SCIENTIFIC PUBLICATIONS OF THE STATE UNIVERSITY OF NOVI PAZAR

SER. A: APPL. MATH. INFORM. AND MECH. vol. 3, 1 (2011), 19-26.

Introducing Afne Invariance to IFS
Lj. M. Koci c, E. Hadzieva, S. Gegovska Zajkova
Abstract: The original denition of the IFS with afne contractive mappings is an important
and handy tool for constructive approach to fractal sets. But, in spite of clear denition, the
concept of IFS does not allow many possibilities in the sense of modeling of such sets, typically
being fairly complicated. One step in direction of improving the concept of IFS consists in
introducing AIFS, a variant of IFS that permits afne invariance property which is vital from
the point of modeling. The theory is supported by comprehensive examples.
Keywords: fractals, IFS, AIFS, CAGD properties
1 Introduction
Let S ={R
m
; w
1
, w
2
, . . . , w
n
} be a hyperbolic iterated function system (IFS) dened on the
Euclidean metric space (R
m
, d
E
), where
w
i
(x) = A
i
x+b
i
, x R
m
, i = 1, 2, ..., n,
A
i
is an mm real matrix, b
i
is an m-dimensional real vector and s
i
< 1, i = 1, 2, . . . , n are
the corresponding Lipschitz factors of w
i
s. Let H (R
m
) be the space of nonempty compact
subsets of R
m
. Let h stand for Hausdorff metric, induced by d
E
, i.e.
h(A, B) = max
_
max
aA
min
bB
d
E
(a, b), max
bB
min
aA
d
E
(b, a)
_
, A, B H (R
m
),
then (H (R
m
), h) is a complete metric space ([1]). Associated with the IFS S is so called
Hutchinson operator W
S
dened on this space by
W
S
(B) =
n

i=1
w
i
(B) , B H (R
m
).
Manuscript received September 27 2010; revised December 22 2010; accepted February 22 2011
Lj. M. Koci c, is with the University of Nis, Faculty of Electronic Engineering, Nis, Serbia,; E. Hadzieva
and S. Gegovska Zajkova are with the University Saints Cyril and Methodius, Faculty of Electrical Engi-
neering and Information Technologies, Skopje, Macedonia,
19
20 Lj. M. Koci c, E. Hadzieva, S. Gegovska Zajkova
Note that W
S
is a contractive mapping on the space (H (R
m
), h) with contractivity factor
s = max
i
{ s
i
}, therefore (by the contraction mapping theorem) W
S
has a unique xed point
A, called the attractor of the IFS S .
Here are some technical notations. A (non-degenerate) m-dimensional simplex

P
m
(or just m-simplex) is the convex hull of a set P
m
, of m+1 afnely independent points
p
1
, p
2
, . . . , p
m+1
in the Euclidean space of dimension m,

P
m
= conv{p
1
, p
2
, . . . , p
m+1
}.
Analogously, the standard m-simplex is

T
m
= conv T
m
= conv{e
1
, e
2
, . . . , e
m+1
} R
m+1
where e
i
, i = 1, 2, . . . , m+1, are the unit vectors of (m+1)-dimensional orthogonal frame
(have all zero-coordinates except 1 at the i-th place). The afne space dened by the stan-
dard m-simplex will be denoted by V
m
, i.e. V
m
= aff {e
1
, e
2
, . . . , e
m+1
} R
m+1
.
Let S = [s
i j
]
m+1
i, j=1
be an (m+1)(m+1) row-stochastic real matrix (the matrix which rows
sum up to 1).
Denition 1.1 We refer to the linear mapping L : R
m+1
R
m+1
, such that L(x) = S
T
x
as the linear mapping associated with S.
Denition 1.2 Let

P
m
be a non-degenerate simplex and let {S
i
}
n
i=1
be a set of real square
nonsingular row-stochastic matrices of order m+1. The system(

P
m
) ={

P
m
; S
1
, S
2
, . . . , S
n
}
is called (hyperbolic) Afne invariant IFS (AIFS), provided that the linear mappings asso-
ciated with S
i
are contractions in (R
m
, d
E
) ([2]-[4]). The AIFS dened in the afne space
induced by the standard simplex {

