Continued Fractions, Discrete Groups and Complex Dynamics

Computational Methods and Function Theory
Volume 1 (2001), No. 2, 535–594
Continued Fractions, Discrete Groups and Complex

Dynamics
Alan F. Beardon
Abstract. It is well-known that a continued fraction may be regarded as a se-

quence of Möbius maps. In this partly expository paper we consider continued
fractions by examining the action of Möbius maps in hyperbolic space, and in
all dimensions. We obtain a general version of the Stern-Stolz theorem valid
in all dimensions, and we draw comparisons between the theory of continued
fractions and the theories of discrete groups of Möbius maps, and complex
dynamics.
Keywords. Continued fractions, Möbius maps, complex dynamics, discrete

groups, hyperbolic geometry, inversive geometry.
2000 MSC. Primary 40A15; Secondary 30B70, 11J70.
Contents
1. Introduction 536
2. Continued fractions 540
3. Complex Möbius maps 542
4. The stability of continued fractions 553
5. Möbius maps and hyperbolic geometry 559
6. General convergence 564
7. Strong divergence 568
8. Möbius maps in higher dimensions 573
9. Backward orbits 581
10. Forward orbits 585
11. Self-maps of a domain 586
12. Concluding remarks 590
References 591
Received April 29, 2002.
c 2001 Heldermann Verlag
ISSN 1617-9447/$ 2.50 °
536 A. F. Beardon CMFT
1. Introduction
This paper is about the connections between continued fractions, Möbius maps,
discrete groups, complex dynamics, hyperbolic and inversive geometries.
Traditionally, an infinite continued fraction is considered to be an expression of
the form
a1
(1.1) K(an |bn ) = ,
a2
b1 +
a3
b2 +
b3 + · · ·
where the ai and bj are infinite sequences of complex numbers with each ai 6= 0.
The continued fraction (1.1) is said to converge to, or to have value, K if the
sequence of truncated continued fractions
a1 a1 a1
(1.2) , , ,...
b1 a2 a2
b1 + b1 +
b2 a3
b2 +
b3
converges to K. More generally, one can consider continued fractions a0 +
K(an |bn ), but there is no harm in taking a0 = 0 as we shall do throughout
this paper.
From an abstract point of view, the continued fraction (1.1) is simply the pair of
complex sequences an and bn in which no ai is zero. Given two such sequences
we can form the sequence sn of Möbius maps defined by
an
(1.3) sn (z) = ,
bn + z
and sn (∞) = 0. Conversely, as any Möbius map s with s(∞) = 0 can be
expressed uniquely in the form s(z) = a/(b + z), we can identify the class of con-
tinued fractions with the class of sequences sn of Möbius maps with sn (∞) = 0
for all n. With this identification the truncated continued fractions in (1.2) are
s1 (0), s1 s2 (0), . . ., where, in general, f g denotes the map z 7→ f (g(z)), so that
the convergence of the continued fraction is equivalent to the convergence of
the complex sequence s1 · · · sn (0). As Möbius maps exist in all dimensions, the
identification of continued fractions with sequences of Möbius maps immediately
provides us with the concept of continued fractions in all dimensions, namely
sequences sn of Möbius maps with sn (∞) = 0. In higher dimensions the op-
eration of division is essentially replaced by inversion in the unit sphere, and
the arguments become geometric rather than algebraic. In this sense, continued
fraction theory may be regarded as a part of inversive geometry in n dimen-
sions. As Möbius maps acting in a higher dimension satisfy most of the familiar
properties enjoyed by complex Möbius maps, we might hope that many results
1 (2001), No. 2 Continued Fractions, Discrete Groups and Complex Dynamics 537
on continued fractions will be true regardless of the dimension. In fact, we al-

ready know that some of the basic results on continued fractions (for example,
Pringsheim’s Theorem, the Hillam-Thron Theorem and the Parabola Theorem)
do generalize to all dimensions, so this observation suggests other directions for
exploration. This paper was motivated by the desire to understand fully the role
of the condition sn (∞) = 0 in the theory. Our tentative conclusion is that the
basic convergence theory of continued fractions holds for all generally convergent
sequences of Möbius maps (these will be defined later), in all dimensions, and
that the condition sn (∞) = 0 is not essential.
The idea that a continued fraction may be regarded as a sequence of complex
Möbius maps is not new and, according to Thron ([75]), the germ of this idea can
be found in the works of Isenkrahe (1888), Netto (1892), Schur (1917), and Hamel
(1918). In 1942 J. F. Paydon and H. S. Wall [65] published a paper with the
title “The continued fraction as a sequence of linear transformations”, and this
perhaps marks the time from which continued fractions were consistently viewed
as sequences of Möbius maps. As is stated in the opening paragraph of [53], this
contrasts with the earlier view of continued fractions as the limit of a sequence
An /Bn , where the An and Bn are solutions of a certain second order recurrence
relation. The first text on continued fractions which adopted this analytic point
of view was Wall’s text [83] published in 1948, and it is interesting to note
that this text is dedicated to L. R. Ford who made a significant contribution to
the theory of discrete groups of Möbius maps and automorphic functions (see
[38]). In fact, Ford also published papers on continued fractions (for example,
[35], [36], [37]), and he considered continued fractions from the point of view
of the Modular group and the Picard group (the groups of Möbius maps whose
coefficients are integers, and Gaussian integers, respectively). These are groups of
the isometries of the hyperbolic plane, and hyperbolic 3-space, respectively, and
these ideas have been used to good effect recently in connection with dynamical
systems. The use of Möbius maps in continued fraction theory opened the way
to apply theorems in complex analysis, and the groundwork for this was laid
much earlier by Stieltjes who, in his long paper [73] (see also [59]) on continued
fractions, introduced ideas about normal families of analytic maps before they
were named and studied (albeit in greater generality and detail) by Montel.
The introduction of Möbius maps into the theory of continued fractions raises the
following issue. Any Möbius map is either of the form (1.3) or is the composition
of two maps of this form for, if a 6= 0, then
Ã ! Ã !Ã !
a b 0 a 0 ad−bc
a
= b
.
c d 1 c 1 a
This implies that any given sequence f1 , f2 , . . . of Möbius transformations arises

as approximately every second term in the sequence of approximants of some
continued fraction. Given this, surely there is a obligation to see whether any
given result on continued fractions is really a result about a general sequence

of Möbius maps, and if it is not, to explain fully the particular role played by
continued fractions. A few papers have addressed this issue, but it does not seem
to have been considered in a systematic way.
Möbius maps have been studied extensively in the theory of discrete groups since
the time of Poincaré and Klein, and it was Beltrami, Poincaré and Klein who
showed us that the natural action for Möbius maps takes place in hyperbolic space,
for it is there that these maps act as isometries. The extended complex plane C∞
is the boundary of hyperbolic three-space, and if we restrict our attention to the
action of Möbius maps in C∞ then we see only a very small part of the overall
picture. More generally, the action of Möbius maps on Euclidean n-dimensional
space is just their action on the boundary of hyperbolic (n + 1)-dimensional
space. It seems fair to say that the geometric theory of Möbius maps and the
huge amount of work, initiated by Beltrami, Poincaré and Klein, and developed
more recently by Ahlfors and others, on the study of discrete groups of Möbius
maps and their relationship to hyperbolic geometry (in all dimensions) seems
to have been largely ignored in the context of continued fractions. It seems
reasonable, then, to try to apply the material already available in the theory of
discrete groups to our study of continued fractions.
Although the connection between continued fractions and Möbius maps is now
firmly established, it is almost entirely restricted to the action of Möbius maps
in the extended complex plane, and a great deal of emphasis is still placed on
the algebraic manipulation of the complex coefficients of Möbius maps. It is
possible to give an algebraic treatment of Möbius maps in all dimensions if we
use the Clifford algebras (a generalization of quaternions), and this opens up the
possibility of continued fractions with coefficients in a Clifford algebra. Although
this is likely to be difficult, there might be something to be gained in the special
case of quaternions; indeed, Hamilton himself initiated a preliminary study of
continued fractions with quaternions as coefficients ([44], [45]) as long ago as
1852. Alternatively, we can try to use the description of hyperbolic geometry in
terms of the quadratic form x1 y1 + · · · xn yn − xn+1 yn+1 , for here the isometries
are matrices. Examples from both of these points of view will be given later, but
for the most part we shall try to develop a view of continued fractions that does
not explicitly use the coefficients of the maps.
Finally, we remark that there is a very close connection between the iteration of
rational maps in the complex plane (that is, complex dynamics) and the theory
of discrete Möbius groups, so we should also look at complex dynamics to see
whether the ideas there can contribute to the study of continued fractions. We
notice, for example, that the tail sequences that are used in continued fraction
theory are the same as the inverse orbits in complex dynamics. The inverse orbits
of a rational map play a fundamental role in complex dynamics; for example, they
accumulate precisely at the Julia set of the rational map, and we shall see that a
similar phenomenon holds in continued fraction theory. Moreover, other results

from complex dynamics (and discrete group theory) suggest possible results in
continued fraction theory. In keeping with our notation f g for the composite
map, f n denotes the n-th iterate of the map f .
In this paper we shall attempt to show how the ideas described above can be
used to illuminate the general theory of continued fractions. Given a result on
continued fractions, we ask the following five questions.
(1) Is there a geometric proof of this result?
(2) Can the result be generalized to arbitrary sequences of Möbius maps?
(3) Is the result and proof valid in all dimensions?
(4) Can the result be given in algebraic terms in higher dimensions?
(5) Is there an analogous result in complex dynamics or discrete group theory?
We have written this partly expository paper primarily for those interested in
continued fractions, and it is written in the hope of stimulating interest in these
questions. We shall attempt to show that there is a rich source of material on
the geometric properties of Möbius maps which has not yet been fully exploited,
and indeed hardly touched, in the context of continued fractions, and the paper
contains many questions which might be investigated further. Throughout, our
aim is to illustrate a geometric approach to the theory, and we have not included
any proofs that follow well known arguments; in particular, we shall not use
difference equations anywhere in this paper. A large part of continued fraction
theory is concerned with numerical questions, and rates of approximation, and
we acknowledge that the ideas in this paper are unlikely to have an immediate
impact on this area (although it should be pointed out that it is well known in
discrete group theory that, for example, Diophantine approximation is intimately
connected to the geometric behaviour of geodesics on a Riemann surface). It is
not yet clear how much of the theory of continued fractions can be successfully
developed in the manner we are advocating here; however, it seems clear enough
that the achievements so far warrant further investigation along these lines.
Section 2 contains a very brief summary of the classical view of continued frac-
tions; this merely sets the scene for what follows. Given the aims in this paper, it
seems appropriate to give a survey of what is known about Möbius maps. We dis-
cuss complex Möbius maps in Section 3 and follow this with applications of these
ideas in Section 4. In Section 5, we discuss the action of complex Möbius maps
on three-dimensional hyperbolic space, with applications in Sections 6 and 7.
Section 8 contains a brief discussion of Möbius maps acting in higher dimen-
sional spaces, followed by more applications in the later sections of the paper.
We make no apology for giving such a lengthy discussion of Möbius maps for our
view is that this underpins much of the basic theory of continued fractions.
For the general theory of continued fractions, we recommend [51], [55], [67], [68]
and [83]. For further information on Möbius maps in all dimensions, we refer
the reader to [4], [5], [12], [23], [38], [62], [64] and [70]. Finally, for the general
theory of complex iteration, we recommend [14], [25] and [74].
2. Continued fractions
Continued fractions have been studied since at least 1572 when R. Bombelli,
and
√ later P.√ Cataldi, used finite continued fractions to obtain approximations to
13 and 18, respectively. The first infinite continued fraction expansion (of
4/π) was given by Lord Brouckner around 1659, who also gave (in 1657) the first
systematic method for solving Pell’s equation by using continued fractions (see
[30, p. 108]). In 1695 John Wallis published Opera Mathematica, a text which
contained the first systematic study of the general theory of continued fractions
and in which the term continued fraction was first introduced. Euler gave a sys-
tematic exposition in De Fractionlous Continious (1737), and Jacobi, Hermite,
Gauss, Cauchy, Stieltjes, Ramanujan, and many others have all contributed to
the subject.
We have already seen that the continued fraction (1.1) can be viewed as the
sequence of Möbius maps sn in (1.3). We shall often refer to the continued
fraction (1.1) as the continued fraction generated by the sn , or simply as the
continued fraction sn . We shall also use the notation [s1 , s2 , . . .] for this continued
fraction; thus if sn (z) = an /(bn + z), then K(an |bn ) and [s1 , s2 , . . .] are different
notations for the same continued fraction. In particular, [s, s, . . .] denotes a
periodic continued fraction (of period one). We write Sn = s1 · · · sn , and we also
say that the sj generate the sequence Sj . We shall reserve the notation an , bn , sn
and Sn as used here throughout this paper.
In general, given a sequence of maps fj of a set into itself, we can construct the
sequences Fn = f1 · · · fn and Gn = fn · · · f1 . We say that Fn is the inner compo-
sition sequence, and that Gn is the outer composition sequence, of the fn . For a
continued fraction, Sn is the inner composition sequence of its generators sn , and
many convergence theorems on continued fractions are special cases of general
results on the convergence of inner composition sequences of analytic maps. We
shall not consider outer composition sequences in this paper; usually, these are
easier to handle than inner composition sequences.
The sequence of truncated continued fraction in (1.2) is the sequence Sn (0);
thus the continued fraction (1.1) converges if and only if the sequence of maps
S1 , S2 , . . . converges at 0. The convergence of a continued fraction is unaffected by
the addition or removal of a finite number of the generators sj for s1 · · · sn (0) → α
if and only if s2 · · · sn (0) → s−1 1 (α). It is easy to see that a given sequence
z1 , z2 , . . . in C∞ is the sequence Sn (0) for some continued fraction (1.1) if and
only if z1 6= 0 and zn 6= zn+1 (see, for example, [51, p. 34] or [55, p. 71]). This
implies that virtually every complex sequence zn is the sequence of approximants
Sn (0), and as these approximants may ‘wander’ over the complex plane in an
arbitrary manner, we can certainly see the need for conditions that guarantee
convergence. Despite the historical precedent, it makes no sense to to study
the convergence of a sequence of maps at just one point, and we must concern
ourselves with the possible convergence (pointwise, and uniform) of the Sn on
subsets of C∞ .
Originally, continued fractions were considered in terms of recurrence relations.
Given the continued fraction (1.1), we define sequences An and Bn by
µ ¶ µ ¶
A−1 A0 1 0
= ,
B−1 B0 0 1
µ ¶ µ ¶µ ¶
An+1 An+2 An An+1 0 an+2
= ,
Bn+1 Bn+2 Bn Bn+1 1 bn+2
where n ≥ −1, or, equivalently, by the recurrence relations
An = bn An−1 + an An−2 ,
(2.1)
Bn = bn Bn−1 + an Bn−2 .
As
µ ¶ µ ¶ µ ¶
An−1 An 0 a1 0 an
(2.2) = ···
Bn−1 Bn 1 b1 1 bn
we can take determinants of both sides and obtain
µ ¶
An−1 An
(2.3) det = (−1)n a1 · · · an 6= 0.
Bn−1 Bn
Now consider the homomorphism
µ ¶
a b az + b
(2.4) Φ : A 7→ gA , A= , gA (z) = ,
c d cz + d
from the group GL(2, C) of 2 × 2 non-singular matrices onto the group M of
Möbius maps (this is the first of several occasions on which the reader will be
urged, with good reason, to distinguish clearly between Möbius maps and ma-
trices). Applying Φ to both sides of (2.2) we see that
An−1 z + An
(2.5) Sn (z) = .
Bn−1 z + Bn
It follows that the sequence of truncated continued fractions in (1.2) or, equiva-
lently, the sequence of approximants Sn (0), is An /Bn . The recurrence relations
(2.1) will be very familiar to those readers who have studied continued fractions
before. In fact, we shall not use these at all in this paper.
Many results on continued fractions are proved by considering the behaviour of
solutions of the recurrence relations (2.1). In general, one cannot solve these
equations; thus progress depends on obtaining estimates of solutions, and there
are very many papers in which this is the central issue. The following classical
theorem is proved in this way, and here the required estimates are easily obtained.
P
Theorem (The Stern-Stolz Theorem). If n |bn | converges then K(1|bn ) di-
verges.
The motivation for this result is clear, for if s(z) = 1/z, then the continued
fraction [s, s, . . .] has approximants ∞, 0, ∞, 0, . . . and so diverges. In the Stern-
Stolz Theorem, sn (z) = 1/(bn + z), and bn → 0, so that sn → s, and we might
expect that [s1 , s2 , . . .] diverges, at least if bn → 0 sufficiently quickly.
This brief discussion raises the question of what is meant by the convergence
gn → g of Möbius maps, and how one should measure the rate of this convergence.
The classical way to measure the rate at which sn → s, where sn (z) = an /(bn +z)
and s(z) = a/(b + z), is in terms of the quantities |an − a| and |bn − b|. This is
equivalent to using the norm k(z, w)k = |z|+|w| on C×C, and it is not at all clear
that this is the best measure to use. Indeed, and as we shall see later, there are
other norms that are much more closely tied to the geometric action of Möbius
maps and which generally provide much shorter and more elegant proofs of this
type of result. For example, the Möbius group M carries a natural metric σ0
with respect to which it is a topological group, and it is surely more natural to
measure the rate at which sn → s directly in terms of σ0 . Note that once again,
we prefer to distinguish clearly between Möbius maps and matrices. In any
event, if we are to measure the rate in terms of the coefficients then, because the
matrix norm is geometrically significant (whereas the individual terms |a n − a|
and |bn − b| are not), it must surely be better to pass from Möbius maps to
matrices and then use the usual matrix norm kAk on the vector space of 2 × 2
matrices. To add support to this view, we end this section with the remark that
we shall see that if sn (z) = 1/(bn + z) (as in the Stern-Stolz Theorem), then |bn |
is essentially the distance that sn moves the distinguished point in our model of
hyperbolic 3-space. The proof of the Stern-Stolz Theorem is then nothing more
than an application of the triangle inequality between hyperbolic distances, and
this proof is valid in all dimensions, and to other continued fractions. From the
matrix point of view, sn is repesented by the unimodular matrix
µ ¶
0 i
An = ,
i ibn
and the crucial fact is that kAn k is close to kIk, where I is the 2 × 2 identity
matrix. In fact, we shall see that an analogue of the Stern-Stolz Theorem is
true in all dimensions for any sequence sn of Möbius maps sn , providing only
that their matrix norms approach kIk sufficiently rapidly. It is not necessary for
the sn themselves to converge.
3. Complex Möbius maps

The extended complex plane C ∪ {∞} is denoted by C∞ and, as usual, we map
this onto the unit sphere in R3 by stereographic projection, which we denote
by ϕ. This map provides us with the chordal metric on C∞ , namely

