FRACTAL n-GONS AND THEIR MANDELBROT SETS

CHRISTOPH BANDT AND NGUYEN VIET HUNG Abstract. We consider self-similar sets in the plane for which a cyclic group acts transitively on the pieces. Examples like n-gon Sierpiski gaskets, Gosper n snowake and terdragon are well-known, but we study the whole family. For each n our family is parametrized by the points in the unit disk. Due to a connectedness criterion, there are corresponding Mandelbrot sets which are used to nd various new examples with interesting properties. The Mandelbrot sets for n > 2 are regular-closed, and the open set condition holds for all parameters on their boundary, which is not known for the case n = 2.

MSC classication: Primary 28A80, Secondary 52B15, 34B45

1. Introduction Most of the classical fractals, like Cantors middle-third set, von Kochs curve, Sierpiskis gasket and carpet, Mengers sponge etc. [20, 8, 12] are self-similar sets n with certain symmetries. That is, they are compact sets which full an equation A = f1 (A) ... fn (A) (1) where f1 , ..., fn are contracting similarity maps on Rd . And there is at least one symmetry map s with s(A) = A which permutes the pieces Ak = fk (A), so that s(Ak ) = Aj where j depends on k and s. Self-similarity and symmetry make the shapes look attractive, but what is more, they simplify the description and mathematical treatment of such fractals. An analytic theory including Brownian motion [19], Dirichlet forms, spectrum of the Laplacian [16] and geodesics [25] has been developed only on symmetric fractals. Recently, Falconer and OConnor [13] have classied and counted certain symmetric fractals. This paper studies a large family of fractals with rich symmetry. A compact subset A of the complex plane is called a fractal n-gon if A fulls equation (1) for similarity mappings fk (z) = k z + ck , k = 1, ..., n, and there is a rotation in the plane which acts transitively on the pieces, permuting them in an n-cycle. Some fractal n-gons which generalize the Sierpiski gasket for arbitrary n have n been considered by several authors [21, 25]. Some others like twindragon, terdragon, Gospers snowake and the fractal cross [20, Ch. 6] are known as tiles. Figures 1 and 2 show some examples which seem to be new. For each n, all fractal n-gons are parametrized by one complex parameter running through the unit disk (Section 2). Since connectedness of A means that neighboring pieces intersect (Section 3), we can dene a Mandelbrot set Mn for fractal
1

CHRISTOPH BANDT AND NGUYEN VIET HUNG

Figure 1. 3-gon fractals where pieces intersect in one point. Upper row: The point has addresses 0 10 , 0 1020 (Example 1). 1 2 1 2 Lower row: Corresponding reverse fractals, see Section 5. n-gons in Section 4. M2 was introduced by Barnsley and Harrington in 1985 [9, 8] and studied by several authors [10, 11, 2, 22, 23, 24]. We shall see that Mn has similar properties. In section 5, we describe a fast algorithm to generate Mn , and discuss the overlap set of n-gons which determines their geometry. Finally, we consider the simplest n-gons with one-point intersection set, classify them and derive the basic analysis for this innite family. 2. Parametrization We want to derive a very simple form for the mappings fk which generate a fractal n-gon. We start with the simple observation that a similarity map, that is, a change of the Cartesian coordinate system, will not change the structure of a self-similar set. Remark 1. Let A = fk (A) be a self-similar set in Rd and h : Rd Rd a similarity map. Then B = h(A) is the self-similar set generated by the mappings gk = hfk h1 . Proof. B = h(A) = hfk (A) = hfk h1 (B)

In the complex plane, when h consists of translation, scaling and rotation, i.e. h(z) = z + c, then gk will have the same factor k as fk . If h includes a reection and thus reverses orientation, h(z) = z + c, then the gk will have the conjugate factors k . In the present paper we consider all similitudes h as isomorphisms and do not distinguish between a self-similar set and its mirror image.

FRACTAL n-GONS AND THEIR MANDELBROT SETS

3
n k=1

Proposition 1. Any fractal n-gon is similar to a self-similar set A = where fk (z) = z + bk for k = 1, ..., n where b = cos

fk (A)

2 2 + i sin and || < 1. n n

Thus for each n, the points in the open unit disk parametrize fractal n-gons. Proof. By denition, a fractal n-gon A has n pieces Ak , k = 1, .., n, and a rotation s transforms the pieces cyclically into each other. We choose the numbering so that s(Ak ) = Ak+1 where + is taken modulo n. In the sequel we shall often write A0 and f0 instead of An , fn . By applying a translation h(z) = z +v we push the origin of our coordinate system to the xed point of s so that s(z) = bz where bn = 1 since sn preserves all pieces of A and so must be the identity map. Since bk runs through all roots of unity, it is no loss of generality (but requires another renumbering of the Ak ) to assume that b = cos 2 + i sin 2 . n n We took the fk to be orientation-preserving, so f0 (z) = z + u for some , u C with || < 1. Applying the rotation h(z) = z/u we transform A into a position where f0 (z) = z + 1. Now we note that sk f0 sk (A) = sk f0 (A) = fk (A) for k = 1, ..., n 1. Thus A = n1 sk f0 sk (A) so due the uniqueness of self-similar sets [12, 8] the set A is k=0 the self similar set with respect to the mappings sk f0 sk (z) = z+bk , k = 0, ..., n1. These mappings will now be called fk . The points z in a self-similar set are often described by their addresses u1 u2 ..., sequences of symbols um {0, ..., n 1} which denote pieces and subpieces of A to which z belongs [8, 12, 5]. We have z = p(u1 u2 ...) = lim fu1 fum (0)
m

(2)

where p : {0, ..., n 1} A is the so-called address map [8]. In our case, the address map has a particularly simple form. Remark 2. The point z with address u1 u2 ..., um {0, ..., n 1} in the fractal ngon A() has the representation z = bum m1 . Thus A() is the support of a m=1 random series of powers of with coecients chosen from the n-th roots of unity. Proof. fu1 fum (0) = bu1 + bu2 + + m1 bum .

