Hilberts Fifth Problem and Related

Topics

Terence Tao

This is a preliminary version of the book Hilberts Fifth Problem and Related Topics published by the
American Mathematical Society (AMS). This preliminary version is made available with the permission of
the AMS and may not be changed, edited, or reposted at any other website without explicit written
permission from the author and the AMS.

Hilberts fifth problem and related topics

Terence Tao
Department of Mathematics, UCLA, Los Angeles, CA 90095
E-mail address: [email protected]

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

To Garth Gaudry, who set me on the road;

To my family, for their constant support;
And to the readers of my blog, for their feedback and contributions.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Contents

Preface

ix

Notation

Acknowledgments

xi

Chapter 1.

Hilberts fifth problem

1.1.

Introduction

1.2.

Lie groups, Lie algebras, and the Baker-Campbell-Hausdorff

formula

22

1.3.

Building Lie structure from representations and metrics

48

1.4.

Haar measure, the Peter-Weyl theorem, and compact or

abelian groups

67

1.5.

Building metrics on groups, and the Gleason-Yamabe theorem 90

1.6.

The structure of locally compact groups

114

1.7.

Ultraproducts as a bridge between hard analysis and soft

analysis

126

1.8.

Models of ultra approximate groups

152

1.9.

The microscopic structure of approximate groups

181

1.10.

Applications of the structural theory of approximate groups 201

Chapter 2.

Related articles

213

2.1.

The Jordan-Schur theorem

214

2.2.

Nilpotent groups and nilprogressions

217

2.3.

Ados theorem

228

2.4.

Associativity of the Baker-Campbell-Hausdorff-Dynkin law

237
vii

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

viii

Contents

2.5.

Local groups

243

2.6.

Central extensions of Lie groups, and cocycle averaging

254

2.7.

The Hilbert-Smith conjecture

265

2.8.

The Peter-Weyl theorem and nonabelian Fourier analysis

271

2.9.

Polynomial bounds via nonstandard analysis

280

2.10.

Loeb measure and the triangle removal lemma

283

2.11.

Two notes on Lie groups

294

Bibliography

299

Index

305

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Preface

Hilberts fifth problem, from his famous list of twenty-three problems in

mathematics from 1900, asks for a topological description of Lie groups,
without any direct reference to smooth structure. As with many of Hilberts
problems, this question can be formalised in a number of ways, but one commonly accepted formulation asks whether any locally Euclidean topological
group is necessarily a Lie group. This question was answered affirmatively
by Montgomery-Zippin [MoZi1952] and Gleason [Gl1952]; see Theorem
1.1.13. As a byproduct of the machinery developed to solve this problem, the
structure of locally compact groups was greatly clarified, leading in particular to the very useful Gleason-Yamabe theorem (Theorem 1.1.17) describing
such groups. This theorem (and related results) have since had a number of
applications, most strikingly in Gromovs celebrated theorem [Gr1981] on
groups of polynomial growth (Theorem 1.1.30), and in the classification of
finite approximate groups (Theorem 1.1.29). These results in turn have applications to the geometry of manifolds, and on related topics in geometric
group theory.
In the fall of 2011, I taught a graduate topics course covering these topics, developing the machinery needed to solve Hilberts fifth problem, and
then using it to classify approximate groups and then finally to develop applications such as Gromovs theorem. Along the way, one needs to develop
a number of standard mathematical tools, such as the Baker-CampbellHausdorff formula relating the group law of a Lie group to the associated
Lie algebra, the Peter-Weyl theorem concerning the representation-theoretic
structure of a compact group, or the basic facts about ultrafilters and ultraproducts that underlie nonstandard analysis.

ix

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Preface

This text is based on the lecture notes from that course, as well as from
some additional posts on my blog at terrytao.wordpress.com on further
topics related to Hilberts fifth problem. The first chapter of this text can
thus serve as the basis for a one-quarter or one-semester advanced graduate course, depending on how much of the optional material one wishes to
cover. The material here assumes familiarity with basic graduate real analysis (such as measure theory and point set topology), as covered for instance
in my texts [Ta2011], [Ta2010], and including topics such as the Riesz representation theorem, the Arzela-Ascoli theorem, Tychonoffs theorem, and
Urysohns lemma. A basic understanding of linear algebra (including, for
instance, the spectral theorem for unitary matrices) is also assumed.
The core of the text is Chapter 1. The first part of this chapter is devoted
to the theory surrounding Hilberts fifth problem, and in particular in fleshing out the long road from locally compact groups to Lie groups. First, the
theory of Lie groups and Lie algebras is reviewed, and it is shown that a Lie
group structure can be built from a special type of metric known as a Gleason metric, thanks to tools such as the Baker-Campbell-Hausdorff formula.
Some representation theory (and in particular, the Peter-Weyl theorem) is
introduced next, in order to classify compact groups. The two tools are
then combined to prove the fundamental Gleason-Yamabe theorem, which
among other things leads to a positive solution to Hilberts fifth problem.
After this, the focus turns from the soft analysis of locally compact
groups to the hard analysis of approximate groups, with the useful tool of
ultraproducts serving as the key bridge between the two topics. By using this
bridge, one can start imposing approximate Lie structure on approximate
groups, which ultimately leads to a satisfactory classification of approximate
groups as well. Finally, Chapter 1 ends with applications of this classification
to geometric group theory and the geometry of manifolds, and in particular
in reproving Gromovs theorem on groups of polynomial growth.
Chapter 2 contains a variety of additional material that is related to one
or more of the topics covered in Chapter 1, but which can be omitted for
the purposes of teaching a graduate course on the subject.

Notation
For reasons of space, we will not be able to define every single mathematical
term that we use in this book. If a term is italicised for reasons other than
emphasis or for definition, then it denotes a standard mathematical object,
result, or concept, which can be easily looked up in any number of references.
(In the blog version of the book, many of these terms were linked to their
Wikipedia pages, or other on-line reference pages.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Acknowledgments

xi

Given a subset E of a space X, the indicator function 1E : X R is

defined by setting 1E (x) equal to 1 for x E and equal to 0 for x 6 E.
The cardinality of a finite set E will be denoted |E|. We will use the
asymptotic notation X = O(Y ), X  Y , or Y  X to denote the estimate
|X| CY for some absolute constant C > 0. In some cases we will need
this constant C to depend on a parameter (e.g. d), in which case we shall
indicate this dependence by subscripts, e.g. X = Od (Y ) or X d Y . We
also sometimes use X Y as a synonym for X  Y  X. (Note though
that once we deploy the machinery of nonstandard analysis in Section 1.7,
we will use a closely related, but slightly different, asymptotic notation.)

Acknowledgments
I am greatly indebted to my students of the course on which this text was
based, as well as many further commenters on my blog, including Marius
Buliga, Tony Carbery, Nick Cook, Alin Galatan, Pierre de la Harpe, Ben
Hayes, Richard Hevener, Vitali Kapovitch, E. Mehmet Kiral, Allen Knutson,
Mateusz Kwasnicki, Fred Lunnon, Peter McNamara, William Meyerson,
Joel Moreira, John Pardon, Ravi Raghunathan, David Roberts, David Ross,
Olof Sisask, David Speyer, Benjamin Steinberg, Neil Strickland, Lou van den
Dries, Joshua Zelinsky, Pavel Zorin, and several anonymous commenters.
These comments can be viewed online at
terrytao.wordpress.com/category/teaching/254a-hilberts-fifth-problem/
The author was supported by a grant from the MacArthur Foundation,
by NSF grant DMS-0649473, and by the NSF Waterman award. Last, but
not least, I thank Emmanuel Breuillard and Ben Green for introducing me
to the beautiful interplay between geometric group theory, additive combinatorics, and topological group theory that arises in this text.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Chapter 1

Hilberts fifth problem

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1. Hilberts fifth problem

1.1. Introduction
This text focuses on three related topics:
Hilberts fifth problem on the topological description of Lie groups,
as well as the closely related (local) classification of locally compact
groups (the Gleason-Yamabe theorem, see Theorem 1.1.17);
Approximate groups in nonabelian groups, and their classification
[Hr2012], [BrGrTa2011] via the Gleason-Yamabe theorem; and
Gromovs theorem [Gr1981] on groups of polynomial growth, as
proven via the classification of approximate groups (as well as some
consequences to fundamental groups of Riemannian manifolds).
These three families of results exemplify two broad principles (part of
what I like to call the the dichotomy between structure and randomness
[Ta2008]):
(Rigidity) If a group-like object exhibits a weak amount of regularity, then it (or a large portion thereof) often automatically exhibits
a strong amount of regularity as well.
(Structure) Furthermore, this strong regularity manifests itself either as Lie type structure (in continuous settings) or nilpotent type
structure (in discrete settings). (In some cases, nilpotent should
be replaced by sister properties such as abelian, solvable, or
polycyclic.)
Let us illustrate these two principles with two simple examples, one in
the continuous setting and one in the discrete setting. We begin with a
continuous example. Given an n n complex matrix A Mn (C), define the
matrix exponential exp(A) of A by the formula
exp(A) :=

X
Ak
k=0

k!

=1+A+

1 2
1
A + A3 + . . .
2!
3!

which can easily be verified to be an absolutely convergent series.

Exercise 1.1.1. Show that the map A 7 exp(A) is a real analytic (and
even complex analytic) map from Mn (C) to Mn (C), and obeys the restricted
homomorphism property
(1.1)

exp(sA) exp(tA) = exp((s + t)A)

for all A Mn (C) and s, t C.

Proposition 1.1.1 (Rigidity and structure of matrix homomorphisms). Let
n be a natural number. Let GLn (C) be the group of invertible n n complex
matrices. Let : R GLn (C) be a map obeying two properties:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

(1) (Group-like object) is a homomorphism, thus (s)(t) = (s+t)

for all s, t R.
(2) (Weak regularity) The map t 7 (t) is continuous.
Then:
(i) (Strong regularity) The map t 7 (t) is smooth (i.e. infinitely
differentiable). In fact it is even real analytic.
(ii) (Lie-type structure) There exists a (unique) complex n n matrix
A such that (t) = exp(tA) for all t R.
Proof. Let be as above. Let > 0 be a small number (depending only
on n). By the homomorphism property, (0) = 1 (where we use 1 here
to denote the identity element of GLn (C)), and so by continuity we may
find a small t0 > 0 such that (t) = 1 + O() for all t [t0 , t0 ] (we use
some arbitrary norm here on the space of n n matrices, and allow implied
constants in the O() notation to depend on n).
The map A 7 exp(A) is real analytic and (by the inverse function
theorem) is a diffeomorphism near 0. Thus, by the inverse function theorem,
we can (if is small enough) find a matrix B of size B = O() such that
(t0 ) = exp(B). By the homomorphism property and (1.1), we thus have
(t0 /2)2 = (t0 ) = exp(B) = exp(B/2)2 .
On the other hand, by another application of the inverse function theorem
we see that the squaring map A 7 A2 is a diffeomorphism near 1 in GLn (C),
and thus (if is small enough)
(t0 /2) = exp(B/2).
We may iterate this argument (for a fixed, but small, value of ) and conclude
that
(t0 /2k ) = exp(B/2k )
for all k = 0, 1, 2, . . . . By the homomorphism property and (1.1) we thus
have
(qt0 ) = exp(qB)
whenever q is a dyadic rational, i.e. a rational of the form a/2k for some
integer a and natural number k. By continuity we thus have
(st0 ) = exp(sB)
for all real s. Setting A := B/t0 we conclude that
(t) = exp(tA)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1. Hilberts fifth problem

for all real t, which gives existence of the representation and also real analyticity and smoothness. Finally, uniqueness of the representation (t) =
exp(tA) follows from the identity
d
A=
exp(tA)|t=0 .
dt

Exercise 1.1.2. Generalise Proposition 1.1.1 by replacing the hypothesis
that is continuous with the hypothesis that is Lebesgue measurable.
(Hint: use the Steinhaus theorem, see e.g. [Ta2011, Exercise 1.6.8].) Show
that the proposition fails (assuming the axiom of choice) if this hypothesis
is omitted entirely.
Note how one needs both the group-like structure and the weak regularity in combination in order to ensure the strong regularity; neither is
sufficient on its own. We will see variants of the above basic argument
throughout the course. Here, the task of obtaining smooth (or real analytic structure) was relatively easy, because we could borrow the smooth (or
real analytic) structure of the domain R and range Mn (C); but, somewhat
remarkably, we shall see that one can still build such smooth or analytic
structures even when none of the original objects have any such structure
to begin with.
Now we turn to a second illustration of the above principles, namely
Jordans theorem [Jo1878], which uses a discreteness hypothesis to upgrade
Lie type structure to nilpotent (and in this case, abelian) structure. We
shall formulate Jordans theorem in a slightly stilted fashion in order to
emphasise the adherence to the above-mentioned principles.
Theorem 1.1.2 (Jordans theorem). Let G be an object with the following
properties:
(1) (Group-like object) G is a group.
(2) (Discreteness) G is finite.
(3) (Lie-type structure) G is a subgroup of Un (C) (the group of unitary
n n matrices) for some n.
Then there is a subgroup G0 of G such that
(i) (G0 is close to G) The index |G/G0 | of G0 in G is On (1) (i.e.
bounded by Cn for some quantity Cn depending only on n).
(ii) (Nilpotent-type structure) G0 is abelian.
A key observation in the proof of Jordans theorem is that if two unitary
elements g, h Un (C) are close to the identity, then their commutator
[g, h] = g 1 h1 gh is even closer to the identity (in, say, the operator norm

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

kkop ). Indeed, since multiplication on the left or right by unitary elements

does not affect the operator norm, we have
k[g, h] 1kop = kgh hgkop
= k(g 1)(h 1) (h 1)(g 1)kop
and so by the triangle inequality
(1.2)

k[g, h] 1kop 2kg 1kop kh 1kop .

Now we can prove Jordans theorem.

Proof. We induct on n, the case n = 1 being trivial. Suppose first that
G contains a central element g (i.e. an element that commutes with every
element in G) which is not a multiple of the identity. Then, by definition,
G is contained in the centraliser Z(g) := {h Un (C) : gh = hg} of g, which
by the spectral theorem is isomorphic to a product Un1 (C) Unk (C)
of smaller unitary groups. Projecting G to each of these factor groups and
applying the induction hypothesis, we obtain the claim.
Thus we may assume that G contains no central elements other than
1
multiples of the identity. Now pick a small > 0 (one could take = 10n
in fact) and consider the subgroup G0 of G generated by those elements of
G that are within of the identity (in the operator norm). By considering
a maximal -net of G we see that G0 has index at most On, (1) in G. By
arguing as before, we may assume that G0 has no central elements other
than multiples of the identity.
If G0 consists only of multiples of the identity, then we are done. If not,
take an element g of G0 that is not a multiple of the identity, and which is
as close as possible to the identity (here is where we crucially use that G
is finite). Note that g is within of the identity. By (1.2), we see that if
is sufficiently small depending on n, and if h is one of the generators of
G0 , then [g, h] lies in G0 and is closer to the identity than g, and is thus a
multiple of the identity. On the other hand, [g, h] has determinant 1. Given
that it is so close to the identity, it must therefore be the identity (if is
small enough). In other words, g is central in G0 , and is thus a multiple of
the identity. But this contradicts the hypothesis that there are no central
elements other than multiples of the identity, and we are done.

Commutator estimates such as (1.2) will play a fundamental role in many
of the arguments we will see in this text; as we saw above, such estimates
combine very well with a discreteness hypothesis, but will also be very useful
in the continuous setting.
Exercise 1.1.3. Generalise Jordans theorem to the case when G is a finite
subgroup of GLn (C) rather than of Un (C). (Hint: The elements of G are

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1. Hilberts fifth problem

not necessarily unitary, and thus do not necessarily preserve the standard
Hilbert inner product of Cn . However, if one averages that inner product by
the finite group G, one obtains a new inner product on Cn that is preserved
by G, which allows one to conjugate G to a subgroup of Un (C). This
averaging trick is (a small) part of Weyls unitary trick in representation
theory.)
Remark 1.1.3. We remark that one can strengthen Jordans theorem further by relaxing the finiteness assumption on G to a periodicity assumption;
see Section 2.1.
Exercise 1.1.4 (Inability to discretise nonabelian Lie groups). Show that
if n 3, then the orthogonal group On (R) cannot contain arbitrarily dense
finite subgroups, in the sense that there exists an = n > 0 depending
only on n such that for every finite subgroup G of On (R), there exists a ball
of radius in On (R) (with, say, the operator norm metric) that is disjoint
from G. What happens in the n = 2 case?
Remark 1.1.4. More precise classifications of the finite subgroups of Un (C)
are known, particularly in low dimensions. For instance, it is a classical
result that the only finite subgroups of SO3 (R) (which SU2 (C) is a double
cover of) are isomorphic to either a cyclic group, a dihedral group, or the
symmetry group of one of the Platonic solids.
1.1.1. Hilberts fifth problem. One of the fundamental categories of
objects in modern mathematics is the category of Lie groups, which are
rich in both algebraic and analytic structure. Let us now briefly recall the
precise definition of what a Lie group is.
Definition 1.1.5 (Smooth manifold). Let d 0 be a natural number. A ddimensional topological manifold is a Hausdorff topological space M which
is locally Euclidean, thus every point in M has a neighbourhood which is
homeomorphic to an open subset of Rd .
A smooth atlas on an d-dimensional topological manifold M is a family
( )A of homeomorphisms : U V from open subsets U of M to
open subsets V of Rd , such that the U form an open cover of M , and for
any , A, the map 1
is smooth (i.e. infinitely differentiable) on
the domain of definition (U U ). Two smooth atlases are equivalent if
their union is also a smooth atlas; this is easily seen to be an equivalence
relation. An equivalence class of smooth atlases is a smooth structure. A
smooth manifold is a topological manifold equipped with a smooth structure.
A map : M M 0 from one smooth manifold to another is said to be
smooth if 0 1
is a smooth function on the domain of definition V
1
1
(U (U )) for any smooth charts , 0 in any the smooth atlases

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

of M, M 0 respectively (one easily verifies that this definition is independent

of the choice of smooth atlas in the smooth structure).
Note that we do not require manifolds to be connected, nor do we require them to be embeddable inside an ambient Euclidean space such as Rn ,
although certainly many key examples of manifolds are of this form. The
requirement that the manifold be Hausdorff is a technical one, in order to
exclude pathological examples such as the line with a doubled point (formally, consider the double line R {0, 1} after identifying (x, 0) with (x, 1)
for all x R\{0}), which is locally Euclidean but not Hausdorff1.
Remark 1.1.6. It is a plausible, but non-trivial, fact that a (non-empty)
topological manifold can have at most one dimension d associated to it;
thus a manifold M cannot both be locally homeomorphic to Rd and locally
0
homeomorphic to Rd unless d = d0 . This fact is a consequence of Brouwers
invariance of domain theorem; see Exercise 1.6.2. On the other hand, it
is an easy consequence of the rank-nullity theorem that a smooth manifold
can have at most one dimension, without the need to invoke invariance of
domain; we leave this as an exercise.
Definition 1.1.7 (Lie group). A Lie group is a group G = (G, ) which is
also a smooth manifold, such that the group operations : G G G and
()1 : G G are smooth maps. (Note that the Cartesian product of two
smooth manifolds can be given the structure of a smooth manifold in the
obvious manner.) We will also use additive notation G = (G, +) to describe
some Lie groups, but only in the case when the Lie group is abelian.
Remark 1.1.8. In some literature, Lie groups are required to be connected
(and occasionally, are even required to be simply connected), but we will
not adopt this convention here. One can also define infinite-dimensional Lie
groups, but in this text all Lie groups are understood to be finite dimensional.
Example 1.1.9. Every group can be viewed as a Lie group if given the
discrete topology (and the discrete smooth structure). (Note that we are
not requiring Lie groups to be connected.)
Example 1.1.10. Finite-dimensional vector spaces such as Rd are (additive) Lie groups, as are sublattices such as Zd or quotients such as Rd /Zd .
However, non-closed subgroups such as Qd are not manifolds (at least with
1In some literature, additional technical assumptions such as paracompactness, second countability, or metrisability are imposed to remove pathological examples of topological manifolds such
as the long line, but it will not be necessary to do so in this text, because (as we shall see later)
we can essentially get such properties for free for locally Euclidean groups.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1. Hilberts fifth problem

the topology induced from Rd ) and are thus not Lie groups; similarly, quotients such as Rd /Qd are not Lie groups either (they are not even Hausdorff).
Also, infinite-dimensional topological vector spaces (such as RN with the
product topology) will not be Lie groups.
Example 1.1.11. The general linear group GLn (C) of invertible n n
complex matrices is a Lie group. A theorem of Cartan (Theorem 1.3.2)
asserts that any closed subgroup of a Lie group is a smooth submanifold of
that Lie group and is in particular also a Lie group. In particular, closed
linear groups (i.e. closed subgroups of a general linear group) are Lie groups;
examples include the real general linear group GLn (R), the unitary group
Un (C), the special unitary group SUn (C), the orthogonal group On (R), the
special orthogonal group SOn (R), and the Heisenberg group

1 R R
0 1 R
0 0 1
of unipotent upper triangular 3 3 real matrices. Many Lie groups are
isomorphic to closed linear groups; for instance, the additive group R can
be identified with the closed linear group


1 R
,
0 1
the circle R/Z can be identified with SO2 (R) (or U1 (C)), and so forth.
However, not all Lie groups are isomorphic to closed linear groups. A somewhat trivial example is that of a discrete group with cardinality larger than
the continuum, which is simply too large to fit inside any linear group. A
less pathological example is provided by the Weil-Heisenberg group

1 R R/Z
1 R R
1 0 Z
R := 0 1 R / 0 1 0
(1.3)
G := 0 1
0 0
1
0 0 1
0 0 1
which is isomorphic to the image of the Heisenberg group under the Weil
representation, or equivalently the group of isometries of L2 (R) generated
by translations and modulations. Despite this, though, it is helpful to think
of closed linear groups and Lie groups as being almost the same concept
as a first approximation. For instance, one can show using Ados theorem
(Theorem 2.3.2) that every Lie group is locally isomorphic to a linear local
group (a concept we will discuss in Section 1.2.1).
An important subclass of the closed linear groups are the linear algebraic
groups, in which the group is also a real or complex algebraic variety (or at
least an algebraically constructible set). All of the examples of closed linear
groups given above are linear algebraic groups, although there exist closed

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

linear groups that are not isomorphic to any algebraic group; see Proposition
2.11.2.
Exercise 1.1.5 (Weil-Heisenberg group is not linear). Show that there is
no injective homomorphism : G GLn (C) from the Weil-Heisenberg
group (1.3) to a general linear group GLn (C) for any finite n. (Hint: The
centre [G, G] maps via to a circle subgroup of GLn (C); diagonalise this
subgroup and reduce to the case when the image of the centre consists of
multiples of the identity. Now, use the fact that commutators in GLn (C)
have determinant one.) This fact was first observed by Birkhoff.
Hilberts fifth problem, like many of Hilberts problems, does not have a
unique interpretation, but one of the most commonly accepted interpretations of the question posed by Hilbert is to determine if the requirement of
smoothness in the definition of a Lie group is redundant. (There is also an
analogue of Hilberts fifth problem for group actions, known as the HilbertSmith conjecture; see Section 2.7.) To answer this question, we need to relax
the notion of a Lie group to that of a topological group.
Definition 1.1.12 (Topological group). A topological group is a group G =
(G, ) that is also a topological space, in such a way that the group operations
: GG G and ()1 : G G are continuous. (As before, we also consider
additive topological groups G = (G, +) provided that they are abelian.)
Clearly, every Lie group is a topological group if one simply forgets the
smooth structure, retaining only the topological and group structures. Furthermore, such topological groups remain locally Euclidean. It was established by Montgomery-Zippin [MoZi1952] and Gleason [Gl1952] that the
converse statement holds, thus solving at least one formulation of Hilberts
fifth problem:
Theorem 1.1.13 (Hilberts fifth problem). Let G be an object with the
following properties:
(1) (Group-like object) G is a topological group.
(2) (Weak regularity) G is locally Euclidean.
Then
(i) (Lie-type structure) G is isomorphic to a Lie group.
Exercise 1.1.6. Show that a locally Euclidean topological group is necessarily Hausdorff (without invoking Theorem 1.1.13).
We will prove this theorem in Section 1.6. As it turns out, Theorem
1.1.13 is not directly useful for many applications, because it is often difficult
to verify that a given topological group is locally Euclidean. On the other

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

10

1. Hilberts fifth problem

hand, the weaker property of local compactness, which is clearly implied by

the locally Euclidean property, is much easier to verify in practice. One can
then ask the more general question of whether every locally compact group
is isomorphic to a Lie group. Unfortunately, the answer to this question is
easily seen to be no, as the following examples show:
Example 1.1.14 (Trivial topology). A group equipped with the trivial
topology is a compact (hence locally compact) group, but will not be Hausdorff (and thus not Lie) unless the group is also trivial. Of course, this is
a rather degenerate counterexample and can be easily eliminated in practice. For instance, we will see later that any topological group can be made
Hausdorff by quotienting out the closure of the identity.
Example 1.1.15 (Infinite-dimensional torus). The infinite-dimensional torus
(R/Z)N (with the product topology) is an (additive) topological group,
which is compact (and thus locally compact) by Tychonoff s theorem. However, it is not a Lie group.
Example 1.1.16 (p-adics). Let p be a prime. We define the p-adic norm
kkp on the integers Z by defining knkp := pj , where pj is the largest power
of p that divides n (with the convention k0kp := 0). This is easily verified to
generate a metric (and even an ultrametric) on Z; the p-adic integers Zp are
then defined as the metric completion of Z under this metric. This is easily
seen to be a compact (hence locally compact) additive group (topologically,
it is homeomorphic to a Cantor set). However, it is not locally Euclidean
(or even locally connected ), and so is not isomorphic to a Lie group.
One can also extend the p-adic norm to the ring Z[ p1 ] of rationals of the
form a/pj for some integers a, j in the obvious manner; the metric completion of this space is then the p-adic rationals Qp . This is now a locally
compact additive group rather than a compact one (Zp is a compact open
neighbourhood of the identity); it is still not locally connected, so it is still
not a Lie group.
One can also define algebraic groups such as GLn over the p-adic rationals Qp ; thus for instance GLn (Qp ) is the group of invertible n n matrices
with entries in the p-adics. This is still a locally compact group, and is
certainly not Lie.
Exercise 1.1.7 (Solenoid). Let p be a prime. Let G be the solenoid group
G := (Zp R)/Z , where Z := {(n, n) : n Z} is the diagonally embedded
copy of the integers in Zp R. (Topologically, G can be viewed as the set
Zp [0, 1] after identifying (x + 1, 1) with (x, 0) for all x Zp .) Show that G
is a compact additive group that is connected but not locally connected (and
thus not a Lie group). Thus one cannot eliminate p-adic type behaviour from
locally compact groups simply by restricting attention to the connected case

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

11

(although we will see later that one can do so by restricting to the locally
connected case).
We have now seen several examples of locally compact groups that are
not Lie groups. However, all of these examples are almost Lie groups in
that they can be turned into Lie groups by quotienting out a small compact
normal subgroup. (It is easy to see that the quotient of a locally compact
group by a compact normal subgroup is again a locally compact group.) For
instance, a group with the trivial topology becomes Lie after quotienting out
the entire group (which is small in the sense that it is contained in every
open neighbourhood of the origin). The infinite-dimensional torus (R/Z)N
can be quotiented into a finite-dimensional torus (R/Z)d (which is of course
a Lie group) by quotienting out the compact subgroup {0}d (R/Z)N ; note
from the definition of the product topology that these compact subgroups
shrink to zero in the sense that every neighbourhood of the group identity
contains at least one (and in fact all but finitely many) of these subgroups.
Similarly, with the p-adic group Zp , one can quotient out by the compact
(and open) subgroups pj Zp (which also shrink to zero, as discussed above) to
obtain the cyclic groups Z/pj Z, which are discrete and thus Lie. Quotienting
out Qp by the same compact open subgroups pj Zp also leads to discrete
(hence Lie) quotients; similarly for algebraic groups defined over Qp , such
as GLn (Qp ). Finally, with the solenoid group G := (Zp R)/Z , one can
quotient out the copy of pj Zp {0} in G for j = 0, 1, 2, . . . (which are
another sequence of compact subgroups shrinking to zero) to obtain the
quotient group (Z/pj Z R)/Z , which is isomorphic to a (highly twisted)
circle R/Z and is thus Lie.
Inspired by these examples, we might be led to the following conjecture:
if G is a locally compact group, and U is a neighbourhood of the identity,
then there exists a compact normal subgroup K of G contained in U such
that G/K is a Lie group. In the event that G is Hausdorff, this is equivalent
to asserting that G is the projective limit (or inverse limit) of Lie groups.
This conjecture is true in several cases; for instance, one can show using
the Peter-Weyl theorem (which we will discuss in Section 1.4) that it is true
for compact groups, and we will later see that it is also true for connected
locally compact groups (see Theorem 1.6.1). However, it is not quite true
in general, as the following example shows.
Exercise 1.1.8. Let p be a prime, and let T : Qp Qp be the automorphism T x := px. Let G := Qp oT Z be the semidirect product of Qp and Z
twisted by T ; more precisely, G is the Cartesian product Qp Z with the
product topology and the group law
(x, n)(y, m) := (x + T n y, n + m).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

12

1. Hilberts fifth problem

Show that G is a locally compact group which is not isomorphic to a Lie

group, and that Zp {0} is an open neighbourhood of the identity that
contains no non-trivial normal subgroups of G. Conclude that the conjecture
stated above is false.
The difficulty in the above example was that it was not easy to keep a
subgroup normal with respect to the entire group Qp oT Z. Note however
that G contains a large (and more precisely, open) subgroup Qp {0}
which is the projective limit of Lie groups. So the above examples do not
rule out that the conjecture can still be salvaged if one passes from a group
G to an open subgroup G0 . This is indeed the case:
Theorem 1.1.17 (Gleason-Yamabe theorem). [Gl1951, Ya1953b] Let G
obey the following hypotheses:
(1) (Group-like object) G is a topological group.
(2) (Weak regularity) G is locally compact.
Then for every open neighbourhood U of the identity, there exists a subgroup G0 of G and a compact normal subgroup K of G0 with the following
properties:
(i) (G0 /K is close to G) G0 is an open subgroup of G, and K is contained in U .
(ii) (Lie-type structure) G0 /K is isomorphic to a Lie group.
The proof of this theorem will occupy the next few sections of this text,
being finally proven in Section 1.5. As stated, G0 may depend on U , but one
can in fact take the open subgroup G0 to be uniform in the choice of U ; we
will show this in later sections. Theorem 1.1.13 can in fact be deduced from
Theorem 1.1.17 and some topological arguments involving the invariance of
domain theorem; this will be shown in Section 1.6.
The Gleason-Yamabe theorem asserts that locally compact groups are
essentially Lie groups, after ignoring the very large scales (by restricting
to an open subgroup) and also ignoring the very small scales (by allowing
one to quotient out by a small group). In special cases, the conclusion of the
theorem can be simplified. For instance, it is easy to see that an open subgroup G0 of a topological group G is also closed (since the complement G\G0
is a union of cosets of G0 ), and so if G is connected, there are no open subgroups other than G itself. Thus, in the connected case of Theorem 1.1.17,
one can take G = G0 . In a similar spirit, if G has the no small subgroups
(NSS) property, that is to say that there exists an open neighbourhood of
the identity that contains no non-trivial subgroups of G, then we can take
K to be trivial. Thus, as a special case of the Gleason-Yamabe theorem, we
see that all connected NSS locally compact groups are Lie; in fact it is not

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

13

Figure 1. A schematic description of the steps needed to establish the

Gleason-Yamabe theorem. The annotation O, Q on an arrow indicates
that one has to pass to an open subgroup, and then quotient out a
compact normal subgroup, in order to obtain the additional structure
at the end of the arrow.

difficult to then conclude that any locally compact NSS group (regardless of
connectedness) is Lie. Conversely, this claim (which we isolate as Corollary
1.5.8) turns out to be a key step in the proof of Theorem 1.1.17, as we shall
see later. (It is also not difficult to show that all Lie groups are NSS; see
Exercise 1.5.6.)
The proof of the Gleason-Yamabe theorem proceeds in a somewhat
lengthy series of steps in which the initial regularity (local compactness)
on the group G is gradually upgraded to increasingly stronger regularity
(e.g. metrisability, the NSS property, or the locally Euclidean property) until one eventually obtains Lie structure; see Figure 1. A key turning point
in the argument will be the construction of a metric (which we call a Gleason metric) on (a large portion of) G which obeys a commutator estimate
similar to (1.2).
While the Gleason-Yamabe theorem does not completely classify all locally compact groups (as mentioned earlier, it primarily controls the mediumscale behaviour, and not the very fine-scale or very coarse-scale behaviour,

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

14

1. Hilberts fifth problem

of such groups), it is still powerful enough for a number of applications, to

which we now turn.
1.1.2. Approximate groups. We now discuss what appears at first glance
to be an unrelated topic, namely that of additive combinatorics (and its noncommutative counterpart, multiplicative combinatorics). One of the main
objects of study in either additive or multiplicative combinatorics are approximate groups - sets A (typically finite) contained in an additive or multiplicative ambient group G that are almost groups in the sense that they
are almost closed under either addition or multiplication. (One can also
consider abstract approximate groups that are not contained in an ambient
genuine group, but we will not do so here.)
There are several ways to quantify what it means for a set A to be
almost closed under addition or multiplication. Here are some common
formulations of this idea (phrased in multiplicative notation, for sake of
concreteness):
(1) (Statistical multiplicative structure) For a large proportion of
pairs (a, b) A A, the product ab also lies in A.
(2) (Small product set) The size of the product set A A := {ab :
a, b A} is comparable to the size of the original set A. (For
technical reasons, one sometimes uses the triple product A A A :=
{abc : a, b, c A} instead of the double product.)
(3) (Covering property) The product set A A can be covered by a
bounded number of (left or right) translates of the original set
A.
Of course, to make these notions precise one would have to precisely quantify the various terms in quotes. Fortunately, the basic theory of additive
combinatorics (and multiplicative combinatorics) can be used to show that
all these different notions of additive or multiplicative structure are essentially equivalent; see [TaVu2006, Chapter 2] or [Ta2008b] for more
discussion.
For the purposes of this text, it will be convenient to focus on the use of
covering to describe approximate multiplicative structure. More precisely:
Definition 1.1.18 (Approximate groups). Let G be a multiplicative group,
and let K 1 be a real number. A K-approximate subgroup of G, or Kapproximate group for short, is a subset A of G which contains the identity,
is symmetric (thus A1 := {a1 : a A} is equal to A) and is such that AA
can be covered by at most K left-translates (or equivalently by symmetry,
right translates) of A, thus there exists a subset X of G of cardinality at
most K such that A A X A.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

15

In most combinatorial applications, one only considers approximate groups

that are finite sets, but one could certainly also consider countably or uncountably infinite approximate groups also. We remark that this definition
is essentially from [Ta2008b] (although the definition in [Ta2008b] places
some additional minor constraints on the set X which have turned out not
to be terribly important in practice).
Example 1.1.19. A 1-approximate subgroup of G is the same thing as a
genuine subgroup of G.
Example 1.1.20. In the additive group of the integers Z, the symmetric arithmetic progression {N, . . . , N } is a 2-approximate group for any
N 1. More generally, in any additive group G, the symmetric generalised
arithmetic progression
{a1 v1 + + ar vr : a1 , . . . , ar Z, |ai | Ni i = 1, . . . , r}
with v1 , . . . , vr G and N1 , . . . , Nr > 0, is a 2r -approximate group.
Exercise 1.1.9. Let A be a convex symmetric subset of Rd . Show that A
is a 5d -approximate group. (Hint: greedily pack 2A with disjoint translates
of 21 A.)
Example 1.1.21. If A is an open precompact 2 symmetric neighbourhood of
the identity in a locally compact group G, then A is a K-approximate group
for some finite K. Thus we see some connection between locally compact
groups and approximate groups; we will see a deeper connection involving
ultraproducts in Section 1.7.
Example 1.1.22. Let G be a d-dimensional Lie group. Then G is a smooth
manifold, and can thus be (non-uniquely) given the structure of a Riemannian manifold. If one does so, then for sufficiently small radii r, the ball
B(1, r) around the identity 1 will be a Od (1)-approximate group.
Example 1.1.23 (Extensions). Let : G H be a surjective group homomorphism (thus G is a group extension of H by the kernel ker() of ).
If A is a K-approximate subgroup of H, then 1 (A) is a K-approximate
subgroup of G. One can think of 1 (A) as an extension of the approximate
group A by ker().
The classification of approximate groups is of importance in additive
combinatorics, and has connections with number theory, geometric group
theory, and the theory of expander graphs. One can ask for a quantitative classification, in which one has explicit dependence of constants on the
approximate group parameter K, or one can settle for a qualitative classification in which one does not attempt to control this dependence of constants.
2A subset of a topological space is said to be precompact if its closure is compact.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

16

1. Hilberts fifth problem

In this text we will focus on the latter question, as this allows us to bring
in qualitative tools such as the Gleason-Yamabe theorem to bear on the
problem.
In the abelian case when the ambient group G is additive, approximate
groups are classified by Freimans theorem for abelian groups 3 [GrRu2007].
As before, we phrase this theorem in a slightly stilted fashion (and in a
qualitative, rather than quantitative, manner) in order to demonstrate its
alignment with the general principles stated in the introduction.
Theorem 1.1.24 (Freimans theorem in an abelian group). Let A be an
object with the following properties:
(1) (Group-like object) A is a subset of an additive group G.
(2) (Weak regularity) A is a K-approximate group.
(3) (Discreteness) A is finite.
Then there exists a finite subgroup H of G, and a subset P of G/H, with
the following properties:
(i) (P is close to A/H) 1 (P ) is contained in 4A := A + A + A + A,
where : G G/H is the quotient map, and |P | K |A|/|H|.
(ii) (Nilpotent type structure) P is a symmetric generalised arithmetic
progression of rank OK (1) (see Example 1.1.20).
Informally, this theorem asserts that in the abelian setting, discrete
approximate groups are essentially bounded rank symmetric generalised
arithmetic progressions, extended by finite groups (such extensions are also
known as coset progressions). The theorem has a simpler conclusion (and is
simpler to prove) in the case when G is a torsion-free abelian group (such
as Z), since in this case H is trivial.
We will not discuss the proof of Theorem 1.1.24 from [GrRu2007] here,
save to say that it relies heavily on Fourier-analytic methods, and as such,
does not seem to easily extend to a general non-abelian setting. To state the
non-abelian analogue of Theorem 1.1.24, one needs multiplicative analogues
of the concept of a generalised arithmetic progression. An ordinary (symmetric) arithmetic progression {N v, . . . , N v} has an obvious multiplicative
analogue, namely a (symmetric) geometric progression {aN , . . . , aN } for
some generator a G. In a similar vein, if one has r commuting generators
a1 , . . . , ar and some dimensions N1 , . . . , Nr > 0, one can form a symmetric
generalised geometric progression
(1.4)

P := {an1 1 . . . anr r : |ni | Ni 1 i r},

3The original theorem of Freiman [Fr1973] obtained an analogous classification in the case
when G was a torsion-free abelian group, such as the integers Z.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

17

which will still be an 3r -approximate group. However, if the a1 , . . . , ar do

not commute, then the set P defined in (1.4) is not quite the right concept to
use here; for instance, there it is no reason for P to be symmetric. However,
it can be modified as follows:
Definition 1.1.25 (Noncommutative progression). Let a1 , . . . , ar be elements of a (not necessarily abelian) group G = (G, ), and let N1 , . . . , Nr > 0.
We define the noncommutative progression P = P (a1 , . . . , ar ; N1 , . . . , Nr ) of
rank r with generators a1 , . . . , ar and dimensions N1 , . . . , Nr to be the collec1
tion of all words w composed using the alphabet a1 , a1
1 , . . . , ar , ar , such
that for each 1 i r, the total number of occurrences of ai and a1
i
combined in w is at most Ni .
Example 1.1.26. P (a, b; 1, 2) consists of the elements
1, a , b , b a , a b , b2 , b2 a , b a b , a b2 ,
where each occurrence of can independently be set to + or ; thus
P (a, b; 1, 2) can have as many as 31 elements.
Example 1.1.27. If the a1 , . . . , ar commute, then the noncommutative progression P (a1 , . . . , ar ; N1 , . . . , Nr ) simplifies to (1.4).
Exercise 1.1.10. Let G be the discrete Heisenberg group

1 Z Z
(1.5)
G = 0 1 Z
0 0 1
and let

1 1 0
1 0 0
e1 := 0 1 0 , e2 := 0 1 1
0 0 1
0 0 1
be the two generators of G. Let N 1 be a sufficiently large natural number.
Show that the noncommutative
progression P (e1 , e2 ; N, N ) contains all the

1 a c
group elements 0 1 b of G with |a|, |b| N and |c| N 2 for a
0 0 1
sufficiently small absolute constant > 0; conversely,
show that all elements

1 a c
of P (e1 , e2 ; N, N ) are of the form 0 1 b with |a|, |b| CN and |c|
0 0 1
2
CN for some sufficiently large absolute constant C > 0. Thus, informally,
we have

1 O(N ) O(N 2 )
1
O(N ) .
P (e1 , e2 ; N, N ) = 0
0
0
1

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

18

1. Hilberts fifth problem

It is clear that noncommutative progressions P (a1 , . . . , ar ; N1 , . . . , Nr )

are symmetric and contain the identity. However, if the a1 , . . . , ar do not
have any commutative properties, then the size of these progressions can
grow exponentially in N1 , . . . , Nr and will not be approximate groups with
any reasonable parameter K. However, the situation changes when the
a1 , . . . , ar generate a nilpotent group 4:
Proposition 1.1.28. Suppose that a1 , . . . , ar G generate a nilpotent group
of step s, and suppose that N1 , . . . , Nr are all sufficiently large depending on
r, s. Then P (a1 , . . . , ar ; N1 , . . . , Nr ) is an Or,s (1)-approximate group.
We will prove this proposition in Section 2.2.
We can now state the noncommutative analogue of Theorem 1.1.24,
proven in [BrGrTa2011]:
Theorem 1.1.29 (Freimans theorem in an arbitrary group). Let A be an
object with the following properties:
(1) (Group-like object) A is a subset of a multiplicative group G.
(2) (Weak regularity) A is a K-approximate group.
(3) (Discreteness) A is finite.
Then there exists a finite subgroup H of G, and a subset P of N (H)/H
(where N (H) := {g G : gH = Hg} is the normaliser of H), with the
following properties:
(i) (P is close to A/H) 1 (P ) is contained in A4 := AAAA, where
: N (H) N (H)/H is the quotient map, and |P | K |A|/|H|.
(ii) (Nilpotent type structure) P is a noncommutative progression of
rank OK (1), whose generators generate a nilpotent group of step
OK (1).
The proof of this theorem relies on the Gleason-Yamabe theorem (Theorem 1.1.17), and will be discussed in Section 1.8. The key connection will
take some time to explain properly, but roughly speaking, it comes from
the fact that the ultraproduct of a sequence of K-approximate groups can
be used to generate a locally compact group, to which the Gleason-Yamabe
theorem can be applied. This in turn can be used to place a metric on approximate groups that obeys a commutator estimate similar to (1.2), which
allows one to run an argument similar to that used to prove Theorem 1.1.2.
4A group G is nilpotent if the lower central series G := G, G := [G, G ], G := [G, G ], . . . ,
1
2
1
3
2
etc. eventually becomes trivial.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

19

1.1.3. Gromovs theorem. The final topic of this chapter will be Gromovs theorem on groups of polynomial growth [Gr1981]. This theorem
is analogous to Theorem 1.1.17 or Theorem 1.1.29, but in the category of
finitely generated groups rather than locally compact groups or approximate
groups.
Let G be a group that is generated by a finite set S of generators;
for notational simplicity we will assume that S is symmetric and contains
the origin. Then S defines a (right-invariant) word metric on G, defined
by setting d(x, y) for x, y G to be the least natural number n such that
x S n y. One easily verifies that this is indeed a metric that is right-invariant
(thus d(xg, yg) = d(x, y) for all x, y, g G). Geometrically, this metric
describes the geometry of the Cayley graph on G formed by connecting x to
sx for each x G and s S. (See [Ta2011c, 2.3] for more discussion of
using Cayley graphs to study groups geometrically.)
Let us now consider the growth of the balls B(1, R) = S bRc as R ,
where bRc is the integer part of R. On the one hand, we have the trivial
upper bound
|B(1, R)| |S|R
that shows that such balls can grow at most exponentially. And for typical
non-abelian groups, this exponential growth actually occurs; consider the
case for instance when S consists of the generators of a free group (together
with their inverses, and the group identity). However, there are some groups
for which the balls grow at a much slower rate. A somewhat trivial example
is that of a finite group G, since clearly |B(1, R)| will top out at |G| (when
R reaches the diameter of the Cayley graph) and stop growing after that
point. Another key example is the abelian case:
Exercise 1.1.11. If G is an abelian group generated by a finite symmetric
set S containing the identity, show that
|B(1, R)| (1 + R)|S| .
In particular, B(1, R) grows at a polynomial rate in R.
Let us say that a finitely group G is a group of polynomial growth if one
has |B(1, R)| CRd for all R 1 and some constants C, d > 0.
Exercise 1.1.12. Show that the notion of a group of polynomial growth
(as well as the rate d of growth) does not depend on the choice of generators
S; thus if S 0 is another set of generators for G, show that G has polynomial
growth with respect to S with rate d if and only if it has polynomial growth
with respect to S 0 with rate d.
Exercise 1.1.13. Let G be a finitely generated group, and let G0 be a finite
index subgroup of G.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

20

1. Hilberts fifth problem

(i) Show that G0 is also finitely generated. (Hint: Let S be a symmetric

set of generators for G containing the identity, and locate a finite
integer n such that S n+1 G0 = S n G0 . Then show that the set S 0 :=
G0 S 2n+1 is such that S n+1 S n S 0 . Conclude that S n meets
every coset of hS 0 i (or equivalently that G = S n hS 0 i), and use this
to show that S 0 generates G0 .)
(ii) Show that G has polynomial growth if and only if G0 has polynomial
growth.
(iii) More generally, show that any finitely generated subgroup of a
group of polynomial growth also has polynomial growth. Conclude
in particular that a group of polynomial growth cannot contain the
free group on two generators.
From Exercise 1.1.10 we see that the discrete Heisenberg group (1.5) is
of polynomial growth. It is in fact not difficult to show that more generally,
any nilpotent finitely generated group is of polynomial growth. By Exercise
1.1.13, this implies that any virtually nilpotent finitely generated group is of
polynomial growth.
Gromovs theorem asserts the converse statement:
Theorem 1.1.30 (Gromovs theorem). [Gr1981] Let G be an object with
the following properties:
(1) (Group-like object) G is a finitely generated group.
(2) (Weak regularity) G is of polynomial growth.
Then there exists a subgroup G0 of G such that
(i) (G0 is close to G) The index |G/G0 | is finite.
(ii) (Nilpotent type structure) G0 is nilpotent.
More succinctly: a finitely generated group is of polynomial growth if
and only if it is virtually nilpotent.
Groups of polynomial growth are related to approximate groups by the
following observation.
Exercise 1.1.14 (Pigeonhole principle). Let G be a finitely generated group
of polynomial growth, and let S be a symmetric set of generators for G
containing the identity.
(i) Show that there exists a C > 1 such that |B(1, 5R/2)| C|B(1, R/2)|
for a sequence R = Rn of radii going to infinity.
(ii) Show that there exists a K > 1 such that B(1, R) is a K-approximate
group for a sequence R = Rn of radii going to infinity. (Hint: Argue
as in Exercise 1.1.9.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.1. Introduction

21

In Section 1.9 we will use this connection to deduce Theorem 1.1.30 from
Theorem 1.1.29. From a historical perspective, this was not the first proof
of Gromovs theorem; Gromovs original proof in [Gr1981] relied instead on
a variant of Theorem 1.1.13 (as did some subsequent variants of Gromovs
argument, such as the nonstandard analysis variant in [vdDrWi1984]), and
a subsequent proof of Kleiner [Kl2010] went by a rather different route,
based on earlier work of Colding and Minicozzi [CoMi1997] on harmonic
functions of polynomial growth. (This latter proof is discussed in [Ta2009,
1.2] and [Ta2011c, 2.5].) The proof we will give in this text is more
recent, based on an argument of Hrushovski [Hr2012]. We remark that
the strategy used to prove Theorem 1.1.29 - namely taking an ultralimit
of a sequence of approximate groups - also appears in Gromovs original
argument5. We will discuss these sorts of limits more carefully in Section
1.7, but an informal example to keep in mind for now is the following: if
one takes a discrete group (such as Zd ) and rescales it (say to N1 Zd for a
large parameter N ), then intuitively this rescaled group converges to a
continuous group (in this case Rd ). More generally, one can generate locally
compact groups (or at least locally compact spaces) out of the limits of
(suitably normalised) groups of polynomial growth or approximate groups,
which is one of the basic observations that tie the three different topics
discussed above together.
As we shall see in Section 1.10, finitely generated groups arise naturally
as the fundamental groups of compact manifolds. Using the tools of Riemannian geometry (such as the Bishop-Gromov inequality), one can relate
the growth of such groups to the curvature of a metric on such a manifold.
As a consequence, Gromovs theorem and its variants can lead to some nontrivial conclusions about the relationship between the topology of a manifold
and its geometry. The following simple consequence is typical:
Proposition 1.1.31. Let M be a compact Riemannian manifold of nonnegative Ricci curvature. Then the fundamental group 1 (M ) of M is virtually nilpotent.
We will discuss this result and some related results (such as a relaxation
of the non-negative curvature hypothesis to an almost non-negative curvature hypothesis) in Section 1.10. We also remark that the above proposition
can also be proven (with stronger conclusions) by more geometric means,
but there are some results of the above type which currently have no known
proof that does not employ some version of Gromovs theorem at some point.
5Strictly speaking, he uses Gromov-Hausdorff limits instead of ultralimits, but the two types
of limits are closely related, as we shall see in Section 1.7.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

22

1. Hilberts fifth problem

1.2. Lie groups, Lie algebras, and the

Baker-Campbell-Hausdorff formula
In this section, we describe the basic analytic structure theory of Lie groups,
by relating them to the simpler concept of a Lie algebra. Roughly speaking,
the Lie algebra encodes the infinitesimal structure of a Lie group, but
is a simpler object, being a vector space rather than a nonlinear manifold.
Nevertheless, thanks to the fundamental theorems of Lie, the Lie algebra
can be used to reconstruct the Lie group (at a local level, at least), by
means of the exponential map and the Baker-Campbell-Hausdorff formula.
As such, the local theory of Lie groups is completely described (in principle,
at least) by the theory of Lie algebras, which leads to a number of useful
consequences, such as the following:
(1) (Local Lie implies Lie) A topological group G is Lie (i.e. it is
isomorphic to a Lie group) if and only if it is locally Lie (i.e. the
group operations are smooth near the origin).
(2) (Uniqueness of Lie structure) A topological group has at most one
smooth structure on it that makes it Lie.
(3) (Weak regularity implies strong regularity, I) Lie groups are automatically real analytic. (In fact one only needs a local C 1,1
regularity on the group structure to obtain real analyticity.)
(4) (Weak regularity implies strong regularity, II) A continuous homomorphism from one Lie group to another is automatically smooth
(and real analytic).
The connection between Lie groups and Lie algebras also highlights the
role of one-parameter subgroups of a topological group, which will play a
central role in the solution of Hilberts fifth problem (cf. Figure 1).
Remark 1.2.1. There is also a very important algebraic structure theory
of Lie groups and Lie algebras, in which the Lie algebra is split into solvable
and semisimple components, with the latter being decomposed further into
simple components, which can then be completely classified using Dynkin
diagrams. This classification is of fundamental importance in many areas
of mathematics (e.g. representation theory, arithmetic geometry, and group
theory), and many of the deeper facts about Lie groups and Lie algebras are
proven via this classification (although in such cases it can be of interest to
also find alternate proofs that avoid the classification). However, it turns
out that we will not need this theory here, and so we will not discuss it
further (though it can of course be found in any graduate text on Lie groups
and Lie algebras, e.g. [Bo1968]).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

23

1.2.1. Local groups. The connection between Lie groups and Lie algebras
will be local in nature - the only portion of the Lie group that will be of
importance will be the portion that is close to the group identity 1. To
formalise this locality, it is convenient to introduce the notion of a local
group and a local Lie group, which are local versions of the concept of a
topological group and a Lie group respectively. We will only set up the
barest bones of the theory of local groups here; a more detailed discussion
is given in Section 2.5.
Definition 1.2.2 (Local group). A local topological group G = (G, , , 1, , ()1 ),
or local group for short, is a topological space G equipped with an identity
element 1 G, a partially defined but continuous multiplication operation
: G for some domain G G, and a partially defined but continuous inversion operation ()1 : G, where G, obeying the following
axioms:
(1) (Local closure) is an open neighbourhood of G {1} {1} G,
and is an open neighbourhood of 1.
(2) (Local associativity) If g, h, k G are such that (gh)k and g(hk)
are both well-defined in G, then they are equal. (Note however that
it may be possible for one of these products to be defined but not
the other.)
(3) (Identity) For all g G, g 1 = 1 g = g.
(4) (Local inverse) If g G and g 1 are well-defined in G, then6 g
g 1 = g 1 g = 1. (In particular this, together with the other
axioms, forces 11 = 1.)
We will sometimes use additive notation for local groups if the groups
are abelian (by which we mean the statement that if g + h is defined, then
h + g is also defined and equal to g + h.)
A local group is said to be symmetric if = G, i.e. if every element g
in G has an inverse g 1 that is also in G.
A local Lie group is a local group that is also a smooth manifold, in such
a fashion that the partially defined group operations , ()1 are smooth on
their domain of definition.
Clearly, every topological group is a local group, and every Lie group is
a local Lie group. We will sometimes refer to the former concepts as global
topological groups and global Lie groups in order to distinguish them from
their local counterparts. One could also consider local discrete groups, in
6Here we adopt the convention that any mathematical sentence involving an undefined operation is automatically false, thus for instance g g 1 = 1 is false unless g g 1 is well-defined,
so that (g, g 1 ) .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

24

1. Hilberts fifth problem

which the topological structure is just the discrete topology, but we will not
need to study such objects in here.
A model class of examples of a local (Lie) group comes from restricting a
global (Lie) group to an open neighbourhood of the identity. Let us formalise
this concept:
Definition 1.2.3 (Restriction). If G is a local group, and U is an open
neighbourhood of the identity in G, then we define the restriction G U
of G to U to be the topological space U with domains U := {(g, h)
: g, h, g h U } and U := {g : g, g 1 U }, and with the group
operations , ()1 being the restriction of the group operations of G to U ,
U respectively. If U is symmetric (in the sense that g 1 is well-defined
and lies in U for all g U ), then this restriction G U will also be symmetric.
If G is a global or local Lie group, then G U will also be a local Lie group.
We will sometimes abuse notation and refer to the local group G U simply
as U .
Thus, for instance, one can take the Euclidean space Rd , and restrict it
to a ball B centred at the origin, to obtain an additive local group Rd B . In
this group, two elements x, y in B have a well-defined sum x + y only when
their sum in Rd stays inside B. Intuitively, this local group behaves like the
global group Rd as long as one is close enough to the identity element 0,
but as one gets closer to the boundary of B, the group structure begins to
break down.
It is natural to ask the question as to whether every local group arises as
the restriction of a global group. The answer to this question is somewhat
complicated, and can be summarised as essentially yes in certain circumstances, but not in general; see Section 2.5.
A key example of a local Lie group arises from pushing forward a Lie
group via a coordinate chart near the origin:
Example 1.2.4. Let G be a global or local Lie group of some dimension d,
and let : U V be a smooth coordinate chart from a neighbourhood U
of the identity 1 in G to a neighbourhood V of the origin 0 in Rd , such that
maps 1 to 0. Then we can define a local group G U which is the set
V (viewed as a smooth submanifold of Rd ) with the local group identity 0,
the local group multiplication law defined by the formula
x y := (1 (x) 1 (y))
defined whenever 1 (x), 1 (y), 1 (x) 1 (y) are well-defined and lie in
U , and the local group inversion law ()1 defined by the formula
x1 := (1 (x)1 )

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

25

defined whenever 1 (x), 1 (x)1 are well-defined and lie in U . One easily
verifies that G U is a local Lie group. We will sometimes denote this local
Lie group as (V, ), to distinguish it from the additive local Lie group (V, +)
arising by restriction of (Rd , +) to V . The precise distinction between the
two local Lie groups will in fact be a major focus of this section.
Example 1.2.5. Let G be the Lie group GLn (R), and let U be the ball
U := {g GLn (R) : kg 1kop < 1}. If we then let V Mn (R) be the ball
V := {x Mn (R) : kxkop < 1} and be the map (g) := g 1, then
is a smooth coordinate chart (after identifying Mn (R) with Rnn ), and by
the construction in the preceding exercise, V = G U becomes a local Lie
group with the operations
x y := x + y + xy
(defined whenever x, y, x + y + xy all lie in V ) and
x1 := (1 + x)1 1 = x x2 + x3 . . .
(defined whenever x and (1+x)1 1 both lie in V ). Note that this Lie group
structure is not equal to the additive structure (V, +) on V , nor is it equal
to the multiplicative structure (V, ) on V given by matrix multiplication,
which is one of the reasons why we use the symbol instead of + or for
such structures.
Many (though not all) of the familiar constructions in group theory can
be generalised to the local setting, though often with some slight additional
subtleties. We will not systematically do so here, but we give a single such
generalisation for now:
Definition 1.2.6 (Homomorphism). A continuous homomorphism : G
H between two local groups G, H is a continuous map from G to H with
the following properties:
(i) maps the identity 1G of G to the identity 1H of H: (1G ) = 1H .
(ii) If g G is such that g 1 is well-defined in G, then (g)1 is welldefined in H and is equal to (g 1 ).
(iii) If g, h G are such that g h is well-defined in G, then (g) (h)
is well-defined and equal to (g h).
A smooth homomorphism : G H between two local Lie groups G, H is
a continuous homomorphism that is also smooth.
A (continuous) local homomorphism : U H between two local
groups G, H is a continuous homomorphism from an open neighbourhood
U of the identity in G to H. Two local homomorphisms are said to hbe
equivalent if they agree on a (possibly smaller) open neighbourhood of the

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

26

1. Hilberts fifth problem

identity. One can of course define the notion of a smooth local homomorphism similarly.
It is easy to see that the composition of two continuous homomorphisms
is again a continuous homomorphism, and that the identity map on a local
group is automatically a continuous homomorphism; this gives the class of
local groups the structure of a category. Similarly, the class of local Lie
groups with their smooth homomorphisms is also a category.
Example 1.2.7. With the notation of Example 1.2.4, : U V is a
smooth homomorphism from the local Lie group G U to the local Lie group
G U . In fact, it is a smooth isomorphism, since 1 : V U provides
the inverse homomorphism.
Let us say that a word g1 . . . gn in a local group G is well-defined in
G (or well-defined, for short) if every possible way of associating this word
using parentheses is well-defined from applying the product operation. For
instance, in order for abcd to be well-defined, ((ab)c)d, (a(bc))d, (ab)(cd),
a(b(cd)), and a((bc)d) must all be well-defined. For instance, in the additive
local group {9, . . . , 9} (with the group structure restricted from that of
the integers Z), 2 + 6 + 5 is not well-defined because one of the ways of
associating this sum, namely 2 + (6 + 5), is not well-defined (even though
(2 + 6) + 5 is well-defined).
Exercise 1.2.1 (Iterating the associative law).
(i) Show that if a word g1 . . . gn in a local group G is well-defined, then
all ways of associating this word give the same answer, and so we
can uniquely evaluate g1 . . . gn as an element in G.
(ii) Give an example of a word g1 . . . gn in a local group G which has
two ways of being associated that are both well-defined, but give
different answers. (Hint: the local associativity axiom prevents
this from happening for n 3, so try n = 4. A small discrete local
group will already suffice to give a counterexample; verifying the
local group axioms are easier if one makes the domain of definition
of the group operations as small as one can get away with while
still having the counterexample.)
Exercise 1.2.2. Show that the number of ways
to associate a word g1 . . . gn
is given by the Catalan number Cn1 := n1 2n2
n1 .
Exercise 1.2.3. Let G be a local group, and let m 1 be an integer. Show
that there exists a symmetric open neighbourhood Um of the identity such
that every word of length m in Um is well-defined in G (or more succinctly,
m is well-defined). (Note though that these words will usually only take
Um

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

27

values in G, rather than in Um , and also the sets Um tend to become smaller
as m increases.)
1.2.2. Some differential geometry. To define the Lie algebra of a Lie
group, we must first quickly recall some basic notions from differential geometry associated to smooth manifolds (which are not necessarily embedded
in some larger Euclidean space, but instead exist intrinsically as abstract
geometric structures). This requires a certain amount of abstract formalism
in order to define things rigorously, though for the purposes of visualisation,
it is more intuitive to view these concepts from a more informal geometric
perspective.
We begin with the concept of the tangent space and related structures.
Definition 1.2.8 (Tangent space). Let M be a smooth d-dimensional manifold. At every point x of this manifold, we can define the tangent space
Tx M of M at x. Formally, this tangent space can be defined as the space of
all continuously differentiable curves : I G defined on an open interval
I containing 0 with (0) = x, modulo the relation that two curves 1 , 2 are
considered equivalent if they have the same derivative at 0, in the sense that
d
d
(1 (t))|t=0 = (2 (t))|t=0
dt
dt
where : U V is a coordinate chart of G defined in a neighbourhood of
x; it is easy to see from the chain rule that this equivalence is independent
of the actual choice of . Using such a coordinate chart, one can identify
the tangent space Tx M with the Euclidean space Rd , by identifying with
d
dt ((t))|t=0 . One easily verifies that this gives Tx M the structure of a ddimensional vector space, in a manner which is independent of the choice
of coordinate chart . Elements of Tx M are called tangent vectors of M at
x. If : I G is a continuously differentiable curve with (0) = x, the
equivalence class of in Tx M will be denoted 0 (0).
S
The space T M := xM ({x} Tx M ) of pairs (x, v), where x is a point
in M and v is a tangent vector of M at x, is called the tangent bundle.
If : M N is a smooth map between two manifolds, we define the
derivative map D : T M T N to be the map defined by setting
D((x, 0 (0))) := ((x), ( )0 (0))
for all continously differentiable curves : I G with (0) = x for
some x M ; one can check that this map is well-defined. We also write
((x), D(x)(v)) for D(x, v), so that for each x M , D(x) is a map
from Tx M to T(x) N . One can easily verify that this latter map is linear.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

28

1. Hilberts fifth problem

We observe the chain rule 7

(1.6)

D( ) = (D) (D)

for any smooth maps : M N , : N O.

Observe that if V is an open subset of Rd , then T V may be identified
with V Rd . In particular, every coordinate chart : U V of M gives
rise to a coordinate chart D : T U V Rd of T M , which gives T M the
structure of a smooth 2d-dimensional manifold.
Remark 1.2.9. Informally, one can think of a tangent vector (x, v) as an
infinitesimal vector from the point x of M to a nearby point x + v + O(2 )
on M , where > 0 is infinitesimally small; a smooth map then sends
x + v + O(2 ) to (x) + D(x)(v) + O(2 ). One can make this informal
perspective rigorous by means of nonstandard analysis, but we will not do
so here.
Once one has the notion of a tangent bundle, one can define the notion
of a smooth vector field:
Definition 1.2.10 (Vector fields). A smooth vector field on M is a smooth
map X : M T M which is a right inverse for the projection map :
T M M , thus (by slight abuse of notation) X maps x to (x, X(x)) for some
X(x) Tx M . The space of all smooth vector fields is denoted (T M ). It is
clearly a real vector space. In fact, it is a C (M )-module: given a smooth
vector field X (T M ) and a smooth function f C (M ) (i.e. a smooth
map f : M R), one can define the product f X in the obvious manner:
f X(x) := f (x)X(x), and one easily verifies the axioms for a module.
Given a smooth function f C (M ) and a smooth vector field X
(T M ), we define the directional derivative X f C (M ) of f along X
by the formula
d
X f (x) := f ((t))|t=0
dt
whenever : I M is a continuously differentiable function with (0) = x
and 0 (0) = X(x); one easily verifies that X f is well-defined and is an
element of C (M ).
Remark 1.2.11. One can define X f in a more co-ordinate free manner
as
X f = Df X,
where : T R R is the projection map to the second coordinate of
T R R R; one can also view X f as the Lie derivative of f along X
7Indeed, one can view the tangent operator T and the derivative operator D together as a
single covariant functor from the category of smooth manifolds to itself, although we will not
need to use this perspective here.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

29

(although, in most texts, the latter definition would be circular, because the
Lie derivative is usually defined using the directional derivative).
Remark 1.2.12. If V is an open subset of Rd , a smooth vector field on
V can be identified with a smooth map X : V Rd from V to Rd . If
X : M T M is a smooth vector field on M and : U V is a coordinate
chart of M , then the pushforward X := D X 1 : V T V of X by
is a smooth vector field of V . Thus, in coordinates, one can view vector
fields as maps from open subsets of Rd to Rd . This perspective is convenient
for quick and dirty calculations; for instance, in coordinates, the directional
derivative X f is the same as the familiar directional derivative X f from
several variable calculus. If however one wishes to perform several changes
of variable, then the more intrinsically geometric (and coordinate-free)
perspective outlined above can be more helpful.
There is a fundamental link between smooth vector fields and derivations
of C (M ):
Exercise 1.2.4 (Correspondence between smooth vector fields and derivations). Let M be a smooth manifold.
(i) If X (T M ) is a smooth vector field, show that X : C (M )
C (M ) is a derivation on the (real) algebra C (M ), i.e. a (real)
linear map that obeys the Leibniz rule
(1.7)

X (f g) = f X g + (X f )g
for all f, g C (M ).
(ii) Conversely, if d : C (M ) C (M ) is a derivation on C (M ),
show that there exists a unique smooth vector field X such that
d = X .

We see from the above exercise that smooth vector fields can be interpreted as a purely algebraic construction associated to the real algebra
C (M ), namely as the space of derivations on that vector space. This can
be useful for analysing the algebraic structure of such vector fields. Indeed,
we have the following basic algebraic observation:
Exercise 1.2.5 (Commutator of derivations is a derivation). Let d1 , d2 :
A A be two derivations on an algebra A. Show that the commutator
[d1 , d2 ] := d1 d2 d2 d1 is also a derivation on A.
From the preceding two exercises, we can define the Lie bracket [X, Y ]
of two vector fields X, Y (T M ) by the formula
[X,Y ] := [X , Y ].

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

30

1. Hilberts fifth problem

This gives the space (T M ) of smooth vector fields the structure of an

(infinite-dimensional) Lie algebra:
Definition 1.2.13 (Lie algebra). A (real) Lie algebra is a real vector space
V (possibly infinite dimensional), together with a bilinear map [, ] : V V
V which is anti-symmetric (thus [X, Y ] = [Y, X] for all X, Y V , or
equivalently [X, X] = 0 for all X V ) and obeys the Jacobi identity
(1.8)

[[X, Y ], Z] + [[Y, Z], X] + [[Z, X], Y ] = 0

for all X, Y, Z V .
Exercise 1.2.6. If M is a smooth manifold, show that (T M ) (equipped
with the Lie bracket) is a Lie algebra.
Remark 1.2.14. This is the abstract definition of a Lie algebra. A more
concrete definition would be to let V be a subspace of an algebra of operators,
and to define the Lie bracket as the commutator. The relation between the
two notions of a Lie algebra is explored in Section 2.3.
1.2.3. The Lie algebra of a Lie group. Let G be a (global) Lie group.
By definition, G is then a smooth manifolds, so we can thus define the
tangent bundle T G and smooth vector fields X (T G) as in the preceding
section. In particular, we can define the tangent space T1 G of G at the
identity element 1.
If g G, then the left multiplication operation left
g : x 7 gx is, by definition of a Lie group, a smooth map from G to G. This creates a derivative
map Dleft
g : T G T G from the tangent bundle T G to itself. We say that
a vector field X (T G) is left-invariant if one has (left
g ) X = X for all
left
left
g G, or equivalently if (Dg ) X = X g for all g G.
Exercise 1.2.7. Let G be a (global) Lie group.
(i) Show that for every element x of T1 G there is a unique left-invariant
vector field X (T G) such that X(1) = x.
(ii) Show that the commutator [X, Y ] of two left-invariant vector fields
is again a left-invariant vector field.
From the above exercise, we can identify the tangent space T1 G with
the left-invariant vector fields on T G, and the Lie bracket structure on the
latter then induces a Lie bracket (which we also call [, ]) on T1 G. The vector
space T1 G together with this Lie bracket is then a (finite-dimensional) Lie
algebra, which we call the Lie algebra of the Lie group G, and we write as
g.
Remark 1.2.15. Informally, an element x of the Lie algebra g is associated
with an infinitesimal perturbation 1 + x + O(2 ) of the identity in the Lie

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

31

group G. This intuition can be formalised fairly easily in the case of matrix
Lie groups such as GLn (C); for more abstract Lie groups, one can still
formalise things using nonstandard analysis, but we will not do so here.
Exercise 1.2.8.
(i) Show that the Lie algebra gln (C) of the general linear group GLn (C)
can be identified with the space Mn (C) of n n complex matrices,
with the Lie bracket [A, B] := AB BA.
(ii) Describe the Lie algebra un (C) of the unitary group Un (C).
(iii) Describe the Lie algebra sun (C) of the special unitary group SUn (C).
(iv) Describe the Lie algebra on (R) of the orthogonal On (R).
(v) Describe the Lie algebra son (R) of the special orthogonal SOn (R).

1 R R
(vi) Describe the Lie algebra of the Heisenberg group 0 1 R.
0 0 1
Exercise 1.2.9. Let : G H be a smooth homomorphism between
(global) Lie groups. Show that the derivative map D(1G ) at the identity
element 1G is then a Lie algebra homomorphism from the Lie algebra g
of G to the Lie algebra h of H (thus this map is linear and preserves the
Lie bracket). (From this and the chain rule (1.6), we see that the map
7 D(1G ) creates a covariant functor from the category of Lie groups to
the category of Lie algebras.)
We have seen that every global Lie group gives rise to a Lie algebra.
One can also associate Lie algebras to local Lie groups as follows:
Exercise 1.2.10. Let G be a local Lie group. Let U be a symmetric neighbourhood of the identity in G. (It is not difficult to see that least one
such neighbourhood exists.) Call a vector field X (T U ) left-invariant
left is the leftif, for every g U , one has (left
g ) X(g) = X(g), where g
multiplication map x 7 gx, defined on the open set {x U : gx U }
(where we adopt the convention that gx U is shorthand for g x is
well-defined and lies in U ).
(i) Establish the analogue of Exercise 1.2.7 in this setting. Conclude
that one can give T1 G the structure of a Lie algebra, which is
independent of the choice of U .
(ii) Establish the analogue of Exercise 1.2.9 in this setting.
Remark 1.2.16. In the converse direction, it is also true that every finitedimensional Lie algebra can be associated to either a local or a global Lie
group; this is known as Lies third theorem. However, this theorem is somewhat tricky to prove (particularly if one wants to associate the Lie algebra

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

32

1. Hilberts fifth problem

with a global Lie group), requiring the non-trivial algebraic tool of Ados
theorem (discussed in Section 2.3); see Exercise 1.2.23 below.
1.2.4. The exponential map. The exponential map x 7 exp(x) on the
reals R (or its extension to the complex numbers C) is of course fundamental
to modern analysis. It can be defined in a variety of ways, such as the
following:
(i) exp : R R is the differentiable map obeying the ODE
exp(x) and the initial condition exp(0) = 1.

d
dx

exp(x) =

(ii) exp : R R is the differentiable map obeying the homomorphism property exp(x+y) = exp(x) exp(y) and the initial condition
d
dx exp(x)|x=0 = 1.
(iii) exp : R R is the limit of the functions x 7 (1 + nx )n as n .
P
xn
(iv) exp : R R is the limit of the infinite series x 7
n=0 n! .
We will need to generalise this map to arbitrary Lie algebras and Lie
groups. In the case of matrix Lie groups (and matrix Lie algebras), one can
use the matrix exponential, which can be defined efficiently by modifying
definition (iv) above, and which was already discussed in Section 1.1. It
is however difficult to use this definition for abstract Lie algebras and Lie
groups. The definition based on (ii) will ultimately be the best one to use
for the purposes of this text, but for foundational purposes (i) or (iii) is
initially easier to work with. In most of the foundational literature on Lie
groups and Lie algebras, one uses (i), in which case the existence and basic
properties of the exponential map can be provided by the Picard existence
theorem from the theory of ordinary differential equations. However, we
will use (iii), because it relies less heavily on the smooth structure of the Lie
group, and will therefore be more aligned with the spirit of Hilberts fifth
problem (which seeks to minimise the reliance of smoothness hypotheses
whenever possible). Actually, for minor technical reasons it is slightly more
n
convenient to work with the limit of (1 + 2xn )2 rather than (1 + nx )n .
We turn to the details. It will be convenient to work in local coordinates,
and for applications to Hilberts fifth problem it will be useful to forget
almost all of the smooth structure. We make the following definition:
Definition 1.2.17 (C 1,1 local group). A C 1,1 local group is a local group
V that is an open neighbourhood of the origin 0 in a Euclidean space Rd ,
with group identity 0, and whose group operation obeys the estimate
(1.9)

x y = x + y + O(|x||y|)

for all sufficiently small x, y, where the implied constant in the O() notation
can depend on V but is uniform in x, y.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

33

Example 1.2.18. Let G be a local Lie group of some dimension d, and let
: U V be a smooth coordinate chart that maps a neighbourhood U
of the group identity 1 to a neighbourhood V of the origin 0 in Rd , with
(1) = 0. Then, as explained in Example 1.2.4, V = (V, ) = G U is a
local Lie group with identity 0; in particular, one has
0 x = x 0 = x.
From Taylor expansion (using the smoothness of ) we thus have (1.9) for
sufficiently small x, y. Thus we see that every local Lie group generates a
C 1,1 local group when viewed in coordinates.
Remark 1.2.19. In real analysis, a (locally) C 1,1 function is a function
f : U Rm on a domain U Rn which is continuously differentiable (i.e.
in the regularity class C 1 ), and whose first derivatives f are (locally) Lipschitz (i.e. in the regularity class C 0,1 ) the C 1,1 regularity class is slightly
weaker (i.e. larger) than the class C 2 of twice continuously differentiable
functions, but much stronger than the class C 1 of singly continuously differentiable functions. See [Ta2010, 1.14] for more on these sorts of regularity
classes. The reason for the terminology C 1,1 in the above definition is that
C 1,1 regularity is essentially the minimal regularity for which one has the
Taylor expansion
f (x) = f (x0 ) + f (x0 ) (x x0 ) + O(|x x0 |2 )
for any x0 in the domain of f , and any x sufficiently close to x0 ; note that
the asymptotic (1.9) is of this form.
We now estimate various expressions in a C 1,1 local group.
Exercise 1.2.11. Let V be a C 1,1 local group. Throughout this exercise,
the implied constants in the O() notation can depend on V , but not on
parameters such as x, y, , k, n.
(i) Show that there exists an > 0 such that one has

X
(1.10)
x1 xk = x1 + + xk + O
|xi ||xj |
1i<jk

P
whenever k 1 and x1 , . . . , xk V are such that ki=1 |xi | ,
and the implied constant is uniform in k. Here and in the sequel we
adopt the convention that a statement such as (1.10) is automatically false unless all expressions in that statement are well-defined.
(Hint: induct on k using (1.9). It is best to replace the asymptotic
O() notation by explicit constants C in order to ensure that such

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

34

1. Hilberts fifth problem

constants remain uniform in k.) In particular, one has the crude

estimate
k
X
x1 xk = O(
|xi |)
i=1

under the same hypotheses as above.

