An Introduction to Limit Groups

H J R Wilton
March 3, 2005
1 Introduction
It is very natural in the study of discrete groups to ask for a description of the
variety Hom(G, H), for nitely generated groups G and H. For example:
1. Recall that a group G is Hopan if every epimorphism G G is an
automorphism, and co-Hopan if every monomorphism G G is an
automorphism. For example, free groups are Hopan, but not co-Hopan.
Setting G = H and understanding Hom(G, G) could lead to a proof that
G is Hopan or co-Hopan.
2. An algorithmic understanding of Hom(G, H) could be useful in solving
the isomorphism problem.
3. Consider a system of equations over a group H, given by words
w
i
(x
1
, . . . , x
n
)
in n unknowns. Then its easy to see that solutions to are precisely in
bijection with Hom(G(), H), where G() is the group with presentation
x
1
, . . . , x
n
[w
1
, w
2
, . . .).
This so-called algebraic geometry over H is also the starting point for
the study of the rst-order theory of H.
One of the oldest and hardest problems in the rst-order theory of groups is
the Tarski problem, which asks which groups have the same rst-order theory
as the free group. This problem was recently solved by Zlil Sela (see [18], [19],
[12], [13], [14], [15] and [16]). The aim of this series of talks is to explain his
description of Hom(G, F), for F a free group of rank at least 2. (In fact, much of
the theory carries through to describe Hom(G, H) for H torsion-free hyperbolic.)
This description takes the form of a Makanin-Razborov diagram, which will be
dened in the second talk.
Limit groups are a crucial feature of Makanin-Razborov diagrams. There
are many dierent denitions of limit groups, and another aim of these talks is
1
to explain all the dierent denitions, why they are equivalent, and how they
t into the theory of Hom(G, F).
Rather than plunging straight into Makanin-Razborov diagrams, therefore,
we start with a class of groups that will turn out to be equivalent to limit groups,
but are easier to dene and to work with.
2 Fully residually free groups
2.1 Denition and easy examples
Fix F a free group of rank r > 1.
Denition 2.1 A nitely generated group G is residually free if, for any non-
trivial g G, there exists a homomorphism f : G F with f(g) ,= 1.
G is fully residually free or -residually free if, for any nite subset X G,
there exists a homomorphism f : G F whose restriction to X is injective.
Note that the choice of F does not matter. Here are some easy examples.
Example 2.2 (Free groups) If F is a free group then F F. In particular,
F is fully residually free.
Example 2.3 (Free abelian groups) If A is a free abelian group, then any
nite subset can be embedded by a homomorphism in Z. In particular, A is fully
residually free. The proof is left as an exercise.
2.2 Elementary properties
A group G is called commutative transitive if the centralizer of any element is
abelian; equivalently, for a, b, g G, if a commutes with g and g commutes with
b then a commutes with b.
Recall that a subgroup H G is malnormal if, whenever g GH,
gHg
1
H = 1.
A group G is completely separated abelian (CSA) if every maximal abelian sub-
group is malnormal.
Exercise 2.4 A CSA group is commutative transitive.
Its easy to see some elementary properties of fully residually free groups.
Proposition 2.5 Let G be a fully residually free group.
1. Any nitely generated subgroup of G is fully residually free.
2. G is torsion-free.
2
3. Any pair of elements of G generates either a free group or a free abelian
group.
4. G is commutative transitive.
5. G is CSA.
Proof: Property 1 is trivial.
Let g G. Then there exists a homomorphism f : G G with f(g)
non-trivial. So f(g
n
) ,= 1 for all n, so g
n
,= 1. This proves 2.
To prove 3, consider g, h G, and assume [g, h] ,= 1. Then there is an
epimorphism
f : g, h) F
2
.
Therefore g, h generate a free non-abelian group.
Consider a, b, g G with [g, a] = [g, b] = 1. There exists a homomorphism
f : G F which is injective on the set
1, g, [a, b].
Now f([g, a]) = f([g, b]) = 1 so f(a) and f(b) and f(g) must all lie in the same
cyclic subgroup if F; in particular, f([a, b]) = 1. Therefore [a, b] = 1. This gives
4.
Let H G be a maximal abelian subgroup, consider g G, and suppose
there exists non-trivial h gHg
1
H. Let f : G F be injective on the set
1, g, h, [g, h].
Then f([h, ghg
1
]) = 1, which implies that f(h) and f(ghg
1
) lie in the same
cyclic subgroup. But in a free group, this is only possible if f(g) also lies in
that cyclic subgroup; so f([g, h]) = 1, and hence [g, h] = 1. By 4 it follows that
g commutes with every element of H, so g H. This proves 5. QED
These properties immediately give some examples of groups that arent fully
residually free.
By 2, any group with torsion is not fully residually free.
By 3, the fundamental group of the Klein bottle is not fully residually
free.
Direct products are not fully residually free. Specically, suppose G =
AB, where A is non-trivial and B is non-abelian. Then B is contained
in the centralizer of any element in A, so G is not commutative transitive
and hence not fully residually free.
A slightly less trivial non-example is the fundamental group of the surface
of Euler characteristic -1, which has presentation
a, b, c[a
2
b
2
c
2
).
By work of Lyndon (see [9]) that, whenever three elements of F satisfy a
2
b
2
c
2
=
1, in fact abc = 1. It follows that
1
() is not fully residually free.
We shall see that, apart from the fundamental groups of the three simplest
non-orientable surfaces, all surface groups are fully residually free.
3
2.3 Dicult properties
Fully residually free groups have other properties that are much less obvious.
They are summarized in the following theorem.
Theorem 2.6 Suppose G is fully residually free.
1. G is nitely presented. Indeed, there exists a nite K(G, 1).
2. All abelian subgroups of G are nitely generated.
3. If G is non-abelian then it has a non-trivial cyclic splitting.
4. G is CAT(0) with isolated ats.
We shall prove 1, 2 and 3 in this series of talks, following Sela. There is a
simpler independent proof that fully residually free groups are nitely presented
due to Guirardel, who shows that fully residually free groups act freely on R
n
-
trees (see [8]). Alibegovic and Bestvina proved 4 in [2]; Alibegovic (in [1]) and
Dahmani (in [7]) had already independently proved that fully residually free
groups were hyperbolic relative to their maximal abelian subgroups.
Note that, by property 3 of the theorem, the only 3-manifold groups that
are limit groups are free products of Z and Z
3
.
2.4 A criterion in free groups
To prove that a group G is fully residually free, it suces to show that for any
nite X G1 there exists a homomorphism f : G F with 1 / f(X). So
a criterion to show that an element of F is not the identity will be useful.
In the short term, this will make it possible to prove that surface groups are
fully residually free. Eventually, it will give a complete constructive characteri-
zation of limit groups.
Lemma 2.7 Let z F 1, and consider an element g of the form
g = u
0
z
i
1
u
1
z
i
2
u
2
. . . u
n1
z
i
n
u
n
where n 1 and, whenever 0 < k < n, [u
k
, z] ,= 1. Then [g, z] ,= 1 whenever
min
k
[i
k
[ is suciently large. In particular, g ,= 1.
Choose a generating set for F so the corresponding Cayley graph is a tree
T, and x x T. An element a F species a geodesic [x, ax] T. Likewise,
a string of elements a
0
, a
1
, . . . , a
n
F denes a path
[x, a
0
x] [a
0
x, a
0
a
1
x] . . . [a
0
. . . a
n1
x, a
0
. . . a
n
x]
in T, where denotes concatenation of paths.
Here is the idea behind the lemma. In the path corresponding to the dening
expression for g, a subarc corresponding to a power of z lies in a translate of
4
PSfrag replacements
x Axis(z)

