Nonlinear Analysis 36 (1999) 721 – 745

The metallic means family and multifractal

spectra
Vera W. de Spinadel
Centro de MatemÂatica y Diseño MAyDI, Facultad de Arquitectura, Diseño y Urbanismo,
Universidad de Buenos Aires, JosÂe M. Paz 1131-1602 Florida, Buenos Aires, Argentina

Received 2 February 1998

Keywords: Fibonacci sequences; Lucas numbers; Hyperbolic geometry; Farey trees; Multi-
fractal spectra

Fractals are geometric or physical congurations with self-similarity, that is, con-
gurations that remain invariant in the presence of “scale changes”. Multifractals, are
nontrivial structures that possess a spectrum of scaling indices, instead of the simple
scaling structure shown in fractals. Usually, these non-regular fractal sets appear in
physics, biology, chemistry, etc. Its characterization has been based on a multifractal
spectral decomposition technique introduced by Procaccia [8].
The purpose of this paper is to show how, bridging between continued fractions
expansions, generalized secondary Fibonacci sequences, hyperbolic geometry, Fuchsian
groups and some members of the family of metallic means family, recently introduced
by the author (see [4]), it is possible to state a mathematical model for the analysis
of fractal and multifractal spectra.

1. Fibonacci sequences

The Fibonacci sequence is a sequence of natural numbers formed by taking each

number equal to the sum of the two precedent terms. For this reason, this type of
sequences is called a “secondary Fibonacci sequence”, to distinguish them from the
ternary Fibonacci sequences, in which each term is a linear combination of the three
precedent terms.

∗ Tel.=fax: +541-795-3246; e-mail: [email protected]; internet: http:==www.fadu.uba.ar=maydi.

0362-546X/99/$19.00 ? 1999 Elsevier Science Ltd. All rights reserved.

PII: S 0 3 6 2 - 5 4 6 X ( 9 8 ) 0 0 1 2 3 - 0
722 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Beginning with F(0) = 1; F(1) = 1, we have the following secondary Fibonacci

sequence:

1; 1; 2; 3; 5; 8; 13; 21; 34; 55; 89; 144; : : : (1.1)

where

F(n + 1) = F(n) + F(n − 1): (1.2)

Secondary Fibonacci sequences can be generalized, originating what is known as

“generalized secondary Fibonacci sequences” GSFS,

a; b; pb + qa; p(pb + qa) + qb; : : :

that satisfy relations of the type

G(n + 1) = pG(n) + qG(n − 1) (1.3)

with p and q natural numbers.

From Eq. (1.3), we get
G(n + 1) G(n − 1) q
=p + q =p + :
G(n) G(n) G(n)
G(n − 1)
Taking limits on both sides of this equation and assuming that limn→∞ G(n + 1)=G(n)
exists and is equal to a real number x – the fact that will be proved in the next theorem
–, we have x = p + q=x or x 2 − px − q = 0, whose positive solution is
p
p + p2 + 4q
x= :
2
This means that
p
G(n + 1) p + p2 + 4q
lim = (1.4)
n→∞ G(n) 2

Now, let us prove the existence of this limit:

Theorem 1. Given a generalized secondary Fibonacci sequence (GSFS)

a; b; pb + qa; p(pb + qa) + qb; : : :

such that

G(n + 1) = pG(n) + qG(n − 1)

with p; q natural numbers; then there exists limn→∞ G(n + 1)=G(n) =  and is a real
positive number.
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 723

Proof. To nd an expression for the nth term of the GSFS, let us rewrite Eq. (1.3)
in the form
G(n + 1) = pG(n) + qH (n); (1.5)
where
H (n + 1) = G(n):
This last equation can be put in matricial form by introducing the matrices
   
G(n) p q
G(n) = ; A= :
H (n) 1 0

Then it is easy to prove that G(n + 1) = A:G(n) or

    
G(n + 1) p q G(n)
= :
G(n) 1 0 G(n − 1)
1


Let us assume that G(0) = G(1) = 1 for simplicity. If G(1) = 1 then

G(n + 1) = An G(1) (1.6)

and the problem is reduced to nding the nth power of the matrix A. The characteristic
equation of the matrix A is
 
p− q 
  = 2 − p − q = 0;
 1 − 
which gives the two eigenvalues
p p
p + p2 + 4q 0 p − p2 + 4q
= ;  = :
2 2
To diagonalize A so as to transform it into
 
 0
Ad = ;
0 0
we shall use the change of base matrix
 
 0
P= :
1 1
The nth power of A is calculated applying the similarity transformation
 
n −1 n n −1  0
A =(PAd P ) = PAd P = P P −1 :
0 0
Finally, we have
 
1 n+1 − 0(n+1) 0 (0n − n )
An = :
 − 0 n − 0n 0 0(n−1)
 ( − (n−1) )
724 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Replacing this value in Eq. (1.6) and performing all the operations, we arrive at the
following expression:
1
G(n + 1) = [n+1 (1 − 0 ) − 0(n+1) (1 − )]:
 − 0
p
Replacing  − 0 = p2 + 4q; 0 = −q= we have