(ii) Show that one has
x1 = x + O(|x|2 )
for x sufficiently close to the origin.
(iii) Show that
x y x1 y 1 = O(|x||y|)
for x, y sufficiently close to the origin. (Hint: first show that xy =
yx+O(|x||y|), then express xy as the product of xyx1 y 1
and y x.)
(iv) Show that
x y x1 = y + O(|x||y|)
whenever x, y are sufficiently close to the origin.
(v) Show that
y x1 , x1 y = O(|x y|)
whenever x, y are sufficiently close to the origin.
(vi) Show that there exists an > 0 such that
x 1 x k = y1 yk + O

k
X

!
|xi yi |

i=1

whenever k 1 and x1 , . . . , xk , y1 , . . . , yk are such that

.

Pk

i=1 |xi |,

Pk

j=1 |yi |

(vii) Show that there exists an > 0 such that

1
|n||x y| |xn y n | 2|n||x y|
2
for all n Z and x, y Rd such that |nx|, |ny| , where xn =
x x is the product of n copies of x (assuming of course that
this product is well-defined) for n 0, and xn := (xn )1 .
(viii) Show that there exists an > 0 such that


(xy)n = xn y n + O |n|2 |x||y|
for all n Z and x, y Rd such that |nx|, |ny| . (Hint: do the
case when n is positive first. In that case, express xn (xy)n as
the product of n conjugates of y by various powers of x.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

35

We can now define the exponential map exp : V 0 V on this C 1,1 local
group by defining

2n
1
(1.11)
exp(x) := lim
x
n 2n
for any x in a sufficiently small neighbourhood V 0 of the origin in V .
Exercise 1.2.12. Let V be a local C 1,1 group.
(i) Show that if V 0 is a sufficiently small neighbourhood of the origin
in V , then the limit in (1.11) exists for all x V 0 . (Hint: use the
n
previous exercise to estimate the distance between ( 21n x)2 and
n+1
1
x)2 .) Establish the additional estimate
( 2n+1
exp(x) = x + O(|x|2 ).

(1.12)

(ii) Show that if : I G is a smooth curve with (0) = 1, and 0 (0)

is sufficiently small, then
n

exp( 0 (0)) = lim (1/2n )2 .

n

(iii) Show that for all sufficiently small x, y, one has the bilipschitz property
1
|(exp(x) exp(y)) (x y)| |x y|.
2
0
Conclude in particular that for V sufficiently small, exp is a homeomorphism between V 0 and an open neighbourhood exp(V 0 ) of the
origin. (Hint: To show that exp(V 0 ) contains a neighbourhood of
the origin, use (1.12) and the contraction mapping theorem.)
(iv) Show that
(1.13)

exp(sx) exp(tx) = exp((s + t)x)

for s, t R and x Rd with sx, tx sufficiently small. (Hint: first
handle the case when s, t Z[ 21 ] are dyadic numbers.)

(v) Show that for any sufficiently small x, y Rd , one has

(1.14)

exp(x + y) = lim (exp(x/2n ) exp(y/2n ))2 .

n

Then conclude the stronger estimate

(1.15)

exp(x + y) = lim (exp(x/n) exp(y/n))n .

n

(vi) Show that for any sufficiently small x, y Rd , one has

exp(x + y) = exp(x) exp(y) + O(|x||y|).
(Hint: use the previous part, as well as Exercise 1.2.11(viii).)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

36

1. Hilberts fifth problem

Let us say that a C 1,1 local group is radially homogeneous if one has
sx tx = (s + t)x

(1.16)

whenever s, t R and x Rd are such that sx, tx are sufficiently small.

(In particular, this implies that x1 = x for sufficiently small x.) From
the above exercise, we see that any C 1,1 local group V can be made into
a radially homogeneous C 1,1 local group V 0 by first restricting to an open
neighbourhood exp(V 0 ) of the identity, and then applying the logarithmic
homeomorphism exp1 . Thus:
Corollary 1.2.20. Every C 1,1 local group has a neighbourhood of the identity which is isomorphic (as a topological group) to a radially homogeneous
C 1,1 local group.
Now we study the exponential map on global Lie groups. If G is a
global Lie group, and g is its Lie algebra, we define the exponential map
exp : g G on a global Lie group G by setting
n

exp( 0 (0)) := lim (1/2n )2

n

whenever : I G is a smooth curve with (0) = 1.

Exercise 1.2.13. Let G be a global Lie group.
(i) Show that the exponential map is well-defined. (Hint: First handle the case when 0 (0) is small, using the previous exercise, then
bootstrap to larger values of 0 (0).)
(ii) Show that for all x, y g and s, t R, one has
(1.17)

exp(sx) exp(tx) = exp((s + t)x)

and

(1.18)

exp(x + y) = lim (exp(x/n) exp(y/n))n .

n

(Hint: again, begin with the case when x, y are small.)

(iii) Show that the exponential map is continuous.
(iv) Show that for each x g, the function t 7 exp(tx) is the unique
homomorphism from R to G that is differentiable at t = 0 with
derivative equal to x.
Proposition 1.2.21 (Lies first theorem). Let G be a Lie group. Then the
exponential map is smooth. Furthermore, there is an open neighbourhood U
of the origin in g and an open neighbourhood V of the identity in G such
that the exponential map exp is a diffeomorphism from U to V .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

37

Proof. We begin with the smoothness. From the homomorphism property

we see that


d
x
exp(tx) = left
exp(tx)
dt

for all x g and t R. If x and t are sufficiently small, and one uses a
coordinate chart near the origin, the function f (t, x) := (exp(tx)) then
satisfies an ODE of the form
d
f (t, x) = F (f (t, x), x)
dt
for some smooth function F , with initial condition f (0, x) = 0; thus by the
fundamental theorem of calculus we have
Z t
(1.19)
f (t, x) =
F (f (t0 , x), x) dt0 .
0

Now let k 0. An application of the contraction mapping theorem (in

k
the function space L
t Cx localised to small region of spacetime) then shows
k
that f lies in L
t Cx for small enough t, x, and by further iteration of the
integral equation we then conclude that f (t, x) is k times continuously differentiable for small enough t, x. By (1.13) we then conclude that exp is
smooth everywhere.
Since
d
exp(tx)|t=0 = x
dt
we see that the derivative of the exponential map at the origin is the identity
map on g. The second claim of the proposition thus follows from the inverse
function theorem.

In view of this proposition, we see that given a vector space basis
X1 , . . . , Xd for the Lie algebra g, we may obtain a smooth coordinate chart
: U V for some neighbourhood U of the identity and neighbourhood V
of the origin in Rd by defining

(exp(t
1 X1 + + td Xd )) := (t1 , . . . , td )
for sufficiently small t1 , . . . , td R. These are known as exponential coordinates of the first kind. Although we will not use them much here,
we also note that there are exponential coordinates of the second kind, in
which the expression exp(t1 X1 + + td Xd ) is replaced by the slight variant
exp(t1 X1 ) . . . exp(td Xd ).
Using exponential coordinates of the first kind, we see that we may identify a local piece U of the Lie group G with the radially homogeneous C 1,1
local group V . In the next section, we will analyse such radially homogeneous C 1,1 groups further. For now, let us record some easy consequences of
the existence of exponential coordinates. Define a one-parameter subgroup

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

38

1. Hilberts fifth problem

of a topological group G to be a continuous homomorphism : R G from

R to G.
Exercise 1.2.14 (Classification of one-parameter subgroups). Let G be a
Lie group. For any X g, show that the map t 7 exp(tX) is a oneparameter subgroup. Conversely, if : R G is a one-parameter subgroup,
there exists a unique X g such that (t) = exp(tX) for all t R. (Hint:
mimic the proof of Proposition 1.1.1.)
Proposition 1.2.22 (Weak regularity implies strong regularity). Let G, H
be global Lie groups, and let : G H be a continuous homomorphism.
Then is smooth.
Proof. Since is a continuous homomorphism, it maps one-parameter subgroups of G to one-parameter subgroups of H. Thus, for every X g, there
exists a unique element L(X) h such that
(exp(tX)) = exp(tL(X))
for all t R. In particular, we see that L is homogeneous: L(sX) = sL(X)
for all X g and s R. Next, we observe using (1.14) and the fact that
is a continuous homomorphism that for any X, Y g and t R, one has


n
(exp(t(X + Y ))) = lim (exp(tX/2n ) exp(tY /2n ))2
n

= lim ((exp(tX/2n ))(exp(tY /2n )))2

n

= lim (exp(tL(X)/2n ) exp(tL(Y )/2n ))2

= exp(t(L(X) + L(Y )))

and thus L is additive:
L(X + Y ) = L(X) + L(Y ).
We conclude that L is a linear transformation from the finite-dimensional
vector space g to the finite-dimensional vector space h. In particular, L is
smooth. On the other hand, we have
(exp(X)) = exp(L(X)).
Since exp : g G and exp : h H are diffeomorphisms near the origin,
we conclude that is smooth in a neighbourhood of the identity. Using
the homomorphism property (and the fact that the group operations are
smooth for both G and H) we conclude that is smooth everywhere, as
required.

This fact has a pleasant corollary:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

39

Corollary 1.2.23 (Uniqueness of Lie structure). Any (global) topological

group can be made into a Lie group in at most one manner. More precisely,
given a topological group G, there is at most one smooth structure one can
place on G that makes the group operations smooth.
Proof. Suppose for sake of contradiction that one could find two different
smooth structures on G that make the group operations smooth, leading
to two different Lie groups G0 , G00 based on G. The identity map from G0
to G00 is a continuous homomorphism, and hence smooth by the preceding
proposition; similarly for the inverse map from G00 to G0 . This implies that
the smooth structures coincide, and the claim follows.

Note that a general high-dimensional topological manifold may have
more than one smooth structure, which may even be non-diffeomorphic to
each other (as the example of exotic spheres [Mi1956] demonstrates), so
this corollary is not entirely vacuous.
Exercise 1.2.15. Let G be a connected (global) Lie group, let H be another
(global) Lie group, and let : G H be a continuous homomorphism
(which is thus smooth by Proposition 1.2.22). Show that is uniquely
determined by the derivative map D(1) : g h. In other words, if 0 :
G H is another continuous homomorphism with D(1) = D0 (1), then
= 0 . (Hint: first prove this in a small neighbourhood of the origin.
What group does this neighbourhood generate?) What happens if G is not
connected?
Exercise 1.2.16 (Weak regularity implies strong regularity, local version).
Let G, H be local Lie groups, and let : G H be a continuous homomorphism. Show that is smooth in a neighbourhood of the identity in
G.
Now we can establish the final stage, at least, of the program outlined
in Figure 1:
Exercise 1.2.17 (Local Lie implies Lie). Let G be a global topological
group. Suppose that there is an open neighbourhood U of the identity
such that the local group G U can be given the structure of a local Lie
group. Show that G can be given the structure of a global Lie group. (Hint:
We already have at least one coordinate chart on G; translate it around
to create an atlas of such charts. To show compatibility of the charts and
global smoothness of the group, one needs to show that the conjugation
maps x 7 gxg 1 are smooth near the origin for any g G. To prove this,
use Exercise 1.2.16.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

40

1. Hilberts fifth problem

1.2.5. The Baker-Campbell-Hausdorff formula. We now study radially homogeneous C 1,1 local groups in more detail, in particular filling in
some of the last few steps in the program in Figure 1. We will show
Theorem 1.2.24 (Baker-Campbell-Hausdorff formula, qualitative version).
Let V Rd be a radially homogeneous C 1,1 local group. Then the group
operation is real analytic near the origin. In particular, after restricting
V to a sufficiently small neighbourhood of the origin, one obtains a local Lie
group.
We will in fact give a more precise formula for , known as the BakerCampbell-Haudorff-Dynkin formula, in the course of proving Theorem 1.2.24.
This formula is usually proven just for Lie groups, but it turns out that
the proof of the formula extends without much difficulty to the C 1,1 local
group setting (the main difference being that continuous operations, such
as Riemann integrals, have to be replaced by discrete counterparts, such as
Riemann sums).
Remark 1.2.25. In the case where V comes from viewing a general linear
group GLn (C) in local exponential coordinates, the group operation is
given by x y = log(exp(x) exp(y)) for sufficiently small x, y Mn (C).
Thus, a corollary of Theorem 1.2.24 is that this map is real analytic.
We begin the proof of Theorem 1.2.24. Throughout this section, V Rd
is a fixed radially homogeneous C 1,1 local group. We will need some variants
of the basic bound (1.9).
Exercise 1.2.18 (Lipschitz bounds). If x, y, z V are sufficiently small,
establish the bounds
(1.20)

x y = x + y + O(|x + y||y|)

(1.21)

x y = x + y + O(|x + y||x|)

(1.22)

x y = x z + O(|y z|)

and
(1.23)

y x = z x + O(|y z|).

(Hint: to prove (1.20), start with the identity (x y) (y) = x.)

Now we exploit the radial homogeneity to describe the conjugation operation y 7 x y (x) as a linear map:
Lemma 1.2.26 (Adjoint representation). For all x sufficiently close to the
origin, there exists a linear transformation Adx : Rd Rd such that x y
(x) = Adx (y) for all y sufficiently close to the origin.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

41

Remark 1.2.27. Using the matrix example from Remark 1.2.25, we are
asserting here that
exp(x) exp(y) exp(x) = exp(Adx (y))
for some linear transform Adx (y) of y, and all sufficiently small x, y. Indeed,
using the basic matrix identity exp(AxA1 ) = A exp(x)A1 for invertible A
(coming from the fact that the conjugation map x 7 AxA1 is a continuous
ring homomorphism) we see that we may take Ad(x) = exp(x)y exp(x)
here.
Proof. Fix x. The map y 7 x y (x) is continuous near the origin, so
it will suffice to establish additivity, in the sense that
x (y + z) (x) = (x y (x)) + (x z (x))
for y, z sufficiently close to the origin.
Let n be a large natural number. Then from (1.16) we have


1
1 n
(y + z) =
y+ z
.
n
n
Conjugating this by x, we see that



n
1
1
x (y + z) (x) = x
y + z (x)
n
n




1
1
y + z (x) .
=n x
n
n
But from (1.9) we have
1
1
1
1
y+ z = y z+O
n
n
n
n

1
n2

and thus (by Exercise 1.2.18)



 
1
1
1
1
1
x
y + z (x) = x y z (x) + O
.
n
n
n
n
n2
But if we split x n1 y n1 z(x) as the product of x n1 y(x) and x n1 z(x)
and use (1.9), we have
 
1
1
1
1
1
x y z (x) = x y (x) + x z (x) + O
.
n
n
n
n
n2
Putting all this together we see that

 
1
1
1
x (y + z) (x) = n x y (x) + x z (x) + O
n
n
n2
 
1
= x y (x) + x z (x) + O
;
n

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

42

1. Hilberts fifth problem

sending n we obtain the claim.

From (1.9) we see that

k Adx Ikop = O(|x|)
for x sufficiently small. Also from the associativity property we see that
(1.24)

Adxy = Adx Ady

for all x, y sufficiently small. Combining these two properties (and using
(1.20)) we conclude in particular that
(1.25)

k Adx Ady kop = O(|x y|)

for x, y sufficiently small. Thus we see that Ad is a (locally) continuous

linear representation. In particular, t 7 Adtx is a (locally) continuous
homomorphism into a linear group, and so (by Proposition 1.1.1) we have
the Hadamard lemma
Adx = exp(adx )
for all sufficiently small x, where adx : Rd Rd is the linear transformation
d
adx =
Adtx |t=0 .
dt
From (1.24), (1.25), (1.9) we see that
Adtx Adty = Adt(x+y) +O(|t|2 )
for x, y, t sufficiently small, and so by the product rule we have
adx+y = adx + ady .
Also we clearly have adtx = t adx for x, t small. Thus we see that adx is
linear in x, and so we have
(1.26)

adx y = [x, y]

for some bilinear form [, ] : Rd Rd .

One can show that this bilinear form in fact defines a Lie bracket (i.e. it
is anti-symmetric and obeys the Jacobi identity), but for now, all we need
is that it is manifestly real analytic (since all bilinear forms are polynomial
and thus analytic). In particular adx and Adx depend analytically on x.
We now give an important approximation to x y in the case when y is
small:
Lemma 1.2.28. For x, y sufficiently small, we have
x y = x + F (Adx )y + O(|y|2 )
where
F (z) :=

z log z
.
z1

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

43

Proof. If we write z := x y x, then z = O(|y|) (by (1.9)) and

(x) (x + z) = y.
We will shortly establish the approximation
(x) (x + z) =

(1.27)

1 exp( adx )
z + O(|z|2 );
adx

inverting
1 exp( adx )
Adx 1
=
adx
Adx log Adx
we obtain the claim.
It remains to verify (1.27). Let n be a large natural number. We can
expand the left-hand side of (1.27) as a telescoping series
 
 

n1
X  j + 1  j + 1
j+1
j
j
j
(1.28)

x
x+
z x
x+ z .
n
n
n
n
n
n
j=0

Using (1.16), the first summand can be expanded as


 


j
x x
z
j
j
x

x+ z .
n
n
n n
n
n




From (1.20) one has nx nx + nz = nz + O( n|z|2 ), so by (1.22), (1.23) we
can write the preceding expression as



 

z
j
j
|z|
j
x+ z +O
x
n
n
n
n
n2
which by definition of Ad can be rewritten as

 

 

z
j
j
j
|z|
(1.29)
Ad j x
x
x+ z +O
.
n
n
n
n
n
n2
From (1.20) one has
 


j
j
j
x
x + z = O(|z|)
n
n
n
while from (1.25) one has Ad j x nz = O(|z|/n), hence from (1.9) we can
n
rewrite (1.29) as

 

 2
 
z
j
j
j
|z|
|z|
Ad j x + x
x+ z +O
+O
.
n
n
n
n
n
n
n2
Inserting this back into (1.28), we can thus write the left-hand side of (1.27)
as

 
n1
X
z
|z|
2

Ad j x + O(|z| ) + O
.
n
n
n
j=0

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

44

1. Hilberts fifth problem



Writing Ad j x = exp nj adx , and then letting n , we conclude
n
(from the convergence of the Riemann sum to the Riemann integral) that
Z 1
exp(t adx )z dt + O(|z|2 )
(x) (x + z) =
0

and the claim follows.

Remark 1.2.29. In the matrix case, the key computation is to show that
exp(x) exp(x + z) = 1 +

1 exp( adx )
z + O(|z|2 ).
adx

To see this, we can use the fundamental theorem of calculus to write the
left-hand side as
Z 1
d
1+
(exp(tx) exp(t(x + z))) dt.
0 dt
d
d
Since dt
exp(tx) = exp(tx)(x) and dt
exp(t(x + z)) = (x + z) exp(t(x +
z)), we can rewrite this as
Z 1
1+
exp(tx)z exp(t(x + z)) dt.
0

Since exp(t(x + z)) = exp(tx) + O(|z|), this becomes

Z 1
1+
exp(tx)z exp(tx) dt + O(|z|2 );
0

since exp(tx)z exp(tx) = exp(t adx )z, we obtain the desired claim.
We can integrate the above formula to obtain an exact formula for :
Corollary 1.2.30 (Baker-Campbell-Hausdorff-Dynkin formula). For x, y
sufficiently small, one has
Z 1
xy =x+
F (Adx Adty )y dt.
0

The right-hand side is clearly real analytic in x and y, and Theorem

1.2.24 follows.
Proof. Let n be a large natural number. We can express x y as the
telescoping sum


 
n1
X
j+1
j
x+
x
y x
y .
n
n
j=0

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

45

From (1.16) followed by Lemma 1.2.28 and (1.24), one has



 
j+1
j
y
x
y =x
y
n
n
n
 
 

y
1
j
.
y + F Adx Ad j y
+O
=x
n
n
n
n2
We conclude that
 
n1

1X 
1
xy =x+
.
F Adx Ad j y y + O
n
n
n
j=0

Sending n , so that the Riemann sum converges to a Riemann integral,

we obtain the claim.

Remark 1.2.31. It is not immediately obvious from this formula alone
why should be associative. A derivation of associativity from the BakerCampbell-Hausdorff-Dynkin formula is given in Section 2.4.
Exercise 1.2.19. Use the Taylor-type expansion
1/z 1 (1/z 1)2 (1/z 1)3
+

+ ...
2
3
4
to obtain the explicit expansion
F (z) = 1

xy =x+

X
(1)m
n=0

n+1
(ady )r1 (adx )s1 . . . (ady )rn (adx )sn
y
r1 !s1 ! . . . rn !sn !(r1 + + rn + 1)

X
ri ,si 0

(ri ,si )6=(0,0)

where m := n + r1 + + rn + s1 + + sn + 1, and show that the series

is absolutely convergent for x, y small enough. Invert this to obtain the
alternate expansion
xy =y+

X
(1)n
n=0

n+1

X
ri ,si 0

(adx )r1 (ady )s1 . . . (adx )rn (ady )sn

x.
r1 !s1 ! . . . rn !sn !(r1 + + rn + 1)

(ri ,si )6=(0,0)

Exercise 1.2.20. Let V be a radially homogeneous C 1,1 local group. By

Theorem 1.2.24, an open neighbourhood of the origin in V has the structure
of a local Lie group, and thus by Exercise 1.2.10 is associated to a Lie
algebra. Show that this Lie algebra is isomorphic to Rd and the Lie bracket
[, ] is given by (1.24). Note that this establishes a posteriori the fact that

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

46

1. Hilberts fifth problem

the bracket [, ] occurring in (1.24) is anti-symmetric and obeys the Jacobi

identity.
We now record some consequences of the Baker-Campbell-Hausdorff formula.
Exercise 1.2.21 (Lie groups are analytic). Let G be a global Lie group.
Show that G is a real analytic manifold (i.e. one can find an atlas of smooth
coordinate charts whose transition maps are all real analytic), and that the
group operations are also real analytic (i.e. they are real analytic when
viewed in the above-mentioned coordinate charts). Furthermore, show that
any continuous homomorphism between Lie groups is also real analytic.
Exercise 1.2.22 (Lies second theorem). Let G, H be global Lie groups, and
let : g h be a Lie algebra homomorphism. Show that there exists an
open neighbourhood U of the identity in G and a homomorphism : U H
from the local Lie group G U to H such that D(1) = . If G is connected
and simply connected, show that one can take U to be all of G.
Exercise 1.2.23 (Lies third theorem). Ados theorem asserts that every
finite-dimensional Lie algebra is isomorphic to a subalgebra of gln (R) for
some n. This (somewhat difficult) theorem and its proof is discussed in Section 2.3. Assuming Ados theorem as a black box, conclude the following
claims:
(i) (Lies third theorem, local version) Every finite-dimensional Lie
algebra is isomorphic to the Lie algebra of some local Lie group.
(ii) Every local or global Lie group has a neighbourhood of the identity
that is isomorphic to a local linear Lie group (i.e. a local Lie group
contained in GLn (R) or GLn (C) for some n).
(iii) (Lies third theorem, global version) Every finite-dimensional Lie
algebra g is isomorphic to the Lie algebra of some global Lie group.
(Hint: from (i) and (ii), one may identify g with the Lie algebra
of a local linear Lie group. Now consider the space of all smooth
curves in the ambient linear group that are everywhere tangent
to this local linear Lie group modulo homotopy, and use this to
build the global Lie group.)
(iv) (Lies third theorem, simply connected version) Every finite-dimensional
Lie algebra g is isomorphic to the Lie algebra of some global connected, simply connected Lie group. Furthermore, this Lie group
is unique up to isomorphism.
(v) Show that every local Lie group G has a neighbourhood of the
identity that is isomorphic to a neighbourhood of the identity of a

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.2. Lie theory and Baker-Campbell-Hausdorff

47

global connected, simply connected Lie group. Furthermore, this

Lie group is unique up to isomorphism.
Remark 1.2.32. One does not need the full strength of Ados theorem to
establish conclusion (i) of the above exercise. Indeed, it suffices to show that
the operation defined in Exercise 1.2.19 is associative near the origin. To
do this, it suffices to verify associativity in the sense of formal power series;
and then by abstract nonsense one can lift up to the free Lie algebra on d
generators, and then down to the free nilpotent Lie algebra on d generators
and of some arbitrary finite step s, which one can verify to be a finite
dimensional Lie algebra. Applying Ados theorem for the special case of
nilpotent Lie algebras (which is easier to establish than the general case of
Ados theorem, as discussed in Section 2.3), one can identify this nilpotent
Lie algebra with a subalgebra of gn (R) for some n, and then one can argue
as in the above exercise to conclude. See also Section 2.4 for an alternate
way to establish associativity of . However, I do not know how to establish
conclusions (ii), (iii) or (iv) without using Ados theorem in full generality
(and (ii) is in fact equivalent to this theorem, at least in characteristic 0).
Remark 1.2.33. Lies three theorems can be interpreted as establishing
an equivalence between three different categories: the category of finitedimensional Lie algebras; the category of local Lie groups (or more precisely,
the category of local Lie group germs, formed by identifying local Lie groups
that are identical near the origin); and the category of global connected,
simply connected Lie groups. See Section 2.5 for further discussion.
The fact that we were able to establish the Baker-Campbell-Hausdorff
formula at the C 1,1 regularity level will be useful for the purposes of proving
results related to Hilberts fifth problem. In particular, we have the following
criterion for a group to be Lie (very much in accordance with the rigidity
principle from the introduction):
Lemma 1.2.34 (Criterion for Lie structure). Let G be a topological group.
Then G is Lie if and only if there is a neighbourhood of the identity in G
which is isomorphic (as a topological group) to a C 1,1 local group.
This gives the last three steps of the program in Figure 1.
Proof. The only if direction is trivial. For the if direction, combine
Corollary 1.2.20 with Theorem 1.2.24 and Exercise 1.2.17.

Remark 1.2.35. Informally, Lemma 1.2.34 asserts that C 1,1 regularity can
automatically be upgraded to smooth (C ) or even real analytic (C ) regularity for topological groups. In contrast, note that a locally Euclidean
group has neighbourhoods of the identity that are isomorphic to a C 0 local

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

48

1. Hilberts fifth problem

group (which is the same concept as a C 1,1 local group, but without the
asymptotic (1.9)). Thus we have reduced Hilberts fifth problem to the task
of boosting C 0 regularity to C 1,1 regularity, rather than that of boosting C 0
regularity to C regularity.
Exercise 1.2.24. Let G be a Lie group with Lie algebra g. For any X, Y
g, show that
2

exp([X, Y ]) = lim (exp(X/n) exp(Y /n) exp(X/n) exp(Y /n))n .

n

1.3. Building Lie structure from representations and metrics

Hilberts fifth problem concerns the minimal hypotheses one needs to place
on a topological group G to ensure that it is actually a Lie group. In Section
1.2, we saw that one could reduce the regularity hypothesis imposed on G
to a C 1,1 condition, namely that there was an open neighbourhood of G
that was isomorphic (as a local group) to an open subset V of a Euclidean
space Rd with identity element 0, and with group operation obeying the
asymptotic
x y = x + y + O(|x||y|)
for sufficiently small x, y. We will call such local groups (V, ) C 1,1 local
groups.
We now reduce the regularity hypothesis further, to one in which there
is no explicit Euclidean space that is initially attached to G; this will flesh
out another two steps of the diagram in Figure 1. Of course, Lie groups
are still locally Euclidean, so if the hypotheses on G do not involve any
explicit Euclidean spaces, then one must somehow build such spaces from
other structures. One way to do so is to exploit an ambient space with
Euclidean or Lie structure that G is embedded or immersed in. A trivial
example of this is provided by the following basic fact from linear algebra:
Lemma 1.3.1. If V is a finite-dimensional vector space (i.e. it is isomorphic to Rd for some d), and W is a linear subspace of V , then W is also a
finite-dimensional vector space.
We will establish a non-linear version of this statement, known as Cartans theorem. Recall that a subset S of a d-dimensional smooth manifold M
is a d0 -dimensional smooth (embedded) submanifold of M for some 0 d0 d
if for every point x S there is a smooth coordinate chart : U V of a
0
neighbourhood U of x in M that maps x to 0, such that (U S) = V Rd ,
0
0
0
where we identify Rd Rd {0}dd with a subspace of Rd . Informally,
0
S locally sits inside M the same way that Rd sits inside Rd .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

49

Theorem 1.3.2 (Cartans theorem). If H is a (topologically) closed subgroup of a Lie group G, then H is a smooth submanifold of G, and is thus
also a Lie group.
Note that the hypothesis that H is closed is essential; for instance, the
rationals Q are a subgroup of the (additive) group of reals R, but the former
is not a Lie group even though the latter is.
Exercise 1.3.1. Let H be a subgroup of a locally compact group G. Show
that H is closed in G if and only if it is locally compact.
A variant of the above results is provided by using (faithful) representations instead of embeddings. Again, the linear version is trivial:
Lemma 1.3.3. If V is a finite-dimensional vector space, and W is another
vector space with an injective linear transformation : W V from W to
V , then W is also a finite-dimensional vector space.
Here is the non-linear version:
Theorem 1.3.4 (von Neumanns theorem). If G is a Lie group, and H is a
locally compact group with an injective continuous homomorphism : H
G, then H also has the structure of a Lie group.
Actually, it will suffice for the homomorphism to be locally injective
rather than injective; related to this, von Neumanns theorem localises to
the case when H is a local group rather a group. The requirement that H be
locally compact is necessary, for much the same reason that the requirement
that H be closed was necessary in Cartans theorem.
Example 1.3.5. Let G = (R/Z)2 be the two-dimensional torus, let H = R,
and let : H G be the map (x) := (x, x), where R is a fixed real
number. Then is a continuous homomorphism which is locally injective,
and is even globally injective if is irrational, and so Theorem 1.3.4 is
consistent with the fact that H is a Lie group. On the other hand, note that
when is irrational, then (H) is not closed; and so Theorem 1.3.4 does not
follow immediately from Theorem 1.3.2 in this case. (We will see, though,
that Theorem 1.3.4 follows from a local version of Theorem 1.3.2.)
As a corollary of Theorem 1.3.4, we observe that any locally compact
Hausdorff group H with a faithful linear representation, i.e. a continuous
injective homomorphism from H into a linear group such as GLn (R) or
GLn (C), is necessarily a Lie group. This suggests a representation-theoretic
approach to Hilberts fifth problem. While this approach does not seem to
readily solve the entire problem, it can be used to establish a number of
important special cases with a well-understood representation theory, such

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

50

1. Hilberts fifth problem

as the compact case or the abelian case (for which the requisite representation theory is given by the Peter-Weyl theorem and Pontryagin duality
respectively). We will discuss these cases further in later sections.
In all of these cases, one is not really building up Euclidean or Lie
structure completely from scratch, because there is already a Euclidean or
Lie structure present in another object in the hypotheses. Now we turn to
results that can create such structure assuming only what is ostensibly a
weaker amount of structure. In the linear case, one example of this is is the
following classical result in the theory of topological vector spaces.
Theorem 1.3.6. Let V be a locally compact Hausdorff topological vector
space. Then V is isomorphic (as a topological vector space) to Rd for some
finite d.
Remark 1.3.7. The Banach-Alaoglu theorem asserts that in a normed vector space V , the closed unit ball in the dual space V is always compact in the
weak-* topology. Of course, this dual space V may be infinite-dimensional.
This however does not contradict the above theorem, because the closed
unit ball is not a neighbourhood of the origin in the weak-* topology (it is
only a neighbourhood with respect to the strong topology).
The full non-linear analogue of this theorem would be the GleasonYamabe theorem, which we are not yet ready to prove in this section. However, by using methods similar to that used to prove Cartans theorem and
von Neumanns theorem, one can obtain a partial non-linear analogue which
requires an additional hypothesis of a special type of metric, which we will
call a Gleason metric:
Definition 1.3.8. Let G be a topological group. A Gleason metric on G is
a left-invariant metric d : G G R+ which generates the topology on G
and obeys the following properties for some constant C > 0, writing kgk for
d(g, id):
(1) (Escape property) If g G and n 1 is such that nkgk
kg n k C1 nkgk.

1
C,

then

(2) (Commutator estimate) If g, h G are such that kgk, khk

then
(1.30)

1
C,

k[g, h]k Ckgkkhk,

recalling that [g, h] := g 1 h1 gh is the commutator of g and h.

Exercise 1.3.2. Let G be a topological group that contains a neighbourhood of the identity isomorphic to a C 1,1 local group. Show that G admits
at least one Gleason metric.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

51

Theorem 1.3.9 (Building Lie structure from Gleason metrics). Let G be a

locally compact group that has a Gleason metric. Then G is isomorphic to
a Lie group.
We will rely on Theorem 1.3.9 (which represents the last five steps of
Figure 1) to solve Hilberts fifth problem; this theorem reduces the task of
establishing Lie structure on a locally compact group to that of building a
metric with suitable properties. Thus, much of the remainder of the solution
of Hilberts fifth problem will now be focused on the problem of how to
construct good metrics on a locally compact group.
In all of the above results, a key idea is to use one-parameter subgroups to
convert from the nonlinear setting to the linear setting. Recall from Section
1.2 that in a Lie group G, the one-parameter subgroups are in one-to-one
correspondence with the elements of the Lie algebra g, which is a vector
space. In a general topological group G, the concept of a one-parameter
subgroup (i.e. a continuous homomorphism from R to G) still makes sense;
the main difficulties are then to show that the space of such subgroups
continues to form a vector space, and that the associated exponential map
exp : 7 (1) is still a local homeomorphism near the origin.
Exercise 1.3.3. The purpose of this exercise is to illustrate the perspective
that a topological group can be viewed as a non-linear analogue of a vector
space. Let G, H be locally compact groups. For technical reasons we assume
that G, H are both -compact (i.e. the countable union of compact sets)
and metrisable.
(i) (Open mapping theorem) Show that if : G H is a continuous
homomorphism which is surjective, then it is open (i.e. the image
of open sets is open). (Hint: mimic the proof of the open mapping
theorem for Banach spaces, as discussed for instance in [Ta2010,
1.7]. In particular, take advantage of the Baire category theorem.)
(ii) (Closed graph theorem) Show that if a homomorphism : G H
is closed (i.e. its graph {(g, (g)) : g G} is a closed subset of
G H), then it is continuous. (Hint: mimic the derivation of
the closed graph theorem from the open mapping theorem in the
Banach space case, as again discussed in [Ta2010, 1.7].)
(iii) Let : G H be a homomorphism, and let : H K be a
continuous injective homomorphism into another Hausdorff topological group K. Show that is continuous if and only if is
continuous.
(iv) Relax the condition of metrisability to that of being Hausdorff.
(Hint: Now one cannot use the Baire category theorem for metric

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

52

1. Hilberts fifth problem

spaces; but there is an analogue of this theorem for locally compact

Hausdorff spaces.)
1.3.1. The theorems of Cartan and von Neumann. We now turn
to the proof of Cartans theorem. As indicated in the introduction, the
fundamental concept here will be that of a one-parameter subgroup:
Definition 1.3.10 (One-parameter subgroups). Let G be a topological
group. A one-parameter subgroup of G is a continuous homomorphism
: R G. The space of all such one-parameter subgroups is denoted
L(G).
Remark 1.3.11. Strictly speaking, the terminology one-parameter subgroup is a misnomer, because it is the image (G) of which is a subgroup of G, rather than itself. Note that we consider reparameterisations
t 7 (t) of a one-parameter subgroup t 7 (t), where is a non-zero real
number, to be distinct from when 6= 1, even though both one-parameter
subgroups have the same image.
We recall Exercise 1.2.14 from the preceding section, which we reformulate here as a lemma:
Lemma 1.3.12 (Classification of one-parameter subgroups). Let G be a Lie
group, with Lie algebra g. Then if X is an element of g, then t 7 exp(tX)
is a one-parameter subgroup; conversely, if is a one-parameter subgroup,
then there is a unique X g such that (t) = exp(tX) for all t R. Thus
we have a canonical one-to-one correspondence between g and L(G).
Now let H be a closed subgroup of a Lie group G. Every one-parameter
subgroup of H is clearly also a one-parameter subgroup of G, which by the
above lemma can be viewed as an element of g:
L(H) L(G) g.
Thus we can think of L(H) as a subset h of g:
h := {X g : exp(tX) H for all t R}.
We claim that h is in fact a linear subspace of g. Indeed, it contains the
zero element of g (which corresponds to the trivial one-parameter subgroup
t 7 1), and from reparameterisation we see that if X h, then X h for all
R. Finally, if X, Y h, then by definition we have exp(tX), exp(tY )
H for all t R. But recall from Exercise 1.2.13(ii) that
n

exp(t(X + Y )) = lim (exp(tX/2n ) exp(tY /2n ))2 .

n

Since H is a group, we see that (exp(tX/2n ) exp(tY /2n ))2 lies in H. Since
H is closed, we conclude that exp(t(X + Y )) H for all t R, which

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

53

implies that X + Y h. Thus h is closed under both addition and scalar

multiplication, and so it is a vector space. (It turns out that h is in fact a
Lie algebra, but we will not need this fact yet.)
The next step is to show that h is large enough to serve as the Lie
algebra of H. To illustrate this type of fact, let us first establish a simple
special case.
Lemma 1.3.13. Suppose that the identity 1 is not an isolated point of H
(i.e. H is not discrete). Then h is non-trivial (i.e. it does not consist solely
of 0).
Proof. As 1 is not isolated, there exists a sequence hn 6= 1 of elements of
H that converge to 1. As exp : g G is a local homeomorphism near the
identity, we may thus find a sequence Xn 6= 0 of elements of g converging to
zero such that exp(Xn ) = hn for all sufficiently large n.
Let us arbitrarily endow the finite-dimensional vector space g with a
norm (it will not matter which norm we select). Then the sequence Xn /kXn k
lies on the unit sphere with respect to this norm, and thus by the HeineBorel theorem (and passing to a subsequence) we may assume that Xn /kXn k
converges to some element of norm 1.
Let t be any positive real number. Then Xn bt/kXn kc converges to t,
and so exp(Xn )bt/kXn kc converges to exp(t). As exp(Xn ) = hn lies in H,
so does exp(Xn )bt/kXn kc ; as H is closed, we conclude that exp(t) H for
all positive t R, and hence for all t R. We conclude that h, and the
claim follows.

Now we establish a stronger version of the above lemma:
Lemma 1.3.14. There exists a neighbourhood U of the identity in H, and
a neighbourhood V of the origin in h, such that exp : V U is a homeomorphism.
Proof. Let V be a neighbourhood of the origin in h such that exp : V
exp(V ) is a homeomorphism (this exists since exp is a local homeomorphism
in a neighbourhood of the origin in g. Clearly exp(V ) lies in H and contains
1. If exp(V ) contains a neighbourhood of the 1 in H then we are done, so
suppose that this is not the case. Then we can find a sequence hn 6 exp(V )
of elements in H that converge to 1. We may write hn = exp(Xn ) for some
Xn 6 V converging to zero in g.
As h is a subspace of the finite-dimensional vector space g, we may write
g = h + k for some vector space k transverse to h (i.e. h k = 0). (We do not
require k to be a Lie algebra.) From the inverse function theorem, the map
(Y, Z) 7 exp(Y ) exp(Z) from h k to G is a local homeomorphism near

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

54

1. Hilberts fifth problem

the identity. Thus we may write exp(Xn ) = exp(Yn ) exp(Zn ) for sufficiently
large n, where Yn h and Zn k both go to zero as n . Since Xn 6 V ,
we see that Zn is non-zero for n sufficiently large.
We arbitrarily place a norm on k. As before, we may pass to a subsequence and assume that Zn /kZn k converges to some limit in the unit
sphere of k; in particular, 6 h.
Since exp(Xn ) and exp(Yn ) both lie in H, exp(Zn ) does also. By arguing
as in the proof of Lemma 1.3.13 we conclude that exp(t) lies in H for all
t R, and so h, yielding the desired contradiction.

From the above lemma we see that H locally agrees with exp(V ) near the
identity, and thus locally agrees with exp(V )h near h for every h H. This
implies that H is a smooth submanifold of G; since it is also a topological
group, it is thus a Lie group. This establishes Cartans theorem.
Remark 1.3.15. Observe a posteriori that h is the Lie algebra of H, and
in particular is closed with respect to Lie brackets. This fact can also be
established directly using Exercise 1.2.24.
There is a local version of Cartans theorem, in which groups are replaced
by local groups:
Theorem 1.3.16 (Local Cartans theorem). If H is a locally compact local
subgroup of a local Lie group G, then there is an open neighbourhood H 0 of
the identity in H that is a smooth submanifold of G, and is thus also a local
Lie group.
The proof of this theorem follows the lines of the global Cartans theorem, with some minor technical changes, and we set this proof out in the
following exercise.
Exercise 1.3.4. Define a local one-parameter subgroup of a local group H
to be a continuous homomorphism : (, ) H from the (additive)
local group (, ) to H. Call two local one-parameter subgroups equivalent
if they agree on a neighbourhood of the origin, and let L(H) be the set
of all equivalence classes of local one-parameter subgroups. Establish the
following claims:
(i) If H is a global group, then there is a canonical one-to-one correspondence that identifies this definition of L(H) with the definition
of L(H) given previously.
(ii) In the situation of Theorem 1.3.16, show that L(H) can be identified with a linear subspace h of g, namely
h := {X g : exp(tX) H for all sufficiently small t}.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

55

(iii) Let the notation and assumptions be as in (ii). For any neighbourhood H 0 of the identity in H, there is a neighbourhood V of the
origin in h such that exp(V ) H 0 .
(iv) Let the notation and assumptions be as in (ii). There exists a
neighbourhood U of the identity in H, and a neighbourhood V of
the origin in h, such that exp : V U is a homeomorphism.
(v) Prove Theorem 1.3.16.
One can then use Theorem 1.3.16 to establish von Neumanns theorem,
as follows. Suppose that H is a locally compact group with an injective
continuous homomorphism : H G into a Lie group G. As H is locally
compact, there is an open neighbourhood U of the origin in H whose closure
U is compact. The map from U to (U ) is a continuous bijection from a
compact set to a Hausdorff set, and is therefore a homeomorphism (since it
maps closed (and hence compact) subsets of U to compact (and hence closed)
subsets of (U )). The set (U ) is then a locally compact local subgroup of
G and thus has a neighbourhood of the identity which is a local Lie group,
by Theorem 1.3.16. Pulling this back by , we see that some neighbourhood
of the identity in H is a local Lie group, and thus H is a global Lie group
by Exercise 1.2.17.
Exercise 1.3.5. State and prove a local version of von Neumanns theorem,
in which G and H are local groups rather than global groups, and the global
injectivity condition is similarly replaced by local injectivity.
1.3.2. Locally compact vector spaces. We will now turn to the study
of topological vector spaces, which we will need to establish Theorem 1.3.9.
We begin by recalling the definition of a topological vector space.
Definition 1.3.17 (Topological vector space). A topological vector space
is a (real) vector space V equipped with a topology that makes the vector
space operations + : V V V and : R V R (jointly) continuous.
(In particular, (V, +) is necessarily a topological group.)
One can also consider complex topological vector spaces, but the theory
for such spaces is almost identical to the real case, and we will only need the
real case for what follows. In the literature, it is often common to restrict
attention to Hausdorff topological vector spaces, although this is not a severe
restriction in practice, as the following exercise shows:
Exercise 1.3.6. Let V be a topological vector space. Show that the closure
W := {0} of the origin is a closed subspace of V , and the quotient space
V /W is a Hausdorff topological vector space. Furthermore, show that a set
is open in V if and only if it is the preimage of an open set in V /W under
the quotient map : V V /W .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

56

1. Hilberts fifth problem

An important class of topological vector spaces are the normed vector

spaces, in which the topology is generated by a norm kk on the vector space.
However, not every topological vector space is generated by a norm. See
[Ta2010, 1.9] for some further discussion.
We emphasise that in order to be a topological vector space, the vector
space operations +, need to be jointly continuous; merely being continuous
in the individual variables is not sufficient to qualify for being a topological
vector space. We illustrate this with some non-examples of topological vector
spaces:
Example 1.3.18. Consider the one-dimensional vector space R with the
cocompact topology (a non-empty set is open iff its complement is compact
in the usual topology). In this topology, the space is a T1 space 8 (though not
Hausdorff), the scalar multiplication map : RR R is jointly continuous
as long as one excludes the scalar zero, and the addition map + : RR R
is continuous in each coordinate (i.e. translations are continuous), but not
jointly continuous; for instance, the set {(x, y) R : x + y 6 [0, 1]} does
not contain a non-trivial Cartesian product of two sets that are open in the
cocompact topology. So this is not a topological vector space. Similarly for
the cocountable or cofinite topologies on R (the latter topology, incidentally,
is the same as the Zariski topology on R).
Example 1.3.19. Consider the topology of R inherited by pulling back the
usual topology on the unit circle R/Z. This pullback topology is not quite
Hausdorff, but the addition map + : R R R is jointly continuous (so
that this gives R the structure of a topological group). On the other hand,
the scalar multiplication map : R R R is not continuous at all. A
slight variant of this topology comes from pulling back the usual topology
on the torus (R/Z)2 under the map x 7 (x, x) for some irrational ; this
restores the Hausdorff property, and addition is still jointly continuous, but
multiplication remains discontinuous.
Example 1.3.20. Consider R with the discrete topology; here, the topology
is Hausdorff, addition is jointly continuous, and every dilation is continuous,
but multiplication is not jointly continuous. If one instead gives R the halfopen topology, then again the topology is Hausdorff and addition is jointly
continuous, but scalar multiplication is only jointly continuous once one
restricts the scalar to be non-negative.
These examples illustrate that a vector space such as R can have many
topologies on it (and many topological group structures), but only one topological vector space structure. More precisely, we have
8A T space is a topological space in which all points are closed.
1

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

57

Theorem 1.3.21. Every finite-dimensional Hausdorff topological vector space

has the usual topology.
Proof. Let V be a finite-dimensional Hausdorff topological vector space,
with topology F. We need to show that every set which is open in the usual
topology, is open in F, and conversely.
Let v1 , . . . , vn be a basis for the finite-dimensional space V . From the
continuity of the vector space operations, we easily verify that the linear
map T : Rn V given by
T (x1 , . . . , xn ) := x1 v1 + + xn vn
is continuous. From this, we see that any set which is open in F, is also
open in the usual topology.
Now we show conversely that every set which is open in the usual topology, is open in F. It suffices to show that there is a bounded open neighbourhood of the origin in F, since one can then translate and dilate this
open neighbourhood to obtain a (sub-)base for the usual topology. (Here,
bounded refers to the usual sense of the term, for instance with respect to
an arbitrarily selected norm on V (note that on a finite-dimensional space,
all norms are equivalent).)
We use T to identify V (as a vector space) with Rn . As T is continuous,
every set which is compact in the usual topology, is compact in F. In
particular, the unit sphere S n1 := {x Rn : kxk = 1} (in, say, the
Euclidean norm kk on Rn ) is compact in F. Using this and the Hausdorff
assumption on F, we can find an open neighbourhood U of the origin in F
which is disjoint from S n1 .
At present, U need not be bounded (note that we are not assuming V
to be locally connected a priori ). However, we can fix this as follows. Using
the joint continuity of the scalar multiplication map, one can find another
open neighbourhood U 0 of the origin and an open interval (, ) around 0
such that the product set (, )U 0 := {tx : t (, ); x U 0 } is contained
in U . In particular, since U avoids the unit sphere S n1 , U 00 must avoid the
region {x Rn : kxk > 1/} and is thus bounded, as required.

We isolate one important consequence of the above theorem:
Corollary 1.3.22. In a Hausdorff topological vector space V , every finitedimensional subspace W is closed.
Proof. It suffices to show that every vector x V \W is in the exterior
of W . But this follows from Theorem 1.3.21 after restricting to the finitedimensional space spanned by W and x.


Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

58

1. Hilberts fifth problem

We can now prove Theorem 1.3.6. Let V be a locally compact Hausdorff

space, thus there exists a compact neighbourhood K of the origin. Then the
dilate 12 K is also a neighbourhood of the origin, and so by compactness K
can be covered by finitely many translates of 12 K, thus
1
KS+ K
2
for some finite set S. If we let W be the finite-dimensional vector space
generated by S, we conclude that
1
K W + K.
2
Iterating this we have
K W + 2n K
for any n 1. On the other hand, if U is a neighbourhood of the origin,
then for every x V we see that 2n x U for sufficiently large n. By
compactness of K (and continuity of the scalar multiplication map at zero),
we conclude that 2n K U for some sufficiently large n, and thus
K W +U
for any neighbourhood U of the origin; thus K is in the closure of W . By
Corollary 1.3.22, we conclude that
K W.
But K is a neighbourhood of the origin, thus for every x V we have
2n x K for all sufficiently large n, and thus x 2n W = W . Thus
V = W , and the claim follows.
Exercise 1.3.7. Establish the Riesz lemma: if V = (V, kk) is a normed
vector space, W is a proper closed subspace of V , and > 0, then there
exists a vector x in V with kxk = 1 and dist(x, V ) 1 . (Hint: pick
an element y of V not in W , and then pick z W that nearly minimises
ky zk. Use these two vectors to construct a suitable x.) Using this lemma
and the Heine-Borel theorem, give an alternate proof of Theorem 1.3.6 in
the case when V is a normed vector space.
1.3.3. From Gleason metrics to Lie groups. Now we prove Theorem
1.3.9. The argument will broadly follow the lines of Cartans theorem, but
we will have to work harder in many stages of the argument in order to
compensate for the lack of an obvious ambient Lie structure in the initial
hypotheses. In particular, the Gleason metric hypothesis will substitute for
the C 1,1 type structure enjoyed by Lie groups, which as we saw in Section
1.2 was needed to obtain good control on the exponential map.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

59

Henceforth, G is a locally compact group with a Gleason metric d (and

an associated norm kgk = d(g, id)). In particular, by the Heine-Borel
theorem, G is complete with this metric.
We use the asymptotic notation X  Y in place of X CY for some
constant C that can vary from line to line (in particular, C need not be
the constant appearing in the definition of a Gleason metric), and write
X Y for X  Y  X. We also let > 0 be a sufficiently small constant
(depending only on the constant in the definition of a Gleason metric) to be
chosen later.
Note that the left-invariant metric properties of d give the symmetry
property
kg 1 k = kgk
and the triangle inequality
n
X
kg1 . . . gn k
kgi k.
i=1

From the commutator estimate (1.30) and the triangle inequality we also
obtain a conjugation estimate
kghg 1 k khk
whenever kgk, khk . Since left-invariance gives
d(g, h) = kg 1 hk
we then conclude an approximate right invariance
d(gk, hk) d(g, h)
whenever kgk, khk, kkk . In a similar spirit, the commutator estimate
(1.30) also gives
d(gh, hg)  kgkkhk

(1.31)
whenever kgk, khk .

This has the following useful consequence, which asserts that the power
maps g 7 g n behave like dilations:
Lemma 1.3.23. If n 1 and kgk, khk /n, then
d(g n hn , (gh)n )  n2 kgkkhk
and
d(g n , hn ) nd(g, h).
Proof. We begin with the first inequality. By the triangle inequality, it
suffices to show that
(1.32)

d((gh)i g ni hni , (gh)i+1 g ni1 hni1 )  nkgkkhk

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

60

1. Hilberts fifth problem

uniformly for all 0 i < n. By left-invariance and approximate rightinvariance, the left-hand side is comparable to
d(g ni1 h, hg ni1 ),
which by (1.31) is bounded above by
 kg ni1 kkhk  nkgkkhk
as required.
Now we prove the second estimate. Write g = hk, then kkk = d(g, h)
2/n. We have
d(hn k n , hn ) = kk n k nkkk
thanks to the escape property (shrinking if necessary). On the other hand,
from the first inequality, we have
d(g n , hn k n )  n2 khkkkk.
If is small enough, the claim now follows from the triangle inequality. 
Remark 1.3.24. Lemma 1.3.23 implies (by a standard covering argument)
that the group G is locally of bounded doubling, though we will not use
this fact here. The bounds above should be compared with the bounds in
Exercise 1.2.11. Indeed, just as the bounds in that exercise were used in the
previous sections to build the exponential map for Lie groups, the bounds
in Lemma 1.3.23 are crucial for controlling the exponential function on the
locally compact group G equipped with the Gleason metric d.
Now we bring in the space L(G) of one-parameter subgroups. We give
this space the compact-open topology, thus the topology is generated by balls
of the form
{ L(G) : sup d((t), 0 (t)) < r}
tI

for 0 L(G), r > 0, and compact I. Actually, using the homomorphism

property, one can use a single compact interval I, such as [1, 1], to generate
the topology if desired, thus making L(G) a metric space.
Given that G is eventually going to be shown to be a Lie group, L(G)
must be isomorphic to a Euclidean space. We now move towards this goal
by establishing various properties of L(G) that Euclidean spaces enjoy.
Lemma 1.3.25. L(G) is locally compact.
Proof. It is easy to see that L(G) is complete. Let 0 L(G). As 0 is
continuous, we can find an interval I = [T, T ] small enough that k0 (t)k
for all t [T, T ]. By the Heine-Borel theorem, it will suffice to show that
the set
B := { L(G) : sup d((t), 0 (t)) < }
t[T,T ]

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

61

is totally bounded. By the Arzel

a-Ascoli theorem, it suffices to show that
the family of functions in B is equicontinuous.
By construction, we have k(t)k 2 whenever |t| T . By the escape
property, this implies (for small enough, of course) that k(t/n)k  /n for
all |t| T and n 1, thus k(t)k  |t|/T whenever |t| T . From the homomorphism property, we conclude that d((t), (t0 ))  |tt0 |/T whenever
|t|, |t0 | T , which gives uniform Lipschitz control and hence equicontinuity
as desired.

We observe for future reference that the proof of the above lemma also
shows that all one-parameter subgroups are locally Lipschitz.
Now we put a vector space structure on L(G), which we define by analogy with the Lie group case, in which each tangent vector X generates a
one-parameter subgroup t 7 exp(tX). From this analogy, the scalar multiplication operation has an obvious definition: if L(G) and c R, we
define c L(G) to be the one-parameter subgroup
c(t) := (ct)

(1.33)

which is easily seen to actually be a one-parameter subgroup.

Now we turn to the addition operation. In the Lie group case, one
can express the one-parameter subgroup t 7 exp(t(X + Y )) in terms of the
one-parameter subgroups t 7 exp(tX), t 7 exp(tY ) by the limiting formula
exp(t(X + Y )) = lim (exp(tX/n) exp(tY /n))n ;
n

cf. Exercise 1.2.13. In view of this, we would like to define the sum +
of two one-parameter subgroups , L(G) by the formula
( + )(t) := lim ((t/n)(t/n))n .
n

Lemma 1.3.26. If , L(G), then + is well-defined and also lies in

L(G).
Proof. To show well-definedness, it suffices to show that for each t, the
sequence ((t/n)(t/n))n is a Cauchy sequence. It suffices to show that
sup d (((t/n)(t/n))n , ((t/nm)(t/nm))nm ) 0
m1

as n . We will in fact prove the slightly stronger claim



0
0
sup sup d ((t/n)(t/n))n , ((t/nm)(t/nm))n m 0.
m1 1n0 n

Observe from continuity of multiplication that to prove this claim for a given
t, it suffices to do so for t/2; thus we may assume without loss of generality
that t is small.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

62

1. Hilberts fifth problem

Let > 0 be a small number to be chosen later. Since , are locally

Lipschitz, we see (if t is sufficiently small depending on ) that
k(t/n)k, k(t/n)k  /n
for all n. From Lemma 1.3.23, we conclude that
d ((t/n)(t/n), ((t/nm)(t/nm)m ))  m2 (/nm)(/nm) = 2 /n2
if m 1 and n is sufficiently large. Another application of Lemma 1.3.23
then gives



0
0
d ((t/n)(t/n))n , (t/nm)(t/nm)n m
 n0 2 /n2  2 /n
if m 1, n is sufficiently large, and 1 n0 n. The claim follows.
The above argument in fact shows that ((t/n)(t/n))n is uniformly
Cauchy for t in a compact interval, and so the pointwise limit + is in
fact a uniform limit of continuous functions and is thus continuous. To prove
that + is a homomorphism, it suffices by density of the rationals to show
that
( + )(at)( + )(bt) = ( + )((a + b)t)
and
( + )(t) = ( + )(t)1
for all t R and all positive integers a, b. To prove the first claim, we
observe that
( + )(at) = lim ((at/n)(at/n))n
n

= lim ((t/n)(t/n))an
n

and similarly for ( + )(bt) and ( + )((a + b)t), whence the claim. To
prove the second claim, we see that
( + )(t)1 = lim ((t/n)(t/n))n
n

= lim ((t/n)(t/n))n ,
n

((t/n)(t/n))n

but
is ((t/n)(t/n))n conjugated by (t/n), which goes
to the identity; and the claim follows.

L(G) also has an obvious zero element, namely the trivial one-parameter
subgroup t 7 id.
Lemma 1.3.27. L(G) is a topological vector space.
Proof. We first show that L(G) is a vector space. It is clear that the zero
element 0 of L(G) is an additive and scalar multiplication identity, and
that scalar multiplication is associative. To show that addition is commutative, we again use the observation that ((t/n)(t/n))n is ((t/n)(t/n))n

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

63

conjugated by an element that goes to the identity. A similar argument

shows that () + () = ( + ), and a change of variables argument
shows that (a) + (a) = a( + ) for all positive integers a, hence for
all rational a, and hence by continuity for all real a. The only remaining
thing to show is that addition is associative, thus if , , L(G), that
(( + ) + )(t) = ( + ( + ))(t) for all t R. By the homomorphism
property, it suffices to show this for all sufficiently small t.
An inspection of the argument used to establish (1.3.26) reveals that
there is a constant > 0 such that
d (( + )(t), ((t/n)(t/n))n )  2 /n
for all small t and all large n, and hence also that
d (( + )(t/n), (t/n)(t/n))  2 /n2
(thanks to Lemma 1.3.23). Similarly we have (after adjusting if necessary)
d ((( + ) + ) (t), (( + )(t/n)(t/n))n )  2 /n.
From Lemma 1.3.23 we have
d ((( + )(t/n)(t/n))n , ((t/n)(t/n)(t/n))n )  2 /n
and thus
d ((( + ) + ) (t), ((t/n)(t/n)(t/n))n )  2 /n.
Similarly for + ( + ). By the triangle inequality we conclude that
d ((( + ) + ) (t), ( + ( + )) (t))  2 /n;
sending t to zero, the claim follows.
Finally, we need to show that the vector space operations are continuous. It is easy to see that scalar multiplication is continuous, as are the
translation operations; the only remaining thing to verify is that addition
is continuous at the origin. Thus, for every  > 0 we need to find a > 0
such that supt[1,1] k( + )(t)k  whenever supt[1,1] k(t)k and
supt[1,1] k(t)k . But if , are as above, then by the escape property (assuming small enough) we conclude that k(t)k, k(t)k  |t|
for t [1, 1], and then from the triangle inequality we conclude that
k( + )(t)k  for t [1, 1], giving the claim.

Exercise 1.3.8. Show that for any L(G), the quantity
kk := lim nk(1/n)k
n

exists and defines a norm on L(G) that generates the topology on L(G).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

64

1. Hilberts fifth problem

As L(G) is both locally compact, metrisable, and a topological vector

space, it must be isomorphic to a finite-dimensional vector space Rn with
the usual topology, thanks to Theorem 1.3.21.
In analogy with the Lie algebra setting, we define the exponential map
exp : L(G) G by setting exp() := (1). Given the topology on L(G), it
is clear that this is a continuous map.
Exercise 1.3.9. Show that the exponential map is locally injective near the
origin. (Hint: from Lemma 1.3.23, obtain the unique square roots property:
if g, h G are sufficiently close to the identity and g 2 = h2 , then g = h.)
We have proved a number of useful things about L(G), but at present we
have not established that L(G) is large in any substantial sense; indeed, at
present, L(G) could be completely trivial even if G was large. In particular,
the image of the exponential map exp could conceivably be quite small.
We now address this issue. As a warmup, we show that L(G) is at least
non-trivial if G is non-discrete (cf. Lemma 1.3.13):
Proposition 1.3.28. Suppose that G is not a discrete group. Then L(G)
is non-trivial.
Of course, the converse is obvious; discrete groups do not admit any
non-trivial one-parameter subgroups.
Proof. As G is not discrete, there is a sequence gn of non-identity elements
of G such that kgn k 0 as n . Writing Nn for the integer part
of /kgn k, then Nn as n , and we conclude from the escape
property that kgnNn k for all n.
We define the approximate one-parameter subgroups n : [1, 1] G
by setting
n (t) := gnbtNn c .
Then we have kn (t)k  |t| + Nn for |t| 1, and we have the approximate
homomorphism property
d(n (t + s), n (t)n (s)) 0
uniformly whenever |t|, |s|, |t + s| 1. As a consequence, n is asymptotically equicontinuous on [1, 1], and so by (a slight generalisation of) the
Arzela-Ascoli theorem, we may pass to a subsequence in which n converges
uniformly to a limit : [1, 1] G, which is a genuine homomorphism that
is genuinely continuous, and is thus can be extended to a one-parameter
subgroup. Also, kn (1)k = kgnNn k for all n, and thus k(1)k ; in
particular, is non-trivial, and the claim follows.


Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.3. Lie structure from representations and metrics

65

We now generalise the above proposition to a more useful result (cf.

Lemma 1.3.14).
Proposition 1.3.29. For any neighbourhood K of the origin in L(G),
exp(K) is a neighbourhood of the identity in G.
Proof. We use an argument9 of Hirschfeld [Hi1990]. By shrinking K if
necessary, we may assume that K is a compact star-shaped neighbourhood,
with exp(K) contained in the ball of radius around the origin. As K is
compact, exp(K) is compact also.
Suppose for contradiction that exp(K) is not a neighbourhood of the
identity, then there is a sequence gn of elements of G\K such that kgn k 0
as n . By the compactness of K, we can find an element hn of K that
minimises the distance d(gn , hn ). If we then write gn = hn kn , then
kkn k = d(gn , hn ) d(gn , id) = kgn k
and hence khn k, kkn k 0 as n .
Let Nn be the integer part of n /kkn k, then Nn as n , and
for all n.

kknNn k

Let n : [1, 1] G be the approximate one-parameter subgroups defined as

n (t) := knbtNn c .
As before, we may pass to a subsequence such that n converges uniformly
to a limit : [1, 1] G, which extends to a one-parameter subgroup
L(G).
In a similar vein, since hn exp(K), we can find n K such that
n (1) = hn , which by the escape property (and the smallness of K implies
that kn (t)k  tkhn k for |t| 1. In particular, n goes to zero in L(G).
We now claim that exp(n + N1n ) is close to gn . Indeed, from Lemma
1.3.23 we see that





1
1
1
d exp n +
, exp(n ) exp


khn k.
Nn
Nn
Nn
Since exp(n ) = hn , we conclude from the triangle inequality and leftinvariance that







1
1
1
d exp n +
, gn 
khn k + d kn , exp

.
Nn
Nn
Nn
But from Lemma 1.3.23 again, one has





1
1
d kn , exp


d knNn , exp() = o(1/Nn )
Nn
Nn
9The author thanks Lou van den Dries and Isaac Goldbring for bringing this argument to
his attention.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

66

1. Hilberts fifth problem

and thus




1
d exp n +
, gn = o(1/Nn ).
Nn
But for n large enough, n + N1n lies in K, and so the distance from gn to
K is o(1/Nn ) = o(d(gn , hn )). But this contradicts the minimality of hn for
n large enough, and the claim follows.

If K is a sufficiently small compact neighbourhood of the identity in
L(G), then exp : K exp(K) is bijective by Lemma 1.3.9; since it is also
continuous, K is compact, and exp(K) is Hausdorff, we conclude that exp :
K exp(K) is a homeomorphism. The local group structure G exp(K) on
exp(K) then pulls back to a local group structure on K.
Exercise 1.3.10. If we identify L(G) with Rd for some d, show that the
exponential map exp : K exp(K) is bilipschitz.
Proposition 1.3.30. K is a radially homogeneous C 1,1 local group (as defined in Definition 1.2.17 and (1.16)), after identifying L(G) with Rd for
some finite d.
Proof. The radial homogeneity is clear from (1.33) and the homomorphism
property, so the main task is to establish the C 1,1 property
x y = x + y + O(|x||y|)
for the local group law on K. By Exercise 1.3.10, this is equivalent to the
assertion that
d((1)(1), ( + )(1))  k(1)kk(1)k
for , sufficiently close to the identity in L(G). By definition of + , it
suffices to show that
d((1)(1), ((1/n)(1/n))n )  k(1)kk(1)k
for all n; but this follows from Lemma 1.3.23 (and the observation, from the
escape property, that k(1/n)k  k(1)k/n and k(1/n)k  k(1)k/n).

Combining this proposition with Lemma 1.2.34, we obtain Theorem
1.3.9.
Exercise 1.3.11. State and prove a version of Theorem 1.3.9 for local
groups. (In order to do this, you must first decide how to define an analogue
of a Gleason metric on a local group.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

67

1.4. Haar measure, the Peter-Weyl theorem, and compact or

abelian groups
In Sections 1.2, 1.3, we have been steadily reducing the amount of regularity
needed on a topological group in order to be able to show that it is in fact
a Lie group, in the spirit of Hilberts fifth problem. Now, we will work
on Hilberts fifth problem from the other end, starting with the minimal
assumption of local compactness on a topological group G, and seeing what
kind of structures one can build using this assumption. (For simplicity we
shall mostly confine our discussion to global groups rather than local groups
for now.) In view of the preceding sections, we would like to see two types
of structures emerge in particular:
(1) representations of G into some more structured group, such as a
matrix group GLn (C); and
(2) metrics on G that capture the escape and commutator structure of
G (i.e. Gleason metrics).
To build either of these structures, a fundamentally useful tool is that
of (left-) Haar measure - a left-invariant Radon measure on G. (One
can of course also consider right-Haar measures; in many cases (such as for
compact or abelian groups), the two concepts are the same, but this is not
always the case.) This concept generalises the concept of Lebesgue measure
on Euclidean spaces Rd , which is of course fundamental in analysis on those
spaces.
Haar measures will help us build useful representations and useful metrics on locally compact groups G. For instance, a Haar measure gives rise
to the regular representation : G U (L2 (G, d)) that maps each element
g G of G to the unitary translation operator (g) : L2 (G, d) L2 (G, d)
on the Hilbert space L2 (G, d) of square-integrable measurable functions on
G with respect to this Haar measure by the formula
(g)f (x) := f (g 1 x).
(The presence of the inverse g 1 is convenient in order to obtain the homomorphism property (gh) = (g) (h) without a reversal in the group multiplication.) In general, this is an infinite-dimensional representation; but
in many cases (and in particular, in the case when G is compact) we can
decompose this representation into a useful collection of finite-dimensional
representations, leading to the Peter-Weyl theorem, which is a fundamental tool for understanding the structure of compact groups. This theorem
is particularly simple in the compact abelian case, where it turns out that
the representations can be decomposed into one-dimensional representations
: G U (C) S 1 , better known as characters, leading to the theory of

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

68

1. Hilberts fifth problem

Fourier analysis on general compact abelian groups. With this and some
additional (largely combinatorial) arguments, we will also be able to obtain
satisfactory structural control on locally compact abelian groups as well.
The link between Haar measure and useful metrics on G is a little more
complicated. Firstly, once one has the regular representation : G
U (L2 (G, d)), and given a suitable test function : G C, one can then
embed G into L2 (G, d) (or into other function spaces on G, such as Cc (G)
or L (G)) by mapping a group element g G to the translate (g) of
in that function space. (This map might not actually be an embedding if
enjoys a non-trivial translation symmetry (g) = , but let us ignore this
possibility for now.) One can then pull the metric structure on the function
space back to a metric on G, for instance defining an L2 (G, d)-based metric
d(g, h) := k (g) (h)kL2 (G,d)
if is square-integrable, or perhaps a Cc (G)-based metric
(1.34)

d(g, h) := k (g) (h)kCc (G)

if is continuous and compactly supported (with kf kCc (G) := supxG |f (x)|

denoting the supremum norm). These metrics tend to have several nice
properties (for instance, they are automatically left-invariant), particularly
if the test function is chosen to be sufficiently smooth. For instance, if we
introduce the differentiation (or more precisely, finite difference) operators
g := 1 (g)
(so that g f (x) = f (x) f (g 1 x)) and use the metric (1.34), then a short
computation (relying on the translation-invariance of the Cc (G) norm) shows
that
d([g, h], id) = kg h h g kCc (G)
for all g, h G. This suggests that commutator estimates, such as those
appearing in Definition 1.3.8, might be available if one can control second
derivatives of ; informally, we would like our test functions to have a
C 1,1 type regularity.
If G was already a Lie group (or something similar, such as a C 1,1 local
group) then it would not be too difficult to concoct such a function by
using local coordinates. But of course the whole point of Hilberts fifth
problem is to do without such regularity hypotheses, and so we need to
build C 1,1 test functions by other means. And here is where the Haar
measure comes in: it provides the fundamental tool of convolution
Z
(x) :=
(xy 1 )(y)d(y)
G

between two suitable functions , : G C, which can be used to build

smoother functions out of rougher ones. For instance:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

69

Exercise 1.4.1. Let , : Rd C be continuous, compactly supported

functions which are Lipschitz continuous. Show that the convolution
using Lebesgue measure on Rd obeys the C 1,1 -type commutator estimate
kg h ( )kCc (Rd ) Ckgkkhk
for all g, h Rd and some finite quantity C depending only on , .
This exercise suggests a strategy to build Gleason metrics by convolving together some Lipschitz test functions and then using the resulting
convolution as a test function to define a metric. This strategy may seem
somewhat circular because one needs a notion of metric in order to define
Lipschitz continuity in the first place, but it turns out that the properties
required on that metric are weaker than those that the Gleason metric will
satisfy, and so one will be able to break the circularity by using a bootstrap
or induction argument.
We will discuss this strategy - which is due to Gleason, and is fundamental to all currently known solutions to Hilberts fifth problem - in later
sections. In this section, we will construct Haar measure on general locally
compact groups, and then establish the Peter-Weyl theorem, which in turn
can be used to obtain a reasonably satisfactory structural classification of
both compact groups and locally compact abelian groups.

1.4.1. Haar measure. For technical reasons, it is convenient to not work

with an absolutely general locally compact group, but to instead restrict
attention to those groups that are both -compact and Hausdorff, in order
to access measure-theoretic tools such as the Fubini-Tonelli theorem and
the Riesz representation theorem without bumping into unwanted technical difficulties. Intuitively, -compact groups are those groups that do not
have enormously large scales - scales are too coarse to be seen by any
compact set. Similarly, Hausdorff groups are those groups that do not have
enormously small scales - scales that are too small to be seen by any
open set. A simple example of a locally compact group that fails to be compact is the real line R = (R, +) with the discrete topology; conversely,
a simple example of a locally compact group that fails to be Hausdorff is
the real line R with the trivial topology.
As the two exercises below show, one can reduce to the -compact Hausdorff case without much difficulty, either by restricting to an open subgroup
to eliminate the largest scales and recover -compactness, or to quotient out
by a compact normal subgroup to eliminate the smallest scales and recover
the Hausdorff property. (This gives the very first step in Figure 1.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

70

1. Hilberts fifth problem

Exercise 1.4.2. Let G be a locally compact group. Show that there exists
an open subgroup G0 which is locally compact and -compact. (Hint: take
the group generated by a compact neighbourhood of the identity.)
Exercise 1.4.3. Let G be a locally compact group. Let H = {id} be the
topological closure of the identity element.
(i) Show that given any open neighbourhood U of a point x in G,
there exists a neighbourhood V of x whose closure lies in U . (Hint:
translate x to the identity and select V so that V 2 U .) In other
words, G is a regular space.
(ii) Show that for any group element g G, that the sets gH and H
are either equal or disjoint.
(iii) Show that H is a compact normal subgroup of G.
(iv) Show that the quotient group G/H (equipped with the quotient
topology) is a locally compact Hausdorff group.
(v) Show that a subset of G is open if and only if it is the preimage of
an open set in G/H.
Now that we have restricted attention to the -compact Hausdorff case,
we can now define the notion of a Haar measure.
Definition 1.4.1 (Radon measure). Let X be a -compact locally compact
Hausdorff topological space. The Borel -algebra B[X] on X is the -algebra
generated by the open subsets of X. A Borel measure is a countably additive
non-negative measure : B[X] [0, +] on the Borel -algebra. A Radon
measure is a Borel measure obeying three additional axioms:
(i) (Local finiteness) One has (K) < for every compact set K.
(ii) (Inner regularity) One has (E) = supKE,K compact (K) for
every Borel measurable set E.
(iii) (Outer regularity) One has (E) = inf U E,U open (U ) for every
Borel measurable set E.
Definition 1.4.2 (Haar measure). Let G = (G, ) be a -compact locally
compact Hausdorff group. A Radon measure is left-invariant (resp. rightinvariant) if one has (gE) = (E) (resp. (Eg) = (E)) for all g G
and Borel measurable sets E. A left-invariant Haar measure is a non-zero
Radon measure which is left-invariant; a right-invariant Haar measure is
defined similarly. A bi-invariant Haar measure is a Haar measure which is
both left-invariant and right-invariant.
Note that we do not consider the zero measure to be a Haar measure.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

71

Example 1.4.3. A large part of the foundations of Lebesgue measure theory

(e.g. most of [Ta2011, 1.2]) can be summed up in the single statement
that Lebesgue measure is a (bi-invariant) Haar measure on Euclidean spaces
Rd = (Rd , +).
Example 1.4.4. If G is a countable discrete group, then counting measure
is a bi-invariant Haar measure.
Example 1.4.5. If is a left-invariant Haar measure on a -compact locally
compact Hausdorff group G, then the reflection
defined by
(E) := (E 1 )
is a right-invariant Haar measure on G, and the scalar multiple is a leftinvariant Haar measure on G for any 0 < < .
Exercise 1.4.4. If is a left-invariant Haar measure on a -compact locally
compact Hausdorff group G, show that (U ) > 0 for any non-empty open
set U .
Let be a left-invariant Haar measure on a -compact locally compact
Hausdorff group. Let Cc (G) be the space of all continuous, compactly supported complex-valued functions f : G C; then f is absolutely integrable
with respect to (thanks to local finiteness), and one has
Z
Z
f (gx) d(x) =
f (x) dx
G

for all g G (thanks to left-invariance). Similarly for right-invariant Haar

measures (but now replacing gx by xg).
The fundamental theorem regarding Haar measures is:
Theorem 1.4.6 (Existence and uniqueness of Haar measure). Let G be a compact locally compact Hausdorff group. Then there exists a left-invariant
Haar measure on G. Furthermore, this measure is unique up to scalars:
if , are two left-invariant Haar measures on G, then = for some
scalar > 0.
Similarly if left-invariant is replaced by right-invariant throughout.
(However, we do not claim that every left-invariant Haar measure is automatically right-invariant, or vice versa.)
This theorem gives half of the second step of Figure 1. (The other half is
the Birkhoff-Kakutani theorem, see Theorem 1.5.2 and Exercise 1.5.4, which
also requires passage to a subquotient.)
To prove Theorem 1.4.6, we will rely on the Riesz representation theorem
(see e.g. [Ta2010, 1.10] for a proof):
Theorem 1.4.7 (Riesz representation theorem). Let X be a -compact locally compact Hausdorff space. Then to every linear functional I : Cc (X)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

72

1. Hilberts fifth problem

R which is non-negative (thus I(f ) 0 whenever

R f 0), one can associate
a unique Radon measure such that I(f ) = X f d for all f R Cc (X).
Conversely, for each Radon measure , the functional I : f 7 X f d is
a non-negative linear functional on Cc (X).
We now establish the uniqueness component of Theorem 1.4.6. We
shall just prove the uniqueness of left-invariant Haar measure, as the rightinvariant case is similar (and also follows from the left-invariant case by
Example 1.4.5). Let , be two left-invariant Haar measures on G. We
need to prove that is a scalar multiple of . From the Riesz representation
theorem, it suffices to show that I is a scalar multiple of I . Equivalently,
it suffices to show that
I (f )I (g) = I (f )I (g)
for all f, g Cc (G).
To show this, the idea is to approximate both f and g by superpositions
of translates of the same function . More precisely, fix f, g Cc (G),
and let > 0. As the functions f and g are continuous and compactly
supported, they are uniformly continuous, in the sense that we can find an
open neighbourhood U of the identity such that |f (xy) f (x)| and
|g(xy) g(x)| for all x G and y U ; we may also assume that the
U are contained in a compact set that is uniform in . By Exercise 1.4.4
and Urysohns lemma, we can then find
R an approximation to the identity
Cc (U ) supported in U such that G (y) d(y) = 1. Since
f (xy) = f (x) + O()
for all y in the support of , we conclude that
Z
f (xy) (y) d(y) = f (x) + O()
G

uniformly in x G; also, the left-hand side has uniformly compact support

in . If we integrate against , we conclude that
Z Z
f (xy) (y) d(y)d(x) = I (f ) + O()
G

where the implied constant in the O() notation can depend on , , f, g but
not on . But by the left-invariance of , the left-hand side is also
Z Z
f (y) (x1 y) d(y)d(x)
G

which by the Fubini-Tonelli theorem is

Z

Z
1
f (y)
(x y) d(x) d(y)
G

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

73

which by the left-invariance of is

Z

Z
1
f (y)
(x ) d(x) d(y)
G

which simplifies to I (f )

G
1
G (x )

Z
I (f ) = I (f )

d(x). We conclude that

(x1 ) d(x) + O()

and similarly
Z
I (g) = I (g)

(x1 ) d(x) + O()

which implies that

I (f )I (g) I (f )I (g) = O().
Sending 0 we obtain the claim.
Exercise 1.4.5. Obtain another proof of uniqueness of Haar measure by
investigating the translation-invariance properties of the Radon-Nikodym
d
derivative d(+)
of with respect to + .
Now we show existence of Haar measure. Again, we restrict attention
to the left-invariant case (using Example 1.4.5 if desired). By the Riesz
representation theorem, it suffices to find a functional I : Cc (G)+ R+
from the space Cc (G)+ of non-negative continuous compactly supported
functions to the non-negative reals obeying the following axioms:
(1) (Homogeneity) I(f ) = I(f ) for all > 0 and f Cc (G)+ .
(2) (Additivity) I(f + g) = I(f ) + I(g) for all f, g Cc (G)+ .
(3) (Left-invariance) I( (x)f ) = I(f ) for all f Cc (G)+ and x G.
(4) (Non-degeneracy) I(f0 ) > 0 for at least one f0 Cc (G)+ .
Here, (x) is the translation operation (x)f (y) := f (x1 y) as discussed in
the introduction.
We will construct this functional by an approximation argument. Specifically, we fix a non-zero f0 Cc (G)+ . We will show that given any finite
number of functions f1 , . . . , fn Cc (G)+ and any > 0, one can find a
functional I = If1 ,...,fn , : Cc (G)+ R+ that obeys the following axioms:
(1) (Homogeneity) I(f ) = I(f ) for all > 0 and f Cc (G)+ .
(2) (Approximate additivity) |I(fi + fj ) I(fi ) I(fj )| for all
1 i, j n.
(3) (Left-invariance) I( (x)f ) = I(f ) for all f Cc (G)+ and x G.
(4) (Uniform bound) For each f Cc (G)+ , we have I(f ) K(f ),
where K(f ) does not depend on f1 , . . . , fn or .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

74

1. Hilberts fifth problem

(5) (Normalisation) I(f0 ) = 1.