2|z − w|
(3.1) σ(z, w) = p p = |ϕ(z) − ϕ(w)|
1 + |z|2 1 + |w|2
(with the obvious interpretation when z or w is ∞), and (C∞ , σ) is a compact
metric space. A related metric is the infinitesimal form of the chordal metric,
namely the spherical metric given by ds = 2|dz|/(1 + |z|2 ). We shall not use the
spherical metric in this paper, but we should not rule out the possibility that it
might be preferable to the chordal metric.
A complex Möbius map is a map g of the form
az + b
g(z) = , ad − bc 6= 0;
cz + d
later we shall consider Möbius maps acting in other dimensions. Each complex
Möbius map g is a homeomorphism
¡ of¢ C∞ onto itself. A map f of C∞ to itself
is a chordal isometry if σ f (z), f (w) = σ(z, w) for all z and w, and every
chordal isometry is a Möbius map (of z or z). Later we shall describe all chordal
isometries but, for the moment, we note that z 7→ 1/z is a chordal isometry.
Each Möbius map is a Lipschitz map with respect to the chordal metric. Indeed,
if g(z) = (az + b)/(cz + d), where ad − bc = 1, then
¡ ¢ ¶s
σ g(z), g(w) |g(z) − g(w)| (1 + |z|2 )(1 + |w|2 )
µ
=
σ(z, w) |z − w| (1 + |g(z)|2 )(1 + |g(w)|2 )
s
1 (1 + |z|2 )(1 + |w|2 )
=
|(cz + d)(cw + d)| (1 + |g(z)|2 )(1 + |g(w)|2 )
p p
1 + |z|2 1 + |w|2
= p p .
|az + b|2 + |cz + d|2 |aw + b|2 + |cw + d|2
Now from the Cauchy-Schwarz inequality (applied to each term) we have
|az + b|2 + |cz + d|2 ≤ (|a|2 + |b|2 + |c|2 + |d|2 )(|z|2 + 1).
We define the norm kgk of g by
kgk2 = |a|2 + |b|2 + |c|2 + |d|2
(when ad − bc = 1) so that
¡ ¢
σ g(z), g(w) 1
≥ .
σ(z, w) kgk2
We now write u = g(z), v = g(w); then, noting that kgk = kg −1 k (as ad−bc = 1),
we obtain
¡ ¢
σ g −1 (u), g −1 (v) σ(z, w)
= ¡ ¢ ≤ kgk2 = kg −1 k2 .
σ(u, v) σ g(z), g(w)
As this holds for all u and v, and all Möbius g −1 , we see (after changing the
variables again) that for all z, w and g,
¡ ¢
σ g(z), g(w)
≤ kgk2 .
σ(z, w)
The best Lipschitz constant of a Möbius map g is L(g), where
¡ ¢
σ g(z), g(w)
L(g) = sup ;
z6=w σ(z, w)
thus
(3.2) L(g) ≤ kgk2 = |a|2 + |b|2 + |c|2 + |d|2 .
In Section 5 we shall give an explicit formula for L(g) in terms of a, b, c and d
which is slightly better than (3.2). However, (3.2) is asymptotically correct as
kgk → +∞.
The chordal metric σ enables us to introduce a metric σ0 on the group M of
complex Möbius maps, namely the metric of uniform convergence on C∞ . This
metric is defined by
¡ ¢
(3.3) σ0 (f, g) = sup σ f (z), g(z) ,
z∈C∞
and by gn → g we shall always mean σ0 (gn , g) → 0; that is, the uniform con-
vergence of gn to g on C∞ . We note that σ0 satisfies the two elementary, but
extremely useful, properties
(3.4) σ0 (f h, gh) = σ0 (f, g), σ0 (hf, hg) ≤ L(h) σ0 (f, g).
The first of these says that σ0 is a right-invariant metric. The second property
is a direct consequence of the inequalities
¡ ¢ ¡ ¢
σ hf (z), hg(z) ≤ L(h) σ f (z), g(z) ≤ L(h) σ0 (f, g)
that hold for all z. We recall the following basic facts about the metric space
(M, σ0 ).
Theorem 3.1. The space (M, σ0 ) is a complete metric space, and also a topo-
logical group.
To say that (M, σ0 ) is a topological group means that the maps f 7→ f −1 and
(f, g) 7→ f g are continuous maps of M and M × M into M, respectively, and
this is easily proved. First, as we are considering a metric space, we can use
sequences to establish continuity. For any f and g we have
σ0 (f −1 , g −1 ) = σ0 (f −1 g, I) = σ0 (f −1 g, f −1 f ) ≤ L(f −1 ) σ0 (g, f ),
so that if gn → f then gn−1 → f −1 . Likewise,
σ0 (fn gn , f g) ≤ σ0 (fn gn , f gn ) + σ0 (f gn , f g) ≤ σ0 (fn , f ) + L(f ) σ0 (gn , g),
so that the map (f, g) 7→ f g is continuous. Results about the convergence of

Möbius maps are usually best viewed in these terms, and we shall see how the
simple properties (3.4) lead to a great simplification of many proofs. In general,
we should probably concentrate on σ0 and place less emphasis on arguments
about coefficients (which inevitably lead to the consideration of special cases).
It is worth mentioning that the Möbius group M supports three natural topolo-
gies, namely
(i) the topology induced by σ0 ,

(ii) the compact-open topology (the standard topology on the class of contin-
uous maps between any two given topological spaces), and
(iii) the quotient topology induced by the map Φ : SL(2, C) → M, where
SL(2, C) is given its usual norm.
As each of these three topologies arises out of a standard topological construction,

it is a relief to find that they are identical. We shall not dwell now on the
compact-open topology here (although we shall use it later), and it suffices to say
that in these circumstances, this topology is metrizable, and that the sequential
convergence of fn to f in this topology is equivalent to uniform convergence on
compact subsets of C∞ . As C∞ is compact, it follows that the topologies (i)
and (ii) are the same.
Let us return to the homomorphism Φ given in (2.4), and (as is customary)
restrict this to the homomorphism from the group SL(2, C) of 2 × 2 unimodular
matrices (matrices with determinant one) to M. In the earlier literature the
two spaces SL(2, C) and M were often rather casually identified, but there are
good reasons to distinguish between them. For example, as metric spaces they
are not isometrically equivalent, and from an algebraic point of view, there are
Möbius groups that are not isomorphic to any subgroup of SL(2, C) (for example,
a Möbius group G that contains g(z) = −z cannot be isomorphic to a matrix
group Γ because if A is any unimodular matrix that represents g then A2 = −I
so that if A ∈ Γ then so are I and −I). Nevertheless, we are familiar with the
fact that exactly two matrices in SL(2, C) correspond under Φ to a given element
of M, and that (with sufficient care) we can move freely between Möbius maps
and matrices. It is best to pause and put these ideas on a firm foundation.
The vector space of all 2 × 2 matrices (singular and non-singular) is equipped

with the norm
µ ¶
2 2 2 2 2 a b
kAk = |a| + |b| + |c| + |d| , A= ,
c d
and kABk ≤ kAk kBk for all A and B. Convergence in the metric kA − Bk
derived from this norm is the same as ‘elementwise’ convergence of matrices.
Obviously, SL(2, C) with the metric kA − Bk is a topological group. From a
group theoretic point of view, M is isomorphic (as a topological group) to the
topological quotient group SL(2, C)/{±I}. The freedom to pass easily between
unimodular matrices and Möbius maps is justified by reference to the basic topo-
logical properties of the map Φ, namely that SL(2, C) is a covering space of M
with covering map Φ.
Theorem 3.2. Let Φ : SL(2, C) → M be the map given in (2.4), where these
spaces have the norm k·k and metric σ0 , respectively. Then Φ is an open, con-
√
tinuous, surjective map, and the restriction of Φ to any open ball of radius 2
is injective (and hence a homeomorphism).
An immediate consequence of this is that the topology induced on M by the

metric σ0 coincides with the quotient topology induced by the map Φ (see [12,
p. 4]). Theorems 3.1 and 3.2 contain all of the elementary results on convergence
of maps and matrices that are often quoted from (or used implicitly in) the ear-
lier literature. In particular, if the matrices An converge to A then (because Φ is
continuous) the Möbius maps Φ(An ) converge to Φ(A) uniformly on C∞ . Con-
versely, given any Möbius f we can choose a matrix A representing
√ f , and then
let N be the open ball in SL(2, C) with centre A and radius 2. As Φ is an open
map, Φ(N ) is an open neighbourhood of f . If fn → f uniformly on C∞ , then
fn ∈ Φ(N ) for sufficiently large n and, for these n we can define matrices An by
An = Φ−1 (fn ), and then An → A.
The advantage of this approach to Möbius maps and matrices is not just that
the proofs are generally more transparent and more elegant than the proofs that
avoid the topological issues, but that we can give explicit quantitative versions of
the statements made above. In fact, we have the following two results, where here
(and elsewhere) we use I for both the identity matrix and the identity Möbius
map.
Theorem 3.3. Suppose that √ a Möbius map g is represented by a unimodular
matrix G. Then σ0 (g, I) ≤ 6kG − Ik.
Theorem 3.4. Any Möbius map g can be represented by a unimodular matrix G
such that
2 4σ0 (g, I)2
kG − Ik ≤ .
4 − σ0 (g, I)2
The importance of these results is that they provide a link between neighbour-
hoods of the identities in the groups M and SL(2, C). Theorem 3.3 occurs in
[12, p. 79], but, by using a more detailed argument, Gehring and Martin ([40,
p. 48]) have shown that
σ0 (g, I) ≤ kG − G−1 k ≤ 2kG − Ik,
the second inequality holding because for any A in SL(2, C),
kA − A−1 k ≤ kA − Ik + kI − A−1 k = 2kA − Ik.
Theorem 3.4 is [39, Theorem 4.1].
Now in any topological group, we can transfer results about a neighbourhood
of the identity to results about a neighbourhood of any other point; thus we
can extend Theorems 3.3 and 3.4 as follows. First, if F and G are unimodular
matrices, and f = Φ(F ) and g = Φ(G), then
σ0 (f, g) = σ0 (I, gf −1 ) ≤ 2kI − GF −1 k
= 2k(F − G)F −1 k ≤ 2kF − Gk kF −1 k
= 2kF − Gk kF k.
Moreover, as we can interchange F and G in this last inequality, we have proved
that if F and G are unimodular matrices, and f = Φ(F ) and g = Φ(G), then
¡ ¢
(3.5) σ0 f, g ≤ 2 min{kF k, kGk} kF − Gk.
This prompts the question: what is the smallest constant µ such that for all
unimodular matrices F and G,
(3.6) σ0 (f, g) ≤ µ min{kF k, kGk} kF − Gk ?
We have just seen that µ ≤ 2, and we shall now show that µ ≥ 1.
Let ε be any positive number, and let
1
1 1ε
µ ¶ µ ¶
1+ε ε
F = , G= 1 .
0 1 0 1+ε
Then kF k ≤ kGk, and kF k2 kF − Gk2 → 2 as ε → 0. Now let f = Φ(F ) and

g = Φ(G), and z = −1/ε. Then
¡ ¢ ¡ ¢ √
σ0 (f, g) ≥ σ f (z), g(z) = σ 0, −(1 + ε) → 2,
as ε → 0. This shows that µ ≥ 1 in (3.6).
To extend Theorem 3.4 to a neighbourhood of a point other than I, we simply
observe that
2kF kσ0 (gf −1 , I) 2kF kσ0 (g, f )
(3.7) kF − Gk ≤ kI − GF −1 k kF k ≤ p =p
4 − σ0 (gf −1 , I)2 4 − σ0 (g, f )2
Sketch of the proof of Theorem 3.2. First, (3.5) shows √ that Φ is continu-
ous. The fact that Φ is injective on any open ball of radius 2 follows directly
from the Cauchy-Schwarz inequality (see [12, p. 78]). It remains to prove that Φ
is an open map. Let U be any non-empty open subset of SL(2, C), take any
matrix F in U and let f = Φ(F ). We want to show that f is an interior point of
Φ(U ). As U is an open neighbourhood of F , there is some positive δ such that
{X ∈ SL(2, C) : kX − F k < δ} ⊂ U . It is clear from (3.7) that there is some
positive η such that if σ0 (g, f ) < η, then g can be represented by a unimodular
matrix G such that kG−F k < δ. Thus Φ(U ) contains the open ball with centre f
and radius η; hence Φ is an open map.
The key step in the proof of Theorems 3.3 and 3.4 is that the unitary matrices act
(via Möbius transformations) as isometries with respect to the chordal metric,
and also (by conjugation) as isometries with respect to the metric kA − Bk. Let
us expand on this statement. A 2 × 2 complex matrix A is unitary if and only if
AĀt = I or, equivalently, if and only if kAk2 = 2 (see [12, p. 17]), and we have
the following result.
Theorem 3.5. Suppose that A is a 2 × 2 unitary matrix. Then the conjuga-
tion X 7→ AXA−1 is an isometry of the vector space of 2 × 2 complex matrices
with respect to the metric kX − Y k. Now let f be the Möbius transformation
Φ(A); then f is a chordal isometry. In particular, for all Möbius g and h,
σ0 (f gf −1 , f hf −1 ) = σ0 (g, h), so that the conjugation g 7→ f gf −1 is an isome-
try of (M, σ0 ). Finally, every chordal isometry is of the form Φ(A) for some
unitary A.
For example, the map z 7→ 1/z is represented by the unimodular matrix A with
entries 0, i, i, 0, respectively, and this is unitary as kAk2 = 2. We remark that
the crucial hypothesis in the Stern-Stolz Theorem is not that the maps sn are of
the form an /(bn + z), nor that they are close to the map z 7→ 1/z; the conclusion
depends on the fact that the sn are Möbius maps (without restriction) that are
close to some unitary map.
This is an appropriate point to mention the following familiar result that is
closely related to Theorem 3.4.
Theorem 3.6. If a sequence gn of Möbius maps converges at three distinct points
of C∞ to three distinct values, then the sequence gn converges uniformly on C∞
to some Möbius map g.
This result is not as transparent as it might seem at first sight for it is, in fact,
the combination of two separate results. The assumption that gn converges at
three distinct points is sufficient to guarantee in all dimensions that the family
{g1 , g2 , . . .} is a normal (equicontinuous) family. Now any limit function arising
from any such family is necessarily Möbius, but the fact that the entire sequence
converges uniformly to some Möbius map is only true in dimension two because it
is only here that the values at three points are sufficient to completely determine
the Möbius map. We shall consider the higher dimensional case later. A proof
of Theorem 3.6 is given in, for example, [31], but the following proof (which
illustrates the use of the metric σ0 ) is much neater.
Proof. Suppose first that gn (w) → w as n → ∞ when w ∈ {0, 1, ∞}. We may

assume that gn (0) and gn (1) are finite and define
z − gn (0)
fn (z) = .
gn (1) − gn (0)
Now fn gn fixes 0 and 1 so that we can represent it by the matrix
µ ¶
an 0
An = , an dn = 1, an = cn + dn ,
cn d n
where (by replacing An by −An if necessary) we may assume that Re[dn ] ≥ 0.
Now d2n = 1 − (1/fn gn (∞)) → 1, so that dn → 1. Thus an → 1 and cn → 0.
Now Theorem 3.3 shows that fn gn → I and fn → I; thus (as M is a topological
group), gn = fn−1 (fn gn ) → I 2 = I. Finally, suppose that gn (zj ) → wj , j = 1, 2, 3,
for distinct z1 , z2 , z3 , and distinct w1 , w2 , w3 . Let u be the Möbius map taking
0, 1, ∞ to z1 , z2 , z3 , respectively, and let v be the Möbius map taking 0, 1, ∞
to w1 , w2 , w3 , respectively. Then v −1 gu(z) → z when z ∈ {0, 1, ∞} and so, by
the first part of the argument, gn = v(v −1 gn u)u−1 → vIu−1 = vu−1 .
In [31] DePree and Thron proved various results about the convergence of Möbius
maps and matrices; in particular, they obtained the following theorems (which
are expressed here in our notation).
P
Theorem 3.7. Let sn be a sequence of Möbius maps. If n σ0 (sn , I) converges,
then s1 · · · sn converges uniformly on C∞ to some Möbius map.
Theorem 3.8. Let An be matrices in SL(2, C) with kAn − Ik ≤ k(−An ) − Ik,
and let fn be the Möbius map corresponding to An . Then the three series
X X X
(3.8) σ0 (fn , I), kAn − Ik, kPn − Pn−1 k
n n n
converge or diverge together, where Pn = A1 · · · An .
Again we illustrate the value of the metric σ0 by giving proofs that are much
shorter than those in [31].
Proof of Theorem 3.7. As σ0 (sn , I) = σ0 (I, s−1 n ), our assumption implies that
−1
P
n σ0 (sn , I) converges. The right-invariance of σ0 now implies that
X
σ0 (s−1 −1 −1 −1
n · · · s1 , sn−1 · · · s1 )
n
converges, and hence that (s1 · · · sn )−1 is a Cauchy sequence. Thus this sequence,
and hence the sequence of its inverses, namely s1 · · · sn , converges in M. This
proof is valid in all dimensions.
Proof of Theorem 3.8. Theorems 3.3 and 3.4 implyPthat the first two series in
(3.8) converge or diverge together. Suppose now that n kPn − Pn−1 k converges.
Then Pn is a Cauchy sequence and hence bounded (as SL(2, C) is complete). As
−1
kAn − Ik = kPn−1 (Pn − Pn−1 )k ≤ kPn−1 k kPn − Pn−1 k,
P
we see that n kAn − Ik converges.
P P
Finally, suppose that n kAn − Ik converges. Then n σ0 (fn , I) converges. Let
pn = f1 · · · fn . Then, by right invariance,
σ0 (fn , I) = σ0 (fn−1 , I) = σ0 (p−1 −1
n , pn−1 );
−1
thus n σ0 (p−1
P
n , pn−1 ) converges. As (M, σ0 ) is complete, the Cauchy sequence
p−1
n converges, and as M is a topological group, pn → q, say, for some Möbius q.
Let Q be a unimodular matrix representing q; then there is a sequence εn , with
εn = ±1, such that εn Pn → Q. We deduce that kPn k ≤ M , say, and as
kPn − Pn−1 k = kPn−1 (An − I)k ≤ kPn−1 k kAn − Ik,
P
we see that n kPn − Pn−1 k converges.
One of the most important techniques in studying Möbius maps is that of taking
conjugates. Conjugation by g (that is, the map f 7→ gf g −1 ) is a homeomor-
phism of M onto itself, and most of the important properties of Möbius maps
are invariant under conjugation. For example, in many proofs it is sufficient to
replace a given Möbius map by a (much simpler) conjugate map and then com-
plete the proof for this map; specifically, it is often sufficient to only consider the
maps z 7→ z + 1 and z 7→ kz. This technique is used constantly in the literature
on discrete groups; it is acknowledged, but rarely exploited, in the literature on
continued fractions. Consider, for example, the classification of Möbius maps.
It seems customary in continued fraction theory to classify Möbius maps by the
position of their fixed points. For example, one reads that if s(z) = a/(b + z),
then s has two fixed points, say z1 and z2 , and s is parabolic if z1 = z2 , elliptic
if z1 6= z2 but |z1 | = |z2 |, and loxodromic if |z1 | 6= |z2 |. We remark that this
classification is only valid for maps of the form a/(b + z) and, more importantly,
it is not invariant under conjugation. By contrast, the classification in terms of
the trace function is applicable to all Möbius maps, it is invariant under conju-
gation, and it is simpler in that it does not require one to find the fixed points.
If we are to extend results on continued fractions to general Möbius maps then
we must work with general tools, and there is ample evidence of the importance
of the trace function from discrete group theory. In fact, this function actually
parametrizes the conjugacy classes, and its importance extends far beyond the
classification of Möbius maps (see, for example, [81]).
Definition 3.9. Given a Möbius g we define trace2 (g) to be [trace(A)]2 , where A

is any one of the (two) unimodular matrices that represent g; thus if
az + b
g(z) = , ad − bc = 1,
cz + d
then trace2 (g) = (a + d)2 . Now suppose that g is not the identity map. Then
(1) g is parabolic if and only if (a + d)2 = 4;
(2) g is elliptic if and only if (a + d)2 ∈ [0, 4);
(3) g is loxodromic if and only if (a + d)2 ∈ / [0, 4].
Obviously, M is the disjoint union of {I} and the three classes of parabolic,
elliptic and loxodromic maps. The following well known result illustrates the
importance of the trace function.
Theorem 3.10. Let f and g be two Möbius maps neither of which is the identity.
Then f and g are conjugate in M if and only if trace2 (f ) = trace2 (g).
The class of loxodromic maps divides into two subclasses: a loxodromic map g (as
above) is hyperbolic if and only if (a + d)2 ∈ (4, +∞), and it is strictly loxodromic
if and only if (a + d)2 ∈ / [0, +∞). The set of hyperbolic maps is an important
subclass of the set of loxodromic maps because a loxodromic map g is hyperbolic
if and only if it has an invariant disc D (that is, such that g(D) = D), and also
if and only if it can be written as the composition of exactly two inversions (most
Möbius maps require four inversions). Now maps of the form s(z) = 1/(b + z),
where b 6= 0, are of interest in number theory and continued fraction theory. As
(trace)2 (s) = −b2 we see that if b is real then s is loxodromic but not hyperbolic,
so that s has no invariant disc in C∞ . However, as we shall see later, s does
have an invariant disc in hyperbolic 3-space, and that this disc is a model of the
hyperbolic plane that happens to lie in this higher dimensional space rather than
in C∞ .
Certain questions about the convergence and stability of Möbius maps can be
settled for loxodromic maps, but not for elliptic and parabolic maps, and the
reason for this is that the set L of loxodromic maps is an open subset of M. If
we make a sufficiently small perturbation of a loxodromic map the resulting map
is loxodromic; a sufficiently small perturbation of an elliptic or parabolic map
can result in an elliptic or a loxodromic map. To see that L is an open subset
of M we consider Φ : SL(2, C) → M¡ and the¢ map τ : SL(2, C) → C given by
τ (A) = [trace(A)]2 . Then L = Φτ −1 C\[0, 4] , and as τ is continuous, and Φ
is an open map, L is an open subset of M. Note also that (in some sense) the
generic Möbius map is loxodromic.
It is easy to characterize the classification given above in terms of the geometric
action of the maps, and the following is standard. A Möbius map g is
(1) parabolic if it is conjugate to z 7→ z + 1;
(2) elliptic if it is conjugate to z 7→ kz for some k with |k| = 1, k 6= 1;

