Hyperbolic Groups

Lecture Notes
James Howie
Heriot-Watt University,
Edinburgh EH14 4AS
Scotland
[email protected]
Introduction
This is a revised and slightly expanded set of notes based on a course of 4 lectures
given at the postgraduate summer school Groups and Applications at the Univer-
sity of the Aegean in July 1999. I am very grateful to the University of the Aegean
- in particular to Vasileios Metaftsis - for the opportunity to give the lectures and
to publish the notes, and for their warm hospitality during the summer school.
The subject matter is hyperbolic groups - one of the main objects of study in
geometric group theory. Geometric group theory began in the 1980s with work
of Cannon, Gromov and others, applying geometric techniques to prove algebraic
properties for large classes of groups. In this, the subject follows on from its ancestor,
Combinatorial Group Theory, which has roots going back to the 19th century
(Fricke, Klein, Poincare). It adds yet another layer of geometric insight through the
idea of treating groups as metric spaces, which can be a very powerful tool.
In a short lecture course I could not hope to do justice to this large and im-
portant subject. Instead, I aimed to give a gentle introduction which would give
some idea of the avour. I have tried to prepare these notes in the same spirit.
One diculty one faces when approaching this subject is the fact that there are
several (equivalent) denitions. To give them all would be time-consuming, to prove
equivalence more so. I have settled for giving just two denitions, each motivated
by the corresponding geometric properties of the hyperbolic plane, and ignoring the
question of equivalence.
The rst lecture deals in general with groups seen as metric spaces, introduces
the idea of quasi-isometry, and illustrates the ideas using the study of growth of
groups. The second lecture gives the thin-triangles denition of hyperbolic group,
and uses it to give a simple proof that hyperbolic groups are nitely presented.
(This is a consequence of a more far-reaching result of Rips, to which we return
later.) The third lecture introduced Dehn diagrams and isoperimetric inequalities,
gives the linear isoperimetric inequality denition of hyperbolic groups, and indi-
cates how to use this to obtain solutions of the word and conjugacy problems for
hyperbolic groups. The nal lecture was designed to give a glimpse of two slightly
more advanced aspects of the subject, namely the Rips complex and the boundary
of a hyperbolic group. In practice, I ran out of time and settled for discussing only
the Rips complex. However, I have included a section on the hyperbolic boundary
in these notes for completeness.
I hope that these notes will encourage readers to learn more about the subject.
The principal references in this area are the original texts of Gromov [7, 8, 9], but
several authors have worked on producing more accessible versions. I found [1, 2, 4]
useful sources.
2
Lecture 1: Groups as metric spaces
Geometric group theory is the study of algebraic objects (groups) by regarding them
as geometric objects (metric spaces). This idea seems unusual at rst, but in fact
is very powerful, and enables us to prove many theorems about groups that satisfy
given geometric conditions.
How can we make a group G into a metric space? Choose a set S of generators
for G. Then every element of G can be expressed as a word in the generators:
g = x

1
1
x

2
2
. . . x

n
n
, where x
1
, x
2
, . . . , x
n
S and
1
,
2
, . . . ,
n
= 1. The natural
number n is called the length of this word. If g, h G then we dene d
S
(g, h) to be
the length of the shortest word representing g
1
h.
Lemma 1 d
S
is a metric on the set G.
Proof. By denition, d
S
(g, h) N. In particular, d
S
(g, h) 0. Moreover, d
S
(g, h) =
0 if and only if g
1
h is represented by the empty word (of length 0). But the empty
word represents the identity element of G, so d
S
(g, h) = 0 g = h.
If x

1
1
x

2
2
. . . x

n
n
is a word of minimum length representing g
1
h, then h
1
g =
x

n
n
. . . x

2
2
x

1
1
, so d
S
(h, g) d
S
(g, h). Similarly, d
S
(g, h) d
S
(h, g), so d
S
(h, g) =
d
S
(g, h).
Let x

1
1
x

2
2
. . . x

n
n
and y

1
1
y

2
2
. . . y

m
m
be words of minimum length representing
g
1
h and h
1
k respectively. Then
x

1
1
x

2
2
. . . x

n
n
y

1
1
y

2
2
. . . y

m
m
is a word (not necessarily of minimum length) representing g
1
h h
1
k = g
1
k.
Hence
d
S
(g, k) d
S
(g, h) +d
S
(h, k)
(the triangle inequality).
Remarks
1. The metric d
S
on G is called the word metric on G with respect to S. It takes
values in N. This distinguishes it from the metrics associated to more standard
geometric objects (euclidean or hyperbolic space, surfaces, manifolds), which
take values in R
+
. However, if the units of measurement are very small, or
equivalently if we look at G from a great distance, then we cannot distinguish
between the discrete-valued metric d
S
and some continuous-valued approxima-
tion to it. This can all be made precise and used to compare G with more
familiar and continuous metric spaces such as euclidean or hyperbolic space.
3
2. The metric d
S
is related to the Cayley graph (G, S) in a natural way: we can
identify G with the set of vertices of (G, S), and two vertices g, h (g ,= h)
are adjacent in if and only if g
1
h S or h
1
g S, in other words if and
only if d
S
(g, h) = 1. More generally, if g, h are joined by a path of length n in
(G, S), then we can express g
1
h as a word of length n in S, so d
S
(g, h) n.
The converse is also true: if g
1
h can be expressed as a word of length n in
S, then g, h can be joined by a path of length n in (G, S). Hence d
S
(g, h) is
precisely the length of a shortest path (a geodesic) in (G, S) from g to h.
3. The metric d
S
depends in an essential way on the choice of generating set S.
For example, if we take S = G then d
S
is just the discrete metric: d
S
(g, h) = 1
whenever g ,= h. This is not an interesting metric, and can not be expected to
give interesting algebraic information about G. To avoid this kind of problem,
we restrict attention to nite generating sets S. In particular, all groups from
now on will be nitely generated.
4. Even with the restriction to nite generating sets, the metric depends on the
choice of S. In particular, for any g, h G, we can choose a generating set
S that contains g
1
h, so that d
S
(g, h) 1. However, despite such obvious
problems, the dependence can be shown to be limited in a very real sense, so
that if we look at G from a distance then the eects of changing generating
set become less apparent. In other words, there are many properties of the
metric space (G, d
S
) that are independent of the choice of S. These properties
are the objects of study in geometric group theory.
Quasi-isometry
An isometry from one metric space (X, d) to another metric space (X

, d

) is a map
f : X X

such that
d

(f(x), f(y)) = d(x, y) x, y X.

