You are on page 1of 12
( ig Ie EAST JOURNAL ON APPROXIMATIONS Volume 3, Number 4 (1997), 381-392 INTERPOLATION BY FRACTAL FUNCTIONS PRESERVING MONOTONICITY AND CONVEXITY S.N. VASILYEV Department of Mathematics and Mechanics, Ural State University, Pr. Lenina 51, Ekaterinburg 620083, RUSSIA ‘The aim of this paper is to determine conditions under which the fractal interpolating function preserves monotonicity or convexity of the inter- polated data. Necessary and sufficient conditions are found for these cases of shape-preserving. The results in monotonicity case generalize those obtained in [4]. 1. Introduction We consider here the standard interpolation problem: Given a data a={(*)} ; (=) ER’, 2) >2;-1 fori=1,2,...,n, vi i=0 vi of points {x;} and values {y;}, construct a function f from a preassigned class which satisfies the conditions f(z;) = y; fori = 0,...,n. Fractal interpolation is used to obtain functions with complicated, highly irregular structure, such as Cantor’s staircase. Such functions are used in computer aided geomet- ric design (CAGD), especially for obtaining realistic contours of mountain ranges, skylines over a forest, sea coast shapes, courses of rivers etc. One of the specific demands on interpolation (it is also a CAGD demand) is that the interpolating function must preserve the “shape” of the data being in- terpolated. The simplest mathematical descriptions of a “shape” of a certain function are its monotonicity and convexity. So, there is a task of constructing a fractal interpolating function f preserving monotonicity and/or convexity of the data. 382 INTERPOLATION BY FRACTAL FUNCTIONS In [1, 2] Barnsley has suggested a method for interpolation by fractal functions. Namely, to a given data A a set of affine mappings W1,W2,...)W_ :R? > R? is associated, so that w; has the form 2 aze + bj ) 1 4; = % (3) (cetay te with {a;,;, ci, dj, e;}%, satisfying the conditions 0; a) = (F) meee (=) - (5). Yo vi-1 Yn vi For each i = 1,...,n, four conditions are imposed on five coefficients. Thus, considering one of them, say d;, as a real parameter, one can determine the others by a = SOF b T— aiTn, Tn — Zo a) KM. Yn — Yo i Yin in Zn — Zo ae’ ci & = Vi-Citn— didn In [2] a metric on the class of subsets of [z0, zn] x R is described such that in this metric the contraction (Lipschitz) factor of w; is |d;| and the contraction factor of the Hutchinson operator W(A) = Cea) (AcR’) i=1 isd= ax {|d;|}. It is assumed that 0 0. It is clear that any graph of an increasing (“increasing” here means the same as “non-decreasing”) function is an increasing set according to this definition. Moreover, if a gtaph of some function h:R>R Gh = {(#) :2e€DC Rn} is an in- creasing set, then for any 21,22 € D, 2; # 22, it follows from (4) that A(z1) - h(z2) " Hence, the graph of a function is an increasing set if and only if the func- tion is increasing. Thus, for an increasing data set A, in order to construct a monotonicity-preserving interpolating fractal function, such an IFS Sq should be constructed that its attractor Gy is an increasing set. Note that from w; (=) = (F) and w; (=) = (=) it follows that Yo Yi-1 Yn Yi 4 Cc W(A),. Furthermore, since the fractal function interpolates the data A, we have A C Gy, and by (3), W*(A) c W*(Gy) = Gy for any k = 0,1,2, Therefore, > m(Gi) > 0 which implies that A is an increasing function. (5) AcW(A) CWA) C++ C WHA) C++ Gy. Let us denote by mg and M,, respectively, the minimal and maximal slope of WEA), i.e. my = m(W#(A)) and My := M(W#(A)). It follows from (4) and (5) that mo =m, > m2 >...> mM, >... > MGs), 6 fo MoS My <2. My S...5 MG). This is the base which makes it possible to prove the following. S.N. VASILYEV 385 Lemma 1. Gy is increasing if and only if W*(A) is increasing for any k =0,1,2,... Proof. If for some k the set W*(A) is not increasing (m, <0), then m(Gys) < mx < 0, hence Gy is not increasing. On the other hand, if Gy is not increasing, then there exist two points uv €Gy such that up > v, and ty = f(s) < f(vs) = vy. Since Gy is a graph of a continuous function f, there exists ¢ >0 such that for any PE (Ur —€,tUy +e) andg €(vz—£,0z +), f(p) < f(q). Note that the max- imal distance between the