T
m
; S
1
, S
2
, . . . , S
n
} is called the standard AIFS. The cor-
responding Hutchinson operator is W(P) =

i
_
S
T
i
P
_
, for any m-simplex P.
The concept of AIFS is closely related to afne geometry of barycentric (areal) co-
ordinates and subdivision phenomena ([2], [3], [4], [5], [6]). To clear the things up, the
following mathematical apparatus is of vital importance.
The afne space V
m
can be treated as a vector space with origin e
m+1
provided that
V
m
is parallel with the vector space V
m
= span{u
1
, u
2
, . . . , u
m
} R
m+1
, u
i
= e
i
e
m+1
,
i =1, 2, ..., m. So, fromnowon, the V
m
will be treated as a vector space unless it is otherwise
specied.
The set U
m
= {u
1
, u
2
, . . . , u
m
} of linearly independent vectors that spans the vector
space V
m
is neither orthogonal nor normalized, since

u
i
, u
j

=
_
1, i = j,
2, i = j.
By means
of the Gram-Schmidt procedure the set U
m
is transformed into an orthonormal basis V
m
=
{v
1
, v
2
, . . . , v
m
} of V
m
. This new orthonormal basis is represented by the m(m+1) matrix
V
m
=
_

_
v
11
v
12
v
1,m+1
v
21
v
22
v
2,m+1
.
.
.
.
.
.
v
m,1
v
m,2
v
m,m+1
_

_
=
_

_
v
T
1
v
T
2
.
.
.
v
T
m
_

_
. (1)
The relation between the barycentric coordinates of a point r V
m
w.r.t. the afne basis
{e
1
, e
2
, . . . , e
m+1
}, r = (
1
,
2
, . . . ,
m+1
), (
m+1
= 1
m
i=1

i
) and the coordinates in the
Introducing Afne Invariance to IFS 21
orthonormal basis V
m
, x = (x
1
, x
2
, . . . , x
m
), of the same point is given by
_

2
.
.
.

m
_

_
=
_

_
v
11
v
21
v
m,1
v
12
v
22
v
m,2
.
.
.
.
.
.
v
1,m
v
2,m
v
m,m
_

_
_

_
x
1
x
2
.
.
.
x
m
_

_
, i.e. r = Q
m
x, (2)
where Q
m
= V
m
T
, V
m
is the truncated matrix V
m
, i.e., the matrix V
m
given by (1), with the
last column dropped and r is a truncated vector r. The inverse transformation of (2) exists
and x = Q
1
m
r ([5]). In [5], the items of the matrices Q
m
= [q
i j
]
m
i, j=1
and Q
1
m
=
_
q

i j
_
m
i, j=1
are explicitly calculated. This result leads to the following
Lemma 1.1 The V
m
-coordinates of the vertices t
1
, t
2
, . . . , t
m+1
of the standard simplex

T
m
are given by the matrix T
T
m
= [t
1
t
2
. . . t
m+1
] =
_
Q
1
m
| 0

. More precisely, T
m
= [t
i j
]
m, m+1
i=1, j=1
,
where
t
i j
=
_

_
1

i(i+1)
, j < i m,

i+1
i
, i = j m,
0, i < j , i = m+1.
(3)
2 Relation between classical IFS and standard AIFS
Let x V
m
be transformed by an afne transform dened on V
m
, given by an mm matrix
A and translation vector b, so that
x

= Ax + b. (4)
By (2), r

= Q
m
x

, which by (4) gives r

= Q
m
x

= Q
m
(Ax + b). Since x = Q
1
m
r, the
penultimate formula is changed into
r

= Q
m
(AQ
1
m
r + b) = Q
m
AQ
1
m
r +Q
m
b. (5)
On the other hand, it is easy to see that, if r = [
1

2
. . .
m+1
]
T
is the full areal coordinate
vector, then r and r relate to each other as follows:
r = K
m
r, (6)
r = J
m
r + e
m+1
, (7)
where
K
m
=
_
I
m
| 0

and J
m
=
_
I
m
1
_
, (8)
22 Lj. M. Koci c, E. Hadzieva, S. Gegovska Zajkova
are block-matrices obtained by the identity matrix by adding a zero-column or (1) - row.
Also e
m+1
is the ultimate unit vector in R
m+1
. So, combining (5), (6) and (7) gives
r

=

Ar +

b, (9)
where

A = J
m
Q
m
AQ
1
m
K
m
(10)
is the (m+1)(m+1) matrix, and

b = J
m
Q
m
b+e
m+1
(11)
is the (m+1)-dimensional vector. Now the following theorem is valid.
Theorem 2.1 Let the pair (A, b) denes the afne mapping x Ax + b, x V
m
. Then
the corresponding linear transformation r S
T
r, r V
m
is dened by the matrix
S
T
=

A+[

b

b ...

b
. .
m+1
], (12)
where

A and

b are given by (10) and (11).
Proof. Introducing r = e
i
, i = 1, 2, ..., m+1 into (9), the standard simplex vertices
{e
1
, e
2
, . . . , e
m+1
} are being transformed into {
1
,
2
, . . . ,
m+1
},

i
=

Ae
i
+

b, i = 1, ..., m+1,
or, in matrix form
[
1

2
...
m+1
] =

A[e
1
e
2
... e
m+1
] +[

b

b ...