(3) hyperbolic if it is conjugate to z 7→ kz for some k with k > 0 and k 6= 1;
(4) loxodromic if it is conjugate to z 7→ kz for some k with |k| 6= 0, 1.
The dynamics of the iterates of a Möbius map g are now obvious, and require no
calculations. If g is parabolic, then there is some h such that hgh−1 (z) = z + 1,
and as (hgh−1 )n = hg n h−1 , we see that g n converges to its unique fixed point ζ,
the convergence being pointwise on C∞ and uniform on compact subsets of the
complement of {ζ}. If g is loxodromic it has an attracting fixed point α, and a
repelling fixed point β. Then there is some h such that hgh−1 (z) = kz, where
0 < |k| < 1, so we see that g n converges locally uniformly to α on the complement
of {β}. Of course, g n (β) = β for all n. Finally, if g is elliptic, then there is some h
such that (hgh−1 )(z) = kz, where |k| = 1 but k 6= 1, so that then g n (z) diverges
except when z is one of the two fixed points of g. This discussion illustrates the
efficient use of conjugation, and it also contains (as a special case) a proof of the
standard result on periodic continued fractions. If s(z) = a/(b + z) then 0 is not
a fixed points of s so sn (0) converges if and only if s is not elliptic. Thus we have
Theorem 3.11. The periodic continued fraction generated by s(z) = a/(b + z)
converges if and only if −b2 /a ∈
/ [0, 4).
Of course, this result is well known, but it often stated in a less concise way
(in terms of the fixed points of s). Moreover, its history provides evidence of
a lack of communication between different fields within this general area. Stolz
is sometimes given credit for first analyzing periodic continued fractions; see,
for example, [51, Chapter 3] where some of the history of this topic is given.
Also, see [53], where a reference is made to a proof by Stolz in 1886, an im-
proved proof by Pringsheim in 1900, and a shorter proof by Perron in 1905 “in
which the ideas related to the transformation point of view were virtually su-
pressed”. It is interesting to compare this with the state of discrete groups at
this time. In 1880 Poincaré submitted an eighty page paper on Fuchsian groups
and hyperbolic geometry to the French Academy, Klein’s Erlangen Program was
announced in 1872, and by 1887 Schottky was already studying the geometry
of free products of cyclic groups of Möbius maps. These works may not have
contained specific references to continued fractions, but they went far beyond the
study of cyclic groups and it is certain that the essential content of the results
about periodic continued fractions were very well known long before 1886. Even
more surprisingly, credit for a description of periodic continued fractions purely
in terms of Möbius maps is sometimes given to Schwerdtfeger ([72]) as late as
1946. In the light of the interest in complex dynamics and Julia sets in recent
years, we should perhaps note that [72] contains an incorrect definition of an
attracting fixed point, and a criticism of Julia’s definition of an indifferent fixed
point.
We end this section by analyzing Möbius maps in terms of inversions across circles
and lines in the plane. This can be done in a canonical way using isometric circles
which were introduced by L. R. Ford in 1927. These give a particularly simple
decomposition of a Möbius map g into inversions, and use only the smallest
number of inversions that are needed (two or four). Consider the Euclidean
distortion of the map
az + b
g(z) = , ad − bc = 1, c 6= 0.
cz + d
The distortion at z is |g 0 (z)| or, equivalently, |cz + d|−2 . According to Ford, the
isometric circle of g is the set of points at which this Euclidean distortion is 1:
thus the isometric circle of g is given by |cz + d| = 1 (providing that ad − bc = 1)
and this is indeed a circle. We denote this circle by I(g); its centre is g −1 (∞)
and its radius is 1/|c|. What is the image of I(g) under g? It must be a circle,
say C, and as the restriction of g to I(g) is a Euclidean isometry, the restriction
of g −1 to C must also be a Euclidean isometry. We conclude that C is the circle
I(g −1 ); this has centre g(∞) and the same radius as I(g). In short, g maps I(g)
given by |cz + d| = 1 onto I(g −1 ) which is given by |cz − a| = 1.
Now let α denote inversion in I(g), and let β denote reflection across the Eu-
clidean bisector of the points g(∞) and g −1 (∞) (henceforth, we shall use the
words ‘reflection’ and ‘inversion’ interchangeably). Then both g and βα are
Möbius maps that map I(g) onto I(g −1 ) and ∞ onto the centre of I(g −1 ). It
follows that g(βα)−1 leaves I(g −1 ) invariant, and as it fixes ∞, it also fixes the
centre of I(g −1 ) (the inverse point to ∞). Thus g(βα)−1 is some rotation γ
about g(∞). This shows that g = γβα, and hence that every Möbius map can be
expressed as the composition of at most four inversions (namely α, β, and two
reflections for γ). If the rotation γ is trivial then only two inversions are needed.
The classification of a Möbius map g is partly (but not entirely) reflected in
the relative position of the isometric circles I(g) and I(g −1 ). If g is parabolic,
then these two circles are tangent; if g is elliptic, these two circles cross; if g is
hyperbolic, these two circles are exterior to each other. In each of these cases, g
is the inversion in I(g) followed by the reflection in the Euclidean bisector of the
points g −1 (∞) and g(∞) (and the rotation γ is trivial). If g is strictly loxodromic
then there is no restriction on the location of the circles I(g) and I(g −1 ), and g is
the composition of the two inversions described above followed by a non-trivial
rotation about the point g(∞) by an angle −2 arg(a + d) (see [38, p. 28]).
4. The stability of continued fractions

In this section we shall consider the stability of the continued fraction K(a n |bn )
with respect to small changes in the coefficients an and bn . We shall take two
results of this kind from the theory of continued fractions and show that the first
is a special case of a result about topological groups, while the second is a special
case of a result about analytic functions. Despite this increased generality, our
arguments are simpler (and shorter) than those given for continued fractions,
and they enable us to describe the stability in terms of explicit inequalities. We
begin with the following result ([56, Theorem 2.2]).
Theorem 4.1. Given a continued fraction K(an |bn ), there exist positive numbers
r1 , r2 , . . . and s1 , s2 , . . . such that if |αn − an | < rn and |βn − bn | < sn for all n,
then the continued fractions K(αn |βn ) and K(an |bn ) are both convergent or both
divergent.
We shall show that this is a special case of the following result on topological
groups, and this generalization implies that the analogous result to Theorem 4.1
holds for any inner composition sequence of Möbius maps in any dimension.
Theorem 4.2. Suppose that G is a topological group whose topology is derived
from a right-invariant metric σ0 , and that (G, σ0 ) is complete. Let f1 , f2 , . . . be
any sequence of elements of G. Then, for each k, there is a neighbourhood Nk of
fk such that if, for all j, gj ∈ Nj , then (g1 · · · gn )(f1 · · · fn )−1 converges to some
element h of G.
Proof. We write Fn = f1 · · · fn , Gn = g1 · · · gn , and hn = Gn Fn−1 . For any k,
σ0 (h−1 −1 −1 −1
k , hk+1 ) = σ0 (hk Gk+1 , hk+1 Gk+1 ) = σ0 (Fk gk+1 , Fk fk+1 ),
so that if m < n,
n−1
X n−1
X
σ0 (h−1 −1
m , hn ) ≤ σ0 (h−1 −1
k , hk+1 ) = σ0 (Fk gk+1 , Fk fk+1 ).
k=m k=m
Now let θk : G → G be the map defined as left multiplication by Fk . Then θk
is continuous, so that if ∆k+1 denotes the open ball with centre θk (fk+1 ) and
radius 1/2k+1 , then ¡there is an open neighbourhood, say Nk+1 , of fk+1 such that
if g ∈ Nk+1 then σ0 θk (g), θk (fk+1 ) < 1/2k+1 , that is, σ0 (Fk g, Fk fk+1 ) < 1/2k+1 .
It follows that if gk+1 ∈ Nk+1 for all k, then h−1 n is a Cauchy sequence and so
−1
converges. As x 7→ x is continuous, the sequence hn also converges.
As the metric σ0 of uniform convergence on C∞ is a right invariant metric on the
Möbius group M, and as (M, σ0 ) is complete, we can apply Theorem 4.2 to M.
Now (by right invariance) for each z,
¡ ¢
σ Gn (z), hFn (z) ≤ σ0 (Gn , hFn ) = σ0 (hn , h),
so we obtain the next result (which contains Theorem 4.1).
Corollary 4.3. Let f1 , f2 , . . . be any sequence of Möbius maps. Then, for each k,
there is a neighbourhood Nk of fk such that if¡ gj ∈ Nj , j = 1, 2, . . ., then
¢ there
is some Möbius map h such that for all z, σ g1 · · · gn (z), hf1 · · · fn (z) → 0 as
n → ∞. In particular, for each point z, limn g1 · · · gn (z) exists if and only if
limn f1 · · · fn (z) exists.
We remark that the argument given here extends to Möbius maps acting in any
dimension. Also, even though it is not needed for our application to Möbius
maps, the following remark may be of interest. A fundamental system of neigh-
bourhoods of the identity e in a topological group G is a sequence of neighbour-
hoods U1 , U2 , . . . of e such that
(i) ∩n Un = {e}, and
(ii) for any neighbourhood V of e, Un ⊂ V for some n.
If such a system exists, then the topology on G is derived from a right invariant
metric on G (see [32, p. 39]).
Theorem 4.1 was used in [56] to prove that if Ω is a simple convergence set for
continued fractions, then so is its closure Ω, and it is clear that Corollary 4.3 can
also be used in this way. Also, Theorem 4.1 extends earlier results in [50], where
estimates of rn and sn were given for certain classes of continued fractions. No
explicit estimates were obtained in [56]. However, despite the generality of our
ideas, we can give explicit examples of the neighbourhoods Nk in the case of all
complex Möbius maps. Using (3.2) and (3.5) (and temporarily using norms of
Möbius maps when we really mean norms of their matrix representations) we
have
σ0 (Fk v, Fk fk+1 ) ≤ L(Fk )σ0 (v, fk+1 )
≤ L(f1 ) · · · L(fk ) 2 kfk+1 k kv − fk+1 k
< 2kf1 k2 · · · kfk k2 kfk+1 k kv − fk+1 k.
Thus, in the case of complex Möbius maps, we may take Nk to be the set of
Möbius g that satisfy
1
kg − fk k < k+2 .
2 kf1 k · · · kfk−1 k2 kfk k
2
It is possible to obtain estimates of the rn and sn in Theorem 4.1 from this, but
we suggest that the Nk are the more natural neighbourhoods to use.
We turn now to discuss the second result on the stability of continued fractions.
By Theorem 3.11, the periodic continued fraction [s, s, . . .] converges when s
is loxodromic. The continued fraction [s1 , s2 , . . .] is a limit periodic continued
fraction of loxodromic type if sn → s for some loxodromic s, and it is natural
to expect that any such continued fraction should also converge. This is true
and for a discussion of this, and of Perron’s contribution to this topic, see [51,
p. 108]. However, this result has nothing to do with continued fractions, or
even with Möbius maps (as we shall see later). First, Mandell and Magnus [61]
took this result out of the context of continued fractions by showing that if s
is loxodromic with repelling fixed point β, and if sn → s uniformly on C∞ , then
s1 · · · sn converges to a constant on C∞ \{β}. It is not necessary to assume that
the sn are loxodromic here (as is done in [61]) because the set L of loxodromics
is an open subset of M, so that sn must be loxodromic for all sufficiently large n.
In fact, because L is open, the real issue here is one of stability, not convergence,
and we have the following result.
Theorem 4.4. Suppose that s is loxodromic with repelling fixed point β. Then
there is a neighbourhood N of s such that if sn ∈ N for all n, then s1 · · · sn
converges to a constant on C∞ \{β}.
Example 4.5. It is easy to see that in Theorem 4.4, the sequence s1 · · · sn (β)
may, but need not, converge. It will certainly converge if we take sn = s for all
n, and we now give an example in which this sequence is dense in C∞ . First,
we write the set of numbers (−1)m /m + (−1)n i/n, n, m = 1, 2, . . ., as a sequence
w1 , w2 , . . . in such a way that the partial sums of the series w1 + w2 + · · · is dense
in C∞ . Now let sn (z) = 2z + wn /2n−1 , and s(z) = 2z. Then sn → s uniformly
on C∞ (because wn → 0), and it is easy to see (by induction) that
√ ¡
s1 · · · sn (z) = 2n z + 2 w1 + · · · + wn ).
This implies that the sequence s1 · · · sn (0) is dense in C∞ .
Theorem 4.4 has probably been known for a long time, but in any event it is a
straightforward corollary of a result in function theory. The crucial fact is that
if α and β are the attracting and repelling fixed points of s, respectively, and
if D = {z : |z − α| < k|z − β|}, where k > 0, then s(D) is a compact subset
of D. Now let Ω be a subdomain of D such that s(D) ⊂ Ω ⊂ Ω ⊂ D. If g is any
Möbius map such that σ0 (g, s) is less than the chordal distance between s(D)
and ∂Ω, then g(D) ⊂ Ω ⊂ Ω and Theorem 4.4 follows directly from the next
(known) result.
Theorem 4.6. Suppose that K is a compact subset of a simply connected subdo-
main D of C∞ , and that f1 , f2 , . . . are analytic maps of D into K. Then f1 · · · fn
converges locally uniformly on D to some constant.
The essential feature here is the existence of a compact subset K of a domain D,
and a family of analytic maps of D into K. As we have already mentioned,
this implies the existence of a relatively compact subdomain Ω of D that con-
tains K, and this is precisely the situation that is used in the construction of
the compact-open topology on function spaces. This observation leads to the
following very general description of the phenomenon that is illustrated by limit
periodic loxodromic continued fractions.
Let X and Y be topological spaces, and let C(X, Y ) be the space of continuous
maps f : X → Y . Then the compact-open topology on C(X, Y ) (which is also
known as the topology of locally uniform convergence on X) is the topology
generated by the sets
[E, U ] = {g ∈ C(X, Y ) : g(E) ⊂ U },
where E is a compact subset of X, and U is an open subset of Y . Now let D be
a domain and consider any analytic f : D → D with the property that f (D) lies
in a given compact subset K of D. Next, construct subdomains Ω1 and Ω2 of D

such that
K ⊂ Ω1 ⊂ Ω1 ⊂ Ω2 ⊂ Ω2 ⊂ D,
and let N = [Ω2 , Ω1 ]. Then N is an open set in the compact-open topology on
the class of analytic maps from D into itself, and as f (Ω2 ) ⊂ f (D) ⊂ K ⊂ Ω1 , we
see that N is an open neighbourhood of f in this topology. Clearly, if g ∈ N then
g(Ω2 ) ⊂ Ω1 . Suppose now that we can remove the hypothesis that D is simply
connected from Theorem 4.6 (and we can), and suppose that f1 , f2 , . . . are in N .
Then, by this more general form of Theorem 4.6, we see that f1 · · · fn converges
locally uniformly on Ω2 to some constant ζ, say. If {f1 , f2 , . . .} is a normal
family in D then f1 · · · fn converges to ζ locally uniformly in D. This proves the
following substantial generalization of the stability result for loxodromic periodic
continued fractions.
Theorem 4.7. Let D be a subdomain of C∞ such that C∞ \D contains at least
three points, and suppose that f is an analytic map of D into itself such that
f (D) lies in a compact subset K of D. Then there is a neighbourhood N of f (in
the compact-open topology on the space of analytic maps of D into itself ) such
that if f1 , f2 , . . . are in N , then f1 · · · fn converges locally uniformly in D to a
constant.
The proof of Theorem 4.7 depends on removing the hypothesis of simple connec-
tivity from Theorem 4.6, so we need to examine the proof of Theorem 4.6. Now
proofs of Theorem 4.6 are given in, for example, [10], [46] and [54], but these do
not seem to generalize to give Theorem 4.7. In fact, the real reason why Theo-
rem 4.6 is true is that D supports a hyperbolic metric and as this has nothing to
with the connectivity of D, a proof based on this should yield Theorem 4.7. To
illustrate the main ideas here we shall give a proof of Theorem 4.6 in the case
when D is a disc; this proof uses only the hyperbolic metric and the standard
Contraction Mapping Theorem.
Proof of Theorem 4.6 when D is a disc. We may take D to be the open

unit disc D, and K to be the closed disc {z : |z| ≤ k}. Let Ω be the disc
{z : |z| < `}, where 0 ≤ k < ` < 1. The hyperbolic densities on D and Ω are
2 2`
λD (z) = , λΩ (z) = ,
1 − |z|2 `2 − |z|2
respectively, and the hyperbolic distances, say ρD and ρΩ , are obtained in the
usual way by minimizing the integrals of the densities over paths. The crucial
fact is that for all z in Ω, λD (z) ≤ `λΩ (z).
Now suppose that f is analytic in D, and that f (D) ⊂ K, and construct a
domain Ω with f (D) ⊂ Ω ⊂ Ω ⊂ D. Note that as Ω carries a hyperbolic metric
but K does not, we must focus on Ω, and not K. Now take any two points z
and w in D, and let γ be the geodesic that joins f (z) to f (w) in the metric space
(Ω, ρΩ ). Then
Z Z
¡ ¢ ¡ ¢
ρD f (z), f (w) ≤ λD (z) |dz| ≤ ` λΩ (z) |dz| = ` ρΩ f (z), f (w) .
γ γ
Now the general form of the Schwarz-Pick Lemma says that as f : (D, ρD ) →
(Ω, ρΩ ) is analytic, it is contracting in the two hyperbolic metrics; thus, for any
analytic map f : D → Ω, we have
¡ ¢ ¡ ¢
ρD f (z), f (w) ≤ ` ρΩ f (z), f (w) ≤ ` ρD (z, w).
As 0 < ` < 1, this shows that f acts on (D, ρ) as a strict contraction in the sense
of the familiar Contraction Mapping Theorem. Note that the scaling factor `
depends on D and Ω, but not on f ; thus the set of analytic maps f : D → Ω
satisfy a uniform contraction property with respect to the hyperbolic metric ρ D .
Now take any sequence f1 , f2 , . . . of analytic maps from D to Ω. Then, exactly
as in the standard proof of the Contraction Mapping Theorem, we have
ρD f1 · · · fn (z), f1 · · · fn fn+1 (z) ≤ `n ρD z, fn+1 (z) .
¡ ¢ ¡ ¢
As fn+1 (z) lies in K, and K is compact, the term ρD (z, fn+1 (z)) is (for each z)
bounded as a function of n, and the usual argument shows that f1 · · · fn (z) is a
Cauchy sequence and so converges to some value, say g(z). Finally, g is constant
because
ρD f1 · · · fn (z1 ), f1 · · · fn (z2 ) ≤ k n ρD (z1 , z2 ) → 0
¡ ¢
as n → ∞ so that g(z1 ) = g(z2 ).
We have given Theorem 4.6 in the form that it is usually quoted, but this ar-
gument suggests that it would be more revealing to state it in the following
equivalent form: if Ω is a relatively compact subdomain of D, and if each f n
is an analytic map of D into Ω, then f1 · · · fn converges to a constant on D.
It is clear that essentially the same proof will work if we now assume that we
are dealing with a family of analytic maps f : D → Ω, where the subdomain Ω
has the property that all such maps f satisfy a uniform Lipschitz condition with
respect to the hyperbolic metric on D (notice that there is no compact set in this
formulation). In order to apply this general proof we need to characterize those
domains Ω that have this uniform Lipschitz property. This is done in [17], where
the generalization of Theorem 4.6 that is used above is proved.
We end this section with an example that illustrates how the ideas described
above can be used to give an explicit construction of the neighbourhood N of
a loxodromic Möbius map s that occurs in Theorem 4.4. It is not necessary to
insist that s(∞) = 0 here but in order to illustrate the effectiveness of argu-
ments involving conjugation even when considering the special case of continued
fractions we shall take s(∞) = 0.
√
Example 4.8. Let s(z) = 1/(z + 3i/ 2). Then s is loxodromic √ with attracting
√
fixed point α and repelling fixed point β, where α = −i/ 2 and β = −i 2.
Thus
s(z) − α z−α
=k ,
s(z) − β z−β
where k = s0 (α) = 1/2. Now let h(z) = (z − α)/(z − β), and let s̃ = hsh−1 . Thus
s̃(z) = 1/2z.
Now s̃ maps D into {z : |z| ≤ 1/2}. As σ(1/2, 1) = 2/5 > 1/2, we can define
p
a number R in (1/2, 1) by σ(1/2, R) = 1/2, and then let Q = {z : |z| ≤ R}. It

follows that if σ0 (g, s̃) < 1/2 then g(D) ⊂ Q.
Now suppose that σ0 (fj , s) < 1/(2khk2 ) for j = 1, 2, . . ., and let f˜j = hfj h−1 .
Then
1
σ0 (f˜j , s̃) = σ0 (hfj h−1 , hsh−1 ) = σ0 (hfj , hs) ≤ L(h)σ0 (fj , s) < ,
2
2 ˜ ˜
because L(h) ≤ khk . It follows that f1 · · · fn converges locally uniformly to
a constant on D, and hence that f1 · · · fn converges uniformly to a constant
on some neighbourhood of α. By using normal families, we see that f1 · · · fn
actually converges
√ locally √uniformly on C∞ \{β} to some constant. Finally, √ as
α = −i/ 2 and β = −i 2, a computation shows that khk2 = 9/ 2; thus
f1 · · · fn converges to a constant on the complement of β providing that σ0 (fj , s) <
√
2/18. √ Theorem 3.3 could now be used to find conditions on |an − 1| and
|bn − 3i/ 2|, where fn (z) = an /(bn + z), for f1 · · · fn to converge to a constant,
but (once again) this will lead to a weaker result.
5. Möbius maps and hyperbolic geometry

