Quasi-Isometries

Kevin Whyte
Berkeley Fall 2007

Lecture 1
Theorem 1. If G acts geometrically on X and Y (proper geodesic metric spaces)
then X and Y are quasi-isometric.

A geometric action is a group action that is cocompact, isometric, and properly

discontinuous. The basic example of such an action is when K is compact, G = π1 (K)
and X = K e where K e is the universal cover of K. The basic idea of a quasi-isometry
is that if we hit a space with a quasi-isometry then the basic shape stays the same or
if we stand far enough back the spaces look the same. One example is shown below
with X = R2 , K = T 2 , G = Z2 .

R2 1111111
0000000 Z2
0000000
1111111
Q.I.
' 0000000
1111111
0000000
1111111
0000000
1111111
0000000
1111111
Proof. Fix G, X and pick p ∈ X. Then since the action of G is cocompact there

X
g2 p

g1 p

x0 = a
xn = b
xi

1
2

exists an R such that the G translates of BR (p) are the whole space X. Now if a
and b are distinct points in X then subdivide the geodesic connecting a, b so that you
get a sequence of points {xi } such that a = x0 , x1 , . . . , xn = b, d(xi , xi+1 ) ≤ 1 and
n ≤ d(a, b) + 1.
We can find gi ∈ G such that d(gi p, xi ) ≤ R. Since the action is isometric we
know that
d(gi p, gi+1 p) ≤ 2R + 1 ⇐⇒ d(p, gi−1 gi+1 p) ≤ 2R + 1
Since the action is properly discontinuous and the space is proper we see that

S = {g ∈ G|d(p, gp) ≤ 2R + 1}

is finite. Look at Y with a base point q. Then

Supg∈S d(q, gq) = R0 < ∞

since we are taking the Sup of a finite set. Consider the path in Y that connects
(gi q, gi+1 q)

gn q
g1 q

g0 q
g2 q

and notice that each segment has length less than R0 and hence

dy (g0 q, gn q) ≤ total length ≤ R0 n ≤ R0 (d(a, b) + 1).

Definition 2 (Quasi-isometry version 1). X, Y are quasi-isometric if there exists an

f : X → Y and g : Y → X such that for some (K, C)

dy (f (a), f (b)) ≤ Kdx (a, b) + C and dx (g(a0 ), g(b0 )) ≤ Kdy (a0 , b0 ) + C

for all a, b ∈ X and a0 , b0 ∈ Y and in addition

d(g ◦ f (a), a) ≤ C ∀a ∈ X and d(f ◦ g(b), b) ≤ C ∀b ∈ Y.

Definition 3 (Quasi-isometry version 2). There exists an f : X → Y so that

d(a, b) − C
≤ d(f (a), f (b)) ≤ Kd(a, b) + C
K
and for all y ∈ Y there exists x ∈ X so that d(f (x), y) ≤ C.
 3

Cayley Graphs
Let C(G, S) be the Cayley graph of the group G with respect to the generating set
S. Then the vertex set corresponds to the elements of G and the edges are given
by (g, g 0 ) ⇐⇒ g −1 g 0 ∈ S. If a space X “looks like” C(G, S) for some choice of
generating set then X is quasi-isometric to C(G, S). This implies that for any two
finite generating sets for the group G the cooresponding Cayley graphs are quasi-
isometric.
Now think about G = π1 (S) for a closed surface S and note that the universal
cover is the same up to quasi-isometry. Thus using the previous theorem we see that
Q.I.
G ' Se = S 2 , R2 , or H2
and consequently there are only three quasi-isometry classes for π1 (S).
Question 4. What things share the same geometry as an arbitrary group G?
Some examples of quasi-isometries are
1. G is quasi-isometric to any finite index subgroup G0 ≤ G.
2. G is quasi-isometric to any quotient G/N if N if a finite and normal subgroup
of G.
Sequences of the above two examples produce an equivalence relation on groups G
called Weak Commensurability. These two examples put together with the theorem
give the following results.

1. Any 2 groups that have isomorphic finite index subgroups are quasi-isomorphic.
2. The genus 2 surface is quasi-isometric to higher genus surfaces.