It follows that f is continuous and injective. If f is also surjective, then f
1
:
X

X is also an isometry, and in this case we say that the metric spaces (X, d)
and (X

, d

) are isometric. This is an equivalence relation between metric spaces.

Isometric metric spaces are regarded as being the same, just as isomorphic groups
or rings, or homeomorphic topological spaces, are the same.
Quasi-isometry is a weaker equivalence relation between metric spaces that meets
the requirements of geometric group theory by neglecting ne detail and concentrat-
ing on the large picture seen from a distance as mentioned in the remarks above. It
is dened in an analogous way. Let , k be positive real numbers. A map f : X X

4
is a (, k)-quasi-isometry if
1

d(x, y) k d

(f(x), f(y)) d(x, y) +k x, y X.

Now f is not in general continuous or injective. If f is almost surjective, in the sense
that every point of X

is a bounded distance from the image of f, then there is a

(

, k

)-quasi-isometry f

: X

X for some

, k

, that is almost an inverse to f.

In this case we say that the metric spaces (X, d) and (X

, d

) are quasi-isometric.
Examples
1. (Z, d) and (R, d) are quasi-isometric, where d is the usual metric: d(x, y) =
[x y[. The natural embedding Z R is an isometry, so a (1, 0)-quasi-
isometry. It is not surjective, but each point of R is at most
1
2
away from Z.
We can dene a (1,
1
2
)-quasi-isometry f : R Z by f(x) = x rounded to the
nearest integer.
2. We can generalise the above example. Let G be a group with a nite generating
set S, and let = (G, S) be the corresponding Cayley graph. We can regard
as a topological space in the usual way, and indeed we can make it into
a metric space by identifying each edge with a unit interval [0, 1] R and
dening d(x, y) to be the length of the shortest path joining x to y. This
coincides with the path-length metric d
S
when x and y are vertices. Since
every point of is in the
1
2
-neighbourhood of some vertex,we see that (G, d
S
)
and ((G, S), d) are quasi-isometric for this choice of d.
3. Every bounded metric space is quasi-isometric to a point.
4. Z Z is quasi-isometric to the euclidean plane E
2
.
5. If S and T are nite generating sets for a group G, then (G, d
S
) and (G, d
T
)
are quasi-isometric. Indeed, let be the maximum length of any element of
S expressed as a word in T or vice versa. Then the identity map G G is a
(, 0)-quasi-isometry form (G, d
S
) to (G, d
T
) and vice versa. Hence, when we
are discussing quasi-isometry in the context of nitely generated groups, we
can omit mention of the particular generating set, and make statements like
G is quasi-isometric to H without ambiguity.
Growth
Suppose that G is a nitely generated group, and that S is a nite generating set
for G. We dene the growth function =
S
: N N for G with respect to S by
(n) = [ g G [ d
S
(g, 1) n [.
5
In other words, (n) is the number of points contained in a ball of radius n in
(G, d
S
). The growth function of a nitely generated group clearly depends on the
choice of generating set, but only in a limited way. Suppose that S and T are two
nite generating sets for a group G. Let k be an integer such that every element
of S can be expressed as a word of length k or less in T. Then for each integer n,
the n-neighbourhood of 1 in G (with respect to the metric d
S
) is contained in the
kn-neighbourhood of 1 (with respect to d
T
). Hence

S
(n)
T
(kn) n N.
Similarly, there is an integer k

such that

T
(n)
S
(k

n) n N.
Thus the asymptotic behaviour of
S
(n) as n is independent of S. This
asymptotic behaviour is what is known as the growth of G
Similar arguments show the following.
Lemma 2 Let G be a nitely generated group, H a subgroup of nite index, and ,
the growth functions of G, H respectively, with respect to suitable choices of nite
generating set. Then there exists a constant C > 0 such that
(n) (Cn), (n) (Cn) n N.
Thus the asymptotic behaviour of the growth function is the same for a subgroup
of nite index.
More generally, if G and H are quasi-isometric nitely generated groups, then the
asymptotic growth rates of G and H (with respect to any choice of nite generating
sets) are the same. In other words, the asymptotic growth rate is a quasi-isometry
invariant.
Examples
1. If G contains an innite cyclic subgroup of nite index, then G has linear
growth. It is enough to consider the growth function of G = Z with respect
to the standard generating set S = 1. But clearly
S
(n) = 2n + 1, a linear
function of n.
2. If G contains a free abelian group of rank r as a subgroup of nite index, then
the growth of G is polynomial of degree r. Again, it is enough to consider
G = Z
r
, with S a basis. A simple calculation shows that
S
(n) is a polynomial
of degree r in n.
6
3. If G contains a free subgroup of rank greater than or equal to 2, then G has
exponential growth. To see this, rst note that, when S is a basis for a free
group F of rank r, then the number of elements of F of length m in S is
2r(2r 1)
m1
(for all m 1). Summing over m = 1, . . . , n, we see that

S
is exponential in n. Since every nitely generated group is a homomorphic
image of a free group of nite rank, no group can grow faster than a free group.
Thus no group grows faster than exponentially. Conversely, if G contains a
free group F of rank r 2, then we can choose a nite generating set T for
G such that T contains a subset S that is a basis for F. Then
T
(n)
S
(n)
grows exponentially.
The denitive result on growth of groups is the following, due to M Gromov [6].
(See also [10] for a survey on groups of polynomial growth, and [3] for an alternative
proof of Gromovs Theorem.)
Theorem 1 Let G be a nitely generated group. Then G has polynomial growth if
and only if G has a subgroup of nite index that is nilpotent.
There exist groups whose growth functions are intermediate (faster than any
polynomial, but slower than any exponential). The rst examples of these were due
to R I Grigorchuk [5].
On the other hand, it is known that any group with growth bounded above by a
polynomial function actually has polynomial growth. The degree of the polynomial
can be computed from the lower central series of the nilpotent group.
Here is the simplest nonabelian example.
Example The Heisenberg group is the group H of 33 matrices with integer entries
of the form

1 x y
0 1 z
0 0 1

It is nilpotent of class 2 with centre Z(H) = [H, H] innite cyclic. Its growth is
polynomial of degree 4. To see this, we choose a system of three generators a, b, c,
where
a =

1 1 0
0 1 0
0 0 1

, b =

1 0 0
0 1 1
0 0 1

, c =

1 0 1
0 1 0
0 0 1

.
Here c = [a, b] is central in H: [a, c] = [b, c] = 1. Now it is not dicult to show
that every element of H has a unique normal form expression if the form a