Although we shall assume that the reader is familiar with a modest amount of
plane hyperbolic geometry, we shall begin with a brief description of this (see,
for example, [12], [23], [34] and [63]). We take the upper half-plane H (given by
y > 0, where z = x + iy) to be the hyperbolic plane. The hyperbolic metric ρ
on H is derived from the line element ds = |dz|/ Im[z], and the Möbius maps
that map H onto itself are the conformal isometries of H. The geodesics are the
semi-circles in H whose centres lie in R, together with the ‘vertical’ half-lines
whose initial point lies in R. An alternative model of the hyperbolic plane is the
open unit disc D with the hyperbolic metric (which we also denote by ρ) derived
from the line element ds = 2|dz|/(1 − |z|2 ). Here the geodesics are the arcs of
circles that are orthogonal to ∂D, together with all diameters of D. The map
z 7→ (z − i)/(z + i) is an isometry from (H, ρ) to (D, ρ).
Without going into details here, the reader should note that there is an exten-
sive collection of trigonometric formulae available for the hyperbolic plane, and
that it is never necessary (or desirable) to give Euclidean arguments in hyper-
bolic geometry. We illustrate this with just two formulae. the first of which is
Pythagoras’ Theorem: if a right-angled triangle has hypotenuse of length c, and

other sides of lengths a and b, then cosh a cosh b = cosh c. The second result is
concerned with the situation in which exactly one vertex of a hyperbolic triangle
is on the boundary of hyperbolic space. In this case the angle at this vertex is
zero, and the two sides ending at this vertex have infinite length. This classical
result is known as the angle of parallelism: if a hyperbolic triangle has angles 0,
π/2, θ, and sides of lengths d, +∞ and +∞, then
(5.1) cosh d sin θ = 1.
For more details of hyperbolic trigonometry see, for example, [12], [23] and [70].
We shall now describe how the action of a Möbius map on C∞ extends (geomet-
rically and algebraically) to R3 ∪ {∞} so as to act as an isometry of hyperbolic
three-space. For the moment we take hyperbolic three-space to be the upper half
H3 = {(x1 , x2 , x3 ) ∈ R3 : x3 > 0}
of Euclidean 3-space with the hyperbolic metric ρ given by the line element
|dx|
ds = , x = (x1 , x2 , x3 ).
x3
This is a direct generalization of the model H of the hyperbolic plane. A geodesic
in H3 is a semi-circle whose centre lies in C and which is orthogonal to C, or that
part of a ‘vertical’ line that lies in H3 . The hyperbolic planes are the hemispheres
in H3 whose centres lie in C, together with those parts of the ‘vertical’ planes
that lie in H3 . Notice that a circle in C∞ is the boundary of a hyperbolic plane
in H3 , and conversely, the boundary of a hyperbolic plane is a circle (as usual,
we treat lines as circles). In particular, we will be able to see the effect of Möbius
maps on circles in C∞ as the effect of applying isometries on planes in H3 .
Each Möbius map g extends naturally to act on R3 ∪ {∞} in the following way.
First, we express the action of g on C∞ as a composition α1 · · · αm , where αj is
the inversion across a circle Cj in C∞ . Each Cj is part of a unique sphere (or
plane) Σj in R3 that is orthogonal to C, and we now regard αj as the inversion
across Σj . In this way every Möbius map extends to a bijection of R3 ∪ {∞} onto
itself, and this extension leaves both C∞ and H3 invariant. As each Möbius map
can be expressed as a composition of inversions in many ways, it is necessary to
prove that this extension to H3 is independent of which inversions are used. This
can be done; thus the action of g on H3 is uniquely determined by the action
of g on C∞ .
There is another, algebraic, way to carry out this extension of g. First, we identify
the point (x1 , x2 , x3 ) in R3 with the quaternion x1 + ix2 + jx3 + 0k, where 1, i, j
and k are the usual basis for quaternions. It is simpler to write x1 + ix2 as z
(a complex number) and then to consider the points of H3 as the points z + tj,
where t > 0. We can now define the action of g on H3 by the formula
¡ ¢¡ ¢−1
g(z + tj) = a[z + tj] + b c[z + tj] + d ,
where this calculation is carried out in the algebra of quaternions. It is not

immediately obvious that the resulting answer will have zero k-component and
positive j-component, but it does, and after a little more work one can see that
this action agrees with the geometric action described in the previous paragraph.
Explicitly, if g(z) = (az + b)/(cz + d), where ad − bc 6= 0, then the action of g on
H3 is given by
(az + b)(cz + d) + act2 + |ad − bc|tj
(5.2) g(z + tj) =
|cz + d|2 + |c|2 t2
(see [12, p. 58] for details, but see also [33]). The advantage that stems from
this construction is that the complex Möbius maps now act as isometries of a
complete metric space (which is naturally embedded in a compact metric space).
The following result, then, is fundamental.
Theorem 5.1. The Möbius group M is the group of all conformal orientation-
preserving isometries of H3 .
Given this, the most penetrating way to study the Möbius group must surely
be to study its action on H3 even if one is only interested in the action on C∞ .
Indeed, the action of the Möbius group on C∞ is the action on the boundary of
H3 , and much information is lost if we confine our attention to the action on the
boundary.
The two most familiar geometric properties of complex Möbius maps are that
they preserve circles, and inverse points. From the point of view of three-
dimensional geometry, these properties are obvious. A hyperbolic plane is the set
of points equidistant from two given points; thus each Möbius map g maps each
hyperbolic plane into a hyperbolic plane. As the boundary of a hyperbolic plane
is a circle, g must map circles to circles. Two points z and w are inverse points
with respect to a circle C if and only if the geodesic with end-points z and w is
orthogonal to the hyperbolic plane bounded by C. Because this orthogonality
(in H3 ) is preserved under Möbius maps, so too is the property of being inverse
points. Finally, the modulus of the ring domain (familiar to complex analysts)
bounded by two circles is a simple function of the hyperbolic distance (in H3 )
between the two hyperbolic planes bounded by these circles.
There is another reason why it is useful to consider H3 . A Möbius map g may
have an invariant circle, say C, in C∞ but no invariant disc (for the two half-
planes determined by C may be interchanged by g). If g has an invariant disc,
say D, then we can take D to be a model of the hyperbolic plane and g is then
an isometry of this hyperbolic plane. If g has an invariant circle C, but no
invariant disc, then g preserves, and hence acts as an isometry of, the hyperbolic
plane in H3 whose boundary is C. Thus we can now say that in all cases, if g
has an invariant circle, then g is the isometry of some hyperbolic plane. For
example, suppose that g(z) = 1/(b + z), where b > 0. In this case g has an
invariant circle, namely R ∪ {∞}, but no invariant disc in C∞ . However, in its
action in H3 , g leaves the hyperbolic plane bounded by R ∪ {∞} invariant so

we can indeed regard g as an isometry of a hyperbolic plane. In this case the
invariant hyperbolic plane is that part of H3 that lies above the real axis, and
this is indistinguishable from the usual upper half-plane model of the hyperbolic
plane. Indeed, in terms of co-ordinates, it is simply the set of points x+tj, where
j 2 = −1, and so we see that the algebra on this plane is identical to that in the
model x + iy, where i2 = −1. Thus we see that there might be some advantages
to be gained by studying the maps z 7→ 1/(z + b), where b > 0, by their action
on the vertical half-plane lying above R.
We shall now consider the geometry associated with the norm kgk of a Möbius
map g. There are two basic results about the norm of a Möbius map, and these
confirm that there is a strong and inescapable link between the action of a Möbius
map in C∞ and its action in H3 . In these results (and in our later work) the
point (0, 0, 1) in H3 plays a vital role. To emphasize this, we shall from now on,
write
j = (0, 0, 1)
(in boldface type) instead of the quaternion j. We now give a formal statement
of the two results (which can be found in, for example, [12]).
Theorem 5.2. Let g be any complex Möbius map, and let I be the identity map.
Then
1 ¡
kgk2 = kIk2 + 4 sinh2 ρ j, g(j) .
¢
(5.3)
2
¡ ¢
As kIk2 = 2 we can also write (5.3) in the form kgk2 = 2 cosh ρ j, g(j) .
Theorem 5.3. Let g be any complex Möbius map. Then the best Lipschitz
constant (relative the chordal metric) for g is given by
£ ¤
(5.4) L(g) = exp ρ(j, g(j) .
Combining (5.3) and (5.4) we obtain (as promised earlier) a formula for L(g) in
terms of the coefficients of g, namely
1¡ p
L(g) = kgk2 + kgk4 − 4 ;
¢
(5.5)
2
in particular,
1
(5.6) kgk2 ≤ L(g) ≤ kgk2 ,
2
2 −1
¡ ¢
¡ −1L(g) ¢∼ kgk as kgk → +∞. Next, as g is an isometry, ρ j, g(j) =
and
ρ g (j), j , so we see from (5.4) that, for all g,
(5.7) L(g −1 ) = L(g) ≥ 1.
¡ ¢
The situation in which the terms kgk, L(g) and ρ j, g(j) simultaneously take
their minimum values is of interest.
Theorem 5.4. For any Möbius g the following are equivalent:

(a) kgk2 = 2;
(b) L(g)
¡ = 1; ¢
(c) ρ j, g(j) = 0;
(d) g is a chordal isometry.
Proof. It is easy to see that (5.3) and (5.4) show that (a), (b) and (c) are
equivalent. Further it is clear that (d) implies (b). Now suppose that (b) holds.
Then, equality holds in (5.7)) and we have
σ g(z), g(w) ≤ L(g) σ(z, w) = σ(z, w) = σ g −1 g(z), g −1 g(w)
¡ ¢ ¡ ¢
≤ L(g −1 ) σ g(z), g(w) = σ g(z), g(w) .

¡ ¢ ¡ ¢
It is now clear that (d) follows.
It is useful to have the ball-model of hyperbolic three-space available, and this is

the open unit ball B3 in R3 with the metric (which we also denote by ρ) given by
ds = 2|dx|/(1 − |x|2 ). Just as in the two-dimensional case there is a Möbius map
of H3 onto B3 that maps j to 0 (see [1] for a quaternionic description of this).
This map is the reflection α1 in the√complex plane followed by inversion α2 in
the sphere with centre j and radius 2 (see Section 8). In fact,
(a) α2 α1 , is an isometry from (H3 , ρ) to (B3 , ρ), and
(b) the restriction of α2 α1 to C∞ is the stereographic projection ϕ of C∞ .
As a consequence of these facts we see that a Möbius map g is unitary (that is,
it is represented by a unitary matrix) if and only if its corresponding action on
∂B3 , namely ϕgϕ−1 , is a rotation of the unit sphere (and of R3 ). One of the basic
differences between Möbius maps in various dimensions is that the stabilizer of a
point in the hyperbolic plane is abelian, whereas the stabilizer of a point in higher
dimensions is non-abelian. Clearly this will have an impact on any attempt to
generalize results about elliptic maps (for example, the Stern-Stolz Theorem)
from continued fraction theory to higher dimensions.
The last result in this section gives information on the Lipschitz constants of two
maps that are close together. Suppose that gn → g. Then, by (3.7), kgn k → kgk,
and it is now clear from (5.5) that L(gn ) → L(g). We remark that this cannot be
entirely trivial for in the case of continuous maps of R into itself the corresponding
result is false. The following lemma gives a weak version of this result which we
shall use later.
√
Lemma 5.5. Suppose that the Möbius maps f and g satisfy σ0 (f, g) < 2. Then
L(g) ≤ 18 L(f ).
Proof. We recall that kgk = kGk, where G is any unimodular matrix represen-
tation of g. Now (3.7) implies that (for suitable choices of the signs of F and G)
√
if σ0 (f, g) < 2, then kG − F k ≤ 2kF k, so that kGk ≤ kG − F k + kF k ≤ 3kF k.
The inequality (5.6) now gives
L(g) ≤ kgk2 = kGk2 ≤ 9kF k2 = 9kf k2 ≤ 18L(f )
as required.
We end this section with a brief sketch of a beautiful application of these ideas;
this is related to our generalization of continued fractions in as far as it describes
all finite groups of unitary maps. In this paragraph, a Möbius transformation
shall be any composition sequence of inversions in spheres (or planes) in R3 ∪{∞}.
Consider any finite Möbius group G. The orbit of the point j is a finite set E
in H3 , and it is not difficult to see that there is therefore a unique smallest closed
hyperbolic ball B that contains E. As E is invariant under G, so is B, and
hence so is the centre ζ of B. We can now map H3 onto B3 by a Möbius map h,
say, with h(ζ) = 0. Thus hGh−1 is a group of Möbius maps each element of
which fixes the origin and leaves B3 invariant. However, the only Möbius maps
that do this are the orthogonal maps (given by orthogonal matrices); thus we
have shown that every finite complex Möbius group is conjugate (in the group
of Möbius maps acting on R3 ∪ {∞}) to a finite group of rotations of R3 , and
hence to the symmetry group of one of the five Platonic solids, or to a cyclic or
dihedral group.
6. General convergence
The value of the continued fraction (1.1) generated by the sj is limn→∞ Sn (0).
We know that Sn (∞) = Sn−1 (0) and, presumably with this in mind, in 1957
Piranian and Thron proved the following result [66] and expressed the hope that
this might be useful in continued fraction theory.
Theorem 6.1. Suppose that a sequence gn of Möbius transformations converges
to the same value, say w, at two distinct points of C∞ . Then, with the possible
exception of one particular value of z, if gn (z) converges then it converges to w.
The sequence gn (z) = nz shows that the exceptional value of z in Theorem 6.1
can exist. We shall give a proof of Theorem 6.1 that is based on hyperbolic
geometry; this proof is short, free of coefficients, and valid in all dimensions. We
shall discuss a generalization of the results in [66] to all dimensions in Section
10.
A geometric proof of Theorem 6.1. It is sufficient to assume that there are
four distinct points z1 , z2 , z3 , z4 and three distinct values w1 , w2 , w3 such that
gn (z1 ) and gn (z2 ) tend to w1 , while gn (z3 ) → w2 and gn (z4 ) → w3 , and reach a
contradiction. By considering a conjugate sequence ggn g −1 for a suitable g, we
may suppose that no wj is ∞. Let γ12 be the geodesic in H3 with endpoints
z1 and z2 , let γ34 be the geodesic with endpoints z3 and z4 , and let d be the
hyperbolic distance between the geodesics γ12 and γ34 (we remark that given
any two hyperbolic geodesics with no end-points in common, there is a unique
hyperbolic segment that is orthogonal to both, and the length of this segment is
the hyperbolic distance between the two geodesics). As each gn is an isometry
of H3 , the distance between gn (γ12 ) and gn (γ34 ) is also d. However for large n,
gn (γ12 ) lies in small Euclidean neighbourhood of w1 , and gn (γ34 ) is approximately
the geodesic with endpoints w2 and w3 (when w2 6= w3 ) or it lies in a small
Euclidean neighbourhood of w2 (when w2 = w3 ). In both cases, the distance
between gn (γ12 ) and gn (γ34 ) tends to +∞ as n → ∞ and this is our contradiction.
At first sight Theorem 6.1 seems to justify the use of the point 0 in the definition
of convergence of a continued fraction, for if Sn (0) converges then so does Sn (∞)
and so, with possibly one exceptional value of z, limn→∞ Sn (z) (when it exists) is
independent of z. Despite this, there is a very good reason why the definition of
convergence of continued fractions should be modified. The problem is that Sn (z)
may converge (to a constant value) at many z, but perhaps not when z is 0 or ∞.
For example, consider any continued fraction in which the sequence s1 , s2 , . . . has
period three, and in which s1 s2 s3 is a loxodromic transformation g with attracting
and repelling fixed points α and β, respectively. Then the sequence g n of iterates
of g converges to α on C∞ \{β}, and it follows from this that Sn → α locally
uniformly on the complement of the three points β, s−1 −1
1 (β) and (s1 s2 ) (β). As
convergence is a topological rather than an arithmetic concept, it seems beyond
question that in this case one should regard the continued fraction as being
convergent, and assign the value α to it, regardless of whether or not Sn converges
at 0 (or at ∞). We now give an explicit example of this behaviour (this is a trivial
simplification of that given in [48]).
Example 6.2. Consider the continued fraction in which the Möbius sequence
s1 , s2 , s3 , . . . is s, t, s, s, t, s, s, t, s, . . . (of period three), where s(z) = 2/(2 + z)
and t(z) = 2/(−1 + z). A calculation shows that sts(z) = −1/2z; thus
µ ¶m
m 1
(sts) (z) = − z,
2
µ ¶m µ ¶
m 1 2
(sts) s(z) = − ,
2 2+z
µ ¶m µ
z−1
¶
m 1
(sts) st(z) = − .
2 z
Clearly, this continued fraction diverges. However, given any compact subset K
of C∞ \{0, −2, ∞} there is a constant k such that each of the functions (sts)m ,
(sts)m s and (sts)m st are bounded by k/2m on K; thus Sn → 0 locally uni-
formly on C∞ \{0, −2, ∞}. Surely, then, this continued fraction should (in any
reasonable definition of convergence) converge to 0?
As ‘almost all’ Möbius transformations are loxodromic the dynamics described

in Example 6.2 must be typical of periodic continued fractions. What makes
Example 6.2 interesting is that the point 0, which is chosen as the indicator of
convergence in the classical setting, is one of the exceptional points in the dynam-
ics. Example 6.2 shows that the classical definition of convergence of a continued
fraction should be replaced by a more general definition, and then followed by
by deeper study of pointwise, and locally uniform, convergence of Möbius maps.
There have been several attempts to obtain a satisfactory definition of conver-
gence of continued fractions, and for an account of the evolution of the definition
we refer the reader to [78], where the following definition was proposed.
Definition 6.3. The continued fraction (1.1) is said to converge strongly to a
value α in C if
(i) there are distinct u and v in C∞ such that Sn (u) → α and Sn (v) → α as
n → ∞, and
(ii) the set {Sn−1 (∞) : n = 1, 2, 3, . . .}, is not dense in C∞ .
The role of (ii) will become clearer in Section 9. Finally, in 1986, Lisa Jacobsen
(now L. Lorentzen) proposed the following definition ([48, p. 480]).
Definition 6.4. The continued fraction K(an |bn ) in (1.1) converges generally,
or is generally convergent, to a value α in C∞ if there exist sequences un and vn
in C∞ such that
lim Sn (un ) = lim Sn (vn ) = α, lim inf σ(un , vn ) > 0.
n→∞ n→∞ n→∞
If we let un = 0 and vn = ∞ then we see immediately that if a continued fraction