G(n + 1) n+1 + (−q=) n+1

lim = lim =
n→∞ G(n) n→∞ n + (−q=) n
and the proof is completed for the GSFS: 1; 1; p + q; p(p + q) + q; : : :
Finally, if instead of choosing G(0) = G(1) = 1, we begin the GSFS with two arbi-
trary values a and b, it is easy to prove that the result is the same. Indeed, given the
GSFS: a; b; pb + qa; p(pb + qa) + qb; : : : we have to evaluate the quotient
G(n)
pb+ qa
G(n + 1) pnG(n) + qaG(n − 1) G(n − 1)
= = qa −→ ;
G(n) pbG(n − 1) + qaG(n − 2) pb + n→∞
G(n − 1)
G(n − 2
and the proof is complete, independent of the selected value of the rst two terms of
the GSFS.

2. The sequence of metallic means

Let us now introduce a new family of positive irrational numbers, obtained by putting
G(0) = G(1) = 1 and considering dierent possibilities for the coecients of Eq. (1.4).

Deÿnition 2. The set of positive eigenvalues of the matricial equation

    
G(n + 1) p q G(n)
= ; (2.1)
G(n) 1 0 G(n − 1)

for dierent values of the parameters p and q (natural numbers) is the family of
metallic means (MMF).
All the members of this family are positive quadratic irrational numbers that are the
positive solutions of quadratic equations of the type

x 2 − px − q = 0: (2.2)

Let us begin our analysis by considering rst the simpler quadratic equation

x 2 − px − 1 = 0: (2.3)

Then, it is very easy to express the members of the MMF satisfying Eq. (2.3) by
expanding them in a continued fraction. Indeed, if p − q − 1, we get x 2 = x + 1, which
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 725

can be written, dividing both members by x as x = 1 + 1=x. Replacing iteratively the x

of the second member, we get
1
x=1 + ;
1
1+
.
1 + ..

that is x = [1; 1; 1; : : :] = [1], a purely periodic continued fraction, that denes the golden
mean
√
1+ 5
x= =  = [1] : (2.4)
2

If p = 2 and q = 1, the sequence has the form

1; 1; 3; 7; 17; 41; 99; 140; : : : (2.5)

where

G(n + 1) = 2G(n) + G(n − 1); (2.6)

and from Eq. (1.4) we get the silver mean

G(n + 1)
Ag = lim = [2]; (2.7)
n→∞ G(n)
another purely periodic continued fraction.
Analogously, if p = 3 and q = 1, we get the bronze mean
√
G(n + 1) 3 + 13
Br = lim = = [3]; (2.8)
n→∞ G(n) 2
another purely periodic continued fraction.
For p = 4; q = 1, the metallic mean pq is
√
41 = 2 + 5 = [4] = 3 ;

a striking result that is related to the continued fraction expansion of odd powers of
the golden mean. As is easily proved, the following metallic means are
√ √
1 5 + 29 1
√ 1 7 + 53
5 = = [5]; 6 = 3 + 10 = [6]; 7 = = [7];
2 2
√
√ 9 + 85 √
81 = 4 + 17 = [8]; 91 = = [9]; 101
= 5 + 26 = [10];
2
and so on. All the members of the MMF that satisfy Eq. (2.3) are of the form [ n ], a
purely periodic continued fraction expansion.
726 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Obviously, among all these metallic means, the slowest converging one is the golden
mean, since its denominators are the smallest possible, e.g. “ones”. An elegant way of
expressing this result is to state:

The golden mean  is the most irrational of all irrational numbers.

If we consider, instead, the equation

x 2 − x − q = 0; (2.9)

we have, for q = 1, again the golden mean.

If p = 1 and q = 2, the sequence is
1; 1; 3; 5; 11; 21; 43; 85; : : : (2.10)

where
G(n + 1) = G(n) + 2G(n − 1)

and we get the copper mean

Cu = 2 = [2; 0]; (2.11)

a periodic continued fraction. If p = 1 and q = 3, we get the nickel mean

√
G(n + 1) 1 + 13
Ni = lim = = [2; 3]; (2.12)
n→∞ G(n) 2
another periodic continued expansion. In a similar way, we get

14 = [2; 1; 1; 3]; 15 = [2; 1; 3]; 16 = 3 = [3; 0]; 17 = [3; 5]; 18 = [3; 2; 1; 2; 5];

