Professional Documents
Culture Documents
Transformations Between Selfreferential Sets
Transformations Between Selfreferential Sets
Self-Referential Sets
Michael F. Barnsley
F := {X; f 1 , f 2 , . . . , f N }
is called a hyperbolic iterated function system or, briefly, an IFS.
A transformation f n : X X is strictly contractive when there is a number ln
[0, 1) such that d( f n (x), f n (y)) ln d(x, y) for all x, y X. The number ln is called
April 2009]
291
exists, is independent of x, and depends continuously on . Furthermore, the convergence to the limit is uniform in x, for x in any compact subset of X. See for example
[3, Theorem 3]. Let
AF = {F ( ) : }.
Then AF X is called the attractor of F . The continuous function
F : AF
is called the address function of F . We call F1 ({x}) = { : F ( ) = x} the set
of addresses of the point x AF .
Clearly AF is compact and nonempty, and has the property
AF = f 1 (AF ) f 2 (AF ) f N (AF ).
That is, the attractor of an IFS is a self-referential set.
If we define H(X) to be the set of nonempty compact subsets of X, and we define
F : H(X) H(X) by
(2.1)
for all S H(X), then AF can be characterized as the unique fixed point of F ; see [9,
Section 3.2], [8], and [19].
IFSs may be used to represent diverse subsets of R2 . For example, let A, B, and C
denote three noncollinear points in R2 . Let c denote a point on the line segment AB, let
a denote a point on the line segment BC, and let b denote a point on the line segment
C A, such that {a, b, c} {A, B, C} = ; see panel (i) of Figure 1.
Let f 1 : R2 R2 denote the unique affine transformation such that
f 1 (ABC) = ca B,
by which we mean that f 1 maps A to c, B to a, and C to B. Using the same notation,
let affine transformations f 2 , f 3 , and f 4 be the ones uniquely defined by
f 2 (ABC) = Cab, f 3 (ABC) = c Ab, and f 4 (ABC) = cab.
292
Figure 1. (i) The points used to define the affine transformations of the IFS F,, = {R2 ; f 1 , f 2 , f 3 , f 4 };
(ii) sketch of the attractor of F,, ; (iii) sketch of the attractor of the IFS {R2 ; f 1 , f 2 , f 3 }. Here = 0.65,
= 0.3, and = 0.4.
an
bn
1 +
1
2
3
4
12
+ 12
1
2
2
1
+
+
cn
dn
12 + 12 14
1
2
34 + 12 + 12 14
1
4
en
qn
0
1
2
1 + + 12
1
2
+ 12
0
1
0
293
The sketches in panels (ii) and (iii) of Figure 1 were computed using the chaos
game. At each iteration the function f n was selected with probability proportional to
the area of the triangle f n (ABC), for n = 1, 2, 3, 4.
In Section 5 we show how the chaos game may be modified to calculate examples
of the fractal transformations that are the subject of this article. Hopefully you will be
inspired to try this new application of the chaos game.
4. THE TOPS FUNCTION. So far we have covered material that is well known to
the fractal geometry community; indeed, some of it forms part of various undergraduate curricula.
Here we introduce the idea of the top of a self-referential set. This concept exploits
the fact that code space possesses an order relation, which allows us to assign a unique
address to each point of the attractor of an IFS. This provides us with a right-hand
inverse of the mapping F with the remarkable property that its domain is shift invariant.
We order the elements of according to
< iff k > k ,
where k is the least index for which k = k . This is a linear ordering, sometimes
called the lexicographic ordering.
Notice that all elements of are less than or equal to 1 = 11111 . . . and greater
than or equal to N = N N N N N . . . . Also, any pair of distinct elements of is such
that one member of the pair is strictly greater than the other. In particular, F1 ({x})
possesses a unique largest element: let 1 be the smallest first coordinate of any address
of x, let 2 be the smallest second coordinate of any address of x whose first coordinate
is 1 , etc; then the continuity of F : AF implies that = 1 2 is an address
of x.
Definition 2. Let F be a hyperbolic IFS with attractor AF and address function F :
AF . Let
F (x) = max{ : F ( ) = x}
for all x AF . Then
F := {F (x) : x AF }
is called the tops code space and
F : A F F
is called the tops function, for the IFS F .
In Section 7 we explain why F is a shift invariant subspace of .