converges in the classical sense then it converges generally (to the same value).
Example 6.2 shows that the converse is false, In fact, Definition 6.4 gives the
correct definition of general convergence in all dimensions although the higher
dimensional case was not considered in [48]. It is clear that Definition 6.4 is
meaningful for any sequence gn of Möbius maps, and accordingly, we have a
definition of general convergence for arbitrary sequences of Möbius maps.
Definition 6.5. A sequence gn of Möbius maps converges generally, or is gen-
erally convergent, to a value α in C∞ if there exist sequences un and vn such
that
(6.1) lim gn (un ) = lim gn (vn ) = α, lim inf σ(un , vn ) > 0.
n→∞ n→∞ n→∞
Of course, the convergence is in terms of the chordal metric. Also, by rejecting

a finite number of initial maps, we may assume that inf n σ(un , vn ) > 0. After
giving Definitions 6.3 and 6.4, Jacobsen verified the first crucial step in any
limiting process, namely that a sequence of Möbius maps can converge generally
to at most one value. This can be proved using the standard recurrence relations
in continued fraction theory (but this proof is not valid for general sequences of
Möbius maps, nor in the higher dimensional situation) and by using cross-ratios.
However, the uniqueness of the limit is an immediate consequence of the following
geometric result which fully justifies the introduction of hyperbolic space, and
which may now be taken as the definition of general convergence.
Theorem 6.6. A sequence gn of Möbius maps converges generally to α in C∞
if and only if gn → α pointwise on H3 .
It is well known, and easy to see, that a sequence of Möbius maps converges
pointwise on H3 to a point α in C∞ if and only if it converges to α uniformly
on compact subsets of H3 . This is a direct consequence of the fact that each
Möbius map is an isometry of H3 and that the Euclidean length of a ‘ruler’ of
fixed hyperbolic length shrinks to zero as the ruler approaches the boundary
of hyperbolic space (as is beautifully illustrated in two dimensions by Escher’s
well-known tesselations of the unit disc). Our proof of Theorem 6.6 is based on
the following geometric lemma.
Lemma 6.7. Let γ be a geodesic in H3 whose endpoints u and v in C∞ are at a
chordal distance δ apart. Then γ passes within a hyperbolic distance cosh −1 (2/δ)
of the point j, where j = (0, 0, 1).
Proof. We map H3 onto the open unit ball B3 in R3 by a Möbius map that
is an extension of the stereographic projection of C∞ onto ∂B3 . This map is a
hyperbolic isometry, and it maps j to 0 in B3 . Let u0 and v 0 be the points on ∂B3
that correspond under this map to u and v in C∞ . Then (by the definition of the
chordal metric) σ(u, v) = |u0 −v 0 |. We now apply the angle of parallelism formula
(5.1) to Figure 6.1 and obtain cosh d sin θ/2 = 1. As σ(u, v) = |u0 −v 0 | = 2 sin θ/2,
we see that the distance d between j and γ is given by cosh d = 2/σ(u, v) as
required.
Proof of Theorem 6.6. Suppose first that gn converges generally to α, where

we may assume that α ∈ C. Let un and vn be as in Definition 6.5 with, say,
σ(un , vn ) ≥ δ > 0. It follows from Lemma 6.7 that each geodesic γn (with
endpoints un and vn ) contains a point xn such that, for all n, ρH (xn , j) ≤ d,
where cosh d = 2/δ. Now gn (un ) and gn (vn ) converge to α, and it is clear that
this implies that gn (xn ) → α (in the Euclidean metric on R3 ). Finally, as
¡ ¢
ρH gn (xn ), gn (j) = ρH (xn , j) ≤ d,
we see (as in Escher’s pictures) that gn (j) → α (in the Euclidean metric on R3 ).
Clearly the same argument holds with j replaced by any other chosen point of
H; thus gn → α pointwise on H.
Now suppose that gn (j) → α. Take any hyperbolic geodesic through j, and let
its endpoints be, say, z and z 0 . As gn (j) → α we see that
min{|gn (z) − α|, |gn (z 0 ) − α|} → 0.
0r
v0
@ θ/2
@
d@
u0
Figure 6.1
Let un be z if gn (z) is closer to α than gn (z 0 ), and z 0 otherwise; thus gn (un ) → α.

Now repeat this with another geodesic through j, with z, z 0 and un replaced by
w, w0 and vn , respectively, where z, z 0 , w and w0 are distinct. Then gn (un ) and
gn (vn ) both tend to α, and
σ(un , vn ) ≥ min{σ(z, w), σ(z, w 0 ), σ(z 0 , w), σ(z 0 , w0 )} > 0,
and this completes the proof of Theorem 6.6.
Despite the fact that we have not yet discussed Möbius maps acting in higher
dimensions (see Section 8), we end this section by giving a formal definition of a
continued fraction in RN .
Definition 6.8. A continued fraction in RN is a sequence of Möbius maps
s1 , s2 , . . ., that act on RN ∪ {∞}, and that satisfy sn (∞) = 0 for all n. The
continued fraction [s1 , s2 , . . .] converges to α if the sequence gn converges gener-
ally to α.
With some minor modifications, Theorem 6.6 and its proof remain valid in all
dimensions, and so we see that the general convergence of a continued fraction
in RN is equivalent to the pointwise convergence of the sequence Sn on HN +1 .
7. Strong divergence
As convergence (whether classical or general) includes convergence to ∞, the
divergence of a sequence of Möbius maps implies an oscillatory behaviour. In
this section we introduce a stronger notion of divergence which can only be

understood by considering the action of Möbius maps on H3 . We have seen that
a sequence gn of Möbius maps converges generally to α in C∞ if and only if
gn → α pointwise on H3 (Theorem 6.6). Now an extreme case in which this
fails is when the the sequence gn (j) does not even escape to the boundary of
hyperbolic space; this is our notion of strong divergence.
Definition 7.1. A sequence gn of complex Möbius maps is strongly divergent
if for one (and hence all) x in H3 , the sequence gn (x) lies in a compact subset
of H3 . A strongly divergent sequence cannot converge generally.
Presumably there are examples in which the sequence gn (j) accumulates both
at points of C∞ and points in H3 , and also examples in which gn (j) accumu-
lates at different points in C∞ but at no point of H3 . Such cases are likely to
be difficult to analyze, and they are closely related to the idea of a restrained
sequence of Möbius maps. A sequence gn of Möbius maps is restrained if and
only if kgn k → +∞ (although this was not the original definition of a restrained
sequence given in [49]); equivalently, by (5.4), if and only if the sequence gn (j)
acumulates only on the boundary of H3 . Such sequences have been examined
in detail by Beardon and Lorentzen (see [18]). By contrast, strong divergence is
easily analyzed. Because
1 ¡
kgn k2 = kIk2 + 4 sinh2 ρ j, gn (j) ,
¢
(7.1)
2
we have the following result.
Theorem 7.2. A sequence gn of complex Möbius maps is strongly divergent if
and only if the sequence kgn k is bounded.
In terms of continued fractions, we recall the numbers An and Bn from the

classical treatments of the subject, and the expression (2.5) for Sn . Then, from
(2.3) (which we need to express Sn in terms of a unimodular matrix), we have
Corollary 7.3. The continued fraction K(an |bn ) is strongly divergent if and only
if there is some constant M such that for all n,
|An−1 |2 + |Bn−1 |2 + |An |2 + |Bn |2 ≤ M |a1 · · · an |.
We can now give a simple criterion for the inner composition sequence g1 · · · gn
to be strongly divergent, and as this is directly in terms of the generators g n of
the composition sequence this is more useful than Theorem 7.2 or Corollary 7.3.
The result is as follows.
P ¡ ¢1/2
Theorem 7.4. Let g1 , g2 , . . . be Möbius maps. If n kgn k2 − 2 converges,
then the sequence g1 · · · gn is strongly divergent.
Proof. It is clear from (7.1), and the fact that sinh x ∼ x as x → 0, that the
three series
X¡ ¢1/2 X 1 X ¡
kgn k2 − 2
¢
, sinh ρ(j, gn (j), ρ j, gn (j) ,
n n
2 n
P ¡ ¢
converge or diverge together. Thus, by assumption, n ρ j, gn (j) = M , say,
where M < +∞. Let G0 = I, and Gn = g1 · · · gn . As the gj are isometries of H3 ,
we have
n−1
X n−1
¡ ¢ ¡ ¢ X ¡ ¢
ρ j, Gn (j) ≤ ρ Gm (j), Gm+1 (j) = ρ j, gm+1 (j) ≤ M.
m=0 m=0
This means that the points Gn (j) lie in some compact subset of H3 and so the
sequence Gn is strongly divergent.
Theorem 7.4 contains the Stern-Stolz Theorem and much more. If we apply it
to the sequence sn , where sn (z) = 1/(z + bn ) then (after representing sn by a
unimodular matrix) we find that ksn k2 = 2 + |bn |2 , and the Stern-Stolz Theorem
follows immediately. Note that in this case we find that
1 ¡ ¢
|bn | = 2 sinh ρ j, sn (j) .
2
We stress that this is an exact result (and not an estimate of |bn |), and that (as
far as I know) no such comparable interpretation is available if we confine our
attention to C∞ . In fact, we have obtained more than the classical Stern-Stolz
Theorem for we have shown that K(an |bn ) is strongly divergent. In addition,
we have used an argument that is valid in all dimensions. At this point it is
appropriate to mention a problem raised by Wall. In the footnote on [83, p. 33],
Wall raises the Pquestion of whether in the Stern-Stolz Theorem the conditional
convergence of n bn is sufficient to guarantee the divergence of the continued
fraction, As far as I know, this question has not yet been resolved. It may be that
the geometric approach taken here can shed more light on this matter; indeed, it
seems plausible that that the answer to Wall’s question may lie in understanding
the way that the orbit Sn (j) lies in H3 .
Let us now apply Theorem 7.4 to the general continued fraction generated by the
sequence sn (z) = an /(bn + z). In this case sn is represented by the unimodular
matrix Ã √ !
0 i an
i ibn
,
√ √
an an
so that
2 (1 − |an |)2 |bn |2
ksn k − 2 = + .
|an | |an |
P ¡ ¢1/2
Now¯ it is easy¯ to see that n ksn k2 −2 converges if and only if the two series
P P
¯1 − |an |¯ and
n n |bn | converge; thus Theorem 7.4 contains the following
result.
P ¯ ¯ P
Corollary 7.5. If n ¯1 − |an |¯ and n |bn | converge, then K(an |bn ) is strongly
divergent.
It is worthwhile to look at Theorem 7.4 in a different light. The Möbius maps
that fix j are the unitary maps; that is, g(j) = j if and only if kgk2 = 2. Thus
Theorem 7.4 says that the inner composition sequence g1 · · · gn is strongly diver-
gent if and only if (in some sense) the gn approach the subgroup of unitary maps
sufficiently rapidly. Thus, in this result we do not require that the gn converge,
but merely that each gn looks increasingly like some unitary map. With contin-
ued fractions in mind we note that the map s(z) = a/(b + z) is unitary if and
only if |an | = 1 and bn = 0, and this is the special case given in Corollary 7.5.
Corollary 7.5 can be derived from the Stern-Stolz Theorem (from the theory of
equivalent continued fractions), and it occurs in [57], but (as we shall see later)
Theorem 7.4 applies to all Möbius maps in all dimensions.
We shall now discuss the following two results which we express in terms of
matrix norms (rather than sequences of coefficients as is done in [57] and [61]).
Theorem 7.6. Suppose
P that s given by s(z) = a/(b + z) is elliptic, that sn (z) =
an /(bn + z) and n ksn − sk converges. Then K(an |bn ) diverges. Further, if
z0 is a fixed point of s, then Sn (z0 ) converges to some limit, say w, and then
Sn−1 (w) → z0 .
P
Theorem 7.7. Suppose that tn → t, where each tn and t is elliptic. If n ktn −tk
converges then t1 · · · tn converges to distinct values at the two fixed points of t,
and diverges elsewhere.
The following result is equivalent to Theorem 7.7 in the case of complex Möbius
maps, but it (and the proof) generalizes to higher dimensions. Moreover, it is
easier to prove if we use the metric σ0 rather than the matrix norm (or the
sequences of coefficients).
P
Theorem 7.8. Let sn and s be complex Möbius maps. Suppose that n σ0 (sn , s)
converges, and that s lies in a compact subgroup G of M. Then s1 · · · sn converges
at the two fixed points of s, and diverges elsewhere.
The crucial step in the proof of Theorem 7.8 depends directly on the compactness
of G rather than the nature of the elliptic maps. Although it is true that every
elliptic map lies in a compact subgroup of M, and that every element of a
compact subgroup of M is elliptic, it is compactness that is the decisive feature
for the following reason. By (5.5), the map g 7→ L(g) is continuous and positive,
so that if G is a compact subgroup of M, then
(7.2) L(G) = sup{L(g) : g ∈ G} < +∞.
Thus a compact subgroup of M is a uniformly Lipschitz family.
Proof of Theorem 7.8. The triangle inequality gives
σ0 (s1 · · · sn , sn ) ≤ σ0 (s1 · · · sn , ss2 · · · sn )

+σ0 (ss2 · · · sn , s2 s3 · · · sn ) + · · · + σ0 (sn−1 sn , sn );
thus, by right invariance,
σ0 (s1 · · · sn , sn ) ≤ σ0 (s1 , s) + L(s)σ0 (s2 , s) + · · · + L(sn−1 )σ0 (sn , s).
With (7.2) this yields
σ0 (s1 · · · sn , sn ) ≤ L(G) σ0 (s1 , s) + σ0 (s2 , s) + · · · + σ0 (sn , s) .

£ ¤
Of course, we can apply this to sk , sk+1 , . . . and thereby obtain the inequality
σ0 (sk sk+1 · · · sk+n−1 , sn ) ≤ L(G) σ0 (sk , s) + σ0 (sk+1 , s)

£
(7.3) ¤
+ · · · + σ0 (sk+n−1 , s) .
We know that s is elliptic. Suppose now that z is not a fixed point of s; then the
sequence sn (z) diverges so that there is a positive d such that for every m the
set {sn (z) : n ≥ m} has chordal diameter at least d. Now choose k such that
∞
X d
L(G) σ0 (sk , s) < .
n=k
4
Then, from (7.3), we have σ0 (sk sk+1 · · · sk+n−1 , sn ) < d/4. It is now obvious
that the sequence sk · · · sn (z), n = k, k + 1, . . ., diverges and hence so does the
sequence s1 · · · sn (z), n = 1, 2, . . .. This shows that s1 · · · sn diverges at every z
that is not fixed by s.
Now let w be a fixed point of s. We shall show that s1 · · · sn (w) is a Cauchy
sequence and so converges. First, from (7.3), we see that for k ≥ K, say,
√
σ0 (sk sk+1 · · · sk+n−1 , sn ) ≤ 2
for all n. This combined with Lemma 5.5 then shows that for all m,
L(s1 · · · sK+m ) ≤ L(s1 · · · sK−1 )L(sK · · · sK+m )

≤ 18L(s1 · · · sK−1 )L(sm )
≤ 18L(s1 · · · sK−1 )L(G).
It is now clear that there is some constant M such that for all n, L(s1 · · · sn ) ≤ M ;
in other words, the family of s1 · · · sn is uniformly Lipschitz.
It is now easy to complete the proof, for we have

¡ ¢ ¡ ¢
σ s1 · · · sm (w), s1 · · · sm+n (w) ≤ L(s1 · · · sm )σ w, sm+1 · · · sm+n (w)
¡ ¢
≤ M σ w, sm+1 · · · sm+n (w)
= M σ sn (w), sm+1 · · · sm+n (w)
¡ ¢
≤ M σ0 sn , sm+1 · · · sm+n
¡ ¢
£ ¤
≤ L(G) M σ0 (sm+1 , s) + · · · + σ0 (sm+n , s) ,
P
the last inequality here following from (7.3) with k = m + 1. As n σ0 (sn , s)
converges we conclude that s1 · · · sn (w) is a Cauchy sequence and this completes
the proof.
Essentially the same argument (in which we replace the uniform Lipschitz con-
dition by the inequality L(sn ) ≤ L(s)n ) yields the following result.
Theorem 7.9. Let sn and s be complex Möbius maps. Suppose that the series
n n
P
n σ0 (sn , s)L(s) converges. If the sequence s (z) diverges then so does the
sequence s1 · · · sn (z). Further, s1 · · · sn converges at the two fixed points of s.
There are other results of this type to be found in, for example, [41] and [61],
and it may be that these and other results of a similar nature can be proved by
geometric means. For example, Thron and Waadeland have proved the following
result ([77, Proposition 1, p. 652]).
Theorem 7.10. Let s(z) = a/(b + z) have fixed points x and y, and suppose
that r satisfies 0 < r < |x/y| < 1/r. Then there is a positive A such that if
|an − a| ≤ Ar n and |bn − b| ≤ Ar n , then the sequence s1 · · · sn (x) converges,
where sn (z) = an /(bn + z).
The following similar result follows directly from Theorem 7.9 (and this too could
be phrased in terms of matrix norms or coefficients).
Theorem 7.11. Let sn and s be Möbius maps, and suppose that µ < L(s). If
there is some A such that for all n, σ0 (sn , s) ≤ Aµn , then the sequence s1 · · · sn (x)
converges for each fixed point x of s.
8. Möbius maps in higher dimensions

We begin with a brief discussion of three spaces, namely Rn with the chordal
metric σ, the upper half-space HN +1 of RN +1 (hyperbolic space), and the open
unit ball BN +1 (our second model of hyperbolic space). These are all direct
generalizations of the two and three-dimensional cases. First, we embed RN in
RN +1 by identifying (x1 , . . . , xn ) with (x1 , . . . , xn , 0). We use |x| for the length
of a vector x (in any dimension), and we let e1 , . . . , eN +1 be the usual Euclidean
basis of RN +1 .
As usual, we adjoin a point ∞ to RN +1 to form the space RN +1