Once one has established the existence of these approximately additive
functionals If1 ,...,fn , , one can then construct the genuinely additive functional I (and thus a left-invariant Haar measure) by a number of standard
compactness arguments. For instance:
(1) One can observe (from Tychonoff s theorem) that the space of
all functionals I : Cc (G)+ R+ obeying the uniform bound
+
I(f ) K(f ) is a compact subset of the product space (R+ )Cc (G) ;
in particular, any collection of closed sets in this space obeying the
finite intersection property has non-empty intersection. Applying
this fact to the closed sets Ff1 ,...,fn , of functionals obeying the homogeneity, approximate additivity, left-invariance, uniform bound,
and normalisation axioms for various f1 , . . . , fn , , we conclude that
there is a functional I that lies in all such sets, giving the claim.
(2) If one lets C be the space of all tuples (f1 , . . . , fn , ), one can use the
Hahn-Banach theorem to construct a bounded real linear functional
: ` (C) R that maps the constant sequence 1 to 1. If one then
applies this functional to the If1 ,...,fn , one can obtain a functional
I with the required properties.
(3) One can also adopt a nonstandard analysis approach, taking an
ultralimit of all the If1 ,...,fn , and then taking a standard part to
recover I.
(4) A closely related method is to obtain I from the If1 ,...,fn , by using
the compactness theorem in logic.
(5) In the case when G is metrisable (and hence separable, by -compactness),
then Cc (G) becomes separable, and one can also use the Arzel
aAscoli theorem in this case. (One can also try in this case to directly ensure that the If1 ,...,fn , converge pointwise, without needing
to pass to a further subsequence, although this requires more effort
than the compactness-based methods.)
These approaches are more or less equivalent to each other, and the choice
of which approach to use is largely a matter of personal taste.
It remains to obtain the approximate functionals If1 ,...,fn , for a given
f0 , f1 , . . . , fn and . As with the uniqueness claim, the basic idea is to
approximate all the functions f0 , f1 , . . . , fn by translates (y) of a given
function . More precisely, let > 0 be a small quantity (depending on
f0 , f1 , . . . , fn and ) to be chosen later. By uniform continuity, we may find
a neighbourhood U of the identity such that fi (xy) = fi (x) + O() for all
x G and y U . Let Cc (G)+ be a function, not identically zero, which
is supported in U .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

75

To motivate the argument that follows, pretend temporarily

R that we
have a left-invariant Haar measure available, and let := G d be
the integral of with respect to this measure. Then 0 < < , and by
left-invariance one has
Z
(y)(x) d(x) = ,
G

and thus
Z X
K

ck (yk )(x) d =

G k=1

K
X

ck

k=1

for any scalars c1 , . . . , cK R+ and y1 , . . . , yK G. In particular, if we

introduce the covering number
)
(K
K
X
X
ck (yk )(x) for all x G
ck : c1 , . . . , cK R+ ; f (x)
[f : ] := inf
k=1

k=1

of a given function f Cc

(G)+

by , we have

Z
f d [f : ].
G

This suggests using a scalar multiple of f 7 [f : ] as the approximate

linear functional (noting that [f : ] can be defined without reference to any
existing Haar measure); in view of the normalisation I(f0 ) = 1, it is then
natural to introduce the functional
[f : ]
I(f ) :=
.
[f0 : ]
(This functional is analogous in some ways to the concept of outer measure
or the upper Darboux integral in measure theory.) Note from compactness
that [f : ] is finite for every f Cc (G)+ , and from the non-triviality of f0
we see that [f0 : ] > 0, so I is well-defined as a map from Cc (G)+ to R.
It is also easy to verify that I obeys the homogeneity, left-invariance, and
normalisation axioms. From the easy inequality
(1.35)

[f : ] [f : f0 ][f0 : ]

we also obtain the uniform bound axiom, and from the infimal nature of
[f : ] we also easily obtain the subadditivity property
I(f + g) I(f ) + I(g).
To finish the construction, it thus suffices to show that
I(fi + fj ) I(fi ) + I(fj )
for each 1 i, j n, if > 0 is chosen sufficiently small depending on
, f0 , f1 , . . . , fn .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

76

1. Hilberts fifth problem

Fix fi , fj . By definition, we have the pointwise bound

fi (x) + fj (x)

(1.36)

K
X

ck (yk )(x)

k=1

for some c1 , . . . , cK with

K
X

(1.37)

k=1



[f0 : ].
ck I(fi + fj ) +
2

If we then write ck = c0k + c00k where

c0k := ck

fi (yk ) +
fi (yk ) + fj (yk ) + 2

c00k := ck

fj (yk ) +
fi (yk ) + fj (yk ) + 2

and

then we claim that

K
X

fi (x)

(1.38)

c0k (yk )(x) + 4

k=1

and
fj (x)

(1.39)

K
X

c00k (yk )(x) + 4

k=1

if is small enough. Indeed, we have

K
X

c0k (yk )(x) =

K
X
k=1

k=1

ck (yk1 x)

fi (yk ) +
.
fi (yk ) + fj (yk ) + 2

(yk1 x)

If
is non-zero, then by the construction of and U , one has |fi (yk )
fi (x)| and |fj (yk ) fj (x)| , which implies that
fi (yk ) +
fi (x)
=
.
fi (yk ) + fj (yk ) + 2
fi (x) + fj (x) + 4
Using (1.36) we thus have
K
X
k=1

c0k (yk )(x) + 4

fi (x)
(fi (x) + fj (x)) + 4
fi (x) + fj (x) + 4

which gives (1.38); a similar argument gives (1.39). From the subadditivity
(and monotonicity) of I, we conclude that
PK 0
c
I(fi ) k=1 k + 4I(g)
[f0 : ]

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

77

and
PK
I(fj )

00
k=1 ck

[f0 : ]

+ 4I(g)

where g Cc (G) equals 1 on the support of fi , fj . Summing and using

(1.37), we conclude that

I(fi ) + I(fj ) I(fi + fj ) + + 8I(g)

2
and the claim follows by taking small enough. This concludes the proof of
Theorem 1.4.6.
Exercise 1.4.6. State and prove a generalisation of Theorem 1.4.6 in which
the hypothesis that G is Hausdorff and -compact are dropped. (This requires extending concepts such as Borel -algebra, Radon measure, and
Haar measure to the non-Hausdorff or non--compact setting. Note that
different texts sometimes have inequivalent definitions of these concepts in
such settings; because of this (and also because of the potential breakdown
of some basic measure-theoretic tools such as the Fubini-Tonelli theorem),
it is usually best to avoid working with Haar measure in the non-Hausdorff
or non--compact case unless one is very careful.)
Remark 1.4.8. An important special case of the Haar measure construction
arises for compact groups G. Here, we can normalise the Haar measure by
requiring that (G) = 1 (i.e. is a probability measure), and so there is
now a unique (left-invariant) Haar probability measure on such a group. In
Exercise 1.4.8 we will see that this measure is in fact bi-invariant.
Remark 1.4.9. The above construction, based on the Riesz representation
theorem, is not the only way to construct Haar measure. Another approach
that is common in the literature is to first build a left-invariant outer measure
and then use the Caratheodory extension theorem. Roughly speaking, the
main difference between that approach and the one given here is that it
is based on covering compact or open sets by other compact or open sets,
rather than covering continuous, compactly supported functions by other
continuous, compactly supported functions. In the Rcompact case, one can
also construct Haar probability measure by defining G f d to be the mean
of f , or more precisely the unique constant function that is an average of
translates of f . See [Ta2010, Exercise 1.12.6] for further discussion (the
text there focuses on the abelian case, but the argument extends to the
nonabelian setting).
Exercise 1.4.7 (Alternate proof of uniqueness of Haar measure). Let , be
two left-invariant Haar measures on a -compact locally compact Hausdorff
group G.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

78

1. Hilberts fifth problem

(i) For any functions f, g Cc (G), establish the identity

Z

Z
 Z

1
f (x)
g(yx) d(y) d(x) =
f (y )d(y)
g(x) d(x) .

Z
G

(Hint: use the Fubini-Tonelli theorem twice and left-invariance

twice.)
R
R
(ii) Show that for any g Cc (G), G g(x) d(x) = 0 implies G g(y) d(y) =
0 and vice versa.
(iii) Conclude that one has = c for some scalar c R+ .
The following exercise explores the distinction between left-invariance
and right-invariance.
Exercise 1.4.8. Let G be a -compact locally compact Hausdorff group,
and let be a left-invariant Haar measure on G.
(i) Show that for each y G, there exists a unique positive real c(y)
(independent of the choice ofR) such that (Ey) = c(y)(E)
for all
R
Borel measurable sets E and G f (xy 1 ) d(x) = c(y) G f (x) d(x)
for all absolutely integrable f . In particular, a left-invariant Haar
measure is right-invariant if and only if c(y) = 1 for all y G.
(ii) Show that the map y 7 c(y) is a continuous homomorphism from
G to the multiplicative group R+ = (R+ , ). (This homomorphism
is known as the modular function, and G is said to be unimodular
if c is identically equal to 1.)
R
R
Show that for any f Cc (G), one has G f (x1 ) d(x)R =R G c(x)1 f (x) d(x).
(Hint: take another function g Cc (G) and evaluate G G g(yx)c(x)1 f (x1 ) d(x)d(y)
in two different ways, one of which involves replacing x by y 1 x.)
In
inRa unimodular group one has (E 1 ) = (E) and
R particular,
1
G f (x ) dx = G f (x) dx for any Borel set E and any f Cc (G).
(iii) Show that G is unimodular if it is compact.
(iv) If G is a Lie group with Lie algebra g, show that c(g) = | det Adg |,
where Adg : g g is the adjoint representation of g, defined by
requiring exp(t Adg X) = g exp(tX)g 1 for all X g (cf. Lemma
1.2.26).
(v) If G is a connected Lie group with Lie algebra g, show that G is
unimodular if and only if tr adX = 0 for all X g, where adX :
Y 7 [X, Y ] is the adjoint representation of X.
(vi) Show that G is unimodular if it is a connected nilpotent Lie group.
(vii) Let G be a connected Lie group whose Lie algebra g is such that
[g, g] = g (where [g, g] is the linear span of the commutators [X, Y ]

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

79

with X, Y g). (This condition is in particular obeyed when the

Lie algebra g is semisimple.) Show that G is unimodular.
(viii) Let G be the group of pairs (a, b) R+ R with the composition
law (a, b)(c, d) := (ac, ad + b). (One can interpret G as the group of
orientation-preserving affine transformations x 7 ax+b on the real
line.) Show that G is a connected Lie group that is not unimodular.
In the case of a Lie group, one can also build Haar measures by starting with a non-invariant smooth measure, and then correcting it. Given
a smooth manifold M , define a smooth measure on M to be a Radon
measure which is a smooth multiple of Lebesgue measure when viewed in
coordinates, thus for any smooth coordinate chart : U V , the pushforward measure ( U ) takes the form f (x) dx V for some smooth function
f : V R+ , thus
Z
f (x) dx

(E) =
(E)

for all E U . We say that the smooth measure is nonvanishing if f is

non-zero on V for every coordinate chart : U V .
Exercise 1.4.9. Let G be a Lie group, and let be a nonvanishing smooth
measure on G.
(i) Show that for every g G, there exists a unique smooth function
g : G R+ such that
Z
Z
1
f (g x) d(x) =
f (x)g (x) d(x).
G

(ii) Verify the cocycle equation gh (x) = g (x)h (gx) for all g, h, x G.
(iii) Show that the measure defined by
Z
(E) :=
x (id)1 d(x)
E

is a left-invariant Haar measure on G.

There are a number of ways to generalise the Haar measure construction.
For instance, one can define a local Haar measure on a local group G. If U is
a neighbourhood of the identity in a -compact locally compact Hausdorff
local group G, we define a local left-invariant Haar measure on U to be
a non-zero Radon measure on U with the property that (gE) = (E)
whenever g G and E U is a Borel set such that gE is well-defined and
also in U .
Exercise 1.4.10 (Local Haar measure). Let G be a -compact locally compact Hausdorff local group, and let U be an open neighbourhood of the
identity in G such that U is symmetric (i.e. U 1 is well-defined and equal

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

80

1. Hilberts fifth problem

to U ) and U 10 is well-defined in G. By adapting the arguments above, show

that there is a local left-invariant Haar measure on U , and that it is unique
up to scalar multiplication. (Hint: a new technical difficulty is that there
are now multiple covering numbers of interest, namely the covering numbers
[f, g]U m associated to various small powers U m of m. However, as long as
one keeps track of which covering number to use at various junctures, this
will not cause difficulty.)
One can also sometimes generalise the Haar measure construction from
groups G to spaces X that G acts transitively on.
Definition 1.4.10 (Group actions). Given a topological group G and a
topological space X, define a (left) continuous action of G on X to be a
continuous map (g, x) 7 gx from G X to X such that g(hx) = (gh)x and
id x = x for all g, h G and x X.
This action is said to be transitive if for any x, y X, there exists g G
such that gx = y, and in this case X is called a homogeneous space with
structure group G, or homogenous G-space for short.
For any x0 X, we call Stab(x0 ) := {g G : gx0 = x0 } the stabiliser of
x0 ; this is a closed subgroup of G.
If G, X are smooth manifolds (so that G is a Lie group) and the action
(g, x) 7 gx is a smooth map, then we say that we have a smooth action of
G on X.
Exercise 1.4.11. If G acts transitively on a space X, show that all the
stabilisers Stab(x0 ) are conjugate to each other, and X is homeomorphic to
the quotient spaces G/ Stab(x0 ) after weakening the topology of the quotient
space (or strengthening the topology of the space X).
If G and X are -compact, locally compact, and Hausdorff, a (left) Haar
measure is a non-zero Radon measure on X such that (gE) = (E) for all
Borel E X and g G.
Exercise 1.4.12. Let G be a -compact, locally compact, and Hausdorff
group (left) acting continuously and transitively on a -compact, locally
compact, and Hausdorff space X.
(i) (Uniqueness up to scalars) Show that if , are (left) Haar measures
on X, then = for some > 0.
(ii) (Compact case) Show that if G is compact, then X is compact too,
and a Haar measure on X exists.
(iii) (Smooth unipotent case) Suppose that the action is smooth (so
that G is a Lie group and X is a smooth manifold). Let x0 be a
point of X. Suppose that for each g Stab(x0 ), the derivative map

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

81

Dg(x0 ) : Tx0 X Tx0 X of the map g : x 7 gx at x0 is unimodular

(i.e. it has determinant 1). Show that a Haar measure on X
exists.
(iv) (Smooth case) Suppose that the action is smooth. Show that any
Haar measure on X is necessarily smooth. Conclude that a Haar
measure exists if and only if the derivative maps Dg are unimodular.
(v) (Counterexample) Let G be the ax+b group from Example 1.4.8(viii),
acting on R by the action (a, b)x := ax + b. Show that there is no
Haar measure on R. (This can be done either through (iv), or by
an elementary direct argument.)
1.4.2. The Peter-Weyl theorem. We now restrict attention to compact
groups G, which we will take to be Hausdorff for simplicity (although the
results in this section will easily extend to the non-Hausdorff case using
Exercise 1.4.3). By the previous discussion, there is a unique bi-invariant
Haar probability measure on G, which gives rise in particular to the Hilbert
space L2 (G) = L2 (G, d) of square-integrable functions f : G C on G
(quotiented out by almost everywhere equivalence, as usual), with norm
Z
1/2
2
kf kL2 (G) :=
|f (x)| d(x)
G

and inner product

Z
hf, giL2 (G) :=

f (x)g(x) dx.
G

For every group element y G, the translation operator (y) : L2 (G)

L2 (G) is defined by
(y)f (x) := f (y 1 x).
One easily verifies that (y 1 ) is both the inverse and the adjoint of (y),
and so (y) is a unitary operator. The map : y 7 (y) is then a continuous
homomorphism from G to the unitary group U (L2 (G)) of L2 (G) (where we
give the latter group the strong operator topology), and is known as the
regular representation of G.
For our purposes, the regular representation is too big of a representation to work with because the underlying Hilbert space L2 (G) is usually
infinite-dimensional. However, we can find smaller representations by locating left-invariant closed subspaces V of L2 (G), i.e. closed linear subspaces of
L2 (G) with the property that (y)V V for all y G. Then the restriction
of to V becomes a representation V : G U (V ) to the unitary group of
V . In particular, if V has some finite dimension n, this gives a representation
of G by a unitary group Un (C) after expressing V in coordinates.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

82

1. Hilberts fifth problem

We can build invariant subspaces from applying spectral theory to an

invariant operator, and more specifically to a convolution operator. If f, g
L2 (G), we define the convolution f g : G C by the formula
Z
f g(x) =
f (y)g(y 1 x) d(y).
G

Exercise 1.4.13. Show that if f, g L2 (G), then f g is well-defined and

lies in C(G), and in particular also lies in L2 (G).
For g L2 (G), let Tg : L2 (G) L2 (G) denote the right-convolution
operator Tg f := f g. This is easily seen to be a bounded linear operator
on L2 (G). Using the properties of Haar measure, we also observe that Tg
will be self-adjoint if g obeys the condition
(1.40)

g(x1 ) = g(x)

and it also commutes with left-translations:

Tg (y) = (y)Tg .
In particular, for any C, the eigenspace
V := {f L2 (G) : Tg f = f }
will be a closed invariant subspace of L2 (G). Thus we see that we can
generate a large number of representations of G by using the eigenspace of
a convolution operator.
Another important fact about these operators, is that the Tg are compact, i.e. they map bounded sets to precompact sets. This is a consequence
of the following more general fact:
Exercise 1.4.14 (Compactness of integral operators). Let (X, ) and (Y, )
be -finite measure spaces, and let K L2 (X Y, ). Define an integral
operator T : L2 (X, ) L2 (Y, ) by the formula
Z
:=
T f (y)
K(x, y)f (x) d(x).
X

(i) Show that T is a bounded linear operator, with operator norm

kT kop bounded by kKkL2 (XY,) . (Hint: use duality.)
(ii) Show that T is a compact linear operator. (Hint: approximate
K by a linear combination of functions of the form a(x)b(y) for
a L2 (X, ) and b L2 (Y, ), plus an error which is small in
L2 (X Y, ) norm, so that T becomes approximated by the
sum of a finite rank operator and an operator of small operator
norm.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

83

Note that Tg is an integral operator with kernel K(x, y) := g(x1 y);

from the invariance properties of Haar measure we see that K L2 (G G)
if g L2 (G) (note here that we crucially use the fact that G is compact,
so that (G) = 1). Thus we conclude that the convolution operator Tg is
compact when G is compact.
Exercise 1.4.15. Show that if g Cc (R) is non-zero, then Tg is not compact on L2 (R). This example demonstrates that compactness of G is needed
in order to ensure compactness of Tg .
We can describe self-adjoint compact operators in terms of their eigenspaces:
Theorem 1.4.11 (Spectral theorem). Let T : H H be a compact selfadjoint operator on a complex Hilbert space H. Then there exists an at most
countable sequence 1 , 2 , . . . of non-zero reals that converge to zero and an
orthogonal decomposition
M
H = V0
Vn
n

of H into the 0 eigenspace (or kernel) V0 of T , and the n -eigenspaces Vn ,

which are all finite-dimensional.
Proof. From self-adjointness we see that all the eigenspaces V are orthogonal
to each other, and only non-trivial for real. If r > 0, then
L
R:||>r V has an orthonormal basis of eigenfunctions v, each of which
is enlarged by a factor of at least r by T . In particular, this basis cannot
be infinite, because otherwise the image of this basis by TLwould have no
convergent subsequence, contradicting compactness. Thus R:||>r V is
finite-dimensional for any r, which implies that V is finite-dimensional for
every non-zero , and those non-zero with non-trivial V can be enumerated to either be finite, or countable and go to zero.
L
Let W be the orthogonal complement of V0 n Vn . If W is trivial,
then we are done, so suppose for sake of contradiction that W is non-trivial.
As all of the V are invariant, and T is self-adjoint, W is also invariant, with
T being self-adjoint on W . As W is orthogonal to the kernel V0 of T , T has
trivial kernel in W . More generally, T has no eigenvectors in W .
Let B be the unit ball in W . As T has trivial kernel and W is non-trivial,
kT kop > 0. Using the identity
(1.41)

kT kop =

sup

|hT x, xi|

W :kxk1

valid for all self-adjoint operators T (see Exercise 1.4.16 below). Thus, we
may find a sequence xn of vectors of norm at most 1 such that
hT xn , xn i

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

84

1. Hilberts fifth problem

for some = kT kop . Since kT xn k2 kT k2op kxn k2 2 , we conclude that

0 kT xn xn k2 = kT xn k2 + 2 kxn k2 2hT xn , xn i 22 2hT xn , xn i
and hence
T xn xn 0;

(1.42)
applying T we conclude that

T (T xn ) T xn 0.
By compactness of T , we may pass to a subsequence so that T xn converges
to a limit y, and thus T y y = 0. As T has no eigenvectors, y must be
trivial; but then hT xn , xn i converges to zero, a contradiction.

Exercise 1.4.16. Establish (1.42) whenever T : W W is a bounded
self-adjoint operator on W . (Hint: Bound |hT x, yi| by the right-hand side
of (1.41) whenever x, y are vectors of norm at most 1, by playing with
hT (ax + by), (ax + by)i for various choices of scalars a, b, in the spirit of the
proof of the Cauchy-Schwarz inequality.)
This leads to the consequence that we can find non-trivial finite-dimensional
representations on at least a single non-identity element:
Theorem 1.4.12 (Baby Peter-Weyl theorem). Let G be a compact Hausdorff group with Haar measure , and let y G be a non-identity element
of G. Then there exists a finite-dimensional invariant subspace of L2 (G) on
which (y) is not the identity.
Proof. Suppose for contradiction that (y) is the identity on every finitedimensional invariant subspace of L2 (G), thus (y)1 annihilates every such
subspace. By Theorem 1.4.11, we conclude that (y) 1 has range in the
kernel of every convolution operator Tg with g L2 , thus Tg ( (y) 1)f = 0
for any f, g L2 (G) with g obeying (1.40), i.e.
(y)(f g) = (f g)
for any such f, g. But one may easily construct f, g such that f g is nonzero at the identity and vanishing at y (e.g. one can set f = g = 1U where
U is an open symmetric neighbourhood of the identity, small enough that y
lies outside U 2 ). This gives the desired contradiction.

Remark 1.4.13. The full Peter-Weyl theorem describes rather precisely
all the invariant subspaces of L2 (G). Roughly speaking, the theorem asserts
that for each irreducible finite-dimensional representation : G U (V )
of G, dim(V ) different copies of V (viewed as an invariant G-space) appear

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

85

in L2 (G), and that they are all orthogonal and make up all of L2 (G); thus,
one has an orthogonal decomposition
M dim(V )

L2 (G)
V

of G-spaces. Actually, this is not the sharpest form of the theorem, as it only
describes the left G-action and not the right G-action; see Section 2.8 for a
precise statement and proof of the Peter-Weyl theorem in its strongest form.
This form is of importance in Fourier analysis and representation theory, but
in this text we will only need the baby form of the theorem (Theorem 1.4.12),
which is an easy consequence of the full Peter-Weyl theorem (since, if g is
not the identity, then (g) is clearly non-trivial on L2 (G) and hence on at
least one of the V factors).
The Peter-Weyl theorem leads to the following structural theorem for
compact groups:
Theorem 1.4.14 (Gleason-Yamabe theorem for compact groups). Let G
be a compact Hausdorff group, and let U be a neighbourhood of the identity.
Then there exists a compact normal subgroup H of G contained in U such
that G/H is isomorphic to a linear group (i.e. a closed subgroup of a general
linear group GLn (C)).
Note from Cartans theorem (Theorem 1.3.2) that every linear group is
Lie; thus, compact Hausdorff groups are almost Lie in some sense.
Proof. Let g be an element of G\U . By the baby Peter-Weyl theorem, we
can find a finite-dimensional invariant subspace V of L2 (G) on which (g)
is non-trivial. Identifying such a subspace with Cn for some finite n, we
thus have a continuous homomorphism : G GLn (C) such that (g)
is non-trivial. By continuity, (g) will also be non-trivial for some open
neighbourhood of g. Using the compactness of G\U , one can then find
a finite number 1 , . . . , k of such continuous homomorphisms i : G
GLni (C) such that for each g G\U , at least one of 1 (g), . . . , k (g) is
non-trivial. If we then form the direct sum
k
k
M
M
:=

i : G
GLni (C) GLn1 ++nk (C)
i=1

i=1

then is still a continuous homomorphism, which is now non-trivial for

any g G\U ; thus the kernel H of is a compact normal subgroup of G
contained in U . There is thus a continuous bijection from the compact space
G/H to the Hausdorff space (G), and so the two spaces are homeomorphic.
As (G) is a compact (hence closed) subgroup of GLn1 ++nk (C), the claim
follows.


Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

86

1. Hilberts fifth problem

Exercise 1.4.17. Show that the hypothesis that G is Hausdorff can be

omitted from Theorem 1.4.14. (Hint: use Exercise 1.4.3.)
Exercise 1.4.18. Show that any compact Lie group is isomorphic to a linear
group. (Hint: first find a neighbourhood of the identity that is so small that
it does not contain any non-trivial subgroups.) The property of having no
small subgroups will be an important one in later sections.
One can rephrase the Gleason-Yamabe theorem for compact groups in
terms of the machinery of inverse limits (also known as projective limits).
Definition 1.4.15 (Inverse limits of groups). Let (G )A be a family of
groups G indexed by a partially ordered set A = (A, <). Suppose that for
each < in A, there is a surjective homomorphism : G G
which obeys the composition law = for all < <
. (If one wishes, one can take a category-theoretic perspective and view
these surjections as describing a functor from the partially ordered set A
to the category of groups.) We then define the inverse Q
limit G = lim G
to be the set of all tuples (g )A in the product set A G such that
(g ) = g for all < ; one easily verifies that this is also a group. We
let : G G denote the coordinate projection maps : (g )A 7 g .
If the G are topological groupsQand the are continuous, we can
give G the topology induced from A G ; one easily verifies that this
makes G a topological group, and that the are continuous homomorphisms.
Exercise 1.4.19 (Universal description of inverse limit). Let (G )A be
a family of groups G with the surjective homomorphisms as in Definition 1.4.15. Let G = lim G be the inverse limit, and let H be another
group. Suppose that one has homomorphisms : H G for each A
such that = for all < . Show that there exists a unique
homomorphism : H G such that = for all A.
Establish the same claim with group and homomorphism replaced
by topological group and continuous homomorphism throughout.
Exercise 1.4.20. Let p be a prime. Show that Zp is isomorphic to the
inverse limit lim Z/pn Z of the cyclic groups Z/pn Z with n N (with the
usual ordering), using the obvious projection homomorphisms from Z/pm Z
to Z/pn Z for m > n.
Exercise 1.4.21. Show that every compact Hausdorff group is isomorphic
(as a topological group) to an inverse limit of linear groups. (Hint: take
the index set A to be the set of all non-empty finite collections of open
neighbourhoods U of the identity, indexed by inclusion.) If the compact

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

87

Hausdorff group is metrisable, show that one can take the inverse limit to
be indexed instead by the natural numbers with the usual ordering.
Exercise 1.4.22. Let G be an abelian group with a homomorphism :
G 7 U (V ) into the unitary group of a finite-dimensional space V . Show
that V can be decomposed as the vector space sum of one-dimensional Ginvariant spaces. (Hint: By the spectral theorem for unitary matrices, any
unitary operator T on V decomposes V into eigenspaces, and any operator
commuting with T must preserve each of these eigenspaces. Now induct on
the dimension of V .)
Exercise 1.4.23 (Fourier analysis on compact abelian groups). Let G be a
compact abelian Hausdorff group with Haar probability measure . Define
a character to be a continuous homomorphism : G 7 S 1 to the unit circle
be the collection of all such characters.
S 1 := {z C : |z| = 1}, and let G
(i) Show that for every g G not equal to the identity, there exists a
character such that (g) 6= 1. (Hint: combine the baby PeterWeyl theorem with the preceding exercise.)
(ii) Show that every function in C(G) is the limit in the uniform topology of finite linear combinations of characters. (Hint: use the
Stone-Weierstrass theorem.)
form an orthonormal basis
(iii) Show that the characters for G
2
of L (G, d).
1.4.3. The structure of locally compact abelian groups. We now use
the above machinery to analyse locally compact abelian groups. We follow
some combinatorial arguments of Pontryagin, as presented in the text of
Montgomery and Zippin [MoZi1974].
We first make a general observation that locally compact groups contain
open subgroups that are finitely generated modulo a compact set. Call
a subgroup of a topological group G cocompact if the quotient space is
compact.
Lemma 1.4.16. Let G be a locally compact group. Then there exists an
open subgroup G0 of G which has a cocompact finitely generated subgroup .
Proof. Let K be a compact neighbourhood of the identity. Then K 2 is also
compact and can thus be covered by finitely many copies of K, thus
K 2 KS
for some finite set S, which we may assume without loss of generality to be
contained in K 1 K 2 . In particular, if is the group generated by S, then
K 2 K.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

88

1. Hilberts fifth problem

Multiplying this on the left by powers of K and inducting, we conclude that

K n K
for all n 1. If we then let G0 be the group generated by K, then lies in
G0 and G0 K G0 . Thus G0 / is the image of the compact set K under
the quotient map, and the claim follows.

In the abelian case, we can improve this lemma by combining it with
the following proposition:
Proposition 1.4.17. Let G be a locally compact Hausdorff abelian group
with a cocompact finitely generated subgroup. Then G has a cocompact discrete finitely generated subgroup.
To prove this proposition, we need the following lemma.
Lemma 1.4.18. Let G be a locally compact Hausdorff group, and let g G.
Then the group hgi generated by g is either precompact or discrete (or both).
Proof. By replacing G with the closed subgroup hgi we may assume without
loss of generality that hgi is dense in G.
We may assume of course that hgi is not discrete. This implies that the
identity element is not an isolated point in hgi, and thus for any neighbourhood of the identity U , there exist arbitrarily large n such that g n U ;
since g n = (g n )1 we may take these n to be large and positive rather than
large and negative.
Let U be a precompact symmetric neighbourhood of the identity, then
U 3 (say) is covered by a finite number gj U of left-translates of U . As hgi
is dense, we conclude that U 3 is covered by a finite number of translates
g nj U 2 of left-translates of U by powers of g. Using the fact that there are
arbitrarily large n with g n U , we may thus cover U 3 by a finite number
of translates g mj U 3 of U 3 with mj > 0. In particular, if g n U 3 , then
there exists an mj such that g nmj U 3 . Iterating this, we see that the set
{n Z : g n U 3 } is left-syndetic, in that it has bounded gaps as one goes
to . Similarly one can argue that this set is right-syndentic and thus
syndetic. This implies that the entire group hgi is covered by a bounded
number of translates of U 3 and is thus precompact as required.

Now we can prove Proposition 1.4.17.
Proof. Let us say that a locally compact Hausdorff abelian group has rank
at most r if it has a cocompact subgroup generated by at most r generators.
We will induct on the rank r. If G has rank 0, then the cocompact subgroup
is trivial, and the claim is obvious; so suppose that G has some rank r 1,
and the claim has already been proven for all smaller ranks.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.4. Haar measure and the Peter-Weyl theorem

89

By hypothesis, G has a cocompact subgroup generated by r generators

e1 , . . . , er . By Lemma 1.4.18, the group her i is either precompact or discrete.
If it is discrete, then we can quotient out by that group to obtain a locally
compact Hausdorff abelian group G/her i of rank at most r 1; by induction
hypothesis, G/her i has a cocompact discrete subgroup, and so G does also.
Hence we may assume that her i is precompact, and more generally that hei i
is precompact for each i. But as we are in an abelian group, is the product
of all the hei i, and is thus also precompact, so is compact. But G/ is
a quotient of G/ and is also compact, and so G itself is compact, and the
claim follows in this case.

We can then combine this with the Gleason-Yamabe theorem for compact groups to obtain
Theorem 1.4.19 (Gleason-Yamabe theorem for abelian groups). Let G be
a locally compact abelian Hausdorff group, and let U be a neighbourhood of
the identity. Then there exists a compact normal subgroup H of G contained
in U such that G/H is isomorphic to a Lie group.
Proof. By Lemma 1.4.16 and Proposition 1.4.17, we can find an open subgroup G0 of G and discrete cocompact subgroup of G0 . By shrinking U
as necessary, we may assume that U is symmetric and U 2 only intersects
at the identity. Let : G0 G0 / be the projection to the compact
abelian group G0 /, then (U ) is a neighbourhood of the identity in G0 /.
By Theorem 1.4.14, one can find a compact normal subgroup H 0 of G0 /
in (U ) such that (G0 /)/H 0 is isomorphic to a linear group, and thus to a
Lie group. If we set H := 1 (H 0 ) U , it is not difficult to verify that H
is also a compact normal subgroup of G0 . If : G0 G0 /H is the quotient
map, then () is a discrete subgroup of G0 /H and from abstract nonsense
one sees that (G0 /H)/() is isomorphic to the Lie group (G/)/H 0 . Thus
G0 /H is locally Lie. Since G0 is an open subgroup of the abelian group G,
G/H is locally Lie also, and is thus G/H is isomorphic to a Lie group by
Exercise 1.2.17.

Exercise 1.4.24. Show that the Hausdorff hypothesis can be dropped from
the above theorem.
Exercise 1.4.25 (Characters separate points). Let G be a locally compact
Hausdorff abelian group, and let g G be not equal to the identity. Show
that there exists a character : G S 1 (see Exercise 1.4.23) such that
(g) 6= 1. This result can be used as the foundation of the theory of Pontryagin duality in abstract harmonic analysis, but we will not pursue this
here; see for instance [Ru1962].

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

90

1. Hilberts fifth problem

Exercise 1.4.26. Show that every locally compact abelian Hausdorff group
is isomorphic to the inverse limit of abelian Lie groups.
Thus, in principle at least, the study of locally compact abelian group is
reduced to that of abelian Lie groups, which are more or less easy to classify:
Exercise 1.4.27.
(i) Show that every discrete subgroup of Rd is iso0
morphic to Zd for some 0 d0 d.
(ii) Show that every connected abelian Lie group G is isomorphic to
0
Rd (R/Z)d for some natural numbers d, d0 . (Hint: first show
that the kernel of the exponential map is a discrete subgroup of the
Lie algebra.) Conclude in particular the divisibility property that
if g G and n 1 then there exists h G with hn = g.
(iii) Show that every compact abelian Lie group G is isomorphic to
(R/Z)d H for some natural number d and a H which is a finite
product of finite cyclic groups. (You may need the classification of
finitely generated abelian groups, and will also need the divisibility
property to lift a certain finite group from a certain quotient space
back to G.)
(iv) Show that every abelian Lie group contains an open subgroup that
0
00
is isomorphic to Rd (R/Z)d Zd H for some natural numbers
d, d0 , d00 and a finite product H of finite cyclic groups.
Remark 1.4.20. Despite the quite explicit description of (most) abelian
Lie groups, some interesting behaviour can still occur in locally compact
abelian groups after taking inverse limits; consider for instance the solenoid
example (Exercise 1.1.7).

1.5. Building metrics on groups, and the Gleason-Yamabe

theorem
In this section we will be able to finally prove the Gleason-Yamabe theorem (Theorem 1.1.17). In the next section, we will combine the GleasonYamabe theorem with some topological analysis (and in particular, using
the invariance of domain theorem) to establish some further control on locally compact groups, and in particular obtaining a solution to Hilberts
fifth problem.
To prove the Gleason-Yamabe theorem, we will use three major tools
developed in previous sections. The first is Theorem 1.3.9, which provided
a criterion for Lie structure in terms of a special type of metric, namely a
Gleason metric (Definition 1.3.8). The second tool is the existence of a leftinvariant Haar measure on any locally compact group; see Theorem 1.4.6.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

91

Finally, we will also need the compact case of the Gleason-Yamabe theorem
(Theorem 1.4.14), which was proven via the Peter-Weyl theorem.
To finish the proof of the Gleason-Yamabe theorem, we have to somehow use the available structures on locally compact groups (such as Haar
measure) to build good metrics on those groups (or on suitable subgroups
or quotient groups). The basic construction is as follows:
Definition 1.5.1 (Building metrics out of test functions). Let G be a topological group, and let : G R+ be a bounded non-negative function.
Then we define the pseudometric d : G G R+ by the formula
d (g, h) := sup | (g)(x) (h)(x)|
xG

= sup |(g 1 x) (h1 x)|

xG

and the semi-norm kk : G R+ by the formula

kgk := d (g, id).
Note that one can also write
kgk = sup |g (x)|
xG

where g (x) := (x) (g 1 x) is the derivative of in the direction g.

Exercise 1.5.1. Let the notation and assumptions be as in the above definition. For any g, h, k G, establish the metric-like properties
(1) (Identity) d (g, h) 0, with equality when g = h.
(2) (Symmetry) d (g, h) = d (h, g).
(3) (Triangle inequality) d (g, k) d (g, h) + d (h, k).
(4) (Continuity) If Cc (G), then the map d : G G R+ is
continuous.
(5) (Boundedness) One has d (g, h) supxG |(x)|. If Cc (G) is
supported in a set K, then equality occurs unless g 1 h KK 1 .
(6) (Left-invariance) d (g, h) = d (kg, kh). In particular, d (g, h) =
kh1 gk = kg 1 hk .
In particular, we have the norm-like properties
(1) (Identity) kgk 0, with equality when g = id.
(2) (Symmetry) kgk = kg 1 k .
(3) (Triangle inequality) kghk kgk + khk .
(4) (Continuity) If Cc (G), then the map kk : G R+ is continuous.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

92

1. Hilberts fifth problem

(5) (Boundedness) One has kgk supxG |(x)|. If Cc (G) is

supported in a set K, then equality occurs unless g KK 1 .
We remark that the first three properties of d in the above exercise
ensure that d is indeed a pseudometric.
To get good metrics (such as Gleason metrics) on groups G, it thus suffices to obtain test functions that obey suitably good regularity properties. We will achieve this primarily by means of two tricks. The first
trick is to obtain high-regularity test functions by convolving together two
low-regularity test functions, taking advantage of the existence of a leftinvariant Haar measure on G. The second trick is to obtain low-regularity
test functions by means of a metric-like object on G. This latter trick may
seem circular, as our whole objective is to get a metric on G in the first
place, but the key point is that the metric one starts with does not need
to have as many good properties as the metric one ends up with, thanks
to the regularity-improving properties of convolution. As such, one can use
a bootstrap argument (or induction argument) to create a good metric
out of almost nothing. It is this bootstrap miracle which is at the heart
of the proof of the Gleason-Yamabe theorem (and hence to the solution of
Hilberts fifth problem).
The arguments here are based on the nonstandard analysis arguments
used to establish Hilberts fifth problem by Hirschfeld [Hi1990] and by
Goldbring [Go2010] (and also some unpublished lecture notes of Goldbring
and van den Dries). However, we will not explicitly use any nonstandard
analysis in this section.
1.5.1. Warmup: the Birkhoff-Kakutani theorem. To illustrate the
basic idea of using test functions to build metrics, let us first establish a
classical theorem on topological groups, which gives a necessary and sufficient condition for metrisability. Recall that a topological space is metrisable
if there is a metric on that space that generates the topology.
Theorem 1.5.2 (Birkhoff-Kakutani theorem). A topological group is metrisable if and only if it is Hausdorff and first countable10.
Remark 1.5.3. The group structure is crucial; for instance, the long line
is Hausdorff and first countable, but not metrisable.
This theorem, together with the existence of Haar measure from the
previous section (and the argument in Exercise 1.5.4 below), completes the
second arrow of Figure 1.
10A topological space is first countable if every point has a countable neighbourhood base.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

93

We now prove Theorem 1.5.2 (following the arguments in [MoZi1974]).

The only if direction is easy, so it suffices to establish the if direction.
The key lemma is
Lemma 1.5.4 (Urysohn-type lemma). Let G be a Hausdorff first countable
group. Then there exists a bounded continuous function : G [0, 1] with
the following properties:
(i) (Unique maximum) (id) = 1, and (x) < 1 for all x 6= id.
(ii) (Neighbourhood base) The sets {x G : (x) > 1 1/n} for
n = 1, 2, . . . form a neighbourhood base at the identity.
(iii) (Uniform continuity) For every > 0, there exists an open neighbourhood U of the identity such that |(gx) (x)| for all
g U and x G.
Note that if G had a left-invariant metric, then the function (x) :=
max(1 d(x, id), 0) would suffice for this lemma, which already gives some
indication as to why this lemma is relevant to the Birkhoff-Kakutani theorem.
Exercise 1.5.2. Let G be a Hausdorff first countable group, and let be
as in Lemma 1.5.4. Show that d is a metric on G (so in particular, d (g, h)
only vanishes when g = h) and that d generates the topology of G (thus
every set which is open with respect to d is open in G, and vice versa).
In view of the above exercise, we see that to prove the Birkhoff-Kakutani
theorem, it suffices to prove Lemma 1.5.4, which we now do. By first countability, we can find a countable neighbourhood base
V1 V2 {id}
of the identity. As G is Hausdorff, we must have

Vn = {id}.

n=1

Using the continuity of the group operations, we can recursively find a sequence of nested open neighbourhoods of the identity
(1.43)

U1 U1/2 U1/4 {id}

such that each U1/2n is symmetric (i.e. g U1/2n if and only if g 1 U1/2n ),
is contained in Vn , and is such that U1/2n+1 U1/2n+1 U1/2n for each n 0.
In particular the U1/2n are also a neighbourhood base of the identity with
(1.44)

U1/2n = {id}.

n=1

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

94

1. Hilberts fifth problem

For every dyadic rational a/2n in (0, 1), we can now define the open sets
Ua/2n by setting
Ua/2n := U1/2nk U1/2n1
where a/2n = 2n1 + + 2nk is the binary expansion of a/2n with 1
n1 < < nk . By repeated use of the hypothesis U1/2n+1 U1/2n+1 U1/2n
we see that the Ua/2n are increasing in a/2n ; indeed, we have the inclusion
(1.45)

U1/2n Ua/2n U(a+1)/2n

for all n 1 and 1 a < 2n .

We now set
n
o
a
(x) := sup 1 n : n 1; 1 a < 2n ; x Ua/2n
2
with the understanding that (x) = 0 if the supremum is over the empty
set. One easily verifies using (1.45) that is continuous, and furthermore
obeys the uniform continuity property. The neighbourhood base property
follows since the U1/2n are a neighbourhood base of the identity, and the
unique maximum property follows from (1.44). This proves Lemma 1.5.4,
and the Birkhoff-Kakutani theorem follows.
Exercise 1.5.3. Let G be a topological group. Show that G is completely
regular, that is to say for every closed subset F in G and every x G\F ,
there exists a continuous function f : G R that equals 1 on F and
vanishes on x.
Exercise 1.5.4 (Reduction to the metrisable case). Let G be a locally
compact group, let U be an open neighbourhood of the identity, and let G0
be the group generated by U .
(i) Construct a sequence of open neighbourhoods of the identity
U U1 U2 . . .
2
U
with the property that Un+1
Un and Un+1
Un for all n 1,
B
b
b
1
where A := {a : a A, b B} and a := b ab.
T
0
(ii) If we set H :=
n=1 Un , show that H is a closed normal subgroup G
in U , and the quotient group G0 /H is Hausdorff and first countable
(and thus metrisable, by the Birkhoff-Kakutani theorem).

(iii) Conclude that to prove the Gleason-Yamabe theorem (Theorem

1.1.17), it suffices to do so under the assumption that G is metrisable.
The above arguments are essentially in [Gl1952].
Exercise 1.5.5 (Birkhoff-Kakutani theorem for local groups). Let G be a
local group which is Hausdorff and first countable. Show that there exists
an open neighbourhood V0 of the identity which is metrisable.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

95

1.5.2. Obtaining the commutator estimate via convolution. We now

return to the main task of constructing Gleason metrics (filling out the remaining arrows in Figure 1). The first thing we will do is dispense with
the commutator property (1.30). To this end, let us temporarily define a
weak Gleason metric on a topological group G to be a left-invariant metric
d : G G R+ which generates the topology on G and obeys the escape
property for some constant C > 0, thus one has
1
1
nkgk whenever g G, n 1, and nkgk .
C
C
In this section we will show
(1.46)

kg n k

Theorem 1.5.5. Every weak Gleason metric is a Gleason metric (possibly

after adjusting the constant C).
This theorem represents the leftmost arrow on the second row of Figure
1.
We now prove Theorem 1.5.5. The key idea here is to involve a bump
function formed by convolving together two Lipschitz functions. The
escape property (1.46) will be crucial in obtaining quantitative control of
the metric geometry at very small scales, as one can study the size of a
group element g very close to the origin through its powers g n , which are
further away from the origin.
Specifically, let > 0 be a small quantity to be chosen later, and let
Cc (G) be a non-negative Lipschitz function supported on the ball B(0, )
which is not identically zero. For instance, one could use the explicit function


kxk
(x) := 1
+
where y+ := max(y, 0), although the exact form of will not be important
for our argument. Being Lipschitz, we see that
(1.47)

kg kCc (G)  kgk

for all g G (where we allow implied constants to depend on G, , and ),

where kkCc (G) denotes the sup norm.
Let be a left-invariant Haar measure on G, the existence of which was
established in Theorem 1.4.6. We then form the convolution := ,
with convolution defined using the formula
Z
:=
(1.48)
f g(x)
f (y)g(y 1 x) d(y).
G

This is a continuous function supported in B(0, 2), and gives a metric d

and a norm kk as usual.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

96

1. Hilberts fifth problem

We now prove a variant of the commutator estimate (1.30), namely that

kg h kCc (G)  kgkkhk

(1.49)

whenever g, h B(0, ). To see this, we first use the left-invariance of Haar

measure to write
h = (h ) ,

(1.50)
thus

(h )(y)(y 1 x) d(y).

h (x) =
G

We would like to similarly move the g operator over to the second factor,
but we run into a difficulty due to the non-abelian nature of G. Nevertheless,
we can still do this provided that we twist that operator by a conjugation.
More precisely, we have
Z
(1.51)
g h (x) = (h )(y)(gy )(y 1 x) d(y)
G

gy

y 1 gy

:=
where
is g conjugated by y. If h B(0, ), the integrand is only
non-zero when y B(0, 2). Applying (1.47), we obtain the bound
kg h kCc (g)  khk

sup

kg y k.

yB(0,2)

To finish the proof of (1.49), it suffices to show that

kg y k  kgk
whenever g B(0, ) and y B(0, 2).
We can achieve this by the escape property (1.46). Let n be a natural
number such that nkgk , then kg n k and so g n B(0, ). Conjugating
by y, this implies that (g y )n B(0, 5), and so by (1.46), we have kg y k  n1
(if is small enough), and the claim follows.
Next, we claim that the norm kk is locally comparable to the original
norm kk. More precisely, we claim:
(1) If g G with kgk sufficiently small, then kgk  kgk .
(2) If g G with kgk sufficiently small, then kgk  kgk.
Claim 2 follows easily from (1.50) and (1.47), so we turn to Claim 1. Let
g G, and let n be a natural number such that
nkgk < kkCc (G) .
Then by the triangle inequality
kg n k < kkCc (G) .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

97

This implies that and gn have overlapping support, and hence g n lies in
B(0, 4). By the escape property (1.46), this implies (if is small enough)
that kgk  n1 , and the claim follows.
Combining Claim 2 with (1.49) we see that
kg h kCc (G)  kgk khk
whenever kgk , khk are small enough. Now we use the identity
k[g, h]k = k ([g, h]) kCc (G)
= k (g) (h) (h) (g)kCc (G)
= kg h h g kCc (G)
and the triangle inequality to conclude that
k[g, h]k  kgk khk
whenever kgk , khk are small enough. Theorem 1.5.5 then follows from
Claim 1 and Claim 2.
1.5.3. Building metrics on NSS groups. We will now be able to build
metrics on groups using a set of hypotheses that do not explicitly involve
any metric at all. The key hypothesis will be the no small subgroups (NSS)
property:
Definition 1.5.6 (No small subgroups). A topological group G has the no
small subgroups (or NSS) property if there exists an open neighbourhood
U of the identity which does not contain any subgroup of G other than the
trivial group.
Exercise 1.5.6. Show that any Lie group is NSS.
Exercise 1.5.7. Show that any group with a weak Gleason metric is NSS.
For an example of a group which is not NSS, consider the infinitedimensional torus (R/Z)N . From the definition of the product topology,
we see that any neighbourhood of the identity in this torus contains an
infinite-dimensional subtorus, and so this group is not NSS.
Exercise 1.5.8. Show that for any prime p, the p-adic groups Zp and Qp
are not NSS. What about the solenoid group R Zp /Z ?
Exercise 1.5.9. Show that an NSS group is automatically Hausdorff. (Hint:
use Exercise 1.4.3.)
Exercise 1.5.10. Show that an NSS locally compact group is automatically
metrisable. (Hint: use Exercise 1.5.4.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

98

1. Hilberts fifth problem

Exercise 1.5.11 (NSS implies escape property). Let G be a locally compact

NSS group. Show that if U is a sufficiently small neighbourhood of the
identity, then for every g G\{id}, there exists a positive integer n such
that g n 6 U . Furthermore, for any other neighbourhood V of the identity,
there exists a positive integer N such that if g, . . . , g N U , then g V .
We can now prove the following theorem (first proven in full generality by
Yamabe [Ya1953b]), which is a key component in the proof of the GleasonYamabe theorem and in the wider theory of Hilberts fifth problem, and a
crucial arrow in Figure 1.
Theorem 1.5.7. Every NSS locally compact group admits a weak Gleason
metric. In particular, by Theorem 1.5.5 and Theorem 1.3.9, every NSS
locally compact group is isomorphic to a Lie group.
In view of this theorem and Exercise 1.5.6, we see that for locally compact groups, we obtain
Corollary 1.5.8. A locally compact group is NSS if and only if it is isomorphic to a Lie group.
This is a major advance towards both the Gleason-Yamabe theorem and
Hilberts fifth problem, as it has reduced the property of being a Lie group
into a condition that is almost purely algebraic in nature.
We now prove Theorem 1.5.7. An important concept will be that of an
escape norm associated to an open neighbourhood U of a group G, defined
by the formula


1
2
n
(1.52)
kgke,U := inf
: g, g , . . . , g U
n+1
for any g G, where n ranges over the natural numbers (thus, for instance
kgke,U 1, with equality iff g 6 U ). Thus, the longer it takes for the orbit
g, g 2 , . . . to escape U , the smaller the escape norm.
Strictly speaking, the escape norm is not necessarily a norm, as it need
not obey the symmetry, non-degeneracy, or triangle inequalities; however,
we shall see that in many situations, the escape norm behaves similarly to a
norm, even if it does not exactly obey the norm axioms. Also, as the name
suggests, the escape norm will be well suited for establishing the escape
property (1.46).
It is possible for the escape norm kgke,U of a non-identity element g G
to be zero, if U contains the group hgi generated by U . But if the group G
has the NSS property, then we see that this cannot occur for all sufficiently
small U (where sufficiently small means contained in a suitably chosen
open neighbourhood U0 of the identity). In fact, more is true: if U, U 0

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

99

are two sufficiently small open neighbourhoods of the identity in a locally

compact NSS group G, then the two escape norms are comparable, thus we
have
(1.53)

kgke,U  kgke,U 0  kgke,U

for all g G (where the implied constants can depend on U, U 0 ).

By symmetry, it suffices to prove the second inequality in (1.53). By
(1.52), it suffices to find an integer m such that whenever g G is such that
g, g 2 , . . . , g m U , then g U 0 . But this follows from Exercise 1.5.11. This
concludes the proof of (1.53).
Exercise 1.5.12. Let G be a locally compact group. Show that if d is a leftinvariant metric on G obeying the escape property (1.46) that generates the
topology, then G is NSS, and kgk is comparable to kgke,U for all sufficiently
small U and for all sufficiently small g. (In particular, any two left-invariant
metrics obeying the escape property and generating the topology are locally
comparable to each other.)
Henceforth G is a locally compact NSS group. We now establish a
metric-like property on the escape norm kke,U0 .
Proposition 1.5.9 (Approximate triangle inequality). Let U0 be a sufficiently small open neighbourhood of the identity. Then for any n and any
g1 , . . . , gn G, one has
kg1 . . . gn ke,U0 

n
X

kgi ke,U0

i=1

(where the implied constant can depend on U0 ).

Of course, in view of (1.53), the exact choice of U0 is irrelevant, so long
as it is small. It is slightly convenient to take U0 to be symmetric (thus
U0 = U01 ), so that kgke,U0 = kg 1 ke,U0 for all g.
Proof. We will use a bootstrap argument. Assume to start with that we
somehow already have a weaker form of the conclusion, namely
(1.54)

kg1 . . . gn ke,U0 M

n
X

kgi ke,U0

i=1

for all n, g1 , . . . , gn and some huge constant M ; we will then deduce the same
estimate with a smaller value of M . Afterwards we will show how to remove
the hypothesis (1.54).
Now suppose we have (1.54) for some M . Motivated by the argument
in the previous section, we now try to convolve together two Lipschitz

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

100

1. Hilberts fifth problem

functions. For this, we will need some metric-like functions. Define the
modified escape norm kgk,U0 by the formula
)
( n
X
kgk,U0 := inf
kgi ke,U0 : g = g1 . . . gn
i=1

where the infimum is over all possible ways to split g as a finite product of
group elements. From (1.54), we have
1
kgke,U0 kgk,U0 kgke,U0
M
and we have the triangle inequality
(1.55)

kghk,U0 kgk,U0 + khk,U0

for any g, h G. We also have the symmetry property kgk,U0 = kg 1 k,U0 .
Thus kk,U0 gives a left-invariant semi-metric on G by defining
dist,U0 (g, h) := kg 1 hk,U0 .
We can now define a Lipschitz function : G R by setting
(x) := (1 M dist,U0 (x, U0 ))+ .
On the one hand, we see from (1.55) that this function takes values in [0, 1]
obeys the Lipschitz bound
(1.56)

|g (x)| M kgke,U0

for any g, x G. On the other hand, it is supported in the region where

dist,U0 (x, U0 ) 1/M , which by (1.55) (and (1.52)) is contained in U02 .
We could convolve with itself in analogy to the preceding section, but
in doing so, we will eventually end up establishing a much worse estimate
than (1.54) (in which the constant M is replaced with something like M 2 ).
Instead, we will need to convolve with another function , that we define
as follows. We will need a large natural number L (independent of M ) to
be chosen later, then a small open neighbourhood U1 U0 of the identity
(depending on L, U0 ) to be chosen later. We then let : G R be the
function
j
(x) := sup{1 : x U1j U0 ; j = 0, . . . , L} {0}.
L
Similarly to , we see that takes values in [0, 1] and obeys the Lipschitztype bound
1
(1.57)
|g (x)|
L
for all g U1 and x G. Also, is supported in U1L U0 , and hence (if U1 is
sufficiently small depending on L, U0 ) is supported in U02 , just as is.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

101

The functions , need not be continuous, but they are compactly supported, bounded, and Borel measurable, and so one can still form their convolution := , which will then be continuous and compactly supported;
indeed, is supported in U04 .
We have a lower bound on how big is, since
(0) (U0 )  1
(where we allow implied constants to depend on , U0 , but remain independent of L, U1 , or M ). This gives us a way to compare kk with kke,U0 .
Indeed, if nkgk < (0), then (as in the proof of Claim 1 in the previous
section) we have g n U08 ; this implies that
kgke,U08  kgk
for all g G, and hence by (1.53) we have
kgke,U0  kgk

(1.58)

also. In the converse direction, we have

kgk = kg ( )kCc (G)
= k(g ) kCc (G)

(1.59)

 M kgke,U0
thanks to (1.56). But we can do better than this, as follows. For any
g, h G, we have the analogue of (1.51), namely
Z
g h (x) = (h )(y)(gy )(y 1 x) d(y)
G

If h U0 , then the integrand vanishes unless y U03 . By continuity, we can

find a small open neighbourhood U2 U1 of the identity such that g y U1
for all g U2 and y U03 ; we conclude from (1.56), (1.57) that
M
khke,U0 .
L
whenever h U0 and g U2 . To use this, we observe the telescoping identity
|g h (x)| 

gn = ng +

n1
X

g g i

i=0

for any g G and natural number n, and thus by the triangle inequality
!
n1
X
n
(1.60)
kg k = nkgk + O
kg gi kCc (G) .
i=0

We conclude that


M
kg k = nkgk + O n kgke,U0
L
n

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

102

1. Hilberts fifth problem

whenever n 1 and g, . . . , g n U2 . Using the trivial bound kg n k = O(1),

we then have
1 M
kgk  +
kgke,U0 ;
n
L
optimising in n we obtain
M
kgk  kgke,U2 +
kgke,U0
L
and hence by (1.53)


M
kgk 
+ OU2 (1) kgke,U0
L
where the implied constant in OU2 (1) can depend on U0 , U1 , U2 , L, but is
crucially independent of M . Note the essential gain of L1 here compared
with (1.59). We also have the norm inequality
kg1 . . . gn k

n
X

kgi k .

i=1

Combining these inequalities with (1.58) we see that

X

n
1
M + OU2 (1)
kg1 . . . gn ke,U0 
kgi ke,U0 .
L
i=1

Thus we have improved the constant M in the hypothesis (1.54) to O( L1 M )+

OU2 (1). Choosing L large enough and iterating, we conclude that we can
bootstrap any finite constant M in (1.54) to O(1).
Of course, there is no reason why there has to be a finite M for which
(1.54) holds in the first place. However, one can rectify this by the usual
trick of creating an epsilon of room. Namely, one replaces the escape norm
kgke,U0 by, say, kgke,U0 + for some small > 0 in the definition of kk,U0
and in the hypothesis (1.54). Then the bound (1.54) will be automatic with
a finite M (of size about O(1/)). One can then run the above argument
with the requisite changes and conclude a bound of the form
kg1 . . . gn ke,U0 

n
X

(kgi ke,U0 + )

i=1

uniformly in ; we omit the details. Sending 0, we have thus shown

Proposition 1.5.9.

Now we can finish the proof of Theorem 1.5.7. Let G be a locally compact
NSS group, and let U0 be a sufficiently small neighbourhood of the identity.
From Proposition 1.5.9, we see that the escape norm kke,U0 and the modified
escape norm kk,U0 are comparable. We have seen d,U0 is a left-invariant
pseudometric. As G is NSS and U0 is small, there are no non-identity

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

103

elements with zero escape norm, and hence no non-identity elements with
zero modified escape norm either; thus d,U0 is a genuine metric.
We now claim that d,U0 generates the topology of G. Given the leftinvariance of d,U0 , it suffices to establish two things: firstly, that any open
neighbourhood of the identity contains a ball around the identity in the
d,U0 metric; and conversely, any such ball contains an open neighbourhood
around the identity.
To prove the first claim, let U be an open neighbourhood around the
identity, and let U 0 U be a smaller neighbourhood of the identity. From
(1.53) we see (if U 0 is small enough) that kk,U0 is comparable to kke,U 0 , and
U 0 contains a small ball around the origin in the d,U0 metric, giving the
claim. To prove the second claim, consider a ball B(0, r) in the d,U0 metric.
For any positive integer m, we can find an open neighbourhood Um of the
1
m U , and hence kgk
identity such that Um
0
e,U0 m for all g Um . For m
large enough, this implies that Um B(0, r), and the claim follows.
To finish the proof of Theorem 1.5.7, we need to verify the escape property (1.46). Thus, we need to show that if g G, n 1 are such that
nkgk,U0 is sufficiently small, then we have kg n k,U0  nkgk,U0 . We may of
course assume that g is not the identity, as the claim is trivial otherwise. As
kk,U0 is comparable to kke,U0 , we know that there exists a natural number
m  1/kgk,U0 such that g m 6 U0 .
Let U1 be a neighbourhood of the identity small enough that U12 U0 .
We have kg i k,U0 nkgk,U0 for all i = 1, . . . , n, so g i U1 and hence
m > n. Let m + i be the first multiple of n larger than n, then i n and so
g i U1 . Since g m 6 U0 , this implies g m+i 6 U1 . Since m + i is divisible by
n
 nkgk,U0 , and the claim follows from
n, we conclude that kg n ke,U1 m+i
(1.53).
1.5.4. NSS from subgroup trapping. In view of Theorem 1.5.7, the only
remaining task in the proof of the Gleason-Yamabe theorem is to locate big
subquotients G0 /H of a locally compact group G with the NSS property. We
will need some further notation. Given a neighbourhood V of the identity
in a topological group G, let Q[V ] denote the union of all the subgroups of
G that are contained in V . Thus, a group is NSS if Q[V ] is trivial for all
sufficiently small V .
We will need a property that is weaker than NSS:
Definition 1.5.10 (Subgroup trapping). A topological group has the subgroup trapping property if, for every open neighbourhood U of the identity,
there exists another open neighbourhood V of the identity such that Q[V ]
generates a subgroup hQ[V ]i contained in U .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

104

1. Hilberts fifth problem

Clearly, every NSS group has the subgroup trapping property. Informally, groups with the latter property do have small subgroups, but one
cannot get very far away from the origin just by combining together such
subgroups.
Example 1.5.11. The infinite-dimensional torus (R/Z)N does not have the
NSS property, but it does have the subgroup trapping property.
It is difficult to produce an example of a group that does not have the
subgroup trapping property; the reason for this will be made clear in the
next section. For now, we establish the following key result (another arrow
of Figure 1):
Proposition 1.5.12 (From subgroup trapping to NSS). Let G be a locally
compact group with the subgroup trapping property, and let U be an open
neighbourhood of the identity in G. Then there exists an open subgroup G0 of
G, and a compact subgroup N of G0 contained in U , such that G0 /N is locally
compact and NSS. In particular, by Theorem 1.5.7, G0 /N is isomorphic to
a Lie group.
Intuitively, the idea is to use the subgroup trapping property to find
a small compact normal subgroup N that contains Q[V ] for some small
V , and then quotient this group out to get an NSS group. Unfortunately,
because N is not necessarily contained in V , this quotienting operation
may create some additional small subgroups. To fix this, we need to pass
from the compact subgroup N to a smaller one. In order to understand
the subgroups of compact groups, the main tool will be Gleason-Yamabe
theorem for compact groups (Theorem 1.4.14).
For us, the main reason why we need the compact case of the GleasonYamabe theorem is that Lie groups automatically have the NSS property,
even though G need not. Thus, one can view Theorem 1.4.14 as giving the
compact case of Proposition 1.5.12.
We now prove Proposition 1.5.12, using an argument of Yamabe [Ya1953].
Let G be a locally compact group with the subgroup trapping property, and
let U be an open neighbourhood of the identity. We may find a smaller
neighbourhood U1 of the identity with U12 U , which in particular implies that U1 U ; by shrinking U1 if necessary, we may assume that U1 is
compact. By the subgroup trapping property, one can find an open neighbourhood U2 of the identity such that hQ(U2 )i is contained in U1 , and thus
H := hQ(U2 )i is a compact subgroup of G contained in U . By shrinking U2
if necessary we may assume U2 U1 .
Ideally, if H were normal and contained in U2 , then the quotient group
hU2 i/H would have the NSS property. Unfortunately H need not be normal,

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

105

and need not be contained in U2 , but we can fix this as follows. Applying
Theorem 1.4.14, we can find a compact normal subgroup N of H contained
in U2 H such that H/N is isomorphic to a Lie group, and in particular
is NSS. In particular, we can find an open symmetric neighbourhood U3
of the identity in G such that U3 N U3 U2 and that the quotient space
(U3 N U3 H) has no non-trivial subgroups in H/N , where : H H/N
is the quotient map.
We now claim that N is normalised by U3 . Indeed, if g U3 , then the
conjugate N g := g 1 N g of N is contained in U3 N U3 and hence in U2 . As
N g is a group, it must thus be contained in Q(U2 ) and hence in H. But
then (N g ) is a subgroup of H/N that is contained in (U3 N U3 H), and
is hence trivial by construction. Thus N g N , and so N is normalised by
U3 . If we then let G0 be the subgroup of G generated by N and U3 , we see
that G0 is an open subgroup of G, with N a compact normal subgroup of
G0 .
To finish the job, we need to show that G0 /N has the NSS property. It
suffices to show that U3 N U3 /N has no nontrivial subgroups. But any subgroup in U3 N U3 /N pulls back to a subgroup in U3 N U3 , hence in U2 , hence
in Q(U2 ), hence in H; since (U3 N U3 H)/N has no nontrivial subgroups,
the claim follows. This concludes the proof of Proposition 1.5.12.
1.5.5. The subgroup trapping property. In view of Theorem 1.5.7,
Proposition 1.5.12, and Exercise 1.5.4, we see that the Gleason-Yamabe
theorem (Theorem 1.1.17) now reduces to the following claim.
Proposition 1.5.13. Every locally compact metrisable group has the subgroup trapping property.
This proposition represents the final two arrows of Figure 1.
We now prove Proposition 1.5.13, which is the hardest step of the entire
proof and uses almost all the tools already developed. In particular, it
requires both Theorem 1.4.14 and Gleasons convolution trick, as well as
some of the basic theory of Hausdorff distance; as such, this is perhaps the
most infinitary of all the steps in the argument.
The Gleason-type arguments can be encapsulated in the following proposition, which is a weak version of the subgroup trapping property:
Proposition 1.5.14 (Finite trapping). Let G be a locally compact group,
let U be an open precompact neighbourhood of the identity, and let m 1
be an integer. Then there exists an open neighbourhood V of the identity
with the following property: if Q Q[V ] is a symmetric set containing the
identity, and n 1 is such that Qn U , then Qmn U 8 .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

106

1. Hilberts fifth problem

Informally, Proposition 1.5.14 asserts that subsets of Q[V ] grow much

more slowly than large sets such as U . We remark that if one could replace
U 8 in the conclusion here by U , then a simple induction on n (after first
shrinking V to lie in U ) would give Proposition 1.5.13. It is the loss of 8 in
the exponent that necessitates some non-trivial additional arguments.
Proof of Proposition 1.5.14. Let V be small enough to be chosen later,
and let Q, n be as in the proposition. Once again we will convolve together
two Lipschitz functions , to obtain a good bump function =
which generates a useful metric for analysing the situation. The first bump
function : G R will be defined by the formula
j
(x) := sup{1 : x Qj U ; j = 0, . . . , n} {0}.
n
Then takes values in [0, 1], equals 1 on U , is supported in U 2 , and obeys
the Lipschitz type property
1
(1.61)
|q (x)|
n
for all q Q. The second bump function : G R is similarly defined by
the formula
j
4
: x (V U )j U ; j = 0, . . . , M } {0},
(x) := sup{1
M
4

where V U := {g 1 xg : x V, g U 4 }, where M is a quantity depending

on m and U to be chosen later. If V is small enough depending on U and
4
m, then (V U )M U , and so also takes values in [0, 1], equals 1 on U , is
supported in U 2 , and obeys the Lipschitz type property
1
(1.62)
|g (x)|
M
4

for all g V U .
Now let := . Then is supported on U 4 and kkCc (G)  1 (where
implied constants can depend on U , ). As before, we conclude that g U 8
whenever kgk is sufficiently small.
Now suppose that q Q[V ]; we will estimate kqk . From (1.60) one has
kqk 

1 n
kq k + sup kqi q kCc (G)
n
0in

(note that qi and q commute). For the first term, we can compute
kq n k = sup |qn ( )(x)|
x

and

Z
qn ( )(x) =

(y)(qn )y (y 1 x)d(y).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

107

Since q Q[V ], q n V , so by (1.62) we conclude that

1
kq n k 
.
M
For the second term, we similarly expand
Z
qi qi (x) = (q )(y)(qn )y (y 1 x)d(y).
G

Using (1.62), (1.61) we conclude that

|qi qi (x)| 

1
.
Mn

Putting this together we see that

1
Mn
for all q Q[V ], which in particular implies that
m
kgk 
M
for all g Qmn . For M sufficiently large, this gives Qmn U 8 as required.

kqk 

We will also need the following compactness result in the Hausdorff

distance
dH (E, F ) := max(sup dist(x, F ), sup dist(E, y))
xE

yF

between two non-empty closed subsets E, F of a metric space (X, d).