b
. .
m+1
].
Since [
1

2
...
m+1
] = S
T
and [e
1
e
2
... e
m+1
] = I
m+1
, the (12) follows.
Theorem 2.2 Given the row stochastic matrix S, dening the linear mapping r S
T
r,
r V
m
. The corresponding linear transformation x Ax + b, x V
m
is then given by
A = Q
1
m
K
m
S
T
J
m
Q
m
, b = Q
1
m
K
m
S
T
e
m+1
. (13)
Proof. Let r

= S
T
r. Combining it with (6) gives r

= K
m
S
T
r. On the other hand,
x

= Q
1
m
r

, so that x

= Q
1
m
K
m
S
T
r. By inserting (7) one gets
x

= Q
1
m
K
m
S
T
J
m
r + Q
1
m
K
m
S
T
e
m+1
and by (2) x

=
_
Q
1
m
K
m
S
T
J
m
Q
m
_
x + Q
1
m
K
m
S
T
e
m+1
, which gives (13).
Introducing Afne Invariance to IFS 23
3 Applications
The Theorems 2.1 and 2.2 are helpful in connecting two spaces, the designing space R
m
and the afne space V
m
. Suppose that our initial model consists of the set of simplices {P;
P
1
, . . . , P
n
} from R
m
, where P
i
= A
i
(P) is the image of the simplex P upon the afne
mapping A
i
, i = 1,. . . , n. This set accomplishes our construction and fully denes our
AIFS (See Def. 1.1 and 1.2), providing the set of linear mappings L
i
: x S
T
i
x, where
L
i
: P P
i
, is known. On the other hand, the set of subdivision matrices {S
1
, S
2
, . . . ,
S
n
}, can be found by using Theorem 2.1, once the transformation T
i
m
= A

i
(T
m
) is being
determined. Here, T
m
contains standard simplexs vertices, given as in Lemma 1.1, and
T = A(P).
Let P, P
i
be any pair of non-degenerate m-simplices and T is the standard m-simplex;
A and A
i
be afne operators dened by T = A(P), P
i
= A
i
(P); T
i
= A(P
i
) and

A
i
is
dened by T
i
=

A
i
(T). If the areal coordinates are in use, then the operatorsA, A
i
and

A
i
can be replaced by linear transformations represented by row-stochastic matrices S, S
i
and

S
i
respectively. The next theorem gives relations between afne and linear operators.
Theorem 3.1 The afne mappings A, A
i
and

A
i
are connected by the relation
A
i
A

A
1
i
A
1
= I, (14)
where I is identity mapping, and ( f g)() = g( f ()). In addition,
S
1

S
1
i
S S
i
= I
m+1
. (15)
Proof. By denition, T
i
= A(P
i
) = A(A
i
(P)). On the other hand, T
i
=

A
i
(T) or
T
i
= A
i
(A(P)). So, A(A
i
(P)) =A
i
(A(P)), wherefrom
A(P) =

A
1
i
(A(A
i
(P))) P =A
1
_

A
1
i
(A(A
i
(P)))

,
which, combined with supposition that P is an arbitrary simplex, yields (14). The relation
(15) is direct consequence of (14).
Example 3.1 (Variations on Sierpinski triangle). Consider a non-degenerate 4-simplex T =
[P
1
P
2
P
3
P
4
P
5
]
T
in R
4
and its images T
1
, T
2
and T
3
under three different linear mappings
performed by three row-stochastic matrices S
1
, S
2
and S
3
given in Fig. 1.
The case T = [(0, 0) (2, 1) (0, 3) (2, 1) (0, 0)]
T
is shown in Figure 1a (in all cases the
R
2
projections are used). Different congurations of the simplex T grossly inuence the
shape of the attractor A. The Fig. 1b is obtained by choosing T = [(4, 1) (2.75, 2.5) (3.5,
4.3) (2.25, 3.5) (1, 1)]
T
while the attractor at Fig. 1c is generated for T = [(6, 1) (6, 6)
(3.5, 6) (1, 6) (3, 1)]
T
. But, in all the cases, the basic structure of the fractal set, known as
Sierpinski triangle (or gasket).
24 Lj. M. Koci c, E. Hadzieva, S. Gegovska Zajkova
S
1
=
_

_
1 0 0 0 0
1
2
1
2
0 0 0
63
188
21
47
7
47
1
47
9
188
3
8
3
16
3
16
1
16
3
16
1
2
0 0 0
1
2
_

_
,
S
2
=
_

_
63
188
21
47
7
47
1
47
9
188
0
1
2
1
2
0 0
0 0 1 0 0
0 0
1
2
1
2
0
9
188
1
47
7
47
21
47
63
188
_