∞ . Exactly as in the
complex case, we can define a stereographic projection ϕ : RN ∞ → S , where S
N N
is the boundary of the unit ball B N +1

in R N +1
, and ϕ(∞) = (0, . . . , 0, 1); then ϕ
is a bijection of RN
∞ onto S N
. To find the formula for ϕ we let x = (x1 , . . . , xN , 0);
then ¡ ¢
ϕ(x) = x + t eN +1 − x ,
where t is chosen so that ϕ(x) ∈ SN or, equivalently, |ϕ(x)|2 = 1. This gives a
quadratic equation in t and, as t 6= 1, we obtain the formula
µ 2
|x| − 1
µ ¶ ¶
2
ϕ(x) = x+ eN +1
|x|2 + 1 |x|2 + 1
(8.1)
2xN |x|2 − 1
µ ¶
2x1
= ,..., 2 , .
|x|2 + 1 |x| + 1 |x|2 + 1
The chordal distance σ(x, y) between points x and y in RN is, by definition, the
Euclidean distance between the points ϕ(x) and ϕ(y) on SN , and it is easy to
see that the formula is the same as in the complex case, namely
2|x − y|
(8.2) σ(x, y) = ,
(1 + |x|2 )1/2 (1
+ |y|2 )1/2
with the obvious interpretation when x or y is ∞. As division is not available
in RN when N ≥ 3 we cannot define the cross-ratio of four distinct points in RN
∞.
However, we can define the absolute cross-ratio by
σ(a, c) σ(b, d)
(8.3) [a, b, c, d] = ,
σ(a, d) σ(b, c)
and this does not need any of the usual ‘conventions’ about ∞. These conventions
are tied to Euclidean geometry, and providing that they are in force, then we
also have
|a − c| |b − d|
[a, b, c, d] = .
|a − d| |b − c|
There is a useful application of these ideas that seems to be rarely used. Consider
(for simplicity) the extended complex plane C∞ with the chordal metric which,
for the moment, we denote by σ2 . Now each Möbius map g : C∞ → C∞ pre-
serves cross-ratios and hence also chordal cross-ratios. Next, each Möbius map g
extends to a Möbius map of R3∞ onto itself, and this space is equipped with its
own chordal metric, which we denote by σ3 . It is clear from (8.1) that σ2 = σ3
on C∞ , and we shall see later that any Möbius map (acting in any dimension)
preserves cross-ratios. Thus the familiar fact that complex Möbius maps preserve
cross-ratios in C∞ can be embedded in the more general statement that complex
Möbius maps act on R3∞ with its chordal metric, and that in this action chordal
cross-ratios are preserved. Thus, in general, we expect any argument involving
only Möbius maps and cross-ratios in the complex plane automatically to be
valid for the corresponding action on all of R3∞ . As there are many arguments
in continued fraction theory that involve cross-ratios (see, for example, [79] and
[82]) this remark shows how some of these might be generalised to R3∞ . This idea
will be exploited in our discussion of tail sequences in Section 9.
Our first model of hyperbolic space is
HN +1 = {(x1 , . . . , xn+1 ) : xn+1 > 0},
where this is equipped with the hyperbolic metric derived from the line element
ds = |dx|/xN +1 . Observe that the boundary of HN +1 is RN ∞ . Our second model
of hyperbolic space is the unit ball B N +1
in R N +1
with the hyperbolic metric
given by
2|dx|
ds = .
1 − |x|2
For brevity, we shall denote the hyperbolic metric in both of these models, and
in all dimensions, by ρ; it should be clear from the context which space ρ refers
to (and in any case the two models are isometrically equivalent). Finally, we
note that both models of hyperbolic space are complete metric spaces. Shortly,
we shall discuss a third model of hyperbolic space.
We shall now describe how Möbius maps act on these spaces, and, in particular,
we shall show how stereographic projection ties together the three spaces that
we have discussed so far. A Möbius map acting in RN ∞ is a finite composition
γ1 · · · γk , where each γj is an inversion (or a reflection) in some (n−1)-dimensional
hyperplane or hypersphere in RN ∞ . The Möbius group acting on R∞ is the group
N
generated by these inversions; the conformal Möbius group, which we denote by

MN , consists of those Möbius maps that can be expressed as the composition of
an even number of inversions.
For the moment, let α : RN ∞ → R∞ be the inversion in the sphere with centre
N
(a1 , . . . , an ) and radius r, and let β : RN

∞
+1
→ RN∞
+1
be the inversion in the sphere
with centre (a1 , . . . , an , 0) and radius r. It is clear that α is the restriction of β
to RN ∞ , and as the corresponding statement for reflections across hyperplanes is
true, we see that every element of Mn is the restriction of an element of MN +1
to RN ∞ ; equivalently, every element of MN extends uniquely to an elelement of
MN +1 that leaves HN +1 invariant. In this way the group MN is identified with
those elements of MN +1 that leave RN × {0}, and also HN +1 , invariant. Exactly
as in the complex case, it can be shown that MN is the group of all conformal
hyperbolic isometries of HN +1 and, as we have already made clear, it is generally
more profitable to study MN through its action on HN +1 rather than its action
on RN ∞.
Let us now consider the inversion, say α, in the sphere Σ in RN +1 given by

(8.4) x21 + · · · + x2n + (xN +1 − 1)2 = 2.
√
Note that Σ has centre eN +1 , and radius 2, and that SN ⊂ Σ. Now α satisfies
¡ ¢
α(y) = eN +1 + t y − eN +1 ,
where t is chosen so that |α(y) − eN +1 | |y − eN +1 | = 2, and this gives
|y − eN +1 |2 − 2
µ ¶ µ ¶
2
(8.5) α(y) = y+ eN +1 .
|y − eN +1 |2 |y − eN +1 |2
If y ∈ RN , then y is orthogonal to eN +1 , so that |y − eN +1 |2 = |y|2 + 1. In this
case we see from (8.1) and (8.5) that ϕ(y) = α(y), and so we have proved that
the stereographic projection ϕ is just the restriction of α to RN . This confirms
that the chordal metric is inextricably linked with inversions, and hence with the
Möbius group.
As α(RN ∞ ) = S , it follows that α must map H
N N +1
bijectively onto either the
unit ball B N +1
, or the exterior E N +1
of the closure of this ball. Now it is evident
(either from (8.5), or geometrically) that α interchanges 0 and −eN +1 ; thus we
conclude that α(HN +1 ) = EN +1 . By considering the lower half space of RN +1 ,
the following result is now clear.
Theorem 8.1. Let α be the inversion in the sphere given in (8.4), and let β be
the reflection across the plane xN +1 = 0. Then αβ is in MN +1 , αβ = ϕ on RN
∞,
αβ(eN +1 ) = 0, and αβ(H N +1
)=B N +1
.
It is important to realise that the map αβ acts as an isometry in two different

ways, namely
(a) from the hyperbolic space Hk+1 to the hyperbolic space Bk+1 , and
(b) from RN∞ with the chordal metric to S
N
with the Euclidean metric.
We omit the proof of (a), but (b) is precisely the definition of the chordal metric.
Let us summarize how these ideas are often used. Instead of regarding MN as
acting on RN∞ , we first regard it as the group of isometries of H
N +1
, and then
replace this action by the action of the conjugate group (αβ)MN (αβ)−1 that
acts on BN +1 and on its boundary SN +1 . For complex Möbius maps this means
that we can (and often should) study their behaviour relative to the chordal
metric by studying their (transformed) action on the unit ball B3 relative to the
hyperbolic metric.
Let us also briefly observe that many of the usual properties of complex maps
hold for Möbius maps acting in higher dimensions. First, Möbius maps take
spheres to spheres (because inversions do), and they preserve inverse points with
respect to spheres. It is known that the chordal cross-ratio is invariant under
Möbius maps; more generally, a map of RN ∞ into itself is Möbius if and only if it
preserves chordal cross-ratios ([12, Theorem 3.2.7]). To minimize the notation
when using cross-ratios of points g(x), we shall often write gx for g(x), and so
on. For many more details on Möbius maps and the chordal cross-ratio, see [16].
Next, we observe that (as in the complex case) the chordal metric induces the
metric
σ0 (f, g) = sup{σ f (x), g(x) : x ∈ RN
¡ ¢
∞}
on the space of continuous maps of R∞ into itself. The restriction to the space of
N
continuous surjective maps is a right-invariant metric, and σ0 (fn , f ) → 0 if and

only if fn → f uniformly on RN ∞ . We also have the following result.
Theorem 8.2. Suppose that g is a Möbius map acting on RN ∞ . Then g is a

Lipschitz map with respect to the chordal metric, and the best Lipschitz constant
L(g) for g is given by
¡ ¢
L(g) = exp ρ eN +1 , g(eN +1 ) .
In particular, a subfamily F of MN is uniformly Lipschitz with respect to the
chordal metric if and only if {g(eN +1 ) : g ∈ F} lies in a bounded subset of HN +1 .
With this, the same proof as in the complex case yields the following result.
Theorem 8.3. The group MN with the metric σ0 is both a complete metric
space, and a topological group.
We have seen that studying the action of a Möbius map g on RN ∞ in terms of

the chordal metric is equivalent to studying the action of the Möbius map g ∗ ,
where g ∗ = (αβ)g(αβ)−1 , on SN in terms of the Euclidean metric, and that this
is best done by studying the action of g ∗ on BN +1 in terms of to its hyperbolic
metric. As with the complex case, there is a theory of isometric spheres which
can be used to analyze the action of g ∗ . Suppose now that g is a Möbius map
that preserves BN +1 . Then g is an inversion in its isometric sphere Sg followed
by a Euclidean isometry so that, as far as the Euclidean distortion on SN is
concerned, we need only consider the inversion in Sg . The isometric sphere Sg
is orthogonal to SN , and its centre is g −1 (∞) (which lies outside SN ). Because
Möbius maps preserve inverse points, we have ¡ −1 |g −1 (∞)| −1
¢ |g ¡(0)| =¢ 1. Next,
because g is a hyperbolic isometry, we have ρ g (0), 0 = ρ 0, g(0) , and this
implies that |g(0)| = |g −1 (0)|. Finally, because Sg is orthogonal to SN , we have
|g −1 (∞)|2 = 1 + Rg2 , where Rg is the Euclidean radius of Sg . These simple
geometric facts yield the important relationship
1
(8.6) = 1 + Rg 2 ,
|g(0)|2
and this leads to the following result.
Theorem 8.4. Suppose that g is a Möbius transformation in MN , and let g ∗
be the corresponding Möbius transformation (as described above) that preserves
BN +1 . Then for all x and y in RN∞,
µ ¶
¡ ¢ 4
(8.7) σ g(x), g(y) ≤ 4 + 2 σ(x, y),
Rg
where Rg is the radius of the isometric sphere of g ∗ .

We remark that the proof of (8.7) depends on (8.6) and the fact that
¡ ∗
¢ 1 + |g ∗ (0)|
L(G) = exp ρ(0, g (0) = ;
1 − |g ∗ (0)|
see [12, Theorem 3.6.1, p. 42]. As general references to the action of Möbius
maps in the hyperbolic spaces HN and BN , we recommend [4], [5], [12], [13], [23],
[28], [64], [70] and [85].
We turn now to another model of hyperbolic space, namely the hyperboloid model.
This model is important for it enables us to realize each Möbius map acting on
RN as a matrix action, and hence gives us a more algebraic way of thinking about
continued fractions in higher dimensions. For x and y in RN +1 , let
q(x, y) = x1 y1 + · · · + xN yN − xN +1 yN +1 ,
and
HN = {x ∈ RN +1 : q(x, x) = 1, xN +1 > 0};
geometrically, HN is one branch of a hyperboloid of two sheets. It is easy to see
that if γ(t) is a curve on HN , then q(γ̇(t), γ̇(t)) ≥ 0, so we can define a metric
on HN by
(8.8) ds2 = dx1 2 + · · · + dxN 2 − dxN +1 2 .
It is a fact that HN with this metric is also a model of hyperbolic N-space; for
general references to this we cite, for example, [5], [12], [13], [23], [26], [43], [70]
and [85]. We now have the following two important results.
Theorem 8.5. The map (x1 , . . . , xN , xN +1 ) 7→ (1 + xN +1 )−1 (x1 , . . . , xN ) is an
isometry of HN onto BN .
Theorem 8.6. The isometries of HN with the hyperbolic metric given in (8.8)
are the (N +1)×(N +1) matrices that preserve q and also the condition x N +1 > 0.
This is the result which enables us to recapture the fact that the Möbius group is a
matrix group. The complex Möbius group is initially defined as an action on C∞
(or R2∞ ), but these Möbius maps have a natural extension to hyperbolic 3-space
H3 . These, in turn, convert by conjugation, into an action of H3 (lying in R4 ),
and so can be represented by 4 × 4 real matrices that preserve the hyperboloid
given in R4 by x1 2 +x2 2 +x3 2 +1 = x4 2 . This is the classical link between complex
Möbius maps and the Lorentz group, and for explicit details of this, the reader
is referred to [28], [29] and [85]. The important relation (5.3) in Theorem 5.2
extends to all dimensions in the following way (see [13]).
Theorem 8.7. Let g be a Möbius map acting on RN , and hence also on hyper-
bolic space HN +1 . Suppose that g corresponds to the (N + 2) × (N + 2) matrix A
acting on HN +2 ; then
kAk2 = kIk2 + 4 sinh2 ρ eN +1 , g(eN +1 ) .
¡ ¢
Note that kIk2 = N +2 here. If g is a complex Möbius map that is represented by

the 4 × 4 matrix A, then kgk2 = kAk (see [13]), so that kgk4 = 4 cosh2 ρ j, g(j) ,
¡ ¢
which is equivalent to (5.3).
As an application of Theorem 8.7, we have the following generalization of the
Stern-Stolz theorem to all dimensions (and the proof is exactly the same as for
Theorem 7.4).
Theorem (The General Stern-Stolz Theorem). Suppose that g1 , g2 , . . . are Möbius
maps in MN , and that gn is represented by the (N + 2) × (N + 2) matrix An in
the sense of Theorem 8.7. If
∞
X p
kAn k2 − kIk2
n=1
converges, then the sequence g1 · · · gn is strongly divergent.
We end this section with a discussion of a generalization of the way that the
action of complex Möbius maps on H3 can be expressed through the algebra of
quaternions. The Clifford algebra CN is the associative algebra over R generated
by the N elements i1 , i2 , . . . , iN subject to the relations ij 2 = −1 and ij ik =
−ik ij when j 6= k. Explicitly, Pthis means that each element of CN has a unique
representation in the form I aI I, where the aI are real, and where in this
sum I runs over all products iν1 iν2 · · · iνp , where 1 ≤ ν1 < · · · < νp ≤ N . The
empty product is allowed, and it is interpreted as the real number 1. The cases
N = 0, 1, 2 give the real numbers, the complex numbers, and the quaternions,
respectively; for quaternions we have i1 = i, i2 = j, and k = i1 i2 .
Clearly, CN is a real normed vector space of dimension 2N , with the norm |a| and
scalar product a·b defined in the natural way by
³ X ´1/2 X
|a| = aI 2 , a·b= aI b I .
I I
N +1
Let V be the subspace of elements in CN of the form a0 + a1 i1 + · · · + aN iN .
We refer to these elements as vectors, and we identify V N +1 with RN +1 in the
obvious way. We also adjoin ∞ to V N +1 to form V∞N +1 , and this is identified
with RN +1
∞ .
There are three involutions that play a central role in the theory: we shall
define each of these on the basis elements iν1 iν2 · · · iνp , and each is then ex-
tended to act on CN by linearity. The first of these is a 7→ a0 : this is the map
iν1 iν2 · · · iνp 7→ (−iν1 )(−iν2 ) · · · (−iνp ). The second involution is a 7→ a∗ : this
is the map ir1 · · · irs 7→ irs · · · ir1 . Finally, as these two involutions commute,
they induce a third involution defined by x 7→ x̄, where x̄ = (x0 )∗ = (x∗ )0 .
These involutions interact with multiplication in the following way: (ab) 0 = a0 b0 ,
(ab)∗ = b∗ a∗ , and (ab) = ba. Next, all three involutions map V N +1 into itself,
and are deemed to fix ∞. If x is a vector then x = x∗ , x0 = x̄ and

³ X ´³ X ´
xx̄ = xx0 = x0 + x r ir x 0 − xs is = |x|2 .
r s
It follows that each non-zero vector x has a unique two-sided multiplicative in-
verse x−1 , namely x0 /|x|2 , and these facts lead to the following result.
Lemma 8.8. The set of products of non-zero vectors is a multiplicative sub-
group ΓN of CN .
We call ΓN the Clifford group in CN . If N ≥ 3 then CN has divisors of zero;
for example, (i1 + i2 i3 )(i1 − i2 i3 ) = 0. In particular, it is not necessarily true
that |ab| = |a| |b|. However, we do have the following result (see [4], and [19,
Theorem 1]).
Lemma 8.9. Suppose that a, b ∈ CN , and that a or b is in ΓN . Then |ab| = |a| |b|.
We come now to the link between the Clifford algebra CN and Möbius trans-
formations acting in RN +1
∞ . Briefly, the group of Möbius transformations acting
on C∞ is essentially the same as the group of non-singular 2 × 2 matrices and,
with suitable restrictions, the same is true here. A matrix
µ ¶
a b
c d
with entries in ΓN ∪ {0} is said to be a Clifford matrix if ab∗ , cd∗ , c∗ a and d∗ b
are vectors, and ad∗ − bc∗ ∈ R\{0} (these rather complicated conditions arise
because CN is not a field). Finally, any Clifford matrix acts on V̄ N +1 by the rule
x 7→ (ax + b)(cx + d)−1 ,
where this is computed within the algebra of CN . Then (and there is much to
check here) these actions coincide with the actions of the Möbius transforma-
tions defined by repeated inversions in RN +1
∞ . Finally, for more details of the
quaternionic and Clifford algebra approach to Möbius maps, we suggest [1], [5],
[6], [7], [8], [9], [24], [42], [52], [69], [71], and [84].
There is now an obvious question, namely which results in continued fraction
theory remain true in higher dimensions when phrased in terms of Clifford alge-
bras? We know, for example, that Pringsheim’s Theorem is true (and in exactly
the same form) in all dimensions, and this is as follows (see [19]).
Theorem 8.10. Consider a sequence of Clifford matrices
µ ¶
0 bn
, n = 1, 2, . . . ,
cn d n
where for each n, |dn | ≥ |bn | + |cn |, and let these act on RN ∞
+1
by the rule
−1
sn (x) = bn (cn x + dn ) . Then the sequence s1 · · · sn (x) converges to a constant
in the unit ball in RN +1
∞ .
The proof of Theorem 8.10 is geometric, but does an algebraic proof exist?
9. Backward orbits
Theorem 6.1 shows that if a continued fraction converges strongly to α then,
apart from possibly one exceptional value of z, Sn (z) → α whenever this limit
exists. Moreover, the following result ([78, Theorem 2.1]) tells us where this
limit exists, and it also illuminates the role of (ii) in Definition 6.3 of strong
convergence.
Theorem 9.1. Suppose that K(an |bn ) converges strongly to a complex number α.
Then Sn → α on C∞ \L, where L is the derived set of the sequence Sn−1 (∞).
Theorem 9.1 suggests that it might be profitable to study the sequence Sn−1 (w) of
inverse images of a point w for different w but before we do this, let us consider
how the same concept plays a crucial role in complex dynamics and in discrete
Möbius groups. Historically, the theory of discrete groups was developed before
iteration theory, but as most readers will probably be more familiar with the
latter, we shall discuss iteration first.
Let R be a rational function. Then R induces a partition of the complex sphere
C∞ into two sets, namely the Julia set J (R) and the Fatou set F(R). The Julia
set is compact, and the Fatou set is open (but not necessarily non-empty). The
two basic properties of these sets are
(a) apart from possibly one exceptional value of w, the Julia set J (R) is the
derived set of the backward orbit of w, that is of the set
∞
[
{z : Rn (z) = w},
n=1
where Rn denotes the n-th iterate of R, and

(b) the Fatou set F(R) is the largest open set on which the family {R n : n ≥ 1}
of iterates of R is a normal family.
Let us now consider a group G of Möbius maps acting on C∞ (although the
same statements hold in any dimension). The group G is discrete if and only
if the identity map (or any single member of G) is an isolated element of G in
the topology induced by the metric σ0 on M. Any discrete group G induces
a partition of the sphere C∞ into a compact subset Λ(G), called the limit set
of G, and its complement Ω(G), called the ordinary set of G. There is a strong
analogy here with the theory of iteration in that any discrete group is countable,
say G = {g1 , g2 , . . .}, and
(a) for all except possible one exceptional value of w, the limit set Λ(G) is the
derived set of the orbit g1 (w), g2 (w), . . ., and
(b) the ordinary set Ω(G) is the largest open set on which {g1 , g2 , . . .} is a
normal family.
Further, a group is discrete if and only if it acts discontinuously in H3 .