,
, , Z. A nave deduction from this would be that H has the same growth
rate as Z
3
, which is cubic. However, for any m, n Z we have [a
m
, b
n
] = c
mn
, so
that the length of c
n
as a shortest word in the generators grows asymptotically as
7
4

n, rather than n. (On the other hand, a

n
and b
n
are shortest words representing
these elements for all n.) Hence the number of words of length n or less grows
approximately like n
4
/4.
8
Lecture 2: Thin triangles and hyperbolic groups
Hyperbolic groups are so-called because they share many of the geometric properties
of hyperbolic spaces. There are a number of possible ways to dene hyperbolic
groups, which turn out to be equivalent. I will discuss only two of them. In order
to motivate these, let us rst look at some of the properties of the hyperbolic plane
H
2
, where these dier substantially from the euclidean plane E
2
.
Thin triangles in the hyperbolic plane
The incircle of a triangle (in E
2
or H
2
) is the circle contained in which is
tangent to all three sides of . The inradius of is the radius of the incircle. For
example, in E
2
if is an equilateral triangle whose sides have length , then the
inradius of is

2

3
. As , the inradius also tends to .
The situation in H
2
is quite dierent, however. Suppose =
1
is a triangle
in H
2
with vertices x, y, z, and let c be the incentre of (that is, the centre of the
incircle). Now for each t R
+
, let x
t
be the point on the half-line from c through
x, y, z respectively, such that d(c, x
t
) = t d(c, x). Dene y
t
, z
t
in a similar way, and
let
t
be the triangle whose vertices are x
t
, y
t
, z
t
. The inradius of
t
is an increasing
function of t. However, this time it does not tend to as t . The limiting
situation is an ideal triangle

, whose vertices all lie on the boundary of H

2
. Now
all ideal triangles in H
2
are congruent, and have area . This is clearly an upper
bound for the area of the incircle of an ideal triangle, so 1 is an upper bound for the
inradius of

, and hence also for that of .

In the hyperbolic plane H
2
, all triangles are thin, in the sense that there is a
bound R (actually, = 1) such that the inradius of every triangle is less than
or equal to . A consequence of this is that each edge of a hyperbolic triangle is
contained in the 2-neighbourhood of the union of the other two edges: if x is a
point on one edge of , then there is a point y on one of the other edges of such
that d(x, y) 2.
Geodesic and hyperbolic metric spaces
In order to generalise the concept of thin triangles to other metric spaces, and hence
to groups, we need to develop a more general notion of triangle. A geodesic segment
of length in a metric space (X, d) (from x to y) is the image of an isometric
embedding i : [0, ] X with i(0) = x and i() = y. In other words, we have
d(i(a), i(b)) = b a for all 0 a b . A (geodesic) triangle in X (with
vertices x, y, z) is the union of three geodesic segments, from x to y, y to z and z to
x respectively.
9
Note that in a metric space (X, d) there may not in general exist a geodesic
segment from x to y (for example, if X is a discrete metric space). If geodesic
segments exist, they need not be unique. For example, in R
2
with the
1
-metric
d
1
((a
1
, a
2
), (b
1
, b
2
)) = [a
1
b
1
[ +[a
2
b
2
[,
there are innitely many geodesic segments from (0, 0) to (1, 1). A geodesic metric
space is one in which there exist geodesic segments between all pairs of points.
(There is no requirement for these geodesic segments to be unique.)
A geodesic metric space X is hyperbolic if all triangles are thin, in the following
sense: there is a (global) constant such that each edge of each triangle in X is
contained in the -neighbourhood of the union of the other two sides of .
Examples
1. Every bounded geodesic metric space is hyperbolic. If d(x, y) B for all x, y,
then automatically any side of a triangle is contained in the B-neighbourhood
of the union of the other two sides.
2. Every tree is a hyperbolic metric space. It is clearly geodesic, since any two
points are connected by a shortest path. Moreover, any side of a triangle is
contained in the union of the other two sides.
3. The hyperbolic plane H
2
is a hyperbolic metric space, by the thin triangles
property for H
2
described above.
4. More generally, hyperbolic n-space H
n
is a hyperbolic metric space by the thin
triangles property for H
2
(since any geodesic triangle is contained in a plane.
5. Euclidean space E
n
is not a hyperbolic metric space for n 2, since E
2
does
not satisfy the thin triangles property.
Lemma 3 Let (X, d) and (X

, d

) be geodesic metric spaces that are quasi-isometric

to one another. If (X, d) is hyperbolic, then so is (X

, d

) (and conversely).
The proof of this is not dicult, but is quite technical, so I will omit it. Details
can be found, for example, in [4, p. 88]. The key point is that hyperbolicity for
metric spaces is an invariant of quasi-isometry type, which is important because
the metric space dened by a nitely generated group is only well-dened up to
quasi-isometry.
10
Hyperbolic groups
If G is a group, with (nite) generating set S, then (G, d
S
) is not a geodesic metric
space, since d
S
takes values in N. However, the geometric realization of the Cayley
graph K = [(G, S)[ is geodesic, with respect to the natural metric (which is quasi-
isometric to G). If K is hyperbolic as a metric space, then G is called a hyperbolic
group.
The rst thing to note is that this property is independent of choice of generating
set, since being hyperbolic is a quasi-isometry invariant. Also, subgroups of nite
index in hyperbolic groups are hyperbolic. Conversely, groups containing subgroups
of nite index that are hyperbolic are themselves hyperbolic.
Examples
1. Every nite group is hyperbolic, because its Cayley graphs are all bounded.
2. Every free group is hyperbolic, because it has Cayley graphs that are trees.
Moreover, if G has a free subgroup of nite index, then G is quasi-isometric
to a free group, and hence hyperbolic.
3. The fundamental group of a surface of genus g 2 is quasi-isometric to H
2
,
and hence is hyperbolic.
4. Z Z is quasi-isometric to E
2
, and hence is not hyperbolic.
The fact that Z Z is not hyperbolic shows that not every nitely generated
group is hyperbolic. In fact, a stronger statement than this is true: there are 2