Example 1. We explain how to determine for the rst fractal in Figure 1. The intersection point z should have the addresses 0 and 10 where = 111... 1 2 1 10 which by the above remark turns into an This is indicated by the relation 01 2 equality for . (In Section 5 we shall see that this is indeed sucient.) p(0 = 1 + b( + 2 + ...) = 1 + b 1) 2 = p(10 = b + + b2 2) 1 1

CHRISTOPH BANDT AND NGUYEN VIET HUNG

Figure 2. Some 4-gon fractals. In the upper row, pieces intersect in two points, in the lower row in four points. See Example 2. which leads to the quadratic equation 2 (1 + b) 2 + 1 = 0. The solution with || 1 is 3 1 = + (1 )i . 2 2 For the upper-right part of Figure 1 with 0 1020 we have to consider p(1020 = 1 2 2) b + + b2 2 + 3 + b4 4 /(1 ) which leads to the equation z 4 z 3 + z 2 + 2bz = b. Again there is only one solution with || 1 which is determined numerically: 0.5135 + 0.1004i. In both cases 1/ is a Pisot number over Q(b). In the lower row of Figure 1, was replaced with (see Section 10). 3. Connectedness Connectedness of fractals is a very important property, which is well-known from the dynamics of complex quadratic maps. For self-similar sets, it implies local connectedness and arcwise connectedness [14, 6]. A self-similar set with n pieces Ak is connected if and only if the graph with vertices k = 1, ..., n and with edges {j, k} for intersecting pieces Aj Ak = is connected [14, 6]. Thus a self-similar set A with two pieces A0 , A1 is connected if and only if A0 A1 = , otherwise it is a Cantor set [8, Ch. 8]. It is not known whether a similar property holds for self-ane tiles A with respect to the mappings fk (x) = M 1 x + kv, k = 0, 1, ..., n 1 where M is an expanding integer d d matrix with determinant n and v Rd . Using algebraic methods, Kirat, Lau and Rao

FRACTAL n-GONS AND THEIR MANDELBROT SETS

[17, 18], Akiyama and Gjini [1] veried connectedness in this case for dimensions d = 2, 3 and 4. With a topological argument we prove now that the alternative connected or Cantor set also holds for fractal n-gons. Due to the rotational symmetry, A0 A1 = implies Ak Ak+1 = for each k which implies connectedness of A due to the above criterion. The necessity of A0 A1 = is more dicult to prove. Theorem 2. A fractal n-gon A is connected if and only if A0 A1 = . If A is disconnected, then it is totally disconnected, and all pieces Ak are disjoint. Proof. Let B(x, r) denote the ball around x with radius r. Let be the smallest positive number such that A B(0, ) = B0 . For any word u = u1 ...um {0, ..., n 1}m of length |u| = m let fu = fu1 fum . For m = 1, 2, ... let Bm =
|u|=m

fu (B(0, )) =
|u|=m

B(fu (0), ||m )

be the m-th level ball approximation of A. We shall use the following fact. (*) If Bm is connected and f0 (Bm ) f1 (Bm ) = , it follows that fj (Bm ) intersects fj+1 (Bm ) for j = 1, ..., n 1 and Bm+1 = n1 fj (Bm ) is connected. j=0 Now we distinguish two cases. Case one. A0 A1 = . In this case (*) implies by induction that all Bm are connected. Then A must be connected, too: if A would split into two disjoint nonempty closed sets F0 , F1 , then the distance of F0 and F1 is a positive and for ||m < the set Bm could not be connected. Case two. A0 A1 = . Then A0 and A1 have some positive distance , and by the argument just given there is a smallest m with f0 (Bm ) f1 (Bm ) = . By (*), Bl is connected for all l = 0, 1, ..., m. We show that all Aj are disjoint and hence A is totally disconnected. Since Bm is a connected nite union of closed balls of radius ||m , the outer boundary of Bm (those points which can be connected with innity in the complement of Bm ) forms a closed curve consisting of nitely many arcs of circles of this radius. may have nitely many double points where balls touch each other, but no other multiple points. Thus we can consider as a curve oriented in clockwise direction, without self crossings, only touching points. Let j = fj () for j = 0, ..., n 1. We assumed 0 1 = and want to show that all j are disjoint. Now let 1 be the smallest positive number with 0 B(0, 1 ), and C the circle around 0 with radius 1 . Then C intersects 0 in a point c0 and j in the point cj = bj c0 . The origin 0 is not inside 0 , because then it would also be inside 1 , which implies that 0 and 1 intersect. Now there is a largest 2 > 0 such that the open ball B(0, 2 ) does not intersect 0 . Let D be the circle around 0 with radius 2 . Then D intersects 0 in a point d0 and each j in the point dj = bj d0 . In the ring formed by D and C, the j sit like spokes in a wheel, connecting dj with cj . Of course 0 may touch D in other points than d0 , but they all must lie between dn1 and d1 because otherwise the Jordan curves 0 and 1 (or n1 ) would

CHRISTOPH BANDT AND NGUYEN VIET HUNG

Figure 3. 4-gon fractals where pieces intersect in a Cantor set intersect. But this means that 0 is enclosed by 1 , n1 and the arcs dn1 d1 on D and cn1 c1 on C. Since 0 does not intersect the neighbor curves and does not cross the circles D and C, it can not intersect the curves 2 , 3 , ..., n2 . Thus all j , and hence all Aj are pairwise disjoint. When dealing with self-similar sets, it is very natural to require the open set condition, OSC [12, 8, 7]. We say that OSC holds for the mappings f1 , ..., fn if there is a nonempty open set U with fk (U ) U and fj (U ) fk (U ) = for j, k {1, ..., n}, j = k. (3) If all Ak are disjoint, with pairwise distance > 0, then such U exists: we can take U = xA B(x, 2 ). For connected sets A in the plane, Bandt and Rao [7] proved that OSC holds when pieces intersect in a nite set. Thus all examples in Figures 1 and 2 full OSC. Proposition 3. If || >
1 n

for n 2, then OSC fails, and A is connected.

