Constr. Approx.

(1986) 2:303-329
CONSTRUCTIVE
APPROXIMATION
(6) 1986 Slaringer-VerlagNew York Inc.

Fractal Functions and Interpolation

Michael F. Barnsley
Abstract. Let a data set {(x~, y~) ~ I x R; i = 0, 1 , . . . , N} be given, where I =
[Xo, XN]C R. We introduce iterated function systems whose attractors G are
graphs of continuous functions f : I ~ R, which interpolate the data according
to f(x,)=Yl for i~ {0, 1. . . . . N}. Results are presented on the existence, coding
theory, functional equations and moment theory for such fractal interpolation
functions. Applications to the approximation of naturally wiggly functions, which
may show some kind of geometrical self-similarity under magnification, such as
profiles of cloud tops and mountain ranges, are envisaged.

1. Introduction

We introduce some continuous interpolation functions f : I -> R, where I is a real

closed interval, which appear ideally suited for the approximation of naturally
occurring functions which display some kind of geometrical self-similarity under
magnification. The types of functions we want to consider approximating include
not only those which describe profiles of mountain ranges, tops of clouds, stalactite
hung roofs of caves and horizons over forests, but also wilder functions whose
self-similarities may be more of a stochastic origin, such as temperatures in flames
as a function of time, electroencephalograph pen traces and even the minute-by-
minute stock market index.
We refer to the new interpolation functions as fractal because it can occur that
the Hausdorff-Besicovitch dimensions of their graphs are noninteger [M]: typi-
cally these functions belong to Lip 8 for some 0 < 8 < 1, which implies they are
not differentiable. However, they are not stochastic and are distinct from those
treated by Mandelbrot [M] and Fournier, Fussell and Carpenter [FFC], which
have been used to produce computer graphics simulations of landscapes. In the
present situation the function values are determined uniquely by the given data
points together with a number of parameters, and in this way they are analogous
to splines and polynomial interpolations. Because of their ability to extrapolate
patterns from one scale to all scales, their usage can reflect the expectations of
the practitioner concerning the presence of deep-hidden determinism in
apparently disorganized data.

Date received: May 23, 1985. Date revised: May 15, 1986. Communicated by Charles A. Micchelli.
AMS classification: 26A, 28A, 41A, 44A, 60J.
Key words and phrases: Interpolation, Fractais, Moment theory.

303
304 M.F. Barnsley

The fractal interpolation functions introduced here have their roots in the
theory of iterated function systems [H, DS, BD 1, BD 2], which is reviewed
briefly in Section 2. In Section 3 we introduce the interpolation functions them-
selves, and pick out several natural classes which are easy to work with. We also
comment on the use of condensation sets. In Section 4 we describe the associated
coding theory and measure theory. The former yields explicit formulas for
calculating function values, while the latter are exploited in Section 5 in the
evaluation of integrals: for a wide class of fractal interpolation functions f(x)
we show how to evaluate explicitly moment integrals of the form

f x"f(x) m dx, n, m =0, 1, 2,...

I

in terms of the interpolation points and the other parameters which fix f(x).
These numbers can be used in the design of effective approximation schemes.
We also use the invariant measures to discover functional equations obeyed by
Fourier and Laplace transforms of fractal interpolation functions. In Section 6
we compute upper and lower bounds on the Hausdorff-Besicovitch dimension
of fractal interpolation functions and show how these can be methodically
sharpened. We also consider connections with some functions introduced by
Besicovitch and Ursell.

2. Iterated Function Systems

We set up notation and briefly review ideas fixing the framework of the rest of
the paper. The main references are [H, BD 2, DS].
Let K be a compact metric space with distance function d(x, y) for x, y c K.
Let H be the set of all nonempty closed subsets of K. Then H is a compact
metric space with the Hausdorff metric
h(A, B) = Max{Sup Inf d(x, y), Sup Inf d(x, y)},
x~A yEB xe~B y e A

which is defined whenever A and B are subsets of K.

Let wn : K --> K for n c {1, 2 , . . . , N} be continuous. Then we call {K, wn: n =
1, 2 , . . . , N} an iterated function system (i.f.s.). Define W: H--> H by
W(A)=Uw~(A) for A~H,
n

where w,(A) = {w,(x): x ~ A}. Any set G e H such that

w(~) = G

is called an attractor for the i.f.s.; and the i.f.s, always admits at least one attractor.
Indeed, one needs only to start with any S ~ H and form the set of all accumulation
points of sequences {Sm}~=I, where S~ ~ W~ and W(W~ for m =
1, 2, 3 . . . . with W~176 = S.
If, for some 0 - s < l and all h e { I , 2 , . . . , N},
d(w~(x),w~(y))<_s.d(x,y) for all x,y~K,
Fractal Functions and Interpolation 305

hyperbolic.In this case

then the i.f.s, is termed W is a contraction mapping; it obeys

h(W(A), W(B))<-s.h(A,B) for all A,B~H;

and consequently admits a unique attractor

G = Lira
m--~oo
W~ in H.

When the attractor G of an i.f.s, is unique, it can be calculated as follows. Let

p > 0 be a probability vector p = (Pl, P2,. 9 PN) with each Pn > 0 and ~nP, = 1.
Start from Xo~ K and define a sequence {xm} by choosing successively

x,, E {w~(x,,_l), Wz(Xm-O,..., wN(x,,-0} for m e {1, 2, 3,...},

where probability p, is attached to the event x,, = w~(xm-a),Then, almost surely,

G consists of the set of points g ~ K such that each open neighborhood of g
contains infinitely many x,,'s. When K = [0, 1] x [0, 1] c R 2 for example, such
computations are readily effected and displayed on a microcomputer by plotting
{xm: m = 51, 5 2 , . . . , 5000}, say. If the i.f.s, is hyperbolic then the x,~'s will
approach a distribution given by the unique probability measure/~, on K which
is stationary for the discrete time Markov process where the probability of transfer
from x c K to a Borel subset B of K is

P(x, B) =Y.p, Sw~ B).

rl

Here By(B) equals one if y ~ B and zero otherwise.