Now we have committed a lie by omission since we never showed that in fact the three
spaces S 2 , R2 , H2 are not quasi-isometric.
Theorem 5 (Connect the dots). Suppose G acts geometrically on X where X is
contractible then if f : Y → X (Y a geodesic metric space) such that
d(f (y), f (y 0 )) ≤ Kd(y, y 0 ) + C
then there exists an f 0 with
d(f 0 (y), f (y)) ≤ R < ∞
for all y ∈ Y which is continuous and
d(f 0 (y), f 0 (y 0 )) ≤ K 0 d(y, y 0 ) + C 0 .
Corollary 6. If X, Y are both uniformly contractible and quasi-isometric to G then
X, Y are proper homotopy equivalent.
Corollary 7. Zn and Zm are not quasi-isometric for n 6= m.
4

Lecture 2
The general idea of the course is to develop quasi-isometry invariants and techniques
and use them to tell different groups apart. In general there are four types of quasi-
isometry invariants:
1. Topological
2. Geometric
3. Analytic
4. Asymptotic
We would like to use the connect the dots theorem to give a topological invariant,
but first we will sketch a proof.
Sketch of connect the dots in a special case. The idea is to use the existence of a finite
K(G, 1)(K is a finite complex with π1 (K) = G and K e contractible) if it exists. The
2 2 e = R2 .
special case we will look at is when G = Z , K = T , and K
(claim) If f : Z2 → Z2 is a quasi-isometry then it extends to a proper (fˆ−1 (K) =
compact for all compact K ⊂ R2 ) homotopy equivalence fˆ : R2 → R2 (fˆ is also a
quasi-isometry).

c 11
00
00
11
d f(d)
1
0
0000
1111
0
1
f(b)

00
11 b f
0000
1111
a −→
0000
1111
0000
1111
f(c)

f(a)

What the picture describes is the following operation. First you map all the
vertices over using f . Second you map all the edges connecting the vertices together
across via f . This gives you a map of S 1 → Z2 . Thirdly you use the fact that K is a
K(G, 1) and more specifically that Ke is contractible to show that this map actually
extends to a map of the whole ball into the one-skeleton. Iterate this process for
higher dimensional spheres and balls to prove the claim.

Definition 8 (Uniformly Simply-Connected). X is uniformly simply connected if

for all D there exists a D0 so that any continuous map S 1 → X with image of
diameter ≤ D bounds a disk of diameter ≤ D0 .
Definition 9 (Uniformly Contractible). X is uniformly contractible if for all R there
exists an R0 such that for all x ∈ X
BR (x) ,→ BR0 (x)
is null-homotopic.
 5

Example 10 (Contractible but not uniformly contractible). So each of the circles

in the base has the same circumference C and then we are attaching a disk but going
higher and higher up each time. This is clearly contractible, but is not uniformly
contractible since for far out you will have to go arbitrarily high up to contract a loop
of length C.

In the case where G has a finite K(G, 1) then the Cayley graph of G is uniformly
contractible and hence if G and H have finite K(π, 1)0 s and are quasi-isometric then

e ∼
Hci (K) = H i (G, ZG) ∼
= H i (H, ZH) ∼
= Hci (L)
e

where K, L are the Cayley graphs of G, H respectively. The isomorphisms are a

consequence of the previous argument since K,
e Le are properly homotopy equivalent.

Example 11. 
0 i 6= n
Hci (Rn ) =
Z i=n
In particular this says that Zm , Zn are not quasi-isometric if m 6= n. Also note that

Hc1 (T ree) = inf initely generated.

The previous discussion shows that the homotopy type of the ends is a quasi-
isometric invariant. In the case of a tree this shows that the natural boundary is a
Cantor set.

Exercise 12. G is finitely presented ⇐⇒ G is uniformly simply connected (there

exists a uniformly simply connected space on which G acts cocompactly and isometri-
cally). This will show that being finitely presented is a quasi-isometric invariant.

Example 13. [Bad Example of a quasi-isometry from R2 to R2 .] If θ(r) = ln(1 + r)

then f is a quasi-isometry, however the red geodesic ray goes to a logarithmic spiral
and hence f has a winding number of ∞ as a map from the disk to the disk.