k1
y
k1
g
k1
x

k
x
k
g
k1
Axis(z)
g
k1
z
i
k
x

k
y
k g
k
Axis(z)
g
k
x
Figure 1: The proof of lemma 2.7.
Axis(z). A high power of z pushes the path a long way along this translate; the
next u
k
term then pushes the path onto a new translate.
Proof of lemma 2.7: Note that the case where either u
0
or u
n
commutes with z
follows from the case where neither do; therefore assume [u
0
, z], [u
n
, z] ,= 1.
For 0 k n, let
g
k
= u
0
z
i
1
u
1
z
i
2
u
2
. . . u
k1
z
i
k
u
k
.
Note that g
n
= g. Also, set g
1
= 1. Every non-trivial element of F acts
hyperbolically on T; x some x Axis(z) T.
Let be the path in T associated with the expression for g, as above. To
be precise, for 1 k n let

k
= [g
k1
x, g
k1
z
i
k
x]
and

k
= [g
k1
z
i
k
x, g
k
x].
Also write
0
= [x, g
0
x]. Then
=
0

1

1
. . .
n1

n

n
.
Let l
c
(w) denote the cyclically reduced length of an element w F (with respect
to the xed generating set). There are some useful observations to be made
about
k
and
k
.
1. The length of
k
is l
c
(z
i
k
) = [i
k
[l
c
(z). The length of
k
is d(g
k1
z
i
k
x, g
k
x) =
d(x, u
k
x).
2. Each
k
is contained in g
k1
Axis(z).
5
3. No
k
is contained in any single translate of Axis(z). The start point of

k
lies in g
k1
Axis(z). The end point of
k
lies in g
k
Axis(z). Let x
k
be
the last point of
k
in g
k1
Axis(z). Let y
k
be the rst point of
k
in
g
k
Axis(z).
Henceforth, assume [i
k
[l
c
(z) > d(x, u
k1
x) +d(x, u
k
x) whenever 1 k n.
The precise statement from which the lemma follows is
gAxis(z) ,= Axis(z).
Suppose gAxis(z) = Axis(z). Since T is a tree, x
k
= y
k1
for some k; otherwise,
[x, x
0
] [x
0
, y
0
] [y
0
, x
1
] . . . [y
n1
, x
n
] [x
n
, y
n
] [y
n
, x]
is a non-trivial loop.
By denition y
k1
lies in the image of
k1
; since g
k1
x is an end-point of

k1
it is immediate that
d(y
k1
, g
k1
x) d(x, u
k1
x).
Likewise the image of
k
contains x
k
. Therefore
d(x
k
, g
k1
z
i
k
x) d(x, u
k
x).
Now the triangle inequality gives
[i
k
[l
c
(z) = d(g
k1
x, g
k1
z
i
k
x)
d(y
k1
, g
k1
x) +d(x
k
, g
k1
z
i
k
x)
d(x, u
k1
x) +d(x, u
k
x)
contradicting the previous assumption on [i
k
[.QED
2.5 Surface groups
This subsection is devoted to showing that the fundamental groups of closed
orientable surfaces are fully residually free.
In fact, a larger class of groups than just surface groups are shown to be
limit groups. Let F be a free group of rank at least 2, and x a cyclic subgroup
z) F that is closed under taking roots. Consider the amalgam
F
z
F.
Such groups are called pinched word groups.
Example 2.8 Let be the closed surface of genus 2. Then theres an obvious
simple closed curve that realizes
1
() as a pinched word group.
Since all orientable surfaces of higher genus cover the surface of genus 2, it
suces to show that
1
() is fully residually free.
6
Proposition 2.9 Pinched word groups are fully residually free.
The assumption that z) is closed under taking roots is necessary; otherwise
G is not commutative transitive.
A key part of the proof will be the automorphisms naturally associated to
an abelian splitting, called Dehn twists.
Denition 2.10 Suppose G = A
C
B, and z B centralizes C. Then the
Dehn twist associated to z is the automorphism
z
dened by

z
(a) = a
for a A and

z
(b) = zbz
1
for b B.
Similarly, if G = A
C
and z A centralizes C, the Dehn twist associated
to z is the automorphism
z
dened by

z
(a) = a
for a A and

z
(t) = tz
for t the stable element.
Proof of proposition 2.9: Without loss, assume F = F. Let X G 1 be
nite. It suces to show that there exists a homomorphism h : G F such
that 1 / h(X).
Let f : G F be the obvious retraction; let
z
: G G be the Dehn twist
in z which is the identity on the rst factor.
Consider x X. By the structure theorem for amalgamated free products,
x has normal form
x = a
0
z
i
1
b
1
z
j
1
a
1
. . . b
n1
z
j
n1
a
n
z
i
n
b
n
where the a
k
lie in the rst copy of F, the b
k
lie in the second copy, i
k
, j
k
Z,
and furthermore:
1. for k > 0, a
k
F z);
2. for k < n, b
k
F z).
Then for m Z,

m
z
(x) = a
0
z
i
1
+m
b
1
z
j
1
m
a
1
. . . b
n1
z
j
n1
m
a
n
z
i
n
+m
b
n
z
m
,
so
f
m
z
(x) = a
0
z
i
1
+m
b
1
z
j
1
m
a
1
. . . b
n1
z
j
n1
m
a
n
z
i
n
+m
b
n
z
m
7
where all terms are now thought of as elements of F. Since the a
k
and b
k
dont commute with z (for k > 0 and k < n respectively), this expression for
f
m
z
(x) satises the conditions of lemma 2.7. So for suciently large m, the
result follows. QED
Note that this proof has not used the full power of the free group criterion.
The same proof would work for any double
G = L
Z
L
where L is fully residually free and Z is maximal abelian in L. Eventually,
continuing inductively in this vein, we will obtain a complete recursive charac-
terization of fully residually free groups.
3 Makanin-Razborov Diagrams and Algebraic
Limit Groups
3.1 Main results and denitions
We return now to our study of Hom(G, F). To state the main theorem properly,
though, a denition is needed.
Denition 3.1 A sequence of homomorphisms (f
i
: G F) is convergent if,
for any g G, f
i
(g) is eventually either always trivial or always non-trivial. To
a convergent sequence is associated the stable kernel ker

f
i
of elements that are
eventually trivial.
An (algebraic) limit group (over F) is any group of the form
L = G/ker