Proof. Assume OSC holds (which is true if A is disconnected). Then the Hausdor dimension of A is dimH (A) = log n [12]. On the other hand, dimH (A) 2 since A log || 1 is a subset of the plane. It follows that || n . This bound is sharp for n = 2, 3, 4 where we have fractal tiles (twindragon or rectangle, terdragon, square). Proposition 4. If n < 25 and a fractal n-gon A fulls OSC then only consecutive pieces Ak , Ak+1 can intersect. Proof. For n 3 all pieces are neighbors, so let n 4. Assume OSC holds and r r A0 Aj = for some j {2, ..., n 2}. Since A0 B(1, 1r ) and Aj B(bj , 1r ) 2r 2r this implies |1bj | 1r . Since j {2, ..., n2}, it follows that |1b2 | 1r which
|1b |1b 1 can be reformulated as r 2+|1b|2 | . However, calculation shows that 2+|1b|2 | > n for 5 n 24 which contradicts Proposition 3. For n = 4 we just get equality 1 when r = || = 2 , and the balls with centers 1 and b2 = 1 meet in zero. Now 1 1 + m=1 bkm m = 0 is only possible for = 2 which gives the square.
2 2

We conjecture that the condition n < 25 is not needed here.

FRACTAL n-GONS AND THEIR MANDELBROT SETS

Figure 4. Mandelbrot sets for 3- and 4-gons with symmetry lines 4. Mandelbrot set for n-gons To get an overview over all fractal n-gons, we work in the parameter space. Let us dene the Mandelbrot set for fractal n-gons as Mn = { | || < 1 , A() is connected }. The set M2 was introduced 1985 by Barnsley and Harrington [9]. It was discussed in Barnsleys book [8] and investigated by Bousch [10, 11], Bandt [2], Solomyak and Xu [24] and Solomyak [22, 23]. Nevertheless, it is still an open question whether M2 is regular-closed. Moreover, it is not known whether OSC is fullled for all in the boundary M2 . We shall solve these questions for Mn with n > 2, n = 4. So let us consider Mn for n > 2. Figure 4 shows Mn for n = 3, 4 as black subset 1 of the region 0 arg() , 0 || < n . 2 Proposition 5. Elementary properties of Mn for n 2. (i) Mn is a closed subset of the open unit disk. (ii) For even n, it has the dihedral group Dn as symmetry group. For odd n, D2n acts as symmetry group. Proof. (i). A() depends continuously on with respect to the Hausdor metric [8], and the limit of connected closed sets with respect to the Hausdor metric cannot be disconnected. (ii). Since A() is a mirror image of A() (cf. Section 2), all sets Mn are symmetric under reection at the real axis, . Moreover, A() = A(b ) for b = cos 2 + i sin 2 , by an argument in the proof n n of Proposition 1. Thus Mn is also invariant under multiplication with b, that is, rotation around 2/n, as can be seen in Figure 4. For odd n, however, Mn is also invariant under the map and hence under rotation around /n. This is proved in Proposition 10,(i).

CHRISTOPH BANDT AND NGUYEN VIET HUNG

Figure 5. Magnication of M4 in a square of side length 0.003. For = 0.3021 + 0.3375i taken from the upper right hole, the fractal n-gon must be a Cantor set, for nearby it will be connected. M2 has an antenna on the real axis since for real [0.5, 0.6] the set A() will be an interval but for nearby o the real axis it will be a Cantor set. For n > 2 we cannot get intervals A() and we have no antenna on Mn . On the whole, Mn seems to become smoother for larger n. Nevertheless, there are apparent holes in M3 and M4 as were veried for M2 [2]. That means there are Cantor sets A() such that when we decrease continuously to reach = 0 we must get a connected A on the way. See Figure 5. 1 Proposition 3 says that all with || n belong to Mn . This bound can be improved for n 5, and a lower bound for points in Mn can be given which shows that for large n the boundary of Mn lies in a very thin ring. Theorem 6. Bounds for Mn . (i) All with || <
sin n 1+sin n

do not belong to Mn .
sin n cos 2n +sin n

For odd n, a better estimate is || < (ii) Mn includes all with || and ||
sin n

which equals

1 2

for n = 3.

sin n cos2
2n

3 2

sin n

for odd n,

1+

for even n.
2 n

3 2

sin

(i) is proved after Remark 4 below, (ii) after Lemma 1 and 2 in Section 6. In Section 9 and 10 we show that the estimate (i) is sharp for all n which are not multiples of 4.