19 = [3; 1; 1; 5]; 110 = [3; 1; 2; 2; 1; 5]; 111 = [3; 1; 5]; 12 = 4 = [4; 0]

and so on. Evidently, all the members of the MMF that satisfy Eq. (2.9) are of the
form [m; n1 ; n2 ; : : :], that is, a periodic continued fraction. Besides, it is easy to ver-
ify that in this set, integer metallic means appear in quite a regular way. The rst
one for q = 2 (1 + h1i = 2), the second one for q = 6 (3 + h3i = 6), the third for
q = 12 (7 + h5i = 12), the fourth for q = 20 (13 + h7i = 20), the fth for q = 30 (21 +
h9i = 30), and so on. The rst digit of the non-integer metallic means is m = 2 for
the rst three; m = 3 for the second ve; m = 4 for the third seven; m = 5 for the
fourth nine and so on. The continued fraction expansions of the non-integer metallic
means are “palindromic”, i.e., the periods are symmetric about their centres, except
for the last number of the period, which equals 2m − 1. No simple rule is known
that predicts period lengths in general, but we can state that some periods are very
short, for example, the rst ones after an integer solution have period length equal to
1 : q = 3[2; 3]; q = 7[3; 5]; q = 13[4; 7]; q = 21[5; 9]; : : : and the last ones before an inte-
ger solution are of the form: q = 5[2; 1; 3]; q = 11[3; 15]; q = 19[4; 17]; : : : , that is, they
have period length equal to 2. Some other periods are very long and except for the
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 727

cases q = 3; 7; 13; 21; 31; : : : the remaining continued fraction expansions present “stable
cycles” of dierent lengths.
In the general case of the quadratic Eq. (2.2), it can be veried that their positive
solutions have periodic continued fraction expansions [5].

3. Additive properties

Let us now form the sequence of ratios of consecutive terms of sequence (1.1)
1 2 3 5 8 13 21 34 55 89
; ; ; ; ; ; ; ; ; ;::: (3.1)
1 1 2 3 5 8 13 21 34 55
Obviously, this sequence converges directly to the golden mean . This sequence is
very useful as a good approximation: indeed the term u(11) = 233=144 = 1:6180 with
four exact decimals! If we now take a geometric progression of ratio  such as
1 1
:::; ; ; 1; ; 2 ; 3 ; : : :
2 
we can easily verify that this geometric progression is also a GSFS. In fact,
1 1 1+
+ = = 1:
2  2
The same happens for the silver mean Ag , starting from the sequence
1 3 7 17 41 99 140
; ; ; ; ; ; ;:::; (3.2)
1 1 3 7 17 41 99
that converges to Ag . The sequence
1 1 2 3
::: 2
; ; 1; Ag ; Ag ; Ag ;:::
Ag Ag

is a geometric progression of ratio Ag which satises condition (2.6). Indeed

1 2 2 3
+ 2 = Ag ; 1 + 2Ag = Ag ; Ag + 2Ag = Ag ;:::
Ag

Similarly, it is easy to prove that the sequence of ratios

1 4 13 43 142 469
; ; ; ; ; ;:::; (3.3)
1 1 4 13 43 142
√
3+ 13
converges to the bronze mean Br = = [3] and the sequence
2
1 1 2 3
::: 2
; ; 1; Br ; Br ; Br ;:::
Br Br
728 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

is a geometric progression of ratio Br . Indeed,

1 2 2 3
+ 3 = Br ; 1 + 3Br = Br ; Br + 3Br = Br ;:::
Br
Similarly for all GSFS. These numerical sequences (3.1), (3.2), (3.3), and so on,
have to be empirically used as approximations to the values of the members of the
MMF. Furthermore, the sequences formed by taking these members as ratios enjoy the
following unique mathematical property:

They are the only positive quadratic irrational numbers that originate
GSFS (with additive properties) which are, simultaneously, geometric
progressions.

This curious property of satisfying both arithmetic additive and geometric properties,
bestow on all the members of the metallic means family interesting characteristics
to become the basis of some systems of geometric proportions in Design, that are
thoroughly analyzed in [5].

4. Proportions systems
√
The golden mean  = (1 + 5)=2, is indissolubly linked to pentagonal symmetry.
Indeed, if we take a regular pentagon of unitary edge, like the one depicted in Fig. 4.1,
it is easy to prove that its diagonal is equal to . Considering the geometric similarity
of the two isosceles triangles ADC and ABF we have
AD DC
= :
DC AD − FD
Being DC = FD = 1 and calling x = AD, we obtain the quadratic equation x(x − 1) = 1
or x 2 − x − 1 = 0, that is Eq. (2.1) with n = 1 and positive solution x = . It is not
dicult to prove besides the following “golden relations” in the regular pentagon
1 1
GB =  − 1 = = 0; 618 : : : ; GI = FG = 1 − ;
 
1 1
FG = = 0; 382 : : : ; JG = = 0; 236 : : :
2 3

Innumerable are the references to the apparition of the golden mean  in the pro-
portion systems adopted by antique civilizations in their constructions, as well as its
presence in the human body proportions and in Botany.
Instead, the silver mean is linked to octagonal symmetry, as it is shown in Fig. 4.2.
“Silver relations” have been found in many examples, coming from quite dierent
elds of human knowledge (see [2, 3, 10, 12]).
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 729

Fig. 4.1. Golden divisions in a pentagon of unity side.