0
isomorphism classes of nitely generated groups, but only
0
of these are hyperbolic.
How do we know this? Not by examining an uncountable set of groups individu-
ally, but by a simple cardinality argument. The set of all nite group presentations is
countable, by the standard countability argument. The following is a simple version
of a theorem due to E Rips (see Theor`eme 2.3 on page 60 of [2]).
Theorem 2 Every hyperbolic group is nitely presented.
Proof. Let G be a hyperbolic group, and let d = d
S
be the metric on G determined
by some xed nite generating set S. For each n N we dene
X
n
= g G [ (g, 1) n
and
R
n
= xyz [ x, y, z X
n
, xyz = 1 in G xx
1
[ x X
n
F(X
n
).
Then
X
1
X
2
. . .
11
and
R
1
R
2
. . .
so if we dene G
n
= X
n
[ R
n
) then we obtain a sequence of group homomorphisms
G
1
G
2
. . . G

= G.
Note rst that these homomorphisms are surjective. If g X
k+1
X
k
, with
k 1, then there exist elements u, v X
k
with uvg = 1 in G. Since u, v, g X
k+1
it follows that uvg R
k+1
, so uvg = 1 in G
k+1
. Hence the image of G
k
G
k+1
contains the generating set X
k+1
and so it is surjective.
Next we show that G
N
G
N+1
is injective (and hence an isomorphism) for all
suciently large N. It follows that G

= G
N
for all large N, or equivalently that
G = X
N
[ R
N
), so is nitely presented, as claimed.
We x N 2. Suppose that xyz R
N+1
. In other words, x, y, z X
N+1
with
xyz = 1 in G. We have to show that this relation can be deduced from those in
R
N
. Unfortunately, the elements x, y, z do not in general belong to X
N
. To make
sense of this, we rst choose, for each x X
N+1
X
N
, a canonical splitting x = x
1
x
2
with each of d(x
1
, 1) > , d(x
2
, 1) > and d(x, 1) = d(x
1
, 1) + d(x
2
, 1). We then
add the generator x and the relation x
1
x
2
x
1
to the presentation for G
N
to get an
equivalent presentation. Having done this, we now show how to deduce xyz = 1
from the relations in X
N
together with the canonical splitting relations.
Case 1 x, y X
N
, z , X
N
.
Let P be the point of the geodesic segment z corresponding to the canonical
splitting z
1
z
2
. By the thin triangle property, this is within distance of some point
Q on one of the other edges of the geodesic triangle with vertices at 1, x, xy. The
geodesic PQ, together with the geodesic from Q to the vertex of opposite the
edge containing Q, divides the geodesic triangle into three smaller triangles, each of
the edges of which has length N or less. It follows that the relation xyz
1
z
2
= 1 can
be deduced from three relations in R
N
, as required.
Case 2 y, z , X
N
.
By case 1 we can assume all (true) relations of the form abc = 1 with a, b X
N
and c X
N+1
. Here we proceed exactly as in case 1, starting from the canonical
splitting of z. The point Q may lie on an edge of length N + 1, in which case it
corresponds to a (possibly non-canonical) splitting of x or y - say x = x

1
x

2
(but still
with x

1
, x

2
of lengths less than N + 1). Hence the relation x = x

1
x

2
is one we are
allowed to assume. Finally, we divide as before into three smaller triangles. This
time it is possible that one of the three triangles has one side of length N + 1, but
all other sides of the smaller triangles have length N or less. By case 1 we are done.
Similar arguments apply to the relations of the form xx
1
, x X
N+1
X
N
,
completing the proof.

12
Lecture 3: Isoperimetric inequalities and decision
problems
Consider a Jordan curve C in the euclidean plane E
2
. The Jordan curve theorem
tells us that C bounds a compact domain D in E
2
. The isoperimetric inequality
compares the area of D to the length of C. To make sense of this, let us assume that
C is nice (smooth, polygonal, piecewise smooth, . . . ) so that it has a well-dened
(nite) length and D has a well-dened (nite) area. A classical result of calculus
of variations says that, for C of xed length , the area of D is maximised when
C is a circle (of radius r = /2). This maximal area is r
2
=
2
/4. Hence the
isoperimetric inequality for E
2
is:
Area(D)

2
4
.
Note that the right hand side of this inequality is a quadratic function of .
We can look at the same thing in the hyperbolic plane H
2
. Again the maximal
area occurs when C is a circle. In H
2
the length of a circle of radius r is 2 sinh(r),
and its area is

r
0
2 sinh(t)dt = 2(cosh(r) 1) 2 sinh(r)
so in this case we have an isoperimetric inequality
Area(D)
for domains D bounded by curves of length . The important dierence here is that
the right hand side of the inequality is a linear function of .
What is the relevance of this for groups?
Let T : X [ R ) be a nite presentation of a group G. If w F(X) is a word
that represents the identity element 1 G, then it can be expressed, in F(X), as a
product of conjugates of elements of R and their inverses:
w = (u
1
1
r

1
1
u
1
) . . . (u
1
n
r

n
n
u
n
),
where u
i
F(X), r
i
R and
i
= 1. There will in general be innitely many such
expressions. The least value of n amongst all such expressions is called the area of
w, Area(w).
The notion of area for words representing 1 G has a geometric interpretation.
Denition A van Kampen diagram or Dehn diagram over the presentation T con-
sists of the following data:
13
A simply-connected, nite 2-dimensional complex M contained in the plane.
An orientation of each 1-cell of M.
A labelling function that labels each 1-cell of M by an element of X, such
that the composite label of the boundary of each 2-cell (read from a suitable
starting point in a suitable direction) is an element of R. (Here, an edge
labelled x X contributes x to the boundary label of the 2-cell if read in the
direction of its orientation, and x
1
if read in the opposite direction.)
The complement of M in the plane is topologically a punctured disc. It also has
a boundary that is a closed path in the 1-skeleton of M. The label of this path is
called the boundary label of the diagram. Note that it is dened only up to cyclic
permutation and inversion.
Lemma 4 There exists a Dehn diagram with boundary label w if and only if w = 1
in G, in which case the minimum number of 2-cells in all Dehn diagrams with
boundary label w is Area(w).
Example T = x, y, [ [x, y] ), (where [x, y] means xyx
1
y
1
),
w = [x
2
, y
2
] = (x[x, y]x
1
)([x, y])(yx[x, y](yx)
1
)(y[x, y]y
1
).
Then Area(w) = 4.
x
x x
x x
x
y
y
y
y
y
y
The function f : N N dened by
f() = maxArea(w) [ w F(X), w = 1 in G, (w) =
14
is called the Dehn function, or isoperimetric function for the presentation T. A given
nitely presented group has (innitely) many possible nite presentations, which can
have very dierent Dehn functions. However, there are aspects of Dehn functions
that are independent of the choice of presentation, and hence are invariants of the
group G.
Lemma 5 Let T and Q be two nite presentations for a group G, and let f, g be
the corresponding Dehn functions. Then there exist constants A, B, C, D N such
that
f(n) Ag(Bn +C) +D n N.
In particular, if g is bounded above by a function that is linear (or quadratic, or
polynomial, or exponential, . . . ) in n, then the same is true for f. These properties
are thus invariants of the group G.
Denition A nitely presented group G has a linear (quadratic, . . . ) isoperimetric
inequality if for some (and hence for any) nite presentation T with Dehn function
f, there is a linear (quadratic, . . . ) function