An example of a property that is not a quasi-isometry invariant would be possesing

a finite K(G, 1) (K(Z2 , 1) = RP ∞ but K(trivial, 1) = point). It would be interesting
if there was a torsion free counter-example. Having a finite k-skeleton for any k on
the other hand is a quasi-isometric invariant.
6

R2 R2

θ(r)

Question 14. Is R2 quasi-isometric to H2 ? Or equivalently is Z2 quasi-isometric to

π1 (S) where S is a surface of genus greater than two?

These two spaces are homeomorphic so a topological invariant will not give us a
negative answer. A geometric invariant namely growth will give us a negative answer.
In R2 a ball has area that is quadratic as a function of the radius R. In H2 points are

R2 H2

111
000 1010 11
0
000
111 10
R

01
R KR+C

widely separated (IE covered by balls of radius say 1) thus the area behaves like eR .
Let G be a group and 1 ∈ S ⊂ G a finite generating set that is closed under inverses
(makes things easier but not essential). Define

N(G,S) (k) = |S k | = {g ∈ G|g = s1 s2 . . . sj f or si ∈ S and j ≤ k}

which is a count of the vertices in the ball of radius k in the Cayley graph.

Example 15. Let G = Z2 with

       
1 0 −1 0
S= , , ,
0 1 0 −1

then N(G,S) (k) is roughly k 2 .

 7

If f : (G, S) → (H, T ) is a (K,C) quasi-isometry with f (id) = id then

f (Bk (G, S)) ⊂ BKk+C (H, T )

f −1 (Bk (H, T )) ⊂ BKk+C (G, S)

for all k. Now f is not quite injective and not quite surjective thus we have both over
counted and undercounted respectively. To make this precise for some A, B and for
all k we have
N(G,S) (k) ≤ AN(H,T ) (Kk + C) + B

N(H,T ) ( k−c
K
)−B
N(G,S) (k) ≥ .
A
This implies that N(G,S) ≈ A(k) (A(k) is a polynomial of degree d) then so is N(H,t)
which gives us a quasi-isometric invariant. This gives another proof that
QI
Zn  Zm

for n 6= m. We can also see that

QI
n
Z  π1 (Sg )

for g ≥ 2 since π1 (Sg ) is hyperbolic and hence grows exponentially. It is important to

note that these geometric methods can not distinguish H2 and H12 while topological
invariants can. And all the methods that we have developed so far are no help in
distinguishing H3 and H2 × R. The hope is that some blending of these arguments
will work, for example L2 -cohomology which is a quasi-isometric invariant.

Lecture 3
We start with another proof that Zm ∼
6= Z n if m 6= n.

Proof. Suppose f : Zn → Zm was a (K, C) quasi-isometry.

Z2

Z
f
Distance 1

Distance 1
8

Zn ⊂ Rn
”f ”
Zm ⊂ Rm

For any  > 0 let f : Zn → Zm (Z n is Zn with the normal metric rescaled by
a factor of ) and notice f = f .