f
i
for (f
i
) a stable sequence of homomorphisms G F.
Example 3.2 Any fully residually free group is a limit group. Suppose G is
fully residually free, and let
S
1
S
2
. . . G
be an exhaustion of G by nite sets. Then for each i there exists a homomor-
phism f
i
: G F injective on S
i
. Then (f
i
) is a convergent sequence with trivial
stable kernel.
Here are the two main theorems of this series of talks.
Theorem 3.3 (Finite width) For any nitely generated group G that isnt
free there exists a nite collection of proper epimorphisms
q
i
: G L
i

with each L
i
a limit group, such that any homomorphism G F factors through
one of the q
i
, after precomposing with some automorphism of G. If G is not a
limit group, no precomposing is necessary.
8
Theorem 3.4 (Finite length) Any sequence of epimorphisms of limit groups
L
1
L
2
. . .
eventually stabilizes.
Iterating the construction of theorem 3.3 gives a tree of epimorphisms through
through which any homomorphism G F factors, twisting with automor-
phisms at each stage. By theorem 3.4, the branches of the tree end in free
groups after nitely many epimorphisms.
3.2 Finite length
Well start with the proof of theorem 3.4, which, over free groups, admits a neat
simplication.
Lemma 3.5 Consider a sequence of epimorphisms
G
1
G
2
. . . .
The corresponding sequence of monomorphisms
Hom(G
1
, F) Hom(G
2
, F) . . .
eventually stabilizes.
Proof: Identifying F with the fundamental group of a hyperbolic punctured
sphere exhibits an embedding
F PSL
2
(R),
which lifts to an embedding F SL
2
(R); this in turn induces an embedding
Hom(G
i
, F) Hom(G
i
, SL
2
(R))
for each i.
Pick a presentation
s
1
, . . . , s
n
[r
1
, r
2
, . . .)
for G. Any f Hom(G, SL
2
(R)) can be thought of as a set of choices for
f(s
1
), . . . , f(s
n
) that satisfy the relation r
1
, r
2
, . . .. These relations are polyno-
mial conditions in SL
2
(R), so Hom(G
i
, SL
2
(R)) is identied with a subvariety
of SL
2
(R)
n
. The decreasing sequence of subvarieties
Hom(G
1
, SL
2
(R)) Hom(G
2
, SL
2
(R)) . . .
terminates by Hilberts Basis Theorem. QED
Proposition 3.6 Limit groups are fully residually free.
9
Proof: Let L be a limit group, and let G and f
i
be as in the denition. Consider
a (generally innite) sequence of epimorphisms
G = G
0
G
1
. . . L
obtained by adding one relation at a time. Let G

be such that
Hom(G

, F) = Hom(L, F).
Then all but nitely many f
i
factor through G

, since each added relation is

killed by almost all f
i
. Therefore all but nitely many f
i
factor through L, and
each non-trivial element of L is killed by only nitely many f
i
. Therefore L is
fully residually free. QED
This shows that the classes of algebraic limit groups and fully residually free
groups coincide, and allows us to prove theorem 3.4.
Proof of theorem 3.4: Consider a proper epimorphism
q : L L

for L residually free. Then

Hom(L, F) Hom(L

, F);
for given k L with q(k) = 1, there exists f Hom(L, F) with f(k) ,= 1, so
f / Hom(L

, F). The theorem now follows from lemma 3.5 and the fact that
limit groups are residually free. QED
3.3 An attempt to prove nite width
Having proved the nite length of the Makanin-Razborov diagram, we turn to
its nite width. The idea is to compactify Hom(G, F), following ideas contained
in [6].
To this end, let Q(G) be the set of epimorphisms
q : G H,
where q
1
: G H
1
and q
2
: G H
2
are regarded as equivalent if there is an
isomorphism i : H
1
H
2
with q
2
= i q
1
. Alternatively, Q(G) can be regarded
as the set of normal subgroups of G, and we will usually work from this point
of view. Theres an obvious map
Hom(G, F) Q(G)
that sends each homomorphism to its kernel. To compactify Hom(G, F), there-
fore, we start by topologising Q(G).
The interpretation of Q(G) as the normal subgroups of G identies it with
a subset of 2
G
, the power set of G. Equip 2
G
with the product topology
10
the weakest topology with respect to which the projections onto the factors are
continuous. To work with this topology, we need to understand it better. Fix a
nite generating set for G, and let B
n
G be the ball of radius n in the word
metric. For a subset X G, and a positive integer n, consider the set of subsets
U(X, n) = Y G[B
n
X = B
n
Y .
Lemma 3.7 The topology induced by the neighbourhoods U(X, n) is the product
topology.
Proof: The product topology is generated by the sub-basis consisting of the
open sets
V (g) = Y G[g Y
together with
V

(g) = Y G[g / Y
for all g G. Now, for any X G and n,
U(X, n) =

gB
n
X
V (g)

gB
n
X
V

(g).
Conversely, for g G of word-length n,
V (g) =

gXB
n
U(X, n)
and
V

(g) =

g/ XB
n
U(X, n).
QED
These equivalent perspectives make it easy to read o the properties of the
topology.
Theorem 3.8 Consider 2
G
endowed with the product topology.
1. 2
G
is compact.
2. 2
G
is metrizable.
3. A sequence X
i
G converges if and only if every g G is eventually
always in X
i
or eventually not in X
i
. The limit is the set of elements that
are eventually in X
i
.
4. Q(G) 2
G
is closed, hence compact.
Proof: Part 1 is Tychonos theorem.
Given X, Y G, dene
d(X, Y ) = e
n
where n is the greatest integer such that B
n
X = B
n
Y . This gives 2.
11
Part 3 is just exactly what it means to converge in the usual denition of
the product topology.
Suppose now that X
i
is a sequence of normal subgroups, converging to X
G. Then for g, h G, let n be greater than the length of g, h and gh. Then for
all suciently large i,
g, h X B
n
X
i
and
gh X
i
B
n
X,
so X is a subgroup. Similarly, for g G and h X, ghg
1
X, so X is normal.
This proves 4. QED
Note that it follows from 3 that
G/ ker f
i
G/ker

f
i
in Q(G). In particular, the set of limit groups for G could equivalently be dened
as the closure of the free groups in Q(G).
Remark 3.9 In a sense, the philosophically correct interpretation of parts 2,3
and 5 of proposition 2.5 is that being torsion-free, commutative transitive and
CSA are closed properties in Q(G).
We are now ready to attack theorem 3.3, in the case where G is not a limit
group.
Proof of theorem 3.3 when G is not limit group: Consider F(G), the closure
of the set of free groups in Q(G). Since G is not a limit group, every epimor-
phism in F(G) is proper. So V (g)[g G is an open cover for F(G). Let
V (g
1
), . . . , V (g
n
)
be a nite subcover. Then every homomorphism
f : G F
factors through one of the quotients
q
i
: G G/g
i
)).
The only complication is that the quotients q
i
(G) may not be limit groups.
Replacing G with its residually free quotient
G/

f:GF
ker f
it can be assumed that G is residually free. Now Hom(q
i
(G), F) is strictly
contained in Hom(G, F), so one may apply lemma 3.5 and induction to conclude
that each q
i
(G) has a collection of nitely many proper quotients through which
any homomorphism to F factors. QED
12
This argument fails when G is a limit group, since then there exist convergent
sequences with trivial stable kernel, so V (g)[g G is no longer a cover of F(G).
We need to nd a subspace F