5. Generalized Fibonacci and Lucas numbers

Starting with F(0) = F(1) = 1 and keeping the recursion law

F(n + 1) = F(n) + F(n − 1); (5.1)

it is possible to obtain another type of generalized Fibonacci sequence, dened by

L(n + 1) = L(1)F(n) + L(0)F(n − 1): (5.2)

Of course, for L(0) = L(1) = 1, we recover the original Fibonacci sequence (1.1).
For L(0) = 2; L(1) = 1 one obtains another sequence

2; 1; 3; 4; 7; 11; 18; 29; 47; 76; 123; 199; 322; : : : ;

the so-called Lucas sequence, introduced by Edouard Lucas (1877) and mentioned in
Ref. [9]. Since this sequence satises the same recursion law as the Fibonacci numbers,
that is L(n + 2) = L(n + 1) + L(n), it is obvious that the ratio of two successive Lucas
numbers also approaches the Golden Mean,
L(n)
lim = :
n→∞ L(n − 1)
730 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Fig. 4.2. Silver relations in an octagon. Â indicates the silver mean.

A closed form for the Lucas numbers is [14]

 n
n −1
L(n) =  + :


From here, we can obtain an expression for the odd powers of the golden mean,

1 1
 2n+1 = L(2n + 1) + = L(2n + 1) + 1
 2n+1 L(2n + 1) +
..
.

 2n+1 = [L2n+1 ] (5.3)

a purely periodic continued fraction expansion! Indeed,

 = [1]; 3 = [4]; 5 = [11]; 7 = [29]; 9 = [76]; 11 = [199]; : : :

Instead, the even powers of the golden mean have continued fraction expansions that
are periodic of period length 2, satisfying the following curious relation:

 2n = [L(2n) − 1; L(1); L(2n) − 2]; (5.4)

 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 731

where L(1) + L(2n) − 2 = L(2n) − 1. In fact

 2 = [2; 1]; 4 = [6; 1; 5]; 6 = [17; 1; 16];
8 = [46; 1; 45]; 10 = [122; 1; 121]; : : :

6. Powers of the members of the MMF

All the members of the MMF that have purely periodic continued fraction expansions
of the form [n], satisfy similar relations as the golden mean  = [ 1 ]. Indeed
2 3 4 5
Ag = [2]; Ag = [5; 1; 4]; Ag = [14]; Ag = [33; 1; 32]; Ag = [82];
6 7 8 9
Ag = [197; 1; 196]; Ag = [478]; Ag = [1153; 1; 1152]; Ag = [2786]; : : :
Similarly for
2 3 4
Br = [3]; Br = [10; 1; 9]; Br = [36]; Br = [118; 1; 117];
5 6 7
Br = [393]; Br = [1297; 1; 1296]; Br = [4287]; : : :
The conclusion is that for these purely periodic metallic means, the odd power contin-
ued fraction expansions are also purely periodic, while the even powers are periodic
of period length 2, satisfying the curious relation that the rst coecient is equal to
the sum of the next two.
Instead, for the rest of the members of the MMF, whose continued fraction expansion
are only periodic, there are no simple rules. In some cases, the powers are periodic
but in other cases, the period length is so big that it is dicult to conjecture that they
are all periodic.

7. Elements of hyperbolic geometry

As is well known, stating that through a point exterior to a straight line, there are two
parallels and innite non-secants, we get the so-called “hyperbolic geometry”. There
are many models of hyperbolic geometry and we will choose the so-called “PoincarÃe’s
model ”. This model consists in mapping the upper-half complex plane H on the interior
of the Poincare’s unitary disk D:
H = {z ∈ C : Im(z)¿0} → D = {z ∈ C : |z|¡1}; (7.1)
where C is the set of complex numbers.
Straight lines are half-circumferences whose centers are on the x-axis (Fig. 7.1) and,
as a limit case, they can be normal straight lines to the x-axis. These lines are mapped
on D as arcs of circumferences normal to the boundary.
On the hyperbolic plane, the movements are bilinear transformations which map H
to itself, that is, invertible maps of the form
az + b
z0 = ; (7.2)
cz + d
732 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Fig. 7.1. Poincare’s model.

Fig. 7.2. Hyperbolic distance between two points A and B.

where a; b; c; d are real constants. To normalize, we may suppose that

ad − bc = ±1.
It is very easy to prove the following properties: (a) Every straight line is determined
by two points. (b) Every straight line is unlimited, since the x-axis does not pertain
to the hyperbolic plane H : (c) Angles on the Poincare’s circle are measured as on the
Euclidean plane. (d) The distance between points A and B of Fig. 7.2 is given by the
following expression:
 
tan g(ÿ=2)
d = k log :
tan g(=2)

From here we have AU = +∞; AV = −∞, i.e., the x-axis is the innite of the plane.
(e) Given a straight line a and an exterior point A like in Fig. 7.3, there are two
lines that bind A with the two points of innite U and V of the line. This means
through the point A we have two parallels to the line a. (f) Circumferences tangent to
the x-axis are the “horocycles” (Lobachevsky introduced this term that literally means
“limit circles”) and circumferences secant to the x-axis are the “hypercycles”, like it
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 733

Fig. 7.3. Through a point A there are two parallels to the line a.

Fig. 7.4. Hypercyles.

is shown in Fig. 7.4. (g) The angle sum of a triangle can be anything from zero to
180◦ , provided it is strictly less than 180◦ .