Regardless of whether or not the i.f.s, is hyperbolic, there always exists such
a stationary probability measure/x; and it is referred to as a p-balanced measure
of the i.f.s, When the attractor G of the i.f.s, is unique, it follows from [BD 2,
Theorem 2] that the support of/~ is G independently of p > 0.
The stationarity o f / z is equivalent to

(2.1) fKfdtx= fr ~p,cfow, dtz forall f~C,

where C denotes the continuous real-valued functions on K. Let P(K) denote
the set of probability measures on K. For given Vo~ P(K) define ~,,, ~ P(K) by

~',,(B)=~.p,v,,_l(w~(B)) for m e{1,2,3,...}.

Then for hyperbolic i.f.s, v,, converges to ~ in the sense that

,~-~LimfKfdv'=frfdtz for each f~C.

These results have applications to computer graphics, see [DS] and [DHN].
Although the above formulation in terms of a Markov process seems very
stochastic, a completely deterministic description can be made; see the Remark
in Section 4.
306 M.F. Barnsley

3. Fractal Interpolation Functions

Let a set of data points {(xl, Yl)e I x R: i= 0, 1 , . . . , N} be given, where I =
[Xo, xN] c R is a closed interval, see Fig. 1. We are concerned with continuous
functions f : I--, R which interpolate the data according to

f(xi)=y~, i=0,1,2 .... ,N,

see Fig. 2. We focus on the existence, construction, and properties of such

functions f : I--> R whose graphs

O = { ( x , f ( x ) ) : x 6 I}
are attractors ofi.f.s. That is, there is a compact subset K of I x R, and a collection
o.f continuous mappings w, : K -->K such that the unique attractor of the i.f.s.
{K, wn: n = 1, 2 , . . . , N} is G. We shall refer to such functions f as fractal interpola-
tion functions: often the Hausdortt-Besicovitch dimension of G will be nonin-
teger; f may then be H/bider continuous but not differentiable.
We will work in the compact metric space K = I x [a, b] for some - ~ < a < b <
oo, with the Euclidean metric or the equivalent metric

(3.1) d((c~,dl),(c2, d2))=Max{Icl-c2l, ld~-d2l} for (c,,dl),(c2, d2)cK.

Set In = [xn-l, x,] and let L,: 1--> In for n ~ {1, 2 , . . . , N} be contractive homeo-
morphisms such that
Ln(xo) = xn-1, Ln(xN) = x~,

IL.(cl)-Ln(cz)l<-l.lc~-c2l whenever Cl, C2eI ~

for some 0 -< l < 1. Furthermore, let mappings F, : K ~ [a, b] be continuous with,
for some 0 -< q < 1,
Fn (Xo, Yo) = Y,-I, Fn(xN, YN) = Y,,
(3.2) ]Fn(c, d l ) - F~(c, d2)l-< q" I d l - d21
for all c e / , dl and d2E[a, b], and n ~ { 1 , 2 , . . . , N}. Now define functions
w n : K - > K for n e { 1 , 2 , . . . , N} by
wn(x, y) = ( L , ( x ) , Fn(x, y)).
Then {K, wn: n = 1, 2 , . . . , N} is an i.f.s, but may not be hyperbolic.

Theorem 1. The i.f.s. {K, w,: n = 1, 2 , . . . , N } defined above has a unique attractor
G. G is the graph of a continuous function f : I-> [ a, b] which obeys
(3.3) f ( x , ) = y, for i = 0, 1, 2 , . . . , N.

Xo XI X2 X; X4 X5

Fig. I
Fractal F u n c t i o n s a n d I n t e r p o l a t i o n 307

f
-1/1

Y4
Yo
Y2

Y3

~. 9 ~ 9 9 px
XO XI X2 X3 X4

Fig. 2

Proof. Let G be any attractor of the i.f.s. Let

= {x ~ I: (x, y) ~ G for some y ~ [a, b]}.
Then, since G=U,w,(G), it follows that I = U , L , ( I ) . But, {I,L,:n=
1, 2 , . . . , N} is a hyperbolic i.f.s, whose unique attractor is I, whence
i = I = [Xo, xN].
Next, we show that G is the graph of a function f : I-~ [a, b] by proving that
only one y-value corresponds to each x-value. First we show this for the x-values
{x0, x l , . . . , xN}. Let
S~ = {(x, y) e G: x = x~} for i e {0, 1 , . . . , N}.
Since w,(So)n S o = 0 for n # 1, we must have w~(So)=So. But Wl is a strict
contraction on the compact metric space So, whence So = (Xo, Yo). Similarly,
SN = (xN, YN). For I c {1, 2 , . . . , N - 1} we observe that the only points which
can map to S~ are So (under w~_l)and SN (under wi), from which S~ = Wi-l(So) u
w,(SN) = (x,, y,).
Consider
= M a x { I s - tt: (x, s) ~ G, (x, t) ~ G, some x ~ I}.

By the compactness o f G this maximum is achieved at some pair of points (~, s)

and (s t) in O, with

~=ls-tl,
and from the last paragraph we can assume ~ ~ (x~_~, x,) for some n. But then
there exists two points
(L~I(s u) and (L:~(s v) b o t h i n (3,
with
s= F.(L:I(~), u) and t= F.(L~I(~), v).
308 M.F. Barnsley

Hence
8 = Is - t[ = [Fn(L~(~), u) - F. (L~I(.~), V)I
<-q" l u - v l < - q 9 8,
with 0 _ < q < l whence 8 = 0 . Hence G is the graph of a function f : I ~ [ a , b]
which obeys (3.3). G is unique because the union of two attractors is an attractor.
To prove f ( x ) is continuous, let C ( I ) denote the Banach space of continuous
real-valued functions g : I ~ R, with norm IgL= Max{Ig(x)l: x 9 I}. Let Co(I)'-
C ( I ) consist of those g ~ C ( I ) such that g: I ~, [a, b), and which obey g(Xo) =Yo
and g(XN) = YN. Then Co(l) is a complete metric space and the i.f.s, induces a
contraction
T: Co(I)-} Co(I)
defined by
(Tg)(x)=F,,(L~I(x),g(L~I(x)) when x~/~.
We have
[Th - Tgl~o = Max{lFn(L~l(x), h ( L f ( x ) ) - Fn(L-~l(x), g(L~l(x)))l: x E In,
n=l,2,...,N}
---Max{q. I h ( L ~ l ( x ) ) - g ( L ~ l ( x ) ) l : x ~ 1,, n = 1, 2 , . . . , N }
<-q'lh-gl~.