Example 1.5.15. In R with the usual metric, the finite sets { ni : i =
1, . . . , n} converge in Hausdorff distance to the closed interval [0, 1].
Exercise 1.5.13. Show that the space K(X) of non-empty closed subsets
of a compact metric space X is itself a compact metric space (with the
Hausdorff distance as the metric). (Hint: use the Heine-Borel theorem.)
Now we can prove Proposition 1.5.13. Let G be a locally compact group
endowed with some metric d, and let U be an open neighbourhood of the
identity; by shrinking U we may assume that U is precompact. Let Vi be a
sequence of balls around the identity with radius going to zero, then Q[Vi ] is
a symmetric set in Vi that contains the identity. If, for some i, Q[Vi ]n U
for every n, then hQ[Vi ]i U and we are done. Thus, we may assume
for sake of contradiction that there exists ni such that Q[Vi ]ni U and
Q[Vi ]ni +1 6 U ; since the Vi go to zero, we have ni . By Proposition
1.5.14, we can also find mi such that Q[Vi ]mi ni U 8 .
ni

The sets Q[Vi ] are closed subsets of U ; by Exercise 1.5.13, we may

pass to a subsequence and assume that they converge to some closed subset

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

108

1. Hilberts fifth problem

E of U . Since the Q[Vi ] are symmetric and contain the identity, E is also
symmetric and contains the identity. For any fixed m, we have Q[Vi ]mni
U 8 for all sufficiently large i, which on taking Hausdorff limits implies that
E m U 8 . In particular, the group H := hEi is a compact subgroup of G
contained in U 8 .
Let U1 be a small neighbourhood of the identity in G to be chosen later.
By Theorem 1.4.14, we can find a normal subgroup N of H contained in
U1 H such that H/N is NSS. Let B be a neigbourhood of the identity in
H/N so small that B 10 has no small subgroups. A compactness argument
then shows that there exists a natural number k such that for any g H/N
that is not in B, at least one of g, . . . , g k must lie outside of B 10 .
Now let > 0 be a small parameter. Since Q[Vi ]ni +1 6 U , we see
that Q[Vi ]ni +1 does not lie in the -neighbourhood 1 (B) of 1 (B) if is
small enough, where : H H/N is the projection map. Let n0i be the first
0
integer for which Q[Vi ]ni does not lie in 1 (B) , then n0i ni + 1 and n0i
0
as i (for fixed ). On the other hand, as Q[Vi ]ni 1 1 (B) , we see
0
from another application of Proposition 1.5.14 that Q[Vi ]kni ( 1 (B) )8
if i is sufficiently large depending on .
On the other hand, since Q[Vi ]ni converges to a subset of H in the
Hausdorff distance, we know that for i large enough, Q[Vi ]2ni and hence
0
Q[Vi ]ni is contained in the -neighbourhood of H. Thus we can find an
0
element gi of Q[Vi ]ni that lies within of a group element hi of H, but
does not lie in 1 (B) ; thus hi lies inside H\ 1 (B). By construction
of B, we can find 1 ji k such that hji i lies in H\ 1 (B 10 ). But hji i
0
also lies within o(1) of giji , which lies in Q[Vi ]kni and hence in ( 1 (B) )8 ,
where o(1) denotes a quantity depending on that goes to zero as 0.
We conclude that H\ 1 (B 10 ) and 1 (B 8 ) are separated by o(1), which
leads to a contradiction if is sufficiently small (note that 1 (B 8 ) and
H\ 1 (B 10 ) are compact and disjoint, and hence separated by a positive
distance), and the claim follows.
Exercise 1.5.14. Let X be a compact metric space, Kc (X) denote the
space of non-empty closed and connected subsets of X. Show that Kc (X)
with the Hausdorff metric is also a compact metric space.
Exercise 1.5.15. Show that the metrisability condition in Proposition 1.5.13
can be dropped; in other words, show that every locally compact group has
the subgroup trapping property.
1.5.6. The local group case. In [Go2009], [Go2010], [vdDrGo2010],
the above theory was extended to the setting of local groups. In fact, there
is relatively little difficulty (other than some notational difficulties) in doing
so, because the analysis in the previous sections can be made to take place

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

109

on a small neighbourhood of the origin. This extension to local groups is

not simply a generalisation for its own sake; it will turn out that it will be
natural to work with local groups when we classify approximate groups in
later sections.
One technical issue that comes up in the theory of local groups is that
basic cancellation laws such as gh = gk = h = k, which are easily verified
for groups, are not always true for local groups. However, this is a minor
issue as one can always recover the cancellation laws by passing to a slightly
smaller local group, as follows.
Definition 1.5.16 (Cancellative local group). A local group G is said to
be symmetric if the inverse operation is always well-defined. It is said to be
cancellative if it is symmetric, and the following axioms hold:
(i) Whenever g, h, k G are such that gh and gk are well-defined
and equal to each other, then h = k. (Note that this implies in
particular that (g 1 )1 = g.)
(ii) Whenever g, h, k G are such that hg and kg are well-defined and
equal to each other, then h = k.
(iii) Whenever g, h G are such that gh and h1 g 1 are well-defined,
then (gh)1 = h1 g 1 . (In particular, if U G is symmetric
and U m is well-defined in G for some m 1, then U m is also
symmetric.)
Clearly, all global groups are cancellative, and more generally the restriction of a global group to a symmetric neighbourhood of the identity s
cancellative. While not all local groups are cancellative, we have the following substitute:
Exercise 1.5.16. Let G be a local group. Show that there is a neighbourhood U of the identity which is cancellative (thus, the restriction G U of G
to U is cancellative).
Note that any symmetric neighbourhood of the identity in a cancellative
local group is again a cancellative local group. Because of this, it turns out
in practice that we may restrict to the cancellative setting without much
loss of generality.
Next, we need to localise the notion of a quotient G/H of a global
group G by a normal subgroup H. Recall that in order for a subset H
og a global group G to be a normal subgroup, it has to be symmetric,
contain the identity, be closed under multiplication (thus h1 h2 H whenever
h1 , h2 H), and closed under conjugation (thus hg := g 1 hg H whenever
h H and g G). We now localise this concept as follows:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

110

1. Hilberts fifth problem

Definition 1.5.17 (Sub-local groups). Given two symmetric local groups

G0 and G, we say that G0 is a sub-local group of G if G0 is the restriction
of G to a symmetric neighbourhood of the identity, and there exists an
open neighbourhood V of G0 with the property that whenever g, h G0 are
such that gh is defined in V , then gh G0 ; we refer to V as an associated
neighbourhood for G0 . If G0 is also a global group, we say that G0 is a
subgroup of G.
If G0 is a sub-local group of G, we say that G0 is normal if there exists
an associated neighbourhood V for G0 with the additional property that
whenever g 0 G0 , h V are such that hg 0 h1 is well-defined and lies in V ,
then hg 0 h1 G0 . We call V a normalising neighbourhood of G0 .
Example 1.5.18. If G, G0 are the (additive) local groups G := {2, 1, 0, +1, +2}
and G0 := {1, 0, +1}, then G0 is a sub-local group of G (with associated
neighbourhood V = G0 ). Note that this is despite G0 not being closed with
respect to addition in G; thus we see why it is necessary to allow the associated neighbourhood V to be strictly smaller than G. In a similar vein, the
open interval (1, 1) is a sub-local group of (2, 2).
The interval (1, 1) {0} is also a sub-local group of R2 ; here, one can
take for instance (1, 1)2 as the associated neighbourhood. As all these
examples are abelian, they are clearly normal.
Example 1.5.19. Let T : V V be a linear transformation on a finitedimensional vector space V , and let G := Z nT V be the associated semidirect product. Let G0 := {0} W , where W is a subspace of V that is
not preserved by T . Then G0 is not a normal subgroup of G, but it is a
normal sub-local group of G, where one can take {0} V as a normalising
neighbourhood of G0 .
Example 1.5.20. In the global group G = R2 = (R2 , +), the open interval
H := (1, 1) {0} is a normal sub-local subgroup if one takes (say) V :=
(1, 1) (1, 1) as the normalising neighbourhood.
Example 1.5.21. Let T : (R/Z)Z (R/Z)Z be the shift map T (an )nZ :=
(an1 )nZ , and let ZnT (R/Z)Z be the semidirect product of Z and (R/Z)Z .
Then if H is any (global) subgroup of (R/Z)Z , the set {0} H is a normal
sub-local subgroup of Z nT (R/Z)Z (with normalising neighbourhood {0}
(R/Z)Z ). This is despite the fact that H will, in general, not be normal in
Z nT (R/Z)Z in the classical (global) sense.
It is easy to see that if H is a normal sublocal group of G, then H is itself
a cancellative local group, using the topology and group structure formed
by restriction from G. (Note how the open neighbourhood V is needed to
ensure that the domain of the multiplication map in H remains open.) One

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

111

also easily verifies that if : U H is a local homomorphism from G to

H for some open neighbourhood U of the identity in G, then ker() is a
normal sub-local group of U , and hence of G. Note that the kernel of a local
morphism is well-defined up to local identity. If H is Hausdorff, then the
kernel ker() will also be closed.
As observed by Goldbring [Go2010], one can define the operation of
quotienting a local group by a normal sub-local group, provided that one
restricts to a sufficiently small neighbourhood of the origin:
Exercise 1.5.17 (Quotient spaces). Let G be a cancellative local group, and
let H be a normal sub-local group with normalising neighbourhood V . Let
W be a symmetric open neighbourhood of the identity such that W 6 V .
Show that there exists a cancellative local group W/H and a surjective
continuous homomorphism : W W/H such that, for any g, h W , one
has (g) = (h) if and only if gh1 H, and for any E W/H, one has E
open if and only if 1 (E) is open.
It is not difficult to show that the quotient W/H defined by the above
exercise is unique up to local isomorphism, so we will abuse notation and
talk about the quotient space W/H given by the above construction.
Example 1.5.22. Let G be the additive local group G := (2, 2)2 , and let H
be the sub-local group H := {0} (1, 1), with normalising neighbourhood
V := (1, 1)2 . If we then set W := (0.1, 0.1)2 , then the hypotheses of
Exercise 1.5.17 are obeyed, and W/H can be identified with (0.1, 0.1),
with the projection map : (x, y) 7 x.
Example 1.5.23. Let G be the torus (R/Z)2 , and let H be the sub-local
group H = {(x, x) mod Z2 : x (0.1, 0.1)}, where 0 < < 1 is an irrational number, with normalising neighbourhood (0.1, 0.1)2 mod Z2 . Set
W := (0.01, 0.01)2 mod Z2 . Then the hypotheses of Exercise 1.5.17 are
again obeyed, and W/H can be identified with the interval I := (0.01(1 +
), 0.01(1 + )), with the projection map : (x, y) mod Z2 7 y x for
(x, y) (0.01, 0.01)2 . Note, in contrast, that if one quotiented G by the
global group hHi = {(x, x) mod Z2 : x R} generated by H, the quotient
would be a non-Hausdorff space (and would also contain a dense set of torsion points, in contrast to the interval I which is locally torsion free). It is
because of this pathological behaviour of quotienting by global groups that
we need to work with local group quotients instead.
We can now state the local version of the Gleason-Yamabe theorem, first
proven in [Go2009] (and by a slightly different method in [vdDrGo2010]):
different method:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

112

1. Hilberts fifth problem

Theorem 1.5.24 (Local Gleason-Yamabe theorem). Let G be a locally compact local group. Then there exists an open symmetric neighbourhood G0 of
the identity, and a compact global group H in G0 that is normalised by G0 ,
such that G0 /H is well-defined and isomorphic to a local Lie group.
The proofs of this theorem by Goldbring and Goldbring-van den Dries
were phrased in the language of nonstandard analysis. However, it is possible
to translate those arguments to standard analysis arguments, which closely
follow the arguments given in previous sections11. We briefly sketch the
main points here.
As in the global case, the route to obtaining (local) Lie structure is via
Gleason metrics. On a local group G, we define a local Gleason metric to be
a metric d : U U R+ defined on some symmetric open neighbourhood
U of the identity with (say) U 100 well-defined (to avoid technical issues),
which generates the topology of U , and which obeys the following version
of the left-invariance, escape and commutator properties:
(1) (Left-invariance) If g, h, k U are such that gh, gk U , then
d(h, k) = d(gh, gk).
(2) (Escape property) If g U and nkgk
defined in U and kg n k C1 nkgk.

1
C,

then g, . . . , g n are well-

(3) (Commutator estimate) If g, h U are such that kgk, khk

then [g, h] is well-defined in U and (1.30) holds.

1
C,

One can then verify (by localisation of the arguments in Section 1.3)
that any locally compact local Lie group with a local Gleason metric is locally Lie (i.e. some neighbourhood of the identity is isomorphic to a local
Lie group); see Exercise 1.3.11. Next, one can define the notion of a weak
local Gleason metric by dropping the commutator estimate, and one can
verify an analogue of Theorem 1.5.5, namely that any weak local Gleason
metric is automatically a local Gleason metric, after possibly shrinking the
neighbourhood U and adjusting the constant C as necessary. The proof of
this statement is essentially the same as that in Theorem 1.5.5 (which is
already localised to small neighbourhoods of the identity), but uses a local
Haar measure instead of a global Haar measure, and requires some preliminary shrinking of the neighbourhood U to ensure that all group-theoretic
operations (and convolutions) are well-defined. We omit the (rather tedious)
details.
Now we define the concept of an NSS local group as a local group which
has an open neighbourhood of the identity that contains no non-trivial global
11Actually, our arguments are not a verbatim translation of those in Goldbring and
Goldbring-van den Dries, as we have made a few simplifications in which the role of Gleason
metrics is much more strongly emphasised.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.5. The Gleason-Yamabe theorem

113

subgroups. The proof of Theorem 1.5.7 is already localised to small neighbourhoods of the identity, and it is possible (after being sufficiently careful
with the notation) to translate that argument to the local setting, and conclude that any NSS local group admits a weak Gleason metric on some open
neighbourhood of the identity, and is hence locally Lie. (A typical example
of being sufficiently careful with the notation: to define the escape norm
(1.52), one adopts the convention that a statement such as g, . . . , g n U
is automatically false if g, . . . , g n are not all well-defined. The induction
hypothesis (1.54) will play a key role in ensuring that all expressions involved are well-defined and localised to a suitably small neighbourhood of
the identity.) Again, we omit the details.
The next step is to obtain a local version of Proposition 1.5.12. Here we
encounter a slight difficulty because in a general local group G, we do not
have a good notion of the group hAi generated by a set A of generators in G.
As such, the subgroup trapping property does not automatically translate
to the local group setting as defined in Definition 1.5.10. However, this
difficulty can be easily avoided by rewording the definition:
Definition 1.5.25 (Subgroup trapping). A local group has the subgroup
trapping property if, for every open neighbourhood U of the identity, there
exists another open neighbourhood V of the identity such that Q[V ] is contained in a global subgroup H that is in turn contained in U . (Here, Q[V ]
is, as before, the union of all the global subgroups contained in V .)
Because Q[V ] is now contained in a global group H, the group hQ[V ]i
generated by H is well-defined. As H is in the open neighbourhood U , one
can then also form the closure hQ[V ]i; if we choose U small enough to be
precompact, then this is a compact global group (and thus describable by the
Gleason-Yamabe theorem for such groups, Theorem 1.4.14). Because of this,
it is possible to adapt Proposition 1.5.12 without much difficulty to the local
setting to conclude that given any locally compact local group G with the
subgroup trapping property, there exists an open symmetric neighbourhood
G0 of the identity, and a compact global group H in G0 that is normalised
by G0 , such that G0 /H is well-defined and NSS (and thus locally isomorphic
to a local Lie group).
Finally, to finish the proof of Theorem 1.5.24, one has to establish the
analogue of Proposition 1.5.13, namely that one has to show that every
locally compact metrisable local group has the subgroup trapping property.
(It is not difficult to adapt Exercise 1.5.4 to the local group setting to reduce
to the metrisable case.) The first step is to prove the local group analogue of
Proposition 1.5.14 (again adopting the obvious convention that a statement
such as Qn U is only considered true if Qn is well-defined, and adding the
additional hypothesis that U is sufficiently small in order to ensure that all

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

114

1. Hilberts fifth problem

manipulations are justified). This can be done by a routine modification of

the proof. But then one can modify the rest of the argument in Proposition
1.5.13 to hold in the local setting as well (note, as in the proof of Proposition
1.5.12, that the compact set H generated in the course of this argument
remains a global group rather than a local one, and so one can again use
Theorem 1.4.14 without difficulty). Again, we omit the details.

1.6. The structure of locally compact groups

In the previous sections, we established the Gleason-Yamabe theorem, Theorem 1.1.17. Roughly speaking, this theorem asserts the mesoscopic structure of a locally compact group (after restricting to an open subgroup G0 to
remove the macroscopic structure, and quotienting out by K to remove the
microscopic structure) is always of Lie type.
In this section, we combine the Gleason-Yamabe theorem with some
additional tools from point-set topology to improve the description of locally
compact groups in various situations.
We first record some easy special cases of this. If the locally compact
group G has the no small subgroups property, then one can take K to be
trivial; thus G0 is Lie, which implies that G is locally Lie and thus Lie as well.
Thus the assertion that all locally compact NSS groups are Lie (Theorem
1.5.7) is a special case of the Gleason-Yamabe theorem.
In a similar spirit, if the locally compact group G is connected, then the
only open subgroup G0 of G is the full group G; in particular, by arguing
as in the treatment of the compact case (Exercise 1.4.21), we conclude that
any connected locally compact Hausdorff group is the inverse limit of Lie
groups.
Now we return to the general case, in which G need not be connected or
NSS. One slight defect of Theorem 1.1.17 is that the group G0 can depend
on the open neighbourhood U . However, by using a basic result from the
theory of totally disconnected groups known as van Dantzigs theorem, one
can make G0 independent of U :
Theorem 1.6.1 (Gleason-Yamabe theorem, stronger version). Let G be a
locally compact group. Then there exists an open subgroup G0 of G such that,
for any open neighbourhood U of the identity in G0 , there exists a compact
normal subgroup K of G0 in U such that G0 /K is isomorphic to a Lie group.
We will prove this theorem later in this section. As in previous sections,
if G is Hausdorff, the group G0 is thus an inverse limit of Lie groups (and if
G (and hence G0 ) is first countable, it is the inverse limit of a sequence of
Lie groups).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.6. The structure of locally compact groups

115

Exercise 1.6.1. By working with the locally compact group

1 Qp Qp Qp
0 1 Qp Qp
,
G=
0 0
1 Qp
0 0
0
1
where Qp is a p-adic group, show that one cannot demand in Theorem 1.6.1
that the open subgroup G0 be normal.
It remains to analyse inverse limits of Lie groups. To do this, it helps
to have some control on the dimensions of the Lie groups involved. A basic
tool for this purpose is the invariance of domain theorem:
Theorem 1.6.2 (Brouwer invariance of domain theorem). Let U be an open
subset of Rn , and let f : U Rn be a continuous injective map. Then f (U )
is also open.
We prove this theorem later in this section also. It has an important
corollary:
Corollary 1.6.3 (Topological invariance of dimension). If n > m, and U
is a non-empty open subset of Rn , then there is no continuous injective
mapping from U to Rm . In particular, Rn and Rm are not homeomorphic.
Exercise 1.6.2 (Uniqueness of dimension). Let X be a non-empty topological space. If X is a manifold of dimension d1 , and also a manifold of
dimension d2 , show that d1 = d2 . Thus, we may define the dimension
dim(X) of a non-empty manifold in a well-defined manner.
If X, Y are non-empty manifolds, and there is a continuous injection
from X to Y , show that dim(X) dim(Y ).
Remark 1.6.4. Note that the analogue of the above exercise for surjections is false: the existence of a continuous surjection from one non-empty
manifold X to another Y does not imply that dim(X) dim(Y ), thanks to
the existence of space-filling curves. Thus we see that invariance of domain,
while intuitively plausible, is not an entirely trivial observation.
As we shall see, we can use Corollary 1.6.3 to bound the dimension of
the Lie groups Ln in an inverse limit G = limn Ln by the dimension
of the inverse limit G. Among other things, this can be used to obtain a
positive resolution to Hilberts fifth problem, Theorem 1.1.13. Again, this
theorem will be proven later in this section.
Another application of this machinery is the following variant of Hilberts
fifth problem, which was used in Gromovs original proof of Gromovs theorem on groups of polynomial growth, although we will not actually need it
here:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

116

1. Hilberts fifth problem

Proposition 1.6.5. Let G be a locally compact -compact group that acts

transitively, faithfully, and continuously on a connected topological manifold
X. Then G is isomorphic to a Lie group.
Recall that a continuous action of a topological group G on a topological
space X is a continuous map : G X X which obeys the associativity
law (gh)x = g(hx) for g, h G and x X, and the identity law 1x = x for
all x X. The action is transitive if, for every x, y X, there is a g G
with gx = y, and faithful if, whenever g, h G are distinct, one has gx 6= hx
for at least one x.
The -compact hypothesis is a technical one, and can likely be dropped,
but we retain it for this discussion (as in most applications we can reduce
to this case).
Exercise 1.6.3. Show that Proposition 1.6.5 implies Theorem 1.1.13.
Remark 1.6.6. It is conjectured that the transitivity hypothesis in Proposition 1.6.5 can be dropped; this is known as the Hilbert-Smith conjecture. It
remains open, with the key remaining difficulty being the need to figure out
a way to eliminate the possibility that G is a p-adic group Zp . See Section
2.7 for more discussion.
1.6.1. Van Dantzigs theorem. Recall that a (non-empty) topological
space X is connected if the only clopen (i.e. closed and open) subsets of
X are the whole space X and the empty set ; a non-empty topological
space is disconnected if it is not connected. (By convention, the empty set is
considered to be neither connected nor disconnected, somewhat analogously
to how the natural number 1 is neither considered prime nor composite.)
At the opposite extreme to connectedness is the property of being a
totally disconnected space. This is a space whose only connected subsets are
the singleton sets. Typical examples of totally disconnected spaces include
discrete spaces (e.g. the integers Z with the discrete topology) and Cantor
spaces (such as the standard Cantor set).
Most topological spaces are neither connected nor totally disconnected,
but some intermediate combination of both. In the case of topological groups
G, this rather vague assertion can be formalised as follows.
Exercise 1.6.4.
(i) Define a connected component of a topological
space X to be a maxmial connected set. Show that the connected
components of X form a partition of X, thus every point in X
belongs to exactly one connected component.
(ii) Let G be a topological group, and let G be the connected component of the identity. Show that G is a closed normal subgroup of

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.6. The structure of locally compact groups

117

G, and that G/G is a totally disconnected subgroup of G. Thus,

one has a short exact sequence
0 G G G/G 0
of topological groups that describes G as an extension of a totally
disconnected group by a connected group.
(iii) Conversely, if one has a short exact sequence
0HGK0
of topological groups, with H connected and K totally disconnected, show that H is isomorphic to G , and K is isomorphic
to G/G .
(iv) If G is locally compact, show that G and G/G are also locally
compact.
In principle at least, the study of locally compact groups thus splits into
the study of connected locally compact groups, and the study of totally
disconnected locally compact groups. Note however that even if one has
a complete understanding of the factors H, K of a short exact sequence
0 H G K 0, it may still be a non-trivial issue to fully understand
the combined group G, due to the possible presence of non-trivial group
cohomology. See for instance [Ta2011c, 2.4] for more discussion.
For totally disconnected locally compact groups, one has the following
fundamental theorem of van Dantzig [vDa1936]:
Theorem 1.6.7 (Van Dantzigs theorem). Every totally disconnected locally
compact group G contains a compact open subgroup H (which will of course
still be totally disconnected).
Example 1.6.8. Let p be a prime. Then the p-adic field Qp (with the
usual p-adic valuation) is totally disconnected locally compact, and the padic integers Zp are a compact open subgroup.
Of course, this situation is the polar opposite of what occurs in the
connected case, in which the only open subgroup is the whole group.
To prove van Dantzigs theorem, we first need a lemma from point
set topology, which shows that totally disconnected spaces contain enough
clopen sets to separate points:
Lemma 1.6.9. Let X be a totally disconnected compact Hausdorff space,
and let x, y be distinct points in X. Then there exists a clopen set that
contains x but not y.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

118

1. Hilberts fifth problem

Proof. Let K be the intersection of all the clopen sets that contain x (note
that X is obviously clopen). Clearly K is closed and contains x. Our objective is to show that K consists solely of {x}. As X is totally disconnected,
it will suffice to show that K is connected.
Suppose this is not the case, then we can split K = K1 K2 where
K1 , K2 are disjoint non-empty closed sets; without loss of generality, we may
assume that x lies in K1 . As all compact Hausdorff spaces are normal, we can
thus enclose K1 , K2 in disjoint open subsets U1 , U2 of X. In particular, the
topological boundary U2 is compact and lies outside of K. By definition of
K, we thus see that for every y U2 , we can find a clopen neighbourhood of
x that avoids y; by compactness of U2 (and the fact that finite intersections
of clopen sets are clopen), we can thus find a clopen neighbourhood L of x
that is disjoint from U2 . One then verifies that L\U2 = L\U2 is a clopen
neighbourhood of x that is disjoint from K2 , contradicting the definition of
K, and the claim follows.

Now we can prove van Dantzigs theorem. We will use an argument
from [HeRo1979]. Let G be totally disconnected locally compact (and thus
Hausdorff). Then we can find a compact neighbourhood K of the identity.
By Lemma 1.6.9, for every y K, we can find a clopen neighbourhood
of the identity that avoids y; by compactness of K, we may thus find a
clopen neighbourhood of the identity that avoids K. By intersecting this
neighbourhood with K, we may thus find a compact clopen neighbourhood F
of the identity. As F is both compact and open, we may then the continuity
of the group operations find a symmetric neighbourhood U of the identity
such that U F F . In particular, if we let G0 be the group generated by U ,
then G0 is an open subgroup of G contained in F and is thus compact as
required.
Remark 1.6.10. The same argument shows that a totally disconnected
locally compact group contains arbitrarily small compact open subgroups,
or in other words the compact open subgroups form a neighbourhood base
for the identity.
In view of van Dantzigs theorem, we see that the local behaviour of
totally disconnected locally compact groups can be modeled by the compact
totally disconnected groups, which are better understood. Thanks to the
Gleason-Yamabe theorem for compact groups, such groups are the inverse
limits of compact totally disconnected Lie groups. But it is easy to see that
a compact totally disconnected Lie group must be finite, and so compact
totally disconnected groups are necessarily profinite (i.e. the inverse limit
of finite groups). The global behaviour however remains more complicated,
in part because the compact open subgroup given by van Dantzigs theorem

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.6. The structure of locally compact groups

119

need not be normal, and so does not necessarily induce a splitting of G into
compact and discrete factors.
Example 1.6.11. Let p be a prime, and let G be the semi-direct product
Z n Qp , where the integers Z act on Qp by the map m : x 7 pm x, and we
give G the product of the discrete topology of Z and the p-adic topology
on Qp . One easily verifies that G is a totally disconnected locally compact
group. It certainly has compact open subgroups, such as {0}Zp . However,
it is easy to show that G has no non-trivial compact normal subgroups (the
problem is that the conjugation action of Z on Qp has all non-trivial orbits
unbounded).
We can pull van Dantzigs theorem back to more general locally compact
groups:
Exercise 1.6.5. Let G be a locally compact group.
(i) Show that G contains an open subgroup G0 which is compact-byconnected in the sense that G0 /(G0 ) is compact. (Hint: apply
van Dantzigs theorem to G/G .)
(ii) If G is compact-by-connected, and U is an open neighbourhood of
the identity, show that there exists a compact subgroup K of G in U
such that G/K is isomorphic to a Lie group. (Hint: use Theorem
1.1.17, and observe that any open subgroup of the compact-byconnected group G has finite index and thus has only finitely many
conjugates.) Conclude Theorem 1.6.1.
(iii) Show that any locally compact Hausdorff group G contains an open
subgroup G0 that is isomorphic to an inverse limit of Lie groups
(L )A , which each Lie group L has at most finitely many connected components. Furthermore, each L is isomorphic to G0 /K
for some compact normal subgroup K of G0 , with K K for
< . If G is first countable, show that this inverse limit can be
taken to be a sequence (so that the index set A is simply the natural numbers N with the usual ordering), and the Kn then shrink to
zero in the sense that they lie inside any given open neighbourhood
of the identity for n large enough.
Exercise 1.6.6. Let G be a totally disconnected locally compact group.
Show that every compact subgroup K of G is contained in a compact open
subgroup. (Hint: van Dantzigs theorem provides a compact open subgroup,
but it need not contain K. But is there a way to modify it so that it is
normalised by K? Why would being normalised by K be useful?)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

120

1. Hilberts fifth problem

1.6.2. The invariance of domain theorem. In this section we give a

proof of the invariance of domain theorem. The main topological tool for
this is Brouwers famous fixed point theorem:
Theorem 1.6.12 (Brouwer fixed point theorem). Let f : B n B n be
a continuous function on the unit ball B n := {x Rn : kxk 1} in a
Euclidean space Rn . Then f has at least one fixed point, thus there exists
x B n with f (x) = x.
This theorem has many proofs. We quickly sketch one of these proofs
(based on one from [Ku1998]) as follows:
Exercise 1.6.7. For this exercise, suppose for sake of contradiction that
Theorem 1.6.12 is false, thus there is a continuous map from B n to B n with
no fixed point.
(i) Show that there exists a smooth map from B n to B n with no fixed
point.
(ii) Show that there exists a smooth map from B n to the unit sphere
S n1 , which equals the identity function on S n1 .
(iii) Show that there exists a smooth map from B n to the unit sphere
x
on a neighbourhood of S n1 .
S n1 , which equals the map x 7 kxk
R
(iv) By computing the integral B n det(1 , . . . , n ) in two different
ways (one by using Stokes theorem, and the other by using the n
1-dimensional nature of the sphere S n1 ), establish a contradiction.
Now we prove Theorem 1.6.2. By rescaling and translation invariance,
it will suffice to show the following claim:
Theorem 1.6.13 (Invariance of domain, again). Let f : B n Rn be an
continuous injective map. Then f (0) lies in the interior of f (B n ).
Let f be as in Theorem 1.6.13. The map f : B n f (B n ) is a continuous
bijection between compact Hausdorff spaces and is thus a homeomorphism.
In particular, the inverse map f 1 : f (B n ) B n is continuous. Using the
Tietze extension theorem, we can find a continuous function G : Rn Rn
that extends f 1 .
The function G has a zero on f (B n ), namely at f (0). We can use the
Brouwer fixed point theorem to show that this zero is stable:
: f (B n ) Rn be a continuous
Lemma 1.6.14 (Stability of zero). Let G

has at
function such that kG(y) G(y)k
1 for all y f (B n ). Then G
n

least one zero (i.e. there is a y f (B ) such that G(y) = 0).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.6. The structure of locally compact groups

121

Proof. Apply Theorem 1.6.12 to the function

(x)) = G(f (x)) G(f
(x)).
x 7 x G(f

Now suppose that Theorem 1.6.13 failed, so that f (0) is not an interior
point of f (B n ). We will use this to locate a small perturbation of G that no
longer has a zero on f (B n ), contradicting Lemma 1.6.14.
We turn to the details. Let > 0 be a small number. By continuity of
G, we see (if is chosen small enough) that we have kG(y)k 0.1 whenever
y Rn and ky f (0)k 2.
On the other hand, since f (0) is not an interior point of f (B n ), there
exists a point c Rn with kc f (0)k < that lies outside f (B n ). By
translating f if necessary, we may take c = 0; thus f (B n ) avoids zero,
kf (0)k < , and we have
(1.63)

kG(y)k 0.1 whenever kyk .

Let denote the set := 1 2 , where

1 := {y f (B n ) : kyk }
and
2 := {y Rn : kyk = }.
By construction, is compact but does not contain f (0). Crucially, there
is a continuous map : f (B n ) defined by setting
(1.64)

(y) := max(

, 1)y.
kyk

Note that is continuous and well-defined since f (B n ) avoids zero. Informally, is a perturbation of f (B n ) caused by pushing f (B n ) out a small
distance away from the origin 0 (and hence also away from f (0)), with
being the pushing map.
By construction, G is non-zero on 1 ; since 1 is compact, G is bounded
from below on 1 by some > 0. By shrinking if necessary we may assume
that < 0.1.
By the Weierstrass approximation theorem, we can find a polynomial
P : Rn Rn such that
(1.65)

kP (y) G(y)k <

for all y ; in particular, P does not vanish on 1 . At present, it is

possible that P vanishes on 2 . But as P is smooth and 2 has measure

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

122

1. Hilberts fifth problem

zero, P (2 ) also has measure zero; so12 by shifting P by a small generic

constant we may assume without loss of generality that P also does not
vanish on 2 .
: f (B n ) Rn defined by
Now consider the function G

:= P ((y)).
G(y)
This is a continuous function that is never zero. From (1.65), (1.64) we have

kG(y) G(y)k
<
whenever y f (B n ) is such that kyk > . On the other hand, if kyk ,
then from (1.64), (1.63) we have
kG(y)k, kG((y))k 0.1
and hence by (1.65) and the triangle inequality

kG(y) G(y)k
0.2 + .
Thus in all cases we have

kG(y) G(y)k
0.2 + 0.3

for all y f (B n ). But this, combined with the non-vanishing nature of G,

contradicts Lemma 1.6.14.
1.6.3. Hilberts fifth problem. We now establish Theorem 1.1.13. Let
G be a locally Euclidean group. By Exercise 1.1.6, G is Hausdorff; it is
also locally compact and first countable. Thus, by Exercise 1.6.5, such a
group contains an open subgroup G0 which is isomorphic to the inverse
limit limn Ln of Lie groups Ln , each of which has only finitely many
components. Clearly, G0 is also locally Euclidean. If it is Lie, then G is
locally Lie and thus Lie, by Exercise 1.2.17. Thus, by replacing G with G0
if necessary, we may assume without loss of generality that G is the inverse
limit G = limn Ln , each of which has only finitely many components.
By Exercise 1.6.5, each Ln is isomorphic to the quotient of G by some
compact normal subgroup Kn with Kn+1 Kn . In particular, Ln is isomorphic to the quotient of Ln+1 by a compact normal subgroup Hn Kn /Kn+1 .
By Cartans theorem (Theorem 1.3.2), Hn is also a Lie group. Among other
things, this implies that the quotient homomorphism from the Lie algebra
ln+1 of Ln+1 to the Lie algebra ln of Ln is surjective; indeed, it is the quotient
map by the Lie algebra hn of Hn . This implies that there is a continuous
map from ln to ln+1 that inverts the quotient map; in other words, we have
12If one wishes, one can use an algebraic geometry argument here instead of a measuretheoretic one, noting that P (2 ) lies in an algebraic hypersurface and can thus be generically
avoided by perturbation. A purely topological way to avoid zeroes in 2 is also given in [Ku1998].

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.6. The structure of locally compact groups

123

a continuous map nn+1 : L(Ln ) L(Ln+1 ) from the one-parameter subgroups n : R Ln of Ln to the one-parameter subgroups n+1 : R Ln+1
of Ln+1 , such that nn+1 (n ) mod Hn = n for all n L(Ln ).
Exercise 1.6.8. By iterating these maps and passing to the inverse limit,
conclude that for each n N, there is a continuous map n : L(Ln ) L(G)
such that n (n ) mod Kn = n for all n L(Ln ).
Because Ln is a Lie group, the exponential map n 7 n (1) is a
homeomorphism from a neighbourhood of the origin in L(Ln ) to a neighbourhood of the identity in Ln . We can thus obtain a continuous map
n (1) 7 n (n )(1) from a neighbourhood of the identity in Ln to G. Since
n (n )(1) mod Kn = n (1), this map is injective.
Now we use the hypothesis that G is locally Euclidean (and in particular,
has a well-defined dimension dim(G)). By Exercise 1.6.2, we have
dim(Ln ) dim(G)
for all n. On the other hand, since each Ln is a quotient of the next Lie
group Ln+1 , one has
dim(Ln ) dim(Ln+1 ).
Since there are only finitely many possible values for the (necessarily integral) dimension dim(Ln ) between 0 and dim(G), we conclude that the
dimension must eventually stabilise, i.e. one has
dim(Ln ) = dim(Ln+1 )
for all sufficiently large n. By discarding the first few terms in the sequence
and relabeling, we may thus assume that the dimension is constant for all n.
Since Ln Ln+1 /Hn , this implies that the Lie groups Hn have dimension
zero for all n. As the Hn are also compact, they are thus finite. Thus each
Kn+1 is a finite extension of Kn . As Kn is the inverse limit of the Kn /Km
as m , we conclude that Kn is a profinite group, i.e. the inverse limit
of finite groups. In particular, Kn is totally disconnected.
We now study the short exact squence
0 Kn G Gn 0,
playing off the locally connected nature of the Lie group Gn against the
totally disconnected nature of Kn .
As discussed earlier, we have a continuous injective map n from a neighbourhood Un of the identity in Gn to G that partially inverts the quotient
map. By translation, we may normalise n (1) = 1. As Gn is locally connected, we can find a connected neighborhood Vn of the identity in Gn such
that (Vn Vn1 )2 Un .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

124

1. Hilberts fifth problem

Now consider the set {n (g)n (g 1 ) : g Vn }. On the one hand, this

set is contained in Kn and contains 1; on the other hand, it is connected.
As Kn is totally disconnected, this set must equal {1}, thus n (g 1 ) =
n (g)1 for all g Vn . A similar argument based on consideration of the
set {n (g)n (h)n (gh)1 : g, h Vn } shows that n (gh) = n (g)n (h) for
all g Vn . Thus n is a homomorphism from the local group Vn to G.
Finally, for any k Kn , a consideration of the set {n (g)kn (g)1 :
g Vn } reveals that n (g) commutes with Kn . As a consequence, we see
that the preimage n1 (Vn ) of Vn under the quotient map n : G Gn is
isomorphic as a local group to Vn Kn , after identifying n (g)k with (g, k)
for any g Vn and k Kn .
On the other hand, G is locally Euclidean, and hence Vn Kn is locally
Euclidean also, and in particular locally connected. This implies that Kn
is locally connected; but as Kn is also totally disconnected, it must be
discrete. This G is now locally isomorphic to Vn and hence to Gn , and is
thus locally Lie and hence Lie as required. (Here, we say that two groups
are locally isomorphic if they have neighbourhoods of the identity which are
isomorphic to each other as local groups.)
Exercise 1.6.9. Let G be a locally compact Hausdorff first-countable group
which is finite-dimensional in the sense that it does not contain continuous
injective images of non-trivial open sets of Euclidean spaces Rn of arbitrarily
large dimension. Show that G is locally isomorphic to the direct product L
K of a Lie group L and a totally disconnected compact group K. (Note that
this local isomorphism does not necessarily extend to a global isomorphism,
as the example of the solenoid group RZp /Z from Example 1.1.7 shows.)
Remark 1.6.15. Of course, it is possible for locally compact groups to
be infinite-dimensional; a simple example is the infinite-dimensional torus
(R/Z)N , which is compact, abelian, metrisable, and locally connected, but
infinite dimensional. (It will still be an inverse limit of Lie groups, though.)
Exercise 1.6.10. Show that a topological group is Lie if and only if it
is locally compact, Hausdorff, first-countable, locally connected, and finitedimensional.
Remark 1.6.16. It is interesting to note that this characterisation barely
uses the real numbers R, which are of course fundamental in defining the
smooth structure of a Lie group; the only remaining reference to R comes
through the notion of finite dimensionality. It is also possible, using dimension theory, to obtain alternate characterisations of finite dimensionality
(e.g. finite Lebesgue covering dimension) that avoid explicit mention of the
real line, thus capturing the concept of a Lie group using only the concepts
of point-set topology (and the concept of a group, of course).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.6. The structure of locally compact groups

125

1.6.4. Transitive actions. We now prove Proposition 1.6.5. As this is a

stronger statement than Theorem 1.1.13, it will not be surprising that we
will be using a very similar argument to prove the result.
Let G be a locally compact -compact group that acts transitively, faithfully, and continuously on a connected manifold X. The advantage of transitivity is that one can now view X as a homogeneous space X = G/H of
G, where H = Stab(x0 ) is the stabiliser of a point x0 (and is thus a closed
subgroup of G). Note that a priori, we only know that X and G/H are identifiable as sets, with the identification map : G/H X defined by setting
(gH) := gx0 being continuous; but thanks to the -compact hypothesis, we
can upgrade to a homeomorphism. Indeed, as G is -compact, G/H is also;
and so given any compact neighbourhood of the identity K in G, G/H can
be covered by countably many translates of KH/H. By the Baire category
theorem, one of these translates gK has an image (gKH/H) = gKx0 in X
with non-empty interior, which implies that K 1 Kx0 has x0 as an interior
point. From this it is not hard to see that the map is open; as it is also
a continuous bijection, it is therefore a homeomorphism. (This argument
dates back at least to [Fr1936].)
By the Gleason-Yamabe theorem (Theorem 1.6.1), G has an open subgroup G0 that is the inverse limit of Lie groups. (Note that G is Hausdorff
because it acts faithfully on the Hausdorff space X.) G0 acts transitively on
G0 H/H, which is an open subset of G/H and thus also a manifold. Thus,
we may assume without loss of generality that G is itself the inverse limit
of Lie groups.
As G is -compact, the manifold X is also. As G acts faithfully on X,
this makes G first countable; and so (by Exercise 1.6.5) G is the inverse
limit of a sequence of Lie groups Gn = G/Kn , with each Gn+1 projecting
surjectively onto Gn , and with the Kn shrinking to the identity.
Let Hn be the projection of H onto Gn ; this is a closed subgroup of the
Lie group Gn , and each Hn+1 projects surjectively onto Hn . Then Gn /Hn
are manifolds, and G/H is the inverse limit of the Gn /Hn = G/HKn .
Exercise 1.6.11. Show that the dimensions of the Gn /Hn are be nondecreasing, and bounded above by the dimension of G/H. (Hint: repeat the
arguments of the previous section. The Hn need no longer be compact, but
they are still closed, and this still suffices to make the preceding arguments
go through.)
Thus, for n large enough, the dimensions of Gn /Hn must be constant; by
renumbering, we may assume that all the Gn /Hn have the same dimension.
As each Gn+1 /Hn+1 is a cover of Gn /Hn with structure group Kn /Kn+1 ,
we conclude that the Kn /Kn+1 are zero-dimensional and compact, and thus

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

126

1. Hilberts fifth problem

finite. On the other hand, G/H is locally connected, which implies that
the Kn /Kn+1 are eventually trivial. Indeed, if we pick a simply connected
neighbbourhood U1 of the identity in G1 /H1 , then by local connectedness
of G/H, there exists a connected neighbourhood U of the identity in G/H
whose projection to G1 /H1 is contained in U1 . Being open, U must contain
one of the Kn . If Kn /Km is non-trivial for any m > n, then the projection of
U to Gm /Hm will then be disconnected (as this projection will be contained
in a neighbourhood with the topological structure of U K1 /Km , and its
intersection with the latter fibre is at least as large as Kn /Km . We conclude
that Kn is trivial for n large enough, and so G = Gn is a Lie group as
required.
Note that while the manifold X in Proposition 1.6.5 is initially only
required a priori to be a topological manifold, it automatically acquires a
smooth structure also:
Exercise 1.6.12. Let G and X be as in Proposition 1.6.5. Show that one
can endow X with the structure of a smooth manifold, such that the action
of G is also smooth. (Hint: apply Cartans theorem to the stabiliser of a
point in X.)

1.7. Ultraproducts as a bridge between hard analysis and

soft analysis
Roughly speaking, mathematical analysis can be divided into two major
styles, namely hard analysis and soft analysis. The precise distinction between the two types of analysis is imprecise (and in some cases one may use
a blend the two styles), but some key differences can be listed as follows.
(1) Hard analysis tends to be concerned with quantitative or effective
properties such as estimates, upper and lower bounds, convergence
rates, and growth rates or decay rates. In contrast, soft analysis
tends to be concerned with qualitative or ineffective properties such
as existence and uniqueness, finiteness, measurability, continuity,
differentiability, connectedness, or compactness.
(2) Hard analysis tends to be focused on finitary, finite-dimensional or
discrete objects, such as finite sets, finitely generated groups, finite
Boolean combination of boxes or balls, or finite-complexity functions, such as polynomials or functions on a finite set. In contrast,
soft analysis tends to be focused on infinitary, infinite-dimensional,
or continuous objects, such as arbitrary measurable sets or measurable functions, or abstract locally compact groups.
(3) Hard analysis tends to involve explicit use of many parameters such
as , , N , etc. In contrast, soft analysis tends to rely instead on

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

127

properties such as continuity, differentiability, compactness, etc.,

which implicitly are defined using a similar set of parameters, but
whose parameters often do not make an explicit appearance in arguments.
(4) In hard analysis, it is often the case that a key lemma in the literature is not quite optimised for the application at hand, and one
has to reprove a slight variant of that lemma (using a variant of
the proof of the original lemma) in order for it to be suitable for
applications. In contrast, in soft analysis, key results can often
be used as black boxes, without need of further modification or
inspection of the proof.
(5) The properties in soft analysis tend to enjoy precise closure properties; for instance, the composition or linear combination of continuous functions is again continuous, and similarly for measurability,
differentiability, etc. In contrast, the closure properties in hard
analysis tend to be fuzzier, in that the parameters in the conclusion are often different from the parameters in the hypotheses. For
instance, the composition of two Lipschitz functions with Lipschitz
constant K is still Lipschitz, but now with Lipschitz constant K 2
instead of K. These changes in parameters mean that hard analysis arguments often require more bookkeeping than their soft
analysis counterparts, and are less able to utilise algebraic constructions (e.g. quotient space constructions) that rely heavily on
precise closure properties.
In the text so far, focusing on the theory surrounding Hilberts fifth
problem, the results and techniques have fallen well inside the category of
soft analysis. However, we will now turn to the theory of approximate
groups, which is a topic which is traditionally studied using the methods
of hard analysis. (Later we will also study groups of polynomial growth,
which lies on an intermediate position in the spectrum between hard and
soft analysis, and which can be profitably analysed using both styles of
analysis.)
Despite the superficial differences between hard and soft analysis, though,
there are a number of important correspondences between results in hard
analysis and results in soft analysis. For instance, if one has some sort of
uniform quantitative bound on some expression relating to finitary objects,
one can often use limiting arguments to then conclude a qualitative bound
on analogous expressions on infinitary objects, by viewing the latter objects as some sort of limit of the former objects. Conversely, if one has a
qualitative bound on infinitary objects, one can often use compactness and

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

128

1. Hilberts fifth problem

contradiction arguments to recover uniform quantitative bounds on finitary

objects as a corollary.
Remark 1.7.1. Another type of correspondence between hard analysis and
soft analysis, which is syntactical rather than semantical in nature,
arises by taking the proofs of a soft analysis result, and translating such
a qualitative proof somehow (e.g. by carefully manipulating quantifiers)
into a quantitative proof of an analogous hard analysis result. This type
of technique is sometimes referred to as proof mining in the proof theory
literature, and is discussed in [Ta2008, 1.3]. We will however not employ
systematic proof mining techniques here, although in later sections we will
informally borrow arguments from infinitary settings (such as the methods
used to construct Gleason metrics) and adapt them to finitary ones.
Let us illustrate the correspondence between hard and soft analysis results with a simple example.
Proposition 1.7.2. Let X be a sequentially compact topological space, let
S be a dense subset of X, and let f : X [0, +] be a continuous function
(giving the extended half-line [0, +] the usual order topology). Then the
following statements are equivalent:
(i) (Qualitative bound on infinitary objects) For all x X, one has
f (x) < +.
(ii) (Quantitative bound on finitary objects) There exists M < +
such that f (x) M for all x S.
In applications, S is typically a (non-compact) set of finitary (or finite
complexity) objects of a certain class, and X is some sort of completion
or compactification of S which admits additional infinitary objects that
may be viewed as limits of finitary objects.
Proof. To see that (ii) implies (i), observe from density that every point x
in X is adherent to S, and so given any neighbourhood U of x, there exists
y S U . Since f (y) M , we conclude from the continuity of f that
f (x) M also, and the claim follows.
Conversely, to show that (i) implies (ii), we use the compactness and
contradiction argument. Suppose for sake of contradiction that (ii) failed.
Then for any natural number n, there exists13 xn S such that f (xn ) n.
Using sequential compactness, and passing to a subsequence if necessary, we
may assume that the xn converge to a limit x X. By continuity of f , this
implies that f (x) = +, contradicting (i).

13Here we have used the axiom of choice, which we will assume throughout the text.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

129

Remark 1.7.3. Note that the above deduction of (ii) from (i) is ineffective
in that it gives no explicit bound on the uniform bound M in (ii). Without
any further information on how the qualitative bound (i) is proven, this
is the best one can do in general (and this is one of the most significant
weaknesses of infinitary methods when used to solve finitary problems); but
if one has access to the proof of (i), one can often finitise or proof mine that
argument to extract an effective bound for M , although often the bound
one obtains in the process is quite poor (particularly if the proof of (i) relied
extensively on infinitary tools, such as limits). See [Ta2008, 1.2] for some
related discussion.
The above simple example illustrates that in order to get from an infinitary statement such as (i) to a finitary statement such as (ii), a key step
is to be able to take a sequence (xn )nN (or in some cases, a more general
net (x )A ) of finitary objects and extract a suitable infinitary limit object
x. In the literature, there are three main ways in which one can extract such
a limit:
(1) (Topological limit) If the xn are all elements of some topological
space S (e.g. an incomplete function space) which has a suitable
compactification or completion X (e.g. a Banach space), then
(after passing to a subsequence if necessary) one can often ensure
the xn converge in a topological sense (or in a metrical sense) to
a limit x. The use of this type of limit to pass between quantitative/finitary and qualitative/infinitary results is particularly common in the more analytical areas of mathematics (such as ergodic
theory, asymptotic combinatorics, or PDE), due to the abundance
of useful compactness results in analysis such as the (sequential)
Banach-Alaoglu theorem, Prokhorovs theorem, the Helly selection
theorem, the Arzel
a-Ascoli theorem, or even the humble BolzanoWeierstrass theorem. However, one often has to take care with the
nature of convergence, as many compactness theorems only guarantee convergence in a weak sense rather than in a strong one.
(2) (Categorical limit) If the xn are all objects in some category (e.g.
metric spaces, groups, fields, etc.) with a number of morphisms between the xn (e.g. morphisms from xn+1 to xn , or vice versa), then
one can often form a direct limit lim xn or inverse limit lim xn of
these objects to form a limiting object x. The use of these types of
limits to connect quantitative and qualitative results is common in
subjects such as algebraic geometry that are particularly amenable
to categorical ways of thinking. (We already have seen inverse limits appear in the discussion of Hilberts fifth problem, although in

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

130

1. Hilberts fifth problem

that context they were not really used to connect quantitative and
qualitative results together.)
(3) (Logical limit) If the xn are all distinct spaces (or elements or subsets of distinct spaces), with few morphisms connecting them together, then topological and categorical limits are often unavailable
or unhelpful. In such cases, however, one can still tie together such
objects using an ultraproduct construction (or similar
Q device) to
create a limiting object limn xn or limiting space n xn that
is a logical limit of the xn , in the sense that various properties of
the xn (particularly those that can be phrased using the language
of first-order logic) are preserved in the limit. As such, logical limits are often very well suited for the task of connecting finitary and
infinitary mathematics together. Ultralimit type constructions are
of course used extensively in logic (particularly in model theory),
but are also popular in metric geometry. They can also be used in
many of the previously mentioned areas of mathematics, such as
algebraic geometry (as discussed in [Ta2011b, 2.1]).
The three types of limits are analogous in many ways, with a number
of connections between them. For instance, in the study of groups of polynomial growth, both topological limits (using the metric notion of GromovHausdorff convergence) and logical limits (using the ultralimit construction)
are commonly used, and to some extent the two constructions are at least
partially interchangeable in this setting. (See also [Ta2011c, 4.4-4.5] for
the use of ultralimits as a substitute for topological limits.) In the theory of
approximate groups, though, it was observed by Hrushovski [Hr2012] that
logical limits (and in particular, ultraproducts) are the most useful type of
limit to connect finitary approximate groups to their infinitary counterparts.
One reason for this is that one is often interested in obtaining results on approximate groups A that are uniform in the choice of ambient group G.
As such, one often seeks to take a limit of approximate groups An that lie
in completely unrelated ambient groups Gn , with no obvious morphisms or
metrics tying the Gn to each other. As such, the topological and categorical
limits are not easily usable, whereas the logical limits can still be employed
without much difficulty.
Logical limits are closely tied with non-standard analysis. Indeed, by
applying an ultraproduct construction to standard number systems such as
the natural numbers N or the reals R, one can obtain nonstandard number
systems such as the nonstandard natural numbers N or the nonstandard
real numbers (or hyperreals) R. These nonstandard number systems behave
very similarly to their standard counterparts, but also enjoy the advantage
of containing the standard number systems as proper subsystems (e.g. R is

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

131

a subring of R), which allows for some convenient algebraic manipulations

(such as the quotient space construction to create spaces such as R/R)
which are not easily accessible in the purely standard universe. Nonstandard spaces also enjoy a useful completeness property, known as countable
saturation (or more precisely, 1 -saturation), which is analogous to metric
completeness (as discussed in [Ta2011c, 4.4]) and which will be particularly useful for us in tying together the theory of approximate groups with
the theory of Hilberts fifth problem. See [Ta2008, 1.5] for more discussion
on ultrafilters and nonstandard analysis.
In this section, we lay out the basic theory of ultraproducts and ultralimits (in particular, proving Loss theorem, which roughly speaking asserts
that ultralimits are limits in a logical sense, as well as the countable saturation (or more precisely, 1 -saturation) property alluded to earlier). We also
lay out some of the basic foundations of nonstandard analysis, although we
will not rely too heavily on nonstandard tools in this text. Finally, we apply
this general theory to approximate groups, to connect finite approximate
groups to an infinitary type of approximate group which we will call an ultra approximate group. We will then study these ultra approximate groups
(and models of such groups) in more detail in the next section.
Remark 1.7.4. Throughout this text, we will assume the axiom of choice,
in order to easily use ultrafilter-based tools. If one really wanted to expend
the effort, though, one could eliminate the axiom of choice from the proofs
of the final finitary results that one is ultimately interested in proving,
at the cost of making the proofs significantly lengthier. Indeed, there is a
general result of Godel [Go1938] that any result which can be stated in
the language of Peano arithmetic (which, roughly speaking, means that the
result is finitary in nature), and can be proven in set theory using the
axiom of choice (or more precisely, in the ZFC axiom system), can also be
proven in set theory without the axiom of choice (i.e. in the ZF system). As
this text is not focused on foundations, we shall simply assume the axiom
of choice henceforth to avoid further distraction by such issues.
1.7.1. Ultrafilters. The concept of an ultrafilter will be fundamental. We
will only need to consider ultrafilters on the natural numbers N, although
one can certainly consider ultrafilters on more general sets.
Definition 1.7.5 (Ultrafilter). An filter on the natural numbers is a
collection of sets of natural numbers obeying the following axioms:
(1) (Monotonicity) If E F N, and E , then F .
(2) (Intersection) If E, F , then E F .
(3) (Properness) 6 .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

132

1. Hilberts fifth problem

An ultrafilter on the natural numbers is a filter which obeys an additional

axiom:
(4) (Maximality) If E N, then exactly one of E and N\E lies in .
A nonprincipal ultrafilter (also known as a free ultrafilter ) is an ultrafilter
that obeys one final axiom:
(5) (Nonprincipality) No finite set belongs to . (Equivalently: any
cofinite set will belong to .)
Given a nonprincipal ultrafilter , we say that a set E N is -large if it
is contained in , and -small otherwise. A property P (n) of the natural
numbers n N is said to hold for all n sufficiently close to if it holds in
an -large set.
The most basic fact about nonprincipal ultrafilters is that they exist - a
result due to Tarski, and known as the ultrafilter lemma:
Exercise 1.7.1 (Ultrafilter lemma). Show that there exists at least one
nonprincipal ultrafilter. (Hint: first construct a nonprincipal filter, and
then use Zorns lemma to place it inside a maximal nonprincipal filter.)
Exercise 1.7.2. Call an ultrafilter principal if it is not nonprincipal. Show
that for any natural number n0 , the set {E N : n0 E} is a principal
ultrafilter, and conversely every principal ultrafilter is of this form. Thus,
the space of principal ultrafilters can be identified with N.
The space of all ultrafilters is denoted N, and so by the preceding exercise one can view N as a subset of N, with N\N denoting the nonprincipal
ultrafilters.
Ultrafilters can be interpreted in a number of different ways, as indicated
by the exercises below.
Exercise 1.7.3 (Ultrafilters as finitely additive probability measures). If
is an ultrafilter, show that the map : 2N {0, 1} that maps -large
subsets of N to 1 and -small subsets of N to 0, is a finitely additive probability measure (thus (E) + (F ) = (E F ) whenever E, F N are
disjoint). Conversely, show that every finitely additive probability measure
: 2N {0, 1} that takes values in {0, 1} arises in this manner.
In view of the above exercise, one may view -large as analogous to
full measure, and holding for all n sufficiently close to as analogous
to holding for almost all n, except that the measure is only finitely additive instead of countably additive, and so concepts such as -largeness are
only stable under finitely many intersections, rather than countably many
intersections.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

133

Exercise 1.7.4 (Ultrafilters as Stone-Cech compactification of N). We view

N
N as a subset of the power set 22 , and give it the induced topology from
N
the product topology on 22 .
(i) If is a nonprincipal ultrafilter, show that every neighbourhood of
in N intersects N in an -large set, and that every -large set
arises in this manner. (This explains the notation n sufficiently
close to .)
(ii) Show that this makes N a compactification of N in the sense that
it is compact Hausdorff and contains N as a dense subset.
(iii) Show that N is a universal compactification (or Stone-Cech compactification), thus if X is any other compactification of N, then
there is a unique continuous map f : N X that is the identity
on N.
(iv) If Y is any compact Hausdorff space, show that every function
f : N Y has a unique continuous extension f : N Y to the
space N.
Remark 1.7.6. More discussion on the Stone-Cech compactification can
be found in [Ta2010, 2.5]. The compact space N can also be endowed
with an interesting semigroup structure, which is of importance in Ramsey
theory and ergodic theory; see [Ta2009, 2.3] for more details.
Exercise 1.7.5 (Ultrafilters as Banach limits). Recall that a functional
: ` (N) C from the *-algebra ` (N) of bounded complex sequences
(an )nN to the complex numbers is a *-homomorphism if it is complex-linear
and obeys the conjugation law
((an )nN ) = ((an )nN )
and the homomorphism law
((an bn )nN ) = ((an )nN ) ((bn )nN )
for all bounded sequences an , bn , while mapping the constant sequence (1)nN
to 1.
(i) Show that if is a ultrafilter, show that there exists a unique *homomorphism lim : ` (N) C with the property that for
every E N, one has lim 1E (n) = 1 if and only if E is -large.
(ii) Conversely, show that all *-homomorphisms from ` (N) to C
arise in this fashion, and that a *-homomorphism arises from a
nonprincipal ultrafilter if and only if it extends the limit functional
(thus ((an )nN ) = limn an whenever an is convergent).
Remark 1.7.7. From the above exercise we see that N can be viewed as
the Gelfand spectrum of ` (N).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

134

1. Hilberts fifth problem

Exercise 1.7.6 (Ultrafilters and Arrows theorem). Let S be a finite set

consisting of at least three elements (representing candidates). Let O(S)
denote the space of all total orderings < on S. Define a voting system to
be a function V : O(S)N O(S) from a sequence (<n )nN of orderings
on S (representing the preferences of an infinite number of voters) to an
ordering V ((<n )nN ) on S. If is a nonprincipal ultrafilter, show that there
is a unique voting system V = V : O(S)N O(S) with the property that
for any distinct A, B S, one has AV ((<n )nN )B if and only if A <n B
for all n sufficiently close to . Show furthermore that this voting system
obeys the following axioms for all distinct candidates A, B and all orderings
<n O(S):
(Consensus) If A <n B for all n, then AV ((<n )nN )B.
(Non-dictatorship) There is no n0 with the property that AV ((<n
)nN )B holds if and only if A <n0 B holds for all choices of
A, B, <n .
(Independence of irrelevant alternatives) The truth value of AV ((<n
)nN )B depends only on the truth values of A <n B for all n. Thus,
if (<0n )nN is another sequence of orderings such that A <n B if
and only if A <0n B, then AV ((<n )nN )B holds if and only if
AV ((<0n )nN )B holds.
Conversely, show that any voting system obeying the above axioms arises
from an nonprincipal ultrafilter in this fashion. Thus we see that Arrows
theorem [Ar1950] does not hold for infinite sets due to the existence of
nonprincipal ultrafilters in this setting; or to put it another way, Arrows
theorem is equivalent to the assertion that finite sets have no nonprincipal
ultrafilters. (This connection between ultrafilters and Arrows theorem was
first observed in [KiSo1972]. See [Ta2008, 1.5] for more discussion of the
voting interpretation of an ultrafilter.)
As the above examples indicate, there are many different ways to view
ultrafilters. My personal preference (at least from the perspective of forming ultraproducts and doing nonstandard analysis) is to view a nonprincipal
ultrafilter as a consistent way to take limits of sequences that do not necessarily converge in the classical sense, as per Exercise 1.7.5; but other readers
may prefer to use one of the other interpretations of an ultrafilter instead.
1.7.2. Ultrapowers and ultralimits. We now turn to the formal definition of an ultrapower and ultralimit. In order to define these two terms, we
must first fix two objects:
A non-principal ultrafilter N\N; and

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

135

A standard universe U, which is some some set of objects, the

elements of which we refer to as standard objects. We refer to
subsets of the standard universe (i.e. sets of standard objects) as
standard spaces (or standard sets).
Informally, the standard universe is going to contain all the objects that
we are initially interested in studying. For instance, if we are doing arithmetic, the standard universe might be the natural numbers N. If we are
doing analysis, the standard universe might include the real numbers R, as
well as the set of subsets of R, the set of functions from R to R, and so
forth. If we are studying approximate groups in an ambient group G, the
standard universe might include the elements of G, subsets of G, and similar
such objects. We will place some more constraints on the standard universe
later, but for now, the only requirement we have is that this universe forms
a set. (In particular, the standard universe cannot be so huge to be a proper
class, such as the class of all sets, or the class of all groups, etc., although
in most applications outside of set theory, this is hardly a restriction.)
In practice, the standard universe will contain many familiar mathematical spaces, such as the natural numbers, the real numbers, and so forth.
To emphasise this, we will sometimes refer to the natural numbers N as the
standard natural numbers, the real numbers R as the standard real numbers R, and so forth. This is in order to distinguish such spaces from the
nonstandard natural numbers N, the nonstandard real numbers R, etc.,
which we now define.
Remark 1.7.8. For the applications in this text, the precise choice of nonprincipal ultrafilter will be completely irrelevant; one such ultrafilter will
be just as good for our purposes as another. Basically, we only use the
ultrafilter in order to fix a consistent way to make choices amongst any
sequence of options indexed by the natural numbers; but the precise way in
which these choices are made will not be important, so long as it is consistent (in the sense that the ultrafilter axioms are obeyed). There are other
applications of ultrafilters, however, in which some nonprincipal ultrafilters
are decidedly superior to others; see [Ta2009, 2.3] for an example of this
in dynamical systems.
Definition 1.7.9 (Ultralimits and ultraproducts). Two sequences (xn )nN ,
(yn )nN of standard objects xn , yn U are said to be equivalent if one has
xn = yn for all n sufficiently close to . This is easily seen to be an equivalence relation. We define the ultralimit limn xn to be the equivalence class
of a sequence (xn )nN ; thus limn xn = limn yn if and only if xn = yn
for all n sufficiently close to . Such ultralimits will be called nonstandard
objects.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

136

1. Hilberts fifth problem

Given a sequence
(Xn )nN of standard spaces Xn U, we define the
Q
ultraproduct n Xn to be the space of all ultralimits limn xn , where
xn Xn for each n. Such spaces will be called nonstandard spaces (also
known as nonstandard sets or internal sets).

If X is a standard
Q space, we define the ultrapower X of X to be the

ultraproduct X := n X of the constant sequence X. We embed X

inside X by identifying each element x of X with the ultralimit limn x.
If X is the standard elements of some type, we refer to elements of X
(i.e. ultralimits of sequences in X) as nonstandard elements of the same
type. Thus, for instance, elements of N are nonstandard natural numbers,
elements of R are nonstandard real numbers (also known as hyperreals),
and so forth. The ultrapower U of the standard universe will be called the
nonstandard universe.

Note that one can define an ultralimit limn xn for sequences xn of

standard objects that are onlyQdefined for an -large set of n, and similarly
one can define ultraproducts n Xn on spaces Xn that are only defined
for an -large set of n. From the nonprincipal nature of we observe that
we can modify xn or Xn forQfinitely many n without affecting the ultralimit
limn xn or ultraproduct n Xn .
Remark 1.7.10. The relation between a standard space X and its ultrapower X is analogous in some ways to the relationship between the rationals Q and its metric completion, namely the real line R. Indeed, using the
usual metric completion construction, one can interpret a real number as
an equivalence class of Cauchy sequences of rationals (or as a formal limit
limn qn of one of the Cauchy sequences in this class), and then one identifies each rational q with its formal limit limn q in order to keep the
rationals embedded inside the reals. Note however that whilst in the metric
completion, one is only allowed to take (formal) limits of Cauchy sequences,
in the ultrapower construction one may take limits of arbitrary sequences.
This makes the ultrapower construction more powerful than the metric
completion construction in that no a priori convergence (or even boundedness) hypotheses are needed in order to take limits. See [Ta2011c, 4.4] for
more discussion of ultrapowers as a completion.
Remark 1.7.11. With our conventions, we have the slightly unfortunate
consequence that every standard object is also a nonstandard object; for
instance, every standard natural number is also a nonstandard natural number, every standard real is also a nonstandard real, and so forth. We will
occasionally use the terminology strictly nonstandard to denote a nonstandard object that is not standard. For instance, an element of N\N would
be a strictly nonstandard natural number.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

137

Exercise 1.7.7. If X is a standard space, show that X = X if and only

if X is finite. In particular, if xn X is a sequence with finite (standard)
range X, then limn xn also lives in the same range X.
The above exercise shows that for finite spaces, there is no distinction
between a standard space X and its nonstandard counterpart X. However,
for infinite spaces, the nonstandard version of the space is usually much
larger:
Exercise 1.7.8. Show that N is uncountable. (Hint: adapt the Cantor
diagonal argument.) The precise cardinality of an ultrapower is actually
quite difficult to determine in general, as it can depend on the choice of
ultrafilter and even on the choice of additional set theory axioms such as
the continuum hypothesis.
Exercise 1.7.9 (Ultrapowers preserve Boolean operations). Let X, Y be
standard spaces. Show that (X Y ) = X Y , (X Y ) = X Y , and
(X\Y ) = X \Y . Also show that X Y if and only if X Y .
Strictly speaking, ultrapowers do not preserve Cartesian products: from
the definitions, (X Y ) and X Y are slightly different spaces. However,
there is an obvious bijection between the two sets, and we shall abuse notation and implicitly make the identification (X Y ) = X Y throughout
the rest of this section. More generally,Qgiven sequences XnQ
, Yn of standard
Q spaces, we make the identification n (Xn Yn ) = ( n Xn )
( n Yn ).
Remark 1.7.12. Just as standard
Q spaces X are subsets of the standard
universe U, nonstandard spaces n Xn are subsets of the nonstandard
universe U. However, the converse is not true in general; not every subset of the nonstandard universe is necessarily a nonstandard space. For
instance, we will show later that N is not a nonstandard subset of N (and
as such, is sometimes referred to as an external subset of N). Very roughly
speaking, internal spaces in the nonstandard universe play a role similar to
that of elementary sets (i.e. finite boolean combinations of intervals) in the
real line14; they are closed under various operations, but fall well short of exhausting all possible subsets of the ambient space. This analogy is pursued
further in Section 2.10.
Now we record an important property of nonstandard spaces, analogous
to the finite intersection property formulation of (countable) compactness.
14Actually, an even better analogy would be the clopen subsets of a Cantor space.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

138

1. Hilberts fifth problem

Lemma 1.7.13 (Countable saturation). Let F1 , F2 , . . . be a sequence of

nonstandard spaces such that
TM any finite collection of these spaces has nonempty intersection (thus m=1 Fm 6= forT all M 1). Then the entire
sequence has non-empty intersection (thus
m=1 Fm 6= ).
Q
T
Proof. Write Fm = n Fn,m , then for each m, M
m=1 FmTis the ultraTM
product of m=1 Fm,n (cf. Exercise 1.7.9). In particular, M
m=1 Fm,n is
non-empty for all n in an -large subset EM of the natural numbers. By
shrinking the EM as necessary we may assume that they are monotone, thus
E1 E2 . . . , and such that
T EM does not contain any natural number
less than M ; in particular,
M =1 EM = . For any n E1 , let Mn denote
the largest natural
number
M
such that n EM , and then let xn denote
TMn
an element of m=1 Fm,n . Then by construction, for each m 1, we have
xn Fm,n for all n Em , andTthus the nonstandard object x := limn xn

lies in Fm for all m, and thus
m=1 Fm is non-empty as required.
Remark 1.7.14. Strictly speaking, this property is slightly stronger than
countable saturation, and should technically be referred to as 1 -saturation.
Exercise 1.7.10 (Countable compactness). Show that any countable cover
of a nonstandard space by other nonstandard spaces has a finite subcover.
Exercise 1.7.11 (Bolzano-Weierstrass type theorem). Let X1 , . . . , Xk be
a finite collection of nonstandard spaces, and let Pi : X1 Xk
{true, false} be an at most countable collection of nonstandard k-ary predicates on these spaces. Let x1,m , . . . , xk,m for m N be sequences of nonstandard objects in X1 , . . . , Xk respectively (indexed by the standard natural numbers N). Show that there are subsequences x1,mj , . . . , xk,mj which
converge elementarily to some elements x1 , . . . , xk of X1 , . . . , Xk in the sense
that for each predicate Pi , one has Pi (x1,mj , . . . , xk,mj ) = Pi (x1 , . . . , xk ) for
all sufficiently large j. (See [Ta2011c, 4.4] for more discussion on these
sorts of completeness properties on nonstandard spaces.)
Remark 1.7.15. By replacing the ultraproduct construction with more sophisticated set-theoretic constructions, one can strengthen the countable saturation property by replacing countable by at most cardinality for any
given cardinal ; however, as one increases , the size of the ultraproducttype object must also necessarily increase (although one can make the size
of the model only slightly larger than , if one is willing to use inaccessible
cardinals). Such (huge) saturated models are of importance in model theory, but for our applications we will only need to work with the countably
saturated models provided by the ultraproduct construction.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

139

We have taken ultralimits and ultraproducts of standard objects and

standard spaces; now we perform a similar construction to standard functions and relations.
Definition 1.7.16 (Ultralimits of functions and relations). A standard function is a function f : X Y between two standard spaces X, Y . Given
a sequence fQ
n : Xn YnQof standard functions, we define the ultralimit
limn fn : n Xn n Yn to be the function defined by the formula
( lim fn )( lim xn ) := lim fn (xn )
n

whenever xn Xn . One can easily verify that this does indeed define
a function, which we refer to as a nonstandard function (also called an
internal function in the literature). The ultralimit f := limn f of a
single standard function is thus a nonstandard function from X to Y that
extends f ; by abuse of notation we shall also refer to f as f (analogously
to how we identify limn x with x rather than giving it a different name
such as x).
Similarly, for a (standard) natural number k N, define a standard kary relation (or standard k-ary predicate) to be a standard function R : X1
Xk {true, false} from the product of k standard spaces to the Boolean
space {true, false} (which we will treat as part of the standard universe U).
By the preceding construction (and Exercise 1.7.7), the ultralimit limn Rn
of a sequence of standard k-ary relations Rn : X1,n Xk,n {true, false}
(with k independent of n) is another k-ary relation
lim Rn : lim X1,n lim Xk,n {true, false},

which we will call a nonstandard k-ary relation (also known as an internal

k-ary relation). Thus limn Rn (limn x1,n , . . . , limn xk,n ) is true if and
only if Rn (x1,n , . . . , xk,n ) is true for all n sufficiently close to . Again, we
identify each relation R with its nonstandard extension R := limn R.
In a very similar spirit, given a sequence On : X1,n Xk,n Yn of
standard k-ary operators, one can define the ultralimit
lim On : lim X1,n lim Xk,n lim Yn

which we will call a nonstandard k-ary operator (also known as an internal k-ary operator ). Again, we identify each standard operator O with its
nonstandard extension O := limn O.
Example 1.7.17. Let a = limn an and b = limn bn be two nonstandard natural numbers (thus an , bn are sequences of standard natural numbers). Then, by definition,
(1) a + b is the nonstandard natural number a + b := limn an + bn .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

140

1. Hilberts fifth problem

(2) ab is the nonstandard natural number ab := limn an bn .

