Vous êtes sur la page 1sur 7

Three-Dimensional Fractals

CHRISTOPH BANDT, MAI THE DUY


AND

MATHIAS MESING

ski gasket and the Koch ractals such as the Sierpin curve (Figure 3) have become standard examples in mathematical textbooks, and there are many related two-dimensional structures [4, 13]. In dimension 3, only two examples are well-known: the Menger sponge and the fractal tetrahedron, shown on many web sites: see for instance http://www.mathpaint.blogspot.com (click on April 2008). A.G. Bell, known for the development of the telephone, found the fractal tetrahedron several years ski, in ying experiments with kites (see before Sierpin Figure 2). In 1903, Bell wrote that since 1899 I have been continuously at work upon experiments relating to kites. Why, I do not know, excepting perhaps because of the intimate connection of the subject with the ying-machine problem. We are all of us interested in aerial locomotion; and I am sure that no one who has observed with attention

the ight of birds can doubt for one moment the possibility of aerial ight by bodies specically heavier than the air. Bell disproved the belief that an air-ship would not be possible because the weight-to-surface ratio must increase for larger machines. His fractal kite, which he considered a milestone of progress in the eld, could be increased indenitely without changing the ratio of surface and weight [5]. Some years later, Bells kite could lift a man, but too late: the Wright brothers rst ight took place in December 1903. Since then, fractals have found many other applications, including the structure of neurons and DNA, clouds, rough surfaces, soil, and other porous materials. Though all applications concern subsets of three-dimensional space, the number of simple three-dimensional geometric models has not increased signicantly. Beside the tetrahedron and the cube, the other three regular polyhedra have been studied.

Figure 1. A new fractal construction.

Figure 2. The fractal tetrahedron, presented as a kite by Bells team in 1903. Reprinted from [5].
2010 Springer Science+Business Media, LLC

A2

A3 A4 A21

A1 A14

ski gasket; right: the Koch curve. Figure 3. Left: the Sierpin

Figure 4. In self-similar sets, pieces are similar to the whole set. If the pieces are too small, they do not touch each other and they form a Cantor set (left). If the pieces are too large, their overlaps can obscure the self-similarity (right). Only for special arrangements is a nice geometric structure obtained.

(See Figure 4 for the octahedron, and [9, 11, 12, 15, 17] for icosahedron and dodecahedron.) Whereas computer scientists used these fractals as a testbed for ray-tracing algorithms rendering complicated 3D scenes, mathematicians apparently were driven by the idea that there should be a small number of fractals with a particular regular and simple structure. After introducing the relevant mathematical concepts, we will address this idea, suggest a denition of

simple structure, and present some new examples, such as the fractal shown in Figure 1.

Geometric Self-Similarity
We consider the simplest class of fractals. A set A in Rn is called self-similar if it is the union of sets A1, ... Am, which are geometrically similar to A. These pieces must then contain similar subpieces, and the subpieces must contain still

.........................................................................................................................................................
AUTHORS

CHRISTOPH BANDT is a professor of

MAI THE DUY received his MSc from Hanoi

mathematics at the University of Greifswald, Germany. Beside fractals, he is interested in random phenomena and biomathematics with real data. Institute of Mathematics Arndt University, 17487 Greifswald Germany e-mail: bandt@uni-greifswald.de
THE MATHEMATICAL INTELLIGENCER

National University and is a PhD student in Greifswald, supported by the Ministry of Education and Training of Vietnam. Institute of Mathematics Arndt University, 17487 Greifswald Germany

smaller similar copies, and so on. This is the idea of fractal structure, according to Mandelbrot who coined the term fractal and convinced the world of its relevance [13]. Fractal structure is also apparent when the similarity is not strict, for instance when afne or conformal mappings act between the set and its pieces, as in the Julia sets recently discussed in this journal [16]. Here we focus on strict self-similarity. If the pieces Ai are disjoint, as on the left side of Figure 4, then A is a Cantor set, which is not very interesting from the geometric point of view. The pieces Ai should be allowed to intersect. But their overlap should be small, because otherwise, as on the right side of Figure 4, the self-similar structure can hardly be recognized. Special constructions are needed to guarantee small overlap. P.A.P. Moran [14] dened a self-similar construction by a bounded open set U and geometrically similar subsets Ui , U, i = 1, ..., m, which are disjoint. In each Ui, disjoint similar copies Uij are chosen, and so on. Moran takes F0 U ; the closure of U, as rst approximation, and [ [ F1 Ui ; F2 Uij etc., so that F0 ' F1 ' F2 ' . . .
i i ;j

Figure 5. The fractal octahedron a deated balloon.

