.5

x 0

0.5

1 -1

-0.5

0.5
y

Figure 5.8: Non normalized bba mm2

bba m1
m1

m1(x,y)
1
0.8
0.6
0.4
0.2
0
-1

-0.5

x 0

0.5

1 -1

-0.5

0.5
y

Figure 5.9: Basic belief assignment m1

5.6. CONCLUSION

151

bba m2
m2

m2(x,y)
1
0.8
0.6
0.4
0.2
0
-1

-0.5

x 0

0.5

1 -1

-0.5

0.5
y

Figure 5.10: Basic belief assignment m2

belief b1
b1

b1(x,y)
1
0.8
0.6
0.4
0.2
0
-1

-0.5

x 0

0.5

1 -1

-0.5

0.5
y

Figure 5.11: Belief function b1

belief b2
b2

b2(x,y)
1
0.8
0.6
0.4
0.2
0
-1

-0.5

x 0

0.5

1 -1

-0.5

Figure 5.12: Belief function b2

0.5
y

152

PRE-BOOLEAN ALGEBRA, ORDERED DSMT AND DSM CONTINUOUS MODELS

bba m1+m2
m1+m2

m1+m2(x,y)
1
0.8
0.6
0.4
0.2
0
-1

-0.5

x 0

0.5

1 -1

-0.5

0.5
y

Figure 5.13: Fused bba m1 m2

belief b1+b2
b1+b2

b1+b2(x,y)
1
0.8
0.6
0.4
0.2
0
-1

-0.5

x 0

0.5

1 -1

-0.5

0.5
y

Figure 5.14: Fused bba b1 b2

is still prospective, but applications should be done in the future on localization problems. At
this time, the concept is restricted to one-dimension informations. However, works are now
accomplished in order to extend the method to multiple-dimensions domains.

5.7

References

[1]

Boolean algebra topic at the Wikipedia online encyclopedia.

http://en.wikipedia.org/wiki/Boolean algebra
http://en.wikipedia.org/wiki/Field of sets

[2]

Corgne S., Hubert-Moy L., Mercier G., Dezert J., Application of DSmT for Land Cover
Change Prediction, in [6].

[3]

Dempster A. P. , Upper and lower probabilities induced by a multiple valued mapping, Ann.
Math. Statistics, no. 38, pp. 325339, 1967.

5.7. REFERENCES

153

[4]

Ristic B, Smets Ph., Belief function theory on the continuous space with an application to
model based classification, in Proceedings of Information Processing and Management of
Uncertainty in Knowledge-Based Systems, IPMU 2004, Perugia, Italy, July 49, 2004.

[5]

Shafer G., A Mathematical Theory of Evidence, Princeton Univ. Press, Princeton, NJ,
1976.

[6]

Smarandache F., Dezert J. (Editors), Advances and Applications of DSmT for Information
Fusion, Collected Works, American Research Press, Rohoboth, June 2004.
http://www.gallup.unm.edu/smarandache/DSmT.htm

[7]

Smets Ph., Kennes R., The transferable belief model, Artificial Intelligence, Vol. 66, pp.
191234, 1994.

[8]

Strat T. M., Continuous belief functions for evidential reasoning, AAAI 1984, pp. 308313.

[9]

Tchamova A., Dezert J., Semerdjiev Tz. , Konstanttinova P., Target Tracking with Generalized Data Association based on the General DSm Rule of Combination, in Proceedings
of International Conference on Information Fusion, Fusion 2004, Stockholm, Sweden, June
28 July 1, 2004.