Notice that the tops function F : AF F is one-to-one and onto. It provides a
right inverse to the address function; that is, F F is the identity on AF . The inverse
function,
F1 : F AF ,
is clearly one-to-one and onto; it is also continuous because it is the restriction of F to
F . However, F may not be continuous. For example, if AF is connected and consists
294
of more than one point, then F (AF ) is not pathwise connected. Let F denote the
closure of F , treated as a subset of the metric space (, d ). Let
F1 : F AF
denote the restriction of F to F or, equivalently, the continuous extension of F1 to
F . Then F1 is continuous and onto. Notice that the ranges of F1 and F1 are both
equal to AF because AF is closed.
5. FRACTAL TRANSFORMATIONS OF PICTURES. Here we use tops functions to construct what we call fractal transformations between self-referential sets.
We illustrate these transformations by applying them to pictures. This leads to some
questions.
Let
G = {Y; g1 , g2 , . . . , g N }
denote an IFS with the same number, N , of functions as F . The associated code space
is , the same as for F . We define a transformation from the attractor of F to the
attractor of G by composing the address function for G with the tops function for F .
Definition 3. Let F and G denote IFSs as above. The associated fractal transformation TF G : AF AG from the attractor of F to the attractor of G is defined by
TF G = G F .
How do such transformations behave? To illustrate transformations between subsets
of R2 we use pictures. We define a picture to be a function of the form
P : DP R2 C
where C is a color space. A picture (function) P assigns a unique color to each point
in its domain DP . For example we may have C ={0, 1, . . . , 255}3 and each point
of C may specify the red, green, and blue components of a color. A picture in the
nonmathematical sense may be thought of as a physical representation of the graph of
a picture function whose domain is contained in R2 .
Let TF G : AF AG be a fractal transformation as above. Let P be a picture whose
domain contains AG . Then P TF G is a picture whose domain is AF . We can obtain
insights into the nature of the fractal transformation TF G by comparing the pictures
P and P TF G , where P may be varied. It is straightforward to do this when the
underlying space is two-dimensional because we can use a variant of the chaos game
to compute the picture P TF G .
To see why this is so, we consider the IFS
K = {X Y ; kn , n = 1, 2, . . . , N }
where kn (x, y, ) = ( f n (x), gn (y), sn ( )) with sn ( ) = n (this means put the symbol n at the front of the string ). The metric d on the space X Y is defined
by
d((x, y, ), (x,
y , )) = max{dX (x, x),
dY (y, y ), d (, )}.
April 2009]
295
Then K has contractivity factor lK = max{lF , lG , 1/2} < 1. Its (unique) attractor is
AK = {(x, y, ) : , F ( ) = x, and G ( ) = y}
because this set is nonempty, compact, and obeys AK = n kn (AK ). In particular, the
graph of TF G is the same as
{(x, y) : (x, y, ) AK , for all (x, y , ) AK }.
It follows that we can use the chaos game to approximate AK and thence the graph
of TF G . In practice, when the underlying spaces are two-dimensional, we arrange coordinates so that the domain of the given picture P contains AG . Then we run the chaos
game, starting from an initial point (x0 , y0 , (0) ) AF AG , to yield a sequence
of points {(xk , yk , (k) )}. At each successive step we color the point xk AF with the
value P(yk ), unless it has already been assigned a color. If the point xk has already
been assigned a color, most recently at some earlier step j, then we compare ( j ) and
(k) ; when the latter is greater than the former, the color of xk is updated to P(yk ),
otherwise it is left at its old value.
Imagine that this process is carried out at finite precision on a digital computer.
After the kth step the point whose screen coordinates correspond to xk is plotted in its
latest color. To begin with we will see the picture changing, but after a while it will
stabilize. What do we see? Here are some illustrative examples.
Let
Figure 2. Examples of fractal transformations are illustrated using the three picture functions PG , PF , and
PH shown in the top row, from left to right. The domains of these functions are the attractors AG , AF , AH of
three IFSs G , F , and H, defined in the text. The bottom row illustrates the pictures PG TF G , PF TF F , and
PH TF H .
CF = {F1 ({x}) F : x AF }.
The address structure of an IFS is a certain partition of F . Let CG denote the
address structure of G . Let us write CF CG to mean that for each S CF there is
T CG such that S T . Notice that if CF = CG then F = G . Some examples of
address structures are given in Section 8.
Theorem 1. Let F and G be two hyperbolic IFSs such that CF CG . Then the fractal
transformation TF G is continuous. If CF = CG then TF G is a homeomorphism.
The proof relies on a standard result in topology, Lemma 2 below, which we present
in the context of metric spaces.
Lemma 1 (cf. [13, p. 194]). Let F : X Y be a continuous mapping from a compact
metric space X onto a metric space Y . Then S Y is open if and only if F 1 (S) X
is open.