(3) One has a < b if an < bn for all n sufficiently close to .
(4) a is even if an is even for all n sufficiently close to .
(5) a is prime if an is prime for all n sufficiently close to .
A basic fact about the ultraproduct construction is that any elementary
(or more precisely, first-order ) statement that is true for a standard space
(or a sequence of such standard spaces) is also true for the associated nonstandard space (either an ultrapower or an ultraproduct). We will formalise
this assertion shortly, but let us first illustrate it with some examples.
Exercise 1.7.12 (Examples of Loss theorem).
(i) Show that addition
on the nonstandard natural numbers is both commutative and associative.
(ii) Show that a nonstandard natural number a is even if and only if it
is of the form a = 2b for some nonstandard natural number b.
(iii) Show that a nonstandard natural number a is prime if and only
if it is greater than 1, but cannot be factored as a = bc for some
nonstandard natural numbers b, c > 1.
(iv) Show that the standard Goldbach conjecture (every even standard
natural number larger than 2 is the sum of two standard primes) is
logically equivalent to the nonstandard Goldbach conjecture (every
even nonstandard natural number larger than 2 is the sum of two
nonstandard primes).
(v) Show that the standard twin prime conjecture (for any standard
natural number a, there exists a standard prime p > a such that
p + 2 is also prime) is logically equivalent to the nonstandard twin
prime conjecture (for any nonstandard natural number a, there exists a nonstandard prime p > a such that p+2 is also a nonstandard
prime).
Now we generalise the above examples.
Exercise 1.7.13. Let k 0, and let Pn , Qn : X1,n Xk,n {true, false}
be two sequences of k-ary predicates. For each 1 i k, let Xi :=
Q
n Xi,n , and write P := limn Pn and Q := limn Qn .
(i) (Continuity of Boolean operations) Show that limn (Pn Qn ) =
P Q, limn (Pn Qn ) = P Q, and limn (Pn ) = P .
(ii) (Continuity of existential quantifier) If 1 i k, i Pn denotes the
k 1-ary predicate xi Xi,n : Pn (x1,n , . . . , xk,n ) in the remaining
k 1 variables xj,n Xj,n for j 6= i, and i P denotes the k 1-ary

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

141

predicate xi Xi : P (x1 , . . . , xk ) in the remaining k 1 variables

xj Xj for j 6= i, show that limn i Pn = i P .
(iii) (Continuity of universal quantifier) If 1 i k, i Pn denotes the
k 1-ary predicate xi Xi,n : Pn (x1,n , . . . , xk,n ) in the remaining
k 1 variables xj,n Xj,n for j 6= i, and i P denotes the k 1-ary
predicate xi Xi : P (x1 , . . . , xk ) in the remaining k 1 variables
xj Xj for j 6= i, show that limn i Pn = i P .
By iterating the above exercise, we can obtain Loss theorem [Lo1955],
which we first state informally as follows.
Theorem 1.7.18 (Loss theorem, informal version). Let k, m be natural
numbers. For each natural number n, suppose one has a collection a1,n , . . . , am,n
of standard objects, spaces, operators, and relations (of various arities), and
let a1 , . . . , am be the corresponding nonstandard objects, spaces, operators,
and relations formed via the ultralimit or ultraproduct construction. Let P
be a formal k-ary predicate expressible in first-order logic using m objects,
spaces, operators, and relations. Then P , when quantified over a1 , . . . , am ,
is the ultralimit of P quantified over a1,n , . . . , am,n . (In particular, P quantified over a1 , . . . , am is a nonstandard predicate.)
Specialising to the case k = 0, (so P is a first-order sentence involving
m objects, spaces, operators, and relations), we see that P is true when
quantified over a1 , . . . , am if and only if P is true when quantified over
a1,n , . . . , am,n for all n sufficiently close to .
Corollary 1.7.19 (Special case of Loss theorem, informal version). Any
first-order sentence that holds for a finite collection of standard objects,
spaces, operators, and relations, also holds for their corresponding nonstandard counterparts (using the ultralimit or ultrapower construction).
Before we state the theorem more formally, let is illustrate it with some
more examples.
(1) R is an ordered field when given the usual constants 0, 1, field
operations +, , , ()1 , and order relation <. (Technically, ()1 is
not an operation on R because it is undefined at 0, but this can
easily be fixed by artificial means, e.g. by arbitrarily defining ()1
at 0.). The assertion that R is an ordered field can be expressed as
a (rather lengthy) sentence in first-order logic using the indicated
spaces, objects, operations, and relations. As a consequence, the
ultrapower R (with the same constants 0, 1 and the nonstandard
extensions of the field operations and order relation) is also an
ordered field.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

142

1. Hilberts fifth problem

(2) Let GQ
n = (Gn , n ) be a sequence of groups; then the ultraproduct
G := n Gn (with the ultralimit group operation := limn n ,
and similarly for the group identity and inverse operations) is also
a group, because the group axioms can be phrased in first-order
logic using the indicated structures. If, for each n, gn is an element
of Gn , then the ultralimit limn gn is a central element of G if and
only if gn is a central element of Gn for all n sufficiently close to .
Similarly, if for each n, Hn is a subset of Gn , then the ultraproduct
Q
n Hn is a normal subgroup of G if and only if Hn is a normal
subgroup of Gn for all n sufficiently close to .
(3) The standard natural numbers N obey the axiom of induction: if
P (n) is a predicate definable in the language of Peano arithmetic,
and P (0) is true, and P (n) implies P (n + 1) for all n N, then
P (n) is true for all n N. As a consequence, the same axiom of
induction also holds for the nonstandard natural numbers N. Note
however it is important for the nonstandard axiom of induction that
the predicate be definable in the language of Peano arithmetic. For
instance, the following argument is fallacious: 0 N, and for any
nonstandard natural number n, n N implies n + 1 N. Hence,
by induction, n N for all nonstandard natural numbers n. The
reason is that the predicate n N is not formalisable in the
language of Peano arithmetic.
(4) Exercise 1.7.9 can be interpreted as a special case of Corollary
1.7.19.
Exercise 1.7.14.
(i) Show that N (viewed as a subset of N) is not a nonstandard space.
(Hint: if it were, the fallacious induction mentioned earlier could
now be made valid.)
(ii) Establish the overspill principle: if a nonstandard predicate P (n)
on the nonstandard natural numbers is true for all standard natural
numbers, then it is true for at least one strictly nonstandard natural
number as well.
(iii) Show that if a nonstandard subset of N consists only of standard
natural numbers, then it is finite.
(iv) Show that the nonstandard natural numbers N are not well-ordered.
Why does this not contradict Loss theorem?
Exercise 1.7.15.
Q For each n, let GnQbe a group, and let Sn be a subset of
Gn . Write G := n Gn and S := n Sn .
(i) Give an example in which each Sn generates Gn as a group, but S
does not generate G as a group.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

143

(ii) Show that S generates G as a group if and only if there exists a

(standard) natural number M such that (Sn {1} Sn1 )M = Gn
for all n sufficiently close to . (Hint: apply countable saturation
(Exercise 1.7.10) to the sets (S {1} S 1 )M .)
Now we make Loss theorem more precise, by defining a formal language
for first-order logic. (This section is optional and may be safely omitted by
the reader.)
Suppose we are given a collection L of formal objects, spaces, operators,
and relations, with each object (formally) belonging to one of the spaces, and
with each of the operators and relations formally defined on some collection
of these spaces (and in the case of operators, taking values in one of the
spaces) as well. We will refer to elements of L as primitive terms. Initially,
we do not actually identify these formal terms with concrete counterparts,
whether they be standard or nonstandard. A well-formed formula involving
L and one or more formal free variables x1 , . . . , xn , each of which belonging
to one of the formal spaces, is formed by a finite number of applications of
the following rules:
(1) Each free variable x1 , . . . , xn is a well-formed formula (belonging to
the associated formal space).
(2) Each object in C is a well-formed formula (belonging to the associated formal space).
(3) If a1 , . . . , ak are well-formed formulae, and O is a k-ary operator,
then O(a1 , . . . , ak ) will be a well-formed formula if the a1 , . . . , ak
belong to the formal spaces indicated by the domain of O. Of
course, O(a1 , . . . , ak ) then belongs to the formal space indicated by
the range of O.
A formal predicate P = P (x1 , . . . , xk ) involving L and one or more free
variables x1 , . . . , xk is then formed by a finite number of applications of the
following rules:
If X, Y are well-formed formulae belonging to the same formal
space, then X = Y is a predicate.
If X1 , . . . , Xk are well-formed formulae, and R is a k-ary relation
in L, then R(X1 , . . . , Xk ) is a predicate if the X1 , . . . , Xk belong to
the formal spaces indicated by the domain of R.
If P and Q are predicates, then so are (P Q), (P Q), (P ),
(P = Q), and (P Q).
If P (x1 , . . . , xk , x) is a predicate in x1 , . . . , xk and an additional free
variable x in some formal space X, then (x X : P (x1 , . . . , xk , x))

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

144

1. Hilberts fifth problem

and (x X : P (x1 , . . . , xk , x)) are predicates in x1 , . . . , xk . (Similarly for any permutation in the ordering of x1 , . . . , xk , x.)
In practice, of course, many of the parentheses in the above constructions
can be removed without causing any ambiguity, and one usually does so in
order to improve readability, for instance abbreviating ((P Q) R) as
P Q R. However, for the purposes of studying formal logic, it can be
convenient to insist on these parentheses.
Given a formal predicate involving L and one or more free variables
x1 , . . . , xn , we may quantify (or interpret) the predicate by associating each
formal object in L to an actual object, which may be a standard object,
a nonstandard object, or something else entirely, and similarly associating
an actual space, relation, or operator to each formal space, relation, or
operator, making sure that all the relevant domains and ranges are respected.
When one does so, the formal predicate P becomes a concrete predicate
P : X1 Xk {true, false} on the appropriate quantified spaces by
interpreting all the symbols in P in the obvious fashion. For instance, if the
primitive terms consist of a formal space X and a formal binary operation
: XX X, then P (x) := (y X : xy = yx) is a formal unary predicate,
and if one quantifies this predicate over a concrete group G = (G, ) (which
may be standard, nonstandard, or neither), then one obtains a concrete
unary predicate P : G {true, false}, with P (g) true precisely when g is a
central element of G.
Exercise 1.7.16. With the above definitions, prove Theorem 1.7.18.
Exercise 1.7.17 (Compactness theorem). Let L be an at most countable
collection of formal objects, spaces, operators and relations. Let S1 , S2 , S3 , . . .
be a sequence of sentences involving the symbols in L. Suppose that any
finite collection S1 , . . . , Sn of these sentences is satisfiable in the standard
universe, thus there exists an assignment of standard objects, spaces, operators, or relations to each element of L for which S1 , . . . , Sn are all true when
quantified over these assignments. Show that the entire collection S1 , S2 , . . .
are then satisfiable in the nonstandard universe, thus there exists an assignment of nonstandard objects. (This is the countable case of the compactness
theorem in logic. One can also use ultraproducts over larger sets than N to
prove the general case of the compactness theorem, but we will not do so
here.)
1.7.3. Nonstandard finite sets and nonstandard finite sums. It will
be convenient to extend the standard machinery of finite sets and finite sums
to the nonstandard
setting. Define a nonstandard finite set to be an ultraQ
product A = n An of finite sets, and a nonstandard finite sequence of
reals (xm )mA to be a nonstandard function m 7 xm from a nonstandard

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

145

finite set A to the nonstandard reals, or equivalently an ultralimit of standard finite sequences (xm,n )mAn of reals. We can define the (nonstandard)
cardinality |A| of a nonstandard finite set in the usual manner:
Y
An | := lim |An |.
|
n

Thus, for instance, if N is a nonstandard natural number, then the nonstandard finite set {n N : 1 n N } has nonstandard cardinality
N.
Similarly, if (xm )mA = limn (xm,n )mAn P
is a nonstandard finite sequence of reals, we define the nonstandard sum mA xm as
X
X
xm := lim
xm,n .
mA

mAn

Thus for instance mA 1 = |A|. By using Loss theorem, all the basic laws
of algebra for manipulating standard finite sums carry over to nonstandard
finite sums. For instance, we may interchange summations
X X
X X
X
xm1 ,m2 =
xm1 ,m2 =
xm1 ,m2
m1 A1 m2 A2

m2 A2 m1 A1

(m1 ,m2 )A1 A2

for any nonstandard finite sequence (xm1 ,m2 )(m1 ,m2 )A1 A2 of reals indexed
by a product set, simply because the same assertion is obviously true for
standard finite sequences.
1.7.4. Asymptotic notation. The ultrapower and ultralimit constructions are extremely general, being basically applicable to any collection of
standard objects, spaces, objects, and relations. However, when one applies
these constructions to standard number systems, such as the natural numbers N, the integers Z, the real numbers R, or the complex numbers C to
obtain nonstandard number systems such as N, Z, R, C, then one can
gain an additional tool of use in analysis, namely a clean asymptotic notation
which is closely related to standard asymptotic notation, but has a slightly
different arrangement of quantifiers that make the nonstandard asymptotic
notation better suited to algebraic constructions (such as the quotient space
construction) than standard asymptotic notation.
Definition 1.7.20 (Asymptotic notation). Let R be one of the standard
number systems (N, Z, Q, R, C). A nonstandard number x in R is said
to be bounded if one has |x| C for some standard real number C (or
equivalently, |x| n some standard natural number n), and unbounded
otherwise. If y is a non-negative nonstandard real, we write x = O(y) if
|x| Cy for some standard real number C, thus a nonstandard number is
bounded if and only if it is O(1). We say that x is infinitesimal if |x| for
every standard real number > 0 (or equivalently, if |x| n1 for all standard

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

146

1. Hilberts fifth problem

positive natural numbers n), and write x = o(y) if |x| y for all standard
real numbers > 0, thus x is infinitesimal if and only if it is o(1).
Example 1.7.21. All standard numbers are bounded, and all non-zero standard numbers are non-infinitesimal. The nonstandard natural number N :=
limn n is unbounded, and the nonstandard real number N1 = limn n1 is
non-zero but infinitesimal.
Remark 1.7.22. One can also develop analogous nonstandard asymptotic
notation on other spaces, such as normed vector spaces or locally compact
groups; see for instance [Ta2011c, 4.5] for an example of the former, and
Hirschfelds nonstandard proof [Hi1990] of Hilberts fifth problem for an
example of the latter. We will however not need to deploy nonstandard
asymptotic notation in such generality here.
Note that we can apply asymptotic notation to individual nonstandard
numbers, in contrast to standard asymptotic notation in which one needs
to have the numbers involved to depend on some additional parameter if
one wishes to prevent the notation from degenerating into triviality. In
particular, we can form well-defined sets such as the bounded nonstandard
reals
O(R) := {x R : x = O(1)}
and the infinitesimal nonstandard reals
o(R) := {x R : x = o(1)}.
Exercise 1.7.18 (Standard part). Show that o(R) is an ideal of the commutative ring O(R), and one has the decomposition O(R) = R o(R), thus
every bounded real x = O(R) has a unique decomposition x = st(x) + (x
st(x)) into a standard part st(x) R and an infinitesimal part x st(x)
o(R). Show that the map x 7 st(x) is a ring homomorphism from O(R) to
R.
Exercise 1.7.19. Let N be an unbounded nonstandard natural number.
Write OZ (N ) := {n Z : n = O(N )} and oZ (N ) := {n Z : n = o(N )}.
Show that oZ (N ) is a subgroup of the additive group OZ (N ), and that
the quotient group OZ (N )/oZ (N ) is isomorphic (as an additive group) to
the standard real numbers R. (One could in fact use this construction as
a definition of the real number system, although this would be a rather
idiosyncratic choice of construction.)
Remark 1.7.23. In some texts, the standard part x = st limn xn of
an ultralimit of a bounded sequence of numbers is denoted by limn xn
(with the ultralimit itself not having a specific notation assigned to it). The
assertion x = st limn xn is also sometimes referred to as xn converges to

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

147

x along , or equivalently that every neighbourhood of x contains xn for

-almost every n.
The following exercise should be compared with Proposition 1.7.2. Note
the absence of a continuity hypothesis on the function f .
Exercise 1.7.20. Let f : X R be a standard function, which we extend
to a nonstandard function f : X R in the usual manner. Show that
the following statements are equivalent:
(i) (Qualitative bound on nonstandard objects) For all x X, one
has f (x) = O(1).
(ii) (Quantitative bound on standard objects) There exists a standard
real M < + such that |f (x)| M for all x X.
(Hint: use some version of the countable saturation property.)
One of the first motivations of nonstandard analysis was to obtain a
rigorous theory of infinitesimals that could be used as an alternate (though
logically equivalent) foundation for real analysis. We will not need to use
this foundation here, but the following exercises are intended to give some
indication of one might start setting up such foundations.
Exercise 1.7.21. Let E be a standard subset of the reals R.
(i) Show that E is open if and only if E + o(R) E, or equivalently
if st1 (E) E. (Here, A + B := {a + b : a A, b B} denotes
the sumset of A and B.)
(ii) Show that E is closed if and only if E O(R) E + o(R), or
equivalently if st( E O(R)) E.
(iii) Show that E is bounded if and only if E O(R).
(iv) Show that E is compact if and only if E E + o(R).
Exercise 1.7.22. Let (xn )nN be a standard sequence of reals, and let
(xn )n N be its nonstandard extension. Let L be a real number.
(i) Show that xn converges to L as n if and only if xN = L + o(1)
for all unbounded natural numbers N .
P
(ii) Show that
n=1 xn is conditionally convergent to L if and only if
PN
n=1 xN = L + o(1) for all unbounded natural numbers N .
Exercise 1.7.23. Let f : R R be a standard function, and let f : R
R be its nonstandard extension. Show that the following are equivalent:
(i) f is a continuous function.
(ii) One has f (y) = f (x) + o(1) whenever x R, y R, and y =
x + o(1).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

148

1. Hilberts fifth problem

(iii) One has f (st(x)) = st(f (x)) for all x O(R).

Exercise 1.7.24. Let f : R R be a standard function, and let f : R
R be its nonstandard extension. Show that the following are equivalent:
(i) f is a uniformly continuous function.
(ii) One has f (y) = f (x) + o(1) whenever x, y R and y = x + o(1).
Exercise 1.7.25. Let f : R R be a standard function, and let f : R
R be its nonstandard extension. Show that the following are equivalent:
(i) f is a differentiable function.
(ii) There exists a standard function f 0 : R R such that f (x + h) =
f (x) + f 0 (x)h + o(|h|) for all x R and h = o(1), or equivalently if
(x)
f 0 (x) = st f (x+h)f
for all x R and non-zero h = o(1).
h
Exercise 1.7.26. Let f : R R be a standard continuous function, and
let f : R R be its nonstandard extension. Show that for any standard
reals a < b and any unbounded natural number N , one has
Z b
N
ba
1 X
f (a +
n)
f (x) dx = st
N
N
a
n=1
PN
ba
(note that n=1 f (a + N n) is a nonstandard finite sum). More generally,
show that
Z b
N
1 X
f (x) dx = st
f (xn )(xn xn1 )
N
a
n=1

(xn )1nN

whenever (xn )0nN and

are nonstandard finite sequences with
a xn1 xn xn b and xn = xn1 + o(1) for all 1 n N .
1.7.5. Ultra approximate groups. In this section we specialise the above
ultraproduct machinery to the concept of a finite approximate group, to link
such objects with a type of infinite approximate group which we call an ultra
approximate group. These objects will be studied much more intensively in
the next two lectures, but for now we focus on the correspondence between
ultra approximate groups and their finitary counterparts.
We first recall the notion of an approximate group. For simplicity, we
work for now in the global group setting, although later on we will also need
to generalise to local groups.
Definition 1.7.24 (Approximate group). Let K 1 be a standard real
number. A (global) approximate group is a subset A of a global group
G = (G, ) which is symmetric, contains the identity, and is such that A A
can be covered by at most K left-translates of A.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

149

Definition 1.7.25 (Ultra

Q approximate group). An ultra approximate group
is an ultraproduct A = n An , where each An Gn is a standard finite
K-approximate group for some K independent of n.
Q
Example 1.7.26. If N = n Nn is a nonstandard natural number, then
the nonstandard arithmetic progression {m N : N m N } is an
ultra approximate group, because it is an ultraproduct of standard finite
2-approximate groups {m N : Nn m Nn }.
Q
Example 1.7.27. If N = n Nn is a Q
nonstandard natural number, then
the nonstandard cyclic group Z/N Z = n Z/Nn Z is an ultra approximate group, because it is an ultraproduct of standard cyclic groups. More
generally, any nonstandard finite group (i.e. an ultraproduct of standard
finite groups) is an ultra approximate group.
For any fixed K, the statement of a set A in a group G being a Kapproximate group can be phrased in first-order logic quantified over G
using the group operations and the predicate x A of membership in A.
From this and Loss theorem, we see that any ultraproduct of K-approximate
groups is again a K-approximate group. In particular, an ultra approximate
group is a K-approximate group for some (standard) finite K.
We will be able to make a correspondence between various assertions
about K-approximate groups and about ultra approximate groups, so that
problems about the former can be translated to problems about the latter.
By itself, this correspondence is not particularly deep or substantial; but
the point will be that by moving from the finitary category of finite Kapproximate groups to the infinitary category of ultra approximate groups,
we will be able to deploy tools from infinitary mathematics, and in particular
the topological group theory results (such as the Gleason-Yamabe theorem)
from previous sections, to bear on the problem. We will execute this strategy
in the next few sections, but for now, let us just see how the correspondence
works. We begin with a simple example, in the bounded torsion case. For
a standard natural number r, we say that a group G is r-torsion if every
element has order at most r, thus for each g G one has g n = 1 for some
1 n r.
Proposition 1.7.28 (Correspondence in the torsion case). Let r 1. The
following two statements are equivalent:
(i) (Finitary statement) For all K 1, there exists CK,r 1 such
that, given a finite K-approximate group A in an r-torsion group
G, one can find a finite (genuine) subgroup H of G such that A4
contains H, and A can be covered by at most CK,r left-translates of
H.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

150

1. Hilberts fifth problem

(ii) (Infinitary statement) For any ultra approximate group A in a rtorsion nonstandard group G, one can find a nonstandard finite
subgroup H of G such that A4 contains H, and A can be covered
by finitely many left-translates of H.
Q
Proof. Let us first assume (i) and establish (ii). Let A = nQ
An be an
ultra approximate group in an r-torsion nonstandard group G = n Gn .
Since G is r-torsion, and the property of being r-torsion is a first-order
statement in the language of groups, we see from Loss theorem that Gn is
a r-torsion group for all n sufficiently close to . Since the An are all finite
K-approximate groups for some K independent of n, we conclude from (i)
that for all n sufficiently close to , we can find a finite subgroup Hn of
and An can be covered by at most CK,r leftGn such that A4n contains Hn , Q
translates of Hn . Taking H := n Hn and applying Loss theorem again,
we cnoclude that H is a nonstandard finite subgroup of G, that A4 contains
H, and A can be covered by at most CK,r left-translates of H, giving (ii) as
desired.
Now we assume (ii) and establish (i). Suppose for contradiction that (i)
failed. Carefully negating the quantifiers, we conclude that there exists K
such that for each natural number n, one can find a finite K-approximate
group An in an r-torsion group Gn for which there does not exist any finite
subgroup Hn of Gn such that A4n contains Hn and An can be covered
Q by at
most n left-translates
of Hn . We then form the ultraproducts A := n An
Q
and G := n Gn . By Loss theorem, G is an r-torsion nonstandard
group, and A is an ultra approximate
there is
Q group in G. Thus, by (ii),
4
a nonstandard finite subgroup H = n Hn of G such that A contains
H and A is covered by M left-translates of H for some (standard) natural
number M . By Loss theorem, we conclude that for all n sufficiently close to
, Hn is a finite subgroup of Gn , A4n contains Hn , and An is covered by M
left-translates of Hn . In particular, as is nonprincipal, this assertion holds
for at least one n > M . But this contradicts the construction of An .

Note how the infinitary statement in the above proposition is a qualitative version of the more quantitative finitary statement. In particular,
parameters such as K and CK,r have been efficiently concealed in the infinitary formulation, whereas they must be made explicit in the finitary
formulation. As such, the infinitary approach leads to a reduction in the
amount of epsilon management that one often has to perform in finitary
settings.
In the next section we will use the Gleason-Yamabe theorem to establish
(ii) and hence conclude (i) as a corollary (this argument is essentially due to
Hrushovski [Hr2012]). Interestingly, no purely finitary proof of (i) in full

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.7. Ultraproducts, hard analysis, and soft analysis

151

generality is currently known (other than by finitising the infinitary proof,

which is in principle possible, but would create an extremely long and messy
proof as a consequence).
We move from the bounded torsion setting to another special case,
namely the abelian setting. Here, the finite subgroups need to be replaced
by the more general concept of a coset progression.
Definition 1.7.29. A (standard symmetric) generalised arithmetic progression P of rank r in a (standard) additive group G = (G, +) is a set of the
form
P = P (v1 , . . . , vr ; N1 , . . . , Nr ) := {a1 v1 + +ar vr : |a1 | N1 , . . . , |ar | Nr }
where v1 , . . . , vr G, N1 , . . . , Nr > 0, and the a1 , . . . , ar are constrained to
be integers.
An ultra generalised arithmetic progression is an ultraproduct
Q
:=
P
n Pn of standard generalised arithmetic progressions Pn of fixed
rank r.
A (standard) coset progression Q of rank r in a (standard) additive
group G = (G, +) is a set of the form Q = H + P , where H is a finite
subgroup of G and P is a generalised arithmeticQprogression of rank r. An
ultra coset progression is an ultraproduct Q := n Qn of standard coset
progressions Qn of fixed rank r; equivalently, it is a set of the form H + P ,
where H is a nonstandard finite group in a nonstandard group G, and P is
an ultra generalised arithmetic progression in G.
Exercise 1.7.27. Show that the following two statements are equivalent:
(i) (Finitary abelian Freiman theorem) For all K 1, there exists
CK , rK 1 such that, given a finite K-approximate group A in an
abelian group G = (G, +), one can find a coset progression Q in G
of rank at most rK such that A4 contains Q, and A can be covered
by at most CK left-translates of Q.
(ii) (Infinitary abelian Freiman theorem) For any ultra approximate
group A in a abelian nonstandard group G = (G, +), one can find
an ultra coset progression Q of G such that A4 contains Q, and A
can be covered by finitely many left-translates of Q.
The statement (i) was established by Green and Ruzsa [GrRu2007] by
finitary means (such as the finite Fourier transform). In later sections we
will give an alternate proof of (i) that proceeds via (ii) and the structure
theory of locally compact abelian groups.
Finally, we can give the correspondence in the full nonabelian setting.
Definition 1.7.30. Given a (standard) finite number of generators v1 , . . . , vr
in a (standard) group G = (G, ), and (standard) real numbers N1 , . . . , Nr >

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

152

1. Hilberts fifth problem

0, define the noncommutative progression P (v1 , . . . , vr ; N1 , . . . , Nr ) to be the

set of all words in v11 , . . . , vr1 involving at most Ni copies of vi1 for each
1 i r. We refer to r as the rank of the noncommutative progression. A
nilprogression of rank r and step at most s is a noncommutative progression
that lies in a nilpotent group of step at most s. A coset nilprogression in
G of rank r and step at most s is a set of the form 1 (P ), where H is a
finite group in G, N (H) := {g H : gH = Hg} is the normaliser of H,
: N (H) N (H)/H is the quotient map, and P is a nilprogression of rank
r in N (H)/H.
Q
An ultra coset nilprogression is an ultraproduct Q = n Qn of coset
nilprogressions of rank r and step at most s, for some standard natural
numbers r, s independent of n.
Exercise 1.7.28. Show that the following two statements are equivalent:
(i) (Finitary nonabelian Freiman theorem) For all K 1, there exists
CK , sK , rK 1 such that, given a finite K-approximate group A
in a group G = (G, ), one can find a coset nilprogression Q in G
of rank at most rK and step at most sK such that A4 contains Q,
and A can be covered by at most CK left-translates of Q.
(ii) (Infinitary nonabelian Freiman theorem) For any ultra approximate
group A in a nonstandard group G = (G, ), one can find an ultra
coset nilprogression Q in G such that A4 contains Q, and A can be
covered by finitely many left-translates of Q.
In later sections we will establish (ii), and hence (i), thus giving a (qualitatively) satisfactory description of finite approximate groups in arbitrary
groups.
Exercise 1.7.29. Let n be a standard natural number. Show that every
nonstandard finite subgroup of GLn (C) is virtually abelian.

1.8. Models of ultra approximate groups

In Section 1.7, we Q
introduced the notion of an ultra approximate group - an
ultraproduct A = n An of finite K-approximate groups An for some K
independent of n, where each K-approximate group An may lie in a distinct
ambient group Gn . Although these objects arise initially from the finitary
objects An , it turns out that ultra approximate groups A can be profitably
analysed by means of infinitary groups L (and in particular, locally compact
groups or Lie groups L), by means of certain models 15 : hAi L of A
(or of the group hAi generated by A). We will define precisely what we
15The terminology here comes from additive combinatorics, and is a little different from the
notion of a model from model theory.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

153

mean by a model later, but as a first approximation one can view a model
as a representation of the ultra approximate group A (or of hAi) that is
macroscopically faithful in that it accurately describes the large scale
behaviour of A (or equivalently, that the kernel of the representation is
microscopic in some sense). In the next section we will see how one can
use Gleason lemma technology to convert this macroscopic control of an
ultra approximate group into microscopic control, which will be the key to
classifying approximate groups.
Models of ultra approximate groups can be viewed as the multiplicative
combinatorics analogue of the more well known concept of an ultralimit of
metric spaces, which we briefly review below the fold as motivation.
The crucial observation is that ultra approximate groups enjoy a local
compactness property which allows them to be usefully modeled by locally
compact groups (and hence, through Theorem 1.1.17, by Lie groups also).
As per the Heine-Borel theorem, the local compactness will come from a
combination of a completeness property and a local total boundedness property. The completeness property turns out to be a direct consequence of
the countable saturation property of ultraproducts, thus illustrating one of
the key advantages of the ultraproduct setting. The local total boundedness property is more interesting. Roughly speaking, it asserts that large
bounded sets (such as A or A100 ) can be covered by finitely many translates of small bounded sets S, where small is a topological group sense,
implying in particular that large powers S m of S lie inside a set such as A
or A4 . The easiest way to obtain such a property comes from the following
lemma of Sanders [Sa2009]:
Lemma 1.8.1 (Sanders lemma). Let A be a finite K-approximate group in
a (global) group G, and let m 1. Then there exists a symmetric subset S
of A4 with |S| K,m |A| containing the identity such that S m A4 .
This lemma has an elementary combinatorial proof, and is the key to
endowing an ultra approximate group with locally compact structure. There
is also a closely related lemma of Croot and Sisask [CrSi2010] which can
achieve similar results, and which will also be discussed below. (The locally
compact structure can also be established more abstractly using the much
more general methods of definability theory, as was first done by Hrushovski
[Hr2012], but we will not discuss this approach here.)
By combining the locally compact structure of ultra approximate groups
A with the Gleason-Yamabe theorem, one ends up being able to model a
large ultra approximate subgroup A0 of A by a Lie group L. Such Lie
models serve a number of important purposes in the structure theory of
approximate groups. Firstly, as all Lie groups have a dimension which is

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

154

1. Hilberts fifth problem

a natural number, they allow one to assign a natural number dimension

to ultra approximate groups, which opens up the ability to perform induction on dimension arguments. Secondly, Lie groups have an escape
property (which is in fact equivalent to no small subgroups property): if a
group element g lies outside of a very small ball B , then some power g n of
it will escape a somewhat larger ball B1 . Or equivalently: if a long orbit
g, g 2 , . . . , g n lies inside the larger ball B1 , one can deduce that the original
element g lies inside the small ball B . Because all Lie groups have this
property, we will be able to show that all ultra approximate groups A essentially have a similar property, in that they are controlled by a nearby
ultra approximate group which obeys a number of escape-type properties
analogous to those enjoyed by small balls in a Lie group, and which we will
call a strong ultra approximate group. This will be discussed in the next
section, where we will also see how these escape-type properties can be exploited to create a metric structure on strong approximate groups analogous
to the Gleason metrics studied in Section 1.3, which can in turn be exploited
(together with an induction on dimension argument) to fully classify such
approximate groups (in the finite case, at least).
There are some cases where the analysis is particularly simple. For
instance, in the bounded torsion case, one can show that the associated Lie
model L is necessarily zero-dimensional, which allows for a easy classification
of approximate groups of bounded torsion.
The material in this section is loosely based on the papers [Hr2012],
[BrGrTa2011].
1.8.1. Ultralimits of metric spaces (Optional). Suppose one has a
sequence (Xn , dn ) of metric spaces. Intuitively, there should be a sense
in which such a sequence can (in certain circumstances) converge to a
limit (X, d) that is another metric space. Some informal examples of this
intuition:
(i) The sets {n, . . . , n} (with the usual metric) should converge as
n to the integers Z (with the usual metric).
(ii) The cyclic groups Z/nZ (with the discrete metric d(i, j) := dist(i+
nZ, j + nZ)) should also converge as n to the integers Z
(with the usual metric).
(iii) The sets {0, n1 , . . . , n1
n } (with the usual metric) should converge
as n to the interval [0, 1] (with the usual metric).
(iv) The cyclic groups Z/nZ (with the bounded metric d(i, j) :=
1
n dist(i + nZ, j + nZ)) should converge as n to the unit
circle R/Z (with the usual metric).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

155

(v) A Euclidean circle of radius n (such as {(x, y) R2 : x2 +(yn)2 =

n2 }) should converge as n to a Euclidean line (such as
{(x, 0) R2 : x R}).
Let us now try to formalise this intuition, proceeding in stages. The first
attempt to formalise the above concepts is via the Hausdoff distance, which
already made an appearance in Section 1.5:
Definition 1.8.2 (Hausdorff distance). Let X = (X, d) be a metric space.
The Hausdorff distance dH (E, F ) between two non-empty subsets E, F of
X is defined by the formula
dH (E, F ) := max(sup d(x, F ), sup d(E, y)).
xE

yF

Thus, if dH (E, F ) < r, then every point in E lies within a distance less than
r of a point in F , and every point in F lies within a distance less r of a point
in E (thus E Nr (F ) and F Nr (E), where Nr (F ) is the r-neighbourhood
of E); and conversely, if the latter claim holds, then dH (E, F ) r.
This distance is always symmetric, non-negative, and obeys the triangle
inequality. If one restricts attention to non-empty compact sets E, F , then
one easily verifies that dH (E, F ) = 0 if and only if E = F , so that the
Hausdorff distance becomes a metric. In particular, it becomes meaningful
to discuss the concept of a sequence En of non-empty compact subsets of X
converging in the Hausdorff distance to a (unique) limit non-empty compact
set E. Note that this concept captures the intuitive example (iii) given
above, but not any of the others. Nevertheless, we will discuss it first as it
is slightly simpler than the more general notions we will be using.
Exercise 1.8.1 (Hausdorff convergence and connectedness). Let En be a
sequence of non-empty compact subsets of a metric space X converging to
another non-empty compact set E.
(i) If the En are all connected, show that E is connected also.
(ii) If the En are all path-connected, does this imply that E is pathconnected also? Support your answer with a proof or counterexample.
(iii) If the En are all disconnected, does this imply that E is disconnected also? Support your answer with a proof or counterexample.
One of the key properties of Hausdorff distance in a compact set is that
it is itself compact:
Lemma 1.8.3 (Compactness of the Hausdorff metric). Let En be a sequence
of compact subsets of a compact metric space X. Then there is a subsequence
of the En that is convergent in the Hausdorff metric.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

156

1. Hilberts fifth problem

The proof of this lemma was set as Exercise 1.5.13. It can be proven by
conventional means (relying in particular on the Heine-Borel theorem),
but we will now sketch how one can establish this result using ultralimits
instead. As in Section 1.7, we fix a standard universe U and a non-principal
ultrafilter in order to define ultrapowers.
If X = (X, d) is a (standard) metric space with metric d : X X R+ ,
then the ultrapower X comes with a nonstandard metric d : X X
R+ that extends d. (In the previous sections, we referred to this metric
as d rather than d, but here it will be convenient to use a different symbol
for this extension of d to reduce confusion.) Let us say that two elements
x, y of X are infinitesimally close if d(x, y) = o(1). (The set of points
infinitesimally close to a standard point x is sometimes known as the monad
of x, although we will not use this terminology.)
Exercise 1.8.2. Let (X, d) be a (standard) compact metric space. Show
that for any nonstandard point x X there exists a unique standard point
st(x) X which is infinitesimally close to x, with
d(st(x), st(y)) = st( d(x, y)).
Exercise 1.8.3 (Automatic completeness).
Let (X, d) be a (standard) comQ
pact metric space, and let E = n En be a nonstandard subset of X
(i.e. an ultraproduct of standard subsets En of X), and let st : X X be
the standard part function as defined in the previous exercise. Show that
st(E) is complete (and thus compact, by the Heine-Borel theorem). (Hint:
take advantage of the countable saturation property from Section 1.7.)
Exercise 1.8.4. Let (X, d) be a compact metric space, and let En for
Q nN
be a sequence of non-empty compact subsets of X. Write E := n En
for the ultraproduct of the En , and let st : X X be the standard part
function as defined in the previous exercises. Show that st(E) is a nonempty compact subset of X, and that limn dH (En , st(E)) = o(1). Use
this to give an alternate proof of Lemma 1.8.3.
One can extend some of the above theory from compact metric spaces
to locally compact metric spaces.
Exercise 1.8.5. Let (X, d) be a (standard) locally compact metric space.
Let X O(X) X denote the set of ultralimits limn xn of precompact
sequences xn in X.
(i) Show that O(X) = X if and only if X is compact.
(ii) Show that there is a unique function st : O(X) X such that x is
infinitesimally close to st(x) for all x O(X).
(iii) If E is a nonstandard subset of X, show that st(E O(X)) is
complete.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

157

(iv) Show that if K is a compact set and > 0 is a (standard) real

number, then K E N (En ) and K En N (E) for all n
sufficiently close to .
Now we consider limits of metric spaces (Xn , dn ) that are not necessarily
all embedded in a single metric space. We begin by considering the case of
bounded metric spaces.
Definition 1.8.4 (Gromov-Hausdorff distance). The Gromov-Hausdorff distance dGH (X, Y ) between two bounded metric spaces (X, dX ), (Y, dY ) is the
infimum of the Hausdorff distance dH (XZ (X), Y Z (Y )), ranging over all
isometric embeddings XZ : X Z, Y Z : Y Z from X, Y respectively
to Z.
Exercise 1.8.6.
(i) Show that Gromov-Hausdorff distance is a pseudometric (i.e. symmetric, non-negative, and obeys the triangle inequality). (Hint: to
show that dGH (X, Z) dGH (X, Y ) + dGH (Y, Z), build a metric on
the disjoint union X ] Z which intuitively captures the idea of the
shortest path from X to Z via Y .)
(ii) If E is a dense subset of a bounded metric space X, show that
dGH (E, X) = 0. In particular, any bounded metric space E is at a
zero Gromov-Hausdorff distance from its metric completion E.
(iii) Show that if X, Y are compact metric spaces, then dGH (X, Y ) = 0
if and only if X and Y are isometric. (Hint: If dGH (X, Y ) = 0, find
a sequence of approximate isometries from X to Y and from Y
to X which almost invert each other. Then adapt the Arzel
a-Ascoli
theorem to take a limit (or one can use ultralimits and standard
parts).)
We say that a sequence Xn = (Xn , dn ) of bounded metric spaces converges in the Gromov-Hausdorff sense to another bounded metric space X
if one has dGH (Xn , X ) 0 as n . This is a more general notion
of convergence than Hausdorff convergence, and encompasses examples (iii)
and (iv) at the beginning of this section.
Exercise 1.8.7. Show that every compact metric space is the limit (in the
Gromov-Hausdorff sense) of finite metric spaces.
Now we turn to an important compactness result about Gromov-Hausdorff
convergence. We say that a sequence Xn = (Xn , dn ) of metric spaces is uniformly totally bounded if the diameters of the Xn are bounded, and, for
every > 0, there exists C such that each Xn can be covered by at most

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

158

1. Hilberts fifth problem

C balls of radius in the dn metric. (Of course, this implies that each Xn
is individually totally bounded.)
Proposition 1.8.5 (Uniformly total bounded spaces are Gromov-Hausdorff precompact). Let Xn be a sequence of uniformly totally bounded metric
spaces. Then there exists a subsequence Xnj which converges in the GromovHausdorff space to a compact limit X .
We will prove this proposition using ultrafilters. Let Xn = (Xn , dn ) be a
sequence of uniformly totally bounded metric spaces, which we will assume
to be standard (by defining the standardQuniverse in a suitable fashion).
Then we can form the ultraproduct X := n Xn and the ultralimit d :=
limn dn , thus d : X X R+ is a nonstandard metric (obeying the
nonstandard symmetry, triangle inequality, and positivity properties). As
the Xn are uniformly bounded, d has bounded range. If we then take the
standard part st(d) : X X R+ of d, then st(d) is a pseudometric on X
(i.e. it obeys all the axioms of a metric except possibly for positivity.) We
can then construct a quotient metric space X := X/ in the usual manner,
by declaring two points x, y in X to be equivalent, x y, if st(d)(x, y) = 0
(or equivalently if d(x, y) = o(1)). The pseudometric st(d) then descends to
a genuine metric d : X X R+ on X . The space (X , d ) is
known as a metric ultralimit (or ultralimit for short) of the (Xn , dn ) (note
that this is a slightly different usage of the term ultralimit from what we
have been using previously).
One can easily establish compactness:
Exercise 1.8.8 (Compactness).
(i) Show that (X , d ) is totally bounded. (Hint: use the uniform
total boundedness of the Xn together with Loss theorem.)
(ii) Show that (X , d ) is complete. (Hint: use countable saturation).
In particular, by the Heine-Borel theorem, X is compact.
Now we establish Gromov-Hausdorff convergence, in the sense that for
every standard > 0, one has dGH (Xn , X ) < for all n sufficiently close
to . From this it is easy to extract a subsequence Xnj that converges in
the Gromov-Hausdorff sense to X as required.
Fix a standard > 0. As X is totally bounded, we can cover it
by at most M balls B (x1, , /10), . . . , B (xM, , /10) of radius /10
(say) in the metric d for some standard natural number M . Lifting
back to the ultraproduct X, we conclude that we may cover X by M balls
B(x1 , 2/10), . . . , B(xM , 2/10) in the nonstandard metric d.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

159

Note that d (xi, , xj, ) is the standard part of d(xi , xj ), and in particular
d (xi, , xj, ) /10 < d(xi , xj ) < d (xi, , xj, ) + /10.
Each ball centre xi is an ultralimit xi = limn xi,n of points xi,n in Xn . By
Loss theorem, we conclude that for n sufficiently close to , Xn is covered
by the M balls Bn (x1,n , 2/10), . . . , Bn (xM,n , 2/10) in the metric dn , and
d (xi, , xj, ) /10 < dn (xi,n , xj,n ) < d (xi, , xj, ) + /10
for all 1 i, j M .
Because of this, we can embed both Xn and X in the disjoint union
Xn ] X , with a metric d(n) extending the metrics on Xn and X , with
the cross distances d(n) (x, y) between points x Xn and y X defined by
the formula
d(n) (x, y) = d(n) (y, x) := inf{d (xi, , y) + d (xi,n , x) + /2 : 1 i M };
informally, this connects each ball center xi, to its counterpart xi,n by a
path of length /2. It is a routine matter to verify that d(n) is indeed a
metric, and that Xn and X are separated by a Hausdorff distance less
than  in Xn ] X , and the claim follows.
Exercise 1.8.9. Prove Proposition 1.8.5 without using ultrafilters.
Exercise 1.8.10 (Completeness). Let Xn be a sequence of bounded metric
spaces which is Cauchy in the Gromov-Hausdorff sense (i.e. dGH (Xn , Xm )
0 as n, m ). Show that Xn converges in the Gromov-Hausdorff sense
to some limit X.
Now we generalise Gromov-Hausdorff convergence to the setting of unbounded metric spaces. To motivate the definition, let us first give an equivalent form of Gromov-Hausdorff convergence:
Exercise 1.8.11. Let (Xn , dn ) be a sequence of bounded metric spaces, and
let (X , d ) be another bounded metric space. Show that the following are
equivalent:
(i) (Xn , dn ) converges in the Gromov-Hausdorff sense to (X , d ).
(ii) There exist maps n : X Xn which are asymptotically isometric isomorphisms in the sense that supx,yX |dn (n (x), n (y))
d(x, y)| 0 and supxn Xn distn (xn , n (X )) 0 as n .
Define a pointed metric space to be a triplet (X, d, p), where (X, d) is a
metric space and p is a point in X.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

160

1. Hilberts fifth problem

Definition 1.8.6 (Pointed Gromov-Hausdorff convergence). A sequence of

pointed metric spaces (Xn , dn , pn ) is said to converge in the pointed GromovHausdorff sense to another pointed metric space (X , d , p ) if there exists
a sequence of maps n : X Xn such that
(1) dn (n (p ), pn ) 0 as n ;
(2) (Asymptotic isometry) For each R > 0, one has
sup

|dn (n (x), n (y)) d (x, y)| 0

x,yB (p ,R)

as n ;
(3) (Asymptotic surjectivity) For each R0 > R > 0, one has
sup

distn (x, n (B (p , R0 ))) 0

xBn (pn ,R)

as n .
Exercise 1.8.12. Verify that all the examples (i)-(v) given at the start of
the section are examples of pointed Gromov-Hausdorff convergence, once
one selects a suitable point in each space.
The above definition is by no means the only definition of pointed
Gromov-Hausdorff convergence. Here is another (which is basically the original definition of Gromov [Gr1981]):
Exercise 1.8.13. Let (Xn , dn , pn ) be a sequence of pointed metric spaces,
and let (X , d , p ) be another pointed metric space. Show that the following are equivalent:
(1) (Xn , dn , pn ) converges in the pointed Gromov-Hausdorff sense to
(X , d , p ).
(2) There exist metrics dn on the disjoint union Xn X extending
the metrics dn , d such that for some sequence n 0, one has
dn (pn , p ) n , Bn (pn , 1/n ) Nn (X ), and B (p , 1/n )
Nn (Xn ).
Using this equivalence, construct a pseudometric between pointed metric
spaces that describes pointwise Gromov-Hausdorff convergence.
Exercise 1.8.14. Let (Xn , dn , pn ) be a sequence of pointed metric spaces
which converge in the pointed Gromov-Hausdorff sense to a limit (X , d , p ),
0 , d0 , p0 ). Suppose that these two limit spaces
and also to another limit (X

0 (p0 , R)
are proper, which means that the closed balls B (p , R) and B

are compact. Show that the two limit spaces are pointedly isometric, thus
0 that maps p
0
there is an isometric isomorphism : X X
to p .
In the compact case, Gromov-Hausdorff convergence and pointwise GromovHausdorff convergence are almost equivalent:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

161

Exercise 1.8.15. Let (Xn , dn , pn ) be a sequence of bounded pointed metric

spaces, and let (X , d , p ) be a compact pointed metric space.
(i) Show that if (Xn , dn , pn ) converges in the pointed Gromov-Hausdorff
sense to (X , d , p ), then (Xn , dn ) converges in the GromovHausdorff sense to (X , d ).
(ii) Conversely, if (Xn , dn ) converges in the Gromov-Hausdorff sense to
(X , d ), show that some subsequence (Xnj , dnj , pnj ) converges
in the pointed Gromov-Hausdorff sense to (X , d , q) for some
q X .
There is a compactness theorem for pointed Gromov-Hausdorff convergence analogous to that for ordinary Gromov-Hausdorff convergence (Proposition 1.8.5):
Proposition 1.8.7 (Locally uniformly total bounded spaces are pointed
Gromov-Hausdorff precompact). Let Xn = (Xn , dn , pn ) be a sequence of
pointed metric spaces such that the balls Bn (pn , R) are uniformly totally
bounded in n for each fixed R > 0. Then there exists a subsequence Xnj
which converges in the pointed Gromov-Hausdorff space to a limit X =
(X , d , p ) which is proper (i.e. every closed ball is compact).
We sketch a proof of this proposition in the exercise below.
Exercise 1.8.16. Let Xn beQas in the above proposition; we assume the
:= n Xn be the ultraproduct of the Xn , and
Xn to be standard. Let X Q
Q
define the ultralimits d := n dn and p := n . Let O(X) := {x X :
d(x, p) = O(1)} be the points in X that are a bounded distance from p.
(i) Show that st(d) : O(X) O(X) R+ is a pseudometric on O(X).
(ii) Show that the associated quotient space X := O(X)/ with the
quotient metric d is a proper metric space.
(iii) Show that some subsequence of the (Xn , pn ) converge in the pointed
Gromov-Hausdorff sense to (X , p ), where p is the image of p
under the quotient map, thus establishing Proposition 1.8.7.
Exercise 1.8.17. Establish Proposition 1.8.7 without using ultrafilters.
Exercise 1.8.18. Let G be a locally compact group with a Gleason metric
d. Show that the pointed metric spaces (G, nd, 1) converge in the pointed
Gromov-Hausdorff sense as n to the vector space L(G) with the metric
dL(G) (x, y) := kxyk and distinguished point 0, where the norm kk on L(G)
was defined in Exercise 1.3.8.
Remark 1.8.8. The above exercise suggests that one could attack Hilberts
fifth problem by somehow blowing up the locally compact group G around

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

162

1. Hilberts fifth problem

the origin and extracting a Gromov-Hausdorff limit using ultraproducts.

This approach can be formalised using the language of nonstandard analysis, and in particular can be used to describe Hirschfelds nonstandard
solution [Hi1990] to Hilberts fifth problem, as well as the later work of
Goldbring [Go2010]. However, we will not detail this approach extensively
here (though, on some level, it contains much the same ingredients as the
known standard solutions to that problem, such as the one given in previous sections).
1.8.2. Sanders-Croot-Sisask theory. We now prove Sanders lemma (Lemma
1.8.1), which roughly speaking will be needed to establish an important total boundedness property of ultra approximate groups, which in turn is
necessary to ensure local compactness for models of such groups.
Let A be a K-approximate group for some K > 1, and let m be a (large)
natural number. Our task is to locate a large set S such that the iterated
power S m is contained inside A4 . Sanders strategy for doing this is to pick
a set S that nearly stabilises a set B that is comparable in some sense to
A. More precisely, suppose we have non-empty finite sets S and B with the
property that
1
|B\sB| < |B|
m
for all s S. Then an easy induction shows that
|B\gB| <

k
|B|
m

for all g S k and all k 1. In particular, B and gB are not disjoint

whenever g S m , which means that S m BB 1 . Thus, to prove Sanders
lemma, it suffices to find a non-empty set B A2 with the property that
the set


1
(1.66)
S := s G : |B\sB| < |B|
m
has cardinality K,m |A|. (Note that this set S will automatically be symmetric and contain the origin.)
Remark 1.8.9. The set (1.66) is known as a symmetry set of B and is
sometimes denoted Sym1/m (B). One can also interpret this set as a ball,
1
using the mapping g 7 |B|
1gB of G into `1 (G) to pull back the `1 (G) metric
1
to G, in which case S is the ball of radius m
centred at the origin.
It remains to find a set B for which the set (1.66) is large. To motivate
how we would do this, let us naively try setting B := A2 . If the set (1.66)
associated to this set is large, we will be done, so let us informally consider

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

163

the opposite case in which S is extremely small; in particular, we have

1 2
|A |
m
for most choices of g G. In particular,
|A2 \gA2 |

|A2 gA2 | (1

1
)|A2 |.
m

Now observe that (A gA)A is contained in A2 gA2 , and so

|(A gA)A| (1

1
)|A2 |.
m

We have thus achieved a dichotomy: either the choice B := A2 works,

or else the set B 0 := A0 A is significantly smaller than B = A2 for some
A0 = A gA. We can then try to repeat this dichotomy, to show that the
choice B 0 = A0 A either works, or else the set B 00 := A00 A is significantly
smaller than B 0 = A0 A for some A00 = A0 g 0 A0 . We can keep iterating this
dichotomy, creating ever smaller sets of the form Bn = An A; but on the
other hand, these sets should be at least as large as |A| (provided that we
choose g, g 0 , etc. to prevent the sets A0 , A00 , etc. from becoming completely
empty). So at some point this iteration has to terminate, at which point we
should get a set that works.
The original paper of Sanders [Sa2009] contains a formalisation of the
above argument. We present a slightly different arrangement of the argument below, which is focused on making sure that sets such as A0 do not get
too small.
For any 0 < t 1, define the quantity
 0

|A A|
0
0
f (t) := inf
: A A; |A | t|A| .
|A|
Then f is a non-decreasing function that takes values between 1 and K. By
the pigeonhole principle, we can find t K,m 1 such that
 2 
1
t
(1.67)
f
> (1 )f (t).
2K
m
(The reasons for these particular choices of parameters will become clearer
shortly.) Fix this t, and let A0 attain the infimum for f (t), thus A0 A,
|A0 | t|A|, and
|A0 A| = f (t)|A|.
(Such an A0 exists since the infimum is only being taken over a finite set.)
Set B := A0 A, and let S be the set in (1.66). Observe that if g 6 S, then
|A0 A\gA0 A|

1 0
|A A|,
m

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

164

1. Hilberts fifth problem

and so by arguing as before we see that

1 0
|A A|.
m
Since |A2 | |A| and |A0 A| = f (t)|A|, we thus have
|(A0 gA0 )A| |A0 A|

1
)f (t)|A|.
m
In particular, from (1.67) and the definition of f , we conclude that
|(A0 gA0 )A| (1

|A0 gA0 | <

t2
|A|
2K

for all g 6 S. In other words,



t2
0
0
(1.68)
S g G : |A gA |
|A| .
2K
So to finish the proof of Sanders lemma, it will suffice to obtain a lower
bound on the right-hand side of (1.68). But this can be done by a standard
Cauchy-Schwarz argument: starting with the identity
X X
1aA0 (x) = |A||A0 | t|A|2
xA2 aA

and using the Cauchy-Schwarz inequality and the bound |A2 | K|A|, we
conclude that
X X
t2
1aA0 (x)1bA0 (x) |A|3
K
2
xA a,bA

and thus
X

|aA0 bA0 |

a,bA

t2
|A|3 .
K

By the pigeonhole principle, we may thus find a A such that

X
t2
|aA0 bA0 | |A|2 ,
K
bA

and thus (setting g :=

a1 b)
X
ga1 A

|A0 gA0 |

t2
|A|2 .
K

Since a1 A has cardinality |A|, this forces |A0 gA0 | to exceed

t2
least 2K
|A| values of g, and the claim follows.

t2
2K |A|

for at

Remark 1.8.10. The lower bound on S obtained by this argument is of

the shape |S| exp(K O(m) )|A|. This is not optimal; see Remark 1.8.12
below.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

165

We now present an alternate approach to Sanders lemma, using the

Croot-Sisask theory [CrSi2010] of almost periods. Whereas in the Sanders
argument, one selected the set S to be the set of group elements that almost
stabilised a set B, we now select S to be the set of group elements that
almost stabilise (or are almost periods of) the convolution
X
1A 1A (x) :=
1A (y)1A (y 1 x).
yG

`2 (G)

It is convenient to work in the

metric. Since the function 1A 1A has
an `1 norm of |A|2 and is supported in A2 , which has cardinality at most
K|A|, we see from Cauchy-Schwarz that
k1A 1A k`2 (G) |A|3/2 /K 1/2 .
We now set
(
S :=

g G : k1A 1A (g)1A 1A k`2 (G)

1 |A|3/2
<
m K 1/2

)
.

Exercise 1.8.19. Show that S is symmetric, contains the origin, and that
S m A4 .
To finish the proof of Sanders lemma, it suffices to show that there are
lots of almost periods of 1A 1A in the sense that |S| K,m |A|.
The key observation here is that the translates (g)1A 1A of 1A 1A ,
as g varies in A, range in a totally bounded set. This in turn comes from
a compactness property of the convolution operator f 7 f 1A . when
f is supported on A2 . Croot and Sisask establish this by using a random
sampling argument to approximate this convolution operator by a bounded
rank operator. More precisely, let M 1 be an integer parameter to be
chosen later, and select M sample points y1 , . . . , yM from A2 independently
and uniformly at random (allowing repetitions). We will approximate the
operator
X
T f := f 1A =
f (y)1yA
yA2

for f supported on A2 by the operator

M

|A2 | X
Sf :=
f (yi )1yi A .
M
i=1

The point is that as M gets larger, Sf becomes an increasingly good

approximation to T f :
Exercise 1.8.20. Let f be a function supported on A2 that is bounded in
magnitude by 1. Show that
ESf = T f

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

166

1. Hilberts fifth problem

and
EkSf T f k2`2 (G) K |A|3 /M.
In particular, with probability at least 1/2, one has
kSf T f k`2 (G) K |A|3/2 M 1/2 .
Thus, for any g A, we have
kS1gA T 1gA k`2 (G) K |A|3/2 M 1/2
with probability at least 1/2. By the pigeonhole principle (or the first moment method), we thus conclude that there exists a choice of sample points
y1 , . . . , yM for which
(1.69)

kS1gA T 1gA k`2 (G) K |A|3/2 M 1/2

for at least |A|/2 choices of g A.

Let A0 denote the set of all g A indicated above. For each g A0 ,
2|
the function S1gA is a linear combination of the M functions |A
M 1yi A with
2

|
2
coefficients between 0 and 1. Each of these functions |A
M 1yi A has an ` (G)
norm of OK (|A|3/2 ). Thus, by the pigeonhole principle, one can find a subset
A00 of A0 of size |A00 | K,m,M |A| such that the functions S1gA for g A00
3/2

1 |A|
all lie within 2m
in `2 (G) norm of each other. If M is large enough
K 1/2
depending on K, m, we then conclude from (1.69) and the triangle inequality
3/2

1 |A|
that the functions T 1gA for g A00 all lie within m
in `2 (G) norm of
K 1/2
00
each other. Translating by some fixed element of A , we obtain the claim.

Remark 1.8.11. For future reference, we observe that the above argument
did not need the full strength of the hypothesis that A was an K-approximate
group; it would have sufficed for A to be finite and non-empty with |A2 |
K|A|.
Remark 1.8.12. The above version of the Croot-Sisask argument gives
a lower bound on S of the shape |S| exp(O(m2 K log K))|A|. As was
observed in a subsequent paper of Sanders [Sa2010], by optimising the argument (in particular, replacing `2 with `p for a large value of p, and replacing
1A 1A by 1A 1A2 ), one can improve this to |S| exp(O(m2 log2 K))|A|.
Remark 1.8.13. It is also possible to replace the random sampling argument above by a singular value decomposition of T (restricted to something
like A2 ) to split it as the sum of a bounded rank component and a small
operator norm component, after computing the Frobenius norm of T in order to limit the number of large singular values. (In the abelian setting,
this corresponds to a Fourier decomposition into large and small Fourier
coefficients, which is a fundamental tool in additive combinatorics.) We will
however not pursue this approach here.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

167

As mentioned previously, the Sanders lemma will be useful in building

topological group structure on ultra approximate groups A (and the group
hAi that they generate). The connection can be seen from the simple observation that if U is a neighbourhood of the identity in a topological group
and m 1, then there exists another neighbourhood S of the identity such
that S m U . However, in addition to this multiplicative structure, we
will also need to impose some conjugacy structure as well. (Strangely, even
though the conjugation operation (g, h) 7 ghg 1 can be defined in terms of
the more primitive operations of multiplication (g, h) 7 gh and inversion
g 7 g 1 , it almost seems to be an independent group operation in some
ways, at least for the purposes of studying approximate groups, and the
most powerful results tend to come from combining multiplicative structure
and conjugation structure together. I do not know a fundamental reason for
this product-conjugation phenomenon but it does seem to come up a lot
in this subject.) Given two non-empty subsets A, B of a group G, define
AB := {ab : a A, b B}
where ab := b1 ab is the conjugate of a by b. (Thus, for instance, H B = H
whenever H is a subgroup normalised by B.)
We can now establish a stronger version of the Sanders lemma:
Lemma 1.8.14 (Normal Sanders lemma). Let A be a finite K-approximate
group in a (global) group G, let m 1, and let B be a symmetric subset of
A with |B| |A| for some > 0. Then there exists a symmetric subset N
m
of A4 with |N | K,m, |A| containing the identity such that (N A )m B 4 .
Roughly speaking, one can think of the set S in Lemma 1.8.1 as a finite
index subgroup of A4 , while the set in Lemma 1.8.14 is a finite index
normal subgroup of A4 . To get from the former to the latter, we will
mimic the proof of the following well-known fact:
Lemma 1.8.15. Suppose that H is a finite index subgroup of a group G.
Then there is a finite index normal subgroup N of G that is contained in H.
Proof. Consider all the conjugates H g of the finite index subgroup H as g
ranges over G. As all the elements g from the same right coset Hk give the
same conjugate H g , and H is finite index, we see that there are only finitely
many such conjugates. Since the intersection of two finite indexTsubgroups
is again a finite index subgroup (why?), the intersection N := gG H g is
thus a finite index subgroup of G. But we have N = N g for all g G, and
so N is normal as desired.

Remark 1.8.16. An inspection of the argument reveals that if H had index
k in G, then the normal subgroup N has index at most k k . One can do

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

168

1. Hilberts fifth problem

slightly better than this by looking at the left action of G on the k-element
quotient space G/H, which one can think of as a homomorphism from G to
the symmetric group Sk of k elements. The kernel N of this homomorphism
is then clearly a normal subgroup of G of index at most |Sk | = k! that is
contained in H. However, this slightly more efficient argument seems to
be more fragile than the more robust proof given above, in that it is not
obvious (to me, at least) how to adapt it to the approximate group setting
of the Sanders lemma. (This illustrates the advantages of knowing multiple
proofs for various basic facts in mathematics.)
Inspired by the above argument, we now prove Lemma 1.8.14. The main
difficulty is to find an approximate version of the claim that the intersection of two finite index subgroups is again a finite index subgroup. This is
provided by the following lemma:
Lemma 1.8.17. Let A be a K-approximate group, and let A1 , A2 A be
such that |A1 | 1 |A| and |A2 | 2 |A|. Then there exists a subset B of A
1
with BB 1 A1 A1
1 A2 A2 and |B| 1 2 |A|/K.
1
2
Proof. Since A1
1 A2 A , we have |A1 A2 | K|A|. It follows that there
is some x with at least 1 2 |A|/K representations as a1
1 a2 . Let B be the
set of all values of a2 that appear. Obviously BB 1 A2 A1
2 . Suppose
1
0
0
that a2 , a2 B. Then there are a1 , a1 such that x = a1 a2 = (a01 )1 a02 , and
1 lies in A A1 as well.
0 1

so a01 a1
1 1
1 = a2 a2 . Thus BB

Now we prove Lemma 1.8.14. By Lemma 1.8.1, we may find a symmetric

set S of B 4 containing the identity such that S 8m (say) is contained in B 4
and |S| K,m, |A|. We need the following simple covering lemma:
Exercise 1.8.21 (Ruzsa covering lemma). Let A, B be finite non-empty
subsets of a group G such that |AB| K|B|. Show that A can be covered
by at most K left-translates aBB 1 of BB 1 for various a A. (Hint: find
a maximal disjoint family of aB with a A.)
From the above lemma we see that A can be covered by OK,m (1) left
translates of S 2 . As Am can be covered by OK,m (1) left-translates of A, we
conclude that Am can be covered by OK,m (1) left-translates of S 2 . Since
T
S 2 A8 , we conclude that Am Jj=1 aj S 2 for some J = OK,m (1) and
a1 , . . . , aJ Am+8 .
1
The conjugates a1
all lie in A2m+20 . By many applications of
j Saj
Lemma 1.8.17, we see that we may find a subset D of A2m+4 with |D| K,m
|A| such that
DD1 aj S 4 a1
j

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

169

for all j = 1, . . . , J. In particular, if N := DD1 , then N aj S 4 for all

m
m
j = 1, . . . , J, and thus N A S 8 . Thus (N A )m B 4 , and the claim
follows.
Remark 1.8.18. The quantitative bounds on |N | given by this argument
are quite poor, being of triple exponential type (!) in K. It is likely that
this can be improved by a more efficient argument.
1.8.3. Locally compact models of ultra approximate groups. We
now use the normal Sanders lemma to place a topology on ultra approximate groups. We first build a neighbourhood base associated to an ultra
approximate group:
Exercise 1.8.22. Let A be an ultra approximate group. Show that there
exist a sequence of ultra approximate groups
A4 = A0 A1 A2 . . .
such that

2
(AA
m ) Am1
for all m 1, and such that A4 can be covered by finitely many lefttranslates of Am for each m. (Hint: you will need Lemma 1.8.14, Lemma
1.8.17, Exercise 1.8.21, and some sort of recursive construction.)

Remark 1.8.19. A simple example to keep in mind here is the nonstandard

interval A = [N, N ] for some unbounded nonstandard natural number N ,
with Am := [4N/2m , 4N/2m ].
S
m
This gives a good topology on the group hAi =
m=1 A generated by
A:
Exercise 1.8.23. Let A be an ultra approximate group, and let hAi be
the group generated by A. Let A0 , A1 , . . . be as in the preceding exercise.
Given a subset E of hAi, call a point g in E an interior point of E if one
has gAm E for some (standard) m. Call E open if every element of E is
an interior point.
(i) Show that this defines a topology on E.
(ii) Show that this topology makes hAi into a topological group (i.e.
the group operations are continuous).
(iii) Show that this topology is first countable (and thus pseudo-metrisable
by the Birkhoff-Kakutani theorem, Theorem 1.5.2).
(iv) Show that one can build a left-invariant pseudometric d : hAi
hAi R+ on hAi which generates the topology on hAi, with the
property that
B(1, c2m ) Am B(1, C2m )

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

170

1. Hilberts fifth problem

for all (standard) m 0 and some (standard) C > c > 0. (Hint:

inspect the proof of the Birkhoff-Kakutani theorem.)
(v) Show that hAi with this pseudometric is complete (i.e. every Cauchy
sequence is convergent). (Hint: use the countable saturation property.)
(vi) Show that Am is totally bounded for each standard m (i.e. covered
by a finite number of -balls for each > 0).
(vii) Show that hAi is locally compact.
With this topology, the group hAi becomes a locally compact topological
group. However, in general this group will not be Hausdorff, because the
identity 1 T
need not be closed. Indeed, it is easy to see that the closure of 1
is the set
m=1 Am , which is not necessarily trivial. For instance, using the
example from Remark 1.8.19, hAi is the group {n Z : n = O(N )}, and the
closure of the identity is {n Z : n = o(N )}. However, we may quotient out
by this closure of the identity to obtain a locally compact Hausdorff group L,
which is now metrisable (with the metric induced from the pseudometric d on
hAi). Let : hAi L be the quotient homomorphism. This homomorphism
obeys a number of good properties, which we formalise as a definition:
Definition 1.8.20. Let A be an ultra approximate group. A (global) good
model for A is a homomorphism : hAi L from hAi to a locally compact
Hausdorff group L that obeys the following axioms:
(1) (Thick image) There exists a neighbourhood U0 of the identity in L
such that 1 (U0 ) A and U0 (A). (In particular, the kernel
of lies in A.)
(2) (Compact image) (A) is precompact.
(3) (Approximation by nonstandard sets) Suppose that F U U0 ,
where F is compact and U is open. Then there exists a nonstandard
finite set B (i.e. an ultraproduct of finite sets) such that 1 (F )
B 1 (U ).
We will often abuse notation and refer to L as the good model, rather than
. (Actually, to be truly pedantic, it is the ordered pair (, L) which is the
good model, but we will not use this notation often.)
Remark 1.8.21. In the next section we will also need to consider local good
models, which only model (say) A8 rather than all of hAi, and in which L is
a local group rather than a global one. However, for simplicity we will not
discuss local good models for the moment.
Remark 1.8.22. The thick and compact image axioms imply that the geometry of L in some sense corresponds to the geometry of A. In particular,

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

171

if a is an element of hAi, then a will be small (in the sense that it lies
in A) if (a) is close to the identity, and large (in the sense that it lies
outside A) if (a) is far away from the identity. Thus is faithful in some
coarse or macroscopic sense. The inclusion U0 (A) is a sort of local
surjectivity condition, and ensures that L does not contain any excess or
redundant components. The approximation by nonstandard set axiom is
a technical measurability axiom, that ensures that the model of the ultra
approximate group actually has something nontrivial to say about the finite
approximate groups that were used to build that ultra approximate group
(as opposed to being some artefact of the ultrafilter itself).
Example 1.8.23. We continue the example from Remark 1.8.19. A good
model for A = [N, N ] is provided by the homomorphism : hAi R
given by the formula (x) := st(x/N ). The thick image and compact image
properties are clear. To illustrate the approximation by nonstandard set
property, take F = [r, r] and U = (s, s) for some (standard) real numbers
0 < r < s. The preimages 1 (F ) = {n Z : |n| rN + o(N )} and
1 (U ) = {n Z : |n| < (s )N for some standard > 0} are not
nonstandard finite sets (why? use the least upper bound axiom), but one
can find a nonstandard integer M such that r < st(M/N ) < s, and [M, M ]
will be a nonstandard finite set between F and U .
The following fundamental observation is essentially due to Hrushovski
[Hr2012]:
Exercise 1.8.24 (Existence of good models). Let A be an ultra approximate
group. Show that A4 has a good model by a locally compact Hausdorff
metrisable group, given by the construction discussed previously.
Exercise 1.8.25. The purpose of this exercise is to show why it is necessary
to model A4 , rather than A, in Exercise 1.8.24. Let F2 be the field of two
elements. For each positive integer n, let An be a random subset of F2n
formed by selecting one element uniformly at random from the set x, x + en
for each x F2n1 {0}, and also selecting the identity 0. Clearly, An is a
2-approximate group, since An is contained in F2n = An + {0, en }.
(i) Show that almost surely, for all but finitely many n, An does not
contain any set of the form B + B, with |B| 20.9n . (Hint: use
the Borel-Cantelli lemma.)
Q
(ii) Show that almost surely, the ultraproduct n An cannot be
modeled by any locally compact group.
Let us now give some further examples of good models, beyond that
given by Example 1.8.23.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

172

1. Hilberts fifth problem

Example 1.8.24 (Nonstandard finite groups). Suppose that Q

An is a sequence of (standard) finite groups; then the ultraproduct A := n An is
an ultra approximate group. In this case, A is in fact a genuine group, thus
A = hAi. In this case, the trivial homomorphism : hAi L = {1} to the
trivial group {1} is a good model of A. Conversely, it is easy to see that this
is the only case in which {1} is a good model for A.
Example 1.8.25 (Generalised arithmetic progression). We still work in
the integers Z, but now take An to be the rank two generalised arithmetic
progression
An := P (1, n10 ; n, n) := {a + bn10 : a, b {n, . . . , n}}.
Q
Then the ultraproduct A := n An is the subset of the nonstandard
integers Z of the form
A = P (1, N 10 ; N, N ) = {a + bN 10 : a, b {N, . . . , N }},
where N is the unbounded natural number N := limn n. This is an ultra
approximate group, with
hAi = {a + bN 10 : a, b Z; a, b = O(N )}.
Then hAi can be modeled by the Euclidean plane R2 , using the model maps
: hAi R2 defined for each standard m by the formula


a
b
10
(a + bN ) := st , st
N
N
whenever a, b = O(N ). The image (Am ) is then the square [m, m]2
for any standard m. Note here that while A lives in a one-dimensional
group Z, the model R2 is two-dimensional. This is also reflected in the
volume growth of the powers Am
n of An for small m and large n, which grow
quadratically rather than linearly in m. Informally, A is modeled by the
unit square in R2 .
Exercise 1.8.26. With the notation of Example 1.8.25, show that A cannot
be modeled by the one-dimensional Lie group R. (Hint: If A was modeled
by R, conclude that Am could be covered by O(m) translates of A for each
standard m.)
Exercise 1.8.27 (Heisenberg box, I). We take each An to be the nilbox
n 1 xn zn   1 Z Z 
o
2
0 1 yn 0 1 Z : |xn |, |yn | n, |zn | n
An :=
.
0 0

0 0 1

Q
Consider the ultraproduct A :=  n An ; this is a subset of
 1the
nilpoxz
1 Z Z
tent (nonstandard) group 0 1 Z , consisting of all elements 0 1 y with
0 0

0 0 1

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

173

|x|, |y| N and |z| N 2 , where N := limn n. Thus



1 O(N ) O(N 2 )
hAi = 0 1 O(N ) .
0

Consider now the map

: hAi

1 R R
0 1 R
0 0 1

defined by
(1.70)

 1 x z 
0 1 y
0 0 1

 1 st
:=

0
0

x
N

1
0

z
N2
y
st N

st


.

(i) Show that A A1 is an ultra approximate group.

(ii) Show that is a good model of A A1 .
Exercise 1.8.28 (Heisenberg box, II). This is a variant of the preceding
exercise, in which the An is now defined as
o
n 1 xn zn 
10
0 1 yn : |xn |, |yn | n, |zn | n
(1.71)
An :=
0 0 1
Q
so that the ultralimit A := n An takes the form
o
n 1 x z   1 Z Z 
0 1 y 0 1 Z : |x|, |y| N, |z| N 10
A :=
0 0

0 0 1

and


hAi :=

1 O(N ) O(N 10 )
0 1
O(N )
0 0
1


,

where N := limn n. Now consider the map

: hAi8 R3
defined by

 x
y
z 
= st , st , st 10 .
N
N
N
is an ultra approximate group.

 1 x z 

(i) Show that A

0 1 y
0 0 1
A1

(ii) Show that is a good model of A A1 .

(iii) Show that A A1 cannot be modeled by R2 , or by the Heisenberg
group.
Remark 1.8.26. Note in the above exercise that the homomorphism : A8
R3 is not associated to any exact homomorphisms n from A8n to R3 . Instead, it is only associated to approximate homomorphisms
x y z 
 1 xn zn 
n
n
n
0 1 yn
:=
, , 10
n
n n n
0 0 1
into R3 . Such approximate homomorphisms are somewhat less pleasant
to work with than genuine homomorphisms; one of the main reasons why
we work in the ultraproduct setting is so that we can use genuine group
homomorphisms.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

174

1. Hilberts fifth problem

We also note the sets Am

n for small m and large n grow cubically in m
in Exercise 1.8.27, and quartically in m in Exercise 1.8.28. This reflects the
corresponding growth rates in R3 and in the Heisenberg group respectively.
Finally, we observe that the nonabelian structure of the ultra approximate group A is lost in the model group L, because the nonabelianness is
infinitesimal at the scale of A. More generally, good models can capture
the macroscopic structure of A, but do not directly see the microscopic
structure.
The following exercise demonstrates that model groups need not be simply connected.
Exercise 1.8.29 (Models of Bohr sets). Let be a standard irrational, let
0 < < 0.1 be a standard real number, and let An Z be the sets
An := {m [n, n] : kmkR/Z }
where kxk denotes the distance to the nearest integer. Set A :=

n An .

(i) Show that A is an ultra approximate group.

(ii) Show that R/Z R is a good model for A.
(iii) Show that R2 is not a good model for A. (Hint: consider the
growth of Am , as measured by the number of translates of A needed
to cover this set.)
(iv) Show that R is not a good model for A. (Hint: consider the decay
of the sets {g : g, g 2 , . . . , g m A}, as measured by the number of
translates of this set needed to cover A.)
Exercise 1.8.30 (Haar measure).
Q Let : hAi L be a good model for an
ultra approximate group A = n An by a locally compact group L. For
any continuous compactly supported function f : L R, we can define a
functional I(f ) by the formula
P
F + (a)
I(f ) = inf st aA
|A|
+
+
where F + = limn
P Fn is+the ultralimit of functions Fn : An R, with the
nonstandard real aA F and nonstandard natural number |A| defined in
the usual fashion as
X
X
F + (a) := lim
Fn+ (an )
aA

an An

and
|A| := lim |An |,
n

and the infimum is over all F + for which F + (a) f ((a)) for all a A.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

175

(i) Establish the equivalent formula

P
F (a)
P
I(f ) = sup st aA
aA 1
where the supremum is over all F for which F (a) f ((a)) for
all a A.
(ii) Show that
R there exists a bi-invariant Haar measure on G such that
I(f ) = G f d for all continuous compactly supported f : L R.
(In particular, this shows that L is necessarily unimodular.)
(iii) Show that
(F )|A| |A0 | (U )|A|
whenever F U U1 , F is compact, U is open, and A0 is a
nonstandard set with
1 (F ) A0 1 (U ).
1.8.4. Lie models of ultra approximate groups. In the examples of
good models in the previous section, the model group L was a Lie group. We
give now give some examples to show that the model need not initially be of
Lie type, but can then be replaced with a Lie model after some modification.
Example 1.8.27 (Nonstandard cyclicQgroup, revisited). Consider the nonstandard cyclic group A := Z/2N Z = n Z/2n Z. This is a nonstandard
finite group and can thus be modeled by the trivial group {1} as discussed
in Example 1.8.24. However, it can also be modeled by the compact abelian
group Z2 of 2-adic integers using the model : A Z2 defined by the
formula
(a) := lim a mod 2n
n

where for each standard natural number n, a mod 2n {0, . . . , 2n1 } is the
remainder of a modulo 2n (this is well-defined in A) and the limit is in the
2-adic metric. Note that the image (A) of A is the entire group Z2 , and
conversely the preimage of Z2 in A8 = A is trivially all of Z/2N Z; as such,
one can quotient out Z2 in this model and recover the trivial model of A.
Example 1.8.28 (Nonstandard abelian 2-torsion group). In a similar spirit
to
example, the nonstandard 2-torsion group A := (Z/2Z)N =
Q the preceding
n
N by
n (Z/2Z) can be modeled by the compact abelian group (Z/2Z)
the formula
(a) := lim n (a)
n

where n : A (Z/2Z)n is the obvious projection, and the limit is in the

product topology of (Z/2Z)N . As before, we can quotient out (Z/2Z)N and
model A instead by the trivial group.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

176

1. Hilberts fifth problem

Remark 1.8.29. The above two

Q examples can be generalised to model any
nonstandard finite group G = n Gn equipped with surjective homomorphisms from Gn+1 to Gn by the inverse limit of the Gn .
Exercise 1.8.31 (Lamplighter group). Let F2 be the field of two elements.
G be the lamplighter group Z n F2Z , where Z acts on F2Z by the shift T : F2Z
defined by T (ak )kZ := (ak1 )kZ . Thus the group law in G is given by
(i, x)(j, y) := (i + j, x + T i y).
For each n, we then set An G to be the set
An := {(i, x) G : i {1, 0, +1}; x F2n },
where we identify F2n with the space of elements (ak )kZ of F2Z such that
ak 6= 0Ponly for k {1, . . . , n}. Let F2 ((t)) be the ring of formal Laurent
series kZ ak tk in which all but finitely many of the ak for negative k are
zero. We let G0 be the modified lamplighter group Z n F2 ((t)), where Z
acts on F2Z by the shift T : f 7
P tf . We give F2 ((t)) a topology by assigning
each non-zero Laurent series kZ ak tk a norm of 2k , where k is the least
integer for which ak 6= 0; this induces a topology on G0 via the product
topology construction.
Q
We will model the ultraproduct A := n An Z n (Z/2Z)Z (or
more precisely, the set A A1 , since A is not quite symmetric) by the
group
G0 Z G0 := {((i, x), (j, y)) G0 G0 : i = j}
using the map
X
X
(n)
(n)
(n)
((i, lim (ak )kZ )) := ((i,
lim ak tk ), (i,
lim ank tk )).
n

kZ

kZ

Roughly speaking, (a) captures the behaviour of a at the two ends of

F2N , where N := limn n. We give G0 Z G0 the topology induced from
the product topology on G0 G0 .
(i) Show that A A1 is an ultra approximate group.
(ii) Show that is a good model of A A1 .
(iii) Show that is no longer a good model if one projects G0 Z G0 to
the first or second copy of G0 , or to the base group Z.
(iv) Show that A A1 does not have a good model by a Lie group L.
(Hint: L does not contain arbitrarily small elements of order two,
other than the identity.)
Remark 1.8.30. In the above exercise, one needs a moderately complicated
(though still locally compact) group G Z G to properly model A and its
powers Am . This can also be seen from volume growth considerations: Am
n
grows like 4m for fixed (large n), which is also the rate of volume growth of

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

177

(A) in G Z G, whereas the volume growth in a single factor G would only

grow like 2m , and the volume growth in Z is only linear
However, if
Q in m.
0
0
0
we pass to the large subset A of A defined by A := n An , where
A0n := {(i, x) G : i = 0; x (Z/2Z)n }
then A0 is now a nonstandard finite group (isomorphic to the group (Z/2Z)N
considered in Example 1.8.28) and can be modeled simply by the trivial
group {1}. Thus we see that we can sometimes greatly simplify the modeling of an ultra approximate group by passing to a large ultra approximate
subgroup.
By using the Gleason-Yamabe theorem, one can formalise these examples. Given two ultra approximate groups A0 , we say that A0 is an large ultra
approximate subgroup of A if (A0 )4 A4 and A can be covered by finitely
many left-translates of A0 .
Theorem 1.8.31 (Hrushovskis Lie model theorem). Let A be an ultra
approximate group. Then there exists a large ultra approximate subgroup A0
of A that can be modeled by a connected Lie group L.
Proof. By Exercise 1.8.24, we have a good model 0 : hAi L0 of A4 by
some locally compact group L0 . In particular, there is an open neighbourhood U0 of the identity in L0 such that 01 (U0 ) A4 and U0 0 (A4 ).
Let U1 be a symmetric precompact neighbourhood of the identity such
that U1100 U0 . By the Gleason-Yamabe theorem (Theorem 1.1.17), there
is an open subgroup L00 of L0 , and a compact normal subgroup N of L00
contained in U1 , such that L00 /N is isomorphic to a connected Lie group L.
Let 1 : L00 L be the quotient map.
Write U2 := U1 L00 . As 0 is a good model, we can find a nonstandard
finite set A0 with
1 (U2 ) A0 1 (U22 ).
By replacing A0 with A0 (A0 )1 if necessary, we may take A0 to be symmetric. As U24 can be covered by finitely many left-translates of U2 , we see
that A0 is an ultra approximate group. Since
(A0 )4 1 (U28 ) 1 (U1100 ) 1 (U0 ) A4
and (A4 ) can be covered by finitely many left-translates of U2 , we see that
A0 is a large ultra approximate subgroup of A. It is then a routine matter
to verify that 1 0 : hA0 i L is a good model for A0 .