FRACTAL n-GONS AND THEIR MANDELBROT SETS

5. How to generate Mn To draw Mn , we must know when the pieces A0 and A1 of A() intersect. By Remark 2, a point z A0 A1 can be written as z = 1 + bkm m = b + m=1 jm m m=1 b where km , jm {0, ..., n 1}. Thus Remark 3. belongs to Mn if and only if is the solution of an equation

g() = 1 b +
m=1

dm m = 0

(4)

where all dm are in n = bk bj | j, k {0, 1, ..., n 1} . For xed , every choice of km , jm , m = 1, 2, ... such that the dm = bkm bjm solve the above equation, corresponds to one point in the intersection set A0 A1 . Of course the choice of dl is restricted since the absolute value of the remaining ||l+1 m sum m=l+1 dm is at most 1|| where = max{|d| | d n } 2. If the absolute value of the sum up to degree l is larger than this value, then there is no chance to extend the sum to obtain a series with g() = 0. This argument, with division by at each step, provides an algorithm for generating Mn . We search through a rooted tree with |n | branches at every vertex. Each || vertex is assigned a complex number v. If |v| > 1|| then the vertex will be considered as dead, and the search will not go deeper at this place. We stop the search if either all vertices are dead, in which case Mn , or if we reach a prescribed level lmax in which case belongs to our approximation of Mn . Remark 4. (Algorithm to generate Mn ) For given and lmax we consider words u = d1 d2 d3 ...dl l with l lmax and n d1 = 1 b, and corresponding points vu C according to the rule vu vd1 = d1 = 1 b , + d , d n . (5) vud = || When |vu | > 1|| then u and its successor words are removed. If no word of length lmax remains, does not belong to Mn . If at least one word of length lmax remains, is considered as an element of Mn .
|| In particular, if |vd1 | > 1|| then is not in Mn . Since |vd1 | = 2 sin and = 2 n for even n, while = 2 cos 2n for odd n, this proves Theorem 6,(i).

Remark 5. If the algorithm would be calculated up to innity, it would also decide about OSC for A() : The mappings h = fj1 fk with j = j1 ...jl , k = k1 ...kl are translations h(z) = z + wjk where wjk bk bj 1b 1 and wjj kk = for f1 f0 , + . So the ws dier from the vs above only by the factor . Now OSC holds if and only if 0 is an accumulation point of the ws [4, 3], or equivalently, of the vs. w01 =

10

CHRISTOPH BANDT AND NGUYEN VIET HUNG

In particular, when vu = 0 is obtained after nitely many steps, that is, (4) holds with a nite sum, then OSC cannot be true. In this case, certain pieces Aj and Ak with k1 = 0 and j1 = 1 coincide, so that the overlap set A0 A1 is really big. For n = 2, the algorithm correctly states when some is not in M, and hence OSC holds, but only for the special values just mentioned it can guarantee that they belong to M [2]. We shall see that for n > 2 the algorithm works much better, due to the structure of n . 6. The difference sets n n contains 0, and all sides and diagonals, considered as vectors in either direction, of the regular n-gon formed by the vertices bk , k = 0, 1, ..., n 1. Thus 2 = {2, 0, 2}, and the 8 non-zero vertices 2, 2i, 1 i of 4 lie on a square. 3 consists 0 and 3ei( 6 +k 3 ) , k = 0, 1, ..., 5, vertices of a regular hexagon, all of with modulus 3. Thus we recover Theorem 6,(i): each point M3 fulls || 1 2 || because otherwise |v1 | > 1|| . Moreover, || = 1 can only lead to some v1d with 2 |v1d | |v1 | if = 1 , or = 1 bk . This is just the Sierpiski gasket and its reverse, n 2 2 see section 8. There is a signicant dierence in the structure of n for odd and for even n. Let us start with odd n = 2q + 1. Sides and diagonals of all lengths of the regular n-gon appear in each of the possible 2n directions i, ib1/2 , ..., ib(2n1)/2 exactly once. The length of a side is s = 2 sin , and the diagonals have lengths n 2 sin 2 , ..., 2 sin q = 2 cos 2n . Thus n n t | t = 0, 1, ..., q, l = 0, 1, ..., 2n 1}. n From a geometric viewpoint, n contains q points on 2n equally spaced rays around 0. Let C1 , ..., Cq and D1 , ..., Dq denote the points on two neighboring rays, and let P Q denote the distance between points P and Q. Then 0C1 = 0D1 = s and Ck Dk+1 = Dk Ck+1 = s for k = 1, ..., q 1 since 0Ck Dk+1 has a congruent triangle build from three vertices of the regular n-gon. We shall need an estimate for the size of circular holes in n . n = {2ibl/2 sin Lemma 1. For odd n, the set n intersects each ball B(x, r) with |x| 2 and r 2 sin 2n . The estimate of r is sharp. Proof. If the center of a circle lies in a triangle (or a trapezium), and the vertices are outside the circle, then the circle must be smaller than the circumscribed circle of the triangle (or trapezium). Thus it suces to show that the radius r of the circumcircle of each trapezium Ck Ck+1 Dk+1 Dk equals 2 sin 2n . (The points x with |x| 2 outside the segment Cq Dq between the two rays are covered by the balls of radius 2 sin 2n around Cq and Dq .) Let M denote the center of this circle, and let = 0Dk+1 Ck . In 0Dk+1 Ck we have 0Ck / sin = s/ sin = 2. Since Ck M Dk = 2, we have in 0M Ck a similar n relation: 0Ck / sin = r/ sin 2n . Now r is obtained from the two equations.