(G) F(G), covered by V (G)[g G, such that

any q F(G) is related to some q

(G) by an automorphism. To do this,

we use a geometric analogue of the techniques of this section.
4 Geometric Limit Groups
4.1 The space of trees
Recall the basic denitions and results of the theory of trees.
Denition 4.1 A (real) tree is a geometric metric space (T, d) in which every
geodesic triangle is a tripod. Equivalently, T is geodesic and satises Gromovs
four-point condition:
d(w, x) +d(y, z) min(d(w, y) +d(x, z), d(w, z) +d(x, y))
whenever w, x, y, z T. For details of the equivalence of these denitions see,
for example, [5].
A (real) G-tree is a real tree with an action of G by isometries.
A tree T is non-degenerate if it is not a point. A G-tree T is trivial if G xes a
point of T. T is minimal if it contains no proper G-invariant subtrees.
Lemma 4.2 If T is non-trivial then T contains a separable unique minimal
subtree.
For the proof of this lemma see, for example, [3].
Consider the set A(G) of non-trivial minimal real G-trees. A(G) is endowed
with the equivariant Gromov-Hausdor topology, dened as follows. A sequence
(T
n
, d
n
) of G-trees converges to a G-tree (T, d) if and only if, for any > 0
and any nite subsets K T, P G, there exist K
n
T
n
and bijections
b
n
: K
n
K such that
[d
n
(gb
n
(x
n
), b
n
(y
n
) d(gx
n
, y
n
)[ <
whenever x
n
, y
n
K
n
and g P. This can be thought of as saying that larger
and larger subtrees of T
i
coincide with subtrees of T.
Let PA(G) be the projectivization of A(G), so (T, d) is identied with (T, d)
for all > 0.
Fix a set of generators for F with respect to which the Cayley graph is a
tree T
F
with the word metric d
F
. A homomorphism f : G F denes a G-tree,
with the action given by left multiplication. Since G/ ker f acts freely, if f ,= 1
then T
f
is non-trivial. Let T
f
PA(G) be the equivalence class of the minimal
G-invariant subtree. The space of interest is T(G), the closure of
T
f
[1 ,= f Hom(G, F)
13
in the equivariant Gromov-Hausdor topology.
Compactness of the space T(G) was proved rst in [10], in which the limits
are constructed using convex hulls. A quicker and more general method, though,
is to use non-standard analysis.
4.2 Ultraproducts
An ultralter is a nitely additive probability measure
: 2
N
0, 1
An ultralter is non-principal if, whenever S N is nite, (S) = 0.
Lemma 4.3 Non-principal ultralters exist.
For this and all subsequent results in this section, see chapter I.5 of [5].
Let X be any topological space. For points x
n
, x X, write lim

x
n
= x if,
for any open set x U X,
n N[x
n
U = 1.
The point x is called the ultralimit of the sequence x
n
(with respect to ).
Lemma 4.4 Fix a non-principal ultralter. If X is a compact metric space
then every sequence has an ultralimit (with respect to ).
Proof: Let x
n
be a sequence in X, and suppose x
n
has no ultralimit. Then
every x X has an open neighbourhood U
x
with n N[x
n
U
x
= 0. Now
U
x
[x X is an open cover of X; let U
x
1
, . . . , U
x
n
be a nite subcover. But
then
1 = (N)

i
n[x
n
U
x
i
= 0
a contradiction. QED
Let (X
n
, d
n
, x
n
) be a sequence of pointed metric G-spaces (
n
denotes the
G-action). Let
Y

n
X
n
be the subspace of sequences (y
n
) such that d
n
(x
n
, y
n
) is uniformly bounded.
The space Y admits a pseudometric dened by
D((y
n
), (z
n
)) = lim

d
n
(y
n
, z
n
).
The associated metric space is denoted (X

, d

) and is called the ultraproduct

of the sequence (X
n
, d
n
, x
n
).
Lemma 4.5 Let be an ultralter and (X
n
, d
n
, x
n
), (X

, d

) as above.
1. If the (X
n
, d
n
) are geodesic then so is (X

, d

).
14
2. If each (X
n
, d
n
) is an R-tree then (X

, d

).
Proof: Consider elements y = [(y
n
)], z = [(z
n
)] X

. Assuming each X
n
is
geodesic, let
n
: [0, d
n
(y
n
, z
n
)] X
n
be a geodesic from y
n
to z
n
. Dene
: [0, d

(y, z)] X

by
t

d
n
(y
n
, z
n
)
d

(y, z)
t

.
Note that this is well-dened since, for any point w
n
on
n
,
d
n
(x
n
, w
n
) 2d
n
(x
n
, y
n
) +d
n
(x
n
, z
n
).
Moreover, is a geodesic, since for s, t [0, d

(y, z)],
d

((s), (t)) = lim

d
n

d
n
(y
n
, z
n
)
d

(y, z)
s

,
n

d
n
(y
n
, z
n
)
d

(y, z)
t

= [s t[
as required. This proves 1.
Assertion 2 follows immediately from Gromovs four-point condition. QED
Suppose each (X
n
, d
n
) admits a G-action. Then the action GY Y given
by
g(y
n
) = (g.y
n
)
descends to an action on (X

, d

) by isometries.
Ultraproducts are useful in this context because they provide limits in the
Gromov topology.
Lemma 4.6 If (X
n
) is a sequence of G-spaces let X

be the ultraproduct for

some choice of base-points. Suppose X X

is a separable G-equivariant sub-

space. Then some subsequence of (X
n
) converges to X in the Gromov topology.
Proof: Let S X be a countable dense subset, and let
S
1
S
2
. . . S
be an exhaustion of S by nite subsets. Let
P
1
P
2
. . . G
be an exhaustion of G by nite subsets. Dene I
n
N to consist of those i N
for which
[d
i
(gx
i
, y
i
) d(gx, y)[ <
1
n
for all g P
n
and x, y S
n
. By denition, (I
n
) = 1. Let n
1
be the least
element of I
1
, and inductively dene n
i
to be the least element of I
i
not to be
contained in n
1
, . . . , n
i1
. The subsequence (X
n
i
) now converges to X in the
Gromov topology. QED
15
4.3 Compactness of T(G)
The only remaining tricky detail is to ensure that the ultralimit is non-trivial.
This is achieved by carefully controlling the base-points and scaling the metric,
and is the key trick of the Bestvina-Paulin method.
Let f
n
: G F be a sequence of homomorphisms, and T
n
the corresponding
sequence of minimal G-trees with the usual word metric d
F
. Fix a generating
set S for G. Consider the function
n
: T
n
R given by

n
(x) = max
gS
d
F
(x, f
n
(g)x);
let

n
= inf
xT
n

n
(x).
Each tree T
n
is simplicial, and the function
n
is integer-valued on vertices and
mid-points of edges, and linear in between. Therefore
n
attains its inmum on
T
n
, say at x
n
. Since S is nite, for some g
0
S,
n N[d
n
(x
n
, g
0
x
n
) = 1 = 1;
that is, g
0
-almost always realizes the maximum in the denition of
n
(x
n
).
Equip T
n
with the modied metric d
n
= d
F
/
n
. By the results of the previous
section, the ultralimit (T