first coordinates of neighboring points in W*(A) is (max, (2: ~ i-1))-(,max ai) which tends to 0 as k + oo. Therefore, for a given e there exist k > O and u',v/ € W*(A) such that wl, € (tee, ue-+e) and v, € (V2—€, 02+). Due to W*(A) C Gy we have wy = f(u,) < f(v,) = vy, hence I(u’,v’) <0. This implies that my = m(W#(A)) < U(w',v’) < 0 and thus W*(A) is not an increasing set. o Remark. It can be proved also that lim m, = m(G,) and lim My = M(Gy). keveo k-vo0 For u,v € R?, uz # vz, let us denote di(l(u, v)) = Mwi(u), wiv). One can notice that since w; is an affine transformation, Ci(Ux = Ve) + di(ty = ry a;(uz — Ye) This implies that 4;(l(u,v)) is independent of u and v, and depends on the value I(u, v) only. So 4; can be defined as follows: (7) oi: RR, o() = 24 St, $i(K(u, v)) = Let us take A = W*(A) for some k = 0,1,2,.... One can check that A is a finite set, so we can write A = {u'}i*), ul-! < ui for i=1,...,m, and W(A) = {vA J, of < vb for j = 1,...,me¢1- By Proposition 1, mW(A)) = min {U(v?,v’)} jetentegs = _axin,{ in {uu anlui »3} pele min, {min fe (Ui? wf yy}. 386 INTERPOLATION BY FRACTAL FUNCTIONS Since all ¢;(2) are linear they attain their extremum values at the maxima and minima of their domains. Therefore we get min {min {tr uy} reln = amin, min {456mm (uy) dr aman al w)) = amin min {4 (mx) 5 br (me)} : Thus (8) Mey = min { min{¢i(mx) , #(Mi)}} - Similarly, it is easily shown that ©) Mag = max, { max( (ms) , &(Mi)}} Now we can prove the following. Lemma 2. W*(A) is increasing for any k = 0,1,2,... if and only if at least one of the following conditions is satisfied: (a) There exist s € iv mo] and S € [Mo, +00) such that $;(1) € [s, S] for all 1 € [s, S], ¢ ) There exists s a ‘fo, mo] such that $;(1) € [s,00) for all 1 € [s,co), fala Proof. If condition (a) is satisfied, s < mz and My < S, then on combining this with (8) and (9), we have $$ , min, {inin {6;(me), 4:(Mu)}} = M15 Mit = max, { max{4;(mx), 6e(Mu)}} <5. Therefore, from s < mo and Mg < S$ it follows that 0 < s < m, and My < S for any k = 1,2,... This implies that all W*(A) are increasing. Similarly, if condition (b) is satisfied, then from mo > s it follows that mg, > s > 0 for any k =0,1,2,... Conversely, suppose that W*(A) is increasing (i.e., my > 0) for any k= 0,1,2,... Then (see (6)) we have jim my = may 2 0. It follows from (8) that (10) Gi(Moo) = lim 4i(mu) 2 Him, , min {min {d5(me) > $j(Mx)}} fim mugs = Mo forany @= 1,047 S.N. VASILYEV 387 If im Mg = +oo, then all ¢;(1) are increasing (otherwise, for 4;(!) which is strictly decreasing we would have 0 < my < $3(Mz-1) —> —00).: This 00 implies that for a given s, ¢;(1) € [#i(s),00) whenever J € [s,00). Hence it follows from (10) that condition (b) is satisfied for s = mo. If Jim Mg = Mco < +00, then similarly to (10) it is easily proved that for any i= Ty-..,n We have Gi( Moc) $ Moos ditoc) $ Moo attd 4i(Moo) 2 to. Thus, for any 1 € [™m0o;Moo]s (1) € [4i( 0), Fi(Moo)] S [77005 Moo] when- ever @; is increasing, and (1) € [4;( Moo), $i(™co)] E [7Me0; Moo] whenever ¢; is decreasing. Thus s = mo. and 5 = Mc satisfy condition (a). Let us denote by r; the slope of the segment (52) , (®)} ,and by r i—1 hi to Ty the slope of the segment [( ) ml |: We have yo) \ dn Fe forte nee eet @j- Ti-1 In — Zo By Proposition 1, for any non-collinear data A, mo = min r; max ti i=1)..)n. f=] jn and mn nH ny ~«i-min {5 ue 0) and gi(s) > 8, ie. 2+ T(s—7) 2s from whence condition (b) of Theorem 1 follows. ‘ Fractal interpolating function is invariant under reflection with respect to y-axis. Thus the same conditions apply to decreasing data. The theorem is proved. a Thus, in order to construct monotonicity preserving fractal interpolat- ing function one can take 0 < s < min 17; and S > max 7 (or just isl FT yeast 0 0 such that Se) = FO"), fp*) = F(v*) ppt P-pP whenever pi € (ui,—e,uite) for i= 1,2,3. Again, for a given e there exist k and v!,v?, 0? € WH(A) such that vf € (ui. —e,ui. +) for i= 1,2,3. Due to the relation W#(A) C Gy, we have vi = f(vi), and therefore U(v!, v?) > U(v?, 0°). Thus, if Gy is not convexial, there is a k for which the set W*(A) is not convexial too. a Let us remark that if the data A is convexial, then mo =11 S12 $.