8. Fuchsian groups

In a paper published in 1982 (see [15]), the English mathematician Caroline

Series presented a mathematical linkage among hyperbolic geometry, continued fraction
expansions and ergodic theory. To achieve this goal, she uses discrete groups of hy-
perbolic geometry, the so-called “Fuchsian groups”. A Fuchsian group G is a discrete
group of symmetries described by 2 × 2 matrices of the form
 
a b
c d
that are isolated from each other in the group of all invertible matrices.
Let us consider one important example of a Fuchsian group, the so-called “uni-
modular group” U , where a; b; c; d are integer numbers and ad − bc = 1. Let
az + b
z0 = : (8.1)
cz + d
Fixed points of this transformation are the solutions of the equation z = (az + b)=
(cz + d) or cz 2 + (d − a)z − b = 0. If c 6= 0, this is a quadratic equation and the
projectivities are classied depending on the sign of the discriminant
= (d−a) 2 +4bc.
734 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

•
¿0: In this case there are two real dierent roots and the transformation is called
“hyperbolic”. There are no xed points in the hyperbolic space but there are two xed
points on the boundary, that is, there is a unique geodesic (line of minimum length)
that connects these two points and is preserved by this transformation, receiving the
name of axis of the transformation.
•
= 0: Then there is a double root and the transformation is called “parabolic”.
There are no xed points in the hyperbolic space but there is a xed point on the
boundary of the disc. If we take the innite xed point in the half-plane H , we get
a translation.
•
¡0: In this case there are two complex roots and the transformation is called
“elliptic”. There is a xed point in the hyperbolic space (center of the disc) and the
transformation is a rigid rotation of the disc.
Associated to any discrete group G of symmetries, there is a tessellation or tiling of
the universe of the geometry. If we take into account property (g) of the angle sum of
a triangle in hyperbolic space, it is clear that there are innitely more tessellations
of hyperbolic space than Euclidean space. In Euclidean space you can tile through
triangles, squares or hexagons. Instead, in hyperbolic space you can arrange for as
many hexagons, for example, to meet at a vertex as you like.
Given a discrete group G, one constructs the tessellation by looking for a funda-
mental region for the group. This is a set X such that each transformation in G carries
X into a disjoint set gX , and that each point of the plane is carried into some point
in X by some element of G. Polygonal fundamental regions always exist for Fuchsian
groups. In another words, there is a natural crystalline structure associated with the
group G.

9. Limit sets in the PoincarÃe circle

Let us consider the tiling shown in Fig. 9.1. The region R is a fundamental region
for a certain Fuchsian group G. It is evident that at this tiling, the tiles get smaller
and smaller as we approach the x-axis = @H (boundary of H ) or @D (boundary of the
disk D). In this situation, the “limit set” L of the Fuchsian group G is dened to be
the set of points at innity which are limit points of an orbit. It turns out that L is
independent of the particular orbit chosen and that there are two possibilities:
(a) L is either the whole circle like in Fig. 9.2 or
(b) L is a set of zero Lebesgue measure (with holes), like in Fig. 9.1.
In case (b), the limit set is the proper self-similar fractal obtained by intersecting
the smaller and smaller arcs inside @D. The generators are transformations A; B; C, such
that A 2 = B 2 = C 2 = I (identity) and they re ect the disk D over the inverted circular
mirrors a; b and c. Then, we have a copy of b and c inside a and so on. Consequently,
the fractal limit set has the following coding:

a b c
ab ac ba bc ca cb
abc aba aca acb bab bac bca bcb cab cac cba cbc
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 735

Fig. 9.1. Limit set with holes.

and so on. Note the inÿnite word structure associated with the limit set of G: longer
words code for smaller regions and innite word structures code for points on the
fractal.

10. Continued fractions

Let us consider two transformations P(z) = z + 1; Q(z) = −1=z from the unimodular
group U . The rst is a translation and the second one, an inversion. Their matricial
forms are
   
1 1 0 1
P= 0 1 Q = −1 0 (10.1)
(hyperbolic) (elliptic)

It is easy to see, that P n = z + n; P −n = z − n. Furthermore, the C . Series has also

proved (see [16]) that P and Q generate all U and the proof is based on continued
fraction expansions. Let us now consider a real positive number x given as a continued
fraction
1
x = [n1 ; n2 ; : : : ; nk ; : : :] = n1 + :
1
n2 +
1
n3 +
..
.
736 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Fig. 9.2. Limit set that is the whole circle.

We can associate with each such x the innite product of generators x = P n1

QP −n2 QP n3 QP −n4 · · · In this form, the connection with hyperbolic geometry becomes
evident, since as we have seen, every real number is a limit of a sequence of points
in H , which lie in some orbit of the unimodular group U .