Hence T has a unique fixed point /~c Co(I). The graph of/~ is an attractor of
the i.f.s., whence h =f, whence f is continuous. []

The above theorem should be compared to a theorem of Hutchinson [H, p.

731] concerning i.f.s. (in our language) whose attractors are continuous para-
metrized curves passing through a specified set of points. In the present case, by
restricting the mappings to be generalized shear transformations, the parametrized
curve is discovered to be the graph of a function. Furthermore, the underlying
compact metric space is different: here the i.f.s, has G as its attractor in a bigger
compact metric space (all closed nonempty subsets of I x [a, b] versus all graphs
of continuous functions g : I--, [a, b] such that g(xo) = Yo and g(XN) = YN). In
particular, stochastic iteration can be used to calculate G.

Example 1. We define linearfractal interpolation functions to be those obtained

using affine maps
Ln(x) = xn-1 + (x, - x , - O " (x - XO)/ (XN -- XO) = an" X + hn,
(3.4)
Fn(x,y)= bnx+a," y+kn,
where the real constants bn and kn, depending on the adjustable real parameter
an, are chosen to ensure
Fn(xo, Yo)=y._l and F.(XN, y N ) = y . .
Fraetal Functions and Interpolation 309

' (x.,y.)

(xN, yN )

(Xo, Yo)

lk,
X

Fig. 3

That is, a . ~ ( - 1 , 1) is chosen and then

b. = (y. - y._l - t~.. (YN - Y0))/(xN - Xo),

k,,=y,_l-a,.yo-b,xo for n ~ { 1 , 2 , . . . , N}.

Then w. maps the parallelogram P onto the parallelogram P. in Fig. 3. We

have written simple computer programs in compiled Basic on an IBM PC, for
plotting the graphs of such linear fractal interpolation functions. One inputs the
interpolation points (xi, yi) for i = 0, 1. . . . , N and vertical scaling parameters a .
for n = 1, 2 . . . . , N ; whereupon points on the graph o f f ( x ) are computed using
the stochastic iteration method described in Section 2. The results are displayed
on a video graphics monitor and may be printed with a graphics printer: typically
computation is very fast and a good view of the whole function is quickly obtained.
Also, it is straightforward to magnify up or look closer at a given piece of the curve.
In Fig. 4, we use x~=(i/lO) for i = 0 , 1 , . . . , 1 0 , with a l = a 2 . . . . . aN=or
taking various values which effectively adjust the dimension of f(x).
In Fig. 5, we show a selection of linear fractal interpolations, this time with
different vertical scalings (the a,'s).

Example 2. We draw attention to the special class where

(3.5,) L,(x) = a, . x + h,,

F . ( x , y ) = a . . y+q.(x),
with F.(xo, Yo)=Y.-1 and F.(xN, yN)=y.. By choosing the functions O,,(x)
appropriately one can fix a priori the overall approximate shape of f(x). V,~en
a . = 0 we call q.(x) a condensation function because then the role it plays in the
present set-up is the same as that of condensation sets in the general theory of
i.f.s.: in particular, in this case, we have f ( x ) = q.(x) for x e I.; if we start with
a . ~ 0 and then let a . ~ 0 then the graph o f f ( x ) over I. "condenses" on that of
q,(x).
310 M.F. Bamsley

a = -0.1 a = -0.1

a=0 a =0.1

Fig. 4(a)

Let us w r i t e
q,(x) = h(L,~(x))-a,,, b(x),
where h(x) is a c o n t i n u o u s real f u n c t i o n such t h a t
h(xi)=y, foraU i~{0, 1,,..., N}
a n d b(x) is c o n t i n u o u s , real a n d obeys
b(xo)=Yo a n d b(xN)=yN.
Fractal Functions and Interpolation 311

a = 0.2 a = 0.3

a = 0.4 a = 0.5

Fig. 4(b)

We call b(x) the base function and h(x) the height function because f(x) is
invariant under the following transformation: subtract the base function from
f(x), rescale the result vertically by the factor an, shrink linearly horizontally so
that [Xo, xN] -> [X~-l, x~], then add the result to the height function over I~, for
each n.
In Example 1, h(x) is the piecewise linear interpolation through the data and
b(x) is the linear function through (Xo, yo) and (xN, yN). More generally one can
312 M, F. Barnsley

Fig. 5(a)

take b(x) and h(x) to be splines or polynomials. One might choose b(x) to fix
the broad outlines of a distant mountain range in profile; or, with a trigonometrical
polynomial, one might match the basic rolling shape o f a wave train, upon which
ripples on all scales are to be eventually superimposed. As the Io~,l's are increased
from zero the dimension of the fractal interpolation function increases, see
Section 5.
The reason we draw attention to the special class of fractal interpolation
functions f(x) which can be generated in this example, is that such functions
are explicitly integrable. In Section 4 we show that integrals of the form

f,.m = II xn(f(x))m dx for all m, n ~ {0, 1, 2 . . . . }

can be evaluated explicitly in terms of integrals of the qk(X)'S.

ThUS, we should be able to construct interpolation functions whose overall
shapes, dimensions, and various moments can be prescribed.
Fractal Functions and Interpolation 313

J
J
I
i
==,'
I
!
/
!
/

f
r~
/

J
!

Fig. $(b)

Example 3. Let f ~ Co(I) be given. Let the mappings {w,: n = I, 2 , . . . , N} be

chosen so that

I.f- r~?l~< ~,
where T: Co(I)4 Co(1) is the mapping introduced in the proof of Theorem 1.
Let f be the associated fractal interpolation function,

Tf = f, f ~ Co(I).
Then

I . f - ~ = Lim Ii- Tmil~ <- Lim ~ IT"-lf - T"jI=

n l --~ o o m " ~ ~ 1 7 6n ~ l

Lim q"-~ 9 - ~< .

314 M.F. Barnsley

In other words, if we can choose the w / s so that T f is close to ], then the

corresponding fractal interpolation function will also be close to ]. This observa-
tion is just a slight variant of the Collage theorem [BEHL] which has been
successfully applied in computer graphics [DHN].