The intersection of the decreasing sequence of compact sets Fi is the resulting fractal set A. In 1946, Moran determined the Hausdorff dimension d of A and proved that the d-dimensional Hausdorff measure of A is positive, which provides a natural concept of volume and geometric probability on the set A. For simplicity, we assume that there are m pieces on each level, and that the similarity ratios of the Ui with respect to U are all equal to some positive real number r. In this case the dimension can be explicitly calculated [4, 8, 10, 13] m rd 1 or d log m : log r

For the examples below, d indicates to what extent the fractal lls three-dimensional space. Morans construction is illustrated in Figure 3. Let U be a triangle. Cut a hole H into U, such that the set U nH consists of a union of similar copies of U. In Figure 3, the hole is a closed triangle. In the Menger sponge, the hole is a threedimensional cross, and in the fractal tetrahedron it is a regular octahedron.

The Topology of the Fractal Octahedron


In the fractal octahedron, the holes are tetrahedra. This example has an interesting topological structure. Figure 5 shows that all 8 tetrahedral holes are carved from outside

.........................................................................
MATHIAS MESING completed his PhD in

2008 and is now teaching high school in Greifswald. Institute of Mathematics Arndt University, 17487 Greifswald Germany

into the basic octahedron U, and all faces of the remaining m = 6 octahedra Ui can be reached completely from outside. By self-similarity, on each level k of the construction, all surface points of the 6k octahedra of the residue set can be reached by an arc from outside U that does not contain other points of the residue set. In the limit, each point of the fractal can be reached by such an arc. This means that each of the 8 faces of the fractal octahedron is the image of a triangle under a continuous map a three-dimensional version of the Koch curve (Figure 3). However, unlike in two dimensions, many pyramidal spikes will touch each other along an edge so that this surface has double points and is not homeomorphic to a triangle. Since the diameter of the remaining octahedra on level n converges to zero for n??, no interior points are left between the fractal faces. The fractal octahedron has eight faces, exactly as the ordinary octahedron, but no interior similar to a deated balloon. Any two neighboring fractal faces have as their intersection an ordinary Euclidean triangle! Its three vertices are the two endpoints of their common edge and the center of the original octahedron. Every square, given by 4 vertices of the basic octahedron U that do not enclose a face, belongs completely to A. (To prove all this, use the fact that the edges of U, and hence of all smaller octahedra, are contained in A. Some of these edges form a grid that in the limit becomes a dense subset of the square.) Thus if we look at an octahedral face from the other side, we see an ordinary pyramid, and if we remove the boundary, smaller pyramids will appear below. Altogether they represent a small part of the face however, since triangles have dimension 2, and the dimen6 sion of the fractal face is log log 2 % 2:58:

Similarity Mappings
The method of carving holes H into appropriate open sets U is not very powerful because of our lack of imagination. But in 1981, Hutchinson introduced similarity mappings fi with Ai = fi(A). A mapping f from Rn into itself is a
2010 Springer Science+Business Media, LLC

similarity mapping with factor r [ 0 if jf x f yj r jx yj for all points x ; y ;

that is, all distances |x - y| are contracted by the ratio r. Given f1, ..., fm with factors ri \ 1, the corresponding compact self-similar set A was dened by the equation A f1 A [ . . . [ fm A: It is not difcult to show that this equation always has a unique nonempty solution, even in the case of afne or conformal contractive mappings [10]. Letting Ai fi A fi1 fi2 . . . fin A; we see that this algebraic approach addresses small pieces by words i = (i1i2 ... in) from the alphabet {1, , m} (see Figure 3). This approach made fractals accessible to computers. In a time when the rst personal computers with graphics facilities became available, Hutchinsons concept became extremely popular, and various algorithms for computer visualization were developed. See Barnsley [4] who calls f1, ..., fm an iterated function system (IFS).

isometry. That is, h(x) = Mx + v, with an orthogonal matrix M and a vector v. The map h = hij, called a neighbor map, maps A to a neighbor set h(A), which has the same position relative to A as Aj has to Ai, up to similarity. In other words, fi(A) = Ai and fi(h(A)) = Aj. The set of neighboring positions of A for the Koch curve is shown in Figure 6. The previous criterion says that such neighboring positions must not come arbitrarily close to the position of A.