f such that f(n)

f(n) n N. A
nitely presented group G is hyperbolic if it has a linear isoperimetric inequality.
We have now given two distinct denitions for hyperbolic group. Implicit in this
is an assertion that these two properties are equivalent: the thin triangles condition
is equivalent to the linear isoperimetric inequality condition.
Examples
1. Every nite group is hyperbolic. If G is a nite group, then the Cayley table
for G is a nite presentation for G. In other words, we take G to be the nite
generating set, and the set of all true equations xy = z in G for the set of
dening relations. Given a word w F(G) such that w = 1 in G, what is
Area(w)? If w has length 3, then it is a relation in G. If it has length greater
than 3, then it has the form w = xyu, where x, y G and u is a word. If z G
with z = xy, then w = (xyz
1
)(zu) with zu shorter than w. An inductive
argument shows that Area(w) (w).
2. Every free group is hyperbolic. Indeed, if F = X [ ) is the standard presen-
tation, then the empty word 1 is the only cyclically reduced word representing
the identity in F, and Area(1) = 0.
3. The fundamental group of a surface of genus g 2 is hyperbolic.
This is because of a Theorem of Dehn: let
G = a
1
, b
1
, . . . , a
g
, b
g
[ [a
1
, b
1
] [a
g
, b
g
] )
15
and let w be a word in the generators a
i
, b
i
such that w = 1 in G. Then
there exist cyclic permutations w

= uv of w and r

= us of the relator
r = [a
1
, b
1
] [a
g
, b
g
] or its inverse r
1
, with a common initial segment u of
length greater than half the length of r, that is (u) > 2g. It follows that
Area(w) = Area(uv) Area(s
1
v) + 1, while (w) < (s
1
v). An inductive
argument then shows that Area(w) < (w).
4. ZZ is not hyperbolic. Under the presentation G = x, y [ [x, y] ), the word
w
n
= [x
n
, y
n
] has area at most n
2
. Indeed there is a Dehn diagram for w
n
which is a square of side length n in E
2
, subdivided into n
2
squares of side
length 1. On the other hand, the boundary of this Dehn diagram is a simple
closed path in E
2
bounding a square of area n
2
, so n
2
is a lower bound for the
area of w
2
n
. Now (w
n
) = 4n, so this sequence of words shows that no linear
isoperimetric inequality holds for Z Z.
The word problem
Let G be a group given by a (nite) presentation X [ R ). (Much of what follows
can also be done for certain types of innite presentation, but let us keep things
simple.) The word problem for G is that of deciding algorithmically whether or not
a given word w in the generating set X represents 1 G. A solution to the word
problem is an algorithm that, when an arbitrary word w is input, will output YES
or NO after a nite time, depending on whether or not w = 1 in G.
As with most of what we have been doing, this problem appears at rst sight
to depend on the choice of nite presentation for G. However, it can be shown to
be independent of this choice. Indeed, given two nite presentations X [ R ) and
Y [ S ) of isomorphic groups G and G

, there are functions X F(Y ) and Y

F(X) that induce the isomorphisms. Given an algorithm to solve the word problem
for X [ R ) and a word w

in Y , we can apply the function Y F(X) to rewrite

w

as a word w in X, then apply our solution to decide whether or not w = 1 in G.

Since this is true if and only if w

= 1 in G

, we are done. (NB this solution assumes

the existence of isomorphisms G G

. Although the practical implementation of

the solution uses one of these isomorphisms, it does not assume that we are able to
nd it for ourselves. Indeed, the problem of determining whether or not too given
presentations represent isomorphic groups is another insoluble decision problem,
called the isomorphism problem).
At rst sight the word problem seems trivial. We have a very good criterion for
deciding whether or not w = 1 in G: namely, can we express w as a product of
conjugates of elements of R (and their inverses) in F(X)? Why can we not use this
criterion as an algorithm? In fact, this criterion supplies part of the answer, but the
other part is missing, in general.
16
Since X and R are nite sets, the set F(X) is countable, and hence so is the
set of conjugates of elements of R R
1
by elements of F(X). Hence so also is the
set of all nite products of such conjugates. In principal, the standard proof that
these sets are countable can be transformed into an algorithm to produce an innite
list that contains all products of conjugates of elements of R R
1
(possibly with
repetitions). This can be combined with an algorithm to check each element of the
list for equality with w. Once we recognise that w is in the list, we stop and output
the answer YES.
However, what if w ,= 1 in G? Our algorithm will continue to list words forever,
without ever recognising w. It will never give us the desired output NO! We may,
after a long wait, suspect that w ,= 1, but we can never be completely sure. If
our computer runs for 1000 years without producing an output, this could be for
three reasons: w ,= 1; w = 1 but w appears much later than 1000 years in the
computer-generated list; or possibly hardware error.
In fact, it is known that there are nitely presented groups with unsolvable word
problem. In other words, there is no algorithm of the kind we are looking for.
However, for hyperbolic groups there is indeed an algorithm. Because of the linear
isoperimetric inequality, there is a bound on Area(w) for words w such that w = 1
in G, depending only on (w). Since R is nite, it can easily be shown that there
are only nitely many Dehn diagrams of a given area and length of boundary, and
hence these can be searched in nite time, comparing boundary labels with w. (NB
it is not true that there are only nitely many Dehn diagrams of a given area. For
example, if r R then there is a Dehn diagram of area 2 for each of the innitely
many words [r, u], u F(X). The condition on boundary length is also important
to make this search work.)
Lemma 6 Hyperbolic groups have solvable word problem
This result is not so strong as it might appear. In fact, our argument does not
use the fact that hyperbolic groups have a linear isoperimetric inequality. All that
is required is that the Dehn function be bounded above by some function that we
can compute (a recursive function). This is a much weaker property.
Lemma 7 Let G be a group given by a (nite) presentation with a recursive isoperi-
metric inequality. Then G has soluble word problem.
The conjugacy problem
However, there are other types of decision problem where the stronger hyperbolic
property can be successfully applied.
17
Let G be a group with nite presentation X [ R ). A solution to the conjugacy
problem for G is an algorithm which, when two words w, w