FRACTAL n-GONS AND THEIR MANDELBROT SETS

11

The proof of Theorem 6,(ii) for odd n is quite similar. We show that when fulls the condition, the algorithm of Remark 4 will produce points vu B = B(0, 2 sin ) n for words u of arbitrary length. Since v1 = 1 b B, it is enough to prove that v v v B implies ( +n )B = . The lemma takes care of | | 2. So we assume that the circle C of radius 2 cos 2n does not contain 0. We check under which condition v C intersects B in an arc subtending an angle at M = , because at least one n v point of such an arc belongs to + n . Let D be an intersection point of C with the boundary of B. Since DM 0 increases when the distance y = 0M decreases, it suces to determine y for the case DM 0 = 2n . The cosine formula gives y 2 4y cos2 2n + 4(cos2 2n sin2 ) = 0. Since n 2 sin v y > 2 cos 2n , calculation gives y = 2 cos2 2n + 3 sin . The condition | | || n y n provides the estimate for ||. For even n = 2q, every side and diagonal of the regular n-gon has a parallel side, except for the longest diagonals which are diameters of the unit circle. Again, the length of a side is s = 2 sin , and the diagonals have lengths 2 sin 2 , ..., 2 sin q = 2. n n n However, now there are n directions for the sides and diagonals with odd length 2 sin 3 , 2 sin 5 , ... alternating with n other directions for the even diagonals. For n n n = 4p we get n = {2bl sin t n {2b
1 l+ 2

| l = 0, ..., n 1, t = 0, 2, ..., 2p}

sin t | l = 0, ..., n 1, t = 1, 3, ..., 2p 1} , n

and for n = 4p + 2 we have t = 1, 3, ..., 2p + 1 in the rst part and t = 0, 2, ..., 2p in the second. In both cases, we can label the points on any two neighboring rays, ordered with respect to their distance to 0, as C0 = 0, C1 , ...Cq . Since 0Ck represent the diagonals, we have Ck Ck+1 = s and Ck Ck1 Ck+1 = Ck Ck+1 Ck1 = k for n k = 1, ..., q 1. Again, we can derive an estimate for the circular holes in n . Lemma For even the set n intersects each ball B(x, r) with center |x| 2 2. n, and r 2 sin = s/ 2. For n = 4p the estimate of r is sharp. n Proof. We consider the triangulation of n into isosceles triangles with side length s described above. We can modify it so that the triangles contain no angles > . (Where such angle appears, there is an adjacent triangle so that the union of 2 both triangles is a rhombus, and then we change the diagonal of this rhombus.) Now it suces to show that the radius r of the circumcircle of each triangle is s/ 2. However, this is just the value for a right-angled isosceles triangle with two sides s, and for an acute isosceles triangle the value is smaller. The proof is complete, and we see that the estimate is sharp for n = 4p where Ck Ck+1 Ck1 = k = n 4 happens for k = p. For n = 4p + 2 the largest possible angle between the two equal 2p sides is 2 (p+1) = 2p+1 which leads to the sharp estimate. In particular, for n 2 n = 6 we have r 1/ 3 which is also obvious from the structure of 6 . Proof of Theorem 6,(ii) for even n. As for odd n, we want that v B = v B(0, 2 sin ) implies ( + n ) B = so that the algorithm of Remark 4 gives n

12

CHRISTOPH BANDT AND NGUYEN VIET HUNG

vu B for arbitrary long words u. Thus has to be taken so that

2 sin n B(0, ) n + B. ||

We consider F = B(0, 2) 2n zk + B where zk are the outmost points on the k=1 2n rays of n . By Lemma 2, F n + B. The points with minimal modulus on the boundary F are intersection points of two neighboring outer circles. Let w be the outer intersection point of z1 + B and z2 + B, and assume |z1 | = 2, |z2 | = 2 cos n . Then |z1 z2 | = 2 sin since 0z2 z1 is a right angle. Thus z1 z2 w is an n equilateral triangle. Choose y on the ray 0z1 such that 0yw is a right angle. Since := wz1 y = + , we have by Pythagoras 6 n 2 3 2 2 2 |w| = (2 + 2 sin cos ) + (2 sin sin ) = 4(1 + sin ). n n 2 n If the condition (ii) for is fullled, then
2 sin n ||

|w| and B(0,

2 sin n ) ||

F.

7. OSC holds on the boundary of Mn Now let us improve the algorithm of Remark 4. We give a condition for the vu which guarantees that belongs to Mn , for all n 5 and n = 3. In our pictures of Mn , either this condition or the Cantor set condition of Remark 4 was fullled for each single pixel. To obtain precise gures, we did not need to go far into the recursion: level lmax = 60 was sucient for magnication up to factor 108 . Theorem 7. For n = 2, 4 there is a critical radius rn with the following property. If |vu | rn for some u in the algorithm of Remark 4, where || 1 for n = 3, and 2 sin n || 1+sin for n 5, then is in Mn .
n

1 Proof. Let rn = 2 sin 2n for odd n 3, rn = 2 sin for even n 8, and r6 = 3 . n We assume |vu | rn and show that |vud | rn for some d n . Then we can use induction to extend u to an innite sequence d1 d2 ... which fulls (4). Thus we want u to show vud = v + d belongs to B(0, rn ). u If we verify |x| 2 for x = v then Lemma 1 and 2 say that B(x, rn ) intersects u n . Thus there is b with |b| rn and d n with v + b = d . In other words, vu d = b belongs to B(0, rn ). Since |vu | rn was assumed, it remains to show |rn | 2. For n = 3 we have || 1 r3 = 1 and || 2 . For odd n > 3,
2 sin 2n 1 + sin n |rn | (1 + sin ) = || sin n cos 2n n

which is < 2 for n = 5 and, by monotonicity, for n > 5. For n = 6 we have r6 = and || 1 . For even n > 6, cancellation results in |rn | 2(1 + sin ) < 2. 3 || n