---S t= Mo. This is generalized by the following Proposition 2. A set A = {v°,v!,...,v'+1} C R? with wcu<...3, is convexial if and only if I(vj-1,0;) < U(vj, 0541) for each j = 1,...5h. Proof. It is evident that the conditions (v;_1, vj) < I(vj,0j41) are neces- sary for the set A to be convexial. Conversely, suppose that for any 5 = 1,...,k we have U(vj-1,0;) < U(vj,0;41). Then, for i1,i2,é3 such that 0 m({v9, v4, .., 0) = U(0%, of), 390 INTERPOLATION BY FRACTAL FUNCTIONS Thus Uv, o8) < 1(0 1, v2) < 1(v2, v2) < 1(v, v9) for any 0 < i; < ig < i3 < k +1. This implies that the set A is convexial. O Note that if ¢; is increasing (i.e., if d; > 0), then the affine transformation w preserves convexial property, and if ¢; is strictly decreasing (d; < 0), then the image of any non-collinear convexial set is not convexial (see Proposition 2 and the definition of convexiality). Therefore, in order to secure the property of the Hutchinson operator W to preserve convexiality of a data A (the data is assumed to be non-collinear), at least all ¢; must be increasing. This also implies that (8) and (9) can be reduced to met = | min Gi(me) and Mey = mx, $i( Mx). Moreover, m (wo: (W#(A))) = de(me) and M (w (WA(A))) = di(M). Suppose that A = W*(A) is convexial for some k = 0,1,2,... and all 4; are increasing. Then, by Proposition 2, W*+1(A) = W(A) is convexial (since w;(A) are convexial) if and only if i-1(Mk) < O:(m,) for i= yn. The conditions that all 4; are increasing and $i-1(Mz) < 4i(mx) for i= 2,...,n can be written in the following way: (11) di(me) < b1(Me) < d2(me) < .--< bn-1(Mk) < nlm) < bn( Me). Thus, if W*(A) is convexial, then (11) is a necessary and sufficient con- dition for WE+1(A) to be convexial. Note also that if (11) holds, then muti = $1(mg) and Migr = bn(Me)- Now we can prove the following statement. Lemma 4. Let A be a non-collinear convezial data. Then W*(A) is convezial for any k = 0,1,2,... if and only if all $; are increasing and at least one of the following conditiuns is satisfied: (a) There exist s such that s < d1(s), dn(S) < Sand $i-1(5) < :(s) for i = 2,...,n. (b) $1,62,-+-,@n—1 are constant and $n(r1) > tn-1- (c) dita) tn—1 must be satisfied. It is clear now that it is impossible to have the both equalities Jim my = —co and Jim My = +00 = 00 (otherwise ¢; must be constant for any j = 1,...,n, therefore my = mo = 1m > —0o and My = Mo = t < +00). The proof that s = m.. and $ = M,, satisfy condition (a) is similar to that one in Lemma 2 without any significant changes. a Our main theorem for convexial set follows from the above lemma. Theorem 2. The fractal interpolating function f given by (1), (2) and (3) preserves converity of a data if and only if at least one of the following conditions is satisfied: (a) There exist numbers 8,9 €R, s <1, 5 >t, such that (b) dj =0 for i=1,...,n—1 and 0< dy < ay (c)0 0, 41(s) > 8 bn(S) < S, and for any i= 1,...,n—1, di(S) < di41(s). The last inequality means that a disn e+ 2S = 7) < rig — SE - nt oS —) Sti tn 8) from whence the conditions in (a) follow. 392 INTERPOLATION BY FRACTAL FUNCTIONS Conditions (b) and (c) of Lemma 4 can be reduced trivially to those of Theorem 2. a Remark, It can be shown in the usual way that the fractal dimension (see [3] for definitions) of a graph of a monotone or convex function is 1. Besides, this fact can be obtained from the restrictions given in Theorem 1 or Theorem 2 using the method of Barnsley [2]. Acknowledgement. I wish to thank V.I. Berdyshev for his suggestion to start working on this theme and his continuing interest and comments. I would also like to thank A. V. Mironenko and E. V. Kazantseva, who carefully read the manuscript, for their helpful critics. References {l] M. F. BaRNsLey, Fractal functions and interpolation, Constr. Approz. 2 (1986), 303-329. [2] M. F. BarNsLey, Fractals Everywhere, Academic Press, 1988. 3] J.E. Hurcninson, Fractals and self similarity, Indiana Univ. Math. J. 30, 5 (1981). [4] Lu. M. Kocté, Monotone interpolation by fractal functions, In: Proceeding of International Conference on Approximation and Optimization (Romania), (Cluj-Napoca, July 29 - August 1, 1996), to appear. Received April 23, 1997

You might also like