_
,
S
3
=
_

_
1
2
0 0 0
1
2
3
16
1
16
3
16
3
16
3
8
9
188
1
47
7
47
21
47
63
188
0 0 0
1
2
1
2
0 0 0 0 1
_

_
.
Fig. 1.
Example 3.2 (Self afne tilling of the domain)The AIFS method can be used for tessella-
tion of the nite subset of R
d
, bounded by a closed hyper-polyhedral surface . If has m
vertices and lies into d-dimensional space (d m1), then two possibilities may occur: If
d < m1, is a projection of m-simplex on R
d
; Otherwise, is m-simplex itself. Now,
suppose that the AIFS ={

T; S
1
, S
2
, . . . , S
n
} is given, and that T
i
= S
T
i
T (all i). Let the
AIFS is just touching, which means that

i

T
i
=

T, and

T
i

T
j
(all i, j) may contain
only common border points. Then, the k-th iteration of Hutchinson operator W (see Def.
1.3) applied on

T , W
k
(

T) gives self afne tilling of the domain of

T. Figure 2 illustrates
this process in three different settings of tilling the rectangle {P
1
, P
2
, P
3
, P
4
} in three parts
(rst two rows of pictures in Fig. 2) and four parts (the last row). The difference between
the rst and second partition dened by subdivision matrices S
1
, S
2
and S
3
, is in matrices
S
1
and S
3
. Namely, for the rst setting,
S
1
= [P
T
4
P
T
1
B
T
M
T
]
T
, S
2
= [P
T
2
B
T
M
T
P
T
3
]
T
, S
3
= [P
T
4
P
T
3
M
T
P
T
4
]
T
while for the second,
S
1
= [P
T
1
B
T
M
T
P
T
4
]
T
, S
2
= [P
T
2
B
T
M
T
P
T
3
]
T
, S
3
= [M
T
P
T
3
P
T
4
M
T
]
T
.
The last row is dened by the setting
S
1
= [P
T
1
M
T
P
T
4
P
T
1
]
T
, S
2
= [P
T
2
M
T
P
T
1
P
T
2
]
T
,
S
3
= [P
T
3
M
T
P
T
2
P
T
3
]
T
, S
4
= [P
T
4
M
T
P
T
3
P
T
4
]
T
.
Introducing Afne Invariance to IFS 25
Fig. 2.
The asymmetric subdivision of triangular areas in this case is caused by having two identical
rows in each matrix as a consequence of doubling the vertices.
4 Conclusion.
The main result of this note is given in Theorems 2.1 and 2.2, where two new formulas
are given for transforming afne into linear operators to connect IFS and AIFS, iterated
systems of contractive mappings that are used for dening and constructing fractal sets. A
by-result is Theorem 3.1, that gives a rule for both linear and afne mappings that transform
a design space into referential one and v.v. Two examples are elaborated. The rst one
is showing the exibility of the AIFS concept. The second indicates possible applications
that are not typically fractal in nature.
References
[1] M. F. BARNSLEY, Fractals Everywhere. Academic Press, San Diego, 1993.
[2] LJ. M. KOCI C, A.C. SIMONCELLI, Shape predictable IFS representations, In: Emergent
Nature, (M. M. Novak, ed.), World Scientic, 2002, pp. 435436.
26 Lj. M. Koci c, E. Hadzieva, S. Gegovska Zajkova
[3] LJ. M. KOCI C, A.C. SIMONCELLI, Towards free-form fractal modelling, In: Mathematical
Methods for Curves and Surfaces II, (M. Daehlen, T. Lyche and L. L. Schumaker, eds.),
Vanderbilt University Press, Nashville (TN.), 1998, pp. 287294.
[4] LJ. M. KOCI C, A.C. SIMONCELLI, Stochastic approach to afne invariant IFS, In: Prague
Stochastics98 (Proc. 6th Prague Symp., Aug. 23-28, M. Hruskova, P. Lachout and J.A.
Visek eds), Vol II, Charles Univ. and Academy of Sciences of Czech Republic, Union of
Czech Mathematicians and Physicists, Prague 1998, pp. 317320.
[5] LJ. M. KOCI C, S. GEGOVSKA-ZAJKOVA, E. BABA CE , Orthogonal decomposititon of frac-
tal sets, Approximation and Computation, Vol.42, (Gautschi, W., Mastroianni, G., Rassias,
T. M., eds.), Series: Springer Optimization and Its Application, 2010, pp. 145156.
[6] S. SCHAFAER, D. LEVIN, R. GOLDMAN, Subdivision Schemes and Attractors, Eurograph-
ics Symposium on Geometry Proceedings, (M. Desbrun, H. Pottmann, eds.), 2005.