, d

) of the sequence
(T
n
, d
n
, x
n
)
is a metric G-tree. Furthermore, consider a point y = [(y
n
)] T

. Then
d

(y, g
0
y) = lim

d
n
(y
n
, g
0
y
n
) lim

d
n
(x
n
, g
0
x
n
) = 1;
in particular, g
0
y ,= y, so T

is non-trivial. Let T T

be the minimal G-
invariant subtree, which is separable. Then a subsequence of T
n
converges to T
in the equivariant Gromov-Hausdor topology. This is a limit for the sequence
T
n
in T(G). Henceforth, denote the metric on T by d.
4.4 Geometric limit groups
Denition 4.7 Let T be a real G-tree in T(G) as constructed in the previous
section; let ker T be the kernel of the action of G on T. A geometric limit group
is any group of the form
G/ ker T
for such a T.
To extract information from this geometric picture, a careful analysis of
equivariant Gromov-Hausdor is needed. The results of this analysis are sum-
marized in the following technical theorem.
16
Theorem 4.8 Let T
n
be a sequence of non-trivial minimal G-trees arising from
homomorphisms f
n
: G F as above, converging to a non-trivial minimal tree
T. Then the following hold.
1. If T is not a line then ker

(f
n
) = ker T.
2. The stabilizer in G/ ker T of a tripod is trivial.
3. Stabilizers in G/ ker T of non-degenerate arc in T are free abelian.
4. If J I are non-degenerate arcs in T then Stab
G/ ker T
(I) = Stab
G/ ker T
(J).
5. T is a line if and only if, for all suciently large n, f
n
has non-trivial
abelian image.
The proof of this theorem is omitted.
The rst consequence of the theorem is that the two dierent notions of limit
group coincide.
Corollary 4.9 The set of algebraic limit groups and the set of geometric limit
groups coincide.
Proof: If L is a nitely generated free abelian group then L is both an algebraic
and a geometric limit group. It follows from parts 1 and 5 of the theorem that,
for L non-abelian, the two notions coincide. QED
Furthermore, the theorem ensures that T has the properties we shall require
to apply Rips theory to study the action in detail.
Denition 4.10 A real G-tree T is stable if, for every descending sequence of
non-degenerate subtrees
T T
1
T
2
. . .
the corresponding sequence of pointwise stabilizers
Stab
G
(T) Stab
G
(T
1
) Stab
G
(T
2
) . . .
eventually stabilizes.
A G-tree is very small if it is non-trivial, minimal, stable, has abelian (non-
degenerate) arc stabilizers, and trivial tripod stabilizers.
Corollary 4.11 For a geometric limit group L = G/ ker T, the tree T is a very
small L-tree.
5 The Shortening Argument
5.1 Denitions and the statement
The shortening argument is a dicult trick that closely analyses the action of
G on T, to force the kernel to be non-trivial. In this section, well just try
to outline some of the argument, and complete the proof of the nite width
theorem.
17
Denition 5.1 A generalized abelian decomposition for group G is a nite
graph of groups with abelian edge groups and three classes of vertices.
1. Surface vertices are the fundamental groups of compact surfaces with bound-
ary. It is required that the surface carry a pseudo-Anosov automorphism;
that is, it is either a torus with a single boundary component, or has Eu-
ler characteristic at most -2. Edges adjoining surface vertices are innite
cyclic, and identied with the fundamental groups of boundary components.
2. Abelian vertices are nitely generated abelian groups. For A an abelian
vertex, let P(A) be the subgroup generated by incident edge groups. Then
dene the peripheral subgroup to be
P(A) =

ker(f)[f Hom(A, Z), f(P(A)) = 0.

3. All other vertices are designated rigid.
Denition 5.2 Let be a generalized abelian decomposition for G. The asso-
ciated modular group Mod() is the subgroup of Aut(G) generated by:
1. inner automorphisms of G;
2. Dehn twists of edges of ;
3. unimodular (that is, determinant 1) automorphisms of abelian vertices A,
which are the identity on the peripheral subgroup;
4. automorphisms of surface vertices arising from automorphisms of the un-
derlying surface that x boundary components (note that these induce well-
dened group automorphisms, since the base-point can be taken in a bound-
ary component).
The modular group of G, Mod(G), is the group of automorphisms of G generated
by the modular automorphisms of all generalized abelian decompositions of G.
Denition 5.3 A homomorphism f : G F is equivalent to all homomor-
phisms of the form i f where i Inn(F) and Mod(G).
Fix a generating set S for G. The length of f is dened to be
[f[ = max
gS
l(f(g)),
where l is word-length in F.
A homomorphism f : G F is short if its length is minimal in its equiva-
lence class.
This is precisely the notion we need.
Theorem 5.4 (See [11]) Suppose G is freely indecomposable. Let f
i
: G F
be a convergent sequence of short homomorphisms. Then
ker

f
i
,= 1.
Suppose ker

f
i
= 1. Then G has a very small action on T. The plan is to use
this action to nd an automorphism to shorten the f
i
.
18
5.2 A avour of the proof
The proof relies heavily on the Rips theory of nitely-generated-group actions
on trees. Bestvina and Feighn, in [4], give a more delicate version of the short-
ening argument that only requires Rips theory for nitely presented groups. An
alternative proof would involve quoting Guirardels result that limit groups are
nitely presented, then applying the usual shortening argument, using only the
nitely presented Rips theory.
We start with the simplest examples of non-simplicial free group actions on
trees.
Example 5.5 (Abelian type) Suppose T