4. Codes, Formulas, and Measures

We consider an i.f.s. {K, w,: n = 1, 2 , . . . , N} which gives rise to a fractal interpola-

tion function f : I ~ R, with graph G, as in Theorem 1.
Let 12 denote the set of all half-infinite sequences of N symbols, {1, 2 , . . . , N},
so that ~o ~ 12 if and only if
,o -- ( ol, o 2, ,03,...),
where each oJ~e {1, 2 , . . . , N}. We reserve the notation oJi for the ith component
of co. We give 12 the product topology: with the appropriate metric it becomes a
compact metric space homeomorphic to the classical Cantor set. Just as in the
case of hyperbolic i.f.s., see [H, BD 2], we can construct a continuous onto
mapping ~ : 12 ~ G by defining
(4.1) qb(eo) = Lim w,ol ~ w,~ ow,~3 o . . . . w,om(w, y ) .
m--*co

To see that this limit exists and is independent of (x, y ) e K, we consider the
hyperbolic i.f.s. {/, L,: n = 1, 2 , . . . , N}, for which we can define a continuous
onto mapping 9 : 12-->/, the attractor of the i.f.s., by
~ ( ~ o ) = Lm i~mo o I~ ~ oL
~e~
o Lr . . . . . L,om(x),

where the limit is independent of x ~/. For the moment let (x, y) e K be fixed
and let dp(o~) denote the set of all accumulation points of the sequence on the
right-hand side of (4.1). Then dp(~o)c G, and takes the form
r = s),
where S c [a, b] is the set of all limit points of the sequence
{Fo,,(~(To~)3 F , ~ ( ~ ( T2~o),. . . , F ~ , , ( ~ ( T"~o), y ) . . .) ) } ~ = , ,
where T: 12 -->12 is defined by
T(oJ1, ~02, 0~3,...) = (w2, ~03, oJ4,...).
But by Theorem 1, since ~(eo) is a single point in I, we must have
r = (~(eo),f(~(oJ))),
whence S consists of a single point, as desired. Furthermore, since ~(0~) is
independent of x ~ / , we must have ~(~0) is independent of (x, y ) e K. Finally,
the continuity of ~ : 1 - / ~ G follows from the continuity of @:12~ I and the
continuity o f the map P : I --> G defined by
P(x)=(x,f(x)) for xeL
Fractal Functions and Interpolation 315

In particular, we have established the result

f(',It(to)) = Lina Fo,,(~(Tto), F,o:(~( T2to),.. 9 Fo,~(~(T"to), y ) . . . ) ) ,

for any y e [a, b]. Observe that ~ ( t o ) = ~ ( 0 3 ) whenever to and 03 are equivalent
representations of the same number in base N, but that this leads to no ambiguity
in using the latter formula to calculate f ( x ) .
We consider invariant measures next.

Theorem 2. Let p > 0 be a probability vector, and let { K, wn: n = 1, 2 . . . . . N} be

the i.f.s, of Theorem 1. Then {K, wn: n = 1, 2 , . . . , N} admits a unique p-balanced
measure ~. Let ~ be the p-balanced measure for the i.f.s. {I, Ln:n = 1, 2 . . . . , N},
and let P: I ~ G be the homeomorphism P( x ) = ( x, f ( x ) ), for x ~ L Then
/.~(B) = / z ( P ( B ) )
for all Borel subsets B of I.

Proof. We show that the p-balanced measure on the graph is simply the lift of
the unique invariant measure for the corresponding process on the x-axis. Let
g~(G) and ~ ( I ) denote the spaces of probability measures supported on G and
I, respectively, with the appropriate weak *-topologies. We define a homeomor-
phism M : ~ ( G ) ~ ~ ( I ) by
(Mp,) (/~)= p,(P/~)
for all tz ~ ~ ( G ) and all Borel subsets /~ of I. The inverse is
(M-'tz)(B)=I~(P-1B)
for all/~ e ~ ( I ) and all Borel subsets B of G.
Now let/x be any p-balanced measure on G, so that
N
~(B)= E p.~(w:'(B))
n=l

for all Borel subsets B of G. Then it is readily verified that

N N
(M/z)(/~)= E p,tz(w~'(PB))= ~ pntz(PL~'(B))
n=l n~l

N
= ~ p,(Mix)(L~I(B)),
n=l

which shows Mlz = 12, the unique p-balanced measure for the i.f.s. {I, L.: n =
1, 2 , . . . , N}. Hence ~ = M - ' f i is also unique. 9

Remark. An i.f.s. {K. w,: n = 1, 2 . . . . . , N} complete with a p-balanced measure

/.~ supplies a dynamical system in the following way, see also [P]. Consider the
space M = [0, 1] x K and define a mapping
~:M~M
316 M . F . Barnsley

by

I~(x,y)=(l(x-p~-p2 . . . . . p,-O,w,(y)) if i ~ { 2 , 3 , . . . , N}

for
X ~ ( P l -t-" 9 9 - t - p / _ l , P l +P2+" 99+ P i ) ,
and

(;(x,y)=(lx, w~(y)) if x~[O, pl].

Then the measure to = .~ x/~ on M, where ~ denotes uniform Lebesgue measure

on [0, 1], is invariant under ~. The dynamical system consists of (M, ~, v), and
we refer to it as the one associated with the i.f.s. We say that two i.f.s, are
isomorphic if their associated dynamical systems are isomorphic. Theorem 2 tells
us that the two i.f.s, under discussion there are isomorphic.
The utility of Theorem 2 is that it allows one to change variables in integrals
involving p-balanced measures on the graphs of fractal interpolation functions.
Let H ~ LI(K,/z). Then

I H(x,y) d/z(x,Y)=I H(x,f(x))d/~(x).

In particular, if we choose the L,'s to be affine, as in Examples 1 and 2, and if
we Choose

(4.2) p, = a, - - - ,
x, - xn_1
xN -Xo
then dl~(x)= dx/(xN- Xo) and
(4.3) f,r, H(x'y) dp'(x'y)= - Xo fl H(x,f(x))dx.