in X are input, will

determine in nite time whether or not w, w

represent conjugate elements of G. As

before, this can be shown to be independent of the choice of presentation.
Lemma 8 If G is hyperbolic, then the conjugacy problem is soluble for G.
Proof. The idea of the proof is roughly as follows. Suppose that w, w

are words
that represent conjugate elements of G, and that the conjugating element g G
is represented by some very long word u. Then the relation w

uw
1
u
1
= 1 in G
corresponds to a van Kampen diagram of area bounded above by an a priori given
linear function f of (u). Let x
0
, . . . , x
(u)
and y
0
, . . . , y
(u)
denote the sequences
of vertices along the boundary of the diagram corresponding to the two segments
labelled u. (So that the paths from x
0
to x
(u)
and from y
0
to y
(u)
are both labelled
u). Then, roughly speaking, d(x
i
, y
i
) is bounded above by some constant multiple
of the constant function f

((u)), where d denotes the path-length metric in the

1-skeleton of the Dehn diagram.
Hence there is a constant N such that, whenever (u) > N, two of the pairs
(x
i
, y
i
) and (x
j
, y
j
) of vertices are joined by paths p
i
, p
j
of equal lengths and with
equal labels. Cutting the diagram open along p
i
and p
j
, and then regluing it by
identifying p
i
with p
j
, we get a diagram corresponding to a new relation w

vw
1
v
1
=
1 with (v) < (u). Hence we can restrict our search for conjugating elements to
those of length less than N. Since there are only nitely many of these, it suces
to apply the solution to the word problem to a nite number of words.
The key part of this proof is the notion of the nite bound for the width of the
long strip that corresponds to the Dehn diagram for w

uw
1
u
1
= 1. The details
of this part of the argument are tricky. The bound has to be calculated without
specic reference to w, w

, except that it depends on m = (w) and n = (w

).
For xed m, n, the area of the potential diagrams under consideration will be
bounded by a linear function of (u). Since we are working with a given nite
presentation for G, the lengths of relations are also bounded, and hence the number
of edges in the diagram is bounded by a linear function of (u). If the geodesic
paths connecting the x
i
to the y
i
are not of bounded length, then the sums of their
lengths will grow faster than linearly with (u), leading to arbitrarily complicated
intersections between these paths for very large u, and hence to a contradiction.
An alternative approach to proving this theorem can be found, for example, in
[2, pages 52-56].
18
Lecture 4: Further aspects
The Rips Complex
We have already seen that hyperbolic groups are nitely presented, a result I at-
tributed to Rips. It is in fact a special case of a much stronger theorem of Rips,
which I will try to explain in this section. Topologically, the fact that a group G is
nitely presented means that there is a nite simplicial complex Y with G

=
1
(Y ).
In fact, this condition is both necessary and sucient for G to be nitely presented.
Moreover, we have a lot of freedom in the choice of Y . For example, we can choose
Y to be 2-dimensional, since the fundamental group of a simplicial complex is de-
termined by its 2-skeleton (the union of all the simplices of dimensions 2).
Let Y be a nite simplicial complex with G

=
1
(Y ), and let K be its universal
cover. Then K is a simply connected complex, and there is a simplicial action of G
on K such that Y = K/G is compact. The G-action is free (gx = x g = 1), and
so in particular it is properly discontinuous. The existence of a simply connected
simplicial complex K and a G-action on K with these properties is yet another
necessary and sucient condition for G to be nitely generated. (Here the weaker
properly discontinuous condition on the action is sucient: one can then construct
another simply connected complex K

on which G acts freely, again with K

/G
compact.)
The Rips complex is a particular choice of K for a hyperbolic group G that is
not only simply connected but contractible. Thus the result we want to prove is the
following.
Theorem 3 (Rips) Let G be a hyperbolic group. Then there exists a simplicial
complex K and an action of G on K such that
1. G acts properly discontinuously on K;
2. K/G is compact;
3. K is contractible.
For the proof, we will x a nite generating set S for G, and a positive real num-
ber such that G is -hyperbolic. We will actually construct a simplicial complex
T
n
(G) on which G acts, for every natural number n.
Dene
T
n
(G) = Y G [ Y ,= & diam
S
(Y ) n.
Some explanation is required here. Firstly, diam
S
(Y ) denotes the diameter of
the set Y with respect to the metric d
S
:
diam
S
(Y ) = maxd
S
(x, y) [ x, y Y .
19
Secondly, what we have dened here is a collection of subsets of G. They are all
nite subsets, since each Y T
n
(G) is contained in a ball of radius n, which has only
nitely many elements. They are nonempty sets by denition. We are implicitly
identifying a (k + 1)-element set Y with a k-simplex whose vertex set is Y . Clearly
, = X Y T
n
(G) X T
n
(G),
so our collection of simplices T
n
(G) is closed with respect to faces. It also follows
from this that two simplices of T
n
(G) are either disjoint or intersect in a common
face of both. These are the dening properties for simplicial complexes, so T
n
(G) is
indeed a simplicial complex, as claimed.
There is a natural action of G on T
n
(G): if g G and Y = x
0
, . . . , x
k
T
n
(G),
then
Y g = x
0
g, . . . , x
k
g T
n
(G).
Since G is acting by permuting simplices, the action is simplicial.
We next note that the rst two properties in Rips Theorem hold automatically
for all natural numbers n.
Claim 1 G acts properly discontinuously on T
n
(G) for all n.
To see this, note rst that the set of vertices (or 0-simplices) of T
n
(G) is just
G (or, more correctly, the set of 1-element subsets of G), with G acting by right
multiplication. If Y is a k-simplex and g G with Y Y g ,= , then x = yg for
some x, y Y , so g = y
1
x with x, y Y . There are only nitely many such g.
Hence G acts properly discontinuously, as claimed.
Claim 2 T
n
(G)/G is compact for all n.
It is enough to show that there are only nitely many orbits of simplices in
T
n
(G). If Y = x
0
, . . . , x
k
T
n
(G) then Y

= Y x
1
0
T
n
(G) with 1 Y

. But
then Y

is contained in the n-neighbourhood of 1, which is nite. There are thus

only nitely many possibilities for Y

, so only nitely many orbits of simplices.