= R and G

= Z
n
. There exists
a faithful homomorphism G Isom(R) mapping each generator to something
algebraically independent of the other generators.
Example 5.6 (Surface type) Let be a surface (or, more generally, a 2-
orbifold), and F a minimal foliation on with transverse measure . This
induces a pseudometric on

given by
d(x, y) = inf

()
where lifts to a path from x to y. The associated metric space is a tree, on
which
1
() acts freely.
In fact, these are the only examples we need to worry about.
Theorem 5.7 Let G be a nitely generated, freely indecomposable group with
an very small action on a minimal tree T. Then T is covered by orbits of a
nite collection of subtrees T
1
, . . . , T
n
such that:
1. gT
i
T
j
is at most one point if i ,= j;
2. gT
i
T
i
is either T
i
or at most one point;
3. the action of G
i
= Stab(T
i
) on T
i
is either of abelian type, or of surface
type, or simplicial.
G has a graph of groups decomposition with:
1. vertices corresponding to orbits of points of T with non-trivial stabilizer;
2. vertices corresponding to orbits of T
i
;
3. edges corresponding to orbits of simplicial edges of T
i
;
4. edges corresponding to points of intersection of orbits of T
i
.
The idea is to nd automorphisms that shorten the length of a geodesic
[x, gx], for any generator g. Well just do a couple of cases.
19
Suppose T is a tree of abelian type. Fix > 0. Suppose g
1
is the generator
with the longest translation length, and g
2
has the second longest translation
length. Then there exists k such that
[g
1
+kg
2
[ < [g
1
[.
Now replace g
1
by g
1
+kg
2
. Proceeding in this manner, every generator can be
given translation length less than . Since the translation length of a generator
in the approximating simplicial trees converges to the translation length in the
limit tree, this contradicts the assumption that the f
i
are short.
It is also true that, in the surface case, the translation lengths of generators
in the limit can be made arbitrarily small.
Now suppose T is the Bass-Serre tree of a splitting of the form
G = A
C
B
for abelian C. In this case theres a uniform lower bound on the translation
lengths of elements in T, so we have to shorten in the approximating spaces.
Let e be the edge xed by C. Assume x is xed by A, and y is the vertex
xed by B.
Any generator is of the form
g = a
0
b
1
a
1
b
n
a
n
where [g[ = 2n. Fix z C. The segments approximating e become arbitrarily
close to the axes of z. Then there exist m(n) so that
f
n
(z
m(n)
)y
n
x.
For large n, d
n
(x
n
, f
n
(g)x
n
) is approximately 2n, while d
n
(x
n
, f
n

m(n)
z
(g)x
n
)
is approximately 0.
5.3 The proof of nite width
Armed with Selas shortening argument, we are now in a position to prove the
nite width theorem.
Proof of theorem 3.3: If G is abelian, then any homomorphism to F factors
through projection onto a factor. So it can be assumed that G is non-abelian.
Suppose the theorem is proved for freely decomposable G. If G = G
1
G
2
is
a non-trivial free product decomposition of G then Hom(G, F) Hom(G
1
, F).
By induction, every homomorphism G
1
F factors through some nite set
q
i
: G
1
L
i

of proper factors, so every homomorphism G F factors through

q
i
id
G
2
: G L
i
G
2
.
20
So G can be assumed freely indecomposable.
Let L T(G) be the subspace of linear subtrees. Theorem 4.8 implies that
L(G) is open in T, and
V (g) = T T(G) L[g ker T
is open. Consider the subspace T

(G) T(G), the closure of the space of

trees arising from short homomorphisms G F. By the shortening argument,
U(G)[g GL is an open cover of T

(G). By compactness, therefore, there

exists a nite subcover
U(g
1
), . . . , U(g
n
), L.
Now any short homomorphism G F factors through one of
G G/g
i
)) G G/[G, G].
The argument is now concluded as in our rst attempt.QED
6 JSJ Decompositions
6.1 Splittings of limit groups
A JSJ decomposition is, loosely, a universal splitting for a group G. To con-
struct JSJ splittings for limit groups, therefore, we will need to understand their
splittings, or equivalently, their actions on simplicial trees. Because limit groups
are CSA, the picture is greatly simplied. Here is the key lemma.
Lemma 6.1 Consider a one-edge splitting of a limit group G over an abelian
subgroup, and M G a non-cyclic maximal abelian subgroup.
1. If G = A
C
B then M is conjugate into A or B.
2. If G = A
C
and M is not conjugate into a A then for some conjugate M
g
of M,
G = A
C
M
g
.
Note: we dont yet know that abelian subgroups of G are nitely generated.
Proof: Suppose G = A
C
B, and let T be the Bass-Serre tree. Assume M is
not conjugate into either A or B. Because M is abelian, it follows from a coarse
classication of group actions on trees that H either xes a line in T, or a point
on the boundary. If it xed a point on the boundary, then there would be an
increasing sequence of edge stabilizers
C
1
C
2
. . . M.
But each C
i
is conjugate to C and M is malnormal, a contradiction; so H xes
a line in T, called the axis of M. Conjugating if necessary, C is the stabilizer
of an edge in the axis. But M acts as
M = M

Z
21
where M

xes the axis, so M

C and, by commutative transitivity, C M.

So C xes the whole axis. But there is only one orbit of edges, so there exists
a AC with aCa
1
= C. Since M is malnormal, a M, a contradiction.
Now suppose G = A
C
, and M is not conjugate into A. As before, M
preserves an axis in the Bass-Serre tree T, C can be assumed to lie in M and
x the axis, and A M = C. Now if t is the stable letter of the splitting, then
without loss of generality, tCt
1
= C. Since M is malnormal, it t M and the
result follows. QED
We will see that this lemma tells us that non-cyclic abelian splittings of limit
groups only interact in a very simple way.
Denition 6.2 A simplicial G-tree is k-acylindrical if the xed point set of any
g G has diameter at most k.
Lemma 6.3 If G is a limit group and T is a simplicial G-tree with abelian edge
stabilizers then, without loss of generality, T is 2-acylindrical.
Proof: Let C G be the stabilizer of an edge e of T. Since G is commutative
transitive, C lies in a unique maximal abelian subgroup M. By the previous
lemma, M can be assumed to x a vertex v of T. Now after a slide, e adjoins
v. So every element of C can only x edges adjacent to v. QED
6.2 The denition
For a one-edge splitting , g G is called -elliptic if g acts elliptically on the
associated Bass-Serre tree, or equivalently if g is conjugate into a vertex group
of . If S is a set of splittings of G then g G is S-elliptic if it is -elliptic for
all S.
Let be a generalized abelian decomposition for G. Then g G is -elliptic
if:
1. g is conjugate into a vertex group of ;
2. if the vertex is surface then g is conjugate into a boundary component;
3. if the vertex is abelian then g is conjugate into the peripheral subgroup.
If g is not -elliptic then there is an obvious one-edge splitting of G such that
g is no -elliptic.
Consider the set A of one-edge splittings of G satisfying:
1. the edge group is abelian;
2. the edge group is closed under taking roots.
Denition 6.4 An abelian JSJ decomposition for G is a generalized abelian
decomposition for G such that the set of -elliptics is the set of A-elliptics.
22
The remainder of this section is devoted to explaining why the existence
of abelian JSJ decompositions for limit groups is much simpler than it is for
general groups. In fact, the cyclic splittings are as hard as in the general case,
but the non-cyclic abelian subgroups are covered by lemma 6.1.
By lemma 6.1, every splitting of A is, at least, closely related to a splitting
in the subset A