dZn (f (x), f (y)) = dZm (f (x), f (y)) ≤ (KdZn (x, y) + C) = KdnZ (x, y) + C

thus f is a (K, C) quasi-isometry from Zn → Zm . Taking a limit as  → 0 we

see that Zn , Zm are sitting inside Rn , Rm more densly. Define fˆ : Rn → Rm
via the following: for x ∈ Rn take any sequence xi ∈ i Zn as i → 0 such that
xi → x and let fˆ = lim fi (xi ). This will be a (K, 0) quasi-isometry which implies
that fˆ is K-bilipschitz and thus well-defined. Notice fˆ : Rn → Rm is a K-bilipschitz
homeomorphism. But fˆ is K-bilipschitz and is almost everywhere differentiable via
Rademacher’s Theorem with derivative a linear isomorphism of norm less than or
equal to K which gives a contradiction.
How does this construction generalize for more general groups Γ? We would like
to take a limit of the metric spaces Γ as  → 0. IE take limits of sequences {γi ∈ Γi }
as i → 0. In general this construction yields something called an (could be lots of
different ones) asymptotic cone of Γ. For example, the asymptotic cone of anything
negatively curved is an R-tree. A nice property that we might hope for is the space
being locally compact (any ball (of radius 1) in the limit can be covered by finitely
many (Nδ ) -balls for any δ > 0). Notice that
1
1. a ball of radius 1 in Γ equals a ball of radius 
= R in Γ
δ
2. a ball of radius δ in Γ equals a ball of radius 
= δR in Γ.
Another way to describe Nδ is the number of ball’s needed to cover a bigger ball.
Theorem 16 (Almost Claim). If balls of radius one in the limit space can be covered
by Nδ balls of radius δ, then in Γ, every R-ball can be covered by approximately Nδ
balls of radius δR.
This implies that
V ol(BR )
≤ Nδ
V ol(BδR )
thus R → V ol(BR ) is bounded by a polynomial of degree ln(N lnδ
δ)
. In addition the
degree and the fact that this is a polynomial are quasi-isometric invariants. It is
important to note that the converse is true if the bound is a polynomial and thus
taking nice limits requires polynomial growth.
Theorem 17 (Gromov). Γ has polynomial growth if and only if Γ is virtually nilpo-
tent.
 9

A virtually nilpotent group is virtually a subgroup of n × n upper triangular

matrices with all ones on the diagonal for some n. IE
  
 1 x y 
 0 1 z  | x, y, z ∈ Z
0 0 1
 

Very fast outline. 1. True for linear groups (IE matrices)

2. Build X = lim→0 Γ

3. Show that X is homogeneous (if you can do 2 then you can do 3)

4. Isom(X) is a Lie Group (technically very hard)

5. Γ → Isom(X) is big enough to induct.

Exercise 18. γ ∈ Γ if d(γx, x) ≤ R for all x ∈ Γ then γ commutes with a finite
index subgroup Γ0 ⊂ Γ.
In particular if this is true for all γ ∈ Γ then Γ has a finite index abelian subgroup
which takes care of the base case for the induction.

Example 19. Let   

 1 x z 
Γ=  0 1 y |x, y, z ∈ Z

0 0 1
 

where the multiplication is

(x, y, z) · (x0 , y 0 , z 0 ) = (x + x0 , y + y 0 , z + z 0 + xy 0 ).

Note that the first two terms are abelian and the third term is abelian modulo some
error term not depending on either z or z 0 . Let
     
1 1 0 1 0 0 1 0 1
X =  0 1 0  , Y =  0 1 1  and Z =  0 1 0  .
0 0 1 0 0 1 0 0 1

XY = Y XZ and Z commutes with everybody. This group is not Z3 for two reasons.
The first is that it is non-abelian and the second is that it is generated by two elements
namely X, Y (Z −1 = [X, Y ]) not three. Let
  
 1 x y  homeo
G=  0 1 z  | x, y, z ∈ R ' R3
0 0 1
 

then G/Γ is a compact 3-manifold with universal cover R3 which has the Nil geometry
of Thurston.
10

(3, 2)
X 2 Y XY

Y 3X 3

(0, 0)

The Cayley graph for Γ maps to Z2 by writing any word in X and Y and counting
the number with signs of X’s and Y 0 s that appear. In the diagram both of the words
trace out paths from the origin to (3, 2) but the distance between the actual endpoints
of the words depends on the number of Z’s that you pick up when moving from one
to another. This number of Z 0 s turns out to be exactly the area of the polygon traced
out by the paths in the projection to Z2 which in this case is 5. It is most efficient to
do all the Y ’s first then all the X’s.
2 2
X n Y n = Y n X n Z n ⇒ Z n = X −n Y −n X n Y n
2
this shows that the path from 0 to Z n has length n2 when writing it as a product of
Z’s but the actual word length is at most 4n.
√ This shows that when you are traveling
vertically the word length is growing like n and is not a geodesic. Thus Γ has
polynomial growth like n4 which implies that it is not quasi-isometric to Z3 . It is also
not quasi-isometric to Z4 .

In fact we are quite ignorant about nilpotent groups.

Question 20. Are there known quasi-isometric nilpotent groups that are not com-
mensurable?

Answer 21. Yes. Cocompact lattices in the same nilpotent lie group.

The problem is that individual quasi-isometries can be awful and we don’t know
enough invariants or how to go the other direction. In fact the tools that we have
developed can’t really do more than what we have already shown.