Finite Type
There is one case where this criterion can be checked algorithmically, with the aid of a computer if needed. When the set of neighbor maps is nite, we need only check whether hij 6 id ; or, equivalently, whether two pieces Ai, Aj coincide. All examples here and almost all examples in the literature have a nite number of neighbor maps. In this case A, and the IFS f1, ..., fm, is said to be of nite type. The neighbor maps can be generated recursively, start1 ing with fi fj for i ; j 1; . . .; m; i 6 j : The recursion is given -1 1 -1 by fik fj = fk (fi fj)f. It turns out that for neighbor maps with Ai \ Aj 6 ; only isometries with |v| \ c need to be considered, where c is a small constant, for example, 2 times the diameter of A. The algorithm will stop after nite time when there are nitely many neighbor maps. In this case, the IFS f1, ..., fm, or the generated self-similar set A, is of nite type. The gures that follow were checked with this algorithm [3].

Neighbor Maps
A basic set U is not needed in the mapping construction of A. It turns out, however, that its Hausdorff dimension and measure can only be calculated, and complicated overlaps such as that in Figure 4 can only be avoided if there is an open set U such that the fi(U) are disjoint subsets of U. An algebraic criterion by Bandt and Graf [1] says that U exists if and only if mappings of the form hij fi1 fj ; with i i1 i2 . . .in ; j j1 j2 . . .jn0 and n; n0 2 N; cannot converge to the identity map. It is sufcient to consider those i, j for which the pieces Ai and Aj intersect each other, and have almost the same size. Since ri = r was assumed here, the last condition means n = n0 so that Ai and Aj have exactly the same size, and the map hij is an

Choice of Maps and Symmetry


Our principle was to look for nite type fractals. The number of neighbor types is a measure of the complexity of the geometry [3], and we are searching for the simplest examples possible. This restricts the choice of mappings since composition of rotations in R3 is usually noncommutative. In a series of articles written around 1999, Conway, Radin, and Sadun discussed related questions. Among other things, they found that when two rotations g, h around the origin in R3 with different axes full any algebraic relation such as g2hg-1h3g = id, then special conditions must be fullled that are often connected with the Platonic solids [7]. Since in the case of nite type at least two neighbor maps must be equal, hij hi0 j0 , the rotational parts of the fi must full such a relation. Thus it seemed reasonable to look for mappings associated with regular polyhedra. Actually, symmetry signicantly decreases the number of neighbor types (dened as in the following). If A does not possess symmetries, the geometric position of A and h(A) is determined by the map h. However, if s, t are symmetries of A, that is, isometries of Rn with s(A) = t(A) = A, then g = sht and h determine the same geometric neighbor position, up to symmetry of the pair of sets. In other words, s-1(A) = A and s-1(g(A)) = h(A). Thus a neighbor type of the IFS f1, ..., fm has to be dened as an equivalence class with respect to the relation g*h if there are symmetries s, t with g = sht. The Koch curve has skis gasket as well two types, as seen in Figure 6, and Sierpin

Figure 6. Four possible neighbor positions for the Koch curve. If the bold curve represents A4 in Figure 3, it has only the upper left neighbor, whereas A2 has also the lower right neighbor. A14 and A21 have both upper neighbors. Of course, for each piece of a curve only one neighbor on each end is possible.
THE MATHEMATICAL INTELLIGENCER

as the tetrahedron has only one type (translation along an edge of the triangle) when reections are included as symmetries. Mengers sponge has three types of neighbors those with a common face, a common edge, or only a common point. In the following we take symmetric sets, saving a lot of computation in checking the nite type condition.