The Lie model L need not be unique. For instance, the nonstandard
cyclic group Z/N Z can be modeled both by the trivial group and by the
unit circle R/Z. However, as observed by Hrushovski [Hr2012], it can

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

178

1. Hilberts fifth problem

be shown that after quotienting out the (unique) maximal compact normal
subgroup from the Lie model L, the resulting quotient group (which is also
a Lie group, and in some sense describes the large scale structure of L) is
unique up to isomorphism. The following exercise fleshes out the details of
this observation.
Exercise 1.8.32 (Large-scale uniqueness of the Lie model). Let L, L0 be
connected Lie groups, and let A be an ultra approximate group with good
models : hAi L and 0 : hA0 i L0 .
(i) Show that the centre Z(L) := {g L : gh = hg for all h L} is an
abelian Lie group.
(ii) Show that the connected component Z(L)0 of the identity in Z(L)
0
is isomorphic to Rd (R/Z)d for some d, d0 0.
(iii) Show that the quotient group Z(L)/Z(L)0 is a finitely generated
00
abelian group, and is isomorphic to Zd H for some finite group
H.
(iv) Show that the torsion points of Z(L) are contained in a compact
0
subgroup of Z(L) isomorphic to (R/Z)d H.
(v) Show that any finite normal subgroup of L is central, and thus
lies in the compact subgroup indicated above. (Hint: L will act
continuously by conjugation on this finite normal subgroup.)
(vi) Show that given any increasing sequence N1 N2 . . . of comS
pact normal subgroups of L, the upper bound n Nn is also a compact normal subgroup of L. (Hint: The dimensions of Nn (which
are well-defined by Cartans theorem) are monotone increasing but
bounded by the dimension of L. In particular, the connected components Nn0 must eventually stabilise. Quotient them out and then
use (v).)
(vii) Show that L contains a unique maximal compact normal subgroup
N . Similarly, L0 contains a unique maximal compact normal subgroup N 0 . Show that the quotient groups L/N, L0 /N 0 contain no
non-trivial compact normal subgroups.
(viii) Show that 1 (N ) = ( 0 )1 (N 0 ). (Hint: if g L, then g N iff
the group generated by g and its conjugates is bounded.)
(ix) Show that L/N is isomorphic to L0 /N 0 .
(x) Show that for sufficiently large standard m, Am can be modeled by
a Lie group with no non-trivial compact normal subgroups, which
is unique up to isomorphism.
To illustrate how this theorem is useful, let us apply it in the bounded
torsion case.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.8. Models of ultra approximate groups

179

Exercise 1.8.33. Let A be an ultra approximate group in an r-torsion

nonstandard group G for some standard r 1. Show that A4 contains a
nonstandard finite group H such that A can be covered by finitely many
left translates of H. (Hint: if a Lie group has positive dimension, then it
contains elements arbitrarily close to the identity of arbitrarily large order.)
Combining this exercise with Proposition 1.7.28, we conclude a finitary
consequence, first observed by Hrushovski [Hr2012]:
Corollary 1.8.32 (Freiman theorem, bounded torsion case). Let r, K 1,
and let A be a finite K-approximate subgroup of an r-torsion group G. Then
A4 contains a finite subgroup H such that A can be covered by OK,r (1) left
translates of H.
Exercise 1.8.34 (Commutator self-containment).
(1) Show that if A is an ultra approximate group, then there exists a
large approximate subgroup B of A such that [B, B] B, where
we write [A, B] := {[a, b] : a A, b B}, with [a, b] := a1 b1 ab.
(Note that this is slightly different from the group-theoretic convention, when H and K are subgroups, to define [H, K] to be the
group generated by the commutators [h, k] with h H and k K.)
(2) Show that if A is a finite K-approximate group, then there exists a
symmetric set B containing the origin with B 4 A4 , |B| K |A|,
and [B, B] B.
Remark 1.8.33. I do not know of any proof of Exercise 1.8.34 that does
not go through the Gleason-Yamabe theorem (or some other comparably
deep fragment of the theory of Hilberts fifth problem).
The following result of Hrushovski is a significant strengthening of the
preceding exercise:
Exercise 1.8.35 (Hrushovskis structure theorem). Let A be a finite Kapproximate group, and let F : N N N be a function. Show that there
exist natural numbers L, M, N with N F (L, M ) and L, M K,F 1, and
nested sets
{1} AN A1 A4
with the following properties:
(i) For each 1 n N , An is symmetric;
(ii) For each 1 n < N , A2n+1 An ;
(iii) For each 1 n N , An is contained in M left-translates of An+1 ;
(iv) For 1 n, m, k N with k < n + m, one has [An , Am ] Ak ;
(v) A can be covered by L left-translates of A1 .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

180

1. Hilberts fifth problem

(Hint: first find and prove an analogous statement for ultra approximate
groups, in which the function F is not present.)
There is a finitary formulation of Theorem 1.8.31, but it takes some
effort to state. Let L be a connected Lie group, with Lie algebra l which we
identify using some coordinate basis with Rd , thus giving a Euclidean norm
kk on l. We say that L with this basis has complexity at most M for some
M 1 if
(1) The dimension d of L is at most M ;
(2) The exponential map exp : l L is injective on the ball {x l :
kxk 1/M };
(3) One has k[x, y]k M kxkkyk for all x, y l.
We then define balls BR on L by the formula
BR := {exp(x) : x l; kxk < R}.
Exercise 1.8.36 (Hrushovskis Lie model theorem, finitary version). Let
F : N N be a function, and let A be a finite K-approximate group. Show
that there exists a natural number 1 M K,F 1, a connected Lie group L
of complexity at most M , a symmetric set A0 containing the identity with
(A0 )4 A4 , and a map : (A0 )F (M ) L obeying the following properties:
(i) (Large subgroup) A can be covered by M left-translates of A0 .
(ii) (Approximate homomorphism) One has (1) = 1, and for all a, b
(A0 )F (M ) with ab (A0 )F (M ) , one has
(ab)(b)1 (a)1 B1/F (M ) .
(iii) (Thick image) If a (A0 )F (M ) and (a) B1/M , then a A0 .
Conversely, if g B1/M , then (A0 ) intersects gB1/F (M ) .
M.
(iv) (Compact image) One has (A0 ) BM

(Hint: One needs to argue by compactness and contradiction, carefully

negating all the quantifiers in the above claim, and then use Theorem 1.8.31.)
Exercise 1.8.37. In the converse direction, show that Theorem 1.8.31 can
be deduced from Exercise 1.8.36. (Hint: to get started, one needs a statement to the effect that if an ultra approximate group A is unable to be
modeled by some Lie group of complexity at most M , then there is also
some > 0 for which A cannot be approximately modeled up to error in
some sense by such a Lie group. Once one has such a statement (provable
via a compactness or ultralimit argument), one can use this to build a
suitable function F () which which to apply Exercise 1.8.36.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

181

1.9. The microscopic structure of approximate groups

A common theme in mathematical analysis (particularly in analysis of a
geometric or statistical flavour) is the interplay between macroscopic
and microscopic scales. These terms are somewhat vague and imprecise,
and their interpretation depends on the context and also on ones choice
of normalisations, but if one uses a macroscopic normalisation, macroscopic scales correspond to scales that are comparable to unit size (i.e.
bounded above and below by absolute constants), while microscopic scales
are much smaller, being the minimal scale at which nontrivial behaviour occurs. (Other normalisations are possible, e.g. making the microscopic scale
a unit scale, and letting the macroscopic scale go off to infinity; for instance,
such a normalisation is often used, at least initially, in the study of groups
of polynomial growth. However, for the theory of approximate groups, a
macroscopic scale normalisation is more convenient.)
One can also consider mesoscopic scales which are intermediate between microscopic and macroscopic scales, or large-scale behaviour at scales
that go off to infinity (and in particular are larger than the macroscopic
range of scales), although the behaviour of these scales will not be the main
focus of this section. Finally, one can divide the macroscopic scales into
local macroscopic scales (less than for some small but fixed > 0) and
global macroscopic scales (scales that are allowed to be larger than a given
large absolute constant C). For instance, given a finite approximate group
A:
(1) Sets such as Am for some fixed m (e.g. A10 ) can be considered to
be sets at a global macroscopic scale. Sending m to infinity, one
enters the large-scale regime.
(2) Sets such as the sets S that appear in the Sanders lemma from the
previous section (thus S m A4 for some fixed m, e.g. m = 100)
can be considered to be sets at a local macroscopic scale. Sending
m to infinity, one enters the mesoscopic regime.
(3) The non-identity element u of A that is closest to the identity
in some suitable metric (cf. the proof of Jordans theorem from
Section 1.1) would be an element associated to the microscopic
scale. The orbit u, u2 , u3 , . . . starts out at microscopic scales, and
(assuming some suitable escape axioms) will pass through mesoscopic scales and finally entering the macroscopic regime. (Beyond
this point, the orbit may exhibit a variety of behaviours, such as
periodically returning back to the smaller scales, diverging off to
ever larger scales, or filling out a dense subset of some macroscopic

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

182

1. Hilberts fifth problem

set; the escape axioms we will use do not exclude any of these
possibilities.)
For comparison, in the theory of locally compact groups, properties
about small neighbourhoods of the identity (e.g. local compactness, or the
NSS property) would be properties at the local macroscopic scale, whereas
the space L(G) of one-parameter subgroups can be interpreted as an object at the microscopic scale. The exponential map then provides a bridge
connecting the microscopic and macroscopic scales.
We return now to approximate groups. The macroscopic structure of
these objects is well described by the Hrushovski Lie model theorem (Theorem 1.8.31), which informally asserts that the macroscopic structure of an
(ultra) approximate group can be modeled by a Lie group. This is already
an important piece of information about general approximate groups, but it
does not directly reveal the full structure of such approximate groups, because these Lie models are unable to see the microscopic behaviour of these
approximate groups.
To illustrate this, let us review one of the examples of a Lie model of
an ultra approximate group, namely Exercise 1.8.28. In this example one
studied a nilbox from a Heisenberg group, which we rewrite here in slightly
different notation. Specifically, let G be the Heisenberg group
G := {(a, b, c) : a, b, c Z}
with group law
(a, b, c) (a0 , b0 , c0 ) := (a + a0 , b + b0 , c + c0 + ab0 )
Q
and let A = n An , where An G is the box

(1.72)

An := {(a, b, c) G : |a|, |b| n; |c| n10 };

thus A is the nonstandard box
A := {(a, b, c) G : |a|, |b| N ; |c| N 10 }
where N := limn n. As the above exercise establishes, A A1 is an ultra
approximate group with a Lie model : hAi R3 given by the formula


a
b
c
(a, b, c) := st , st , st 10
N
N
N
for a, b = O(N ) and c = O(N 10 ). Note how the nonabelian nature of G
(arising from the ab0 term in the group law (1.72)) has been lost in the
N2
model R3 , because the effect of that nonabelian term on Nc10 is only O( N
8)
which is infinitesimal and thus does not contribute to the standard part. In

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

183

particular, if we replace G with the abelian group G0 := {(a, b, c) : a, b, c Z}

with the additive group law
(a, b, c) 0 (a0 , b0 , c0 ) := (a + a0 , b + b0 , c + c0 )
and let A0 and 0 be defined exactly as with A and , but placed inside the
group structure of G0 rather than G, then AA1 and A0 (A0 )1 are essentially indistinguishable as far as their models by R3 are concerned, even
though the latter approximate group is abelian and the former is not. The
problem is that the nonabelian-ness in the former example is so microscopic
that it falls entirely inside the kernel of and is thus not detected at all by
the model.
The problem of not being able to see the microscopic structure of a
group (or approximate group) also was a key difficulty in the theory surrounding Hilberts fifth problem that was discussed in Sections 1.3, 1.5. A
key tool in being able to resolve such structure was to build left-invariant
metrics d (or equivalently, norms kk) on ones group, which obeyed useful
Gleason axioms such as the commutator axiom
(1.73)

k[g, h]k  kgkkhk

for sufficiently small g, h, or the escape axiom

(1.74)

kg n k  |n|kgk

when |n|kgk was sufficiently small. Such axioms have important and nontrivial content even in the microscopic regime where g or h are extremely
close to the identity. For instance, in the proof of Jordans theorem (Theorem 1.1.2), a key step was to apply the commutator axiom (1.73) (for the
distance to the identity in operator norm) to the most microscopic element of G, or more precisely a non-identity element of G of minimal norm.
The key point was that this microscopic element was virtually central in
G, and as such it restricted much of G to a lower-dimensional subgroup of
the unitary group, at which point one could argue using an induction-ondimension argument. As we shall see, a similar argument can be used to
place virtually nilpotent structure on finite approximate groups. For instance, in the Heisenberg-type approximate groups A A1 and A0 (A0 )1
discussed earlier, the element (0, 0, 1) will be closest to the origin in a
suitable sense to be defined later, and is centralised by both approximate
groups; quotienting out (the orbit of) that central element and iterating the
process two more times, we shall see that one can express both A A1
and A0 (A0 )1 as a tower of central cyclic extensions, which in particular
establishes the nilpotency of both groups.
The escape axiom (1.74) is a particularly important axiom in connecting the microscopic structure of a group G to its macroscopic structure;

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

184

1. Hilberts fifth problem

for instance, as shown in Section 1.3, this axiom (in conjunction with the
closely related commutator axiom) tends to imply dilation estimates such as
d(g n , hn ) nd(g, h) that allow one to understand the microscopic geometry
of points g, h close to the identity in terms of the (local) macroscopic geometry of points g n , hn that are significantly further away from the identity.
It is thus of interest to build some notion of a norm (or left-invariant
metrics) on an approximate group A that obeys the escape and commutator axioms (while being non-degenerate enough to adequately capture the
geometry of A in some sense), in a fashion analogous to the Gleason metrics that played such a key role in the theory of Hilberts fifth problem.
It is tempting to use the Lie model theorem to do this, since Lie groups
certainly come with Gleason metrics. However, if one does this, one ends
up, roughly speaking, with a norm on A that only obeys the escape and
commutator estimates macroscopically; roughly speaking, this means that
one has a macroscopic commutator inequality
k[g, h]k  kgkkhk + o(1)
and a macroscopic escape property
kg n k  |n|kgk o(|n|)
but such axioms are too weak for analysis at the microscopic scale, and in
particular in establishing centrality of the element closest to the identity.
Another way to proceed is to build a norm that is specifically designed
to obey the crucial escape property. Given an approximate group A in a
group G, and an element g of G, we can define the escape norm kgke,A of g
by the formula


1
: n N; g, g 2 , . . . , g n A .
kgke,A := inf
n+1
Thus, kgke,A equals 1 if g lies outside of A, equals 1/2 if g lies in A but g 2
lies outside of A, and so forth. Such norms had already appeared in Section
1.5, in the context of analysing NSS groups.
As it turns out, this expression will obey an escape axiom, as long as
we place some additional hypotheses on A which we will present shortly.
However, it need not actually be a norm; in particular, the triangle inequality
kghke,A kgke,A + khke,A
is not necessarily true. Fortunately, it turns out that by a (slightly more
complicated) version of the Gleason machinery from Section 1.5 we can
establish a usable substitute for this inequality, namely the quasi-triangle
inequality
kg1 . . . gk ke,A C(kg1 ke,A + + kgk ke,A ),

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

185

where C is a constant independent of k. As we shall see, these estimates

can then be used to obtain a commutator estimate (1.73).
However, to do all this, it is not enough for A to be an approximate
group; it must obey two additional trapping axioms that improve the
properties of the escape norm. We formalise these axioms (somewhat arbitrarily) as follows:
Definition 1.9.1 (Strong approximate group). Let K 1. A strong Kapproximate group is a finite K-approximate group A in a group G with a
symmetric subset S obeying the following axioms:
(1) (S small) One has
(1.75)

(S A )1000K A.

(2) (First trapping condition) If g, g 2 , . . . , g 1000 A100 , then g A.

6K3

(3) (Second trapping condition) If g, g 2 , . . . , g 10

A, then g S.
Q
An ultra strong K-approximate group is an ultraproduct A = n An of
strong K-approximate groups.
The first trapping condition can be rewritten as
kgke,A 1000kgke,A100
and the second trapping condition can similarly be rewritten as
kgke,S 106 K 3 kgke,A .
This makes the escape norms of A, A100 , and S comparable to each other,
which will be needed for a number of reasons (and in particular to close
a certain bootstrap argument properly). Compare this with (1.53), which
used the NSS hypothesis to obtain similar conclusions. Thus, one can view
the strong approximate group axioms as being a sort of proxy for the NSS
property.
Example 1.9.2. Let N be a large natural number. Then the interval A =
[N, N ] in the integers is a 2-approximate group, which is also a strong
2-approximate group (setting S = [106 N, 106 N ], for instance). On the
other hand, if one places A in Z/5N Z rather than in the integers, then the
first trapping condition is lost and one is no longer a strong 2-approximate
group. Also, if one remains in the integers, but deletes a few elements from
A, e.g. deleting b1010 N c from A), then one is still a O(1)-approximate
group, but is no longer a strong O(1)-approximate group, again because the
first trapping condition is lost.
A key consequence of the Hrushovski Lie model theorem is that it allows
one to replace approximate groups by strong approximate groups:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

186

1. Hilberts fifth problem

Exercise 1.9.1 (Finding strong approximate groups).

(i) Let A be an ultra approximate group with a good Lie model : hAi
L, and let B be a symmetric convex body (i.e. a convex open
bounded subset) in the Lie algebra l. Show that if r > 0 is a sufficiently small standard number, then there exists a strong ultra
approximate group A0 with
1 (exp(rB)) A0 1 (exp(1.1rB)) A,
and with A can be covered by finitely many left translates of A0 .
Furthermore, is also a good model for A0 .
(ii) If A is a finite K-approximate group, show that there is a strong
OK (1)-approximate group A0 inside A4 with the property that A
can be covered by OK (1) left translates of A0 . (Hint: use (i),
Hrushovskis Lie model theorem, and a compactness and contradiction argument.)
The need to compare the strong approximate group to an exponentiated
small ball exp(rB) will be convenient later, as it allows one to easily use the
geometry of L to track various aspects of the strong approximate group.
As mentioned previously, strong approximate groups exhibit some of
the features of NSS locally compact groups. In Section 1.5, we saw that the
escape norm for NSS locally compact groups was comparable to a Gleason
metric. The following theorem is an analogue of that result:
Theorem 1.9.3 (Gleason lemma). Let A be a strong K-approximate group
in a group G.
(i) (Symmetry) For any g G, one has kg 1 ke,A = kgke,A .
(ii) (Conjugacy bound) For any g, h A10 , one has kg h ke,A  kgke,A .
(iii) (Triangle inequality) For any g1 , . . . , gk G, one has kg1 . . . gk ke,A K
(kg1 ke,A + + kgk ke,A ).
(iv) (Escape property) One has kg n ke,A  |n|kgke,A whenever |n|kgke,A <
1.
(v) (Commutator inequality) For any g, h A10 , one has k[g, h]ke,A K
kgke,A khke,A .
The proof of this theorem will occupy a large part of this section.
We then aim to use this theorem to classify strong approximate groups.
The basic strategy (temporarily ignoring a key technical issue) follows the
Bieberbach-Frobenius proof of Jordans theorem, as given in Section 1.1, is
as follows.
(1) Start with an (ultra) strong approximate group A.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

187

(2) From the Gleason lemma, the elements with zero escape norm form
a normal subgroup of A. Quotient these elements out. Show that
all non-identity elements will have positive escape norm.
(3) Find the non-identity element g1 in (the quotient of) A of minimal escape norm. Use the commutator estimate (assuming it is
inherited by the quotient) to show that g1 will centralise (most of)
this quotient. In particular, the orbit hg1 i is (essentially) a central
subgroup of hAi.
(4) Quotient this orbit out; then find the next non-identity element g2
in this new quotient of A. Again, show that hg2 i is essentially a
central subgroup of this quotient.
(5) Repeat this process until A becomes entirely trivial. Undoing all
the quotients, this should demonstrate that hAi is virtually nilpotent, and that A is essentially a coset nilprogression.
There are two main technical issues to resolve to make this strategy work.
The first is to show that the iterative step in the argument terminates in
finite time. This we do by returning to the Lie model theorem. It turns out
that each time one quotients out by an orbit of an element that escapes,
the dimension of the Lie model drops by at least one. This will ensure
termination of the argument in finite time.
The other technical issue is that while the quotienting out all the elements of zero escape norm eliminates all torsion from A (in the sense that
the quotient of A has no non-trivial elements of zero escape norm), further
quotienting operations can inadvertently re-introduce such torsion. This
torsion can be re-eradicated by further quotienting, but the price one pays
for this is that the final structural description of hAi is no longer as strong
as virtually nilpotent, but is instead a more complicated tower alternating
between (ultra) finite extensions and central extensions.
Example 1.9.4. Consider the strong O(1)-approximate group
A := {aN 10 + 5b : |a| N ; |b| N 2 }
in the integers, where N is a large natural number not divisible by 5. As
Z is torsion-free, all non-zero elements of A have positive escape norm, and
the nonzero element of minimal escape norm here is g = 5 (or g = 5). But
if one quotients by hgi, A projects down to Z/5Z, which now has torsion
(and all elements in this quotient have zero escape norm). Thus torsion has
been re-introduced by the quotienting operation. (A related observation is
that the intersection of A with hgi = 5Z is not a simple progression, but is
a more complicated object, namely a generalised arithmetic progression of
rank two.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

188

1. Hilberts fifth problem

To deal with this issue, we will not quotient out by the entire cyclic
group hgi = {g n : n Z} generated by the element g of minimal escape
norm, but rather by an arithmetic progression P = {g n : |n| N }, where
N is a natural number comparable to the reciprocal 1/kgke,A of the escape
norm, as this will be enough to cut the dimension of the Lie model down
by one without introducing any further torsion. Of course, this cannot be
done in the category of global groups, since the arithmetic progression P
will not, in general, be a group. However, it is still a local group, and it
turns out that there is an analogue of the quotient space construction in
local groups. This fixes the problem, but at a cost: in order to make the
inductive portion of the argument work smoothly, it is now more natural to
place the entire argument inside the category of local groups rather than
global groups, even though the primary interest in approximate groups A is
in the global case when A lies inside a global group. This necessitates some
technical modification to some of the preceding discussion (for instance,
the Gleason-Yamabe theorem must be replaced by the local version of this
theorem, Theorem 1.5.24); details can be found in [BrGrTa2011], but will
only be sketched here.
1.9.1. Gleasons lemma. Throughout this section, A is a strong K-approximate
group in a global group G. We will prove the various estimates in Theorem
1.9.3. The arguments will be very close to those in Section 1.5; indeed, it
is possible to unify the results here with the results in that section by a
suitable modification of the notation, but we will not do so here.
We begin with the easy estimates. The symmetry property is immediate
from the symmetry of A. Now we turn to the escape property. By symmetry, we may take n to be positive (the n = 0 case is trivial). We may of
course assume that kgke,A is strictly positive, say equal to 1/(m + 1); thus
g, . . . , g m A and g m+1 6 A, and n m. By the first trapping property,
this implies that g j(m+1) 6 A100 for some 1 j 1000.
Let kn be the first multiple of n larger than or equal to j(m + 1), then
kn  m + 1. Since kn j(m + 1) is less than m, we have g knj(m+1) A;
since g j(m+1) 6 A100 , we conclude that g kn 6 A99 . In particular this shows
that kg n ke,A  1/k  n/(m + 1), and the claim follows.
The escape property implies the conjugacy bound:
Exercise 1.9.2. Establish the conjugacy bound. (Hint: one can mimic the
arguments establishing a nearly identical bound in Section 1.5.2.)
Now we turn to the triangle inequality, which (as in Section 1.5) is the
most difficult property to establish. Our arguments will closely resemble the
4
proof of Proposition 1.5.9, with S A and A playing the roles of U1 and U0
from that argument. As in that theorem, we will initially assume that we

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

189

have an a priori bound of the form

kg1 . . . gk ke,A M (kg1 ke,A + + kgk ke,A )

(1.76)

for all g1 , . . . , gk , and some (large) M independent of k, and remove this

hypothesis later. We then introduce the norm
kgk,A := inf{kg1 ke,A + + kgk ke,A : g = g1 . . . gk },
then kgk,A is symmetric, obeys the triangle inequality, and is comparable
to kke,A in the sense that
(1.77)

1
kgke,A kgk,A kgke,A
M

for all g G.
We introduce the function : G R+ by
(x) := (1 M dist,A (x, A))+ ,
where dist,A (x, A) := inf yA kx1 yk,A . Then takes values between 0 and
1, equals 1 on A, is supported on A2 , and obeys the Lipschitz bound
kg k` (G) M kgke,A

(1.78)

for all g, thanks to the triangle inequality for kk,A and (1.77). We also
introduce the function : G R+ by
j
2
: x (S A )j A} {0},
104 K 3
then also takes values between 0 and 1, equals 1 on A, is supported on
A2 , and obeys the Lipschitz bound
(x) := sup{1

(1.79)

khy k` (G)

1
104 K 3

for all h S and y A4 .

Now we form the convolution : G R+ by the formula
1 X
(y)(y 1 x)
(x) :=
|A|
yG

1 X
(xy)(y 1 ).
|A|
yG

By construction, is supported on A4 and at least 1 at the identity. As

or is supported in A2 , which has cardinality at most K|A|, we have the
uniform bound
(1.80)

kk` (G) K 2 .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

190

1. Hilberts fifth problem

Similarly, from the identity

g (x) =

1 X
g (y)(y 1 x)
|A|
yG

and (1.78) we have the Lipschitz bound

(1.81)

kg k` (G) K 2 M kgke,A .

Finally, from the identity

h g (x) =

1 X
g (y)hy (y 1 x)
|A|
yG

and restricting g to A (so that g is supported on A4 , which has cardinality

at most K 3 |A|) we see from (1.78), (1.79) that
1
kh g k` (G) 4 M kgke,A
10
for g A and h S.
We can use this to improve the bound (1.81). Indeed, using the telescoping identity
n1
X
gn = ng +
gi g
i=0

we see that
n1

kg k` (G)

1X
1
kgn k` (G) +
kgi g k` (G)
n
n
i=0

and thus

1
1
+ 4 M kgke,A
n 10
S. Using the second trapping property, this

kg k` (G)
whenever g, g 2 , . . . , g n1
implies that

1
M + OK (1))kgke,A
104
In the converse direction, if kg k` (G) < n1 , then
kg k` (G) (

kgi k` (G) < 1

and thus g i A4 from the support of , for all 1 i n. But then by the
first trapping property, this implies that g i A for all 1 i n/1000. We
conclude that
kgke,A 1000kg k` (G) .
The triangle inequality for kg k` (G) then implies a triangle inequality for
kgke,A ,


1
kg1 . . . gk ke,A
M + OK (1) (kg1 ke,A + + kgk ke,A ) ,
10

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

191

1
which is (1.76) with M replaced by 10
M +OK (1). If we knew (1.76) for some
large but finite M , we could iterate this argument and conclude that (1.76)
held with M replaced by OK (1), which would give the triangle inequality.
Now it is not immediate that (1.76) holds for any finite M , but we can avoid
this problem with the usual regularisation trick of replacing kgke,A with
kgke,A + throughout the argument for some small > 0, which makes (1.76)
automatically true with M = O(1/), run the above iteration argument, and
then finally send to zero.

Exercise 1.9.3. Verify that the modifications to the above argument sketched
above actually do establish the triangle inequality.
A final application of the Gleason convolution machinery then gives the
final estimate in Gleasons lemma:
Exercise 1.9.4. Use the properties of the escape norm already established
(and in particular, the escape property and the triangle inequality) to establish the commutator inequality. (Hint: adapt the argument from Section
1.5.2.)
The proof of Theorem 1.9.3 is now complete.
Exercise 1.9.5. Generalise Theorem 1.9.3 to the setting where A is not
necessarily finite, but is instead an open precompact subset of a locally
compact group G. (Hint: replace cardinality by left-invariant Haar measure
and follow the arguments in Section 1.5 closely.) Note that this already
gives most of one of the key results from that section, namely that any NSS
group admits a Gleason metric, since it is not difficult to show that NSS
groups contain open precompact strong approximate groups.
1.9.2. A cheap version of the structure theorem. In this section we
use Theorem 1.9.3 to establish a cheap version of the structure theorem
for ultra approximate groups. We begin by eliminating the elements of zero
escape norm. Let us say that an approximate group A is NSS if it contains
no non-trivial subgroups of the ambient group, or equivalently if every nonidentity element of A has a positive escape norm. We say that an ultra
approximate group is NSS if it is the ultralimit of NSS approximate groups.
Using the Gleason lemma, we can easily reduce to the NSS case:
Exercise 1.9.6 (Reduction to the NSS case). Let L be a connected Lie
group with Lie algebra l, let B be a bounded symmetric convex body in l,
let r > 0 be a sufficiently small standard real. Let 0 < r0 < r/2, and let A
be an ultra strong approximate group which has a good model : hAi L
with
1 (exp(rB)) A 1 (exp(1.1rB)).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

192

1. Hilberts fifth problem

Let H be the set of all elements in A of zero (nonstandard) escape norm.

Show that H is a normal nonstandard finite subgroup of hAi that lies in
ker(). If : hAi hAi/H is the quotient map, and 0 : hAi/H L is
the map factored through , show that there exists an ultra strong NSS
approximate group A0 in (A) which has 0 as a good model with
( 0 )1 (exp(r0 B)) A0 ( 0 )1 (exp(1.1r0 B)),
and such that A is covered by finitely many left-translates of 1 (A0 ).
Let us now analyse the NSS case. Let L be a connected Lie group, with
Lie algebra l, let B be a bounded symmetric convex body in l, let r > 0 be a
sufficiently small standard real. Let A be an ultra strong NSS approximate
group which has a good model : hAi L with
1 (exp(rB)) A 1 (exp(1.1rB)).
If L is zero-dimensional, then by connectedness it is trivial, and then
(by the properties of a good model) A is a nonstandard finite group; since
it is NSS, it is also trivial. Now suppose that L is not zero-dimensional.
Then A contains non-identity elements whose image under is arbitrarily
close to the identity of L; in particular, A does not consist solely of the
identity element, and thus contains elements of positive escape norm by the
NSS assumption. Let g be a non-identity element of G with minimal escape
norm kuke,A , then (u) must be the identity (so in particular kuke,A is
infinitesimal). (Note that any non-trivial NSS finite approximate group will
contain non-identity elements of minimal escape norm, and the extension of
this claim to the ultra approximate group case follows from Loss theorem.)
From Theorem 1.9.3 one has
k[u, h]ke,A  kuke,A khke,A
for all h A. (Here we are now using the nonstandard asymptotic notation,
thus X  Y means that X CY for some standard C.) In particular, from
the minimality of kuke,A , we see that there is a standard c > 0 such that u
commutes with all elements h with khke,A < c. In particular, if r0 > 0 is a
sufficiently small standard real number, we can find an ultra approximate
subgroup A0 of A with
1 (exp(r0 B)) A0 1 (exp(1.1r0 B))
which is centralised by u.
Now we show that u generates a one-parameter subgroup of the model
Lie group L.
Exercise 1.9.7 (One-parameter subgroups from orbits). Let the notation
be as above. Let g A be such that kgke,A is infinitesimal but non-zero.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

193

(i) Show that (g i ) = 1 whenever i = o(1/kgke,A ).

(ii) Show that the map t 7 (g bt/kgke,A ) is a one-parameter subgroup
in L (i.e. a continuous homomorphism from R to L).
(iii) Show that there exists an element X of 1.1rB\rB such that (g i ) =
exp(st(ikgke,A )X) for all i = O(1/kgke,A ).
Similar statements hold with A, r replaced by A0 , r0 .
We can now quotient out by the centraliser of u and reduce the dimension
of the Lie model:
Exercise 1.9.8. Let Z(u) := {h G : uh = hu} be the centraliser of u in
G, and let H := {un : n Z} be the nonstandard cyclic group generated
by u. (Thus, by the preceding discussion, Z(u) contains A0 , and H is a
central subgroup of Z(u) containing u. It will be important for us that Z(u)
and H are both nonstandard sets, i.e. ultraproducts of standard sets.)
(i) Show that (H Am ) is a compact subset of L for each standard
m.
(ii) Show that (H hAi) is a central subgroup of L that contains (R).
(iii) Show that (H hAi) is a central subgroup of L that is a Lie
group of dimension at least one, and so the quotient group L0 :=
L/(H hAi) is a Lie group of dimension strictly less than the
dimension of L.
(iv) Let : Z(u) Z(u)/H be the quotient map, and let 0 : (hA0 i)
L0 be the obvious quotient of . Let B 0 be a convex symmetric
body in the Lie algebra l0 of L. Show that for sufficiently small
standard r00 > 0, there exists an ultra strong approximate group
( 0 )1 (exp(r00 B)) A00 ( 0 )1 (exp(1.1r00 B))
with 0 as a good model, with A00 (A0 ), and with 0 (A0 ) covered
by finitely many left-translates of A00 .
Note that the quotient approximate group A00 obtained by the above
procedure is not necessarily NSS. However, it can be made NSS by Exercise
1.9.6. As such, one can iterate the above exercise until the dimension of the
Lie model shrinks all the way to zero, at which point the NSS approximate
group one is working with becomes trivial. This leads to a cheap structure
theorem for approximate groups:
Exercise 1.9.9 (Cheap structure theorem). Let A be an ultra approximate
group in a nonstandard group G.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

194

1. Hilberts fifth problem

(i) Show that if A has a good model by a connected Lie group L, then
L is nilpotent. (Hint: first use Exercise 1.9.1, and then induct on
the dimension of L.)
(ii) Show that A is covered by finitely many left translates of a nonstandard subgroup G0 of G which admits a normal series
G0 = G00 B G01 B G02 B B G0k = {1}
for some standard k, where for every 0 i < k, G0i+1 is a normal nonstandard subgroup of G0 and of G0i , and G0i /G0i+1 is either
a nonstandard finite group or a nonstandard central subgroup of
G0 /G0i+1 . Furthermore, if G0i /G0i+1 is not central, then it is contained in the image of A4 G0 in G0 /G0i+1 ; and G0 is generated (as
a nonstandard subgroup) by A4 G0 . (Hint: first use the Lie model
theorem and Exercise 1.9.1, and then induct on the dimension of
L.)
Exercise 1.9.10 (Cheap structure theorem, finite version). Let A be a
finite K-approximate group in a group G. Show that A is covered by OK (1)
left-translates of a subgroup G0 of G which admits a normal series
G0 = G00 B G01 B G02 B B G0k = {1}
for some k = OK (1), where for every 0 i < k, G0i+1 is a normal subgroup
of G0i and of G0 , and G0i /G0i+1 is either finite or central in G0 /G0i+1 . Furthermore, if G0i /G0i+1 is not central, then it is contained in the image of A4 G0
in G0 /G0i+1 , and G0 is generated as a group by A4 G.
Exercise 1.9.11.
(i) Show that any finite extension G of a virtually nilpotent group H
(thus there is a short exact sequence 0 K G H 0 with
K finite) is virtually nilpotent. (Hint: G acts on K by conjugation;
look at the stabiliser of this action.)
(ii) Conclude that the group G0 in the previous exercise is virtually
nilpotent.
One can push the cheap structure theorem a bit further by controlling
the dimension of the nilpotent Lie group in terms of the covering number K
of the ultra approximate group, as laid out in the following exercise.
Exercise 1.9.12 (Nilpotent groups). A Lie algebra g is said to be nilpotent
if the derived series g1 := g, g2 := [g1 , g], g3 := [g2 , g], . . . becomes trivial
after a finite number of steps.
(i) Show that a connected Lie group is nilpotent if and only if its Lie
algebra is nilpotent.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

195

(ii) If g is a finite-dimensional nilpotent Lie algebra, show that there

is a simply connected Lie group G with Lie algebra g, for which
the exponential map exp : g G is a (global) homeomorphism.
Furthermore, any other connected Lie group with Lie algebra g is
a quotient of G by a discrete central subgroup of G.
(iii) If g and G are as in (ii), show that the pushforward of a Haar
measure (or Lebesgue measure) on g is a bi-invariant Haar measure
on G. (Recall from Exercise 1.4.8 that connected nilpotent Lie
groups are unimodular.)
(iv) If g and G are as in (ii), and is a bi-invariant Haar measure on G,
show that (A2 ) 2d (A) for all open precompact A G, where
d is the dimension of G.
is a connected (but not necessarily simply connected) nilpotent
(v) If G
Lie group, and N is the maximal compact normal subgroup of G

(which exists by Exercise 1.8.32), show that N is central, and G/N

is simply connected. As a consequence, conclude that if
is a left2
d

Haar measure of G, then

(A ) 2
(A) for all open precompact
where d is the dimension of G/N .
A G,
(vi) Show that if A is an ultra K-approximate group which has a Lie
model L, and N is the maximal compact normal subgroup of L,
then L/N has dimension at most log2 K.
(vii) Show that if A is an ultra K-approximate group, then there is an
ultra K O(1) -approximate group A0 in A4 that is modeled by a Lie
group L, such that A is covered by finitely many left-translates of
A. (Hint: A4 has a good model : A4 G by a locally compact
group G; by the Gleason-Yamabe theorem, G has an open subgroup
G0 and a normal subgroup N of G0 inside (A4 ) with G0 /N a Lie
group. Set A0 := A4 1 (G0 ).)
(viii) Show that if A is an ultra K-approximate group, then there is an
ultra K O(1) -approximate group A00 in AO(1) that is modeled by a
nilpotent group of dimension O(log K), such that A can be covered
by finitely many left-translates of A.
1.9.3. Local groups. The main weakness of the cheap structure theorem
in the preceding section is the continual reintroduction of torsion whenever
one quotients out by the centraliser Z(u), which can destroy the NSS property. We now address the issue of how to fix this, by moving to the context of
local groups rather than global groups. We will omit some details, referring
to [BrGrTa2011] for details.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

196

1. Hilberts fifth problem

We need to extend many of the notions we have been considering to the

local group setting. We begin by generalising the concept of an approximate
group.
Definition 1.9.5 (Approximate groups). A (local) K-approximate group is
a subset A of a local group G which is symmetric and contains the identity,
such that A200 is well-defined in G, and for which A2 is covered by K left
3
translates of A (by
Q elements in A ). An ultra approximate group is an
ultraproduct A = n An of K-approximate groups.
Note that we make no topological requirements on A or G in this definition; in particular, we may as well give the local group G the discrete
topology. There are some minor technical advantages in requiring the local
group to be symmetric (so that the inversion map is globally defined) and
cancellative (so that gh = gk or hg = kg implies h = k), although these
assumptions are essentially automatic in practice.
The exponent 200 here is not terribly important in practice, thanks to
the following variant of the Sanders lemma:
Exercise 1.9.13 (local Sanders lemma). Let A be a finite K-approximate
group in a local group G, except with only A8 known to be well-defined
rather than A200 . Let m 1. Show that there exists a finite OK,m (1)approximate subgroup A0 in G with (A0 )m well-defined and contained in A4 ,
and with A covered by OK,m (1) left-translates of A0 (by elements in A5 ).
(Hint: adapt the proof of Lemma 1.8.1.)
Just as global approximate groups can be modeled by global locally
compact groups (and in particular, global Lie groups), local approximate
groups can be modeled by local locally compact groups:
Definition 1.9.6 (Good models). Let A be a (local) ultra approximate
group. A (local) good model for A is a homomorphism : A8 L from
A8 to a locally compact Hausdorff local group L that obeys the following
axioms:
(1) (Thick image) There exists a neighbourhood U0 of the identity in
L such that 1 (U0 ) A and U0 (A).
(2) (Compact image) (A) is precompact.
(3) (Approximation by nonstandard sets) Suppose that F U U0 ,
where F is compact and U is open. Then there exists a nonstandard
finite set B such that 1 (F ) B 1 (U ).
We make the pedantic remark that with our conventions, a global good
model : hAi L of a global approximate group only becomes a local good
model of A by L after restricting the domain of to A8 . It is also convenient

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

197

for minor technical reasons to assume that the local group L is symmetric
(i.e. the inversion map is globally defined) but this hypothesis is not of
major importance.
The Hrushovski Lie Model theorem can be localised:
Theorem 1.9.7 (Local Hrushovski Lie model theorem). Let A be a (local)
ultra approximate group. Then there is an ultra approximate subgroup A0
of A (thus (A0 )4 A4 ) with A covered by finitely many left-translates of A0
(by elements in A (A0 )1 ), which has a good model by a connected local Lie
group L.
The proof of this theorem is basically a localisation of the proof of the
global Lie model theorem from Section 1.8, and is omitted (see [BrGrTa2011]
for details). One key replacement is that if A is a local approximate group
rather than a global one, then the global Gleason-Yamabe theorem (Theorem 1.1.17) must be replaced by the local Gleason-Yamabe theorem (Theorem 1.5.24).
One can define the notion of a strong K-approximate group and ultra
strong approximate group in the local setting without much difficulty, since
strong approximate groups only need to work inside A100 , which is welldefined. Using the local Lie model theorem, one can obtain a local version
of Exercise 1.9.1. The Gleason lemma (Theorem 1.9.3) also localises without
much difficulty to local strong approximate groups, as does the reduction to
the NSS case in Exercise 1.9.6.
Now we once again analyse the NSS case. As before, let L be a connected
(local) Lie group, with Lie algebra l, let B be a bounded symmetric convex
body in l, let r > 0 be a sufficiently small standard real. Let A be a
(local) ultra strong NSS approximate group which has a (local) good model
: hAi L with
1 (exp(rB)) A 1 (exp(1.1rB)).
Again, we assume L has dimension at least 1, since A is trivial otherwise.
We let u be a non-identity element of minimal escape norm. As before,
u will have an infinitesimal escape norm and lie in the kernel of . If we
set N := kuke,A , then N is an unbounded natural number, and the map
: t 7 (g btN c ) will be a local one-parameter subgroup, i.e. a continuous
homomorphism from [1, 1] to L. This one-parameter subgroup will be
non-trivial and centralised by a neighbourhood of the identity in L.
In the global setting, we quotiented (the group generated by a large
portion of) A by the centraliser Z(u) of u. In the local setting, we perform
a more gentle quotienting, which roughly speaking arises by quotienting
A by the geometric progression P := {un : cN n cN }, where c > 0

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

198

1. Hilberts fifth problem

is a sufficiently small standard quantity to be chosen later. However, P

is only a local group rather than a global one, and so we will need the
machinery of notion of quotients of local groups from Section 1.5.6; the
reader is encouraged to review that section now.
We now begin the analysis of the NSS ultra strong approximate group
A. We give the ambient local group G the discrete topology.
Exercise 1.9.14. If r0 > 0 is a standard real that is sufficiently small
depending on c, show that there exists an ultra approximate group A0 with
1 (exp(r0 B)) A0 1 (exp(1.1r0 B)),
such that P is a sub-local group of A with normalising neighbourhood (A0 )6
P , that is also centralised by A0 .
By Exercise 1.5.17, we may now form the quotient set A00 := A0 /P . Show
that this is an ultra approximate group that is modeled by U/(c, c), where
U is an open neighbourhood of the identity in L and : [1, 1] L is the
local one-parameter subgroup of L introduced earlier. In particular, A00 is
modeled by a local Lie group of dimension one less than the dimension of L.
Now we come to a key observation, which is the main reason why we
work in the local groups category in the first place:
Lemma 1.9.8 (Preservation of the NSS property). A00 is NSS.
We will in fact prove a stronger claim:
Lemma 1.9.9 (Lifting lemma). If g A00 , then there exists g A0 such
that (
g ) = g and k
g ke,A0  kgke,A00 , where : A0 A00 is the projection
map.
Since A0 is NSS, all non-identity elements g of A0 have non-zero escape
norm, and so by the lifting lemma, all non-identity elements of A00 also have
non-zero escape norm, giving Lemma 1.9.8.
Proof. (Proof of Lemma 1.9.9) We choose g to be a lift of g (i.e. an element
of 1 (
g ) in A0 ) that minimises the escape norm k
g ke,A0 . (Such a minimum
0
exists since A is nonstandard finite, thanks to Loss theorem.) If g is trivial,
then so is g and there is nothing to prove. Therefore we may assume that
g is not the identity and hence, since A0 is NSS, that it has positive escape
norm. Suppose, by way of contradiction, that kgke,A0 /P = o(k
g ke,A0 ). Our
goal will be to reach a contradiction by finding another lift h of g with
strictly smaller escape norm than g. We will do this by setting h = gum for
some suitably chosen m.
We may assume that kgke,A00 /P is infinitesimal, since otherwise there
is nothing to prove; in particular g lies in the kernel of the local model

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.9. The microscopic structure of approximate groups

199

: A0 /P U/(c, c). We may thus find a lift g of g in the kernel of . In

particular, we may assume that g has infinitesimal escape norm.
Set M := 1/k
g ke,A0 , then M is unbounded. By hypothesis, kgke,A0 /P =
n
o(1/M ); thus g A0 /P whenever n = O(M ). In particular, for every
(standard) integer k N, g kn A0 /P . This implies that the group generated by g n lies in A0 /P . In particular, g n lies in the kernel of
, and hence
(
g n ) lies in (c, c) for all M n M .
By (an appropriate local version of) Exercise 1.9.7, we can find X
1.1B\B such that
(1.82)

(
g n ) = exp(st(n/M )r0 X)

whenever |n| rr0 M ; since (

g n ) lies in (c, c) for |n| M , we conclude
that X must be parallel to the generator 0 (0) of . Similarly, we have
(1.83)

(un ) = exp(st(n/N )rY )

whenever |n| 4N (say) for some Y 1.1B\B that is also parallel to 0 (0).
In particular, Y = X for some
1
1.1.
1.1
Since 1/N = kuke,A is the minimal escape norm of non-identity elements
of A, we have k
g ke,A 1/N , and thus gi A2 \A for some 1 i N ; in
i
particular, (
g ) 6 exp(rB). Comparing this with (1.82) we see that
st(i/M )r0 X 6 rB
and thus

st

N
M

1 r
,
1.1 r0

and hence

M
r0
1.3 .
N
r
By the Euclidean algorithm, we can thus find a nonstandard integer number
m such that the quantity
M r
N r0
lies in the interval [0.5, 0.5]. In particular
:= 1 + m

N r0
.
Mr
then (as u commutes with g) we see for all |n| M that
|m| 2

If we set h := gum

hn = gn umn

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

200

1. Hilberts fifth problem

and thus by (1.82), (1.83)

(hn ) = exp((st(n/M )r0 + st(mn/N )r)X)
= exp((st(n/M )r0 X)
exp(r0 B)
for all |n| M . In particular, khke,A0 < 1/M . Since h is also a lift of g, this
contradicts the minimality of k
g ke,A0 = 1/M , and the claim follows.

Because the NSS property is preserved, it is possible to improve upon
Exercise 1.9.9:
Exercise 1.9.15. Strengthen Exercise 1.9.9 by ensuring the final quotient
G0k1 /G0k = G0k1 is nonstandard finite, and all the other quotients G0i /G0i+1
are central in G0 /G0i+1 .
As a consequence, one obtains a stronger structure theorem than Exercise 1.9.9. Call a symmetric subset U containing the identity in a local
group nilpotent of step at most s if every iterated commutator in U of length
s + 1 is well-defined and trivial.
Exercise 1.9.16 (Helfgott-Lindenstrauss conjecture).
(i) Let A be a (local) NSS ultra strong approximate group. Show that
there is a symmetric subset A0 of A containing the identity which
is nilpotent of some finite step, such that A is covered by a finite
number of left translates of A0 .
(ii) Let A be a global NSS ultra strong approximate group with ambient
group G. Show that there is a nonstandard nilpotent subgroup G0
of G such that A is covered by a finite number of left translates of
G0 .
(iii) Let A be an NSS strong K-approximate group in a global group
G. Show that there is a nilpotent subgroup G0 of G of step OK (1)
such that A can be covered by a finite number of left translates of
G0 .
(iv) Let A be a K-approximate group in a global group G. Show that
there exists a subgroup G0 of G and a normal subgroup N of G0
contained in A4 , such that A is covered by OK (1) left-translates of
G0 , and G0 /N is nilpotent of step OK (1).
In fact, a stronger statement is true, involving the nilprogressions defined
in Section 1.7:
Proposition 1.9.10.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.10. Applications

201

(i) If A is an NSS ultra strong approximate group, then there is an

ultra nilprogression Q in G such that A contains Q, and A can be
covered by finitely many left-translates of Q.
(ii) If A is an ultra approximate group, then there is an ultra coset nilprogression Q in G such that A4 contains Q, and A can be covered
by finitely many left-translates of Q.
(iii) For all K 1, there exists CK , sK , rK 1 such that, given a
finite K-approximate group A in a group G = (G, ), one can find
a coset nilprogression Q in G of rank at most rK and step at most
sK such that A4 contains Q, and A can be covered by at most CK
left-translates of Q.
This proposition is established in [BrGrTa2011]. The key point is to
use the lifting lemma to observe that if (with the notation of the preceding
discussion) A0 /P contains a large nilprogression, then A0 also contains a
large nilprogression. One consequence of this proposition is that there is
essentially no difference between local and global approximate groups, at
the qualitative level at least:
Corollary 1.9.11. Let A be a local K-approximate group. Then there exists a OK (1)-approximate subgroup A0 of A, with A covered by OK (1) lefttranslates of A0 , such that A0 is isomorphic to a global OK (1)-approximate
subgroup.
This is because coset nilprogressions (or large fractions thereof) can be
embedded into global groups; again, see [BrGrTa2011] for details.
For most applications, one does not need the full strength of Proposition
1.9.10; Exercise 1.9.16 will suffice. We will give some examples of this in the
next section.

1.10. Applications of the structural theory of approximate

groups
In Section 1.9, we obtained the following structural theorem concerning
approximate groups:
Theorem 1.10.1. Let A be a finite K-approximate group. Then there exists
a coset nilprogression P of rank and step OK (1) contained in A4 , such that
A is covered by OK (1) left-translates of P (and hence also by OK (1) righttranslates of P ).
Remark 1.10.2. Under some mild additional hypotheses (e.g. if the dimensions of P are sufficiently large, or if P is placed in a certain normal form,
details of which may be found in [BrGrTa2011]), a coset nilprogression P

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

202

1. Hilberts fifth problem

of rank and step OK (1) will be an OK (1)-approximate group, thus giving a

partial converse to Theorem 1.10.1. (It is not quite a full converse though,
even if one works qualitatively and forgets how the constants depend on K:
if A is covered by a bounded number of left- and right-translates gP, P g of
P , one needs the group elements g to approximately normalise P in some
sense if one wants to then conclude that A is an approximate group.) The
mild hypotheses alluded to above can be enforced in the statement of the
theorem, but we will not discuss this technicality here, and refer the reader
to the above-mentioned paper for details.
By placing the coset nilprogression in a virtually nilpotent group, we
have the following corollary in the global case:
Corollary 1.10.3. Let A be a finite K-approximate group in an ambient
group G. Then A is covered by OK (1) left cosets of a virtually nilpotent
subgroup G0 of G.
In this section, we give some applications of the above results. The
first application is to replace K-approximate group by sets of bounded
doubling:
Proposition 1.10.4. Let A be a finite non-empty subset of a (global) group
G such that |A2 | K|A|. Then there exists a coset nilprogression P of rank
and step OK (1) and cardinality |P | K |A| such that A can be covered by
OK (1) left-translates of P , and also by OK (1) right-translates of P .
We will also establish (a strengthening of) a well-known theorem of
Gromov [Gr1981] on groups of polynomial growth, as promised back in
Section 1.1, as well as a variant result (of a type known as a generalised
Margulis lemma) controlling the almost stabilisers of discrete actions of
isometries.
The material here is largely drawn from [BrGrTa2011].
1.10.1. Sets of bounded doubling. In this section we will deduce Proposition 1.10.4 from Theorem 1.10.1. This can be done using the general (nonabelian) additive combinatorics machinery from [Ta2008b], but we will give
here an alternate argument relying on a version of the Croot-Sisask lemma
used in Section 1.8 which is a little weaker with regards to quantitative
bounds, but is slightly simpler technically (once one has the Croot-Sisask
lemma).
We recall the Croot-Sisask lemma:
Lemma 1.10.5 (Croot-Sisask). [CrSi2010] Let A be a non-empty finite
subset of a group G such that |A2 | K|A|. Then any M 1, there is a

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.10. Applications

203

symmetric set S containing the origin with |S| K,M |A| such that
1
(1.84)
k1A 1A (g)1A 1A k`2 (G)
|A|3/2
M
for all g S.
Let A be as in Proposition 1.10.4. We apply Lemma 1.10.5 with some
large M depending on K to be chosen later. Then for any g S 100 one has
100 3/2
|A| .
k1A 1A (g)1A 1A k`2 (G)
M
Since 1A 1A has an `1 (G) norm of |A|2 and is supported on the set A2 ,
which has cardinality at most K|A|, we see from Cauchy-Schwarz that
k1A 1A k`2 (G) K |A|3/2
and hence (if M is large enough depending on K)
k (g)1A 1A k`2 (A2 ) K |A|3/2 .
In particular, we have |gA2 A2 | K |A|, thus every element of S 100 has
K |A| representations of the form xy 1 with x, y A2 . As there are at
most K 2 |A|2 pairs (x, y) with x, y A2 , we conclude that |S 100 | K |A|.
In particular, by the Ruzsa covering lemma (Exercise 1.8.21) we see that S 4
can be covered by OK,M (1) left-translates of S 2 , and hence S 2 is a OK,M (1)approximate group.
In view of Theorem 1.10.1, we thus see that to conclude the proof of
Proposition 1.10.4, it suffices to show that A can be covered by OK,M (1)
left-translates (or right-translates) of S 2 if M is sufficiently large depending
on K.
We will just prove the claim for left-translates, as the claim for righttranslates is similar. We will need the following useful inequality:
Lemma 1.10.6 (Ruzsa triangle inequality). Let A, B, C be finite non-empty
1 ||BC 1 |
subsets of a group G. Then |A C 1 | |AB |B|
.
Proof. Observe that if x is an element of A C 1 , so that x = ac1 for
some a A and c C, then x has at least |B| representations of the form
x = yz with y A B 1 and z B C 1 , since ac1 = (ab1 )(bc1 ) for all
b B. As there are only |A B 1 ||B C 1 | possible pairs (y, z) that could
form such representations, the claim follows.

Now we return to the proof of Proposition 1.10.4. From (1.84) and
Minkowskis inequality we see that




1
1
1A 1A
1S 1A 1A

|A|3/2 ,


|S|
M
2
` (G)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

204

1. Hilberts fifth problem

and thus (if M is sufficiently large depending on K)

1

3/2
1S 1A 1A
|S|
2 2 K |A|
` (A )
and in particular


1

1S 1A 1A
|S|

K |A|

` (A2 )

and thus by Youngs inequality

1

1S 1A
K 1,
|S|

` (G)
and so
1S 1A (x) K,M |A|
for some x G. In other words,
|S xA1 | K,M |A|
If we then set B := S xA1 , then
|B S 1 | |S 2 | K,M |A|
and
|A B 1 | |A2 x| = |A2 | K|A|
and hence by the Ruzsa triangle inequality
|A S| = |A S 1 | K,M |A|.
By the Ruzsa covering lemma (Exercise 1.8.21), this implies that A can be
covered by OK,M (1) left-translates of S, as required. This proves Proposition
1.10.4. By placing the coset nilprogression in a virtually nilpotent group,
we obtain a strengthening of Corollary 1.10.3:
Corollary 1.10.7. Let A be a finite non-empty subset of an ambient group
G such that |A2 | K|A|. Then A is covered by OK (1) left cosets (and also
by OK (1) right-cosets) of a virtually nilpotent subgroup G0 of G.
We remark that there is also an off-diagonal version of Proposition
1.10.4:
Proposition 1.10.8. Let A, B be finite non-empty subsets of a (global)
group G such that |AB| K|A|1/2 |B|1/2 . Then there exists a coset nilprogression P of rank and step OK (1) and cardinality |P | K |A| such that A
can be covered by OK (1) left-translates of P , and B can be covered by OK (1)
right-translates of P .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.10. Applications

205

This is a consequence of Theorem 1.10.1 combined with [Ta2008b, Theorem 4.6]; we omit the details. There is also a statistical variant (using
[Ta2008b, Theorem 5.4] instead), based on an additional tool, the (nonabelian) Balog-Szemeredi-Gowers theorem, which will not be discussed in
detail here:
Proposition 1.10.9. Let A, B be finite non-empty subsets of a (global)
group G such that
|{(a, b, a0 , b0 ) A B A B : ab = a0 b0 }| |A|3/2 |B|3/2 /K.
Then there exists a coset nilprogression P of rank and step OK (1) and cardinality |P | K |A| such that A intersects a left-translate of P in a set
of cardinality K |A|, and B intersects a right-translate of P in a set of
cardinality K |A|.
1.10.2. Polynomial growth. The above results show that finite approximate groups A (as well as related objects, such as finite sets of bounded
doubling) can be efficiently covered by virtually nilpotent groups. However,
they do not place all of A inside a virtually nilpotent group. Indeed, this is
not possible in general:
Exercise 1.10.1. Let G be the ax + b group, that is to say the group of
all affine transformations x 7 ax + b on the real line, with a R\{0} and
b R. Show that there exists an absolute constant K and arbitrarily large
finite K-approximate groups A in G that are not contained in any virtually
nilpotent group. (Hint: build a set A which is very long in the b direction
and very thin in the a direction.)
Such counterexamples have the feature of being thin in at least one
of the directions of G. However, this can be fixed by adding a thickness
assumption to the approximate group. In particular, we have the following
result:
Theorem 1.10.10 (Thick sets of bounded doubling are virtually nilpotent).
For every K 1 there exists M 1 such that the following statement holds:
whenever G is a group, S is a finite symmetric subset of G containing the
identity, and A is a finite set containing S M such that |A2 | K|A|, then S
generates a virtually nilpotent group.
Proof. Fix K, and let M be a sufficiently large natural number depending
on K to be chosen later. Let S, A, G be as in the theorem. By Proposition
1.10.4, there exists a virtually nilpotent group H such that A is covered by
OK (1) left-cosets of H. In particular, S m H consists of OK (1) left-cosets
of H for all 0 m M . On the other hand, as S contains the identity,
S m H is nondecreasing in m. If M is large enough, then by the pigeonhole

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

206

1. Hilberts fifth problem

principle we may thus find some 0 m < M such that S m+1 H = S m H. By

induction this implies that S k H = S m H for all k m; we conclude that

[
hSi hSiH =
S k H = S m H.
k=m

In particular, H hSi has finite index in hSi. Since H is virtually nilpotent,

we conclude that hSi is virtually nilpotent also.

This theorem leads to the following Gromov-type theorem:
Exercise 1.10.2 (Gromov-type theorem). Show that for every C, d > 0
there exists M 1 such that the following statement holds: whenever G is
a group generated by a finite symmetric set S of generators containing the
identity, and |S m | Cmd |S| for some m M , then G is virtually nilpotent.
Note that this implies Gromovs original theorem (Theorem 1.1.30) as a
corollary, but it is stronger because (a) one only needs a polynomial growth
bound at a single scale m, rather than at all scales, and (b) the lower bound
required on m does not depend on the size of the generating set S. (A previous result in this direction, which obtained (a) but not (b), was established
in [ShTa2010], by a rather different argument based on an argument of
Kleiner [Kl2010]. The original proof of Gromov [Gr1981] of his theorem
had some features in common with the arguments given here, in particular using the machinery of Gromov-Hausdorff limits as well as some of the
theory surrounding Hilberts fifth problem, and was also amenable to nonstandard analysis methods as demonstrated in [vdDrWi1984], but differed
in a number of technical details.)
Remark 1.10.11. By inspecting the arguments carefully, one can obtain a
slightly sharper description of the group G in Exercise 1.10.2, namely that
G contains a normal subgroup G0 of index Od (1) which is the extension of
a finitely nilpotent group of step and rank O(d) by a finite group contained
in S m . See [BrGrTa2011] for details.
Exercise 1.10.3 (Gap between polynomial and non-polynomial growth).
Show that there exists a function f : N R+ which grows faster than any
polynomial (i.e. f (m)/md as m + for any d), with the property
that |S m | f (m)|S| whenever G is any finitely generated group that is
not virtually nilpotent, and S is any symmetric set of generators of G that
contains the identity.
Remark 1.10.12. No effective bound for the function f in this exercise
is explicitly known, though in principle one could eventually extract such
a bound by painstakingly finitising the proof of the structure theorem for
approximate groups. If one restricts the size of S to be bounded, then one

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.10. Applications

207

can take f (m) to be m(log log m) for some c > 0 and m sufficiently large
depending on the size of S, by the result of my paper with Shalom, but
this is unlikely to be best possible. (In the converse direction, Grigorchuks
construction [Gr1984] of a group of intermediate growth shows that f (m)
cannot grow faster than exp(m ) for some absolute constant < 1 (and it
is believed that one can take = 1/2).)
Exercise 1.10.4 (Infinite groups have at least linear growth). If G is an
infinite group generated by a finite symmetric set S containing the identity,
show that |S m | |S| + m 1 for all m 1.
Exercise 1.10.5 (Linear growth implies virtually cyclic). Let G be an infinite group generated by a finite symmetric set S containing the identity.
Suppose that G is of linear growth, in the sense that |S m | Cm for all
m 1 and some finite C.
(i) Place a left-invariant metric d on G by defining d(x, y) to be the
least natural number m for which x yS m . Define a geodesic to
be a finite or infinite sequence (gn )nI indexed by some discrete
interval I Z such that d(gn , gm ) = |n m| for all n, m I. Show
that there exist arbitrarily long finite geodesics.
(ii) Show that there exists a doubly infinite geodesic (gn )nZ with g0 =
1. (Hint: use (i) and a compactness argument.)
(iii) Show that |S 2m | 2|S m | 1 for all m 1. (Hint: study the balls
of radius m centred at gm and gm .) More generally, show that
|S km | 2k|S m | 2k + 1 for all m, k 1.
(iv) Show that |S m |/m converges to a finite non-zero limit as m ,
thus |S m | = m + o(m) for all m 1, where o(m) denotes a
quantity which, when divided by m, goes to zero as m .
(v) Show that for all m 1, then all elements of S m lie within a
distance at most o(m) of the geodesic (gm , . . . , gm ). (Hint: first
show that all but at most o(m) elements of S m lie within this
distance, using (iv) and the argument used to prove (iii).)
1 lies within distance o(m) of
(vi) Show that for sufficiently large m, gm
gm .

(vii) Show that for sufficiently large m, S m+1 lies within distance m of
1 }.
{1, gm , gm
(viii) Show that G is virtually cyclic (i.e. it has a cyclic subgroup of finite
index).
Exercise 1.10.6 (Nilpotent groups have polynomial growth). Let S be a
finite symmetric subset of a nilpotent group G containing the identity.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

208

1. Hilberts fifth problem

(i) Let s be an element of S that is not the identity, and let S 0 be

the minimal symmetric set containing S\{s, s1 } that is closed under the operations g 7 [g, s1 ]1 . Show that S 0 is also a finite
symmetric subset of G containing the identity, and that every element of S m can be written in the form si h for some |i| m and
2
h (S 0 )O(m ) , where the implied constant can depend on S, G.
(ii) Show that |S m | Cmd for all m 1 and some C, d > 0 depending
on S, G (i.e. G is of polynomial growth).
(iii) Show that every virtually nilpotent group is of polynomial growth.
1.10.3. Fundamental groups of compact manifolds (optional). This
section presupposes some familiarity with Riemannian geometry. Throughout this text, Riemannian manifolds are always understood to be complete
and without boundary.
We now apply the above theory to establish some relationships between
the topology (and more precisely, the fundamental group) of a compact Riemannian manifold, and the curvature of such manifolds. A basic theme
in this subject is that lower bounds on curvature tend to give somewhat
restrictive control on the topology of a manifold. Consider for instance Myers theorem [My1941], which among other things tells us that a connected
Riemannian manifold M with a uniform positive lower bound on the Ricci
curvature Ric is necessarily compact (with an explicit upper bound on the
diameter). In a similar vein we have the splitting theorem, which asserts that
if a connected Riemannian manifold M has everywhere non-negative Ricci
curvature, then it splits as the product of a Euclidean space and a manifold
without straight lines (i.e. embedded copises of R).
To analyse the fundamental group 1 (M ) of a connected Riemannian
manifold M , it is convenient to work with its universal cover:
Exercise 1.10.7. Let M be a connected Riemannian manifold.
(i) (Existence of universal cover) Show that there exists a simply con with the same dimension as M
nected Riemannian manifold M
M which is a local diffeowith a smooth surjective map : M
morphism and a Riemannian isometry (i.e. the metric tensors are
preserved); such a manifold (or more precisely, the pair (M, )) is
to be the space
known as a universal cover of M . (Hint: take M

of all paths from a fixed base point p0 in M , quotiented out by

homotopies fixing the endpoints. Once is constructed, pull back
the smooth and Riemannian structures.)
(ii) (Universality) Show that if M 0 is any smooth connected manifold
with a smooth surjective map f : M 0 M that is a local diffeomor are
phism and Riemannian isometry, and p0 M 0 , p M, p M

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.10. Applications

209

point such that f (p0 ) = (

p) = p, then there exists a unique smooth
M 0 with 0 (
a universal cover
map 0 : M
p) = p0 that makes M
0
of M also.
(iii) (Uniqueness) Show that a universal cover is unique up to isometric
isomorphism.
(iv) (Covering space) Show that for every p M there exists a neighbourhood U of p such that 1 (U ) is isometric (as a Riemannian
manifold) to 1 (M ) U , where we give the fundamental group
1 (M ) the discrete topology. In particular, the fibres 1 ({p}) of
a point p M are discrete and can be placed in bijection with
1 (M ).
(v) (Deck transformations) Show that 1 (M ) acts freely and isometrically on M 0 , in such a way that the orbits of 1 (M ) are the fibres
of . Conversely, show that every isometry on M 0 that preserves
the fibres of arises from an element of 1 (M ).
(vi) (Cocompactness) if M is compact, and 1 ({p}) is a fibre of M ,
show that every element of M 0 lies a distance at most diam(M )
from an element of the fibre 1 ({p}).
, show that the
(vii) (Finite generation) If M is compact and p M
set {g 1 (M ) : dist(gp, p) 2diam(M )} is finite and generates
1 (M ). (Hint: if gp is further than 2diam(M ) from p, use (vi) to
find a factorisation g = hk such that kp is closer to p or gp than p
is to gp.)
(viii) (Polynomial growth) If M is compact, show that the group 1 (M )
is of polynomial growth (thus |S m | Cmd for some generating set
S, some C, d 0, and all m 1) if and only if the universal cover
is of polynomial growth (thus Vol(B(x0 , r)) Crd for some base
M
, some C, d 0, and all r 1).
point x0 M
The above exercise thus links polynomial growth of groups to polynomial
growth of manifolds. To control the latter, a useful tool is the BishopGromov inequality:
Proposition 1.10.13 (Bishop-Gromov inequality). Let M be a Riemannian manifold whose Ricci curvature is everywhere bounded below by some
be the simply connected Riemannian manifold of
constant R. Let M
constant curvature and the same dimension as M (this will be a Euclidean
space for = 0, a sphere for > 0, and hyperbolic space for < 0). Let p
. Then the expression
be a point in M and p be a point in M
vol(B(p, r))
vol(B(
p, r))

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

210

1. Hilberts fifth problem

is monotone non-increasing in r.
We will not prove this proposition here (as it requires, among other
things, a definition of Ricci curvature, which would be beyond the scope of
this text); but see for instance [Pe2006, Chapter 9] or [Ta2009b, 2.10] for
a proof. This inequality is consistent with the geometric intuition that an
increase in curvature on a manifold should correspond to a stunting of the
growth of the volume of balls. For instance, in the positively curved spheres,
the volume of balls eventually stabilises as a constant; in the zero curvature
Euclidean spaces, the volume of balls grows polynomially; and in the negative curvature hyperbolic spaces, the volume of balls grows exponentially.
.
Informally, the balls in M cannot grow any faster than the balls in M
Setting = 0, we conclude in particular that if M has non-negative Ricci
curvature and dimension d, then vol(B(p, r))/rd is non-decreasing in r; in
particular, any manifold of non-negative Ricci curvature is of polynomial
growth. Applying Exercise 1.10.7 and Gromovs theorem, we conclude that
any manifold of non-negative Ricci curvature has a fundamental group which
is virtually nilpotent. (In fact, such fundametal groups can be shown, using
the splitting theorem, to be virtually abelian; see [ChGr1972]. However,
this improvement seems to be beyond the combinatorial methods used here.)
The above monotonicity also shows that whenever M has Ricci curvature
at least zero, we have the doubling bound
vol(B(p, 2r)) 2d vol(B(p, r)).
A continuity argument then shows that for every R > 0 and > 0, there
exists > 0 such that if M has Ricci curvature at least , then one has
vol(B(p, 2r)) (2d + ) vol(B(p, r))
for all 0 < r R.
Exercise 1.10.8. By using the above observation combined with Exercise
1.10.7 and Exercise 1.10.2, show that for every dimension d there exists
> 0 such that if M is any compact Riemannian manifold of diameter at
most 1 and Ricci curvature at least everywhere, then 1 (M ) is virtually
nilpotent.
Remark 1.10.14. This result was first conjectured by Gromov, and was
proven by Cheeger-Colding [ChCo1996] and Kapovitch-Wilking [KaWi2011]
using deep Riemannian geometry tools (beyond just the Bishop-Gromov inequality). (See also the paper [KaPeTu2010], which established the analogous result assuming a lower bound on sectional curvature rather than Ricci
curvature.)
The arguments in these papers in fact give more precise information on
the fundamental group 1 (M ), namely that there is a nilpotent subgroup of

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

1.10. Applications

211

step and rank Od (1) and index Od (1). The methods here (based purely on
controlling the growth of balls in 1 (M )) can give the step and rank bounds,
but appear to be insufficient to obtain the index bound. The exercise cannot be immediately obtained via a compactness-and-contradiction argument
from the (easier) = 0 case mentioned previously, because of the problem
of collapsing: there is no lower bound assumed on the injectivity radius of
M , and as such the space of all manifolds with the indicated diameter is
non-compact even if one bounds the derivatives of the metric to all orders.
(An equivalent way of phrasing the problem is that the orbits of 1 (M ) in
the universal cover may be arbitrarily dense, and so a ball of bounded radius
may correspond to an arbitrarily large subset S of the fundamental
in M
group. For this application it is thus of importance that there is no upper
bound on the size of the sets S or A assumed in Exercise 1.10.2.)
Remark 1.10.15. One way to view the above results is as an assertion that
it is quite rare for a compact manifold to be equippable with a Riemannian
metric with (almost) non-negative Ricci curvature. Indeed, an application of
van Kampens theorem [Se1931], [vKa1833] shows that every fundamental group 1 (M ) of a compact manifold is finitely presented, and conversely
a gluing argument for four-manifolds shows that every finitely presented
group is the fundamental group of some (four-dimensional) manifold. Intuitively, most finitely presented groups are not virtually nilpotent, and so
most compact manifolds cannot have metrics with almost non-negative
Ricci curvature.
1.10.4. A Margulis-type lemma. In Exercise 1.10.7 we saw that the fundamental group 1 (M ) of a connected Riemannian manifold can be viewed
. The
as a discrete group of isometries acting on a Riemannian manifold M

,
curvature properties of M then give doubling properties of the balls in M
and hence of 1 (M ), allowing one to use tools such as Exercise 1.10.2.
It turns out that one can abstract this process by replacing the universal
by a more general metric space X:
cover M
Lemma 1.10.16 (Margulis-type lemma). Let K 1. Let X = (X, d) be
a metric space, with the property that every ball of radius 4 can be covered
by K balls of radius 1. Let be a group of isometries of X, which acts
discretely in the sense that {g : gx B} is finite for every x X
and every bounded set B X. Then if x X and is sufficiently small
depending on K, the set S := {g : d(gx, x) } generates a virtually
nilpotent group.
Proof. We can can cover B(x, 4) by K balls B(xi , 1). If g, h S4 and
gx, hx B(xi , 1), then (by the isometric action of ) we see that h1 g S2 .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

212

1. Hilberts fifth problem

We conclude that S4 can be covered by K right-translates of S2 . Since

M S 2 for all M 1, the claim then follows from Exercise
S22 S4 and S2/M
1.10.2.

Roughly speaking, the above lemma asserts that for discrete actions of
isometries on spaces of bounded doubling, the almost stabiliser of a
point is virtually nilpotent. In the case when X is a Riemannian manifold
with a lower bound on curvature, this result was established in [ChCo1996],
[KaWi2011] (and, as mentioned in the previous section, stronger control
on S was established). The original lemma of Margulis addressed the case
when X was a hyperbolic space, and relied on commutator estimates not
unrelated to the commutator estimates of Gleason metrics and of strong
approximate groups that were used in previous sections.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Chapter 2

Related articles

213

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

214

2. Related articles

2.1. The Jordan-Schur theorem

We recall Jordans theorem, proven in Exercise 1.1.3:
Theorem 2.1.1 (Jordans theorem). Let G be a finite subgroup of the general linear group GLd (C). Then there is an abelian subgroup G0 of G of
index [G : G0 ] Cd , where Cd depends only on d.
Informally, Jordans theorem asserts that finite linear groups over the
complex numbers are almost abelian. The theorem can be extended to other
fields of characteristic zero, and also to fields of positive characteristic so long
as the characteristic does not divide the order of G, but we will not consider
these generalisations here.
The finiteness hypothesis on the group G in this theorem can be relaxed
to the significantly weaker condition of periodicity. Recall that a group G
is periodic if all elements are of finite order. Jordans theorem with finite
replaced by periodic is known as the Jordan-Schur theorem.
The Jordan-Schur theorem can be quickly deduced from Jordans theorem, and the following result of Schur:
Theorem 2.1.2 (Schurs theorem). Every finitely generated periodic subgroup of a general linear group GLd (C) is finite. (Equivalently, every periodic linear group is locally finite.)
Remark 2.1.3. The question of whether all finitely generated periodic
subgroups (not necessarily linear in nature) were finite was known as the
Burnside problem; the answer was shown to be negative in [GoSa1964].
Let us see how Jordans theorem and Schurs theorem combine via a
compactness argument to form the Jordan-Schur theorem. Let G be a periodic subgroup of GLd (C). Then for every finite subset S of G, the group
GS generated by S is finite by Theorem 2.1.2. Applying Jordans theorem,
GS contains an abelian subgroup G0S of index at most Cd .
In particular, given any finite number S1 , . . . , Sm of finite subsets of G,
we can find abelian subgroups G0S1 , . . . , G0Sm of GS1 , . . . , GSm respectively
such that each G0Sj has index at most Cd in GSj . We claim that we may
furthermore impose the compatibility condition G0Si = G0Sj GSi whenever
Si Sj . To see this, we set S := S1 Sm , locate an abelian subgroup
G0S of GS of index at most Cd , and then set G0Si := G0S GSi . As GS is
covered by at most Cd cosets of G0S , we see that GSi is covered by at most
Cd cosets of G0Si , and the claim follows.
Note that for each S, the set of possible G0S is finite, and so the product
space of all configurations (G0S )SG , as S ranges over finite subsets of G,
is compact by Tychonoff s theorem. Using the finite intersection property,

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.1. The Jordan-Schur theorem

215

we may thus locate a subgroup G0S of GS of index at most Cd for all finite
subsets S of G, obeying the compatibility
condition G0T = G0S GT whenever
S
T S. If we then set G0 := S G0S , where S ranges over all finite subsets
of G, we then easily verify that G0 is abelian and has index at most Cd in
G, as required.
2.1.1. Proofs. We now give a proof of Schurs theorem, based on the proof
in [We1973]. We begin with a lemma of Burnside. Given a vector space
V , let End(V ) denote the ring of linear transformations from V to itself.
Lemma 2.1.4. Let V be a finite-dimensional complex vector space, and
let A be a complex algebra with identity in End(V ), i.e. a linear subspace
of End(V ) that is closed under multiplication and contains the identity operator. Then either A = End(V ), or there exists some proper subspace
{0} ( W ( V which is A-invariant, i.e. aW W for all a A.
Proof. Suppose that no such proper A-invariant subspace W exists. Then
for any non-zero v V , the vector space Av must equal all of V , since it
is a non-trivial A-invariant subspace. By duality, this implies that for any
non-zero dual vector u V , the vector space A u must equal all of V .
Let u, v be linearly independent elements of V . We claim that there
exists an element a of A such that au 6= 0 and av = 0. Suppose that this
is not the case; then by the Hahn-Banach theorem, there exists t End(V )
such that au = tav for all a A. In particular, setting a = 1 we obtain
u = tv, and thus atv = tav. Replacing a by ab for some b A, we conclude
that tabv = abtv = atbv, thus taat annihilates all of Av. Since Av = V , we
conclude that at = ta, thus A lies in the centraliser of t. But since u = tv
and u, v are linearly independent, t is not a multiple of the identity, and
thus by the spectral theorem, t has at least one proper eigenspace. But this
eigenspace is fixed by A, a contradiction.
Thus we can find a A such that au 6= 0 and av = 0 for any linearly
independent u, v. Iterating this, we see that for any non-zero u V and
any 0 k dim(V ) 1, we can find a A of corank at least k that does
not annihilate u. In particular, A contains a rank one transformation. Since
Av = V and A u = V for all v V and u V , this implies that A
contains all rank one transformations, and hence contains all of End(V ) by
linearity.