5. Integration of Fractal Interpolation Functions

We continue to discuss an i.f.s. {K, wn: n = 1, 2 . . . . , N} which gives rise to a

fractal interpolation function f : I ~ R, as in Theorems 1 and 2.
The stationarity property (2.1) of the associated p-balanced measure implies
that for H ~ LI(K,/x),

H(x, y) dp.(x, y) = Y, p,,H(w,,(x,y)) dlx(x, y)

G (3 n=l

= 2 p~ N(x, y)) d~(x, y).

O n=l

If we choose the L,(x)'s to be affine and such that (4.2) holds, then (4.3) implies
the remarkable identity

(5.1) Ii H(x,f(x)) dx= n~l

~ a, I, H(L,(x),F,(x,f(x))) dx.
Fractal Functions and Interpolation 317

Theorem 3. (i) Let f(x) be a fractal interpolation function generated by the special
class of i.fs. introduced in Example 2, For each m e {0, 1, 2 , . . . } one can
evaluate explicitly the moment integral

f " = f , xmf(x) dx

in terms off.,_1, f,.-2, . . . , fo, the interpolation points, the sca~ing parameters
{ a, : n = 1,2,..., N } and the moment

Q., -- f , x"Q(x) dx,

where Q : I -~ R is defined by
Q(x) = q, o L~I(x) for x ~ I,.
(ii) Let f(x) be a linear fractal interpolation function as introduced in Example
1. For each I and m in {0, 1, 2,...} one can evaluate explicitly the moment
integrals

f " = f t xm(f(x))' dx

in terms of {flj:j=O, 1 , . . . , m - 1 } u { f ~ , j : j = O , 1 , . . . , l+m-p; p=

0, 1 , . . . , l - 1}, the interpolation points and the scaling parameters {a,: n =
1,2 . . . . , N } .

Proof. (i) Choose H(x, y) = x"y in (5.1), with L, and F, given by (3.5), to obtain

fro=,=1
~ a, f1(a..x+h.)'(a..f(x)+q.(x))dx
k+l rrf--k
= Y. a,.a, 9 h, 9 "fk+Qm,
n=l k=O

where

Qm = n~= l an fl (L~(x))"" q.(x)" dx

=" ~ f ~mq~~ L~'(~) d~l = f x ' O ( x ) dx,

n~l In I
where in the nth integral we have made the change of variable ~ = L.(x). It
follows that

fro= fk" " a, .a,. h,"-k+Q" 1- a, .a, .

= n=l n=l

N
The denominator is positive because ~.=1 a. = 1 with each a . > O ; and each
a,, e ( - 1 , 1).
318 M.F. Barnsley

(ii) Choose H(x, y) = x'ny t in (5.1), with F, and L, given by (3.4), to obtain

Stm= n=l
~ a,, I I (a.. x+hn)'n(an,f(x)+bnx+k.)'dx

where we define the polynomial Pt,,,~p(x) and its coefficients, by

l~l'm'p(X)= n = l ~ ( ; ) a n ( b n " x + k n ) l - p o t P ( a n " x + h n ) m

l+m--p
= ~ K ( l , m , p , j ) x j.
j=0
N an,,+1 " a~ e (-1, 1), we obtain
Noting that K(I, m, 1, m) =Y.n=l

f,.~--
(-~, AJ" (7,) E~ an~, "~o"
~_~
\j=O n=l

+ E 2 K(l,m,p,j)fp, j 1- an "a , 9
p=O J=O 1

Example 4. Consider the linear fractal interpolation function for the three
interpolation points (0, 0), (1, 1), and (2, 0), with vertical scalings a~ = a2 = 89 we
find

~(;)=(~ ~)(,) ~(;)=(i ~

corresponding to Ll(x)= 89 Lz(x)= 89 Fl(x,y) =~x+~y,
1 1 and F2(x,y)=
-89 + 89 + 1; see Fig. 6.
We readily calculate

f.--
;o xnf(x) dx= 89
Io~(~x) (~x + 89
~ n ~ dx

Io
+89 ( 89189189 dx

from which
n(:) 2- 2n~2
f " = m=0
~ 2%'~--
f r2o - t - (n + 1)(n +2)'

whence we calculate successively fo = 2, fl = 2, f 2 _- ~52

, etc.
Equation (5.1) can provide functional equations satisfied by integral transforms
of fractal interpolation functions. These are of interest for a number of reasons.
First, the existence of fractal interpolation functions furnishes existence theorems
for solutions of the functional equations. Second, these equations will involve
renormalization, and thus they supply generalizations and sources of examples
Fractal Functions and Interpolation 319

1
Fig. 6

in statistical physics. We note the successful exploitation by Bellissard [Be] and

others of such equations in the construction of limit periodic operators associated
with interesting quantum mechanical models, and also by Derrida and others,
in connection with partition functions for various Ising models [Der]. Third,
such equations may allow one to investigate smoothness properties of fractal
interpolations by studying the asymptotics of the Fourier transform, for example,
somewhat along the lines in [G].
For the general integral transform

:(,) ~"fI K(X, Of(x) dx

we have from (5.1)

:(,) = ~ an ~ K(Ln(x), t)Fn(x,f(x)) dx.

n=| 3t
Hence, if Fn(x,f(x))= an" f(x)+ qn(x) as in Example 2, we obtain

(5.2) ~ an" a." ~ K(Ln(x), t)f(x) dx+Q(t),

f(t)=
.tl
where Q ( x ) = qn(L~(x)) for x ~ / , . In particular, when K(x, t) depends either
on the product or the difference of x and t, the sum on the right-hand side of
(5.2)' can be re-expressed in terms of f,, as in the following example. Note that
these functional equations are related to, but distinct from, ones obeyed by
transforms of p-balanced measures, eL [BD 2].
320 M.F. Barnsley

Example 5. The Fourier transform is obtained with K (x, t) = e ix, in (5.2). We find
N
f(t)=
t1=1a,:a,,
~ e~h,'.f(allt)+t)(t).
For the numbers in Example 4, we find Q(x)=x for O~_x___ 1 and Q(x)=2-x
for l - x - 2 , so
](t)_t_~(t~ 9 2ei'-e2"-I
--4Jk2] +le"f(2+ l)'t t2

Example 6. The Stieltjes transform is obtained with K(x, t ) = 1~(t-x). We

discover, from (5.2),

J~(t) ~n=l~ +O(t)"

For the numbers in Example 4 we find
](t) ~_1 A
~f(2t)+89^ t ln(t/(t- 1)+ ( t - 2 ) l n ( ( t - 2 ) / ( t - 1)).