11. Farey tessellations

Given a natural number n, if we order all the reduced fractions whose non negative
denominators are less than or equal to n, in increasing magnitude, we obtain the so
called “Farey sequence” of order n, e.g., for n = 5 in the interval [0; 1], the Farey
fractions of order 5 are
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 737

0 1 1 1 2 1 3 2 3 4 1
:
1 5 4 3 5 2 5 3 4 5 1

Note that each fraction is the “mediant” (p + p0 )=(q + q0 ) of its neighbors. For other
intervals, Farey fractions are congruent mod 1 to that indicated for [0; 1], so we will
restrict our attention to this interval. While Farey sequences have many applications
and very nice properties, such as classifying the rational numbers according to the
magnitudes of their denominators, they have a completely irregular behavior if we list
the number of additional fractions in going from Farey sequence of one order to the
following.
A much more regular order is obtained by “Farey trees”, in which the number
of fractions added with each generation is a power of 2. Thus, the Farey tree for the

Fig. 11.1. Farey tree.

Fig. 11.2. Farey tessellation.

738 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Fig. 11.3. Farey triangle.

Fig. 11.4. Farey tessellation in terms of Farey trees.

interval [0; 1] looks as is shown in Fig. 11.1. If we zigzag down the Farey tree from its
upper right corner, we nd fractions whose numerators and denominators are curiously
Fibonacci numbers Fn ! It is easy to see that these ratios Fn =Fn+1 converge to 1= when
n → ∞. Following the zigzag down the Farey tree from its upper left corner, we nd
ratios Fn =Fn+2 , which converges to 1= 2 when n → ∞. The ratios that form the Farey
tree are sometimes called the Farey-Brocot sequence. Following the C Series, we now
introduce the “Farey tessellation” F (Fig. 11.2). The Farey tessellation is a tessellation
of H by triangles S whose vertices lie on R ∪ {∞}. Precisely, the set of vertices is the
set of rationals {∞}. Reduced fractions of the form p=q; p0 =q0 are the endpoints of
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 739

a side of a triangle in F if and only if

 
 p p0 
 
 q q0  = ±1:

The sides of F turn out to be the images of the imaginary axis under G and in
Fig. 11.3, a classical Farey triangle is shown. Such Farey triangles can tessellate all
of H or the disk D. In terms of Farey trees, Farey tessellation may be obtained by
drawing the vertical lines through 0 and then successively joining adjacent points in
each Farey sequence, like in Fig. 11.4.

12. Hyperbolic geometry and multifractal spectra

With all these ingredients, Maria Piacquadio and her co-workers developed a math-
ematical model to characterize non-regular fractal sets appearing in physics, biology,
chemistry, etc. [1, 6, 7]. This characterization is based on the multifractal spectral de-
composition technique introduced by Procaccia. Using hyperbolic empirical ndings, as
we know, for each Fuchsian group, we have a tiling of H or D. The limit set of this
tessellation is a “fractal” at the innity line. These fractal structures can be self-similar
or not, in which case they are “multifractals”.
This approach has to be related to continued fraction expansions, Farey tessellations
and the innite word problem.
Let us return to the classication of matrices from the unimodular group, and let us
put it in another form, more useful to compare with empirical results.
Given a matrix!
a b
A= ∈U ; ad − bc = 1
c d

we have to analyze the sign of the discriminant
= (a + d)2 − 4. Consequently, if

¿0 then (a + d)2 − 4¿0 or |a + d|¿2, that is, the absolute value of the trace of the
matrix A is greater than 2. It is similar in the remaining two cases and so we can state:
(a) Hyperbolic case: |trace(A)|¿2.
(b) Parabolic case: |trace(A)| = 2.
(c) Elliptic case: |trace(A)|¡2.
Given a Fuchsian group G with, say, elliptic generators, then for the limit set L(G)
to be a structure of positive fractal dimension, we need a minimum of three genera-
tors, while when dealing with parabolic generators, we need a minimum of two. Such
a group generated by a minimal number of generators will be called a “minimally
generated group”.
For parabolic generators, we know that the simplest of all parabolic matrices are the
following:
! ! !
1 0 1 0 1 1
I= identity; A = ; P= :
0 1 1 1 0 1

In terms of P and Q we have A = QP −1 Q. The fundamental region is the quadrilateral

T , indicated in Fig. 12.1a. With this fundamental region, we can tessellate H , by
740 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Fig. 12.1(a). Quadrilateral T .

performing the actions P n T , where n is an integer. In Fig. 12.1b it is shown how A,

the only word with one letter, touches 0 and 1; the words of two letters introduce
the number 1=2; the words of three letters introduce 1=3 and 2=3, etc., which are the
Farey–Brocot partition of the segment [0; 1]. Note that with this parabolic partition,
we get the complete Farey tree and proceeding in the same way with all segments
[n; n + 1]; n integer, as we just did with [0; 1], we obtain an object of dimension one.
That is

The limit set L of the parabolic group G is all R

Grynberg and Piacquadio found for the minimum value of the concentration,
√ √
ln 2 ln 2
min = = = 0:71:
2 ln  ln 

What is more relevant about this numerical result is that this theoretical value agrees
with the one obtained in the graphical representation of the spectral curve (; f()).
E.g., Procaccia and Zeitak [13], stretching a rubber band around a DLA (diusion-
limited aggregation) grown o lattice in two dimensions, predicted the value min = 57 =
0:71. Interesting is the fact that the innite word associated with min is the periodic
AP or (AP)∞ .
For elliptic generators, we know that to get a minimally generated group, we need
three elliptic generators. For simplicity, we may choose generators with a minimal
value of |trace|, i.e. zero trace. Besides, let us recall that given a hyperbolic movement
z 0 = (az + b)=(cz + d), the equation |cz + d| = 1 corresponds to the “isometric circles”
of the movement. Let us assume that the centres of the three isometric circles will be
equidistant from one another. The three elliptic matrices are the following:
! ! !
−a −(1 + a2 ) 0 −1 a −(1 + a2 )
T−a = ; T0 = ; Ta = ;
1 a 1 0 1 −a
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 741

Fig. 12.1(b). Farey–Brocot partition.