The similarity is obvious: in both cases there is family of functions ({R n } or G),
and a compact set Λ, such that (roughly speaking) the family acts chaotically
on Λ but is well-behaved on its complement. We shall now show that the same
ideas are applicable to continued fractions, and indeed to any generally conver-
gent sequence of Möbius maps; this was first observed by Aebischer (see [2] and
[3]). We shall also see that certain questions about the convergence of continued
fractions in C∞ can only be answered satisfactorily by looking at the dynamics
that is taking place in hyperbolic space H3 .
To begin, consider a convergent continued fraction K(an |bn ) and, for m =
1, 2, . . ., let
am
Km (an |bn ) = .
am+1
bm +
bm+1 + · · ·
For brevity, we shall write Km for Km (an |bn ), so that K1 = K(an |bn ) (this is the
value of the continued fraction). The terms Km are (for obvious reasons) called
the tails of the continued fraction K(an |bn ), and as Sm (Km+1 ) = K(an |bn ) = K1 ,
we see that the sequence of tails Kn is just the sequence Sn−1 (K1 ). This sequence
is not defined when the continued fraction diverges (because then K1 is not
defined); see, for example, [80]. Quite generally, given any w, we can define the
tail sequence associated with w to be the sequence S1−1 (w), S2−1 (w), . . .. Equally,
given any sequence g1 , g2 , . . . of Möbius maps, acting in any dimension, there is
the inverse orbit g1−1 (w), g2−1 (w), . . . associated to each w and, for this reason,
and also to maintain the strong analogy with the other topics mentioned above,
we prefer to use the term inverse orbit rather than tails.
Definition 9.2. Given any sequence g1 , g2 , . . . of Möbius maps acting in RN
∞ , the
associated inverse orbit of w is the sequence gn−1 (w), n = 1, 2, . . .. The derived
set of the sequence gn−1 (w) (that is, the set of limit points of the sequence) is
called the limit set of w and is denoted by Λ(w).
Naturally, we think of the w in Definition 9.2 to be in RN ∞ , but it is important to

understand that we have many more w at our disposal than this. In particular,
each gn extends to an action on RN ∞
+1
and we can look at the inverse orbit of
any point in this space. It is a striking fact that if the sequence gn is generally
convergent in RN∞ , then the limit set of a point w in R∞
N +1
is actually a subset of
R∞ ; this seems to be the fact that makes generally convergent sequences (and
N
continued fractions) special and, once again, we see the value of extending the
action of Möbius maps from Euclidean space to hyperbolic space. We begin our
discussion of inverse orbits with the following basic lemma.
Lemma 9.3. Suppose that g1 , g2 , . . . is a sequence of Möbius maps acting in RN∞
that converges generally to α. If w ∈ RN ∞
+1
and w 6
= α then g n → α locally
uniformly on RN +1
∞ \Λ(w).
Proof. We shall work in RN +1

∞ , and we recall that the chordal metric σ on this
space is the extension of the chordal metric on RN . If K and L are subsets of
RN∞
+1
we let σ(K, L) be the infimum of the distances σ(z, w), where x ∈ K and
y ∈ L. If K and L are compact and disjoint, then σ(K, L) > 0. We shall also
use the chordal cross-ratio defined in (8.3), and we recall that this is invariant
under Möbius maps.
As g1 , g2 , . . . converges generally to α, there exist sequences an and bn in RN
∞ such
that
(9.1) lim inf σ(an , bn ) > 0, lim gn (an ) = lim gn (bn ) = α,
n→∞ n→∞ n→∞
and we know that gn → α pointwise on HN +1 . Now take any w in RN ∞

+1
other
than α. If gn (an ) = w for infinitely many n then w = α which is false, and
likewise for bn . Thus for all sufficiently large n the points an , bn and gn−1 (w) are
distinct.
Let K be a compact subset of RN +1
∞ \Λ(w), and let x be any point in K. For
sufficiently large n, an 6= bn so that for these n either x 6= an or x 6= bn . For the
moment we suppose that x 6= bn . Then, by the invariance of cross-ratios,
σ(an , bn ) σ(x, gn−1 w) σ(gn an , gn bn ) σ(gn x, w)
−1
= .
σ(an , gn w) σ(x, bn ) σ(gn an , w) σ(gn x, gn bn )
Replacing three of these chordal distances by the upper bound 2, we obtain the
inequality
8 σ(gn an , gn bn )
σ(gn x, gn bn ) ≤ .
σ(an , bn ) σ(gn an , w) σ(x, gn−1 w)
Now
lim inf σ(an , bn ) > 0,
n
lim inf σ(gn an , w) = σ(α, w) > 0,
n
lim inf σ(x, gn−1 w) ≥ σ(K, Λ(w)) > 0.
n
Thus we see that there is a positive M (which depends on w and K, but not on
n) such that for all sufficiently large n,
σ(gn x, gn bn ) ≤ M σ(gn an , gn bn ).
This was proved under the assumption that x 6= bn , but it clearly holds when
x = bn ; thus it holds for all x in K. As (9.1) holds, we see that gn → α uniformly
on K, and hence locally uniformly on RN +1
∞ \Λ(w).
Next, we see that with one exception, Λ(w) is independent of w. If we take w in

RN
∞ then clearly Λ(w) ⊂ R∞ ; thus this must also hold even when w is in R∞ .
N N +1
Lemma 9.4. Suppose that g1 , g2 , . . . is a sequence of Möbius maps acting in RN

∞
that converges generally to α. If u and v are in RN ∞
+1
\{α}, then Λ(u) = Λ(v).
In particular, Λ(u) ⊂ RN∞.
Proof. We may assume that u 6= v. Let us suppose that Λ(u) 6= Λ(v). Then
(by interchanging u and v if necessary) there is some ζ in Λ(u) but not in Λ(v).
Now find a neighbourhood Nα of α that does not contain u, and also a neigh-
bourhood Nζ of ζ whose closure is disjoint from the closed set Λ(v). As v 6= α,
Lemma 9.3 implies that gn → α uniformly on Nζ , so that for all sufficiently
large n, gn (Nζ ) ⊂ Nα . For these n, gn−1 (u) ∈
/ Nζ so that ζ cannot be in Λ(u),
contrary to our assumption.
Lemma 9.4 enables us to make the following definition.
Definition 9.5. Suppose that g1 , g2 , . . . is a sequence of Möbius maps acting in
RN∞ that converges generally to α. Then the limit set Λ of the sequence g n is the
set Λ(w) for all w in RN +1
∞ \{α}.
Theorem 9.6. Suppose that g1 , g2 , . . . is a sequence of Möbius maps acting in RN

∞
that converges generally to α, and let Λ be the limit set of the sequence gn . Then
RN∞ \Λ is the largest open set in R∞
+1 N +1
on which gn converges locally uniformly
to α.
Proof. Lemmas 9.3 and 9.4 imply that Σ contains RN +1
∞ \Λ. It is clear that gn
cannot converge locally uniformly to α on any larger open set than RN +1
∞ \Λ.
Indeed, if x ∈ / R∞ \Λ then x ∈ Λ and every neighbourhood of x will, by
N +1
Lemma 9.4, contains infinitely many points gn−1 (u) for each u other than α.
Clearly, this prevents gn from converging uniformly to α on any neighbourhood
of x.
In addition to the result in Theorem 9.6, it is also true that RN +1
∞ \Λ(w) is
the largest open set in R∞ on which {g1 , g2 , . . .} is a normal family (see [2]).
N +1
Because of this result, and the strong similarity it has with the earlier theorems,
we might also call Λ the Julia set of the sequence g1 , g2 , . . .. As the generally
convergent sequences include convergent continued fractions as a special case,
we should now ask whether any other theorems about Julia sets or limit sets (in
discrete group theory) have counterparts in the theory of continued fractions.
Moreover, we might ask whether Λ is (always) fractal, and does it have any
self-similarity properties? Someone should provides pictures of the sets Λ for
different continued fractions.
There is one more topic in this section, and this alone fully justifies (even in
the context of classical continued fractions) our consideration of the extension
of Möbius maps from RN ∞ to H∞ . Let gn be a sequence of complex Möbius
N +1
maps that converges generally to α. Of course these might be the approximants

of a convergent continued fraction but they need not be. In [78], and in the
context of classical continued fractions, a question is raised about the set C of
points z for which gn (z) → α. It is remarked there that C may contain points
in Λ, so that C includes, but may be larger than, C∞ \Λ. In particular, it is
noted that this fact makes it hard to characterize the set C of convergence of
the gn . Now such a characterization is given in [2] in terms of the so-called

conical limits points in Λ. Briefly, the sequence gn (z), where z ∈ Λ, fails to
converge to α if and only if the inverse orbit g1−1 (w), g2−1 (w), . . . of any w in H3
accumulates at z within in a vertical cone above z. There are several remarks
that can be made to help the reader appreciate this result. First, it should
be clearly understood that this is an example of a natural question about the
convergence of approximants of a continued fraction in the complex plane whose
(beautiful) answer is given entirely in terms of the hyperbolic geometry of H 3 .
There are probably many more questions of this type still to be uncovered. Next,
the cone, or Stolz region, mentioned in this result is a hyperbolic (as opposed to
a Euclidean) object, for it is the set of points in H3 that lie at most a hyperbolic
distance d, say, away from the hyperbolic geodesic in H3 that is vertically above z.
Finally, the condition that orbits may or may not accumulate at a point on the
boundary of hyperbolic space from within a cone is a standard, and extremely
important, condition in discrete group theory and, although we shall not discuss
it here, it is intimately connected with rates of Diophantine approximation of
irrationals by rationals and is best understood by looking at the geometry of
geodesics on the quotient surfaces. Much of this aspect of number theory (and
some parts of continued fraction theory) can be phrased in terms of the Modular
group and then these results generalized to all finitely generated discrete groups
acting on the hyperbolic plane. In this context, the condition that the orbits
accumulate in a cone plays a crucial role in understanding the geometry of the
group action in hyperbolic space; this was first clarified in H3 in [11].
10. Forward orbits

In 1957 Piranian and Thron published a paper ([66]) that is frequently quoted
in papers on continued fractions; indeed, Theorem 6.1 comes from that paper.
The main result in their paper gives a classification of the possible limits of a
sequence of complex Möbius maps and it is as follows.
Theorem 10.1. Let gn be any sequence of complex Möbius maps, and suppose
that the set C of z at which the sequence gn (z) converges is non-empty. Let
g(z) = limn gn (z) on C. Then one of the following possibilities occurs:
(a) g is constant on C;
(b) C = C∞ , and g is a Möbius map;
(c) C = C∞ and g takes exactly two values, one of which is taken exactly once;
(d) C = {z1 , z2 }, say, and g(z1 ) 6= g(z2 ).
In fact, there is a corresponding result in all dimensions which can be obtained by

purely geometric arguments. However, it is clear that some changes are necessary
before Theorem 10.1 can be established in higher dimensions. Consider, for
example, a rotation g by an irrational multiple of π about an axis A in R3 . If gn
is the n-th iterate of g, then gn converges on A ∪ {∞} but nowhere else, and this
case is not covered by Theorem 10.1 in its present form. The following result is
the desired generalization of Theorem 10.1 (see [22]).
Theorem 10.2. Let gn be any sequence of Möbius maps acting on RN ∞ , and
suppose that the gn converge pointwise on the non-empty set C (and nowhere
else) to the function g. Then one of the following occurs:
(a) g is the restriction of some Möbius map to C;
(b) C = RN ∞ , and g takes exactly two values, one of which is taken exactly once;
(c) g is constant on C.
Suppose that we are in case (b); thus there is some x0 such that, say, gn → α
pointwise on RN ∞ \{x0 }, and gn (x0 ) → β, where α 6= β. In these circumstances
one can show that gn → α locally uniformly on RN ∞ \{x0 }. Now consider the
sets C in (a). If N = 2 then, by Theorem 10.1 C is either C∞ or some double-
ton {z1 , z2 }. However, in higher dimensions other possibilities can occur and a
complete description of the possibilities is given in our next result.
Theorem 10.3. Suppose that g and g1 , g2 , . . . are Möbius maps acting on RN
∞,
and that gn converges pointwise to g on, and only on, the non-empty set C. Then
there is a subspace V of RN of dimension d, where d 6= N − 1, and a Möbius
map h, such that C = h(V ∪ {∞}). Further, every set C of this form arises as
the set C for some sequence gn .
These results give insight into the two-dimensional case. If N = 2, then V in
Theorem 10.3 is {0} or C and we obtain the following corollary which explains
part of the conclusion in Theorem 10.1. For other results of this type in all
dimensions, see also [2].
Corollary 10.4. Suppose that g and g1 , g2 , . . . are complex Möbius maps, and
that gn → g on, and only on, the non-empty set C. Then either C = C∞ or C
contains exactly two points.
11. Self-maps of a domain

Many results about the convergence of continued fractions are proved by con-
sidering analytic functions f1 , f2 , . . . that map a domain D into itself, and then
imposing conditions that imply that f1 · · · fn converges locally uniformly on D
to a constant. Theorems 4.6 and 4.7 are general results of this kind, but one
might hope for better results in the special case when the fj are Möbius maps
(see, for example, [60] and [76]). The earliest result of this type was Worpitzky’s
Theorem; if sn (z) = an /(1 + z), where |an | ≤ 1/4, then K(an |1) converges.
Later, this was generalized by Śleszyński, and by Pringsheim, who proved that
if sn (z) = an /(bn + z), where |bn | ≥ 1 + |an |, then K(an |bn ) converges. As the
condition |bn | ≥ |an | + 1 is equivalent to sn (D) ⊂ D, which makes sense in all
dimensions, we prefer the following formulation.
Theorem 11.1. Suppose that sn (z) = an /(bn + z) and sn (D) ⊂ D. Then

K(an |bn ) converges.
In the situation described in Theorem 11.1, we have
∞
\
(11.1) D ⊃ s1 (D) ⊃ s1 s2 (D) ⊃ · · · ⊃ s1 · · · sn (D) = ∆,
n=1
where ∆ is a compact disc of radius, say r∞ , so there are two possibilities:
(a) r∞ > 0 (the limit-circle case), and
(b) r∞ = 0 (the limit-point case).
Recently the author [15] has given the following purely geometric characterization
of those nested sets of discs that arise as the discs s1 · · · sn (D) from an application
of Theorem 11.1.
Theorem 11.2. Suppose that z0 , z1 , . . . is a sequence of points in D with z0 = 0,
and that K0 , K1 , . . . is a sequence of closed discs in D with D = K0 ⊃ K1 ⊃
K2 ⊃ · · · . Let Cn = ∂Kn . Then there are some sn as in Theorem 11.1 with
Kn = s1 · · · sn (D), and zn = s1 · · · sn (0), if and only if for each n,
(i) zn lies between Cn and Cn+1 , and
(ii) zn and zn+1 are inverse points with respect to Cn+1 .
This a result in inversive geometry (that is, the geometry of C∞ with respect
to the group M). However, as each circle in C is the boundary of a plane
in hyperbolic 3-space H3 , it can equally well be thought of as a result about
disjoint planes in hyperbolic 3-space. In [51, p. 93], it is stated that “few cases
are known in which the limit-circle case actually occurs”; however, Theorem
11.2 now enables us to construct as many examples of this type as we wish. It
is perhaps worth remarking that this argument shows that the limit-circle case
is best described in terms of the geometry of the discs, and not in terms of the
coefficients of the sn .
In 1965 Hillam and Thron ([47]) proved the following generalization of Theo-
rem 11.1.
Theorem 11.3. Let D be an open disc in C that contains 0. Suppose that sn ,
n = 1, 2, . . ., are Möbius maps such that for each n, sn (D) ⊂ D and sn (∞) = 0.
Then s1 · · · sn converges locally uniformly on D to some constant.
From a topological viewpoint, it seems reasonable to expect the hypothesis
sn (∞) = 0 to be stable under small perturbations, and even more, to allow
each of these points to lie in given compact subsets that are interior to, and
exterior to, the disc D. This is proved when 0 is replaced by a compact subset
of D in [55], but the result is also true when ∞ is replaced by a compact subset,
and in this more general form it is also true in all dimensions. Thus we have the
following result.
Theorem 11.4. Let B be an open ball in RN , and let K1 and K2 be compact

subsets of the interior, and the exterior of B, respectively. Suppose that for each
n, sn is a Möbius map acting on RN ∞ such that sn (B) ⊂ B and sn (K2 ) ∩ K1 6= ∅.
Then s1 · · · sn converges locally uniformly on B to some constant.
Most of the proof of Theorem 11.4 is given in [20], and the completion of the
proof has been given recently by I. M. Short. The proof uses the same general
idea of the proof of Theorem 11.3 as given in [47], [51] and [55], but the algebraic
approach is replaced by a purely geometric approach (which is shorter and less
computational). Both proofs are divided into two cases depending on whether a
certain infinite product converges or diverges to zero. However, this division does
not seem to correspond to the geometric division into the limit-circle or limit-
point cases, and there is a strong analogy here between this and an important
idea in the theory of discrete Möbius groups which we shall now describe. An
important part of discrete group theory is the study of groups that preserve the
unit ball BN , and any such group G is Q said to be of convergence or divergence
type according to whether the product g∈G |g(0)| converges or diverges. There
is also a division of such groups into groups of the first kind and second kind
depending on a certain geometric criterion (which we shall not go into here). The
point is that there are two important dichotomies: one in terms of an infinite
product and another in terms of a more elementary geometric condition. In many
circumstances it is the classification into convergence or divergence type that is
the more important, and this raises the question of whether the convergence or
divergence of the infinite product arising in the proofs of Theorems 11.3 and 11.4
is perhaps an equally important issue as the distinction between the limit-point
and limit-circle cases.
There is another result of this type in which hyperbolic geometry plays a crucial
role, and in which the conclusion is the uniform convergence of the composition
maps on the closed disc (see [21]).
Theorem 11.5. Suppose that f1 , f2 , . . . is a sequence of Möbius maps each of
which maps the open unit ball BN in RN into itself, and let rj be the hyperbolic
radius of the ball fj (BN ) (possibly rj = +∞). If the sum
X ∞
(11.2) e−rn
n=1
diverges, then f1 · · · fn converges uniformly on BN to a constant. This P result

is best possible in the sense that if Rn are positive numbers such that n e−Rn
converges, then there is a sequence fn of Möbius maps of BN into itself such
that fj (BN ) has hyperbolic radius Rj , and such that f1 · · · fn does not converge
pointwise on BN to a constant.
We remark that if (in the two-dimensional case) f is a Möbius map of D into
a proper subset of itself, then f is not a hyperbolic isometry of D, but as it is
a hyperbolic isometry of H3 we can obtain more information by working in H3

than in D. Of course, if f satisfies f (D) = D, then f is a hyperbolic isometry of
both D and H3 .
As we noted earlier, the class of circles in C∞ may be indentified with the class of
hyperbolic planes in H3 , and it is helpful to work with the (shortest) hyperbolic
distance between two disjoint planes. This distance satisfies the ‘reverse triangle
inequality’ (see [21], [27] and [28]), and it is essentially the same as the modulus
of the ring domain bounded by the two circles. Thus in this way Theorem 11.5
is seen to be closely related to a result of Baker and Rippon ([10, Theorem 3.1])
which gives a sufficient condition for the convergence of f1 · · · fn to a constant
(where in this result the fj need not be Möbius maps). In the context of Möbius
maps, the ‘reverse triangle inequality’ mentioned above is equivalent to the stan-
dard inequality between moduli of ‘adjacent’ ring domains.
The ideas in the proof of Theorem 11.5 suggest that, in the case N = 2, the
main issue is the relative location of the nested sequence of discs f1 · · · fn (B),
and that the best way to consider their relative positions is in H3 . If these planes
have a common orthogonal geodesic (that is a common hyperbolic centre in D),
then the divergence of the series (11.2) is a necessary and sufficient condition for
f1 · · · fn to tend to a constant, and in this case we have equality in the ‘reverse
triangle inequality’ at every stage. However, if the planes are ‘scattered’ in such a
way that the collection of common orthogonal segments that join each successive
pair of planes are far from lying on a single geodesic, then the reverse triangle
inequality is far from equality and the two conditions are no longer equivalent.
Once again, the geometry of H3 reveals information that is harder to find in C∞ .
Theorem 11.5 is rather delicate for it shows that if fn (B) has hyperbolic radius
log n then f1 · · · fn converges to a constant, whereas for any positive ε there are
examples in which fn (B) has radius (1 + ε) log n but f1 · · · fn does not converge
to a constant. Notice that in Theorem 11.5 there is no restriction on where the
fj (B) are located in B; the only thing that matters is their hyperbolic radii. In
particular, we can allow the fj (B) to accumulate very rapidly at, or even be
tangent to, the sphere ∂B. In this discussion we adopt the usual convention that
a horoball is a ball in B that is internally tangent to ∂B and, by convention, this
is a hyperbolic ball of infinite hyperbolic radius whose hyperbolic centre lies at
the point of tangency. We know from the theory of discrete groups that horoballs
play an important role in the hyperbolic geometry of B.
A Euclidean result related to Theorem 11.5 is given in [58, Theorem 1.4A], where
it
Pis proved that if dn is the Euclidean distance between fn (D) and ∂D, and if
n dn = +∞, then f1 · · · fn converges uniformly on D to a constant. However,
Theorem 11.5 is much stronger than this for the hyperbolic radius rn of fn (D)
satisfies rn ≤ log 1/dn and hence exp(−rn ) ≥ dn . Briefly, Theorem 11.5 is
applicable regardless of the location of the fj (D); however, [58, Theorem 1.4A]
is applicable only if the fj (D) move out to ∂D sufficiently slowly.
Finally, we remark that one of the important consequences of the Hillam-Thron