Lecture 4
One question that is both natural and interesting is

Question 22. Given a group G, which groups H are quasi-isometric to G?

Unfortunately, the techniques we have developed can not help so we need some-
thing new. Suppose that G acts geometrically on a nice space X what works in a lot
of cases is to get H to act on X or something similar. Two theorems of this type are:
 11

Theorem 23. If G is quasi-isometric to Hn then G acts geometrically on Hn .

Theorem 24. If G is quasi-isometric to F2 (the free group on two letters) then G
acts geometrically on some (locally finite) tree. Note: acting geometrically on a tree
implies that G is virtually free.
In order to approach the proofs additional information is needed.
Definition 25 (Quasi-Action). A (K,C) quasi-action of G on X is

1. ∀g ∈ G associate a (K,C) quasi-isometry φg : X → X.

2. ∀g, h ∈ G supx d(φg ◦ φh (x)) ≤ C

Definition 26. Define QI(X) = {f : X → X a quasi − isometry}/f ∼ g where

f ∼ g if and only if supx d(f (x), g(x)) < ∞. It can be checked that QI(X) is a group.
A quasi-action of G on X gives a uniform (image of φg all have quasi-isometry
constants (K,C)) homomorphism G → QI(X), but the converse is not true whatso-
ever.
Example 27. Z quasi-acts on R by integer translation, but the induced homomor-
phism Z → QI(R) is trivial. It is completely mysterious what QI(R) really is, but
some subgroups of QI(R) are known. The first is R× via the map x 7→ λx for λ 6= 0.
The second is the space of lipschitz diffeomorphisms of R. However there are bad
things that happen in the topology. Let

 2x |x| ≤ n
fn : R → R given by fn (x) = x+n x>n
x − n x < −n


fn ∼ Id for every n but limn→∞ fn = 2x 6∼ Id which implies that a constant sequence

slope = 1

−n n
slope = 2

slope = 1

of identity maps converge to a non-identity element of the group.

12

There are spaces X(even some nice ones) where Isom(X) → QI(X) is an isomor-
phism (IE higher rank symmetric spaces, buildings, . . . ). There are even some groups
G where QI(G) = G which is about as strong a rigidity result for quasi-isometries
that is possible. It is suspected that most groups are of this form. One example
would be [?].

Claim 28. If G acts geometrically on X and H is quasi-isometric to G then H

quasi-acts on X.
QI QI QI
Proof. Since G ∼ = X and H ∼= G we see that H ∼ = X and call this quasi-isometry f .
H acts on itself isometrically by left multiplication so define the quasi-action to be
φH (x) = f (h · f −1 (x)).

We could have tried a more direct approach: let T : H → G be the given quasi-
isometry and define φh (x) = h · x = (T h) · x. The problem is that T (h1 h2 ) and
T (h1 )T (h2 ) as group elements in G might have nothing to do with one another which
gives no control over the induced action of H on X. This shows that quasi-isometries
do not behave like group homomorphisms.

Proof. (Outline for 23 and 24)

1. Get a quasi-action of H on X. (28)

2. Improve the quasi-action to an isometric action (techniques generally depend

on X, for example different techniques will be utilized in 23 and 24).

For Hn , quasi-isometries of Hn “=” quasi-conformal (mob, symmetric) homeo-

morphisms of S n−1 .

Theorem 29. (Tukia for n > 3 and Sullivan for n = 3) A uniform group of quasi-
conformal homeomorphisms of S n−1 is quasi-conformal conjugate to a group of con-
formal homeomorphisms.

This is a hard theorem, but it completes step two in the proof of 23. Unfortunately
the boundary of the space will not help once the space X no longer has negative
curvature. For example, the quasi-isometry in 13 sends geodesic rays in R2 to a
logarithmic spiral which does not converge to anything nice on the boundary.

Theorem 30. If G is quasi-isometric to Z then G acts properly discontinuously and

cocompactly by isometries on R.

