2. Interpolation

The following theorem supplies the basic elements necessary for construction of generalized fractal interpolation functions.

Theorem 1. Let the interpolating data set D = {Pi = [x y zi]T Î R3, xi < xi+1, i = 0, 1,.., n}(n ³ 2)  be given. Consider the IFS s(D) = {R 3; w1,..., wn}, where  wi are given by (1) with

(4)

and the real numbers  di, hi, li, mare parameters chosen such that wi (i = 1,..., n ) are contractive mappings.  Then, the limiting set

FD= limk ®¥W k( G ),

is the graph of a continuous vector valued function  f : [x0, xn] ® R2 having the interpolation property  f(xi) = [ yzi]T ,  i = 0,1,..., n.

Proof.  It is straightforward to see that, with the special choice of  ei, fi, ggiven here, the transformations in the IFS take the form
 

 wi(P) PiA (P - Pn )                                                        (5)


which give immediately   wi(P0) = Pi-1 and  wi(Pn) = Pi   for any  i = 1,2,..., n. The rest of the proof  follows from Theorems 4.1 and 4.2  in [3, ch. VI ].  à

For  i = 1,..,n, denote by Mi the sub-matrix of  A whose elements are di, hi, li, mi
 

                                               (6)

The conditions that  di, hi, li, m must obey in order that a transformation of the form of  wÎs(D) be a contraction, are given in the following lemma.

Lemma 1.   A transformation whaving  the form (1)with  ai < 1   is a contraction of  R3 if and only if the elements  di, hi, li, mi  of  Ai satisfy one of the following two systems of inequalities
 

                        (di-mi)2 + 4 hi l³ 0      and     dimi- hi li - | di + mi | + 1  >  0 ,                (7a)

     (di-mi)2 + 4 hi li  < 0     and     1 -dimi + hi li > 0 .                                    (7b)

Proof.  Denote the spectral radius of a matrix  by r(M). Then
 

r( Ai ) < 1                                                                       (8)

is a necessary and sufficient condition for the mapping wi to be a contraction w.r.t. a feasible metric d in R 3.

Let  Qi = (di-mi)2 + 4 hi li , and let l1(i) , l2(i) , l3(i)  denote the eigenvalues of  the matrix  A in (2),  more precisely  let l1 (i) = ai, l2(i)= 1/2( di+ mi -  ÖQi ) and l3(i)= 1/2( di+ mi+  ÖQi). Then, (8) is equivalent to

| lk(i)| < 1,      k = 1, 2, 3

and since | l1(i) | = | ai | < 1 holds by assumption, only the behaviour of l2(i) and l3(i) is left to be examined ( notice that these are common  eigenvalues of  Ai and of its submatrix  Mi ) .

Let Q ³ 0 ( this is equivalent to the first inequality in the first system in the assertion ). Then, l2(i) , l3(i) Î R,  and  l2(i) £ l3(i).So the condition {|l2(i)| < 1 Ù |l3(i)| < 1 } is equivalent to { -1 < l2(i Ù  l3(i) < 1 }, or  {ÖQ < di + mi + 2   Ù   ÖQi < -di-mi + 2} which, in turn, is equivalent to the set of inequalities

            ÖQi < 2 - di- mi ,       di+ mi ³ 0,

            ÖQi < 2 + di+ mi ,       di+ mi < 0.

By squaring both sides of  both  left inequalities and  replacing  Qi = (di-mi)2 + 4 hi li, the  above set  reduces to the second  inequality of (7a).

If, instead,  Q < 0, the first inequality in the second system is satisfied. Furthermore, in this case  l2(i) and l3(i) are conjugate complex numbers having common modulus  |l2(i)|  |l3(i)| =  1/2| di + mi + i Ö(-Qi )|  = 1/2 Ö[(di+mi)2- Qi ], therefore  the condition {|l2(i)| < 1  Ù |l3(i)| < 1} yields  Ö[ (di+mi)2- Qi ] < 2  which reduces to Ö(dimi-hi li) < 1,  and to the second  inequality of (7b), which is thereby completely proved. à

The folowing corollary gives the condition on matrices  Mi ( i = 1,..,n) under which the IFS is hyperbolic.

Corollary 1. The IFS s(D) = {R3; w1, ..., wn}, built according to the rules given in Theorem 1,  is hyperbolic if and only if for all  i = 1,..., the elements of the parameter matrix Mi satisfy the conditions (7a) and (7b).

Proof. It is straightforward since the hypotheses made in Theorem 1 about  D yield ai < 1 ( i = 1,.. ,n) and, therefore, Lemma 1 applies. à

Let us denote  by M the set of all the parameter matrices in the IFS  s(D), M = {Mi, i = 1,..,n}. It is clear, now,  that the uniqueness of the attractor of s(D) and, of course, the shape of the attractor itself depend on the parameter matrices M just as well as on the interpolation data D. Therefore, in the following, whenever this dependence will be relevant to our discussion we will introduce the set of parameter matrices in our notation, denoting by  s(D,M) the corresponding IFS.  Also we will denote by the same symbol  the corresponding  interpolation scheme, namely the rule by  which the continuous vector valued function  f , being the unique attractor of  s(D,M), is associated  to D.

Example 1. Consider the data set D= {(0, 0, 0), (1/4, 0, 1/2), (1/2, 1/2, 1), (3/4, 1, 1/2), (1, 1, 0)}. By Theorem 1, a four term IFS s(D) = {R 3; w1, w2, w3, w4}, such that its attractor is the graph of a vector valued fractal interpolation  function f, can be associated to D. If  the items of {Mi , i = 1,..,4} are chosen, for example, to be :  d1 = d2= 1/2, d3 = -1/4, d4 = 0; h1 = h4 = 1/4, h2 = -1/4, h3 = 1/2; l1 = l4 = 1/4,  l2 = l3 = -1/4; m1 = 0, m2 = 1/4,  m3 = -1/4, m4 =1/2, then the affine mappings w1, w2, w3, w4   have the form  (2), with

Direct checking shows that this choice of  di, hi, li, mi  satisfies constraints in  Lemma 1, so the coefficient matrices above define contractive linear mappings R 3® R 3. Indeed, one could also evaluate directly r( A1) = r( A4) = (1+ Ö2 )/4 » 0.603553,   r(A2) = ( 3 + Ö5 )/8 » 0.654508,  and r(A3) = Ö3 /4 » 0.433013,  to discover that all are smaller than 1, which guarantees that the IFS s(D) is hyperbolic and its attractor FD is unique.

In Figure 1 (top) the preattractors W 1(G), W 2(G) and W 3(G) are displayed, where G is the piecewise linear interpolant of D. The rest of Fig. 1 shows the projections of these preattractors on the coordinate planes. The interpolating points are marked in all projections.

Figure 1. Preattractors of  FD and their projections on the coordinate planes

Figure 2 (left)  shows the attractor of the orbit of  G, this is FD =  {[x f(x) y(x)]T, x ÎI= [0,1]} which is the graph of   f (x) = [f(x) y(x)]T . The projections of  FD on the coordinate planes,  f ={ [x f(x)]T, xÎI}, y={[x y(x)]T, xÎI}and the parametric curve  f = (f,y) are shown in Figure 2 (right). The functions f and y are called hidden variable fractal interpolation functions (Barnsley [2], [3], Massopust [14]) meaning that they depend on the "third variable" that is not present in the plane containing their graphs.

In fact they do share the interpolation property of FDbut, while FD is, by its own nature, a self affine set, in the general case the projections f and y are not self affine.

Figure 2. Components of  f .


NEXT