Proof. If S Y is open then F 1 (S) X is open because F : X Y is continuous.
Suppose that F 1 (S) is open. Then X \F 1 (S) is closed. But a closed subset of a
April 2009]
297
compact metric space is compact. The continuity of F now implies that F(X \F 1 (S))
is compact and hence closed. But F(X \F 1 (S)) = Y \S. Hence S is open.
Lemma 2 (cf. [13, Prop. 7.4, p. 195]). Let F : X Y be a continuous mapping
from a compact metric space X onto a metric space Y . Let H : Y Z where Z is a
metric space. Let H F : X Z be continuous. Then H : Y Z is continuous.
Proof. Let O Z be open. Then (H F)1 (O) = F 1 (H 1 (O)) is open. But then
by Lemma 1 H 1 (O) is open. Hence H : Y Z is continuous.
Proof of Theorem 1. F is a compact metric space and F1 : F AF is continuous. Now look at the function
TF G F1 = G F F1 : F AG .
If F , then both and (F F1 )( ) belong to the same set in CF . Since CF CG
it follows that both and (F F1 )( ) belong to the same set in CG . Therefore
(G F F1 )( ) = G ( ) for all F .
But G : AG is continuous. Hence TF G F1 is continuous. Thus, by Lemma 2,
TF G is continuous.
It is readily verified that, when CF = CG , G F : AF AG is one-to-one and
onto and its inverse is F G . Also CF = CG implies CG CF and so, by the first
part of the theorem, F G is continuous. Hence TF G = G F : AF AG is a
homeomorphism.
7. THE TOPS DYNAMICAL SYSTEM. The shift operator S : is defined
by S(1 2 3 ) = 2 3 4 for all 1 2 3 . The shift operator and shift
invariant sets play a central role in the study of symbolic dynamical systems [11], [15].
The calculation of the address structure of an IFS may be simplified by the observation
that the associated tops code space is shift invariant.
Theorem 2.
(i) For any IFS F we have
S(F ) F .
(ii) If f 1 is one-to-one on AF then
S(F ) = F .
Proof. We start by observing that for any j {1, 2, . . . , N } and any ,
f j (F ( )) = F ( j ). The reason is that since f j is continuous, for any z X we
have
f j (F ( )) = f j ( lim f 1 f 2 f k (z))
k
= lim f j f 1 f 2 f k (z) = F ( j ).
k
(i) To see that S( ) F , suppose that there is some > S( ) such that
F () = F (S( )). Then
F (1 ) = f 1 (F ()) = f 1 (F (S( ))) = F (1 S( )) = F ( ).
But 1 > , so this contradicts the fact that F . Therefore S( ) is the
largest address of F (S( )), so S( ) F .
(ii) We show that, when f 1 is invertible on AF , we have 1 F . Suppose that
1
/ F . Then there is some > 1 such that F () = F (1 ), and = 1
where
> . Then
f 1 (F (
)) = F (1
) = F () = F (1 ) = f 1 (F ( )),
so since f 1 is one-to-one, F (
) = F ( ), which leads to a contradiction.
Hence 1 F .
Theorem 2 tells us that we can define what we call the tops dynamical system TF :
AF AF by
TF = F1 S F .
We note that if the address structures of F and G are the same then TF G is a homeomorphism and the two tops dynamical systems TF and TG are topologically conjugate,
with
TG = TF G TF TF1G .
Thus we have a means for conveying ideas from dynamical systems theory to the
analysis of self-referential sets. For example, we might define the entropy of a selfreferential set to be the infimum of the entropies of all associated shift dynamical
systems. We can use the orbits of a tops dynamical system to calculate the tops code
space: for each x AF the value of F (x) = 1 2 is given by
k = min{n {1, 2, . . . , N } : TF(k1) (x) f n (AF )},
where TF0 (x) = x, TF1 (x) = TF (x), TF2 (x) = TF (TF (x)), and so on. This formula is
useful when, as in Example 1 below, the sets f n (AF ) have straight edges and a simple
formula for TF (x) is available.
8. EXAMPLES. We do not entertain you here with (a) an example of IFSs F and
G which have the same address structure, but have different address structures when
the indices in F are permuted; (b) an example where F = G yet AF and AG are
not homeomorphic; (c) an example where f 1 is not one-to-one on AF yet TF (AF ) =
AF . Such examples demonstrate the intriguing nature of fractal transformations and
you can probably find them yourself. Here we demonstrate an area-preserving fractal
homeomorphism applied to a picture of a bird, and a continuous transformation from
a fractal fern onto a filled square.