Fig. 12.2. Isometric circles (from [12]).

and, as it is easy to prove, the three isometric circles of unitary radii, have their
centers at the points −a; 0; a. The limit set on @H or on the boundary @D, is a fractal
of dimension strictly between 0 and 1, and can be visualized as the set of points in
R inside an innity of circles, as is suggested by Fig. 12.2, where, for short, −a; 0; a
indicate the isometric circles of T−a ; T0 and Ta . In this case, min is coded also by an
innite word with a period of two letters Ta T−a .
742 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

Fig. 13.1. Tel’s universal curves (from [16]).

13. Modeling the (; f()) spectral curves

Tamas Tel [18] has classied qualitatively the (; f()) spectral curves that can arise
empirically in dierent multifractal spectra. He classied the universal curves (; f())
into four types, I, II, III and IV as shown in Fig. 13.1. With three elliptic generators of
zero |trace|, it is possible to generate all four Tel types I to IV of functions (; f()):
( ! ! !)
0 1 13 −17 4 −17
GI = ; ; ;
−1 0 10 −13 1 −4
( ! ! !)
−6 −37 0 1 3 −10
GII = ; ; ;
1 6 −1 0 1 −3
( ! ! !)
−M −M 2 − 1 0 1 3 −10
GIII = ; ; ;
1 M −1 0 1 −3
( ! ! !)
a −a2 − 1 0 1 −a −a2 − 1
GIV = ; ; :
1 −a −1 0 1 a

A remarkable fact is that with a minimal number (e.g. two) of hyperbolic generators,
a type not classied by Tel is obtained [17]! This new type is symmetrical like type
IV but with part of the left leg missing. Looking for this sort of curve, we found
that Stanley and Meakin [17] presented similar theoretical and experimental plots of
the function (; f()) for viscous ngering patterns, that look like the ones shown
in Fig. 13.1. Furthermore, Procaccia and Zeitak [13], considered a family of exactly
solvable models which exhibit complex ramied fractal objects with any desired number
of major ngers, whose spectra can be systematically obtained. The results are plotted
in Fig. 13.2a–d.

These facts indicate clearly that the set of TÃel’s universal spectral curves is
included in the set of spectral curves obtainable by using this hyperbolic
model.
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 743

Fig. 13.1*. Stanley and Meakin’s plots (from [17]).

Fig. 13.2. Procaccia and Zeitak’s results (from [15]).

Obviously, the generation of these spectral curves is not unique, e.g. type III can be
generated with the simplest parabolic matrices A and P.

14. The inÿnite word structure

Given a fractal object

, let us assume that we have decomposed it into a disjoint
union of sets
 , each self-similar. If we take into consideration the Fuchsian group
GIV , it is easy to see that words coding for points in
min are of the form T−a Ta T−a : : : ;
i.e.

P a QP −a P −a QP a P a QP −a · · · = P a QP −2a QP 2a QP −2a Q · · ·

It could also be any other word agreeing on an innite tail. The same happens in the
parabolic case: the points in min are coded by innite words agreeing on a tail with
744 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745

PAPAP: : : The
max for the parabolic case is coded by an innite word of the type
AAA: : : with one generator only. However, in the minimally generated hyperbolic case,

min is coded for A∞ , whereas ABABAB: : : codes for

max .
From the innite word structure point of view, there are large dierences between
the elliptic and parabolic cases and the hyperbolic ones, i.e. between |trace| ≤ 2 and
|trace|¿2. This theoretical fact agrees with important empirical results obtained by
Kohmoto et al. (mentioned in [11]), in the study of the electronic energy spectrum of a
quasicrystal: the condition |trace| ≤ 2 corresponds to allowed energies while |trace|¿2
corresponds to forbidden energies.