1 3

FRACTAL n-GONS AND THEIR MANDELBROT SETS

13

Theorem 8. For all n = 2, 4, the open set condition holds for all on the boundary Mn of Mn . Proof. Take Mn , and assume OSC does not hold. Then 0 is an accumulation point of the vu by Remark 5, and so |vu | < rn for some u in the algorithm of Remark 4. Since vu depends continuously on there is a neighborhood B = B(, ) of such that for each B we also have |vu | < rn . By Theorem 6,(i), can be taken so that all B full the lower bound in Theorem 7. (The only exceptions are n = 3 and = 1 , which concerns the Sierpiski gasket and its reverse, see Section 10, and n 2 sin n = 1+sin for even n = 4p + 2 6, which are discussed in Example 4. In both n cases, OSC holds.) Theorem 6 now says that B Mn , so cannot be a boundary point of Mn . We proved that all for which OSC fails are in the interior of Mn . The converse is not true: a few in the interior of Mn are known for which OSC holds, like the 1 twindragon and tame twindragon for n = 2, the terdragon ( = 2 + 63 i) for n = 3, and Gospers snowakes (see Section 9) for n = 6. Question. Are there in the interior of Mn which full OSC and are not related to tilings? In particular, do such parameters exist for n = 5 ? Is the number of such examples nite for each n ? 8. Mn is regular-closed Our next statement, as well as Theorem 8, remains an open problem for n = 2 [22, 23, 24]. Theorem 9. For all n = 2, 4, the set Mn is the closure of its interior.
m Proof. Let Mn , that is, g() = = 0 for certain dm n . m=0 dm Taking any > 0 we have to verify that there is int Mn with | | < . By 1 1 Proposition 3 this is true for n = 3, 3 and for n 5, 2 . Otherwise, since |dm | 2 for all n and |dm | 3 for n = 3, the function g is analytic in . So the zeros of g do not accumulate at , and we nd such 1 1 that in the ball B(, ), only z = fulls g(z) = 0, and moreover |z| < 2 (< 3 in case n = 3). Let = min{|g(z)| | |z | = }, and let M be so large that e h(z) = dm z m < for |z | . Now the well-known theorem of Rouch m=M M 1 m (cf. [22, 23, 2]) implies that the polynomial g(z) h(z) = m=0 dm z has exactly one zero in B(, ). By Remark 5, int Mn .

We do not study connectedness and local connectedness of Mn by the methods of Bousch [10, 11]. See [2, Section 10] for a discussion of these arguments in case n = 2. For odd n 3, the previous results and computer experiments like Figure 6 indicate that the structure of Mn is very simple. Conjecture. For odd n 3, the boundary of Mn is a Jordan curve. For even n 4 it is a countable union of disjoint Jordan curves.

14

CHRISTOPH BANDT AND NGUYEN VIET HUNG

Figure 6. Left: M3 in a square of side length 105 with center (0.479227, 0.276787). This is one of the suspicious places, but further magnication shows no holes. Right: a hole in M6 . Side length 104 , center (0.33643, 0.15616). Remark 6. (The case n = 4) For n = 4, Lemma 2 reads as follows. B(x, r) intersects 4 whenever |x| 5 and r This allows to prove Proposition 7 with r4 = 1, under the 1. assumption || 1/ 5. This proves Theorem 8 for all M4 with modulus 1/ 5, that is, for the region x 0.4, y 0.2 in Figure 4. In this region, M4 is also regular-closed. Thus although n = 4 was excluded from the theorems, there is at least a partial result, as was proved for n = 2 by Solomyak and Xu [24]. 9. Overlap set and examples for even n Now we discuss the size of the overlap set D = A0 A1 . For n = 2 we have b {1, 1} so that d = 0 has two representations bk bj while d = 2 and d = 2 admit a unique representation. If is the root of just one power series g() and this g has only coecients 2 then D is a single point. If N of the coecients are zero, D has cardinality 2N , and if there are innitely many zero coecients, D is a Cantor set [23, 5]. For even n > 2, the situation is very similar. We shall not consider zero coecients since we conjecture that cannot be the root of only one g() with coecients in when there are zero coecients in q. However, in a regular n-gon with even n 4, as it is formed by the roots bk , k = 0, ..., n 1, every side and diagonal, except for the longest ones, has one parallel side or diagonal, and the corresponding d has two representations bk bj . Thus only those power series g() where every dm has the form bkm bkm +n/2 can yield a single intersection point, see Example 4 below. If we have N coecients dm which do not correspond to the longest diagonals, we shall have 2N points in D, see Figure 2. And when there are innitely many such d then D will be a Cantor set, as in Figure 3.
k

Example 2. For n = 4 we have b = i and the bk form a square. The 4-gon in the lower row of Figure 2 was obtained from the relation of addresses u1 u2 ... = 0 13030 = v1 v2 ... Here um = vm + 2 mod 4, corresponding to a diagonal of the 1 3 square, for all m > 1 except m = 3 and m = 5 where (uk , vk ) = (1, 0) which can

FRACTAL n-GONS AND THEIR MANDELBROT SETS

15

Figure 7. Fractal 6-gons with central piece. Note the dierence between Gosper snowakes with = 5+ 3i and 2+7 3i . 14 be replaced by (2, 3) since i 1 = i2 i3 . Thus there are 4 pairs of addresses which give the same equation
5 = 1 + i = p(13030 = i i + 2 i3 + 4 i . p(01) 3) 1 1

which results in 5 4 + 3 2 + (1 + 2i) = i with numerical solution 0.4587 + 0.1598i. Example 3. As a landmark point in M4 , we consider the intersection of the symmetry line in Figure 4 with the boundary of M4 . It turns out that this is = 1+i . 1+ 5

The corresponding fractal, shown on the left of Figure 3, is mirror-symmetric: A() = A(i) = A(). It seems the only mirror-symmetric 4-gon beside the square which fulls OSC. It is possible to determine by identifying addresses u = u1 u2 ... = 010033221 and v = 133221100. The corresponding equation (4) results in a geometric series with , and the resulting polynomial of degree 4 factors into two quadratic polynomials, of which 2z 2 +(1+i)zi has as the root with modulus 1. Thus 1/ is a quadratic Pisot number over Q(i).