6. Dimension of Fractal Interpolation Functions

Given a set of points A c R n, 0 < e < oo, and 0 - p < oo let

M~(A)=inf ~ [8(Ai)]p,
i=l

where A = U~ Ai is a countable decomposition of A in subsets of diameter 8(A~)

less than e. We set [8(A~)] ~ = 0 when A~ is empty and = 1 otherwise. As e increases
we have more sets of decompositions to choose from so M;(A) is a monotone
decreasing function of e. We define
Mp(A) = Lira M;(A) = Sup M~(A).
~-~0 e>0

The p-dimensional measure of A, Mp(A), is a Borel regular measure, but is not

usually bounded on finite sets. For each A there is a unique real number D, the
Hausdorti-Besicovitch dimension of A, such that

Mp(A)={O ifp<D,
if p>D.
MD(A) may take any value in [0, c~]. A reference is IF].
The following theorem develops from [BD 2, Theorem 8], which applies to
the dimension of the attractor G of a hyperbolic i.f.s, such that W~(G) n Wj(G) =
O for i#j. A related result [H, Theorem 5.3(1)] permits the latter condition to
be "slightly" violated in that a certain open set condition must be obeyed (this
condition is satisfied in the case of fractal interpolation functions), but it applies
only when the w.'s are similitudes, which is rarely the case here. A mapping
W, : R 2 ~ R 2 is a similitude when there is s e R so that w , ( x ) - w , ( y ) = s. ( x - y )
for all x, y ~ R 2.
Fractal Functions and Interpolation 321

Theorem 4. Let {K, w.: n = 1, 2 . . . . , N } be a hyper6olic i.f.s., which generates a

fractal interpolation function f with graph G as in Section 3 and such that the
mappings L. : I -->In are affine with L. ( x ) = a. . x + h.. Let bounds on the contrac-
tivities be expressed by numbers 0 < t. < s. < 1 such that for each n
t.d(x,y)<-d(w'.(x),w.(y))<s.d(x,y) forall x, y e K.
where d ( . , .) is a metric on K equivalent to the Euclidean metric. Then the
Hausdorff-Besicovitch dimension D of G is bounded by
Min{2, l} < D -< u,
where I and u are the positive solutions of
N N
y t .I = l and ~ s .w= l;
n=l ~,1~1

and where for the lower bound to hold it must also be true that

t 1 " t N ~-- ( M i n { a l , aN}) tin

n 1

Proof. Without loss of generality we take d ( . , 9) to be the Euclidean metric. In

practice it is convenient to take the metric to be the one in (3.1).
To obtain the upper bound let G = [,.It At be a finite decomposition of G into
subsets of diameter < e, so that ~t [ 8 (At)]P < oo for all 0 - p - oo. A new decompo-
sition is provided by
N
G=U U A o,
i j=l

where A o = wj(Ai). Since

N N
Y ~ [8(ao)]~-<~, ~ Isjl~[8(a,)] ~
i j=l i j=l
N
= ~ [sjl p ~ [8(A,)] p,
j=l i

it follows that whenever F.jN11sjlp < 1 we must have M ~ ( G ) = 0, which implies

Mp(G) = 0 and hence p -> D.
For the lower bound we use, as in [BD 2, Theorem 8], a theorem of H. P.
McKean, see [BIG]. In the present setting it states that i f / t is any probability
measure on G, and if fl E R is such that

fofod~(x)d~(y)< ix_y]~
then/3 -- D.
Let ~'o be the uniform probability measure on K, namely

~o(B) = j'~ d~
(XN -- Xo)(b - a)
322 M.F. Barnsley

where Le denotes uniform Lebesgue measure in the plane. Then it is readily

verified that

Jo*f f dv~ dv~176176 forall 0<-fl < 2 .

r K [x-Yl ~
Now define

J=+l = f f dv,.+,(x)dv,.+,(Y) for m = 0, 1, 2,

K ,,: Ix-yl ~ " " '

where, as in Section 2,
N
Pm+I(B) = E Pn~'m(w'~l(B))
rim|
for each Borel subset B of K. Here p = (Pl,P2, 999 PN) is a prescribed probability
vector; and we know (recall Section 2)that vm converges to the p-balanced
measure ~, supported on G, which is associated with the i.f.s.In particular,the
convergence of {Jm} to a finitelimiting value J as m -->m would imply

so our next aim is to establish values of/3 for which {Jm} converges.
Now

(6.1) Jm+l=fr t lx-yl-~ dVm+l(X)dv,~+|(y)

= ~ ~ PiPj
i=1 j=l
LLIw,(x)-wj(y)l -/3 dvm(X)dvm(y)

= ,=1
~ p~ fr fK ,w,(x)-wj(y),-' dv=(x)dvm(y)

"Of-(
i~'~=l
-~i,j~1)(
pipjIKfK'Wi(X)--wj(Y)l-•dl:m(x)dl:m(Y))
'
not n.n. n.n.

where "n.n." means "nearest neighbors," namely i=j + 1. The first term here is
dominated by C ( / / ) 9 Jm,
where
N
C(/3)= Z
~=l
p~t7~.
The second term is dominated by

d(jS) =(,j~| p,pj) p-/3=p-S,

where
p = Min{lx-y[: x s w~(K),y~ wj(K), i #j, not n.n.}
> M i n { I x i - xjl: i # j, not n.n.} > 0.
Fractal Functions and Interpolation 323

TO dominate the third term we consider for example

(6.2) J~:= f r Ir 'W~(X)- w2(y)l-~ dvm(x) dvm(y)

~- f, I, IL,(x)- L2(y)I-a d~r~(x) d~m(y)

= It, f,2 Ix-yl-~ d~m(L-f'(x))d~m(L-~(y)),

where ~r~ is the measure obtained by projecting Vm onto I according to

~m(/~)= Vm(/tX[a, b]) for Borelsubsets/~ of I.