A of splittings in which every non-cyclic abelian subgroup is

elliptic.
6.3 Non-intersecting splittings
Let
1
,
2
be a pair of non-trivial one-edge splittings of a group G, with free
abelian edge groups C
1
, C
2
respectively.
1
is
2
elliptic if C
1
is conjugate into
a vertex group of
2
; otherwise,
1
is
2
-hyperbolic.
Lemma 6.5 If
1
is
2
-elliptic then there exists a splitting such that the set
of -elliptics is equal to the intersection of the sets of
1
- and
2
-elliptics.
Proof: Assume that
1
is an amalgamated free product; the HNN-extension
case is similar. So
G = A
1

C
1
B
1
= A
2

C
2
B
2
.
Suppose C
1
is conjugate into A
2
; conjugating A
2
if necessary, it can be assumed
that C
1
A
2
.
2
induces splittings of A
1
and B
2
, denoted
A
2
and
B
2
respec-
tively. Without loss of generality, A
1
A
2
can be assumed to be a vertex group
of
2
A
; and likewise B
1
A
2
can be assumed to be a vertex group of
2
B
. Since
C
1
A
1
A
2
, B
1
A
2
,
A
2
and
B
2
can be joined by an edge stabilized by C
1
to produce , as required. QED
The graph of groups is called the renement of
1
by
2
. Note that a
similar construction can be carried out when
1
and
2
have more than one
edge.
The pair (
1
,
2
) is elliptic-elliptic of
1
is
2
-elliptic and vice versa. Like-
wise, such a pair can be elliptic-hyperbolic, hyperbolic-elliptic, and hyperbolic-
hyperbolic. Elliptic-elliptic pairs should be thought of as disjoint; hyperbolic-
hyperbolic pairs should be thought of as intersecting. The other two possibilities
can be ruled out, under certain assumptions.
Lemma 6.6 Let
1
and
2
be as above, and suppose G does not split over
any innite-index subgroups of C
2
. Then (
1
,
2
) is either elliptic-elliptic or
hyperbolic-hyperbolic.
Proof: Suppose
1
is
2
-elliptic, and
2
is
1
-hyperbolic. As before, assume
1
is an amalgamated free product; the HNN-extension case is similar. Let be
the renement of
1
by
2
.
Suppose the graphs of groups
A
2
,
B
2
are both trivial; so A
1
and B
1
are
both conjugate into vertex groups of
2
. These vertex groups must be distinct,
since
2
is a non-trivial splitting; therefore C
1
is conjugate into C
2
, and must
be conjugate to a nite index subgroup; but then C
2
must be conjugate into a
vertex of
1
.
23
Assume, therefore, that
A
2
is non-trivial; an edge group is a conjugate of a
subgroup of C
2
, and also a subgroup of A
1
. This subgroup of C
2
must be of
innite index; since it is also an edge group of , this contradicts the assumption
that G doesnt split over innite-index subgroups. QED
The would like to iteratively apply lemma 6.5 to construct the JSJ decom-
position; but we dont yet know how to deal with hyperbolic-hyperbolic pairs
of splittings.
6.4 Intersecting splittings
A set of splittings is called intersecting if, for any splittings ,

, there exists a
nite chain of splittings
=
1
, . . . ,
n
=

such that (
i
,
i+1
) is a hyperbolic-hyperbolic pair of splittings.
Henceforth, assume G is freely indecomposable. Suppose S A

is a set of
intersecting abelian splittings. If has cyclic edge group and

has non-cyclic
abelian edge group, then

is assumed to be -hyperbolic. Now by lemma 6.6,

is

-hyperbolic.
The model example of an intersecting pair of splittings is a pair of intersect-
ing simple closed curves on a surface. In fact, this is the only example!
Theorem 6.7 Let S be a maximal set of intersecting innite-cyclic splittings of
nitely generated G. Then there exists a graph of groups
S
for G with cyclic
edge groups and an orbifold vertex S such that every splitting in S arises by
splitting S along a simple closed curve, and every edge group is identied with
a boundary component or cone point of S.
The intersection of the sets of -elliptics for all S is the set of elements
that are conjugate into a vertex of
S
other than S.
S is called an enclosing vertex for S. By assumption, every pair of non-cyclic
splittings in A

is elliptic-elliptic.
6.5 Acylindrical accessibility
To construct the decomposition, now, we simply proceed by induction, splitting
repeatedly and replacing intersecting sets of splittings with surface vertices. We
need to know that this process terminates. If we knew G was nitely pre-
sented, then we could apply a relatively simple accessibility result of Bestvina
and Feighn. Since we dont know that G is nitely presented, we have to use a
dierent approach.
Theorem 6.8 (Sela, [17]) Let G be a non-cyclic freely indecomposable nitely
generated group and a k-acylindrical graph of groups for G. Then there is a
uniform bound on the number of vertices in the core of .
By lemma 6.3, every splitting of A

, their renements the graphs of their

enclosing vertices can be assumed to be 2-acylindrical.
24
6.6 The construction
In this subsection, we prove the existence of abelian JSJ decompositions for
limit groups.
Theorem 6.9 Every freely indecomposable limit group G has an abelian JSJ
decomposition.
Proof: Associated to any one-edge splitting A

, there is the maximal inter-

secting set of splittings S. If [S[ > 1 then is a splitting over Z, and there is
the associated graph
S
. Otherwise, set
S
to be .
We construct the JSJ decomposition inductively, as follows. Let
1
be
a one-edge splitting in A

. Let S
1
be the maximal intersecting set of one-edge
splittings containing
1
. Set =
S
1
.
Now let
2
be a one-edge splitting in A

S
1
. Let S
2
be the maximal
intersecting set of splittings containing
2
. Then by denition, and

S
2
are
an elliptic-elliptic pair of splittings. Dene the new to be the core of the
renement of this pair. Now continue in this way iteratively.
Because every splitting of is without loss 2-acylindrical, it follows that this
process eventually terminates. The result is a JSJ-decomposition for G. QED
7 Constructive Limit Groups
7.1 The main theorem
Denition 7.1 Constructive limit groups (CLGs) are nitely generated and
dened inductively. A CLG of level 0 is a nitely generated free group. A group
G is a CLG of level at most n if one of the following holds.
1. G = G
1
G
2
, for G
1
, G
2
CLGs of level at most n 1.
2. G has a generalized abelian decomposition . is assumed to have nitely
generated vertex groups and edge groups; furthermore, each edge group is
assumed to be maximal abelian on one side of the associated one-edge
splitting. There exists a homomorphism : G G