Corollary 2.1.5 (Burnsides theorem). Let V be a complex vector space
of some finite dimension d, and let G be a subgroup of GL(V ) where every
element has order at most r. Then G is finite with cardinality Od,r (1).
Proof. We induct on dimension, assuming the claim has already been proven
for smaller values of d. Let A End(V ) be the complex algebra generated

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

216

2. Related articles

by G (or equivalently, the complex linear span of G). Suppose first that there
is a proper A-invariant subspace W . Then G projects down to GL(W ) and
to GL(V /W ), and by induction hypothesis both of these projections are finite with cardinality Od,r (1). Thus there exists a subgroup G0 of G of index
Od,r (1) whose projections to GL(W ) and GL(V /W ) are trivial; in particular, all elements of G0 are unipotent. But as the complex numbers have zero
characteristic, the only unipotent element of finite order is the identity, and
so G0 is trivial, and the claim follows.
Hence we may assume that V has no proper A-invariant subspace. By
Lemma 2.1.4, A must be all of GL(V ). In particular, one can find d2 linearly
independent elements g1 , . . . , gd2 of G.
For any g G, the element ggi has order at most r, and thus all the
eigenvalues of ggi are roots of unity of order at most r. This means that
there are at most Or,d (1) possible values of the trace tr(ggi ), which is a linear
functional of g. Letting gi vary among the basis g1 , . . . , gd2 of End(V ), we
conclude that there are at most Or,d (1) possible values of g, and the claim
follows.

Remark 2.1.6. The question of whether any finite group with m generators in which all elements of order at most r is necessarily of order Om,r (1)
is known as the restricted Burnside problem, and was famously solved affirmatively by Zelmanov [Ze1990]. (Note however that for certain values of m
and r it is possible for the group to be infinite. Also, while any finite group
is trivially embedded in some linear group, one does not have any obvious
control on the dimension of that group in terms of m and r, so one cannot
immediately solve this problem just from Corollary 2.1.5.)
To prove Schurs theorem (Theorem 2.1.2), it thus suffices to establish
the following proposition:
Proposition 2.1.7. Let k be a finitely generated extension of the field of
rationals Q. Then every periodic element of GLd (k) has order at most
Od,k (1).
Indeed, to obtain Schurs theorem, one applies Proposition 2.1.7 with
k equal to the field generated by the coefficients of the generators of the
finitely generated periodic group G, and then applies Corollary 2.1.5.
Proof. Suppose first that k is a finite extension of Q. If a GLd (k) has
period n, then the field generated by the eigenvalues of a contains a primitive
nth root of unity, and thus contains the cyclotomic field of that order. On the
other hand, this field has degree Od (1) over k, and thus has degree Od,k (1)
over the rationals. Thus n = Od,k (1), and the claim follows. Note that the
bound on n depends only on the degree of k, and not on k itself.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.2. Nilpotent groups and nilprogressions

217

Now we extend from the finite degree case to the finitely generated case.
By using a transcendence basis, one can write k as a finite extension of
Q(z1 , . . . , zm ) for some algebraically independent z1 , . . . , zm over Q. By the
primitive element theorem, one can then write k = Q(z1 , . . . , zm )() where
is algebraic over Q(z1 , . . . , zm ) of some degree D.
Now suppose we have an element a GLd (k) of period n, thus a, . . . , an1 6=
1 and an = 1. Let R be the ring in k generated by the coefficients of a. We
can create a ring homomorphism : R k to a finite extension k of the
rationals by mapping each zi to a rational number qi , and then replace
with a root of the polynomial formed by replacing zi with qi in the minimal
polynomial of . As long as one chooses (q1 , . . . , qm ) generically (i.e. outside of a codimension one subset of Qm ), this operation is well-defined on
R (in that no division by zero issues arise in any of the coefficients of a).
Furthermore, generically one has (a), . . . , (a)n1 6= 1 and (a)n = 1, thus
a has period n. Furthermore, the degree of k over Q is at most the degree
D of over Q(z1 , . . . , zm ) and is thus bounded uniformly in q1 , . . . , qm . The
claim now follows from the finite extension case.


2.2. Nilpotent groups and nilprogressions

In this section we present some of the algebra of nilpotent groups, and as an
application establish that all nilprogressions are approximate groups (Proposition 1.1.28). This turns out to be a moderately lengthy computation; an
easier partial result in this direction appears in [BrGrTa2011], in which
an additional hypothesis is imposed that the nilprogression be in C-normal
form.
There are several ways to think about nilpotent groups; for instance one
can use the model example of the Heisenberg group

1 R R
H(R) := 0 1 R
0 0 1
over an arbitrary ring R (which need not be commutative), or more generally
any matrix group consisting of unipotent upper triangular matrices, and
view a general nilpotent group as being an abstract generalisation of such
concrete groups. (In the case of nilpotent Lie groups, at least, this is quite
an accurate intuition, thanks to Engels theorem.) Or, one can adopt a Lietheoretic viewpoint and try to think of nilpotent groups as somehow arising
from nilpotent Lie algebras; this intuition is rigorous when working with
nilpotent Lie groups (at least when the characteristic is large, in order to
avoid issues coming from the denominators in the Baker-Campbell-Hausdorff
formula), but also retains some conceptual value in the non-Lie setting. In

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

218

2. Related articles

particular, nilpotent groups (particularly finitely generated ones) can be

viewed in some sense as nilpotent Lie groups over Z, even though Lie
theory does not quite work perfectly when the underlying scalars merely
form an integral domain instead of a field.
Another point of view, which arises naturally both in analysis and in
algebraic geometry, is to view nilpotent groups as modeling infinitesimal
perturbations of the identity, where the infinitesimals have a certain finite
order. For instance, given a (not necessarily commutative) ring R without identity (representing all the small elements of some larger ring or
algebra), we can form the powers Rj for j = 1, 2, . . . , defined as the ring
generated by j-fold products r1 . . . rj of elements r1 , . . . , rj in R; this is an
ideal of R which represents the elements which are j th order in some sense.
If one then formally adjoins an identity 1 onto the ring R, then for any s 1,
the multiplicative group G := 1 + R mod Rs+1 is a nilpotent group of step
at most s. For instance, if R is the ring of strictly upper s s matrices (over
some base ring), then Rs+1 vanishes and G becomes the group of unipotent
upper triangular matrices over the same ring, thus recovering the previous
matrix-based example. In analysis applications, R might be a ring of operators which are somehow of order O() or O(~) for some small parameter
or ~, and one wishes to perform Taylor expansions up to order O(s ) or
O(~s ), thus discarding (i.e. quotienting out) all errors in Rs+1 .
From a dynamical or group-theoretic perspective, one can also view
nilpotent groups as towers of central extensions of a trivial group. Finitely
generated nilpotent groups can also be profitably viewed as a special type
of polycylic group; this is the perspective taken for instance in [Ta2010b,
2.13]. Last, but not least, one can view nilpotent groups from a combinatorial group theory perspective, as being words from some set of generators
of various degrees subject to some commutation relations, with commutators of two low-degree generators being expressed in terms of higher degree
objects, and all commutators of a sufficiently high degree vanishing. In particular, generators of a given degree can be moved freely around a word, as
long as one is willing to generate commutator errors of higher degree.
With this last perspective, in particular, one can start computing in
nilpotent groups by adopting the philosophy that the lowest order terms
should be attended to first, without much initial concern for the higher order
errors generated in the process of organising the lower order terms. Only
after the lower order terms are in place should attention then turn to higher
order terms, working successively up the hierarchy of degrees until all terms
are dealt with. This turns out to be a relatively straightforward philosophy
to implement in many cases (particularly if one is not interested in explicit
expressions and constants, being content instead with qualitative expansions

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.2. Nilpotent groups and nilprogressions

219

of controlled complexity), but the arguments are necessarily recursive in

nature and as such can become a bit messy, and require a fair amount of
notation to express precisely. So, unfortunately, the arguments here will be
somewhat cumbersome and notation-heavy, even if the underlying methods
of proof are relatively simple.
2.2.1. Some elementary group theory. Let g, h be two elements of a
group G = (G, ). We define the conjugate g h and commutator 1 [g, h] by the
formulae
(2.1)

g h := h1 gh

and
(2.2)

[g, h] := g 1 h1 gh.

Conjugation by a fixed element h is an automorphism of G, thus

(gk)h = g h k h
and
(g h )1 = (g 1 )h
and
1h = 1
and
[g, k]h = [g h , k h ]
for all g, h, k G. Conjugation is also an action, thus
(g h )k = g hk
and
g 1 = g.
An automorphism of the form g 7 g h is called an inner automorphism.
Conjugation is related to multiplication by the identity
gh = hg h ,
thus one can pull g to the right of h at the cost of twisting (i.e. conjugating)
it by h. Commutation is related to multiplication by the identity
gh = hg[g, h],
1Note that this convention for [g, h] is not universal; for instance, the alternate convention
[g, h] = ghg 1 h1 also appears in the literature. The distinctions between the two conventions
however are quite minor; the conventions here are optimised for pulling group elements to the
right of a word, whereas other conventions may be slightly better for pulling group elements to
the left of a word.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

220

2. Related articles

thus one can pull g to the right of h at the cost of adding an additional
commutator factor [g, h] to the right. Finally, commutation is related to
conjugation by the identity
g h = g[g, h].
The commutator can be viewed as a nonlinear group-theoretic analogue
of the Lie bracket. For instance, in a matrix group G = GLn (C), we observe
that the commutator [1 + A, 1 + B] of two elements 1 + A and 1 + B
close to the identity is of the form 1 + 2 (AB BA) + O(3 ), thus linking
the group-theoretic commutator to the Lie bracket A, B 7 AB BA.
Because of this link, we expect the group-theoretic commutator to obey
some nonlinear analogues of the basic Lie bracket identities, and this is
indeed the case. For instance, one easily observes that the commutator is
antisymmetric in the sense that
(2.3)

[h, g] = [g, h]1

and is approximately odd in the sense that

(2.4)

[g 1 , h] = ([g, h]1 )g

and
(2.5)

[g, h1 ] = ([g, h]1 )h

for any g, h G. We also have the easily verified approximate bilinearity

identities
[g, hk] = [g, h][g, k]h
and
[gh, k] = [g, k]h [h, k]
for any g, h, k G. Finally, we have the approximate Jacobi identity (better
known as the Hall-Witt identity)
(2.6)

[[g, h1 ], k]h [[h, k 1 ], g]k [[k, g 1 ], h]g = 1.

A subgroup H of G is said to be normal if it is preserved by all inner

automorphisms, thus H g = H for all g (writing H g := {hg : h H},
of course), and characteristic if it is preserved by all automorphisms (not
necessarily inner). Thus, all characteristic subgroups are normal, but the
converse is not necessarily true. We write H G or G H if H is a
subgroup of G, and H B G or G C H if H is a normal subgroup of G.
If N is a normal subgroup of G, we write g 7 g mod N for the quotient
map from G to G/N , thus g = h mod N if g hN = N h. Given any other
subgroup H of G, we write H mod N for the image of H under the mod N
map, thus
H mod N HN/N = N H/N H/(H N ).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.2. Nilpotent groups and nilprogressions

221

Given two subgroups H, K of a group G, we define the commutator

group [H, K] to be the group generated by the commutators [h, k] with
h H, k K. We say that H and K commute if [H, K] is trivial, or
equivalently if every element of H commutes with every element of K. Note
that if H, K are normal, then [H, K] is also normal. In this case, one can
view [H, K] as the smallest subgroup of G such that H, K commute modulo
[H, K] (or equivalently, that H mod [H, K] and K mod [H, K] commute).
Similarly, if H, K are characteristic, then [H, K] is also characteristic.
Observe from (2.3) that
[H, K] = [K, H]
for any subgroups H, K. Also that if N is a normal subgroup of G, then (as
mod N is a homomorphism)
HK mod N = (H mod N )(K mod N )
and
(2.7)

[H, K] mod N = [H mod N, K mod N ]

Exercise 2.2.1.
(i) If H, K are normal subgroups of G generated by
subsets A H, B K respectively, show that [H, K] is the normal
subgroup of G generated (as a normal subgroup) by the commutators [a, b] with a A, b B.
(ii) If N, H, K are normal subgroups of G, show that [[N, H], K] lies
in the normal subgroup generated by [[H, K], N ] and [[K, N ], H].
(Hint: use the Hall-Witt identity (2.6).)
Given an arbitrary group G, we define the lower central series G =
G1 C G2 . . . of G by setting G1 := G and Gi+1 := [G, Gi ] for all i 1.
Observe that all of the groups Gi in this series are characteristic and thus
normal. A group G is said to be nilpotent of step at most s if Gs+1 is trivial;
in particular, by (2.7), we see that G mod Gs+1 = G/Gs+1 is nilpotent of
step at most s for any group G. We have a basic inclusion:
Exercise 2.2.2 (Filtration property). We have [Gi , Gj ] Gi+j for all i, j
1. (Hint: use Exercise 2.2.1.)
The commutator structure can be clarified by passing to the top order
components of this commutator, as follows. (Strictly speaking, this analysis is not needed to study nilprogressions, but it is still conceptually useful
nonetheless.) Consider the subquotients Vi := Gi mod Gi+1 of the group
G for i = 1, 2, . . . . As Gi+1 contains [Gi , Gi ], we see that each group Vi is
abelian. To emphasise this, we will write Vi additively instead of multiplicatively, thus we shall denote the group operation on Vi as +.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

222

2. Related articles

Lemma 2.2.1. For each i, j 1, the commutator map [, ] : Gi Gj

[Gi , Gj ] descends to a map [, ]i,j : Vi Vj Vi+j , thus
(2.8)

[g, h] mod Gi+j+1 = [g mod Gi+1 , h mod Gj+1 ]i,j

for g Gi , h Gj .
Proof. We can quotient out by Gi+j+1 , and assume that Gi+j+1 is trivial.
In particular, by Lemma 2.2.2, Gi+1 commutes with Gj , and Gi commutes
with Gj+1 . As such, it is easy to see that if g Gi , h Gj , then [g, h] is
unchanged if one multiplies g (on the left or right) by an element of Gi+1 ,
and similarly for h and Gj+1 . This defines the map [, ]i,j as required.

We refer to the maps [, ]i,j as quotiented commutator maps. The identity
(2.8) asserts that these maps [, ]i,j capture the top order nonlinear behaviour of the group G. As the following exercise shows, these maps behave
like (graded components of) a Lie bracket.
Exercise 2.2.3. Let G be a group with lower central series G = G1 C G2 C
. . . , subquotients Vi , and quotiented commutator maps [, ]i,j : Vi Vj
Vi+j . Let i, j, k 1, let x, x0 Vi , y, y 0 Vj , and z Vk . Establish the
antisymmetry
[x, y]i,j = [y, x]j,i ,
the bihomomorphism properties
[x + x0 , y]i,j = [x, y]i,j + [x0 , y]i,j
[x, y]i,j = [x, y]i,j
[0, y]i,j = 0
and
[x, y + y 0 ]i,j = [x, y]i,j + [x, y 0 ]i,j
[x, y]i,j = [x, y]i,j
[x, 0]i,j = 0
and the Jacobi identity
[[x, y]i,j , z]i+j,k + [[y, z]j,k , x]j+k,i + [[z, x]k,i , y]k+i,j = 0.
Remark 2.2.2. If one wished, one could combine all of these subquotients
Vi and quotiented commutator maps [, ]i,j into a single graded object, namely
the additive group V := i Vi consisting of tuples (vi )
i=1 with vi Vi (and
all but finitely many of the vi trivial), with a bracket [, ] : V V V defined
by
X

[(vi )
[vi , wj ]i,j )
i=1 , (wj )j=1 ] := (
k=1 .
i+j=k

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.2. Nilpotent groups and nilprogressions

223

Then this bracket is an antisymmetric bihomomorphism obeying the Jacobi

identity, thus V is a Lie algebra over the integers Z. One could view this
object as the top order or Carnot component of the Lie algebra of G.
We will however not need this object here.
We can iterate the bilinear commutator operations to create multilinear operations. Given a non-empty finite set A of formal group elements
g, we can define a formal commutator word w on A to be a word that can
be generated by the following rules:
(1) If g is a formal group element, then g is a formal commutator word
on {g}.
(2) If A, B are disjoint finite non-empty sets of formal group elements,
and u, v are formal commutator words on A, B respectively, then
[u, v] is a formal commutator word on A B.
We refer to |A| as the length of the formal commutator word. Thus, for
instance, [[g1 , g2 ], [g3 , g4 ]] is a formal commutator word on {g1 , g2 , g3 , g4 }
of length 4. Given a formal commutator word w on A and an assignment
g : g 7 g(g) of group elements g(g) G to each formal group element
g A, we can define the group element w(g) G by substituting g(g) for
g in w for each g G. This gives a commutator word map w : GA G.
Given an assignment i : g 7 i(g) of a positive integer (or degree) i(g)
to each formal
group element A, Lemma 2.2.1 and induction then gives a
Q
map wi : gA Vi(g) VPgA i(g) which is a multihomomorphism (i.e. a
homomorphism in each variable). From (2.8) one has that
w(g) mod GPgA i(g)+1 = wi (g mod Gi+1 )
where g mod Gi+1 is shorthand for the assignment g 7 g(g) mod Gi(g)+1 .
In particular, using the degree map i(g) := 1, we obtain a multihomomorphism
w1 : V1l Vl
for any commutator word w of length l, such that
w(g) mod Gl+1 = w1 (g mod G2 ).
From the multihomomorphism nature of w1 , we conclude in particular
that
(2.9)

w(g1n1 , . . . , glnl ) = w(g1 , . . . , gl )n1 ...nl mod Gl+1

for any g1 , . . . , gl G and integers n1 , . . . , nl . A variant of this approximate

identity will be key in understanding nilprogressions.
One can use commutator words to generate a nilpotent group G and its
lower central series:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

224

2. Related articles

Exercise 2.2.4. Let G be a nilpotent group that is generated by a set S of

generators.
(i) Show that for every positive integer i, Gi is generated by the words
w(g), where w ranges over formal commutator words of length l at
least i, and g is a collection of l generators from S (possibly with
repetition).
(ii) Suppose further that S is finite (so that there are only finitely many
possible choices for w(g) for any given length l, and let e1 , . . . , en
be all the values of w(g) as w varies over formal commutator words
of length at most s, and g ranges over collections of generators from
S, arranged in non-decreasing length of w (this arrangement is not
unique, and may contain repeated values). Show that for every
positive integer i, any element in Gi can be expressed in the form
a
ej j . . . eann , where ej , . . . , en are the elements of e1 , . . . , en arising
from words w of length at least i, and aj , . . . , an are integers.
2.2.2. Nilprogressions. Recall from Section 1.1 that a noncommutative
progression P (a1 , . . . , ar ; N1 , . . . , Nr ) in a group G with generators a1 , . . . , ar
G and dimensions N1 , . . . , Nr > 0 is the set of all words with alphabet
1
1
are used a
a1
1 , . . . , ar , such that for each 1 i r, the symbols ai , ai
total of at most Ni times. A nilprogression is a noncommutative progression
in a nilpotent group. The objective of this section is to prove Proposition
1.1.28, restated here:
Proposition 2.2.3. Suppose that a1 , . . . , ar G generate a nilpotent group
of step s, and suppose that N1 , . . . , Nr are all sufficiently large depending on
r, s. Then P := P (a1 , . . . , ar ; N1 , . . . , Nr ) is an Or,s (1)-approximate group.
Recall that a subset A of a group G is a K-approximate group if it is
symmetric, contains the origin, and if A2 can be covered by K left-translates
(or equivalently, right-translates) of A. The first two properties are clear, so
it suffices to show that P 2 can be covered by Or,s (1) right-translates of P .
We allow all implied constants to depend on r, s. We pick a small constant > 0 (depending on r, s). It will suffice to show the inclusion
P (a1 , . . . , ar ; 2N1 , . . . , 2Nr ) P (a1 , . . . , ar ; O(N1 + 1), . . . , O(Nr + 1))X
for some set X of cardinality |X|  1.
We will need some auxiliary objects. For i 1 and t > 0, let
(2.10)

Qti (a1 , . . . , ar ; N1 , . . . , Nr ) := P (a01 , . . . , a0R ; tN10 , . . . , tNR0 ),

where a01 , . . . , a0R consists of all group elements of the form w(ai1 , . . . , ail )
where w is a formal commutator word of length l i, 1 i1 , . . . , il r, and
each a0j = w(ai1 , . . . , ail ) is associated to the dimension Nj0 := Ni1 . . . Nil

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.2. Nilpotent groups and nilprogressions

225

(note that we allow some of the a0i to be equal, or to degenerate to the identity). Thus the Qi (a1 , . . . , ar ; N1 , . . . , Nr ) are non-increasing in i and become
trivial for i s. We will usually abbreviate Qti (a1 , . . . , ar ; N1 , . . . , Nr ) as Qti .
Trivially we also have
P (a1 , . . . , ar ; N1 , . . . , Nr ) Qt1 .
0

Observe also that the Qti are symmetric, contain the identity, and Qti Qti
0
Qt+t
.
i
O(1)

Exercise 2.2.5 (Approximate filtration property). If g Qi

Qj (a1 , . . . , ar ; N1 , . . . , Nr ), show that

and h

O(1)

[g, h] Qi+j
and
O(1)

g h Qi

Next, we need the following variant of (2.9).

Lemma 2.2.4. Let w be a formal commutator word of length l 1, and let
1 i1 , . . . , il r. For each 1 j r, let nj be an integer with nj = O(Nij ).
Then
O(1)
w(ani11 , . . . , anill ) w(ai1 , . . . , ail )n1 ...nl Ql+1 .
Proof. When l > s, the expressions involving w collapse to the identity,
and the claim follows, so we may assume that 1 l s. We induct on l.
The claim l = 1 is trivial, so suppose that 1 < l s and that the claim has
been proven for smaller values of l.
We first establish the key case l = 2. In this case, it suffices to show that
(2.11)

O(1)

[an , bm ] [a, b]nm Q3

(a, b; |n|, |m|)

for any a, b G and n, m Z. By using commutator identities (2.4), (2.5)

we may assume that n, m are positive. It suffices to show that
m

O(1)

(an )b an [a, b]nm Q3

bm

(a, b; n, m).

bm

We can write (an ) = (a )n . It is then not difficult to see (and we leave

as an exercise to the reader) that the claim will follow if we can show
m

O(1)

(a)b a[a, b]m Q3

(a, b; 1, m),

or equivalently
O(1)

[a, bm ] [a, b]m Q3

(a, b; 1, m).

Thus we have effectively reduced the problem to the case n = 1. A similar

argument allows us to also obtain the additional reduction m = 1, at which
point the claim is trivial. This completes the treatment of the l = 2 case.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

226

2. Related articles

Now we handle the general case. After some relabeling, we may write
w = [u, v], where u, v are words of length l1 , l2 respectively for some l1 , l2 < l
adding up to l, with
nl

nl

w(ani11 , . . . , anill ) = [u(ani11 , . . . , ail 1 ), v(ail 1+1 , . . . , anill )]

+1

and similarly
w(ai1 , . . . , ail ) = [u(ai1 , . . . , ail1 , v(ail1 +1 , . . . , ail )].
By induction hypothesis, one has
nl

O(1)

u(ani11 , . . . , ail 1 ) u(ai1 , . . . , ail1 )n1 ...nl1 Ql1 +1 .

1

In particular we have
nl

O(1)

u(ani11 , . . . , ail 1 ) Ql1

Similarly we have
nl

O(1)

O(1)

v(ail 1+1 , . . . , anill ) v(ail1 +1 , . . . , ail )nli +1 ...nl Ql2 +1 Ql2

+1

Using (2.2.5) and the approximate homomorphism properties of commutators, we conclude that
nl

nl

O(1)

[u(ani11 , . . . , ail 1 ), v(ail 1+1 , . . . , anill )] [u(ai1 , . . . , ail1 )n1 ...nl1 , v(ail1 +1 , . . . , ail )nli +1 ...nl ]Ql+1 ,
+1

but from (2.11) we have

O(1)

[u(ai1 , . . . , ail1 )n1 ...nl1 , v(ail1 +1 , . . . , ail )nli +1 ...nl ] [u(ai1 , . . . , ail1 ), v(ail1 +1 , . . . , ail )]n1 ...nl Ql+1 .
The claim follows.

Exercise 2.2.6. For positive integers a, b, show that

O(1)

g n an [a, b]nm Q3

(a, b; n, m)

whenever
O(1)

g a[a, b]m Q3

(a, b; 1, m).

Next, we need the following elementary number-theoretic lemma:

Exercise 2.2.7. Let M1 , . . . , Ml 1, and let n be an integer with n =
O(M1 . . . Ml ). Show that n can be expressed as the sum of Ol (1) terms of
the form n1 . . . nl , where the n1 , . . . , nl are integers with ni = O(Mi ) for each
i = 1, . . . , l. (Hint: Despite the superficial similarity here with non-trivial
number-theoretic questions such as the Waring problem, this is actually a
very elementary fact which can be proven by induction on l.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.2. Nilpotent groups and nilprogressions

227

Exercise 2.2.8. Let w be a formal commutator word of length l 1, and

let 1 i1 , . . . , il r. Let n = O(Ni1 . . . Nir ). Show that
O(1)

w(ai1 , . . . , ail )n P O(1) Ql+1 .

Conclude that
O(1)

O(1)
QC
Ql+1
l P

whenever C = O(1), and on iteration conclude that

O(1)
QC
1 P

whenever C = O(1).
A similar argument shows that Q1 P for a sufficiently small >
0 depending only on r, s, assuming that N1 , . . . , Nr are sufficiently large
depending on r, s. Since P 2 Q21 , it thus suffices to show that for any
O()
> 0, that Q21 can be covered by O (1) right-translates of Q1 . But this
then follows by iterating the following exercise:
Exercise 2.2.9. Let > 0, l 1, and C = O(1). Show that QC
l can be
O()
O(1)
covered by O (1) right-translates of Ql
Ql+1 . (Hint: this is similar to
into
words w(ai1 , . . . , ail ) of length
Exercise 2.2.5. Factor an element of QC
l
l, together with words of higher length. Gather all the words of length l into
O(1)
monomials w(ai1 , . . . , ail )n with n = O(Ni1 . . . Nil ), times a factor in Ql+1 .
Split these monomials into a monomial with exponent O(Ni1 . . . Nil ), times
a monomial which can take at most O (1) possible values. Then push the
O(1)
latter monomials (and the Ql+1 factor) to the right.)
Exercise 2.2.10 (Polynomial growth). Suppose that a1 , . . . , ar G generate a nilpotent group of step s, and suppose that N1 , . . . , Nr 1. Show
that
|P (a1 , . . . , ar ; N1 , . . . , Nr )| r,s (N1 . . . Nr )Or,s (1) .
Exercise 2.2.11. Suppose that a1 , . . . , ar G generate a nilpotent group
of step s, and suppose that N1 , . . . , Nr are sufficiently large depending on
r, s. Write P := |P (a1 , . . . , ar ; N1 , . . . , Nr )|. Let a01 , . . . , a0R , N10 , . . . , NR0 be
as in (2.10), arranged in non-decreasing order of the length of the associated
formal commutator words.
(i) Show that every element of P can be represented in the form
(a01 )n1 . . . (a0R )nR
for some integers ni = Or,s (Ni0 ). (We do not claim however that
this representation is unique, and indeed the generators a01 , . . . , a0R
are likely to contain quite a lot of redundancy.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

228

2. Related articles

(ii) Conversely, show that there exists an > 0 depending only on r, s

such that any expression of the form
(a01 )n1 . . . (a0R )nR
with integers ni with |ni | Ni0 , lies in P .

2.3. Ados theorem

Recall from Definition 1.2.13 that a (complex) abstract Lie algebra is a complex vector space g (either finite or infinite dimensional) equipped with a
bilinear antisymmetric form [] : g g g that obeys the Jacobi identity
(2.12)

[[X, Y ], Z] + [[Y, Z], X] + [[Z, X], Y ] = 0.

One can of course define Lie algebras over other fields than the complex
numbers C, but it will be convenieng in this section to work over an algebraically closed field. For most of this text it is preferable to work over the
reals R instead, but in many cases one can pass from the complex theory to
the real theory by complexifying the Lie algebra.
An important special case of the abstract Lie algebras are the concrete
Lie algebras, in which g End(V ) is a (complex) vector space of linear
transformations X : V V on a vector space V (which again can be either
finite or infinite dimensional), and the bilinear form is given by the usual
Lie bracket
[X, Y ] := XY Y X.
It is easy to verify that every concrete Lie algebra is an abstract Lie algebra.
In the converse direction, we have
Theorem 2.3.1. Every abstract Lie algebra is isomorphic to a concrete Lie
algebra.
To prove this theorem, we introduce the useful algebraic tool of the
universal enveloping algebra U (g) of the abstract Lie algebra g. This is
the free (associative, complex) algebra generated by g (viewed as a complex
vector space), subject to the constraints
(2.13)

[X, Y ] = XY Y X.

This algebra is described by the Poincare-Birkhoff-Witt theorem, which asserts that given an ordered basis (Xi )iI of g as a vector space, that a basis
of U (g) is given by monomials of the form
(2.14)

Xia11 . . . Xiamm

where m is a natural number, the i1 < < im are an increasing sequence

of indices in I, and the a1 , . . . , am are positive integers. Indeed, given two

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.3. Ados theorem

229

such monomials, one can express their product as a finite linear combination of further monomials of the form (2.14) after repeatedly applying (2.13)
(which we rewrite as XY = Y X + [X, Y ]) to reorder the terms in this product modulo lower order terms until one all monomials have their indices
in the required increasing order. It is then a routine exercise in basic abstract algebra (using all the axioms of an abstract Lie algebra) to verify that
this is multiplication rule on monomials does indeed define a complex associative algebra which has the universal properties required of the universal
enveloping algebra.
The abstract Lie algebra g acts on its universal enveloping algebra
U (g) by left-multiplication: X : M 7 XM , thus giving a map from g to
End(U (g)). It is easy to verify that this map is a Lie algebra homomorphism (so this is indeed an action (or representation) of the Lie algebra),
and this action is clearly faithful (i.e. the map from g to End(U g) is injective), since each element X of g maps the identity element 1 of U (g) to a
different element of U (g), namely X. Thus g is isomorphic to its image in
End(U (g)), proving Theorem 2.3.1.
In the converse direction, every representation : g End(V ) of a Lie
algebra factors through the universal enveloping algebra, in that it extends
to an algebra homomorphism from U (g) to End(V ), which by abuse of
notation we shall also call .
One drawback of Theorem 2.3.1 is that the space U (g) that the concrete
Lie algebra acts on will almost always be infinite-dimensional, even when
the original Lie algebra g is finite-dimensional. However, there is a useful
theorem of Ado that rectifies this:
Theorem 2.3.2 (Ados theorem). Every finite-dimensional abstract Lie algebra is isomorphic to a concrete Lie algebra over a finite-dimensional vector
space V .
Among other things, this theorem can be used (in conjunction with
the Baker-Campbell-Hausdorff formula) to show that every abstract (finitedimensional) Lie group (or abstract local Lie group) is locally isomorphic
to a linear group, a result also known as Lies third theorem; see Section
1.2. It is well-known, though, that abstract Lie groups are not necessarily
globally isomorphic to a linear group, see for instance Exercise 1.1.5 for a
counterexample.
Ados theorem is surprisingly tricky to prove in general, but some special
cases are easy. For instance, one can try using the adjoint representation
ad : g End(g) of g on itself, defined by the action X : Y 7 [X, Y ]; the Jacobi identity (1.8) ensures that this indeed a representation of g. The kernel
of this representation is the centre Z(g) := {X g : [X, Y ] = 0 for all Y

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

230

2. Related articles

g}. This already gives Ados theorem in the case when g is semisimple, in
which case the center is trivial.
The adjoint representation does not suffice, by itself, to prove Ados
theorem in the non-semisimple case. However, it does provide an important reduction in the proof, namely it reduces matters to showing that every finite-dimensional Lie algebra g has a finite-dimensional representation
: g End(V ) which is faithful on the centre Z(g). Indeed, if one has such
a representation, one can then take the direct sum of that representation
with the adjoint representation to obtain a new finite-dimensional representation which is now faithful on all of g, which then gives Ados theorem for
g.
It remains to find a finite-dimensional representation of g which is faithful on the centre Z(g). In the case when g is abelian, so that the centre
Z(g) is all of g, this is again easy, because g then acts faithfully on g C
by the infinitesimal shear maps X : (Y, t) 7 (tX, 0). In matrix form, this
representation identifies each X in this abelian Lie algebra with an uppertriangular matrix:


0 X
X
.
0 0
This construction gives a faithful finite-dimensional representation of
the centre Z(g) of any finite-dimensional Lie algebra. The standard proof of
Ados theorem (which I believe dates back to work of Harish-Chandra) then
proceeds by gradually extending this representation of the centre Z(g) to
larger and larger sub-algebras of g, while preserving the finite-dimensionality
of the representation and the faithfulness on Z(g), until one obtains a representation on the entire Lie algebra g with the required properties. (For
technical inductive reasons, one also needs to carry along an additional property of the representation, namely that it maps the nilradical to nilpotent
elements, but we will discuss this technicality later.)
This procedure is a little tricky to execute in general, but becomes simpler in the nilpotent case, in which the lower central series g1 := g; gn+1 :=
[g, gn ] becomes trivial for sufficiently large n:
Theorem 2.3.3 (Ados theorem for nilpotent Lie algebras). Let n be a
finite-dimensional nilpotent Lie algebra. Then there exists a finite-dimensional
faithful representation : n End(V ) of n. Furthermore, there exists a natural number k such that (n)k = {0}, i.e. one has (X1 ) . . . (Xk ) = 0 for
all X1 , . . . , Xk n.
The second conclusion of Ados theorem here is useful for induction
purposes. (By Engels theorem, this conclusion is also equivalent to the

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.3. Ados theorem

231

assertion that every element of (n) is nilpotent, but we can prove Theorem
2.3.3 without explicitly invoking Engels theorem.)
Below the fold, we give a proof of Theorem 2.3.3, and then extend the
argument to cover the full strength of Ados theorem. The presentation here
is based on the one in in [FuHa1991].
2.3.1. The nilpotent case. We first prove Theorem 2.3.3. We achieve this
by an induction on the dimension of n. The claim is trivial for dimension
zero, so we assume inductively that n has positive dimension, and that
Theorem 2.3.3 has already been proven for all lower-dimensional nilpotent
Lie algebras.
As noted earlier, the adjoint representation already verifies the claim
except for the fact that it is not faithful on the centre Z(n). (The nilpotency
of n ensures the existence of some k for which ad(n)k = 0.) Thus, it will
suffice to find a finite-dimensional representation : n End(V ) that is
faithful on the center Z(n), and for which (n)k = {0} for some k.
We have already verified the claim when n is abelian (and can take k = 1
in this case), so suppose that n is not abelian; in particular, the centre Z(n)
has strictly smaller dimension. We then observe that n must contain a codimension one ideal a containing Z(n), which will of course again be a nilpotent
Lie algebra. This can be seen by passing to the quotient n0 := n/Z(n) (which
has positive dimension) and then to the abelianisation n0 /[n0 , n0 ] (which still
has positive dimension, by nilpotency), arbitrarily selecting a codimension
one subspace of that abelianisation, and then passing back to n.
If we let h be a one-dimensional complementary subspace of a, then h is
automatically an abelian Lie algebra, and we have the decomposition
n=ah
as vector spaces. This is not necessarily a direct sum of Lie algebras, though,
because a and h need not commute. However, as a is an ideal, we do know
that [h, a] a, thus there is an adjoint action ad : h End(a) of h on a.
By induction hypothesis, we know that there is a faithful representation
0 : a End(V0 ) on some finite-dimensional space V0 with 0 (a)k0 = {0}
for some k0 . We would like to somehow extend this representation to a
finite-dimensional representation : n End(V ) which is still faithful on a
(and hence on Z(n)), and with (n)k = {0} for some k.
To motivate the construction, let us begin not with 0 , but with the
universal enveloping algebra representation A : M 7 AM of a on U (a). The
idea is to somehow combine this action a with an action of h on the same
space U (a) to obtain an action of the full Lie algebra n on U (a). To do this,
recall that we have the adjoint action ad(H) : A 7 [H, A] of h on a. This

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

232

2. Related articles

extends to an adjoint action of h on U (a), defined by the Leibniz rule

(2.15)

[H, A1 . . . Am ] :=

m
X

A1 . . . Ai1 [H, Ai ]Ai+1 . . . Am ;

i=1

one can easily verify that this is indeed an action (and furthermore that
the map ad(H) : M 7 [H, M ] is a derivation on U (a)). One can then
combine the multiplicative action A : M 7 AM of a and the adjoint action
H : M 7 [H, M ] of h into a joint action
(2.16)

(A + H) : M 7 AM + [H, M ]

of a + h = n on U (a). A minor algebraic miracle occurs here, namely that

this combined action is still a genuine action of the Lie algebra a + h. In
particular one has
[A + H, A0 + H 0 ]M = (A + H)(A0 + H 0 )M (A0 + H 0 )(A + H)M
for all A + H, A0 + H 0 a + h, as can be verified by a brief computation.
Also, because the action of a was already faithful on U (a), this enlarged
action of n will still be faithful on a.
Remark 2.3.4. One can explain this miracle by working first in the
larger universal algebra U (n). This space has a left multiplicative action
X : M 7 XM of n, but also has a right multiplicative action X : M 7
M X which commutes with the left multiplicative action. Furthermore,
the splitting n = a + h induces a projection map : n h which will be
a Lie algebra homomorphism because a is an ideal. This gives a projected
right multiplicative action X : M 7 M (X) which still commutes with
the left multiplicative action. In particular, the combined action X : M 7
XM M (X) is still an action of n. One then observes that this action
preserves U (a), and when restricted to that space, becomes precisely (2.16).
Now we need to project the above construction down to a finite-dimensional
representation. It turns out to be convenient not to use the original representation 0 directly, but only indirectly via the property 0 (a)k0 = {0}.
Specifically, we consider the (two-sided) ideal I := h(a)k0 i of U (a) generated
by k0 -fold products of elements in a, and then consider the quotient algebra
U (a)/I. Because 0 (a)k0 = {0}, this algebra still maps to End(V0 ); since a
was faithful in 0 , it must also be faithful in U (a)/I.
The point of doing this quotienting is that whereas U (a) is likely to be
infinite-dimensional, the space U (a)/I is only finite-dimensional. Indeed, it
is generated by those monomials (2.14) whose total degree a1 + + am is
less than k0 ; since a is finite-dimensional, there are only finitely many such
monomials.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.3. Ados theorem

233

From the Leibniz rule (2.15) we see that the adjoint action of h on U (a)
preserves I, and thus descends to an action on U (a)/I. We can combine this
action with the multiplicative action of a as before using (2.16) to create an
action of n on U (a)/I, which is now a finite-dimensional representation which
(as observed previously) is faithful on a.
The only remaining thing to show is that for some sufficiently large k,
that nk annihilates U (a)/I. By linearity, it suffices to show that
X1 . . . Xk A1 . . . Am = 0 mod I
whenever X1 , . . . , Xk lie in either a or h, and A1 . . . Am is a monomial in
U (a). But observe that if one multiplies a monomial A1 . . . Am by an element
A of a, then one gets a monomial AA1 . . . Am of one higher degree; and if
instead one multiplies a monomial A1 . . . Am by an element H of h, then from
(2.15) one gets a sum of monomials in which one of the terms Ai has been
replaced with the higher order term [H, Ai ]. Using the nilpotency of n
(which implies that all sufficiently long iterated commutators must vanish),
we thus see that if k is large enough, then X1 . . . Xk A1 . . . Am will consist
only of terms of degree at least k0 , which automatically lie in I, and the
claim follows.
Remark 2.3.5. The above argument gives an effective (though not particularly efficient) bound on the dimension of V and on the order k of (n) in
terms of the dimension of n. Similarly for the arguments we give below to
prove more general versions of Ados theorem.
2.3.2. The solvable case. To go beyond the nilpotent case one needs to
use more of the structural theory of Lie algebras. In particular, we will use
Engels theorem:
Theorem 2.3.6 (Engels theorem). Let g End(V ) be a concrete Lie
algebra on a finite-dimensional space V which is concretely nilpotent in the
sense that every element of g is a nilpotent operator on V . Then there exists
a basis of V such that g consists entirely of upper-triangular matrices. In
particular, if d is the dimension of V , then gd = {0} (i.e. X1 . . . Xd = 0 for
all X1 , . . . , Xd g).
The proof of Engels theorem is not too difficult, but we omit it here
as we have nothing to add to the textbook proofs of this theorem (such as
the one in [FuHa1991]). Note that this theorem, combined with Theorem
2.3.3, gives a correspondence between concretely nilpotent Lie algebras and
abstractly nilpotent Lie algebras.
Engels theorem has the following consequence:

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

234

2. Related articles

Corollary 2.3.7 (Nilpotent ideals are null). Let g End(V ) be a concrete

Lie algebra on a finite-dimensional space V , and let a be a concretely nilpotent Lie algebra ideal of g. Then for any X1 , . . . , Xm g with at least one
of the Xi in a, one has tr(X1 . . . Xm ) = 0.
Proof. Set Vi := ai V to be the vector space spanned by vectors of the form
A1 . . . Ai v for A1 , . . . , Ai a and v V , then V = V0 Vd = {0} is
a flag, with a mapping Vi to Vi+1 for each i. As a is an ideal of g, we see
that each of the spaces Vi is invariant with respect to g. Thus X1 . . . Xm
also maps Vi to Vi+1 for each i, and so must have zero trace.

This leads to the following curious algebraic fact:
Corollary 2.3.8. Let g End(V ) be a concrete Lie algebra on a finitedimensional space V , and let a, b be ideals of g such that b [a, g]. If [a, b]
is concretely nilpotent, then b is also concretely nilpotent.
Proof. Let B b. We need to show that B is nilpotent. By the CayleyHamilton theorem, it suffices to show that tr(B k ) = 0 for each k 1. Since
B b [a, g], it suffices to show that tr([A, X]B k1 ) = 0 for each k 1,
A a, and X g. But we can use the cyclic property of trace to rearrange
tr([A, X]B k1 ) = tr(X[A, B k1 ]).
We can then use the Leibniz rule to expand
[A, B k1 ] =

k1
X

B i1 [A, B]B ki1 .

i=1

But [A, B] lies in [a, b], and the claim now follows from Corollary 2.3.7.

We can apply this corollary to the radical r of an (abstract) finitedimensional Lie algebra g, defined as the maximal solvable ideal of the Lie
algebra. (One can easily verify that the vector space sum of two solvable
ideals is again a solvable ideal, which implies that the radical is well-defined.)
Theorem 2.3.9. Let g be a finite-dimensional (abstract) Lie algebra, and
let r be a solvable ideal in g. Then [g, g] r is an (abstractly) nilpotent ideal
of g.
Proof. Without loss of generality we may take r to be the radical of g.
Let us first verify the theorem in the case when g is a concrete Lie algebra
over a finite-dimensional vector space V . We let r(i+1) := [r(i) , r(i) ] be the
derived series of r. One easily verifies that each of the r(i) are ideals of g.
We will show by downward induction on i that the ideals [g, g] r(i) are
concretely nilpotent. As r is solvable, this claim is trivial for i large enough.
Now suppose that i is such that the ideal [g, g]r(i+1) is concretely nilpotent.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.3. Ados theorem

235

This ideal contains [r(i) , [r(i) , g]], which is then also concretely nilpotent. By
Corollary 2.3.8, this implies that the ideal [r(i) , g] is concretely nilpotent,
and hence the smaller ideal [g, [g, g] r(i) ] is also concretely nilpotent. By
a second application of Corollary 2.3.8, we conclude that [g, g] r(i) is also
concretely nilpotent, closing the induction. This proves the theorem in the
concrete case.
To establish the claim in the abstract case, we simply use the adjoint
representation, which effectively quotients out the centre Z(g) of g (which
will also be an ideal of r) to convert an finite-dimensional abstract Lie algebra
into a concrete Lie algebra over a finite-dimensional space. We conclude that
the quotient of [g, g] r by the central ideal Z(g) is nilpotent, which implies
that [g, g] r is nilpotent as required.

This has the following corollary. Recall that a derivation D : g g on an
abstract Lie algebra g is a linear map such that D[X, Y ] = [DX, Y ]+[X, DY ]
for each X, Y g. Examples of derivations include the inner derivations
DX := [A, X] for some fixed A g, but not all derivations are inner.
Corollary 2.3.10. Let g be a finite-dimensional (abstract) Lie algebra, and
let r be a solvable ideal in g. Then for any derivation D : g g, Dr is
nilpotent.
Proof. If D were an inner derivation, the claim would follow easily from
Theorem 2.3.9. In the non-inner case, the trick is simply to view D as an
inner derivation coming from an extension of g. Namely, we work in the
enlarged Lie algebra g oD C, which is the Cartesian product g C with Lie
bracket
[(X, s), (Y, t)] := ([X, Y ] + sDy tDx, 0).
One easily verifies that this is an (abstract) Lie algebra which contains g
g {0} as a subalgebra. The derivation D on g then arises from the inner
derivation on g oD C coming from (0, 1). The claim then easily follows by
applying Theorem 2.3.9 to this enlarged algebra.

Remark 2.3.11. I would be curious to know if there is a more direct proof
of Corollary 2.3.10 (which in particular did not need the full strength of
Engels theorem), as this would give a simpler proof of Ados theorem in the
solvable case (indeed, it seems almost equivalent to that theorem, especially
in view of Lies third theorem, Exercise 1.2.23).
Define the nilradical of an (abstract) finite-dimensional Lie algebra g to
be the maximal nilpotent ideal in g. One can verify that the sum of two
nilpotent ideals is again a nilpotent ideal, so the nilradical is well-defined.
One can verify that the radical r is characteristic (i.e. preserved by all

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

236

2. Related articles

derivations), and so Dr is also characteristic; in particular, it is an ideal.

We conclude that Dr is in fact contained in the nilradical of g.
We can now prove Ados theorem in the solvable case:
Theorem 2.3.12 (Ados theorem for solvable Lie algebras). Let r be a finitedimensional solvable Lie algebra. Then there exists a finite-dimensional
faithful representation : r End(V ) of r. Furthermore, if n is the nilradical of r, one can ensure that (n) is nilpotent.
Proof. This is similar to the proof of Theorem 2.3.3. We induct on the
relative dimension r/n of r relative to its nilradical. When this dimension is
zero, the claim follows from Theorem 2.3.3, so we assume that this dimension
is positive, and the claim has already been proven for lesser dimensions.
Using the adjoint representation as before, it suffices to establish a representation which is nilpotent on n and faithful on the centre Z(r). As the
centre is contained in the nilradical, it will suffice to establish nilpotency
and faithfulness on n.
As in the proof of Theorem 2.3.3, we can find a codimension one ideal
a of r that contains n, and so can split r = a h as vector spaces, for
some abelian one-dimensional Lie algebra h. Note that the nilradical of a
is characteristic in a, and hence a nilpotent ideal in r; as a consequence, it
must be identical to the nilradical n of r.
By induction hypothesis, we have a finite-dimensional faithful representation 0 : a End(V0 ) which is nilpotent on n; we extend this representation to the universal enveloping algebra U (a). In particular (by Engels
theorem), there exists a natural number k such that 0 (nk ) = 0.
Let I be the two-sided ideal in the universal enveloping algebra U (a)
generated by the kernel ker(0 ) (in the universal enveloping algebra U (a),
not the Lie algebra) and n. Then the ideal I k is contained in the kernel of
0 . Since 0 is faithful on a, we see that the quotient map from U (a) to
U (a)/I is also faithful on a.
By the Cayley-Hamilton theorem, for every element A of a there is a
monic polynomial P (A) of A that is annihilated by 0 and is thus in I.
The monic polynomial P (A)k is thus in I k . Using this, we see that we can
express any monomial (2.14) with at least one sufficiently large exponent ai
in terms of monomials of lower degree modulo I k . From this we see that
U (a)/I k can be generated by just finitely many such monomials, and is in
particular finite-dimensional.
The Lie algebra h has an adjoint action on a. These are derivations on
a, so by Corollary 2.3.10, they take values in n and in particular in I. If
we extend these derivations to the universal enveloping algebra U (a) by the

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.4. Associativity of the Baker-Campbell-Hausdorff-Dynkin law

237

Leibniz rule (cf. (2.15)), we thus see that these derivations preserve I, and
thus (by the Leibniz rule) also preserve I k . Thus this also gives an action
of h on U (a)/I k by derivations.
In analogy with (2.16), we may now combine the actions of a and h on
U (a)/I k to an action of r which will remain faithful and nilpotent on n, and
the claim follows.

2.3.3. The general case. Finally, we handle the general case. Actually
we can use basically the same argument as in preceding cases, but we need
one additional ingredient, namely Levis theorem:
Theorem 2.3.13 (Levis theorem). Let g be a finite-dimensional (abstract)
Lie algebra. Then there exists a splitting g = r h as vector spaces, where
r is the radical of g and h is another Lie algebra.
We remark that h necessarily has trivial radical and is thus semisimple.
It is easy to see that the quotient g/r is semisimple; the entire difficulty of
Levis theorem is to ensure that one can lift this quotient back up into the
original space g. This requires some knowledge of the structural theory of
semisimple Lie algebras. The proof will be omitted, as I have nothing to
add to the textbook proofs of this result (such as the one in [FuHa1991]).
Using Levis theorem and Theorem 2.3.12, one obtains the full Ado
theorem:
Theorem 2.3.14 (Ados theorem for general Lie algebras). Let g be a finitedimensional Lie algebra. Then there exists a finite-dimensional faithful representation : g End(V ) of g. Furthermore, if n is the nilradical of r,
one can ensure that (n) is nilpotent.
Exercise 2.3.1. Prove Theorem 2.3.14 (and thus also Theorem 2.3.2).
(Hint: deduce this theorem from Theorem 2.3.12 and Theorem 2.3.13 by
using the same argument used to deduce Theorem 2.3.12 from the induction
hypothesis, with the key point again being Corollary 2.3.10 that places the
adjoint action of h on r inside n.)

2.4. Associativity of the Baker-Campbell-Hausdorff-Dynkin

law
Let g be a finite-dimensional Lie algebra (over the reals). Given two sufficiently small elements x, y of g, define the right Baker-Campbell-HausdorffDynkin law
Z 1
(2.17)
Ry (x) := x +
FR (Adx Adty )y dt
0

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

238

2. Related articles

where Adx := exp(adx ), adx : g g is the adjoint map adx (y) := [x, y], and
log z
, which is analytic for z near 1. Similarly,
FR is the function FR (z) := zz1
define the left Baker-Campbell-Hausdorff-Dynkin law
Z 1
(2.18)
Lx (y) := y +
FL (Adtx Ady )x dt
0
z
where FL (z) := log
z1 . One easily verifies that these expressions are welldefined (and depend smoothly on x and y) when x and y are sufficiently
small.

We have the famous Baker-Campbell-Hausdoff-Dynkin formula:

Theorem 2.4.1 (BCH formula). Let G be a finite-dimensional Lie group
over the reals with Lie algebra g. Let log be a local inverse of the exponential
map exp : g G, defined in a neighbourhood of the identity. Then for
sufficiently small x, y g, one has
log(exp(x) exp(y)) = Ry (x) = Lx (y).
See Section 1.2 for a proof of this formula. In particular, one can give
a neighbourhood of the identity in g the structure of a local Lie group by
defining the group operation as
(2.19)

x y := Ry (x) = Lx (y)

for sufficiently small x, y, and the inverse operation by x1 := x (one easily

verifies that Rx (x) = Lx (x) = 0 for all small x).
It is tempting to reverse the BCH formula and conclude (the local form
of) Lies third theorem, that every finite-dimensional Lie algebra is isomorphic to the Lie algebra of some local Lie group, by using (2.19) to define a
smooth local group structure on a neighbourhood of the identity; see Exercise 1.2.23. The main difficulty in doing so is in verifying that the definition
(2.19) is well-defined (i.e. that Ry (x) is always equal to Lx (y)) and locally
associative. The well-definedness issue can be trivially disposed of by using
just one of the expressions Ry (x) or Lx (y) as the definition of (though, as
we shall see, it will be very convenient to use both of them simultaneously).
However, the associativity is not obvious at all.
With the assistance of Ados theorem (Theorem 2.3.2), which places g
inside the general linear Lie algebra gln (R) for some n, one can deduce both
the well-definedness and associativity of (2.19) from the Baker-CampbellHausdorff formula for gln (R). However, Ados theorem is rather difficult
to prove, and it is natural to ask whether there is a way to establish these
facts without Ados theorem, thus giving an independent proof of the local
version of Lies third theorem. This will be the purpose of this section.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.4. Associativity of the Baker-Campbell-Hausdorff-Dynkin law

239

The key is to observe that the right and left BCH laws commute with
each other:
Proposition 2.4.2 (Commutativity). Let g be a finite-dimensional Lie algebra. Then for sufficiently small x, y, z, one has
(2.20)

Ly (Rz (x)) = Rz (Ly (x)).

Note that this commutativity has to hold if (2.19) is to be both welldefined and associative. Assuming Proposition 2.4.2, we can set x = 0 in
(2.20) and use the easily verified identities Rz (0) = z, Ly (0) = y to conclude
that Ly (z) = Rz (y) for small y, z, ensuring that (2.19) is well-defined; and
then inserting (2.19) into (2.20) we obtain the desired (local) associativity.
It remains to prove Proposition 2.4.2. We first make a convenient observation. Thanks to the Jacobi identity, the adjoint representation ad : x 7
adx is a Lie algebra homomorphism from g to the Lie algebra gl(g). As this
latter Lie algebra is the Lie algebra of a Lie group, namely GL(g), the BakerCampbell-Hausdorff formula is valid for that Lie algebra. In particular, one
has
log(exp(adx ) exp(ady )) = Rady (adx ) = Ladx (ady )
for sufficiently small x, y. But as ad is a Lie algebra homomorphism, one
has
Rady (adx ) = adRy (x)
and similarly
Ladx (ady ) = adLx (y) .
Exponentiating, we conclude that
(2.21)

Adx Ady = AdRy (x) = AdLx (y) .

This would already give what we want if the adjoint representation was
faithful. We unfortunately cannot assume this (and this is the main reason,
by the way, why Ados theorem is so difficult), but we can at least use (2.21)
to rewrite the formulae (2.17), (2.18) as
Z 1
Ry (x) = x +
FR (AdRty (x) ) dt
0

and
Z
Lx (y) := y +
0

FL (AdLtx (y) ) dt.

This leads to the important radial homogeneity identities

R(s+t)y (x) = Rsy (Rty (x))
and
L(s+t)x (y) = Lsx (Ltx (y))

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

240

2. Related articles

for all sufficiently small x, y g and 0 s, t 1, as can be seen by a short

computation.
Because of these radial homogeneity identities (together with the smoothness of the right and left BCH laws), it will now suffice to prove the approximate commutativity law
(2.22)

Ly (Rz (x)) = Rz (Ly (x)) + O(|y|2 |z|) + O(|y||z|2 )

for all small x, y, z. Indeed, this law implies that

(2.23)

Ly/n Rz/n = Rz/n Ly/n + O(1/n3 )

for fixed small y, z, a large natural number n, and with the understanding
that the operations are only applied to sufficiently small elements x. From
n , and so a large
radial homogeneity we have Ly = Lny/n and Rz = Rz/n
number (O(n2 ), to be more precise) of iterations of (2.23) (using uniform
smoothness to control all errors) gives
Ly Rz = Rz Ly + O(1/n),
and the claim (2.20) then follows by sending n .
It remains to prove (2.22). When y = 0, then Ly is the identity map and
the claim is trivial; similarly if z = 0. By Taylor expansion, it thus suffices
to establish the infinitesimal commutativity law


Lay (Rbz (x))|a=b=0 =
Rbz (Lay (x))|a=b=0 .
a b
a b
(One can interpret this infinitesimal commutativity as a commutativity of
the vector fields corresponding to the infinitesimal generators of the left
and right BCH laws, although we will not explicitly adopt that perspective
here.) This is a simplification, because the infinitesimal versions of (2.17),
(2.18) are simpler than the non-infinitesimal versions. Indeed, from the
fundamental theorem of calculus one has

Lay (w)|a=0 = FL (Adw )y

a
for any fixed y, w, and similarly

Rbz (v)|b=0 = FR (Adv )z.

b
Thus it suffices (by Clairauts theorem) to show that

FL (AdRbz (x) )y|b=0 =

FR (AdLaz (x) )z|a=0 .
b
a
It will be more convenient to work with the reciprocals FL1 , FR1 of the
functions FL , FR . Recall the general matrix identity
d 1
A (t) = A1 (t)A0 (t)A1 (t)
dt
(2.24)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.4. Associativity of the Baker-Campbell-Hausdorff-Dynkin law

241

for any smoothly varying invertible matrix function A(t) of a real parameter
t. Using this identity, we can write the left-hand side of (2.24) as
1
F (AdRbz (x) )|b=0 )FL (Adx )y.
b L
If we write Y := FL (Adx )y and Z := FR (Adx )z, then from Taylor expansion
we have
Rbz (x) = x + bZ + O(|b|2 )
FL (Adx )(

and so we can simplify the above expression as

FL (Adx )(

1
F (Adx+bZ )|b=0 )Y.
b L

Similarly, the right-hand side of (2.24) is

FR (Adx )(

1
F (Adx+aY )|a=0 )Z.
a R

Since FR (Adx ) = Adx FL (Adx ), it thus suffices to show that

1
FL (Adx+bZ )|b=0 )Y = Adx ( FR1 (Adx+aY )|a=0 )Z.
b
a
Now, we write
Z 1
1
Adtx dt
FL (Adx ) =

(2.25)

and
FR1 (Adx )

Z
=

Adtx dt
0

and thus expand (2.25) as

Z 1
Z 1

(2.26)
( Adtx+tbZ )|b=0 Y dt = Adx
( AdtxtaY )|a=0 Z dt.
b
b
0
0
We write Ad as the exponential of ad. Using the Duhamel matrix identity
Z 1
d
exp(A(t)) =
exp(sA(t))A0 (t) exp((1 s)A(t)) dt
dt
0
for any smoothly varying matrix function A(t) of a real variable t, together
with the linearity of ad, we see that
Z 1

( Adtx+tbZ )|b=0 =
Adstx t adZ Ad(1s)tx ds
b
0
and similarly

( AdtxtaY )|a=0 =
b

Adstx t adY Ad(1s)tx ds.

0

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

242

2. Related articles

Collecting terms, our task is now to show that

(2.27)
Z 1Z 1
Z 1Z 1
Ad(1st)x adY Ad(1s)tx Z tdsdt.
Adstx adZ Ad(1s)tx Y tdsdt =
0

For any x g, the adjoint map adx : g g is a derivation in the sense that
adx [y, z] = [adx y, z] + [y, adx z],
thanks to the Jacobi identity. Exponentiating, we conclude that
Adx [y, z] = [Adx y, Adx z]
(thus each Adx is a Lie algebra homomorphism) and thus
Adx ady Ad1
x = adAdx y .
Using this, we can simplify (2.27) as
Z 1Z 1
Z
adAdstx Z Adtx Y tdsdt =
0

which we can rewrite as

Z 1Z 1
Z
[Adstx Z, Adtx Y ] tdsdt =
0

1Z 1

adAd(1st)x Y Ad(1t)x Y tdsdt

1Z 1

[Ad(1st)x Y, Ad(1t)x Y ] tdsdt.

0

But by an appropriate change of variables (and the anti-symmetry of the

Lie bracket), both sides of this equation can be written as
Z
[Adax Z, Adbx Y ] dadb
0ab1

and the claim follows.

Remark 2.4.3. The above argument shows that every finite-dimensional
Lie algebra g can be viewed as arising from a local Lie group G. It is natural
to then ask if that local Lie group (or a sufficiently small piece thereof) can in
The answer to this is affirmative,
turn be extended to a global Lie group G.
as was first shown by Cartan. I have been unable however to find a proof
of this result that does not either use Ados theorem, the proof method of
Ados theorem (in particular, the structural decomposition of Lie algebras
into semisimple and solvable factors), or some facts about group cohomology
(particularly with regards to central extensions of Lie groups) which are
closely related to the structural decompositions just mentioned. (As noted
by Serre [Se1964], though, a certain amount of this sort of difficulty in the
proof may in fact be necessary, given that the global form of Lies third
theorem is known to fail in the infinite-dimensional case.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.5. Local groups

243

2.5. Local groups

The material here is based in part on [Ol1996], [Go2010].
One of the fundamental structures in modern mathematics is that of
a group. Formally, a group is a set G = (G, 1, , ()1 ) equipped with an
identity element 1 = 1G G, a multiplication operation : G G G, and
an inversion operation ()1 : G G obeying the following axioms:
(1) (Closure) If g, h G, then g h and g 1 are well-defined and lie in
G. (This axiom is redundant from the above description, but we
include it for emphasis.)
(2) (Associativity) If g, h, k G, then (g h) k = g (h k).
(3) (Identity) If g G, then g 1 = 1 g = g.
(4) (Inverse) If g G, then g g 1 = g 1 g = 1.
One can also consider additive groups G = (G, 0, +, ) instead of multiplicative groups, with the obvious changes of notation. By convention, additive
groups are always understood to be abelian, so it is convenient to use additive notation when one wishes to emphasise the abelian nature of the group
structure. As usual, we often abbreviate g h by gh (and 1G by 1) when
there is no chance of confusion.
If furthermore G is equipped with a topology, and the group operations
, ()1 are continuous in this topology, then G is a topological group. Any
group can be made into a topological group by imposing the discrete topology, but there are many more interesting examples of topological groups,
such as Lie groups, in which G is not just a topological space, but is in fact
a smooth manifold (and the group operations are not merely continuous,
but also smooth).
There are many naturally occuring group-like objects that obey some,
but not all, of the axioms. For instance, monoids are required to obey the
closure, associativity, and identity axioms, but not the inverse axiom. If we
also drop the identity axiom, we end up with a semigroup. Groupoids do not
necessarily obey the closure axiom, but obey (versions of) the associativity,
identity, and inverse axioms. And so forth.
Another group-like concept is that of a local topological group (or local
group, for short), defined in Section 1.2.1. A prime example of a local group
can be formed by restricting any global topological group G to an open
neighbourhood U G of the identity, with the domains
:= {(g, h) U : g h U }
and
:= {g U : g 1 U };

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

244

2. Related articles

one easily verifies that this gives U the structure of a local group (which
we will sometimes call G U to emphasise the original group G). If U is
symmetric (i.e. U 1 = U ), then we in fact have a symmetric local group.
One can also restrict local groups G to open neighbourhoods U to obtain a
smaller local group G U by the same procedure (adopting the convention
that statements such as g h U or g 1 U are considered false if the
left-hand side is undefined). (Note though that if one restricts to non-open
neighbourhoods of the identity, then one usually does not get a local group;
for instance [1, 1] is not a local group (why?).)
Finite subsets of (Hausdorff) groups containing the identity can be viewed
as local groups. This point of view turns out to be particularly useful for
studying approximate groups in additive combinatorics, a point which I
hope to expound more on later. Thus, for instance, the discrete interval
{9, . . . , 9} Z is an additive symmetric local group, which informally
might model an adding machine that can only handle (signed) one-digit
numbers. More generally, one can view a local group as an object that behaves like a group near the identity, but for which the group laws (and in
particular, the closure axiom) can start breaking down once one moves far
enough away from the identity.
In many situations (such as when one is investigating the local structure
of a global group) one is only interested in the local properties of a (local or
global) group. We can formalise this by the following definition. Let us call
two local groups G = (G, , , 1G , , ()1 ) and G0 = (G0 , 0 , 0 , 1G0 , , ()1 )
locally identical if they have a common restriction, thus there exists a set
U G G0 such that G U = G0 U (thus, 1G = 1G0 , and the topology
and group operations of G and G0 agree on U ). This is easily seen to be
an equivalence relation. We call an equivalence class [G] of local groups a
group germ.
Let P be a property of a local group (e.g. abelianness, connectedness,
compactness, etc.). We call a group germ locally P if every local group in
that germ has a restriction that obeys P; we call a local or global group G
locally P if its germ is locally P (or equivalently, every open neighbourhood
of the identity in G contains a further neighbourhood that obeys P). Thus,
the study of local properties of (local or global) groups is subsumed by the
study of group germs.
Exercise 2.5.1.
(i) Show that the above general definition is consistent with the usual
definitions of the properties connected and locally connected
from point-set topology.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.5. Local groups

245

(ii) Strictly speaking, the above definition is not consistent with the
usual definitions of the properties compact and local compact
from point-set topology because in the definition of local compactness, the compact neighbourhoods are certainly not required to be
open. Show however that the point-set topology notion of locally
compact is equivalent, using the above conventions, to the notion
of locally precompact inside of an ambient local group. Of course,
this is a much more clumsy terminology, and so we shall abuse notation slightly and continue to use the standard terminology locally
compact even though it is, strictly speaking, not compatible with
the above general convention.
(iii) Show that a local group is locally discrete if and only if it is locally
trivial.
(iv) Show that a connected global group is abelian if and only if it is
locally abelian. (Hint: in a connected global group, the only open
subgroup is the whole group.)
(v) Show that a global topological group is first-countable if and only
if it is locally first countable. (By the Birkhoff-Kakutani theorem
(Theorem 1.5.2), this implies that such groups are metrisable if and
only if they are locally metrisable.)
Let p be a prime. Show that the solenoid group Zp R/Z , where
Zp is the p-adic integers and Z := {(n, n) : n Z} is the diagonal embedding of Z inside Zp R, is connected but not locally
connected.
Remark 2.5.1. One can also study the local properties of groups using
nonstandard analysis. Instead of group germs, one works (at least in the
case when G is first countable) with the monad o(G) of the identity element
1G of G, defined as the nonstandard group elements g = limn gn in G
that are infinitesimally close to the origin in the sense that they lie in every
standard neighbourhood of the identity. The monad o(G) is closely related
to the group germ [G], but has the advantage of being a genuine (global)
group, as opposed to an equivalence class of local groups. It is possible
to recast most of the results here in this nonstandard formulation; see e.g.
[Ro1966]. However, we will not adopt this perspective here.
A useful fact to know is that Lie structure is local; see Exercise 1.2.17.
As with so many other basic classes of objects in mathematics, it is of
fundamental importance to specify and study the morphisms between local
groups (and group germs). Given two local groups G, G0 , we can define the
notion of a (continuous) homomorphism : G G0 between them, defined

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

246

2. Related articles

as a continuous map with

(1G ) = 1G0
such that whenever g, h G are such that gh is well-defined, then (g)(h)
is well-defined and equal to (gh); similarly, whenever g G is such that
g 1 is well-defined, then (g)1 is well-defined and equal to (g 1 ). (In
abstract algebra, the continuity requirement is omitted from the definition
of a homomorphism; we will call such maps discrete homomorphisms to
distinguish them from the continuous ones which will be the ones studied
here.)
It is often more convenient to work locally: define a local (continuous)
homomorphism : U G0 from G to G0 to be a homomorphism from an
open neighbourhood U of the identity to G0 . Given two local homomor G
0 from one pair of locally identical groups
phisms : U G0 , : U
to another pair G0 , G
0 , we say that , 0 are locally identical if they
G, G
0 (note that it
agree on some open neighbourhood of the identity in U U
or both). An
does not matter here whether we require openness in G, in G,
equivalence class [] of local homomorphisms will be called a germ homomorphism (or morphism for short) from the group germ [G] to the group
germ [G0 ].
Exercise 2.5.2. Show that the class of group germs, equipped with the germ
homomorphisms, becomes a category. (Strictly speaking, because group
germs are themselves classes rather than sets, the collection of all group
germs is a second-order class rather than a class, but this set-theoretic technicality can be resolved in a number of ways (e.g. by restricting all global
and local groups under consideration to some fixed universe) and should
be ignored for this exercise.)
As is usual in category theory, once we have a notion of a morphism, we
have a notion of an isomorphism: two group germs [G], [G0 ] are isomorphic
if there are germ homomorphisms : [G] [G0 ], : [G0 ] [G] that invert
each other. Lifting back to local groups, the associated notion is that of
local isomorphism: two local groups G, G0 are locally isomorphic if there
exist local isomorphisms : U G0 and : U 0 G from G to G0 and
from G0 to G that locally invert each other, thus ((g)) = g for g G
sufficiently close to 1G , and ((g)) for g 0 G0 sufficiently close to 1G0 .
Note that all local properties of (global or local) groups that can be defined
purely in terms of the group and topological structures will be preserved
under local isomorphism. Thus, for instance, if G, G0 are locally isomorphic
local groups, then G is locally connected iff G0 is, G is locally compact iff
G0 is, and (by Exercise 1.2.17) G is Lie iff G0 is.
Exercise 2.5.3.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.5. Local groups

247

Show that the additive global groups R/Z and R are locally isomorphic.
Show that every locally path-connected group G is locally isomorphic to a
path-connected, simply connected group.
2.5.1. Lies third theorem. Lies fundamental theorems of Lie theory
link the Lie group germs to Lie algebras. Observe that if [G] is a locally Lie
group germ, then the tangent space g := T1 G at the identity of this germ is
well-defined, and is a finite-dimensional vector space. If we choose G to be
symmetric, then g can also be identified with the left-invariant (say) vector
fields on G, which are first-order differential operators on C (M ). The Lie
bracket for vector fields then endows g with the structure of a Lie algebra. It
is easy to check that every morphism : [G] [H] of locally Lie germs gives
rise (via the derivative map at the identity) to a morphism D(1) : g h
of the associated Lie algebras. From the Baker-Campbell-Hausdorff formula
(which is valid for local Lie groups, as discussed in this previous post) we
conversely see that D(1) uniquely determines the germ homomorphism .
Thus the derivative map provides a covariant functor from the category of
locally Lie group germs to the category of (finite-dimensional) Lie algebras.
In fact, this functor is an isomorphism:
Exercise 2.5.4 (Lies third theorem). For this exercise, all Lie algebras are
understood to be finite dimensional (and over the reals).
(i) Show that every Lie algebra g is the Lie algebra of a local Lie group
germ [G], which is unique up to germ isomorphism (fixing g).
(ii) Show that every Lie algebra g is the Lie algebra of some global
connected, simply connected Lie group G, which is unique up to
Lie group isomorphism (fixing g).
(iii) Show that every homomorphism : g h between Lie algebras
is the derivative of a unique germ homomorphism : [G] [H]
between the associated local Lie group germs.
(iv) Show that every homomorphism : g h between Lie algebras
is the derivative of a unique Lie group homomorphism : G
H between the associated global connected, simply connected, Lie
groups.
(v) Show that every local Lie group germ is the germ of a global connected, simply connected Lie group G, which is unique up to Lie
group isomorphism. In particular, every local Lie group is locally
isomorphic to a global Lie group.
(Hint: use Exercise 1.2.23.)
Lies third theorem (which, actually, was proven in full generality by
Cartan) demonstrates the equivalence of three categories: the category of

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

248

2. Related articles

finite-dimensonal Lie algebras, the category of local Lie group germs, and
the category of connected, simply connected Lie groups.
2.5.2. Globalising a local group. Many properties of a local group improve after passing to a smaller neighbourhood of the identity. Here are
some simple examples:
Exercise 2.5.5. Let G be a local group.
(i) Give an example to show that G does not necessarily obey the
cancellation laws
(2.28)

gk = hk = g = h;

kg = kh = g = h

for g, h, k G (with the convention that statements such as gk =

hk are false if either side is undefined). However, show that there
exists an open neighbourhood U of G within which the cancellation
law holds.
(ii) Repeat the previous part, but with the cancellation law (2.28) replaced by the inversion law
(2.29)

(gh)1 = h1 g 1
for any g, h G for which both sides are well-defined.

(iii) Repeat the previous part, but with the inversion law replaced by
the involution law
(2.30)

(g 1 )1 = g
for any g for which the left-hand side is well-defined.

Note that the counterexamples in the above exercise demonstrate that

not every local group is the restriction of a global group, because global
groups (and hence, their restrictions) always obey the cancellation law (2.28),
the inversion law (2.29), and the involution law (2.30). Another way in which
a local group can fail to come from a global group is if it contains relations
which can interact in a global way to cause trouble, in a fashion which
is invisible at the local level. For instance, consider the open unit cube
(1, 1)3 , and consider four points a1 , a2 , a3 , a4 in this cube that are close to
the upper four corners (1, 1, 1), (1, 1, 1), (1, 1, 1), (1, 1, 1) of this cube
respectively. Define an equivalence relation on this cube by setting x y
if x, y (1, 1)3 and x y is equal to either 0 or 2ai for some i = 1, . . . , 4.
Note that this indeed an equivalence relation if a1 , a2 , a3 , a4 are close enough
to the corners (as this forces all non-trivial combinations 2ai 2aj to lie
outside the doubled cube (2, 2)3 ). The quotient space (1, 1)3 / (which
is a cube with bits around opposite corners identified together) can then
be seen to be a symmetric additive local Lie group, but will usually not

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.5. Local groups

249

come from a global group. Indeed, it is not hard to see that if (1, 1)3 /
is the restriction of a global group G, then G must be a Lie group with
Lie algebra R3 (by Exercise 1.2.17), and so the connected component G
of G containing the identity is isomorphic to R3 / for some sublattice of
R3 that contains a1 , a2 , a3 , a4 ; but for generic a1 , a2 , a3 , a4 , there is no such
lattice, as the ai will generate a dense subset of R3 . (The situation here is
somewhat analogous to a number of famous Escher prints, such as Ascending and Descending, in which the geometry is locally consistent but globally
inconsistent.) We will give this sort of argument in more detail later, when
we prove Proposition 2.5.6.
Nevertheless, the space (1, 1)3 / is still locally isomorphic to a global
Lie group, namely R3 ; for instance, the open neighbourhood (0.5, 0.5)3 /
is isomorphic to (0.5, 0.5)3 , which is an open neighbourhood of R3 . More
generally, Lies third theorem tells us that any local Lie group is locally
isomorphic to a global Lie group.
Let us call a local group globalisable if it is locally isomorphic to a
global group; thus Lies third theorem tells us that every local Lie group
is globalisable. Thanks to Goldbrings solution [Go2010] to the local version of Hilberts fifth problem (Theorem 1.5.24), we also know that locally
Euclidean local groups are globalisable. A modification of this argument
[vdDrGo2010] shows in fact that every locally compact local group is globalisable.
In view of these results, it is tempting to conjecture that all local groups
are globalisable;; among other things, this would simplify the proof of Lies
third theorem (and of the local version of Hilberts fifth problem). Unfortunately, this claim as stated is false:
Theorem 2.5.2. There exists local groups G which are not globalisable.
The counterexamples used to establish Theorem 2.5.2 are remarkably
delicate; the first example I know of appears in [vEsKo1964]. One reason
for this, of course, is that the previous results prevents one from using any
local Lie group, or even a locally compact group as a counterexample. We
will present a (somewhat complicated) example shortly, based on the unit
ball in the infinite-dimensional Banach space ` (N2 ).
Despite such counterexamples, there are certainly many situations in
which we can globalise a local group. For instance, this is the case if one has
a locally faithful representation of that local group inside a global group:
Lemma 2.5.3 (Faithful representation implies globalisability). Let G be
a local group, and suppose there exists an injective local homomorphism
: U H from G into a global topological group H with U symmetric.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

250

2. Related articles

Then U is isomorphic to the restriction of a global topological group to an

open neighbourhood of the identity; in particular, G is globalisable.
We now prove Lemma 2.5.3. Let G, , U, H be as in that lemma. The
set (U ) generates a subgroup h(U )i of H, which contains an embedded
copy (U ) of U . It is then tempting to restrict the topology of H to that of
h(U )i to give h(U )i the structure of a global topological group and then
declare victory, but the difficulty is that (U ) need not be an open subset
of h(U )i, as the following key example demonstrates.
2
Example 2.5.4. Take G = U = (1, 1), H = (R/Z)
, and (t) :=
2
(t, t) mod Z , where is an irrational number (e.g. = 2). Then h(U )i
is the dense subgroup {(t, t) mod Z2 : t R} of (R/Z)2 , which is not locally isomorphic to G if endowed with the topology inherited from (R/Z)2
(for instance, h(U )i is not locally connected in this topology, whereas G
is). Also, (U ), while homeomorphic to U , is not an open subset of h(U )i.
Thus we see that the global behaviour of (U ), as captured by the group
h(U )i, can be rather different from the local structure of (U ).