16

CHRISTOPH BANDT AND NGUYEN VIET HUNG

Figure 8. n-gons for odd n with several series g() : Examples 6,7. Note that every second combination (uk , vk ) can be replaced by one other choice. For instance (0, 3) and (1, 2) give the same d = 1 + i. Thus D = A0 A1 is uncountable, and it is a linear Cantor set. In fact D is a self-similar set with respect 1 to two homotheties with factor r2 = ||2 : For E = f0 (D), inspection of Figure 3 2 shows that E = g1 (E) g2 (E) with g1 (z) = f1 (iz) and g2 (z) = f1 f0 (iz). So the log 4 Hausdor dimensions of A and D are log r and log 2r2 , respectively [12], and the log dimension of D is just one quarter of the dimension of A. Clearly D fulls OSC. For A we can explicitly construct an open set - an octagon with interior angles 3 and with alternating side lengths s, rs for some constant s 4 (for our choice of mappings s = 4r). Drawing this octagon U and the fi (U ), we can determine r in alternative way, by elementary geometry. Example 4: Sierpi ski n-gons for even n. We determine all Mn with n sin n minimal modulus || = 1+sin (cf. Theorem 6,(i)). We use the algorithm of Remark
2|| 4, starting with v = vd1 = 1 b which just has the critical modulus 1|| . Since v | | |v| = 2, there is at most one choice of d2 = bk bj for which |vd1 d2 | is not larger than the critical modulus: d2 must be a diameter of the unit circle which is a diagonal of the regular n-gon and parallel to 1 b. For n = 4p such diagonal does not exist. For n = 4p + 2, the only choice is k = p + 1, j = 3p + 2. This leads to vd1 d2 = vd1 , so that by induction dk = d2 for all k > 2. So the single with minimal sin n modulus is the real number = 1+sin . The formula in [21] is more complicated. n By Remark 5, OSC holds.
n

Remark 7. (Adding a central piece for n = 6) The set 6 is special in the sense that it contains the bk , not only their dierences. Thus whenever OSC holds for some , we can add an seventh map f6 (z) = z, and the extended system of mappings fulls OSC. This follows from Remark 5. Moreover, A() is connected if and only if the extended self-similar set is, because A0 A1 = A6 A2 . Figure 7 shows some examples. The Gosper tiles full || = 1/ 7 which explains the bound in Theorem 6,(ii) for n = 6.

FRACTAL n-GONS AND THEIR MANDELBROT SETS

17

10. Overlap set and examples for odd n A regular n-gon with odd n has no parallel sides, and no parallel diagonals of the same length. So d = 0 is the only vector in which can be represented as d = bk bj in dierent ways - actually in n ways. When we have OSC, however, coecients 0 are unlikely to appear in the power series of g(), at least for n 5. So it seems we have the following alternative: either D is a singleton (Figure 1), or is the root of several power series g() (see Figure 8). Example 5: Sierpi ski n-gons for odd n. We determine all Mn with n sin n minimal modulus || = cos +sin (cf. Theorem 6,(i)). We use the algorithm of 2n n Remark 4, starting with v = vd1 = 1 b which has the critical modulus 2 sin n = || v , where = 2 cos 2n . Since | ||v| = , there is at most one choice of d2 = bk bj 1|| for which |vd1 d2 | is not larger than the critical modulus: d2 must be the longest diagonal of the regular n-gon which is parallel to 1 b. As for even n, we conclude that dk = 2 for all k > 2. So the single with minimal modulus is the real number d sin n = cos +sin . The formula in [21] is more complicated. OSC holds by Remark 5.
2n n

Example 6. Considering 2-gons, Solomyak [23] asked whether there is a which is root of several power series g() and still admits nite intersection set D. For n = 3, Figure 8 shows an example, with address identications 0110022221111000 1022111000022221 and 0111110000222211 1222222211110000 . This is obtained from a landmark point 0.4793 + 0.2767i : the intersection of the 30 degree symmetry line with the boundary of M3 . Supported by experiment, we conjecture that we have the same situation for the corresponding parameter on the symmetry line for every odd n, and we do not understand the algebraic reason. For n = 3, the power series derived from the given identications both lead to 4 3 polynomials with the factor 2z + (3 + 3i)z + 2(1 + 3i)z 2 + 1 3i, and is a root of this polynomial. 1 Also on the symmetry line, but far inside M3 , there is the parameter = 2 + 63 i for the terdragon for which we have lots of power series with g() = 0. Instead of this well-known gure, we discuss a similar example introduced in [3]. Example 7. For the second picture in Figure 8, we have the identication 012 100
1 and many others, since D is a Cantor set [3]. Actually, E = f0 (D) is self-similar: 2 2 2 E = f1 (E) f1 (E) where denotes a clockwise 3 rotation with appropriate center. In this case, 0.5222 0.0893i fulls the equation 1 = (1 b2 )2 . It seems not possible to use this equation directly to replace nitely many symbols in the above address identication. But a calculation shows that we can replace 12 and 00 by 1122 and 0001, respectively, which leads to identications 0(12)k 112 1(00)k 0001 and corresponding qk () = 0 for k = 1, 2, ... Thus we have innitely many power series.