, for G

a CLG of
level at most n 1, satisfying the following properties:
(a) is injective on edge groups;
(b) has non-abelian image on surface vertices;
(c) is injective on the peripheral subgroups of abelian vertices;
(d) is injective on the envelopes of rigid vertices.
Some properties, such as the existence of a nite presentation, are much
easier to prove for CLGs than for limit groups. So the next theorem has some
profound consequences.
Theorem 7.2 The sets of limit groups and constructive limit groups coincide.
25
The proof that all CLGs are limit groups is tricky but not very enlightening.
The idea is to twist with modular automorphisms of , and use lemma 2.7
to prove that G is fully residually free. The other direction, however, makes
heavy use of the shortening argument and JSJ decompositions. The rest of this
section is devoted to this direction of the proof.
7.2 Modular automorphisms
The most important observation is the relation of the abelian JSJ decomposition
to modular automorphisms.
Denition 7.3 Consider a one-edge splitting , of the form
G = A
C
B
or
G = A
C
.
A generalized Dehn twist associated to a splitting is a Dehn twist, or if A is
abelian a unimodular automorphism of A restricting to the identity on the edge
group (and B in the rst case).
Lemma 7.4 For any G, Mod(G) is generated by generalized Dehn twists in
one-edge splittings in A.
Proof: Surface automorphisms are generated by Dehn twists of the surface.
Unimodular automorphisms of abelian vertex groups A are generalized Dehn
twists in the obvious one-edge splittings of the form
G = A
P(A)
B.
Consider a one-edge splitting of G. If the edge group isnt closed under
taking roots on one side then there is an immediate contradiction of property 4
of lemma 2.5. Let A be a vertex group of , and suppose Z = Z
A
(C) C. Let

be the one-edge splitting obtained by expanding Z and contracting C. For

example, if is the splitting G = A
C
B then

is
G = A
Z
(Z
C
B).
Then the edge group of

is closed under taking roots, and any Dehn twist in

arises as a Dehn twist in

. QED
7.3 Abelian subgroups
Its fairly easy to see that abelian subgroups of CLGs are nitely generated. To
show that limit groups are constructive, well need to see that the same is true
of limit groups.
Proposition 7.5 Abelian subgroups of limit groups are nitely generated.
26
Lemma 7.6 If G is a limit group with factor set
q
i
: G G
i

and H G is any subgroup such that every homomorphism H F factors

through some q
i
[H (without pre-composing with automorphisms) then, for some
i, q
i
[H is injective.
Proof: Suppose not. Then there exists non-trivial h
i
ker(q
i
[
H
). But G is a
limit group, so there exists f : G H injective on 1, h
1
, . . . , h
n
. But f[
H
factors through some q
i
[
H
, so f(h
i
) = 1, a contradiction. QED
Proof of proposition 7.5: All limit groups are torsion free, so it remains to
show that any abelian subgroup A G is nitely generated. For Mod(G)
there exists a nitely generated subgroup A

A and a retraction r

: G A

such that [A agrees with an inner automorphism on Aker(r

). For, suppose
A is non-cyclic; then by lemma 6.1, A can be assumed to be conjugate to a
vertex in . Now, if its true for , Mod(G) then setting A

= (A

and
r

= r

, its true for too.

The homomorphism

Mod(G)
r

: G

has nitely generated image, since G is nitely generated. Therefore A = A

0

A
1
, where A
1
is nitely generated and each r

is trivial on A
0
. Now by lemma
7.6, for some homomorphism q in the factor set of G, q[
A
0
is injective. Now by
induction q[
A
0
is nitely generated, hence so are A
0
and A. QED
7.4 Limit groups are constructive
Were now ready to prove that limit groups are constructive.
Let G be a generic limit group, and x a generating set S. Let f
i
be a
sequence in Hom(G, F) such that f
i
is injective on elements of length at most
i in the length metric with respect to S. Then ker

f
i
= 1. Choose

f
i
to be
short maps equivalent to f
i
. By theorem 5.4, : G G

= G/ker

f
i
is a proper
epimorphism, and so by induction assume that G

is a CLG. Let be an abelian

JSJ decomposition for G. The claim is that and satisfy the conditions of
denition 7.1.
Let E be an edge group of the JSJ decomposition. Then it is elliptic in
every one-splitting of A, so all generalized Dehn twists coincide with some inner
automorphism on E. Therefore for g E 1,

f
i
(g) is conjugate to f
i
(g)
which is non-trivial for all suciently large i; so [
E
is injective. Moreover, E
must be maximal abelian in a vertex of the associated one-edge splitting, by
commutative transitivity.
Abelian vertices of are nitely generated and free by proposition 7.5. Let
P(A) be the peripheral subgroup. As in the edge group case, P(A) is elliptic
27
in all relevant one-edge splittings, so Mod(G) acts as inner automorphisms and
[
P(A)
is injective.
Let S be a surface vertex, and suppose (S) is abelian. Then for all su-
ciently large i,

f
i
(S) is abelian. But note that every element of Mod(G) maps
S to (a conjugate of) itself, so eventually f
i
(S) is abelian, contradicting the
triviality of ker

f
i
.
The envelope of a rigid vertex B is elliptic in every splitting of G, so is pre-
served up to conjugacy by Mod(G). Therefore [
B
is injective. This completes
the proof.
References
[1] E. Alibegovic. A combination theorem for relatively hyperbolic groups,
2003. Preprint.
[2] E. Alibegovic and M. Bestvina. Limit groups are CAT(0), 2004. Preprint.
[3] M. Bestvina. R-trees in geometry, topology and group theory. In Handbook
of geometric topology. North-Holland, Amsterdam, 2002.
[4] M. Bestvina and M. Feighn. Notes on Selas work: Limit groups and
Makanin-Razborov diagrams, 2003. Preprint.
[5] M. Bridson and A. Haeiger. Metric spaces of non-positive curvature, vol-
ume 319 of Grundlehren der mathematischen Wissenschaften. Springer,
1999.
[6] C. Champetier and V. Guirardel. Limit groups as limits of free groups:
compactifying the set of free groups, 2004. Preprint.
[7] F. Dahmani. Combination of convergence groups. Geom. Topol., 86, 2003.
[8] V. Guirardel. Limit groups and groups acting freely on R
n
-trees, 2003.
Preprint.
[9] R. S. Lyndon. The equation a
2
b
2
= c
2
in free groups. Michigan Math. J,
6, 1959.
[10] F. Paulin. Topologie de Gromov equivariante, structures hyperboliques et
arbes reels. Invent. Math., 94(1), 1988.
[11] E. Rips and Z. Sela. Structure and rigidity in hyperbolic groups I. GAFA,
4(3), 1994.
[12] Z. Sela. Diophantine geometry over groups III: Rigid and solid solutions.
Preprint.
[13] Z. Sela. Diophantine geometry over groups IV: An iterative procedure for
validation of a sentence. Preprint.
28
[14] Z. Sela. Diophantine geometry over groups V: Quantier elimination.
Preprint.
[15] Z. Sela. Diophantine geometry over groups VI: The elementary theory of
a free group. Preprint.
[16] Z. Sela. Diophantine geometry over groups VIII: The elementary theory of
a hyperbolic group. Preprint.
[17] Z. Sela. Acylindrical accessibility for groups. Invent. Math., 129(3), 1997.
[18] Z. Sela. Diophantine geometry over groups I: Makanin-Razborov diagrams.
Publ. Inst. Hautes

Etudes Sci., 93, 2001.
[19] Z. Sela. Diophantine geometry over groups II: Completions, closures and
formal solutions. Israel J. Math., 134, 2003.
29