Proof. (sketch 1) If f : R → R is a quasi-isometry the f either fixes or flips the ends

{±∞}. Let an = f (n) where n ∈ Z since f is a quasi-isometry

|an+1 − an | = K + C.
 13

If |n − m| > K + C then |f (n) − f (m)| > 1. Take N > K + C such that aN 6= a0

and without loss of generality assume that aN > a0 . For all n ≥ N from the previous
fact |f (n) − f (0)| > 1 and thus all the an must be on the same side of a0 . Thus
a2N → ∞ if N is large enough. If instead aN < a0 then the sequence converges to
−∞ instead.

Claim 31. If f : R → R is a quasi-isometry then there exists D > 0 such that one
of the following holds

1. (Coarsely Increasing) For all x, y with x ≥ y + D, f (x) ≥ f (y) + 1

2. (Coarsely Decreasing) For all x, y with x ≤ y − D, f (x) ≥ f (y) + 1

Corollary 32. If G quasi-acts on R preserving ends then there exists x ∈ R such

that d(fg (x), x) ≥ D which implies g has infinite order in G. This implies that if
f (x) > x then f 2 (x) > f (x) and the process can be iterated.

The point is the following. Suppose fg (x)  x then fg (fg (x))  fg (x) and
fg (fg (x)) is close to fg2 (x). Thus fg2 (x)  fg (x)  x and by induction fgn (x) 
fgn−1 (x) . In particular, the “orbits” of {g n } are unbounded so:
QI
G∼
=R ⇒ G quasi-acts on R.
⇒ G0 ⊂ G of index at most 2 quasi-acts preserving orientation.
⇒ There exists g ∈ G of infinite order
(uses the fact that quasi-action is coarsely cocompact).
⇒ f |{<g>} → R is a quasi-isometry.
⇒ < g >⊂ G is finite index (coarsely properly discontinuous).
To summarize if G is quasi-isometric to Z then there exists Z ⊂ G of finite index.

Lecture 5

$Id$ G
g2

QI

Z x0
−∞ ∞

Last time we saw that if a group G was quasi-isometric to Z then there exists
g ∈ G of infinite order such that · · · < g n−1 x0 < g n x0 < g n+1 x0 < . . . . This gives a
14

quasi-action of G on R that is uniformly proper (∀R > 0 there exists N so that ∀x ∈ X

|{g|d(gx, x) < R}| < N ) and cobounded (∃R > 0 so that ∀x and x0 ∈ X ∃g ∈ G
with d(gx, x0 ) < R) which is the quasi-action analogue of proper and cocompact. An
example of a space and group that have these properties is given by the action of G
on a Cayley graph.
Exercise 33. Check that the above properties are invariant under quasi-isometries. In
fact, anything defined in the coarse uniform language will be quasi-isometric invariant.
Exercise 34. Suppose G quasi-acts properly and coboundedly on X and G0 ⊂ G also
acts coboundedly then G0 has finite index in G.
In the following picture we see that x0 disconnects Z which gives a natural order
on Z. In the case of a general group G quasi-isomorphic to Z a point might not

Z x0
−∞ ∞

disconnect the space.

Example 35. Take G = Z with the generating set {±1, ±2} which has the following
Cayley graph. In this example points do not disconnect the space.
−2 0 2 4

−1 1 3 5

This shows that the property of points disconnecting the space is not a quasi-
isometric invariant. What can be said is that sets of some bounded size disconnect
the space and thus talking about ordering of points that are far enough apart so that
they were not removed in the disconnecting operation can be done. Z is special as a
large scale geometry since it has two ends. We can rephrase this as
Theorem 36. Any group with two ends has a cyclic subgroup of finite index.
Proof. (Idea) Since it is two ended take a subset of the Cayley graph that separates
it. If the translates of the chosen subset are disjoint then make each of the translates
an edge and collapse everything in between to a vertex to get an action of G on R.
If the translates do intersect then the procedure described above does not yield an
action on R. To complete the proof we need to modify Γ into a new separating set
Γ0 in such a way that the translates of Γ0 are disjoint.
 15

g −1 Γ Γ gΓ gΓ Γ

An extension of this technique can be used to prove the following.

Theorem 37. If G is quasi-isometric to a finitely generated free group then G acts

properly discontinuously and cocompactly on some (locally finite) tree.