It is readily verified, much as in Theorem 2, that

N
~m+~(/~) = • p,~;m(L~I(/~)), m = 0, 1, 2 , . . .
n=l

from which it follows ~,~-~/~, the p-balanced measure for the i.f.s. {I, L~: n =
1 , 2 , . . . , N}.
We will obtain an upper bound for (6.2) by replacing the piecewise constant
measures z;m(/~.~-1) by point masses, located at the ends of intervals of constancy.
To do this we will use the notation

x(i-~)j= Li(x~) for i = 1 , 2 , . . . , N; j = 0 , 1 , . . . , N.

Then, for example,

9~i--I = X(i--1),O ~s "~(i--1),1 ~ X(i--1).2 ~ " " " ~ X(i--1).(N--1) ~ X ( i - - 1 ) , N ~ Xi"

Similarly, we define

X(~-~).(l-~),...,(m-~)j = Lk(Xu-~),...,(,~-~),j)

for k, l , . . . , m = 1, 2 , . . . , N and j = 0, 1 , . . . , N. For example,

X0,(N-1) = X0,(N-1),0 = LI(LN(xo)) < X - 0 , ( N - 1 ) , I

X0,(N-1),2 ~ 9 9 . ~ X0,(N-1),(N--I) ( X0.(N-1),N

= LI(LN(xN)) = LI(xN) = Xo.N= xl.

For m = 0 in (6.2) we have for/3 < 2

yl.2= plp2S(xo, xl, x2, fl),

where for a < b < c we define

S(a,b,c, fl)= (y_x)ody dx.

 324 M.F. Barnsley

One readily calculates

S(a, b, c,/3)

l ( c - b) 2-~ + (b - a) 2-~ - ( c - a) ~-~

when /3 < 2 w i t h / 3 # 1

[(c-a) Ln(c-a)-(c-b) Ln(c--b)-(b-a) l n ( b - a)

when /3 = 1.
I, (c-b)(b-a)

For m = 1 in (6.2) we have

N--1 N--I
L2 PkPI+I
J1 <--PIP2 v~ ~ ~-(plp2)(plvpl)S(Xo,N-l,X. Xl.l, fl).
k=~ 1=~ (Xul--Xo,~,) ~

TO see h o w this u p p e r b o u n d is obtained consider Figure 7(a) and (b). In Figure

7(a) we show the actual measures ;~(L~ "~) and ~(L2~), and in Figure 7(b) we
show the modified m e a s u r e s which majorize the integral.

~l(Li -l)
F ~I(L~-1)
r

M~ass~ Mass
\P2 \
I /1 "/ \ ~pN,.~kV//Mass/ /

Xo X0,1 -~0,2 XO,(N--I) XI Xl,I Xl,2 XI,(N--I) X2

y >
(a)
Mass Mass
P2 Mass

'[
PN--1
Mass P2 Mass
Pl Mass PN
P~

Xo,l Xo,2
d'-
X
X0,(N-1) X1
v
XI,1
T
XI,2
y ---
0
XI,(N--1) X2

(b}
Fig. 7 (a) The actual measures ~2(L~-~) and ~z(L21). (b). The measures in (a) modified so that the
integral (6.2),' with m -- 1, increases.
Fractal Functions and Interpolation 325

For m = 2 in (6.2) we have, see Figures 8(a) and (b),

N-, N-, PkP,+,

j~,2 < p,p2 E E
k=l 1=1 (x~.l- Xo,) ~
N-1 N-I PkPl+l
+ PlP2(PNPl) k~= l l=lZ ( X l , o , / -- XO,N_I.k)~
2
+PlP2(PNPl) S(Xo.~N-1),fN-1), xl, x,.o:,/3).
Following the same line in the general case we must obtain
N-I N - ,
11,2 < PkPl+1
-,,, - PiP2 k=lZ t=lZ ( x l , t - Xo,k)~

N-1 N-~ PkPI+I

+p,e~(eNpl) k=l
E l=~
Y. (X1,0,l--Xo,N-l,k)/~
N-1 N-1
PkPI+I ~-...
+P1P2(PNPl)2 ~ E
k=l / = 1 (Xl.o,o, 1 - - XO,N_I,N_I,k) fl

N-1 N-1 PkPl

-t-PlP2(PNPl)m-' Z E :
k=l a=l (Xl.o.....O,I--Xo,(N-l),,...tN-n,k)
m - 1 times ra - 1 t i m e s

+ p,p2(pNp,)mS(xo, N_l...,N_l, X~, Xl.o.....O.,,/3)-

Hence

j~2 PiP2 N-I N-1 PkPt+t . [PNPa~"

E l - p l P 2 ~ - - ~ : S ( x o l , x,, x2, ~),
< ( 1 - - p N p l / a ~) k:l I=1 (Xw--Xo,k) ~
where

a = M i n { a . aN}.

We assume P~PN < a a so the last term goes to zero as m--> o0. For example, if
P , = P N = 1 / N and a = 1 / N then this requirement is equivalent to N a < N 2,
n a m e l y / 3 < 2.
Putting results together, we have that the last term in (6.1) is dominated by

N-2 N-I N-I PkPt+l

2 E P,+1P,+2 E E
Y J~-< ,-o
i,j
#j ( 1 - P~Pl
n.n a ~ /

( P N P l ~ m N--2
+\--aft"/ i=oEp,+,p,+2S(x,x,+,,x,+2,/3)

--- e(/3) <oo for all /3>2,

 326 M.F. Bamsley

ss Mass~

I i
~0 XO,I X0,2 XO,2,1 X0,2,2 XI Xl,0,1 X1,0,2 XI,I X1,2 X2 X2,1 X2,2
X ~-g y
(a)

Mass Mass Mass Mass Mass

Mass P2 P3
P3 P2 Pt P2
Pl P2 Mass
P1173
Mass) ~ _

I I i

CO Xo. ! X0,2 XO,2,1 X0,2.,2J X! XI,O,1 Xl,O,2 XI,1 X1,2 X2 X2,1 X2,2
~ y ~
(b)

Fig. 8. (a) The actual measures ~2(L~-l) and ~:(L~-I), assuming N = 3. (b) Modified of measures in
(a) so that (65.2), with m = 2, increases its value.

where e ( f l ) is independent o f m. It follows that

(6.3) Jm§ for m=0, 1,2,...,

where f ( / 3 ) = d (/3) + e (/3) < oo is a positive constant independent o f m, for 3 < 2,
and
N
c(/3) = Z p n~t , - ~ ,
n=l

together with the condition pNP~ < a ~. We anticipate that we will finally choose
t~
(6.4) P-- N for n = 1,2,..., N;
E t~
n=l

and thus we impose the requirement

(6.5) tltN < Min{ab aN}) i t.) .