Theorem is the Parabola Theorem, and recently I. M. Short has shown that this
too makes sense and is true in higher dimensions.
12. Concluding remarks

There is little doubt that one can best understand the action of Möbius maps
when they are identified as the group of isometries of hyperbolic space. As
continued fractions arise as sequences of Möbius maps, these too must at least
be considered from this point of view. There is a great need for explicit examples
in which the behaviour of a continued fraction can be analyzed and understood
through its action in hyperbolic space and, as we have already mentioned, it
may be that the question of conditional convergence in the Stern-Stolz Theorem
can be better understood from this point of view. There is also a need for more
‘counter-examples’ in the subject to show what is not generally true.
It seems likely that there are few, if any, convergence results in continued fractions
that are not true for a generally convergent sequence of Möbius maps. Of course,
numerical results on rates of convergence, and so on, are not invariant under
conjugation and these are tied specifically to the notion that ∞ is a special
point. However, this is no reason why we should not seek to establish these
results in the context of generally convergent sequences using the chordal metric.
One of the main ways forward in this area would be to establish as much as
possible of the theory of continued fractions to generally convergent sequences of
Möbius maps in all dimensions.
Finally, there is work to be done in relating continued fractions in higher dimen-
sions as discussed in this paper with those already appearing in the literature
(see. for example, [71], [86] and [87]). Real N-dimensional continued fractions are
introduced algebraically as follows. The inversion in the unit sphere is denoted
by x 7→ x∗ = x/|x|2 , and a continued fraction is a repeated application of maps
of the form x 7→ a(x + b)∗ , where a is a real constant and b is a vector. These
continued fractions are a special case of those discussed in this paper, and while
they may (or may not) be easier to handle than those discussed here, this is prob-
ably not the ‘correct’ definition. The map x 7→ a(x + b)∗ is the map x 7→ (x + b)∗
followed by the map x 7→ ax. Now according to the theory outlined in this paper,
this second map should be the most general Möbius map that fixes 0 and ∞,
and this is a map of the form x 7→ aM x, where a > 0 and M is an orthogonal
matrix. Thus from this point of view, the N-dimensional continued fractions
should arise from a repeated composition of maps x 7→ A(x + b)∗ , where b is a
vector, and A is a positive scalar multiple of an orthogonal matrix. Maps of the
form x 7→ a(x + b)∗ also give rise to continued fractions that are in CN , and it
remains to be seen how these may best be interpreted geometrically.
References
1. W. Abikoff, The bounded model for hyperbolic 3-space and a quaternionic Uniformization
Theorem, Math. Scand. 54 (1984), 5–16.
2. B. Aebischer, The limiting behaviour of sequences of Möbius transformations, Math. Zeit.
205 (1990), 49–59.
3. B. Aebischer, Stable convergence of sequences of Möbius transformations, in: Computa-
tional Methods and Function Theory (CMFT ’94), eds. R. M. Ali, St. Ruscheweyh and E.
B. Saff, World Scientific Pub. Co., 1995, 1–21.
4. L. V. Ahlfors, Hyperbolic motions, Nagoya Math. J. 29 (1967), 163–166.
5. L. V. Ahlfors, Möbius Transformations in Several Dimensions, Univ. Minnesota Lecture
Notes, Minnesota, 1981.
6. L. V. Ahlfors, On the fixed points of Möbius transformations in Rn , Ann. Acad. Sci. Fenn.
Ser. A. 1. Math. 10 (1985), 15–27.
7. L. V. Ahlfors, Möbius Transformations and Clifford Numbers, in Differential Geometry
and Complex Analysis, eds. I. Chavel, and H.M. Farkas, Springer-Verlag, Berlin (1985),
65–73.
8. L. V. Ahlfors, Clifford numbers and Möbius transformations in Rn , Clifford algebras and
their applications in mathematical physics, (Canterbury, 1985), 167–175, NATO Adv. Sci.
Inst. Ser. C Math. Phy. Sci., 183, Reidel, Dordrecht, 1986.
9. L. V. Ahlfors, Möbius transformations in Rn expressed through 2 × 2 matrices of Clifford
numbers, Complex Variables Theory Appl. 5 (1986), 215–224.
10. I. N. Baker and P. J. Rippon, Towers of exponents and other composite maps, Complex
Variables 12 (1989), 181–200.
11. A. F. Beardon and B. Maskit, Limit points of Kleinian groups and finite-sided fundamental
polyhedra, Acta Math. 132 (1974), 1–12.
12. A. F. Beardon, The Geometry of Discrete Groups, Springer-Verlag, 1983.
13. A. F. Beardon and J. B. Wilker, The norm of a Mobius transformation, Math. Proc. Camb.
Phil. Soc. 96 (1984), 301–308.
14. A. F. Beardon, Iteration of Rational Functions, Springer-Verlag, 1991.
15. A. F. Beardon, The geometry of Pringsheim’s continued fractions, Geometriae Dedicata
84 (2001), 125–134.
16. A. F. Beardon and D. Minda, Sphere-preserving maps in inversive geometry, Proc. Amer.
Math. Soc. 130 (2002), 987–998.
17. A. F. Beardon, T. K. Carne, D. Minda, and T. W. Ng, Random iteration of analytic maps,
to appear in Ergodic Th. & Dyn. Systems.
18. A. F. Beardon and L. Lorentzen, Continued fractions and restrained sequences of Möbius
maps, to appear in Rocky Mount. J. Math..
19. A. F. Beardon, Continued fractions, Möbius transformations and Clifford algebras, to
appear in Bull. London Math. Soc..
20. A. F. Beardon, The Hillam-Thron Theorem in higher dimensions, to appear in Geom.
Dedicata.
21. A. F. Beardon, Repeated compositions of analytic maps, to appear in Comp. Methods and
Function Theory.
22. A. F. Beardon, The pointwise convergence of Möbius maps, preprint, 2002.
23. R. Benedetti and C. Petronio, Lectures on hyperbolic geometry, Universitext, Springer-
Verlag, 1992.
24. C. Cao and P. L. Waterman, Conjugacy invariants of Möbius groups, Quasiconformal
mappings and analysis (Ann Arbor, MI, 1995), Springer, New York, 1998, 109–139.
25. L. Carleson and T. W. Gamelin, Complex Dynamics, Springer-Verlag, 1993.
26. S. S. Chen and L. Greenberg, Hyperbolic spaces, in: Contributions to Analysis, eds L. V.
Ahlfors, I. Kra, B. Maskit, and L. Nirenberg, Academic Press, 1974.
27. H. S. M. Coxeter, Inversive distance, Ann. Mat. Pura Appl. 71 (1966), 73–83.
28. H. S. M. Coxeter, The inversive plane and hyperbolic space, Hamburger Math. Abh. 29
(1966), 217–242.
29. H. S. M. Coxeter, The Lorentz group and the group of homographies, Proc. Intern. Conf.
Theory of Groups (Canberra 1965), Gordon and Breach, New York, 1967.
30. H. Davenport, The Higher Arithmetic, Sixth Edn, Cambridge Univ. Press, 1995.
31. J. D. de Pree and W. J. Thron, On sequences of Moebius transformations, Math. Zeit. 80
(1962), 184–193.
32. J. Dieudonné, Treatise on Analysis, Vol II, Academic Press, 1970.
33. P. Du Val, Homographies, Quaternions and Rotations, Clarendon Press, Oxford, 1964.
34. W. Fenchel, Elementary Geometry in Hyperbolic Space, De Gruyter Studies in Mathemat-
ics 11, de Gruyter, 1989.
35. L. R. Ford, A geometrical proof of a theorem of Hurwitz, Proc. Edinburgh Math. Soc. 35
(1917), 59–65.
36. L. R. Ford, Rational approximations to irrational complex numbers, Trans. Amer. Math.
Soc. 19 (1918), 1–42.
37. L. R. Ford, Fractions, Amer. Math. Monthly 45 (1938), 586–601.
38. L. R. Ford, Automorphic functions, Chelsea Pub. Co., Second Edition, 1951.
39. F. W. Gehring and G. J. Martin, The Matrix and Chordal Norms of Möbius Transforma-
tions, Complex Analysis, Birkhäuser, Basel, 1988, 51–59.
40. F. W. Gehring and G. J. Martin, Inequalities for Möbius transformations and discrete
groups, J. Reine Angew. Math. 418 (1991), 31–76.
41. J. Gill, Infinite compositions of Möbius transformations, Trans. American Math. Soc. 176
(1973), 479–487.
42. P. G. Gormley, Stereographic projection and the linear fractional group of transformations
of quaternions, Proc. Royal Irish Acad. Sect A 51 6 (1947), 67–85.
43. L. Greenberg, Discrete subgroups of the Lorentz group, Math. Scand. 10 (1962), 85–107.
44. W. R. Hamilton, On continued fractions in quaternions, Phil. Mag., iii (1852), 371–373; iv
(1852), 303; v (1853), 117–118, 236–238, 321–326.
45. W. R. Hamilton, On the connexion of quaternions with continued fractions and quadratic
equations, Proc. Royal Irish Acad. 5 (1853), 219–221, 299–301.
46. P. Henrici, Applied and Computational Complex Analysis, Vol. 2, Special Functions, In-
tegral Transforms, Asymptotics and Continued Fractions, Wiley, New York, 1977.
47. K. L. Hillam and W. J. Thron, A general convergence criterion for continued fractions
K(an /bn ), Proc. Amer. Math. Soc. 16 (1965), 1256–1262.
48. L. Jacobsen, General convergence of continued fractions, Trans. Amer. Math. Soc. 294
(1986), 477–485.
49. L. Jacobsen and W. J. Thron, Limiting structures for sequences of linear fractional trans-
formations, Proc. Amer. Math. Soc. 99 (1987), 141–146.
50. L. Jacobsen, Nearness of continued fractions, Math. Scand. 60 (1987), 129–147.
51. W. B. Jones and W. J. Thron, Continued Fractions: Analytic Theory and Applications,
Encyclopedia of Mathematics and its Applications, 11, Addison-Wesley, Reading, Mass,
(1980). Now distributed by Cambridge Univ. Press, New York.
52. I. L. Kantor and A. S. Solodovnikov, Hypercomplex Numbers, Springer-Verlag, 1989.
53. R. E. Lane, The convergence and values of periodic continued fractions, Bull. Amer. Math.
Soc. 51 (1945) 246–250.
54. L. Lorentzen, Compositions of contractions, J. Comput. Appl. Math. 32 (1990), 169–178.
55. L. Lorentzen and H. Waadeland, Continued Fractions and Some of its Applications, North-
Holland, 1992.
56. L. Lorentzen, The closure of convergence sets for continued fractions are convergence sets,
Proc. Ednburgh Math. Soc., 37 (1993), 39–46.
57. L. Lorentzen, Divergence of continued fractions related to hypergeometric series, Math.
Comp. 62 (1994), 671–686.
58. L. Lorentzen, A Convergence Property for Sequences of Linear Fractional Transformations,
Continued fractions and orthogonal functions (Loen, 1992), 281–304, Lecture Notes in Pure
and Appl. Math., 154, Dekker, New York, 1994.
59. L. Lorentzen, A convergence question inspired by Stieltjes and by value sets in continued
fraction theory, J. Comp. Appl. Math. 65 (1995), 233–251.
60. L. Lorentzen, Convergence of compositions of self-mappings, Ann. Univ. Marie Curie
Sklodowska A 53 13 (1999), 121–145.
61. M. Mandell and A. Magnus, On Convergence of Sequences of Linear Fractional Transfor-
mations, Math. Zeit. 115 (1970), 11–17.
62. J. W. Magnus, Non-Euclidean Tessselations and their Groups, Academic Press, 1974.
63. J. Milnor, Hyperbolic geometry — the first 150 years, Bull. Amer. Math. Soc. 6 (1982),
9–24.
64. P. J. Nicholls, The Ergodic Theory of Discrete Groups, London Math. Soc. Lectures Notes
143, Camb. Univ. Press, 1989.
65. J. F. Paydon and H. S. Wall, The continued fraction as a sequence of linear transformations,
Duke Math. Jour. 9 (1942), 360–372.
66. G. Piranian and W. J. Thron, Convergence properties of sequences of linear fractional
transformations, Michigan Math. J. 4 (1957), 129–135.
67. O. Perron, Die Lehre von den Kettenbrüchen, Band I, Teubner, Stuttgart, 1954.
68. O. Perron, Die Lehre von den Kettenbrüchen, Band II, Teubner, Stuttgart, 1957.
69. I. R. Porteous, Topological Geometry, Van Nostrand, 1969.
70. J. G. Ratcliffe, Foundations of hyperbolic manifolds, Graduate Texts 149, Springer-Verlag,
1994.
71. D. E. Roberts, On a representation of vector continued fractions, J. Comput. Appl. Math.
105 (1999), 453–466.
72. H. Schwerdtfeger, Moebius transformations and continued fractions, Bull. Amer. Math.
Soc. 52 (1946), 307–309.
73. T. J. Stieltjes, Recherches surles fractions continues, Ann. Fac. Sci. Toulouse, 8J (1894),
1–122; 9A (1894), 1–47; Oeuvres, 2, 402–566.
74. N. Steinmetz, Rational Iteration, De Gruyter Studies in Mathematics 16, de Gruyter, 1993.
75. W. J. Thron, Convergence regions for continued fractions and other infinite processes,
Amer. Math. Monthly 68 (1961), 734–750.
76. W. J. Thron, Convergence of Sequences of Linear Fractional transformations and of Con-
tinued Fractions, J. Indian Math. Soc. 27 (1963), 103–127.
77. W. J. Thron and H. Waadeland, Convergence questions for limit periodic continued frac-
tions, Rocky Mountain J. Math. 11 (1981), 641–657.
78. W. J. Thron and H. Waadeland, Modifications of continued fractions, a survey, Analytic
theory of continued fractions (Loen, 1981), 38–66, Lecture Notes in Math.. 932, Springer,
Berlin-New York, 1982.
79. W. J. Thron, Continued fraction identities derived from the invariance of the crossratio
under l.f.t., Analytic theory of continued fractions, III (Redstone, CO, 1988), 124–134,
Lecture Notes in Math.,1406, Springer, Berlin, 1989.
80. H. Waadeland, Tales about tails, Proc. Amer. Math. Soc. 90 (1984), 57–64.
81. M. Wada, Conjugacy invariants of Möbius transformations, Complex Variables Theory

Appl. 15 (1990), 125–133.
82. H. S. Wall, Continued fractions and cross-ratio groups of Cremona transformations, Bull.
Amer. Math. Soc. 40 (1934), 587–592.
83. H. S. Wall, Analytic Theory of Continued Fractions, Chelsea Pub. Co.,1948.
84. P. Waterman, Möbius transformations in several dimensions, Adv. Math. 101 (1993), 87–
113.
85. J. B. Wilker, Inversive geometry, in: The Geometric Vein, eds. C. Davis, B. Grünbaum
and F. A. Scherk, Springer-Verlag, 1981, 379–442.
86. P. Wynn, Continued fractions whose coefficients obey a non-commutative law of multipli-
cation, Arch. Rational Mech. Anal. 12 (1963), 273–312.
87. P. Wynn, Vector continued fractions, Linear Algebra Appl. 1 (1968), 357–395.
Alan F. Beardon E-mail: [email protected],uk

Address: Centre for Mathematical Sciences, Wilberforce Road, Cambridge CB3 0WB, U.K.

Continued Fractions, Discrete Groups and Complex Dynamics

Uploaded by

Copyright:

Available Formats

Continued Fractions, Discrete Groups and Complex Dynamics

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Continued Fractions, Discrete Groups and Complex Dynamics

Uploaded by

Copyright:

Available Formats

Computational Methods and Function Theory

Volume 1 (2001), No. 2, 535–594

Continued Fractions, Discrete Groups and Complex

Abstract. It is well-known that a continued fraction may be regarded as a se-

Keywords. Continued fractions, Möbius maps, complex dynamics, discrete

on continued fractions will be true regardless of the dimension. In fact, we al-

This implies that any given sequence f1 , f2 , . . . of Möbius transformations arises

given result on continued fractions is really a result about a general sequence

similar phenomenon holds in continued fraction theory. Moreover, other results

3. Complex Möbius maps

by ϕ. This map provides us with the chordal metric on C∞ , namely

We define the norm kgk of g by

kgk2 = |a|2 + |b|2 + |c|2 + |d|2

so that if gn → f then gn−1 → f −1 . Likewise,

σ0 (fn gn , f g) ≤ σ0 (fn gn , f gn ) + σ0 (f gn , f g) ≤ σ0 (fn , f ) + L(f ) σ0 (gn , g),

so that the map (f, g) 7→ f g is continuous. Results about the convergence of

(i) the topology induced by σ0 ,

As each of these three topologies arises out of a standard topological construction,

The vector space of all 2 × 2 matrices (singular and non-singular) is equipped

An immediate consequence of this is that the topology induced on M by the

Then kF k ≤ kGk, and kF k2 kF − Gk2 → 2 as ε → 0. Now let f = Φ(F ) and

Proof. Suppose first that gn (w) → w as n → ∞ when w ∈ {0, 1, ∞}. We may

converge or diverge together, where Pn = A1 · · · An .

Definition 3.9. Given a Möbius g we define trace2 (g) to be [trace(A)]2 , where A

(2) elliptic if it is conjugate to z 7→ kz for some k with |k| = 1, k 6= 1;

4. The stability of continued fractions

in a given compact subset K of D. Next, construct subdomains Ω1 and Ω2 of D

Proof of Theorem 4.6 when D is a disc. We may take D to be the open

as n → ∞ so that g(z1 ) = g(z2 ).

a number R in (1/2, 1) by σ(1/2, R) = 1/2, and then let Q = {z : |z| ≤ R}. It

5. Möbius maps and hyperbolic geometry

Pythagoras’ Theorem: if a right-angled triangle has hypotenuse of length c, and

where this calculation is carried out in the algebra of quaternions. It is not

action in H3 , g leaves the hyperbolic plane bounded by R ∪ {∞} invariant so

Theorem 5.4. For any Möbius g the following are equivalent:

≤ L(g −1 ) σ g(z), g(w) = σ g(z), g(w) .

It is now clear that (d) follows.

It is useful to have the ball-model of hyperbolic three-space available, and this is

As ‘almost all’ Möbius transformations are loxodromic the dynamics described

If we let un = 0 and vn = ∞ then we see immediately that if a continued fraction

Of course, the convergence is in terms of the chordal metric. Also, by rejecting

Proof of Theorem 6.6. Suppose first that gn converges generally to α, where

Let un be z if gn (z) is closer to α than gn (z 0 ), and z 0 otherwise; thus gn (un ) → α.

this section we introduce a stronger notion of divergence which can only be

In terms of continued fractions, we recall the numbers An and Bn from the

Thus a compact subgroup of M is a uniformly Lipschitz family.

Proof of Theorem 7.8. The triangle inequality gives

σ0 (s1 · · · sn , sn ) ≤ σ0 (s1 · · · sn , ss2 · · · sn )

thus, by right invariance,

σ0 (s1 · · · sn , sn ) ≤ σ0 (s1 , s) + L(s)σ0 (s2 , s) + · · · + L(sn−1 )σ0 (sn , s).

With (7.2) this yields

σ0 (s1 · · · sn , sn ) ≤ L(G) σ0 (s1 , s) + σ0 (s2 , s) + · · · + σ0 (sn , s) .

σ0 (sk sk+1 · · · sk+n−1 , sn ) ≤ L(G) σ0 (sk , s) + σ0 (sk+1 , s)

L(s1 · · · sK+m ) ≤ L(s1 · · · sK−1 )L(sK · · · sK+m )

It is now easy to complete the proof, for we have

8. Möbius maps in higher dimensions

As usual, we adjoin a point ∞ to RN +1 to form the space RN +1

is the boundary of the unit ball B N +1