For the third property of T
n
(G) the value of n is important. For example, T
0
(G)
is just the discrete set G regarded as a 0-dimensional simplicial complex, so it is
not even connected (unless G is the trivial group). The point of Rips Theorem is
that the hyperbolic property of G ensures that T
n
(G) is contractible for large n. To
specify what we mean by large n, recall that we have xed a metric d = d
S
on G,
with respect to which G is -hyperbolic for some real number .
Proposition 1 If n > 4 + 2, then T
n
(G) is contractible.
Proof. If not, then there is a nonzero homotopy group
t
(T
n
(G)) ,= 0, so a continuous
map f : S
t
T
n
(G) that is not homotopic to a constant. The image of f is compact,
20
so contained in a nite subcomplex K of T
n
(G). It is therefore sucient to show
that every nite subcomplex K is contractible to a point 1 within T
n
(G).
Let X denote the (nite) set of vertices of the (nite) subcomplex K of T
n
(G).
We argue by induction on
=

xX
d(1, x).
Another useful parameter is
= max
xX
d(1, x).
If n/2 then for any x, y X we have
d(x, y) d(x, 1) +d(1, y)
n
2
+
n
2
= n,
so X is the vertex set of a simplex of T
n
(G), which is therefore contractible (to 1),
and K is a union of faces of this simplex. The result is therefore true in this case,
which includes the initial case of the induction.
Suppose then that > n/2, and choose x
1
X with d(1, x
1
) = . There is a
geodesic in the Cayley graph = (G, S) from 1 to x
1
, and we dene x
0
G to
be the point on this geodesic with d(x
0
, x
1
) = [n/2], the integer part of n/2. We
now dene a map f : X G by
f(x) =

x (x ,= x
1
)
x
0
(x = x
1
)
Now f(X) is a nite set whose parameter is less than that of X, so we can apply the
inductive hypothesis to any subcomplex of T
n
(G) with vertex set f(X). To complete
the inductive argument, it is sucient to show that f extends to a simplicial map
f : K T
n
(G) that is homotopic to the identity map on K. If is a simplex of K
that does not contain x
1
, then f is the identity on , so it is enough to consider the
action of f on simplices that contain x
1
. Let be such a simplex, and D its vertex
set. I claim that D x
0
is the vertex set of a simplex

of T
n
(G). Then and
f() will be faces of

, so f will be simplicial and homotopic to the identity on .

To prove the existence of

, we need only check that any two of its vertices

are no more than n apart in the metric d = d
S
. This is true by denition for two
vertices of , so we need only prove that d(x
0
, y) n for every vertex y : in
other words, we must prove that
[d(x
1
, y) n & d(1, y) = d(1, x
1
)] d(x
0
, y) n.
Choose geodesics from 1 to y and from y to x
1
in . Then , , are the
sides of a geodesic triangle in , and x
0
is a point on . By the thin triangle property
there is a point z on such that d(x
0
, z) < .
21
Case 1: z .
Then d(1, z) = d(1, y) d(y, z) d(y, z) and so
= d(1, x
1
) d(1, z) +d(z, x
0
) +d(x
0
, x
1
) d(y, z) + + [n/2].
Thus d(y, z) + [n/2] and so
d(y, x
0
) d(y, z) +d(z, x
0
) + [n/2] + < n.
Case 2: z .
Then [n/2] = d(x
0
, x
1
) d(x
0
, z)+d(z, x
1
) +d(z, x
1
), so d(z, x
1
) [n/2] .
Moreover, d(y, z) +d(z, x
1
) = d(y, x
1
) n, so d(y, z) n[n/2] + [n/2] +1+.
Finally,
d(y, x
0
) d(y, z) +d(z, x
0
) [n/2] + 1 + + n.

Boundary of a hyperbolic group

The ordinary hyperbolic plane, when considered in the Poincare disc model, has a
natural boundary - the circle at innity. Similarly, hyperbolic space of dimension n
has a natural boundary which is an (n1)-sphere. There are analogous constructions
for geodesic metric spaces in general, and for hyperbolic groups in particular.
The purpose of this section is to describe this construction and an application
to the structure of hyperbolic groups. We will omit most of the technical details for
the benet of expositional clarity. We refer the reader to [2, 4] for more rigorous
treatments.
Consider rst the hyperbolic plane H
2
. In the Poincare disc model, this is the
interior of the unit disc in R
2
: x R
2
, [[x[[ < 1. The boundary of the disc,
H
2
= S
1
= x R
2
, [[x[[ = 1 is a compact space. Indeed the union

H
2
= H
2
H
2
is a compact space in which H
2
is an open dense subset, so is a natural compactica-
tion of H
2
. There are various ways of thinking about the boundary S
1
, any of which
can be used to produce the analogous constructions for hyperbolic metric spaces in
general.
I will describe only one of these ideas. Fix a basepoint x
0
H
2
. (In the Poincare
disc model, one should think of x
0
as being the euclidean origin (0, 0).) The set
of geodesic rays starting at x
0
is naturally identied with S
1
E
2
. Each such ray
meets the boundary in a unique point, so we get an identication H
2

= S
1
. The
22
advantage of this approach is that it induces a metric on H
2
, using the euclidean
or spherical metric on S
1
. The metric depends on the choice of base-point, but the
underlying topology does not.
From now on, we consider a hyperbolic group G, with a xed generating set S,
and corresponding Cayley graph = (G, S). For our basepoint x
0
, we make the
canonical choice of the identity element 1 G = V (). We then consider geodesic
rays from x
0
(that is, isometries [0, ) with 0 x
0
).
For some purposes, it is useful to be able to vary the starting point of the geodesic,
so that we relax the condition 0 x
0
.
Two geodesics g, h are said to be equivalent if d(g(t), h(t)), t [0, ) is bounded.
Then G = is dened to be the set of equivalence classes of geodesic rays.
Examples
1. If G is nite, then G = .
2. If G is an innite cyclic group, then G = , +.
3. If G is a Fuchsian group, then the Cayley graph embeds quasi-isometrically
in H
2
. Every geodesic ray in can be approximated by a geodesic ray in
H
2
, and vice versa, so that there is a natural identication between G and
H
2