This is the same theorem as 24 since all finitely generated non-abelian free groups
are quasi-isometric to F2 (finite index subgroups). One way to see the quasi-isometry
to F3 is by collapsing one edge at every vertex in the Cayley graph of F2 . The map π

C(F2 )
C(F3 )

is a quasi-isometry in which we might have collapsed at most half the edges rounded
up on any path thus

d(x, y) − 1
≤ d(π(x), π(y)) ≤ d(x, y).
2
At this point thinking that F2 acts properly discontinuously and cocompactly on
C(F3 ) might seem reasonable, but this is not the case. In general all groups quasi-
isometric to a finitely generated free group act on some tree but we do not pick the
tree, which is unlike when hyperbolic space was picked in the proof of 23. The trees
that we are talking about have infinitely many ends (certainly true for bounded sets)
since larger and larger bounded sets separate more and more.
Proof. We would hope to build a dual tree like we did before. We still have the
problem of overlapping translates of the disconnecting sets and that the vertices that
we would like to define (circles) would all be adjacent in this picture giving a loop
16

and the resulting graph would not be a tree. Suppose G is quasi-isometric to F2 ,

let X be a Cayley graph for G, then there exists D so that π1 (X) is generated by
loops of diameter≤ D. Equivalently we can attach 2-cells of diam ≤ D to X to get
something simply connected. In the picture we are taking steps of bounded size and

maybe circles look close edge crossed by path

eventually the path closes. One operation that we can do is to write the loop as a
union of smaller loops until the quasi-isometry constants take care of things. Why
do we care about the separating set being simply connected? The answer is that we
want to understand the sets that are doing the disconnecting, and hence we want
them to be as simple as possible (why we require that the disconnecting set be simply
connected). An example of a disconnecting set that we want to rule out is shown
below. We have K, a G two complex which is quasi-isometric to our tree T (glued in

discs of bounded size), we may assume the map

K→T

is continuous. Look at the preimage of a point and notice that with Z we could cut
 17

gΓ Γ

an edge in half to make a simpler disconnecting set. The same operation can be done
in a tree. To do this go back to the connect the dots argument.

First map in points and then pick a geodesic to map to.

00
11
0
1
0
1
0
1
What you get is a triangle mapping onto a tripod. Then crush the triangle connecting
the preimages of the center of the tripod to a point and map obviously with the
boundary mapping in geodesically. Point inverses look like the picture below.

We call the disconnecting dual graphs tracks, and all we know about these tracks
is that they are connected. We still have the problem that the tracks might not be
disjoint. So find a graph that separates that are as short as possible. With luck we can
do a cut and paste argument and then the minimal graphs will not intersect. Thus
we get an action on some tree and then repeat the argument so that any unbounded
chunks become manageable.
18 1 LECTURE 6

1 Lecture 6
The following is a summary of what happened last time. What can we know about
some group G that is quasi-isometric to Fk a non-abelian free group. To this end we
did the following.

1. Build K a simply connected 2-complex with properly discontinuous and cocom-

pact G action.

2. Find compact subsets C of K such that K/C has more than one unbounded
component.

3. Use some definition to find “minimal” such C.

4. Hope that by minimality translates are disjoint. The picture to imagine is the
following.

5. Conclude that G acts on the dual tree with finite edge stabilizers.

Question 38. Why is the dual graph a tree?

Answer 39. Suppose not. Then there is some non-trivial loop in the dual graph and
thus removing one of the translates of C does not disconnect the space which is a
contradiction.

Question 40. Is the tree that you obtain necessarily locally finite?

Answer 41. No.

Theorem 42 (Stalling’s Ends Theorem). Suppose that G is finitely presented and G

has more than one end. Then G has a non-trivial splitting G = A ∗C B or G = A∗C
with C finite.

Proof. If G is quasi-isometric to a free group then outline gives a variant of Dun-

woody’s proof of Stalling’s Theorem where the finitely presented assumption is nec-
essary to build the 2-complex and the assumption that G has more than one end
enables us to find minimal compact sets that disconnect.

Question 43. What is the splitting that this theorem gives for Z?