Example 1. An example of an address structure is provided by the IFS F = F,,
introduced at the end of Section 2. We show that
April 2009]
CF,, = CF,,
(8.1)
299
for all , , ,
, ,
(0, 1). This is intuitively obvious, but the shortest proof that
I can find is the following.
Let F = F,, and G = F,,
. We show (i) that F = , (which implies CF =
{F1 ({x}) : x }); and (ii) that, for any pair , , F ( ) = F () if and only
if G ( ) = G ().
) = 1 2 k 4 for any
To prove (i) we show that F1 ( F (1 2 k 4)
1 2 k . (We write to denote the set of finite length strings of the
symbols {1, 2, . . . , N } together with the empty string.) Let = 1 2 k 4 and
lies in the interior of f 4 () and F ( ) =
let n be greater than k. Since F (4)
it follows that F ( ) lies in the interior of the triangle
f 1 f 2 f n (F (4))
f 1 f 2 f n (). Thus if is such that F () = F ( ) then 1 = 1 ,
2 = 2 , . . . , n = n . Since this is true for all n we must have = . It follows that
1 2 k 4 belongs to F for all 1 2 k . The set of such points is dense in
so F = .
To prove (ii) we suppose that F ( ) = F () = x. Then f ,n () f ,n () is a
decreasing sequence of nonempty compact sets which converges to {x}, where f ,n =
f 1 f 2 f n . Now notice that we can define a homeomorphism K n :
by
1
(x) for x f ,n ().
K n (x) = g,n f ,n
(The crucial fact that this does indeed provide a homeomorphism is readily proved by
induction on n.) We now find that for k n,
K n ( f ,k () f ,k ()) = K n ( f ,k ()) K n ( f ,k ()) = g,k () g,k ().
It follows that g,n () g,n () is a decreasing sequence of compact sets. It converges to a singleton {y} . Hence G ( ) = y = G (). We complete the proof of
(ii) by repeating the same argument with F and G swapped. We conclude that CF = CG
and hence, by Theorem 1, TF G : is a homeomorphism.
Figure 3 illustrates the action of TF G for F = F0.5,0.5,0.5 and G = F 0.65,0.3,0.4. The
images were computed using the modified chaos game. The image on the left shows
S, the union of the attractors of the two IFSs {; f 1 , f 3 , f 4 } and {; f 2 , f 3 , f 4 }. The
image on the right shows
S, the union of the attractors of the IFSs {; g1 , g3 , g4 } and
{; g2 , g3 , g4 }. The two sets are related by TF G (S) =
S.
Figure 3. Before, on the left, and after, on the right, a fractal homeomorphism.
300
Figure 4. Two examples of fractal homeomorphisms applied to the picture at the top. The tranformation from
the top image to the one at bottom left is area-preserving.
an
bn
cn
dn
en
qn
1
2
3
4
0.85
0.06
0.17
0.17
0.05
0.02
0.22
0.22
0.125
0.45
0.195
0.805
0.05
0.0
0.22
0.22
0.85
0.165
0.17
0.17
0.039
0.835
0.776
0.776
Table 3.
April 2009]
an
bn
cn
dn
en
qn
1
2
3
4
0.8
0.2
0.2
0.8
0.0
0.0
0.0
0.0
0.0
0.8
1.0
0.0
0.0
0.0
0.0
0.0
0.8
0.8
0.8
0.2
0.0
0.2
0.0
1.0
301
Figure 5. (i) Shows the points i, j, k, l, m and (ii) shows the points I, J, K , L , M.
(8.2)
CF = CF(1) CF(2)
where
(8.3)
g3 (I ) = M, g3 (K ) = L , g4 (I ) = M, g4 (K ) = J,
where the points I, J, K , L , M AG are approximately as labeled in Figure 5.
Note that equations (8.3) are the same as equations (8.2) upon substitution of f 1 , f 2 ,
f 3 , f 4 , i, j, k, l, and m by g1 , g2 , g3 , g4 , I , J , K , L, and M, respectively. It is readily
deduced that K = G (1), I = G (2), M = G (12) = G (21) = G (32) = G (42),
and that G = . As a consequence CF CG : if s CF then either s CF(1) or s
CF(2) ; if s CF(1) then s is a singleton and, since CG is a partition of , there must be
302
t CG such that s t; if s CF(2) then s = { 12, 21, 32, 42} for some ,
and since M = G (12) = G (21) = G (32) = G (42), it follows that CG contains a
set that contains s. Hence CF CG and, by Theorem 1, the fractal transformation TF G
is continuous. Note however that in this case CF = CG because there is a set in CG
which consists of a pair of distinct addresses, whereas all sets in CF contain either one
or four distinct addresses.