18

CHRISTOPH BANDT AND NGUYEN VIET HUNG

The reverse n-gon. For odd n, we call A() the reverse of A(). The Sierpiski gasket is self-reverse, and for Example 4 and the terdragon, the reverse is n mirror-symmetric to A(). If we are not on a symmetry line, however, the appearance of A and its reverse can dier considerably, as Figure 1 shows. Nevertheless, their structure is similar. Since the reverse of the reverse is A itself, if in the following statements means if and only if. Proposition 10. Let n be odd and A = A() a fractal n-gon. (i) A is connected if the reverse is connected. (ii) The cardinalities of the intersection sets D of A and of its reverse coincide. (iii) A is a p.c.f. fractal if the reverse is. (iv) A satises OSC if the reverse does. (v) A is of nite type if the reverse is. Proof. For (i) and (ii) we use Remark 3. D() is described by the sequences d1 , d2 , ... from for which 1 b + dm m = 0, and D() is described by the m=1 sequences d1 , d2 , ... from for which 1 b + dm ()m = 0. Thus between m=1 both sets there is a one-to-one correspondence, given by dm = dm for even m and dm = dm for odd m. If one set is empty or nite, then the other is, too. Moreover, when a point of D() has preperiodic addresses, then the corresponding point of D() also has. So if A is p.c.f. [16], then the reverse is p.c.f. 1 For (iv) and (v) we use neighbor maps fu fv for words u, v {0, 1, ..., n1}m , m = 1, 2, ... [4, 7, 3]. For fractal n-gons we have w = fv (z) = m z + m bvk k1 and k=1 1 fu (w) = w m buk km1 . Thus all neighbor maps are translations: m k=1
m 1 fu fv (z)

=z+
k=1

(bvk buk )km1 .

Now dening u , v by uk = uk , vk = vk for even k and uk = vk , vk = uk we see that an n-gon and its reverse have just the same neighbor maps. OSC means that neighbor maps cannot approach the identity map [4], so (iii) is proved. Finite type 1 says that there are only nitely many neighbor maps fu fv with Au Av = [7]. Au , Av are disjoint i the dk = bvk buk can be extended to a sequence d1 , d2 , ... for which 1 b + dk k = 0. With the above construction, this implies (iv). Acknowledgement. Most of this work was done when Nguyen Viet Hung visited the University of Greifswald, supported by the Ministry of Education and Training of Vietnam and by the DAAD. References
[1] S. Akiyama and N. Gjini, Connectedness of number-theoretic tilings, Discrete Math. Theoret. Computer Science, to appear [2] C. Bandt, On the Mandelbrot set for pairs of linear maps, Nonlinearity 15 (2002), 11271147. [3] C. Bandt, Self-similar measures, Ergodic Theory, Analysis, and Ecient Simulation of Dynamical Systems, B. Fiedler (ed.), Springer (2001), 3146. [4] C. Bandt and S. Graf, Self-similar sets VII. A characterization of self-similar fractals with positive Hausdor measure. Proc. Amer. Math. Soc. 114 (1992), 9951001.

FRACTAL n-GONS AND THEIR MANDELBROT SETS

19

[5] C. Bandt and N.V. Hung, Self-similar sets with open set condition and great variety of overlaps, Preprint, 2007. [6] C. Bandt and K. Keller, Self-similar sets 2. A simple approach to the topological structure of Fractals, Math. Nachr. 154 (1991), 2739. [7] C. Bandt and H. Rao, Topology and separation of self-similar fractals in the plane, Nonlinearity 20 (2007), 1463-1474 [8] M.F. Barnsley, Fractals Everywhere, 2nd ed., Academic Press 1993. [9] M.F. Barnsley and A.N. Harrington, A Mandelbrot set for pairs of linear maps, Physica D 15 (1985), 421432. [10] T. Bousch Sur quelques probl`mes de dynamique holomorphe, Ph.D. thesis, Orsay e (http://topo.math.u-psud.fr/bousch), 1992. [11] T. Bousch, Connexit locale et par chemins hlderiens pour les syst`mes itrs de fonctions, e o e e e Preprint (http://topo.math.u-psud.fr/bousch), 1993. [12] K.J. Falconer, Fractal Geometry, Wiley 1990. [13] K.J. Falconer and J.J. OConnor, Symmetry and enumeration of self-similar fractals, Bull. London Math. Soc. electronic version 27 Feb. 2007 [14] M. Hata, On the structure of self-similar sets, Japan J. Appl. Math. 2 (1985), 281414. [15] J.E. Hutchinson, Fractals and self-similarity, Indiana Univ. Math. J. 30 (1981), 713747. [16] J. Kigami, Analysis on fractals, Cambridge University Press, 2001. [17] I. Kirat and K.-S. Lau, On the connectedness of self-ane tiles, J. London Math. Soc. 62 (2000), 291-304. [18] I. Kirat, K.-S. Lau and H. Rao, Expanding polynomials and connectedness of self-ane tiles, Discrete Comput. Geom. 31 (2004), 275-286. [19] T. Lindstrm, Brownian motion on nested fractals, Mem. Am. Math. Soc. 83 (1990), 1124. [20] B.B. Mandelbrot, The Fractal Geometry of Nature, Freeman, San Francisco 1982. n [21] S. Schlicker and K. Dennis, Sierpiski n-gons, Pi Mu Epsilon J. 10 (1995), 8189. [22] B. Solomyak, Mandelbrot set for pairs of linear maps: the local geometry, Analysis in Theory and Applications 20:2 (2004), 149157. [23] B. Solomyak, On the Mandelbrot set for pairs of linear maps: asymptotic self-similarity, Nonlinearity 18 (2005), 19271943. [24] B. Solomyak and H. Xu, On the Mandelbrot set for a pair of linear maps and complex Bernoulli convolutions, Nonlinearity 16 (2003), 17331749. [25] R.S. Strichartz, Isoperimetric estimates on Sierpiski gasket type fractals, Trans. Amer. n Math. Soc. 351 (1999), 1705-1752.

Christoph Bandt Institute for Mathematics and Informatics Arndt University 17487 Greifswald, Germany bandt@uni-greifswald.de

Nguyen Viet Hung Department of Mathematics Hue University Hue, Vietnam nvh0@yahoo.com