15. Inÿnite words, continued fractions and the metallic means

Let us recall that for√the parabolic Fuchsian group, the minimum value of the con-
centration is min = ln 2= ln . For the elliptic case, instead, the following extreme
values for the concentration or strength  are quoted:
ln 2 2 ln 2
min = √ ; max = √ :
ln(a + a2 + 1) ln((a + a4 + 4)=2)
The parameter a is not necessarily an integer, the only condition imposed on it is
to be a positive value not smaller than one. √
Let√us consider dierent values of the parameter a to calculate a + a2 + 1 and
(a + a4 + 4)=2 and let us state the results as continued fraction expansions (Tables 1
and 2).
If we restrict ourselves to the case where a is a natural number, we have the follow-
ing result: the innite word (Ta ; T−a )∞ codes for min in the elliptic case. This innite
word can be expressed in terms of matrices P and Q in the following way:
P a QP −a P −a QP a P a QP −a · · · = P a QP −2a QP 2a QP −2a Q · · ·
whose codes for the real number x = [a; 2a; 2a; 2a; : : :] = [a; 2a]. The set of√real numbers
with a tail [2a; 2a; : : :] is the set of real numbers [2a; 2a; 2a; : : :] = a + a2 + 1. For
instance, if a = 1, then all silver numbers have min concentration; if a = 2, all 3 -
golden numbers have minimal concentration and so on. This circumstance indicates
that the value of a cannot be restricted to be a natural number.
In such a way, we arrive at the following conclusions:
• For any value of a, natural or irrational, the continued fraction expansions of the
two expressions stated in Tables 1 and 2, determining the values of min and max ,
correspond to members of the metallic means family.
• For a natural, the values of min for the elliptic Fuchsian group, have purely periodic
continued fraction expansions, related to the golden mean  and the Silver Mean
Ag , as well as to its powers.
• The values of max for the elliptic Fuchsian group have periodic and purely periodic
continued fraction expansions, related to the golden mean , the silver mean Ag
and the nickel mean Ni , as well as to its powers.
This indicates the existence of deep underlying relationships among minimally gen-
erated Fuchsian groups, innite word structures, Farey–Brocot sequences, continued
 V.W. de Spinadel / Nonlinear Analysis 36 (1999) 721 – 745 745

Table 2
√
a+ a4 + 4
a
2

1  = [1]
√
2 Ag = [2]
Table 1 √
3 Ni = [2; 3]
√
a a+ a2 + 1 2 3 = [4]
√
5 15 = [2; 1; 3]
1 Ag = [2] √
2 3 = [4] 6 16 = [6]
√ 2 = [5; 1; 4] √
8 Ag 8 18 = [8]
3 [6] 3 [6; 3]
4 [8] 4 [10; 4]
5 [10] 5 [15; 5]
7 3 = [14]
Ag 7 [28; 7]

fraction expansions and metallic means, relationships which open the road for the fu-
ture study and analysis of a galaxy of scientically related phenomena.

References

[1] E. Cesaratto, S. Grynberg, R. Hansen, M. Piacquadio, Multifractal spectra and hyperbolic geometry,
Chaos Solitons Fractals 6 (1995) 75 – 82.
[2] V.W. de Spinadel, El Modulor de Le Corbusier, AREA (3) February 1996.
[3] V.W. de Spinadel, Algo mas de Matematica, AREA (4) August 1996.
[4] V.W. de Spinadel, On characterization of the onset to chaos, Chaos, Solitons Fractals 8 (10) (1997)
1631–1643.
[5] V.W. de Spinadel, From the Golden Mean to Chaos, Nueva Librera, 1998.
[6] S. Grynberg, M. Piacquadio, Hyperbolic geometry and multifractal spectra, Part I, preprint 251; Part II,
preprint 252; Part III, preprint 257, Instituto Argentino de Matematica, 1995a.
[7] S. Grynberg, M. Piacquadio, Geometric structure of the Farey-Brocot sequence, Revista de la Union
Matematica Argentina 39 (1995b) 209 –222.
[8] T.C. Halsey, M.H. Jensen, J.P. Kadano, I. Procaccia, B.I. Shraiman, Fractal measures and their
singularities: the characterization of strange sets, Nucl. Phys. B 2 (Proc. Suppl.) (1987) 501–512.
[9] V.E. Hogatt, Fibonacci and Lucas Numbers, Houghton Miin, Boston, 1969.
[10] Jay Kappra, Musical proportions at the basis of systems of architectural proportion both ancient and
modern, in: Kim Williams (Ed.), NEXUS – Architecture and Mathematics, Edizioni dell’Erba, 1996.
[11] M. Kohmoto, B. Sutherland, C. Tang, Critical wave functions and a Cantor-set spectrum of a one-
dimensional quasi-crystal model, Phys. Rev. B 35 (2) (1987) 1020 –1033.
[12] Le Corbusier, Le Modulor, 1950. Modulor 2, 1954. Paris, France.
[13] I. Procaccia, R. Zeitak, Shape of fractal growth patterns: exactly solvable models and stability
considerations, Phys. Rev. Lett. 60 (24) (1988) 2511–2514.
[14] M.R. Schroeder, Number Theory in Science and Communication, with Applications in Cryptography,
Physics, Digital Information, Computing and Self-Similarity, 3rd ed., Springer, Berlin, 1997.
[15] C. Series, Non-euclidean geometry, continued fractions and ergodic theory, Math. Intell. 4 (1) (1982).
[16] C. Series, The modular surface and continued fractions, J. London Math. Soc. 31 (2) (1985) 69 – 80.
[17] E.H. Stanley, P. Meakin, Multifractal phenomena in physics and chemistry, Nature 335 (1988) 405 – 419.
[18] Tamas Tel, Fractals, multifractals and thermodynamics, Z. Naturforsch. A 43 (1988) 1154 –1174.