Modication of the Fractal Tetrahedron


Next, consider the fractal tetrahedron, Figure 2. As a rotation axis for the fi, take the altitude from the vertex ci (the xed-point of fi) onto the opposite face of the tetrahedron. Rotation around 120 would be a symmetry of A. To obtain something new, combine the fi with a rotation around 60, or, equivalently, around 180. It turns out that the images of the basic tetrahedron will meet along an edge when we increase the factor to r 3 5 : The result is shown in Figure 8. To see how we obtain r, consider the basic tetrahedron T as subset of the unit cube, with vertices c3 = (0, 0, 0), c4 = (1, 1, 0), c1 = (1, 0, 1), and c2 = (0, 1, 1). The diagram in Figure 8 shows the projection of T onto the x1, x2plane. As in Figure 7, let c02 and c03 denote the images of c2, c3 under f1 that are outside T. We have to choose r so that both c02 and c03 are images of vertices cj under two other mappings fi, which is indicated by connecting them to the vertices cj. By symmetry, c02 = (t, t, t) and c03 = (1 - t, 1 2 0 t, t) for some t. The equation jc02 c03 j2 jc 2 p p c1 j gives 3 4 0 t 5 and jc2 c1 j 5 2: Since jc2 c1 j 2; the factor of f1 is 3 5: For the open set U we can take the convex polyhedron with vertices ci and c0i , that is, the convex hull of the union of the four image tetrahedra fi(T). (Incidentally, they enclose an inner tetrahedron of the same size.) The Ai intersect in Cantor sets, and there are only two neighbor types that share a long or short edge, but there are many other types of neighbors that meet at a single point. The Ai intersect the outer faces of U in Koch curves (to prove this, calculate the four mappings fi fj which map a face of U into itself). The log 4 dimension log 5=3 % 2:71 is almost the same as for Figure 7, 20 and also for the Menger sponge, log log 3 % 2:73: This method also applies to the cube, which is the self-similar set with respect to 8 homotheties with factor 1 2 and centres in the vertices. If the mappings are combined with 60 rotations around the space diagonals, and the factor is 5 8 ; Figure 9 is obtained, where the neighboring pieces touch in a single point, similar to the fractal tetrahedron.

Turning Plane Fractals into Space


We now explain the pictures, which were produced with the excellent public domain software chaoscope [6]. We skis triangle with vertices ci, arranged start with Sierpin symmetrically around 0 in the x1, x2-plane: fi(x) = rx + (1 - r)ci for i = 1, 2, 3, where r 1 2 : Now we compose each fi with a 90 rotation around the axis [0, ci]. Explicitly, let c1 = (1, 0, 0), and let t denote the 120 rotation in the x1, x2-plane with t(c1) = c2. Then we take f1 x1 ; x2 ; x3 rx1 1 r ; rx3 ; rx2 and f2 tf1 t 1 ; f3 t 1 f1 t :
2 For r 1 2 we would get a Cantor set, but for r 3 the images of the triangle Dc1, c2, c3 under the fi will intersect in a vertical line segment L through 0. In Figure 7, the triangle and its image under f1 are depicted. Since each altitude in Dc1, c2, c3 is mapped into itself by the corresponding fi, the factor r 2 3 comes from the fact that altitudes in an isosceles triangle intersect in a ratio 1:2. The three pieces Ai of the resulting fractal will meet in a Cantor set on L. As open set U we can take the double pyramid with vertices c1, c2, c3, c02 , and c03 . Altogether, there are 420 neighbor maps. When neighboring pieces meeting in a single point are neglected, and the symmetry reduction is performed with the full symmetry group of A, there remain only two neighbor types. Neighbor pieces can meet Ai along a long or a short edge of the corresponding set Ui. In both cases, the angle between the pieces is log 3 unique. The dimension of this fractal is log 3=2 % 2:71: The construction can be applied to all n-gons with n C 3.