The continuous transformation TF G from the fern-shaped set onto is illustrated
on the left-hand side of Figure 6. This shows a close-up on PG TF G . The picture PG ,
represented in the center of Figure 6, has been chosen to have apparently continuously
varying intensity so that the continuity of TF G is illustrated by the smooth variation of
intensity in the left-hand fern image. You can imagine the left-hand fern to be deformed
without breaks to become the central image!
Figure 6. The ferns on the left and right are both obtained by fractal transformations from fern shaped sets
onto a filled square. The central image is the continuous image of the fern on the left. But the transformation
from the right-hand fern onto the square is not continuous.
an
bn
cn
dn
en
qn
1
2
3
4
0.8
0.2
0.8
0.2
0.0
0.0
0.0
0.0
0.8
1.0
0.0
0.8
0.0
0.0
0.0
0.0
0.8
0.2
0.2
0.8
0.8
1.0
0.8
0.0
April 2009]
303
REFERENCES
1. M. F. Barnsley, Fractals Everywhere, Academic Press, Boston, 1988.
2.
, Superfractals, Cambridge University Press, Cambridge, 2006.
3. M. F. Barnsley and S. G. Demko, Iterated function systems and the global construction of fractals, Proc.
Roy. Soc. London Ser. A 399 (1985) 243275.
4. M. A. Berger, An Introduction to Probability and Stochastic Processes, Springer-Verlag, New York, 1993.
5. R. Devaney, Chaos, Fractals, and Dynamics: Computer Experiments in Mathematics, Addison-Wesley,
Menlo Park, CA, 1989.
6. J. H. Elton, An ergodic theorem for iterated maps, Ergodic Theory Dynam. Systems 7 (1987) 481488.
7. B. Forte and F. Mendivil, A classical ergodic property for IFS: A simple proof, Ergodic Theory Dynam.
Systems 18 (1998) 609611.
8. M. Hata, On the structure of self-similar sets, Japan J. Appl. Math. 2 (1985) 381414.
9. J. E. Hutchinson, Fractals and self-similarity, Indiana Univ. Math. J. 30 (1981) 713747.
10. T. Kaijser, On a new contraction condition for random systems with complete connections, Rev.
Roumaine Math. Pures Appl. 26 (1981) 10751117.
11. A. Katok and B. Hasselblatt, Introduction to the Modern Theory of Dynamical Systems. With a Supplementary Chapter by Katok and Leonardo Mendoza, Cambridge University Press, Cambridge, 1995.
12. B. B. Mandelbrot, The Fractal Geometry of Nature, W. H. Freeman, San Francisco, 1983.
13. B. Mendelson, Introduction to Topology (British edition), Blackie & Son, London, 1963.
14. O. Onicescu and G. Mihok, Sur les chanes de variables statistiques, Bull. Sci. Math. de France 59 (1935)
174192.
15. W. Parry, Symbolic dynamics and transformations of the unit interval, Trans. Amer. Math. Soc. 122 (1966)
368378.
16. D. Peak and M. Frame, Chaos Under Control: The Art and Science of Complexity, W. H. Freeman, New
York, 1994.
17. H.-O. Peitgen et al., Fractals in the Classroom, Vols. I and II, Springer-Verlag, New York, 1992.
Stenflo, Uniqueness of invariant measures for place-dependent random iterations of functions, in Frac18. O.
tals in Multimedia, M. F. Barnsley, D. Saupe, and E. R. Vrscay, eds., IMA Volumes in Mathematics and
its Applications, vol. 132, Springer-Verlag, New York, 2002.
19. R. F. Williams, Composition of contractions, Bol. da Soc. Brasil de Mat. 2 (1971) 5559.
MICHAEL BARNSLEY (B.A. 1968 Oxford University, Ph.D. 1972 University of WisconsinMadison) is a
professor of mathematics at the Australian National University, where he teaches fractal geometry and collaborates with John Hutchinson. He splits his time between Canberra and Atlanta, GA. In previous lives he was an
itinerant post-doc in England, France, and Italy (19731979); a professor at Georgia Tech (1979-1991); and an
entrepreneurthe company he cofounded, Iterated Systems Inc. (19871998), developed fractal image compression technology, used for example in Microsoft Encarta. He is enthusiastic about teaching mathematics
and observing nature.
Department of Mathematics, Australian National University, Canberra, ACT, Australia
michael.barnsley@maths.anu.edu.au, mbarnsley@aol.com
http://www.superfractals.com
304