Z= 1 1
 19

Exercise 44. ?
Why does the splitting process given by Stalling’s Theorem terminate? The an-
swer is that the group G is finitely presented. More precisely
1. If G is torsion free then G = A ∗ B
2. One can try to inductively split A and B until we get a graph of groups de-
composition for G with ≤ 1 − ended vertex groups and finite edge groups (if
G is torsion free this follows from Grushko’s Theorem). But we need finitely
presented to get this process to terminate in general.
In fact there are known examples where this process does not terminate. Dunwoody
has an example of a finitely generated but not finitely presented group where the
process does not stop.
What do we know so far?
1. If G is Gromov hyperbolic then we have ∂G which is quasi-isometry invariant.
2. If G splits you may be able to reduce your question to the pieces. IE analyze
your mystery group in terms of quasi-isometry invariants of the pieces.
3. If G has polynomial growth (almost flat) then Gromov says that G is virtually
nilpotent which reduces your question to nilpotent groups which is a very open
area.
4. What if G has non-positive curvature? Unfortunately there is not a large scale
(ie coarse) notion of non-positive curvature.
What about the specific case of three-manifold geometries? We can cut along tori to
get an analogue of the graph of groups decompositions where the pieces have one of
the following geometries: S 3 , S 2 × S 1 , R3 , H3 , H2 × R, Nil, Sol, or SL
^ 2 R. There are
certain geometries that we can “handle” at this point (handle means know all groups
quasi-isometric to this geometry) namely S 3 (QI to trivial group), S 2 × S 1 (compact),
R3 , H3 , H2 × R (Kleiner spoke about this one), Nil. We are not going to worry about
Sol. SL
^ 2 R turns out to be annoying but we will say some words about it.
An example of a manifold with SL 2 R geometry is M = {unit tangent bundle of a
^
closed surface Σ of genus ≥ 2 }. In this case M is a non-trivial circle bundle over Σ
1 → Z → π1 (M ) → π1 (Σ) → 1
which gives a non-trivial central extension. This almost looks like π1 (M ) = π1 (Σ) × Z
2
and so you might think that SL 2 R is quasi-isometric to H × R, but SL2 R is not non-
^ ^
positively curved. There is a result in this vein which is the following exercise.
Exercise 45. A central extension
1→Z→G→H→1
is quasi-isometric to H × Z if there is a lipschitz section H → G. This is not quite
an if and only if statement but it is very close.
20 1 LECTURE 6

If X is CAT (0) one approach would be to try to control the quasi-flats of maximal
rank (quasi-isometric embeddings of Rn where n is as large as possible) in order to
determine the question of what spaces X is quasi-isometric to.

Example 46. Examine H2 × H2 . Here the maximal flats are isometric copies of R2
given by products of the form l1 × l2 where l1 and l2 are geodesics in different factors
of H2 × H2 . A natural set of questions is where and how do two flats come close to
each other. In each factor they can get close in three distinct ways namely they can
intersect in a point (different endpoints on the boundary), can become close along a
ray (can share one endpoint on the boundary), or they are the same geodesic (sharing
2 points on the boundary). This gives essentially 5 different behaviors:
111
000
000
111 111
000
000
111
000
111
000
111 000
111
000
111 .

The quarter plane is the basic building block of a two-dimensional flat. Build a graph
to record the incidence data. Let the vertices be the vertical and horizontal rays up
to bounded distance which equals ∂H2 × ∂H2 . Connect the vertices by an edge if they
bound a quadrant (this is the analogue to the visual boundary of hyperbolic space).
This give a complete bipartite graph which implies that a quasi-isometry of this space
is a quasi-isometry of each of the factors because of the fact that switching sides is
not allowed.

Example 47. Now let us look at SL3 R/SO(3). Here the graph defined in the same

111
000
way is again bipartite with the basic unit being a sixth of a plane

000
111
000
111
000
111
.

The graph at “∞” is the incidence graph of points and lines in the real projective
plane. It turns out that automorphisms of this graph are (essentially) SL3 R. This
relationship is analogous to the relationship of the curve complex to the mapping class
group.

If you can control the flats then you have a handle on what is going on. Unfortu-
nately, there are “bad” quasi-flats even in H2 × H2 . One example is that you can get
a loop of length six in the bipartite graph. Thus you can glue six quadrants together
to form a quasi-flat. A general way to build quasi-flats is to build them by stringing
together a finite number of flats.