Fractal Functions and Interpolation 327

From (6.3) we find inductively

Jra+l~-(C(~))m+lJo+ ~ (C(fll)Jf(fl) for re=O, 1 , 2 , . . .

j=O

It follows that if O< C(fl)< 1 then {Jm} is uniformly bounded, so

f~fclx-yl-~dlz(x) dtz(y)<m.
Hence
N
D - > S u p { f l < 2 : ~ p~t:tJ<l,p~pN<a~},
n=l

from which we get

A
D-> Min{2, fl},
where/~ is the positive solution of
N
pnt, = 1,
n=l

where we assume PlPN <--a~. Hence choosing the p, as in (6.4), and assuming
(6.5) but with possible equality, we find
D -> Min{2, l},
where ! is the positive solution of
N
E t~=l. 9
n=l

In practice this theorem may not give good numbers in general, especially
when the s,'s and t,'s are evaluated with respect to the Euclidean metric. How-
ever, the bounds can often be tightened by iterating the i.f.s.: i f f corresponds
to {K, w, : n = 1, 2 , . . . , N} then it also corresponds to {K, w, o w,, : n, m =
1, 2 , . . . , N}, and in the obvious notation, for affine maps, we have
tn " t m <-- tn, m and Sn, m ~ Sn " Sra.

When these inequalities are strict, as they typically will be, we observe that if
N
X t~=l,
n=l

then
N N
X X t~tm
' ' = 1,
n=l ra=l

which implies
N N
E E t.,,.
' > 1;
n=l m=l
328 M.F. Barnsley

so the solution [ of
N N
E E t.,,~
r =1
n=l m=l

must obey
I<L
Similarly, the upper bound may be improved.

7. H61der Continuity and Some Functions of Besicovitch

Besicovitch and Ursell [BU] prove the following. Let y = f ( x ) denote a curve in
R 2, with f : I ~ R 2 belonging to the Lipschitz 8-class (LipS), with D the Hausdorff-
Besicovitch dimension of its graph G. Then
1<_D<_2-8.
The bounds are sharp: a curve of class Lip s may have D anywhere in the above
range.
f ( x ) with domain I belongs to Lip s if there is a constant C so that for any
x ~ I ~ there corresponds an interval ( x - h , x + h) such that, for any y in this
interval
If(x) -f(Y)I < C. Ix - y l ~.
In proving sharpness Besicovitch considers special functions f ( x ) of the follow-
ing type. Write ~(x) for the function equal to 2x in 0 -< x -<89and defined elsewhere
by the relations
+ l).
Then define

f ( x ) = ~ a,,O(b,,x), x ~ [0, 1],

n~O

where
a, = b~ a, 0 < 8 < 1.

If b,+l> B. b, where B > 1 then f ( x ) is of class Lip s and not of any higher
Lipschitz class.
An example of such a function is
1
f(x) = a,(2"x).

On the interval [0, 1] this is exactly the linear fractal interpolation function
(Example 1) which passes through (0, 0), (~,
1 ~) and (1, 0), with vertical scalings
al = ce2= 1/28.
Acknowledgments. Helpful discussions related to this work were had with
Stephen Demko and Douglas Hardin. This research was supported in part by
National Science Foundation Grant DMS-8401609
Fractal Functions and Interpolation 329

References
[BD 1] M. Barnsley, S. Demko (1985): Iterated function systems and the global construction of
fractals. Proc. Roy. Soc. London, Ser. A, 399:243-275.
[BD2] M. Barnsley, S. Demko (1984): Rational approximation offractals. In: Rational
Approximation and Interpolation (P. R. Graves-Morris, E. B. Saff, R. S. Varga, eds.).
New York: Springer-Verlag.
[BEHL] M. Barnsley, V. Ervin, D. Hardin, J. Lancaster (1986): Solution of an inverse problem
forfractals and other sets. Proc. Nat. Acad. Sci. U.S.A., 83:1975-1977.
[Be] J. Bellissard (1984): Stability and instability in quantum mechanics. C.N.R.S. (France).
Preprint.
[BIG] R.M. Blumenthal, R. G. Getoor (1960): Some theorems on stable processes. Trans. Amer.
Math. Soc., 95:263-273.
[BU] A.S. Besicovitch, H. D. Ursell (1937): Sets of fractional dimension. J. London Math.
Soc., 12:18-25.
[DHN] S. Demko, L. Hodges, B. Naylor (1985): Construction offractal objects with iterated
function systems. Computer Graphics, 19:271-278.
[Der] B. Derrida (1986): Real space renormalization and Julia sets in statistical physics. In:
Chaotic Dynamics and Fractals (M. F. Bamsley, S. G. Demko, eds.). New York:
Academic Press.
[DS] P. Diaconis, M. Shashahani (1986): Products of random matrices and computer image
generation. Contemp. Math., 50:173-182.
IF] K.J. Falconer (1985): The Geometry of Fractal Sets. London: Cambridge University
Press.
[FFC] A. Fournier, D. Fussell, L. Carpenter (1982): Computer rendering of stochastic models.
Comm. ACM, 25.
[G] A.M. Garcia (1962): Arithmetic properties of Bernoulli convolutions. Trans. Amer. Math.
Soc., 102:409-432.
[H] J. Hutchinson (1981): Fractals and self-similarity. Indiana Univ. Math. J., 30:713-747.
[M] B. Mandelhrot (1982): The Fractal Geometry of Nature. San Francisco: W. H. Freeman.
[P] S. Pelikan (1984): Invariant densities for random maps of the interval. Trans. Amer.
Math. Soc., 281:813-825.
M. F. Barnsley
School of Mathematics
Georgia Institute of Technology
Atlanta
Georgia 30332
U.S.A.