C 2

C2

C3
C3

C 1

skis triangle. The diagram shows the triangle given by the xed points ci of Figure 7. A three-dimensional modication of Sierpin the mappings fi, and the image triangle under f1.
2010 Springer Science+Business Media, LLC

C2 C 2 C 3 C3

C4

C1

Figure 8. The modied fractal tetrahedron touches the faces of U with Koch curves. U is obtained from a tetrahedron T by adding small pyramids on the faces of T. In the diagram, U is depicted as a subset of the unit cube, viewed from above. The vertices c1, ..., c4 of T are vertices of the cube; the other upper vertices of the cube were moved inside to the position of c02 and c03 .

Figure 9. Two views of the modied cube, generated by homotheties combined with 180 rotation around the cubes diagonal.

Figure 10. Two views of the reverse of Figure 7.


THE MATHEMATICAL INTELLIGENCER

The Reverse Fractal


For A generated from the mappings fi, dene the reverse fractal of A by the mappings - fi (cf. [2], Section 10). It is 1 obvious that the neighbor maps hij = fi fj coincide for both families of mappings. Thus when one of the fractals is of nite type or fulls the open set condition, so does the other. When A is centrally symmetric, as in Figures 9 and 5, it coincides with its reverse. In other cases, the appearance can be quite different although the dimension and the number of types remains unchanged. Figure 1 is the reverse of Figure 8. The reverse of Figure 7 is shown in Figure 10. In both cases, the geometry becomes more complicated, and more realistic as a model for natural phenomena. It is also possible to apply the minus sign only to some of the fi, or to some of the three coordinates, but we will stop here and leave it to you to create more examples.
REFERENCES

in tetrahedral kites, see http://www.cit.gu.edu.au/*anthony/ kites/tetra. [6] Chaoscope software developed by N. Desprez, http://www. chaoscope.org/. [7] J. H. Conway, C. Radin, and L. Sadun, Relations in SO(3) Supported by Geodetic Angles, Discrete Comput. Geom. 23 (2000), 453463. [8] K.J. Falconer, Fractal Geometry. Mathematical Foundations and Applications, Wiley 1990. [9] T. A. DeFanti and J. C. Hart, Efcient antialiased rendering of 3-D linear fractals, ACM SIGGRAPH Computer Graphics 25 (3) (1991), 91100. [10] J.E. Hutchinson, Fractals and self-similarity, Indiana Univ. Math. J. 30 (1981), 713747. [11] H. Jones and A. Campa, Fractals Based on Regular Polygons and Polyhedra, N.M. Patrikalakis (ed.) Scientic Vizualization of Physical Phenomena, Springer, New York, 1991, 299-314. ski Polyhedra, Pi Mu [12] A. Kunnen and S. Schlicker, Regular Sierpin Epsilon J. 10 No. 8 (1998), 607619. [13] B.B. Mandelbrot, The Fractal Geometry of Nature, Freeman, San Francisco, 1982. [14] P.A.P. Moran, Additive functions of intervals and Hausdorff measure, Proc. Camb. Phil. Soc. 42 (1946), 1523. [15] A. Norton, Generation and rendering of geometric fractals in 3-D, ACM SIGGRAPH Computer Graphics 16(3) (1982), 6167. [16] K.A. Roth, Julia sets that are full of holes, Math. Intelligencer 30, No. 4 (2008), 5156. rpern, [17] W. Sternemann, Neue Fraktale aus platonischen Ko Spektrum der Wissenschaft 11 (2000), 116118.

[1] C. Bandt and S. Graf, Self-similar sets 7. A characterization of self-similar fractals with positive Hausdorff measure, Proc. Amer. Math. Soc. 114 (1992), 9951001. [2] C. Bandt and N.V. Hung, Fractal n-gons and their Mandelbrot sets, Nonlinearity 21 (2008), 26532670. [3] C. Bandt and M. Mesing, Fractals of nite type, Banach Center Publications 84 (2009), 131148. [4] M.F. Barnsley, Fractals Everywhere, 2nd ed., Academic Press, 1993. [5] A.G. Bell, Tetrahedral Principle in Kite Structure, National Geographic Magazine Vol. XIV, No. 6, 1903. Available at http:// www.fang-den-wind.de/bell_eng.htm . For recent developments

2010 Springer Science+Business Media, LLC

Vous aimerez peut-être aussi