= S
1
.
4. Similarly, if G is a Kleinian group, then G = H
3
= S
2
.
5. If G is a nonabelian free group, then is a tree and G is a Cantor set.
Remark In the above examples, we have made no explicit mention of the particular
generating set of G chosen to determine G. It is not dicult to see that the
denition of G does not depend on the choice of generating set. Indeed, G is an
invariant of quasi-isometry, so for example G = H if H is a subgroup of nite
index in G.
Also implicit in the above examples is the existence of a natural topology on G.
In fact, we can dene a metric on G. As with the hyperbolic plane H
2
, the metric
on G will depend on our choice of base-point, although the resulting topology does
not. This is important for what follows. We use the canonical choice of the identity
element of G to be the base-point. Given two elements x, y G, we dene the
Gromov inner product
x[y) =
d(1, x) +d(1, y) d(x, y)
2
.
A good way to think of this is as follows. Recalling the thin triangles property
of hyperbolic metric spaces, we see that two travellers moving from 1 to x and y
23
respectively along suitable geodesics will remain close together (less than 2 apart)
for a certain distance, before beginning to diverge rapidly. The inner product x[y)
measures approximately the length of time that the two travellers remain close
together. In particular, x[y) = 0 if and only if the travellers diverge immediately,
that is, i there is a geodesic from x to y that passes through 1.
The inner product extends to the boundary by taking limits as x and/or y vary
along geodesic rays. The inner product on G is enough to recover the original metric,
but we have to be slightly more subtle to dene a metric on G.
Denition Choose > 0. Dene

(a, b) = exp(a[b)) for a, b G, and

d

(a, b) = inf
n

i=1

(a
i1
, a
i
),
the inmum being taken over all nite sequences a
0
, . . . , a
n
G with a
0
= a and
a
n
= b.
Then d

is a metric on G for suciently small . In general

is not a metric,
because it does not satisfy the triangle inequality. It turns out that

G = G is
a compact topological space, such that the subspace topology induced on G is the
same as that induced by the metric d

.
To apply the boundary of G to obtain results about its algebraic structure, we
use the fact that elements of G act by isometries on the Cayley graph . This
action extends to

G as an action by homeomorphisms. Not isometries, as we have
not dened a metric on

G! In general, the restriction of the action to G is also not
by isometries, since the metric on G depends on the base-point, which is not xed
by the isometries of in question.
Lemma 9 Any element of G acts on with bounded orbits if and only if it has
nite order. Any element of innite order in G xes a point of G.
Proof. Clearly, if g G has nite order n, then its orbits have the form
h, gh, . . . , g
n1
h,
and so are bounded. Conversely, an orbit of g is bounded if and only if it is nite,
which can happen if and only if g has nite order. If g has innite order, then the
sequence 1, g, g
2
, . . . can be approximated by a geodesic ray, and so tends to a limit
a G.
We can apply this fact to the study of centralisers in G. Suppose x has innite
order, and a = lim
n
x
n
G as in the proof of the Lemma. If y belongs to the
centraliser of x in G, then yx
n
= x
n
y for all n, so
d(yx
n
, x
n
) = d(1, x
1
yx) = d(1, y)
24
is constant. It follows that y maps the geodesic ray that approximates the sequence
1, x, x
2
. . . to one a bounded distance from the original (in other words, another
geodesic ray in the same equivalence class). Hence y xes a G. We can therefore
study the centraliser of G by analysing the dynamics of the stabiliser of a acting on
G.
From the thin triangles property, it is easy to deduce that there is in fact a global
bound on the distance between two equivalent geodesic rays for large t, provided
that one is allowed to vary the starting point of one of the rays. If D is an upper
bound for d(g(t), h(t)), and T t D, then the point h(t) on the geodesic for h(0)
to h(T) must be close to some point y on the geodesic from g(0) to h(T) (since it is
not close to any point of the geodesic from h(0) to g(0)). Similarly, y is close to some
point on the geodesic from g(0) to g(T), ie to g(t +u) for some u. Once one analyses
this argument, one obtains the following global bound. If g, h are equivalent rays,
then there are real numbers C, u such that
d(g(t), h(t u)) 16t > C.
Returning to the study of centralisers, one can show that only nitely many of
elements of the centraliser of x give rise to a given value of u in the above inequality,
from which it follows that x) has nite index in its centraliser.
Theorem 4 The centraliser of an element x of innite order in a hyperbolic group
G contains the cyclic group x) as a subgroup of nite index.
Corollary 1 No hyperbolic group contains a free abelian subgroup of rank 2.
Corollary 2 The group Z Z is not hyperbolic
Corollary 3 The only knot group which is hyperbolic is the innite cyclic group
(the group of the unknot).
Proof. The group of any nontrivial knot contains a peripheral subgroup (that is,
the fundamental group of the boundary of a regular neighbourhood of the knot)
isomorphic to Z Z.
References
[1] J Alonso et al, Notes on negatively curved groups, in: Group theory from a geo-
metrical viewpoint (ed. E Ghys, A Haeiger and A Verjovsky), World Scientic,
1991, pp 3-63.
25
[2] M Coornaert, T Delzant and A Papadopoulos, Geometrie et theorie des groupes
(Les groupes hyperboliques de Gromov), Lecture Notes in Mathematics 1441,
Springer-Verlag, 1990.
[3] L van den Dries and A J Wilkie, Gromovs theorem on groups of polynomial
growth and elementary logic, J Algebra 89 (1984), 349-374.
[4] E Ghys and P de la Harpe (editors), Sur les groupes hyperboliques dapres
Mikhael Gromov, Progress in Mathematics 83, Birkhauser, 1990.
[5] R I Grigorchuk, Milnors problem on the growth of groups, Dokl. Akad. Nauk.
SSSR 271 (1983), 30-33; Soviet Math. Dokl. 28 (1983) 23-26.
[6] M Gromov, Groups of polynomial growth and expanding maps, Pub. Math.
IHES 53 (1981), 53-73.
[7] M Gromov, Innite groups as geometric objects Proceedings of the International
Congress of Mathematicians (Warsaw, 1983), PWN, Warsaw, 1984, pp 385-392.
[8] M Gromov, Hyperbolic groups, in: Essays in group theory (ed. S M Gersten),
MSRI Publications 8, Springer-Verlag, 1987, pp 75-265.
[9] M Gromov, Asymptotic Invariants of Innite Groups, (Geometric Group The-
ory, volume 2, edited by G A Niblo and M A Roller), London Mathematical
Society Lecture Notes Series 182, Cambridge University Press, 1983.
[10] J Tits, Groupes a croissance polynomiale (dapr`es M. Gromov et al.), Seminaire
Bourbaki 1980/81, Lecture Notes in Math 901, Springer (1981), pp 176-188.
26