However, the problem can be easily resolved by giving h(U )i a different

topology, as follows. We use the sets {(W ) : 1 W U, W open} as
a neighbourhood base for the identity in h(U )i, and their left-translates
{g(W ) : 1 W U, W open} as a neighbourhood base for any other
element g of h(U )i. This is easily seen to generate a topology. To show that
the group operations remain continuous in this topology, the main task is to
show that the conjugation operations x 7 gxg 1 are continuous with respect
to the neighbourhood base at the identity, in the sense that for every open
neighbourhood W of the identity in U and every g h(U )i, there exists an
open neighbourhood W 0 of the identity such that (W 0 ) g(W )g 1 . But
for g (U ) this is clear from the injective local homomorphism properties
of (after shrinking W small enough that g(W )g 1 will still fall in (U )),
and then an induction shows the same is true for g in any product set (U )n
of (U ), and hence in all of h(U )i. (It is instructive to follow through this
argument for the example given above.)
There is another characterisation of globalisability, due to Malcev [Ma1941],
which is stronger than Lemma 2.5.3, but this strengthening is usually not
needed in applications. Call a local group G globally associative if, whenever g1 , . . . , gn G and there are two ways to associate the product g1 . . . gn
which are individually well-defined, then the value obtained by these two
associations are equal to each other. This implies but is stronger than local
associativity (which only covers the cases n 3).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.5. Local groups

251

Proposition 2.5.5 (Globalisation criterion). Let G be a symmetric local

group. Then G is isomorphic to (a restriction of ) an open symmetric neighbourhood of the identity in a global topological group if and only if it is
globally associative.
By restriction of an open symmetric neighbourhood of the identity U ,
I mean the local group formed from U by restricting the set U U of
admissible products for the local group law to some open neighbourhood of
{1} U U {1} in U U .
Proof. The only if direction is clear, so now suppose that G is a globally
associative symmetric local group. Let us call a formal product g1 . . . gn with
g1 , . . . , gn G weakly well-defined if there is at least one way to associate
the product so that it can be defined in G (this is opposed to actual welldefinedness of g1 . . . gn , which requires all associations to be well-defined).
By global associativity, the product g1 . . . gn has a unique evaluation in G
whenever it is weakly well-defined.
Let F = (F, ) be the (discrete) free group generated by the elements
of G (now viewed merely as a discrete set), thus each element of F can be
expressed as a formal product g11 . . . gnn of elements g1 , . . . , gn in G
and their formal inverses g11 , . . . , gn1 , where 1 , . . . , n {1, +1}, and
G can be viewed (only as a set, not as a local group) as a subset of F . Let
N be the set of elements in F that have at least one representation (not
necessarily reduced) of the form g11 . . . gnn such that g1 , . . . , gn G and
1 , . . . , n {1, +1} with g11 . . . gnn weakly well-defined and evaluating
to the identity in G. It is easy to see that N is a normal subgroup of
F , and so we may form the quotient group F/N and the quotient map
: F F/N . We claim is injective on G, and so G is isomorphic (as
a discrete local group) to (G). To see this, suppose for contradiction that
there are distinct g, h G such that (g) = h, thus g g11 . . . gnn = h
for some g11 . . . gnn N . As this identity takes place in the free group
F , this implies that the formal word g g11 . . . gnn can be reduced to the
generator h of G by a finite number of operations in which an adjacent pair
of the form k k 1 or k 1 k for some k G, or a singleton of the form 1,
is deleted. In particular, this implies that gg11 . . . gnn is weakly well-defined
and evaluates to h in G. On other hand, as g11 . . . gnn N , we also
see (by associating in a different way) that gg11 . . . gnn is weakly well-defined
and evaluates to g in G. This contradicts global associativity, and the claim
follows.
To conclude the proof, we need to place a topology on F/N that makes
G homeomorphic to G. This can be done by taking the sets x(W ), with
W an open neighbourhood of the identity in G, as a neighbourhood base for

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

252

2. Related articles

each x F/N . By arguing as in the proof of Lemma 2.5.3 one can verify
that this generates a topology that makes F/N a topological group, and
makes G homeomorphic to (G); we leave the details as an exercise.

Exercise 2.5.6. Complete the proof of the above proposition.
Exercise 2.5.7. Use Proposition 2.5.5 to give an alternate proof of Proposition 2.5.3.
2.5.3. A non-globalisable group. We now prove Theorem 2.5.2. We
begin with a preliminary construction, which gives a local group that has a
fixed small neighbourhood that cannot arise from a global group.
Proposition 2.5.6 (Preliminary counterexample). For any m 1, there
exists an equivalence relation m on the open unit ball B(0, 1) of ` (N)
which gives B(0, 1)/ m the structure of a local group, but such that B(0, 1/m)/ m
is not isomorphic (even as a discrete local group) to a subset of a global
group.
Proof. We will use a probabilistic construction, mimicking the three-dimensional
example (1, 1)3 / from the introduction. Fix m, and let N be a large
integer (depending on m) to be chosen later. We identify the N -dimensional
cube (1, 1)N with the unit ball in ` ({1, . . . , N }), which embeds in the unit
ball in ` (N) via extension by zero. Let b1 , . . . , bN +1 {1, 1}N be randomly chosen corners of this cube. A simple application of the union bound
shows that with probability approaching 1 as N , we have bi 6= bj for
all i < j, but also that for any 1 i1 < < i100m N and any choice
of signs 1 , . . . , 100m , the vectors 1 bi1 , . . . , 100m bi100m agree on at least one
coordinate. As a corollary of this, we see that
k

N
+1
X

ni bi k` =

N
+1
X

i=1

for any integers n1 , . . . , nN +1 with

|ni |

i=1

PN +1
i=1

|ni | 100m.

Let > 0 be a small number (depending on m, N ) to be chosen later.

We let ai be an element of (1, 1)N which is within distance of bi , then
we have
(2.31)

N
+1
X
i=1

ni ai k` =

N
+1
X

|ni | + O()

i=1

(allowing implied constants to depend on m, N ) for n1 , . . . , nN +1 as above.

By generically perturbing the a1 , . . . , aN +1 , we may assume that they are
noncommensurable in the sense that span a dense subset of RN . In particular, the a1 , . . . , aN +1 are linearly independent over Z.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.5. Local groups

253

We now define an equivalence relation m on B(0, 1) by defining f m g

whenever
f g =

(2.32)

N +1
1 X
ni ai
2m
i=1

PN +1

for some integers n1 , . . . , nN with i=1 |ni | 100m. Since kf gk`

2, we see from (2.31) (if is small enough) that the equation (2.32) can
P +1
only be true if N
i=1 |ni | 4m. As a consequence, we see that m is a
equivalence relation. We can then form the quotient space B(0, 1)/ m .
Observe that if f1 , f2 , g1 , g2 B(0, 1) are such that f1 m g1 , f2 m g2 ,
and f1 + f2 , g1 + g2 B(0, 1), then we also have f1 + f2 m g1 + g2 .
Thus B(0, 1)/ m has an addition operation +, defined on those equivalence
classes [f ], [g] B(0, 1)/ m for which f + g B(0, 1) for at least one
representative f, g of [f ], [g] respectively. One easily verifies that this gives
B(0, 1)/ m the structure of a local group.
Now suppose for sake of contradiction that B(0, 1/m)/ m is isomorphic
1
to a restriction of a global topological group G. Since 0 m
ai for all
i = 1, . . . , N + 1, we thus have a map : B(0, 1/m) G which annihilates
1
all of the 2m
ai , and is locally additive in the sense that (f +g) = (f )(g)
whenever f, g, f + g B(0, 1/m). In particular, we see that all the elements
of (B(0, 1/2m)) commute with each other. Furthermore the kernel {f
B(0, 1/m) : (f ) = 1} is precisely equal to the set
(
)
N +1
N
+1
X
1 X
ni ai :
|ni | 4m B(0, 1/m).
(2.33)
2m
i=1

i=1

As the a1 , . . . , aN +1 span a dense subset of Rn , we can find integers

n1 , . . . , nN +1 such that
n1 a1 + + nN +1 aN +1 B(0, 1)
and
N
+1
X

|ni | > 4m.

i=1
1
We claim that 2m
(n1 a1 + + nN +1 aN +1 ) lies in the kernel of , contradicting the description (2.33) of that description (and the linear independence of the ai ). To see this, we observe for a sufficiently large natural
number M that the local homomorphism property (and the commutativity)
gives

 NY
nN +1
+1 
1
1

(n1 a1 + + nN +1 aN +1 ) =

aN +1
2mM
2mM
i=1

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

254

2. Related articles

and hence (by further application of local homomorphism and commutativity)


M nN +1
 NY
+1 
1
1

.
(n1 a1 + + nN +1 aN +1 ) =
aN +1

2m
2mM
i=1

But by yet more application of the local homomorphism property, we have

M



1
1

aN +1
=
aN +1 = 1,
2mM
2m
and the claim follows.

Now we glue together the examples in Proposition 2.5.6 to establish

Theorem 2.5.2. We work in the space ` (N2 ), the elements of which we
can think of as a sequence (fm )mN of uniformly bounded functions fm
` (N). The unit ball in this space can then be identified (as a set) with
the product B(0, 1)N , where B(0, 1) is the unit ball in ` (N), though we
caution that the topology on B(0, 1)N is not given by the product topology
(or the box topology).
We can combine the equivalence relations m on B(0, 1) to a relation
on B(0, 1)N , defined by setting (fm )mN (gm )mN iff fm m gm for
all m N. This is clearly an equivalence relation on B(0, 1)N , and so we
can create the quotient space B(0, 1)N / with the quotient topology. One
easily verifies that this gives a local group. The sets B(0, 1/m)N / form
a neighbourhood base of the identity, but none of these sets is isomorphic
(even as a discrete local group) to a subset of a global group, as it contains
a copy of B(0, 1/m)/ m , and the claim follows.

2.6. Central extensions of Lie groups, and cocycle averaging

The theory of Hilberts fifth problem sprawls across many subfields of mathematics: Lie theory, representation theory, group theory, nonabelian Fourier
analysis, point-set topology, and even a little bit of group cohomology. The
latter aspect of this theory is what this section will be focused on. The general question that comes into play here is the extension problem: given two
(topological or Lie) groups H and K, what is the structure of the possible
groups G that are formed by extending H by K. In other words, given a
short exact sequence
0 K G H 0,
to what extent is the structure of G determined by that of H and K?
As an example of why understanding the extension problem would help
in structural theory, let us consider the task of classifying the structure of

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.6. Central extensions of Lie groups, and cocycle averaging

255

a Lie group G. Firstly, we factor out the connected component G of the

identity as
0 G G G/G 0;
as Lie groups are locally connected, G/G is discrete. Thus, to understand
general Lie groups, it suffices to understand the extensions of discrete groups
by connected Lie groups.
Next, to study a connected Lie group G, we can consider the conjugation
action g : X 7 gXg 1 on the Lie algebra g, which gives the adjoint representation Ad : G GL(g). The kernel of this representation consists of all
the group elements g that commute with all elements of the Lie algebra, and
thus (by connectedness) is the center Z(G) of G. The adjoint representation
is then faithful on the quotient G/Z(G). The short exact sequence
0 Z(G) G G/Z(G) 0
then describes G as a central extension (by the abelian Lie group Z(G)) of
G/Z(G), which is a connected Lie group with a faithful finite-dimensional
linear representation.
This suggests a route to Hilberts fifth problem, at least in the case of
connected groups G. Let G be a connected locally compact group that we
hope to demonstrate is isomorphic to a Lie group. As discussed in Section
1.3, we first form the space L(G) of one-parameter subgroups of G (which
should, eventually, become the Lie algebra of G). Hopefully, L(G) has the
structure of a vector space. The group G acts on L(G) by conjugation; this
action should be both continuous and linear, giving an adjoint representation Ad : G GL(L(G)). The kernel of this representation should then
be the center Z(G) of G. The quotient G/Z(G) is locally compact and has
a faithful linear representation, and is thus a Lie group by von Neumanns
theorem (Theorem 1.3.4). The group Z(G) is locally compact abelian, and
so it should be a relatively easy task to establish that it is also a Lie group.
To finish the job, one needs the following result:
Theorem 2.6.1 (Central extensions of Lie are Lie). Let G be a locally
compact group which is a central extension of a Lie group H by an abelian
Lie group K. Then G is also isomorphic to a Lie group.
This result can be obtained by combining a result of Kuranishi [Ku1950]
with a result of Gleason [Gl1950]. It is superceded by the subsequent theory of Hilberts fifth problem (cf. Exercise 2.7.1 below), but can be proven
without using such machinery. The point here is that while G is initially
only a topological group, the smooth structures of H and K can be combined (after a little bit of cohomology) to create the smooth structure on
G required to upgrade G from a topological group to a Lie group. One of
the main ideas here is to improve the behaviour of a cocycle by averaging

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

256

2. Related articles

it; this basic trick is helpful elsewhere in the theory, resolving a number of
cohomological issues in topological group theory. The result can be generalised to show in fact that arbitrary (topological) extensions of Lie groups
by Lie groups remain Lie; this was shown in [Gl1950]. However, the above
special case of this result is already sufficient (in conjunction with the rest
of the theory, of course) to resolve Hilberts fifth problem.

Remark 2.6.2. We have shown in the above discussion that every connected Lie group is a central extension (by an abelian Lie group) of a Lie
group with a faithful continuous linear representation. It is natural to ask
whether this central extension is necessary. Unfortunately, not every connected Lie group admits a faithful continuous linear representation; see Exercise 1.1.5. (On the other hand, the group G in that example is certainly
isomorphic to the extension of the linear group R2 by the abelian group
R/Z.)

2.6.1. A little group cohomology. Let us first ignore the topological or

Lie structure, consider the (central) extension problem for discrete groups
only. Thus, let us suppose we have a (discrete) group G = (G, ) which is
a central extension of a group H by a group K. We view K as a central
subgroup of G (which we write additively to emphasise its abelian nature),
and use : G H to denote the projection map. If k K and g G, we
write g + k for kg = gk to emphasise the central nature of K.
It may help to view G as a principal K-bundle over H, with K being
thought of as the vertical component of G and H as the horizontal
component. Thus G is the union of vertical fibres 1 (h), h H indexed
by the horizontal group H, each of which is a coset (or a torsor ) of the
vertical group H.
As central extensions are not unique, we will need to specify some additional data beyond H and K. One way to view this data is to specify a
section of the extension, that is to say a map : H G that is a rightinverse for the projection map , thus selects one element (h) from each
fibre 1 (h) of the projection map (which is also a coset of K). Such a section can be always chosen using the axiom of choice, though of course there
is no guarantee of any measurability, continuity, or smoothness properties
of such a section if the groups involved have the relevant measure-theoretic,
topological, or smooth structure.
Note that we do not necessarily require the section to map the goup
identity 1H of H to the group identity 1G of G, though in practice it is
usually not difficult to impose this constraint if desired.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.6. Central extensions of Lie groups, and cocycle averaging

257

The group G can then be described, as a set, as the disjoint union of its
fibres (h) + K for h H:
]
G=
(h) + K.
hH

Thus each element g G can be uniquely expressed as (h) + k for some

h H and k K, and can thus be viewed as a system of coordinates of G
(identifying it as a set with H K). Now we turn to the group operations
()1 : G G and : G G G. If h1 , h2 H, then the product (h1 )(h2 )
must lie in the fibre of h1 h2 , and so we have
(h1 )(h2 ) = (h1 h2 ) + (h1 , h2 )
for some function : H H K. By the centrality of K, this describes
the product law for general elements of G in the H K coordinates:
(2.34)

((h1 ) + k1 )((h2 ) + k2 ) = (h1 h2 ) + (k1 + k2 + (h1 , h2 )).

Using the associative law for G, we see that must obey the cocycle equation
(2.35)

(h1 , h2 ) + (h1 h2 , h3 ) = (h1 , h2 h3 ) + (h2 , h3 )

for all h1 , h2 , h3 K; we refer to functions that obey this equation as cocycles (the reason for this terminology being explained in [Ta2009, 1.13]).
2 (H, K), if we
The space of all such cocyles is denoted Z 2 (H, K) (or Zdisc
wish to emphasise that we are working for now in the discrete category,
as opposed to the measurable, topological, or smooth category); this is an
abelian group with respect to pointwise addition. Using (2.34), we also have
a description of the group identity 1G of G in coordinates:
(2.36)

1G = (1H ) (1H , 1H ).

Indeed, this can be seen by noting that 1G is the unique solution of the
group equation g g = g. Similarly, we can describe the inverse operation in
H K coordinates:
(2.37)

((h) + k)1 = (h1 ) + (k (h, h1 ) (1H , 1H )).

Thus we see that the cocycle , together with the group structures on
H and K, capture all the group-theoretic structure of G. Conversely, given
any cocycle Z 2 (H, K), one can place a group structure on the set H K
by declaring the multiplication law as
(h1 , k1 ) (h2 , k2 ) := (h1 h2 , k1 + k2 + (h1 , h2 )),
the identity element as (1H , (1H , 1H )), and the inversion law as


(h, k)1 := h1 , k (h, h1 ) (1H , 1H ) .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

258

2. Related articles

Exercise 2.6.1. Using the cocycle equation (2.35), show that the above
operations do indeed yield group structure on H K (i.e. the group axioms
are obeyed). If arises as the cocycle associated to a section of a group
extension G, we then see from (2.34), (2.36), (2.37) that this group structure
we have just placed on H K is isomorphic to that on G.
Thus we see that once we select a section, we can describe a central
extension of K by H (up to group isomorphism) as a cocycle, and conversely
every cocycle arises in this manner. However, we have some freedom in
deciding how to select this section. Given one section : H G, any other
section 0 : H G takes the form
0 (h) = (h) + f (h)
for some function f : H K; conversely, every such function f can be used
to shift a section to a new section 0 . We refer to such functions f as
gauge functions. The cocycles , 0 associated to , 0 are related by the
gauge transformation
0 (h1 , h2 ) = (h1 , h2 ) + df (h1 , h2 )
where df : H H K is the function
df (h1 , h2 ) := f (h1 ) + f (h2 ) f (h1 h2 ).
We refer to functions of the form df as coboundaries, and denote the space
2 (H, K), if we want to emphaof all such coboundaries as B 2 (H, K) (or Bdisc
sise the discrete nature of these coboundaries). One easily verifies that all
coboundaries are cocycles, and so B 2 (H, K) is a subgroup of Z 2 (H, K). We
2 (H, K)) to be
then define the second group cohomology H 2 (H, K) (or Hdisc
the quotient group
H 2 (H, K) := Z 2 (H, K)/B 2 (H, K),
and refer to elements of H 2 (H, K) as cohomology classes. (There are higher
order group cohomologies, which also have some relevance for the extension
problem, but will not be needed here; see [Ta2009, 1.13] for further discussion. The first group cohomology H 1 (H, K) = Hom(H, K) is just the
space of homomorphisms from H to K; again, this has some relevance for
the extension problem but will not be needed here.)
We call two cocycles cohomologous if they differ by a coboundary (i.e.
they lie in the same cohomology class). Thus we see that different sections of
a single central group extension provide cohomologous cocycles. Conversely,
if two cocycles are cohomologous, then the group structures on H K given
by these cocycles are easily seen to be isomorphic; furthermore, if we restrict
group isomorphism to fix each fibre of , this is the only way in which group
structures generated by such cocycles are isomorphic. Thus, we see that up

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.6. Central extensions of Lie groups, and cocycle averaging

259

to group isomorphism, central group extensions are described by cohomology

classes.
Remark 2.6.3. All constant cocycles are coboundaries, and so if desired
one can always normalise a cocycle : H H K (up to coboundaries) so
that (1H , 1H ) = 0. This can lead to some minor simplifications to some of
the cocycle formulae (such as (2.36), (2.37)).
The trivial cohomology class of course contains the trivial cocycle, which
in turn generates the direct product H K. Non-trivial cohomology classes
generate skew products that will not be isomorphic to the direct product
H K (at least if we insist on fixing each fibre). In particular, we see that
if the second group cohomology H 2 (H, K) is trivial, then the only central
extensions of H by K are (up to isomorphism) the direct product extension.
In general, we do not expect the cohomology group to be trivial: not
every cocycle is a coboundary. A simple example is provided by viewing
the integer group Z as a central extension of a cyclic group Z/N Z by the
subgroup N Z using the short exact sequence
0 N Z Z Z/N Z 0.
Clearly Z is not isomorphic to N ZZ/N Z (the former group is torsion-free,
while the latter group is not), and so H 2 (Z/N Z, N Z) is non-trivial. If we use
the section : Z/N Z Z that maps i mod N to i for i = 0, . . . , N 1, then
the associated cocycle (i mod N, j mod N ) is the familiar carry bit one
learns about in primary school, which equals N when i, j = 0, 1, . . . , N 1
and i + j N , but vanishes for other i, j = 0, 1, . . . , N 1.
However, when the vertical group K is sufficiently Euclidean in
nature, one can start deploying the method of averaging to improve the
behaviour of cocycles. Here is a simple example of the averaging method in
action:
Lemma 2.6.4. Let H be a finite group, and let K be a real vector space.
Then H 2 (H, K) is trivial.
Proof. We need to show that every cocycle is a coboundary. Accordingly,
let Z 2 (H, K) be a cocycle, thus
(h1 , h2 ) + (h1 h2 , h3 ) = (h1 , h2 h3 ) + (h2 , h3 )
for all h1 , h2 , h3 H. We sum this equation over all h3 H, and then
divide by the cardinality |H| of H; note that this averaging operation relies
on the finite nature of H and the vector space nature of K. We obtain as a
consequence
(h1 , h2 ) + f (h1 h2 ) = f (h1 ) + f (h2 )

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

260

2. Related articles

where
f (h) := Eh3 H (h, h3 ) =

1 X
(h, h3 )
|H|
h3 H

is the averaging of in the second variable. We thus have = df , and so

is a coboundary as desired.

It is instructive to compare this argument against the non-triviality of
mentioned earlier. While Z/N Z is still finite, the problem
here is that N Z is not a vector space and so one cannot divide by N to
straighten the cocycle. However, the averaging argument can still achieve
some simplification to such cocycles:
H 2 (Z/N Z, N Z)

Exercise 2.6.2. Let H be a finite group, and let Z 2 (H, Z) is a cocycle.

Show that is cohomologous to a cocycle taking values in {1, 0, +1}.
(Hint: average the cocycle as before using the reals R, then round to the
nearest integer.)
Now we turn from discrete group theory to topological group theory.
Now H, K, G are required to be topological groups rather than discrete
groups, with the projection map : G H continuous. (In particular, if
H is Hausdorff, then K must be closed.) In this case, it is no longer the
2 (H, K) gives rise to a topological
case that every (discrete) cocycle Zdisc
group, because the group structure on H K given by need not be continuous with respect to the product topology of H K. However, if the cocycle
is continuous, then it is clear that the group operations will be continuous,
and so we will in fact generate a topological group. Conversely, if we have a
section : H G which is continuous, then it is not difficult to verify that
the cocycle generated is also continuous. This leads to a slightly different
2 (H, K) of continuous cocycles, and
group cohomology, using the space Ztop
2
also the space Btop (H, K) of coboundaries df arising from continuous gauge
functions f .
However, such global cohomology contains some nontrivial global topological obstructions that limit its usefulness for the extension problem. One
particularly fundamental such obstruction is that one does not expect global
continuous sections : H G of a group extension to exist in general, unless H or K have a particularly simple topology (e.g. if they are contractible
or simply connected). For instance, with the short exact sequence
0 Z R R/Z 0
there is no way to continuously lift the horizontal group R/Z back up to the
real line R, due to the presence of nontrivial monodromy. While these global
obstructions are quite interesting from an algebraic topology perspective,
they are not of central importance in the theory surrounding Hilberts fifth

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.6. Central extensions of Lie groups, and cocycle averaging

261

problem, which is instead more concerned with the local topological and Lie
structure of groups.
Because of this, we shall work with local sections and cocycles instead
of global ones. A local section in a topological group extension
0KGH0
is a continuous map : U G from an open neighbourhood U of the
identity 1H in H to G which is a right inverse of on U , thus (h) 1 (h)
for all h U . (It is not yet obvious why local sections exist at all - there
may still be local obstructions to trivialising the fibre bundle - but certainly
this task should be easier than that of locating global continuous sections.)
Similarly, a local cocycle is a continuous map : U U K defined
on some open neighbourhood U of the identity in H that obeys the cocycle
equation (2.35) whenever h1 , h2 , h3 U are such that h1 h2 , h2 h3 U . We
consider two local sections : U G, 0 : U 0 G to be locally identical if
there exists a neighbourhood U 00 of the identity contained in both U and U 0
such that and 0 agree on U 00 . Similarly, we consider two local cocycles
: U U G and 0 : U 0 U 0 G to be locally identical if there is a
neighbourhood U 00 of the identity in U U 0 such that and 0 agree on
2
U 00 U 00 . We let Ztop,loc
(H, K) be the space of local cocycles, modulo local
identity; this is easily seen to be an abelian group.
One easily verifies that every local section induces a local cocycle (perhaps after shrinking the open neighbourhood U slightly), which is welldefined up to local identity. Conversely, the computations used to show
Exercise 2.6.1 show that every local cocycle creates a local group structure
on H K.
If U is an open neighbourhood the identity of H, f : U K is a continuous gauge function, and U 0 is a smaller neighbourhood such that (U 0 )2 U ,
we can define the local coboundary df : U 0 U 0 K by the usual formula
df (h1 , h2 ) := f (h1 ) + f (h2 ) f (h1 h2 ).
2
This is well-defined up to local identity. We let Btop,loc
(H, K) be the space of
2
local coboundaries, modulo local identity; this is a subgroup of Ztop,loc
(H, K).
2
Thus, one can form the local topological group cohomology Htop,loc (H, K) :=
2
2
Ztop,loc
(H, K)/Btop,loc
(H, K). We say that two local cocycles are locally cohomologous if they differ by a local coboundary.

The relevance of local cohomology to the Lie group extension problem

can be seen from the following lemma.
Lemma 2.6.5. Let G be a central (topological) group extension of a Lie
group H by a Lie group K. If there is a local section : U G whose

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

262

2. Related articles

associated local cocycle : U 0 U 0 K is locally cohomologous to a smooth

local cocycle 0 : U 00 U 00 K, then G is also isomorphic to a Lie group.
Note that the notion of smoothness of a (local) cocycle : U 0 U 0 K
makes sense, because there are smooth structures in place for both H (and
hence U 0 ) and K. In contrast, one cannot initially talk about a smooth
(local) section : U G, because G initially only has a topological structure
and not a smooth one.
Proof. By hypothesis, we locally have = 0 + df for some continuous
gauge function f : U 000 K. Rotating by f , we thus obtain another local
section 0 : U 0000 G whose associated local cocycle is smooth. We use
this section to identify 1 (U (5) ) as a topological space with U (5) K for
some sufficiently small neighbourhood U (5) of the origin. We can then give
1 (U (5) ) a smooth structure induced from the product smooth structure
on U (5) K. By (2.34), (2.37) (and (2.36)), we see that the group operations
are smooth on a neighbourhood of the identity in G.
To finish the job and give G the structure of a Lie group, we need to
extend this smooth structure to the rest of G in a manner which preserves
the smooth nature of the group operations. Firstly, by using the local coordinates of U (5) K we see that G is locally connected. Thus, if G is the
connected component of the identity, then the quotient group G/G .
Next, we extend the smooth structure on a neighbourhood of the identity on G to the rest of G by (say) left-invariance; it is easy to see that
these smooth coordinate patches are compatible with each other. A continuity argument then shows that the group operations are smooth on G .
Each element g of G acts via conjugation by a continuous homomorphism
x 7 gxg 1 on the Lie group G ; by applying Cartans theorem to the graph
of this homomorphism as in the preceding post, we see that that such homomorphisms are smooth. From this we can then conclude that the group
operations are smooth on all of G.

2.6.2. Proof of theorem. Lemma 2.6.5 suggests a strategy to prove Theorem 2.6.1: first, obtain a local section of G, extract the associated local
cocycle, then straighten it to a smooth cocycle. This will indeed be how
we proceed. In obtaining the local section and in smoothing the cocycle we
will take advantage of the averaging argument, adapted to the Lie algebra
setting.
We begin by constructing the section, following [Gl1950]. The idea
is to use the zero set of a certain continuous function F : G V from G
into a finite-dimensional vector space V . We will use topological arguments
to force this function to have a zero at every fibre, and will give F some

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.6. Central extensions of Lie groups, and cocycle averaging

263

vertical non-degeneracy to prevent it from having more than one zero on

any given fibre, at least locally.
Lets see how this works in the special case when the abelian Lie group
K is compact, so that (by the Peter-Weyl theorem, as discussed in this post)
it can be modeled as a closed subgroup of a unitary group U (n), thus we
have an injective continuous homomorphism : K U (n). We view U (n)
2
as a subset of the vector space Cn of n n complex matrices.
The group G is locally compact and also Hausdorff (being the extension
of one Hausdorff group by another). Thus, we may then apply the Tietze
2
extension theorem, and extend the map : K Cn to a continuous (and
2
compactly supported) function F0 : G Cn defined on all of G.
Note that while the original map was a homomorphism, thus
(2.38)

(k + k 0 ) = (k)(k 0 )

for all k, k 0 K, the extension F0 need not obey any analogous symmetry
in general. However, this can be rectified by an averaging argument. Define
2
the function F1 : G Cn by the formula
Z
F1 (g) :=
(k 0 )1 F0 (g + k 0 ) dK (k 0 ),
K

where K is normalised Haar measure on K. Then F1 is still continuous

and still extends ; but, unlike F0 , it now obeys the equivariance property
F1 (g + k) = (k)F1 (g)
for g G and k K, as can easily be seen from (2.38).
The function F1 is a better extension than F0 , but suffers from another
defect; while took values in the closed group (K), F1 takes values in
2
Cn (for instance, F1 is going to be compactly supported). There may well
be global topological reasons that prevent F1 from taking values in (K)
at all points, but as long as we are willing to work locally, we can repair
this issue as follows. Using the inverse function theorem, one can find a
2
manifold W Cn going through the identity matrix I, which is transverse
to (K) in the sense that every matrix M that is sufficiently close to the
identity can be uniquely (and continuously) decomposed as 1 (M )2 (M ),
where 1 (M ) (K) and 2 (M ) W . (Indeed, to build W , one can just
use a complementary subspace to the tangent space (or Lie algebra) of (K)
at the identity.). For g sufficiently close to the identity 1G , we may thus
factor
F1 (g) = F2 (g)F3 (g)
where F2 (g) (K) and F3 (g) W . By construction, F2 will be continuous
and take values in (K) near 1G , and equal on K near 1G ; it will also obey

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

264

2. Related articles

the equivariance
F2 (g + k) = (k)F2 (g)
for g G and k K that are sufficiently close to the identity. From this, we
see that for each h H that is sufficiently close to the identity, there exists
a unique solution (h) in 1 (h) to the equation F2 (g) = I that is also close
to the identity; from the continuity of F2 we see that is continuous near
the identity and is thus a local section of G as required.
Now we argue in the general case, in which the abelian Lie group K need
not be compact. If K is connected, then it is isomorphic to a Euclidean space
quotiented by a discrete subgroup. In particular, it can be quotiented down
to a torus in a manner which is a local homeomorphism near the origin.
Embedding that torus in a unitary group U (n), we obtain an continuous
homomorphism : K U (n) which is still locally injective and has compact
image. One can then repeat the previous arguments to obtain a local section;
we omit the details.
Finally, if K is not connected, we can work with the connected component K of the identity, which is locally the same as K, and repeat the
previous argument again to obtain a local section.
Using this local section, we obtain a local cocycle : U U K; shifting
by a constant, we may assume that (1H , 1H ) = 0. Near the identity, the
abelian Lie group K can be locally identified with a Euclidean space Rn , so
(shrinking U if necessary) induces a Rn -valued local cocycle : U U
Rn .
We will smooth this cocycle by an averaging argument akin to the one
used in Lemma 2.6.4. Instead of a discrete averaging, though, we will now
use a non-trivial left-invariant Haar measure H on H (the existence of
which is guaranteed by Theorem 1.4.6). As we only have a local cocycle,
though, we will need to only average using a smooth, compactly supported
function : H R+ supported on a small Rneighbourhood U 0 of U (small
enough that (U 0 )3 U ), normalised so that H (h3 ) dH (h3 ) = 1. For any
h1 , h2 , h3 U 0 , we have the cocycle equation
1 , h2 ) + (h
1 h2 , h3 ) = (h
1 , h2 h3 ) + (h
2 , h3 );
(h
we average this against (h3 ) dH (h3 ) and conclude that
1 , h2 ) + F (h1 h2 ) = F (h1 ) + F (h2 ) + (h1 , h2 )
(h
for h1 , h2 U 0 , where F : (U 0 )2 Rn is the function
Z
h3 )(h3 ) dH (h3 )
F (h) :=
(h,
H

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.7. The Hilbert-Smith conjecture

265

and : U 0 U 0 Rn is the function

Z
1 , h2 h3 ) (h
1 , h3 ))(h3 ) dH (h3 ).
((h
(h1 , h2 ) :=
H

The function F is continuous, and so is locally cohomologous to . Using

the left-invariance of H , we can rewrite as
Z
1 , h3 )((h1 h3 ) (h3 )) dH (h3 ),
(h1 , h2 ) :=
(h
2
H

from which it becomes clear that is smooth in the h2 variable. A similar averaging argument (now against a bump function on a right-invariant
measure) then shows that is in turn locally cohomologous to another local
cocycle 0 : U 00 U 00 Rn which is now smooth in both the h1 and h2 variables. Pulling Rn back to K, we conclude that is locally cohomologous to
a smooth local cocycle, and Theorem 2.6.1 now follows from Lemma 2.6.5.

2.7. The Hilbert-Smith conjecture

The classical formulation of Hilberts fifth problem asks whether topological
groups that have the topological structure of a manifold, are necessarily Lie
groups. This is indeed the case, thanks to following theorem of Gleason
[Gl1952] and Montgomery-Zippin [MoZi1952]:
Theorem 2.7.1 (Hilberts fifth problem). Let G be a topological group
which is locally Euclidean. Then G is isomorphic to a Lie group.
This theorem was proven in Section 1.6.3 There is however a generalisation of Hilberts fifth problem which remains open, namely the Hilbert-Smith
conjecture, in which it is a space acted on by the group which has the manifold structure, rather than the group itself:
Conjecture 2.7.2 (Hilbert-Smith conjecture). Let G be a locally compact
topological group which acts continuously and faithfully (or effectively) on
a connected finite-dimensional manifold X. Then G is isomorphic to a Lie
group.
Note that Conjecture 2.7.2 easily implies Theorem 2.7.1 as one can pass
to the connected component G of a locally Euclidean group (which is
clearly locally compact), and then look at the action of G on itself by
left-multiplication.
The hypothesis that the action is faithful (i.e. each non-identity group
element g G\{id} acts non-trivially on X) cannot be completely eliminated, as any group G will have a trivial action on any space X. The
requirement that G be locally compact is similarly necessary: consider for
instance the diffeomorphism group Diff(S 1 ) of, say, the unit circle S 1 , which

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

266

2. Related articles

acts on S 1 but is infinite dimensional and is not locally compact (with, say,
the uniform topology). Finally, the connectedness of X is also important:
the infinite torus G = (R/Z)N (with the product topology) acts faithfully
on the disconnected manifold X := R/Z N by the action
(gn )nN (, m) := ( + gm , m).
Note that (R/Z)N contains the p-adic group Zp as an embedded subgroup
(identifying a p-adic integer m with ( pmn mod 1)nN ), so this also gives a
faithful action of Zp on X.
The conjecture in full generality remains open. However, there are a
number of partial results. For instance, it was observed by Montgomery
and Zippin [MoZi1974] that the conjecture is true for transitive actions;
see Section 1.6.4. Another partial result is the reduction of the HilbertSmith conjecture to the p-adic case. Indeed, it is known that Conjecture
2.7.2 is equivalent to
Conjecture 2.7.3 (Hilbert-Smith conjecture for p-adic actions). It is not
possible for a p-adic group Zp to act continuously and effectively on a connected finite-dimensional manifold X.
The reduction to the p-adic case follows from the structural theory of
locally compact groups (specifically, the Gleason-Yamabe theorem, Theorem
1.1.17) and some results of Newman [Ne1931] that sharply restrict the
ability of periodic actions on a manifold X to be close to the identity. This
argument will also be given below (following the presentation in [Le1997]).
Very recently, the three-dimensional case of Conjecture 2.7.3 (and hence
the three-dimensional case of Conjecture 2.7.2) was settled in [Pa2013], by
topological methods; however, we will not discuss the proof of this conjecture
here.
2.7.1. Periodic actions of prime order. We now study periodic actions
T : X X on a manifold X of some prime order p, thus T p = id.
The basic observation to exploit here is that of rigidity: a periodic action
(or more precisely, the orbits of this action) cannot be too close to the
identity, without actually being the identity. More precisely, we have the
following theorem of Newman [Ne1931]:
Theorem 2.7.4 (Newmans first theorem). Let U be an open subset of Rn
containing the closed unit ball B, and let T : U U be a homeomorphism
of some prime period p 1. Suppose that for every x U , the orbit
{T n x : n = 0, 1, . . . , p 1} has diameter strictly less than 1. Then T (0) = 0.
Note that some result like this must be needed in order to establish
the Hilbert-Smith conjecture. Suppose for instance that one could find a

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.7. The Hilbert-Smith conjecture

267

non-trivial transformation T of some period p on the unit ball B that acts

trivially on the boundary of that ball. Then by placing infinitely many
disjoint copies of that ball into Rn , and considering maps that are equal
to some power of T on each such ball, and on the identity outside all the
balls, we can obtain a faithful action of (Z/pZ)N on Rn , contradicting the
Hilbert-Smith conjecture.
To prove the theorem we will need some basic degree theory. Given a
continuously differentiable map : B Rn , we know from Sards theorem
that almost every point x in Rn is a regular point, in that the preimage
1 ({x}) is finite and avoids the boundary B of B, with (x) being nondegenerate at each x. We can define the degree of at the regular point x
to be the number of preimages with orientation-preserving, minus the
number of preimages with orientation-reversing. One can show that
this degree extends to a constant integer-valued function on each connected
component U of Rn \(B); indeed, one can define the degree deg(U ) on
such a component analytically by the formula
Z
deg(U ) =

B

for any volume form on U of total mass 1 (one can show that this definition
is independent of the choice of ). This definition is stable under uniform
convergence of , and thus can be used to also define the degree for maps
that are merely continuous rather than continuously differentiable.
Proof. Suppose for contradiction that T (0) 6= 0. We use an averaging
argument, combined with degree theory. Let : U Rn be the map
p1

(x) :=

1X n
T x.
p
n=0

Then from construction, is continuous and T -invariant (thus (x) =

(T x) for all x U ) and we have |(x) x| < 1 for all x U . In particular,
is non-zero on the boundary B of B, and (B) is contractible to B
in Rn \0 (by taking convex combinations of and the identity). As such,
the degree of on B near the origin is equal to 1.
A similar argument shows that the homeomorphism T has degree 1 on
B near the origin, and in particular is orientation-preserving rather than
orientation reversing.
On the other hand, it can be shown that the degree of must be divisible
by p, leading to a contradiction. This is easiest to see in the case when
is continuously differentiable, for then by Sards theorem we may find a
regular point x arbitrarily close to 0. On the other hand, since T (0) is a
positive distance away from 0, we see that for x sufficiently close to 0, x is

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

268

2. Related articles

not a fixed point of T , and thus (as T has prime order) all elements in the
preimage 1 ({x}) are not fixed points of T either. Thus 1 ({x}) can be
partitioned into a finite number of disjoint orbits of T of cardinality p. As
T is orientation preserving and x is a regular point of , each of these orbits
contributes +p or p to the degree, giving the claim.
The case when is not continuously differentiable is trickier, as the
degree is not as easily computed in this case. One way to proceed is to
perturb T (or more precisely, the graph {(x, T x, . . . , T p1 x) : x U } in U p )
to be piecewise linear near the preimage of 0 (while preserving the periodicity
properties of the graph), so that degree can be computed by hand; this is
the approach taken in Newmans original paper [Ne1931]. Another is to
use the machinery of singular homology, which is more general and flexible
than degree theory; this is the approach taken in [Sm1941], [Dr1969]. 
Note that the above argument shows not only that T fixes the origin
0, but must also fix an open neighbourhood of the origin, by translating B
slightly. One can then extend this open neighbourhood to the entire space
by the following variant of Theorem 2.7.4.
Theorem 2.7.5 (Newmans second theorem). Let X be a connected manifold, and let T : X X be a homeomorphism of some prime order p that
fixes a non-empty open set U . Then T is the identity.
Proof. We need to show that T fixes all points in X, and not just U .
Suppose for sake of contradiction that T just fixes some of the points in X
and not others. By a continuity argument, and applying a homeomorphic
change of variables if necessary, we can find a coordinate chart containing
the ball B, where T fixes all x = (x1 , . . . , xn ) B with x1 0, but does
not fix all x B with x1 > 0. By shrinking B we may assume that the
entire orbit B, T B, . . . , T p1 B stays inside the coordinate chart (and can
thus be viewed as a subset of Rn ). Then the map can be defined as
before. This map is the identity on the left hemisphere {x B : x1 0}.
On the right hemisphere {x B : x1 > 0}, one observes for x small enough
that x, T x, . . . , T p1 x must all stay on the right-hemisphere (as they must
lie in B and cannot enter the left-hemisphere, where T is the identity) and
so stays on the right. This implies that has degree 1 near 0 on a small
ball around the origin, but as before one can argue that the degree must in
fact be divisible by p, leading again to a contradiction.

2.7.2. Reduction to the p-adic case. We are now ready to prove Conjecture 2.7.2 assuming Conjecture 2.7.3. Let G be a locally compact group
acting continuously and faithfully on a connected manifold X; we wish to
show that G is Lie.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.7. The Hilbert-Smith conjecture

269

We first make some basic reductions. We will need the following fact:
Exercise 2.7.1. Show that the extension of a Lie group by another Lie
group is again isomorphic to a Lie group. (Hint: use Corollary 1.5.8.) This
result was first established in [Gl1951]. Note that this result supersedes
Theorem 2.6.1, though it uses more of the theory of Hilberts fifth problem
in its proof.
From the Gleason-Yamabe theorem (Theorem 1.1.17), every locally compact group G contains an open subgroup which is an extension of a Lie group
by a compact subgroup of G. Since a group with a Lie group as an open
subgroup is again Lie (because all outer automorphisms of Lie groups are
smooth), it thus suffices by Exercise 2.7.1 to prove the claim when G is
compact. In particular, all orbits of G on X are also compact.
Let B be a small ball in chart of X around some origin x0 . By continuity,
there is some neighbourhood U of the identity in G such that gB 3 x0 for
all g U . By the Peter-Weyl theorem, there is a compact normal subgroup
G0 of G in U with G/G0 linear (and hence Lie). The set G0 B is then a
G0 -invariant manifold, which is precompact and connected (because all the
shifts gB are connected and share a common point). If we let G00 be the
subgroup of G0 that fixes G0 B, then G00 is a compact normal subgroup of G0 ,
and G0 /G00 acts faithfully on G0 B. By Newmans first theorem (Theorem
2.7.4), we see that if U is small enough, then G0 /G00 cannot contain any
elements of prime order, and hence cannot contain any non-trivial periodic
elements whatsoever; by Conjecture 2.7.3, it also cannot contain a continuously embedded copy of Zp for any p. We claim that this forces G0 /G00 to
be trivial.
As G0 B is precompact, the space of C(G0 B G0 B) of continuous maps
from G0 B to itself (with the compact-open topology) is first countable, which
makes G0 /G00 first-countable as well (since G0 /G00 is homeomorphic to a
subspace of C(G0 B G0 B)). The claim now follows from
Lemma 2.7.6. Let G be a compact first-countable group which does not
contain any non-trivial periodic elements or a continuously embedded copy
of Zp for any p. Then G is trivial.
Proof. Every element g of G is contained in a compact abelian subgroup
of G, namely the closed group hgi generated by g. Thus we may assume
without loss of generality that G is abelian.
As G is both compact and first countable, it can be written (using
Exercise 1.4.21) as the inverse limit lim Gn of a countable sequence of
compact abelian Lie groups Gn , with surjective continuous projection homomorphisms n+1n : Gn+1 Gn between these Lie groups. (Note that

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

270

2. Related articles

without first countability, one might only be an inverse limit of a net of Lie
groups rather than a sequence, as one can see for instance with the example
(R/Z)R .)
Suppose for contradiction that at least one of the Gn , say G1 , is nontrivial. It is a standard fact that every compact abelian Lie group is isomorphic to the direct product of a torus and a finite group. (Indeed, in the
connected case one can inspect the kernel of the exponential map, and then
one can extend to the general case by viewing a compact Lie group as an
extension of a finite group by a connected compact Lie group.) in particular,
the periodic points (i.e. points of finite order) are dense, and so there exists
an element g1 of G1 of finite non-trivial order. By raising g1 to a suitable
power, we may assume that g1 has some prime order p.
We now claim inductively that for each n = 1, 2, . . . , g1 can be lifted
to an element gn Gn of some order pkn , where 1 = k1 k2 . . . are
a non-decreasing set of integers. Indeed, suppose inductively that we have
already lifted g1 up to gn Gn with order pkn . The preimage of gn in Gn+1
is then a dense subset of the preimage of hgn i, which is a compact abelian
0
Lie group, and thus contains an element gn+1
of some finite order. As gn
k
0
n
has order p , gn+1 must have an order divisible by pkn , and thus of the
0
form pkn+1 q for some q coprime to p and some kn+1 kn . By raising gn+1
k
to a multiple of q that equals 1 mod p n , we may eliminate q and obtain a
preimage gn+1 of order pkn+1 , and the claim follows.
There are now two cases, depending on whether kn goes to infinity or
not. If the kn stay bounded, then they converge to a limit k, and the inverse
limit of the gn is then an element g of G of finite order pk , a contradiction.
But if the kn are unbounded, then G contains a continuously embedded
copy of the inverse limit of the Z/pkn Z, which is Zp , and again we have a
contradiction.

Since G0 /G00 and G/G0 are both Lie groups, G/G00 is Lie too. Thus it
suffices to show that G00 is Lie. Without loss of generality, we may therefore
replace G by G00 and assume that B is fixed by G.
Now let be the closure of the interior of the set of fixed points of G.
Then is non-empty, and is also clearly closed. We claim that is open;
by connectedness of X, this implies that = X, which by faithfulness of
G implies that G is trivial, giving the claim. Indeed, let x , and let B
be a small ball around x. As before, GB is then a connected G-invariant
-compact manifold, and so if G0 is the subgroup of G that fixes GB, then
G/G0 is a compact Lie group that acts faithfully on GB fixing a nontrivial
open subset of GB. By Newmans theorem (Theorem 2.7.5), G/G0 cannot
contain any periodic elements; by Conjecture 2.7.3, it also cannot contain

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.8. The Peter-Weyl theorem and nonabelian Fourier analysis

271

any copy of Zp . By Lemma 2.7.6, G/G0 is trivial, and so G fixes all of B.

Thus gives the desired openness of as required.
Remark 2.7.7. An alternate derivation of Conjecture 2.7.2 from Conjecture
2.7.3, suggested to the author by John Pardon, goes as follows. Again we
may assume G to be compact. If G is NSS, then it is Lie (e.g. by PeterWeyl) and we are done, so suppose G is not NSS. By Newmans theorem,
we can find an compact neighbourhood of the identity in G whose action
on X contains no periodic transformations other than the identity. As G
is NSS, we may thus find a non-trivial compact subgroup G0 inside this
neighbourhood, which has no non-trivial periodic elements. Also, as the
space of homeomorphisms on M is metrisable, it is first countable; as the
action maps the compact Hausdorff space G0 continuously and injectively
into the Hausdorff space of homeomorphisms on M , G0 is homeomorphic to
its image and is thus also first countable. By Lemma 2.7.6, G0 then contains
a copy of Zp , and the claim follows.
Remark 2.7.8. Whereas actions of the finite group Z/pZ on manifolds
can be analysed by degree theory, it appears that actions of p-adic groups
Zp require more sophisticated homological tools; the p-adic analogues of the
averaged map now typically have infinite preimages and it is no longer obvious how to compute the degree of such maps. Nevertheless, some progress
has been made along these lines under some additional regularity hypotheses on the action, such as Lipschitz continuity; see for instance [ReSc97].
Note that if T : X X is the action of the generator of Zp , then the powers
n
T p : X X will converge to the identity locally uniformly as n .
This is already enough to rule out generating maps T that are smooth but
non-trivial by use of Taylor expansion; see [MoZi1974]. (In fact, this type
of argument even works for C 1 actions.)

2.8. The Peter-Weyl theorem and nonabelian Fourier

analysis
Let G be a compact group. (Throughout this section, all topological groups
are assumed to be Hausdorff.) Then G has a number of unitary representations, i.e. continuous homomorphisms : G U (H) to the group
U (H) of unitary operators on a Hilbert space H, equipped with the strong
operator topology. In particular, one has the left-regular representation
: G U (L2 (G)), where we equip G with its normalised Haar measure
(and the Borel -algebra) to form the Hilbert space L2 (G), and is the
translation operation
(g)f (x) := f (g 1 x).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

272

2. Related articles

We call two unitary representations : G U (H) and 0 : G U (H 0 )

isomorphic if one has 0 (g) = U (g)U 1 for some unitary transformation
U : H H 0 , in which case we write 0 .
Given two unitary representations : G U (H) and 0 : G U (H 0 ),
one can form their direct sum 0 : G U (H H 0 ) in the obvious manner: 0 (g)(v) := ((g)v, 0 (g)v). Conversely, if a unitary representation
: G U (H) has a closed invariant subspace V H of H (thus (g)V V
for all g G), then the orthogonal complement V is also invariant, leading to a decomposition V V of into the subrepresentations
V : G U (V ), V : G U (V ). Accordingly, we will call a unitary
representation : G U (H) irreducible if H is nontrivial (i.e. H 6= {0})
and there are no nontrivial invariant subspaces (i.e. no invariant subspaces
other than {0} and H); the irreducible representations play a role in the
subject analogous to those of prime numbers in multiplicative number theory. By the principle of infinite descent, every finite-dimensional unitary
representation is then expressible (perhaps non-uniquely) as the direct sum
of irreducible representations.
The Peter-Weyl theorem asserts, among other things, that the same
claim is true for the regular representation:
Theorem 2.8.1 (Peter-Weyl theorem). Let G be a compact group. Then
the regular representation : G U (L2 (G)) is isomorphic to the direct
L
dim(V )
sum of irreducible representations. In fact, one has
,

G
where ( )G is an enumeration of the irreducible finite-dimensional unitary
representations : G U (V ) of G (up to isomorphism). (It is not difficult
to see that such an enumeration exists.)
In the case when G is abelian, the Peter-Weyl theorem is a consequence
of the Plancherel theorem; in that case, the irreducible representations are
of characters
all one dimensional, and are thus indexed by the space G
: G R/Z (i.e. continuous homomorphisms into the unit circle R/Z),
known as the Pontryagin dual of G; see e.g. [Ta2010, 1.12]. Conversely,
the Peter-Weyl theorem can be used to deduce the Plancherel theorem for
compact groups, as well as other basic results in Fourier analysis on these
groups, such as the Fourier inversion formula. A baby version of this theorem
(Theorem 1.4.12) is also an essential component in the solution to Hilberts
fifth problem, as discussed in Section 1.4. In this section, we will upgrade
Theorem 1.4.12 to the full Peter-Weyl theorem.

2.8.1. Proof of the Peter-Weyl theorem. Henceforth in this section,

G is a fixed compact group.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.8. The Peter-Weyl theorem and nonabelian Fourier analysis

273

Let : G U (H) and 0 : G U (H 0 ) be unitary representations. An

(linear) equivariant map T : H H 0 is defined to be a continuous linear
transformation such that T (g) = 0 (g)T for all g G.
A fundamental fact in representation theory, known as Schurs lemma,
asserts (roughly speaking) that equivariant maps cannot mix irreducible
representations together unless they are isomorphic. More precisely:
Lemma 2.8.2 (Schurs lemma for unitary representations). Suppose that
: G U (H) and 0 : G U (H 0 ) are irreducible unitary representations,
and let T : H H 0 be an equivariant map. Then T is either the zero
transformation, or a constant multiple of an isomorphism. In particular, if
6 0 , then there are no non-trivial equivariant maps between H and H 0 .
Proof. The adjoint map T : H 0 H of the equivariant map T is also
equivariant, and thus so is T T : H H. As T T is also a bounded selfadjoint operator, we can apply the spectral theorem to it. Observe that any
closed invariant subspace of T T is G-invariant, and is thus either {0} or H.
By the spectral theorem, this forces T T to be a constant multiple of the
identity. Similarly for T T . This forces T to either be zero or a constant
multiple of a unitary map, and the claim follows. (Thanks to Frederick
Goodman for this proof.)

Schurs lemma has many foundational applications in the subject. For
instance, we have the following generalisation of the well-known fact that
eigenvectors of a unitary operator with distinct eigenvalues are necessarily
orthogonal:
Corollary 2.8.3. Let V : G U (V ) and W : G U (W ) be two
irreducible subrepresentations of a unitary representation : G U (H).
Then one either has V W or V W .
Proof. Apply Schurs lemma to the orthogonal projection from W to V . 
Another application shows that finite-dimensional linear representations
can be canonically identified (up to constants) with finite-dimensional unitary representations:
Corollary 2.8.4. Let : G GL(V ) be a linear representation on a finitedimensional space V . Then there exists a Hermitian inner product h, i on
V that makes this representation unitary. Furthermore, if V is irreducible,
then this inner product is unique up to constants.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

274

2. Related articles

Proof. To show existence of the Hermitian inner product that unitarises ,

take an arbitrary Hermitian inner product h, i0 and then form the average
Z
hv, wi := h(g)v, (g)wi0 d(g).
G

(this is the Weyl averaging trick, which crucially exploits compactness of

G). Then one easily checks (using the fact that V is finite dimensional and
thus locally compact) that h, i is also Hermitian, and that is unitary with
respect to this inner product, as desired. (This part of the argument does
not use finite dimensionality.)
To show uniqueness up to constants, assume that one has two such
inner products h, i, h, i0 on V , and apply Schurs lemma to the identity
map between the two Hilbert spaces (V, h, i) and (V, h, i0 ). (Here, finite
dimensionality is used to establish

A third application of Schurs lemma allows us to express the trace of a
linear operator as an average:
Corollary 2.8.5. Let : G GL(H) be an irreducible unitary representation on a non-trivial finite-dimensional space H, and let T : H H be a
linear transformation. Then
Z
1
trH (T )IH =
(g)T (g) d(g),
dim(H)
G
where IH : H H is the identity operator.
Proof. The right-hand side is equivariant, and hence by Schurs lemma is a
multiple of the identity. Taking traces, we see that the right-hand side also
has the same trace as T . The claim follows.

Let us now consider the irreducible subrepresentations V : G U (V )
of the left-regular representation : G U (L2 (G)). From Corollary 2.8.3,
we know that those subrepresentations coming from different isomorphism
are orthogonal, so we now focus attention on those subrepresenclasses in G
Define the -isotypic component
tations coming from a single class G.
2
L (G) of the regular representation to be the finite-dimensional subspace
of L2 (G) spanned by the functions of the form
f,v,w : g 7 hv, (g)wiV
where v, w are arbitrary vectors in V . This is clearly a left-invariant subspace of L2 (G) (in fact, it is bi-invariant, a point which we will return
to later), and thus induces a subrepresentation of the left-regular representation. In fact, it captures precisely all the subrepresentations of the
left-regular representation that are isomorphic to :

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.8. The Peter-Weyl theorem and nonabelian Fourier analysis

275

Then every irreducible subrepresentation

Proposition 2.8.6. Let G.
V : G U (V ) of the left-regular representation : G U (L2 (G)) that
is isomorphic to is a subrepresentation of L2 (G) . Conversely, L2 (G) is
dim(V )

isomorphic to the direct sum

of dim(V ) copies of : G U (V ).
2
(In particular, L (G) has dimension dim(V )2 ).

Proof. Let V : G U (V ) be a subrepresentation of the left-regular representation that is isomorphic to . Thus, we have an equivariant isometry
: V L2 (G) whose image is V ; it has an adjoint : L2 (G) V .
Let v V and K L2 (G). The convolution
Z
(v) K(g) :=
(v)(gh)K(h1 ) d(h)
G

can be re-arranged as
Z

d(h)
(g 1 )((v))(h)K(h)

L2 (G)
= h (g 1 )((v)), Ki
L2 (G)
= h( (g 1 )v), Ki
V
= h (g 1 )v, Ki

= hv, (g) KiV

where

:= K(g 1 ).
K(g)
In particular, we see that (v) K L2 (G) for every K. Letting K be
a sequence (or net) of approximations to the identity, we conclude that
(v) L2 (G) as well, and so V L2 (G) , which is the first claim.
To prove the converse claim, write n := dim(V ), and let e1 , . . . , en be
an orthonormal basis for V . Observe that we may then decompose L2 (G)
as the direct sum of the spaces
L2 (G),ei := {f,v,ei : v V }
for i = 1, . . . , n. The claim follows.

are
From Corollary 2.8.3, the -isotypic components L2 (G) for G
2
pairwise orthogonal, and so we can form the direct sum G L (G)
dim(G)

G
, which is an invariant subspace of L2 (G) that contains all
the finite-dimensional irreducible subrepresentations (and hence also all the
finite-dimensional representations, period). The essence of the Peter-Weyl
theorem is then the assertion that this direct sum in fact occupies all of
L2 (G):

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

276

2. Related articles

Proposition 2.8.7. We have L2 (G) = G L2 (G) .

Proof. Suppose this is not the case. Taking orthogonal complements, we
conclude that there exists a non-trivial f L2 (G) which is orthogonal to
all L2 (G) , and is in particular orthogonal to all finite-dimensional subrepresentations of L2 (G).
Now let K L2 (G) be an arbitrary self-adjoint kernel, thus K(g 1 ) =
K(g) for all g G. The convolution operator T : f 7 f K is then a
self-adjoint Hilbert-Schmidt operator and is thus compact. (Here, we have
crucially used the compactness of G.) By the spectral theorem, the cokernel
ker(T ) of this operator then splits as the direct sum of finite-dimensional
eigenspaces. As T is equivariant, all these eigenspaces are invariant, and thus
orthogonal to f ; thus f must lie in the kernel of T , and thus f K vanishes
for all self-adjoint K L2 (G). Using a sequence (or net) of approximations
to the identity, we conclude that f vanishes also, a contradiction.

Theorem 2.8.1 follows by combining this proposition with 2.8.6.
the space HS(V ) of
2.8.2. Nonabelian Fourier analysis. Given G,
linear transformations from V to V is a finite-dimensional Hilbert space,
with the Hilbert-Schmidt inner product hS, T iHS(V ) := trV ST ; it has
a unitary action of G as defined by HS(V ) (g) : T 7 (g)T . For any T
HS(V ), the function g 7 hT, (g)iHS(V ) can be easily seen to lie in L2 (G) ,
giving rise to a map : HS(V ) L2 (G) . It is easy to see that this map
is equivariant.
the map dim(V )1/2 is unitary.
Proposition 2.8.8. For each G,
Proof. As HS(V ) and L2 (G) are finite-dimensional spaces with the same
dimension dim(V )2 , it suffices to show that this map is an isometry, thus
we need to show that
1
hF (S), F (T )iL2 (G) =
hS, T iHS(V )
dim(V )
for all S, T HS(V ). By bilinearity, we may reduce to the case when S, T
are rank one operators
S := ab ; T := cd
for some a, b, c, d V , where b : V C is the dual vector b : v 7 hv, bi
to b, and similarly for d. Then we have
hS, T iHS(V ) = trV ab dc = (c a)(b d)
and
hF (S), F (T )iL2 (G)

Z
=

(b (g)a)(d (g)c) d(G).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.8. The Peter-Weyl theorem and nonabelian Fourier analysis

277

The latter expression can be rewritten as

Z
h (g) db (g), ca iHS(V ) d(g).
G

Applying Fubinis theorem, followed by Corollary 2.8.5, this simplifies to



1

,
tr(db )IV , ca
dim(V )
HS(V )
which simplifies to

dim(V ) (c a)(b d),

and the claim follows.

As a corollary of the above proposition, the orthogonal projection of a

function f L2 (G) to L2 (G) can be expressed as
dim(V ) f.
We call

f() := f =

f (g)(g) d(g) HS(V )

G

the Fourier coefficient of f at , thus the projection of f to L2 (G) is the

function
g 7 dim(V )hf(), (g)i
which has an L2 (G) norm of dim(V )1/2 kf()kHS(V ) . From the Peter-Weyl
theorem we thus obtain the Fourier inversion formula
X
f (g) =
dim(V )hf(), (g)i

and the Plancherel identity

kf k2L2 (G) =

dim(V )kf()k2HS(V ) .

We can write these identities more compactly as an isomorphism

M
(2.39)
L2 (G)
dim(V ) HS(V )

where the dilation c H of a Hilbert space H is formed by using the inner

product hv, wicH := chv, wiH . This is an isomorphism not only of Hilbert
spaces, but of the left-action of G. Indeed, it is an isomorphism of the
bi-action of G G on both the left and right of both L2 (G) and HS(V ),
defined by
L2 (G),GG (g, h)(f )(x) := f (g 1 xh)
and
,GG (g, h)(T ) := (g)T (h) .
It is easy to see that each of the HS(V ) are irreducible with respect to
the G G action. Indeed, first observe from Proposition 2.8.8 that is

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

278

2. Related articles

surjective, and thus (g) HS(V ) must span all of HS(V ). Thus, any
bi-invariant subspace of HS(V ) must also be invariant with respect to left
and right multiplication by arbitrary elements of HS(V ), and in particular
by rank one operators; from this one easily sees that there are no non-trivial
bi-invariant subspaces. Thus we can view the Peter-Weyl theorem as also
describing the irreducible decomposition of L2 (G) into G G-irreducible
components.
as being the
Remark 2.8.9. In view of (2.39), it is natural to view G
occuring with multiplicity
spectrum of G, with each frequency G
dim(V ).
In the abelian case, any eigenspace of one unitary operator (g) is automatically an invariant subspace of all other (h), which quickly implies (from
the spectral theorem) that all irreducible finite-dimensional unitary representations must be one-dimensional, in which case we see that the above
formulae collapse to the usual Fourier inversion and Plancherel theorems for
compact abelian groups.
In the case of a finite group G, we can take dimensions in (2.39) to obtain
the identity
X
|G| =
dim(V )2 .

have the same

In the finite abelian case, we see in particular that G and G
cardinality.
Direct computation also shows other basic Fourier identities, such as the
convolution identity

f\
1 f2 () = f1 ()f2 ()
for f1 , f2 L2 (G), thus partially diagonalising convolution into multiplication of linear operators on finite-dimensional vector spaces V . (Of course,
one cannot expect complete diagonalisation in the non-abelian case, since
convolution would then also be non-abelian, whereas diagonalised operators
must always commute with each other.)
Call a function f L2 (G) a class function if it is conjugation-invariant,
thus f (gxg 1 ) = f (x) for all x, g G. It is easy to see that this is equivalent to each of the Fourier coefficients f() also being conjugation-invariant:
(g)f() (g) = f(). By Lemma 2.8.5, this is in turn equivalent to f()
being equal to a multiple of the identity:
f() =

1
1
tr(f())IV =
hf, iL2 (G) IV
dim(V )
dim(V )

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.8. The Peter-Weyl theorem and nonabelian Fourier analysis

279

where the character L2 (G) of the representation is given by the

formula
(g) := trV (g).
The Plancherel identity then simplifies to
X
hf, iL2 (G) ,
f=

thus the form an orthonormal basis for the space L2 (G)G of class functions. Analogously to (2.39), we have
M
C.
L2 (G)G

has the same cardinality as

(In particular, in the case of finite groups G, G
the space of conjugacy classes of G.)
Characters are a fundamentally important tool in analysing finite-dimensional
representations V of G that are not necessarily irreducible; indeed, if V
L
m
, then the character V (g) :=
decomposes into irreducibles as
V
G
trV (g ) then similarly splits as
X
V =
m

and so the multiplicities m of each component V in V can be given by the

formula
m = hV , iL2 (G) .
In particular, these multiplicities are unique: all decompositions of V into
irreducibles have the same multiplicities.
Remark 2.8.10. Representation theory becomes much more complicated
once one leaves the compact case; convolution operators f 7 f K are no
longer compact, and can now admit continuous spectrum in addition to pure
point spectrum. Furthermore, even when one has pure point spectrum, the
eigenspaces can now be infinite dimensional. Thus, one must now grapple
with infinite-dimensional irreducible representations, as well as continuous
combinations of representations that cannot be readily resolved into irreducible components. Nevertheless, in the important case of locally compact
groups, it is still the case that there are enough irreducible unitary representations to recover a significant portion of the above theory. The fundamental theorem here is the Gelfand-Raikov theorem, which asserts that given
any non-trivial group element g in a locally compact group, there exists a
irreducible unitary representation (possibly infinite-dimensional) on which g
acts non-trivially. Very roughly speaking, this theorem is first proven by observing that g acts non-trivially on the regular representation, which (by the

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

280

2. Related articles

Gelfand-Naimark-Segal (GNS) construction) gives a state on the *-algebra

of measures on G that distinguishes the Dirac mass g at g from the Dirac
mass 0 from the origin. Applying the Krein-Milman theorem, one then
finds an extreme state with this property; applying the GNS construction,
one then obtains the desired irreducible representation.

2.9. Polynomial bounds via nonstandard analysis

As discussed in Section 1.7, nonstandard analysis is useful in allowing one
to import tools from infinitary (or qualitative) mathematics in order to establish results in finitary (or quantitative) mathematics. One drawback,
though, to using nonstandard analysis methods is that the bounds one obtains by such methods are usually ineffective: in particular, the conclusions
of a nonstandard analysis argument may involve an unspecified constant
C that is known to be finite but for which no explicit bound is obviously
available2.
Because of this fact, it would seem that quantitative bounds, such as
polynomial type bounds X CY C that show that one quantity X is controlled in a polynomial fashion by another quantity Y , are not easily obtainable through the ineffective methods of nonstandard analysis. Actually,
this is not the case; as I will demonstrate by an example below, nonstandard
analysis can certainly yield polynomial type bounds. The catch is that the
exponent C in such bounds will be ineffective; but nevertheless such bounds
are still good enough for many applications.
Let us now illustrate this by reproving a lemma from a paper of Chang
[Ch2003, Lemma 2.14], which was recently pointed out to me by Van Vu.
Changs paper is focused primarily on the sum-product problem, but she
uses a quantitative lemma from algebraic geometry which is of independent
interest. To motivate the lemma, let us first establish a qualitative version:
Lemma 2.9.1 (Qualitative solvability). Let P1 , . . . , Pr : Cd C be a finite
number of polynomials in several variables with rational coefficients. If there
is a complex solution z = (z1 , . . . , zd ) Cd to the simultaneous system of
equations
P1 (z) = = Pr (z) = 0,
d

then there also exists a solution z Q whose coefficients are algebraic

numbers (i.e. they lie in the algebraic closure Q of the rationals).
2In many cases, a bound can eventually be worked out by performing proof mining on
the argument, and in particular by carefully unpacking the proofs of all the various results from
infinitary mathematics that were used in the argument, as opposed to simply using them as black
boxes, but this is a time-consuming task and the bounds that one eventually obtains tend to be
quite poor (e.g. tower exponential or Ackermann type bounds are not uncommon).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.9. Polynomial bounds via nonstandard analysis

281

Proof. Suppose there was no solution to P1 (z) = = Pr (z) = 0 over Q.

Applying Hilberts nullstellensatz (which is available as Q is algebraically
closed), we conclude the existence of some polynomials Q1 , . . . , Qr (with
coefficients in Q) such that
P1 Q1 + + Pr Qr = 1
as polynomials. In particular, we have
P1 (z)Q1 (z) + + Pr (z)Qr (z) = 1
Cd .

for all z
This shows that there is no solution to P1 (z) = = Pr (z) =
0 over C, as required.

Remark 2.9.2. Observe that in the above argument, one could replace Q
and C by any other pair of fields, with the latter containing the algebraic
closure of the former, and still obtain the same result.
The above lemma asserts that if a system of rational equations is solvable
at all, then it is solvable with some algebraic solution. But it gives no bound
on the complexity of that solution in terms of the complexity of the original
equation. Changs lemma provides such a bound. If H 1 is an integer,
let us say that an algebraic number has height at most H if its minimal
polynomial (after clearing denominators) consists of integers of magnitude
at most H.
Lemma 2.9.3 (Quantitative solvability). Let P1 , . . . , Pr : Cd C be a
finite number of polynomials of degree at most D with rational coefficients,
each of height at most H. If there is a complex solution z = (z1 , . . . , zd ) Cd
to the simultaneous system of equations
P1 (z) = = Pr (z) = 0,
d

then there also exists a solution z Q whose coefficients are algebraic

numbers of degree at most C and height at most CH C , where C = CD,d,r
depends only on D, d and r.
Chang proves this lemma by essentially establishing a quantitative version of the nullstellensatz, via elementary elimination theory (somewhat similar, actually, to approach I took to the nullstellensatz in [Ta2008, 1.15]).
She also notes that one could also establish the result through the machinery of Gr
obner bases. In each of these arguments, it was not possible to
use Lemma 2.9.1 (or the closely related nullstellensatz) as a black box; one
actually had to unpack one of the proofs of that lemma or nullstellensatz
to get the polynomial bound. However, using nonstandard analysis, it is
possible to get such polynomial bounds (albeit with an ineffective value of
the constant C) directly from Lemma 2.9.1 (or more precisely, the generalisation in Remark 2.9.2) without having to inspect the proof, and instead

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

282

2. Related articles

simply using it as a black box, thus providing a soft proof of Lemma 2.9.3
that is an alternative to the hard proofs mentioned above.
Heres how the proof works. Informally, the idea is that Lemma 2.9.3
should follow from Lemma 2.9.1 after replacing the field of rationals Q with
the field of rationals of polynomially bounded height. Unfortunately, the
latter object does not really make sense as a field in standard analysis;
nevertheless, it is a perfectly sensible object in nonstandard analysis, and
this allows the above informal argument to be made rigorous.
We turn to the details. As is common whenever one uses nonstandard
analysis to prove finitary results, we use a compactness and contradiction
argument (or more precisely, an ultralimit and contradiction argument).
Suppose for contradiction that Lemma 2.9.3 failed. Carefully negating the
quantifiers (and using the axiom of choice), we conclude that there exists
D, d, r such that for each natural number n, there is a positive integer H (n)
(n)
(n)
and a family P1 , . . . , Pr : Cd C of polynomials of degree at most D
and rational coefficients of height at most H (n) , such that there exist at least
one complex solution z (n) Cd to
(2.40)

(n)

P1 (z (n) ) = = Pr (z (n) ) = 0,

but such that there does not exist any such solution whose coefficients are
algebraic numbers of degree at most n and height at most n(H (n) )n .
Now we take ultralimits, as in Section 1.7. Let N\N be a nonprincipal ultrafilter. For each i = 1, . . . , r, the ultralimit
(n)
Pi := lim Pi
n

(n)
Pi

of the (standard) polynomials

is a nonstandard polynomial Pi : Cd
C of degree at most D, whose coefficients now lie in the nonstandard rationals Q. Actually, due to the height restriction, we can say more. Let
H := limn H (n) N be the ultralimit of the H (n) , this is a nonstandard
natural number (which will almost certainly be unbounded, but we will not
need to use this). Let us say that a nonstandard integer a is of polynomial
size if we have |a| CH C for some standard natural number C, and say
that a nonstandard rational number a/b is of polynomial height if a, b are of
polynomial size. Let Qpoly(H) be the collection of all nonstandard rationals
of polynomial height. (In the language of nonstandard analysis, Qpoly(H) is
an external set rather than an internal one, because it is not itself an ultraproduct of standard sets; but this will not be relevant for the argument that
follows.) It is easy to see that Qpoly(H) is a field, basically because the sum
or product of two integers of polynomial size, remains of polynomial size. By
construction, it is clear that the coefficients of Pi are nonstandard rationals
of polynomial height, and thus P1 , . . . , Pr are defined over Qpoly(H) .

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.10. Loeb measure and the triangle removal lemma

283

Meanwhile, if we let z := limn z (n) Cd be the ultralimit of the

solutions z (n) in (2.40), we have
P1 (z) = = Pr (z) = 0,
thus P1 , . . . , Pr are solvable in C. Applying Lemma 2.9.1 (or more precisely,
the generalisation in Remark 2.9.2), we see that P1 , . . . , Pr are also solvable
in Qpoly(H) . (Note that as C is algebraically closed, C is also (by Loss
theorem), and so C contains Qpoly(H) .) Thus, there exists w Qpoly(H)
with
P1 (w) = = Pr (w) = 0.

As Qpoly(H) lies in Cd , we can write w as an ultralimit w = limn w(n) of

standard complex vectors w(n) Cd . By construction, the coefficients of w
each obey a non-trivial polynomial equation of degree at most C and whose
coefficients are nonstandard integers of magnitude at most CH C , for some
standard natural number C. Undoing the ultralimit, we conclude that for n
sufficiently close to p, the coefficients of w(n) obey a non-trivial polynomial
equation of degree at most C whose coefficients are standard integers of
magnitude at most C(H (n) )C . In particular, these coefficients have height
at most C(H (n) )C . Also, we have
(n)

P1 (w(n) ) = = Pr(n) (w(n) ) = 0.

(n)

But for n larger than C, this contradicts the construction of the Pi , and
the claim follows. (Note that as p is non-principal, any neighbourhood of p
in N will contain arbitrarily large natural numbers.)
Remark 2.9.4. The same argument actually gives a slightly stronger version of Lemma 2.9.3, namely that the integer coefficients used to define the
algebraic solution z can be taken to be polynomials in the coefficients of
P1 , . . . , Pr , with degree and coefficients bounded by CD,d,r .
Remark 2.9.5. A related application of nonstandard analysis to quantitative algebraic geometry was given in [Sc1989]. (Thanks to Matthias
Aschenbrenner for this reference.)

2.10. Loeb measure and the triangle removal lemma

Formally, a measure space is a triple (X, B, ), where X is a set, B is a
-algebra of subsets of X, and : B [0, +] is a countably additive
unsigned measure on B. If the measure (X) of the total space is one, then
the measure space becomes a probability space. If a non-negative function
f : X [0, +] Ris B-measurable (or measurable for short), one can then
form the integral X f d [0, +] by the usual abstract measure-theoretic
construction (as discussed for instance in [Ta2011, 1.4]).

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

284

2. Related articles

A measure space is complete if every subset of a null set (i.e. a measurable set of measure zero) is also a null set. Not all measure spaces are
complete, but one can always form the completion (X, B, ) of a measure
space (X, B, ) by enlarging the -algebra B to the space of all sets which are
equal to a measurable set outside of a null set, and extending the measure
appropriately.
Given two (-finite) measure spaces (X, BX , X ) and (Y, BY , Y ), one
can form the product space (X Y, BX BY , X Y ). This is a measure
space whose domain is the Cartesian product X Y , the -algebra BX BY
is generated by the rectangles A B with A BX , B BY , and the
measure X Y is the unique measure on BX BY obeying the identity
X Y (A B) = X (A)Y (B).
See for instance [Ta2011, 1.7] for a formal construction of product measure3. One of the fundamental theorems concerning product measure is
Tonellis theorem (which is basically the unsigned version of the more wellknown Fubini theorem), which asserts that if f : X Y [0, +] is BX BY
measurable, then the integral expressions
Z Z
( f (x, y) dY (y)) dX (x)
X Y
Z Z
( f (x, y) dX (x)) dY (y)
Y

and

Z
f (x, y) dXY (x, y)
XY

all exist (thus all integrands are almost-everywhere well-defined and measurable with respect to the appropriate -algebras), and are all equal to each
other; see e.g. [Ta2011, Theorem 1.7.10].
Any finite non-empty set V can be turned into a probability space
(V, 2V , V ) by endowing it with the discrete -algebra 2V := {A : A V }
of all subsets of V , and the normalised counting measure
|A|
(A) :=
,
|V |
where |A| denotes the cardinality of A. In this discrete setting, the probability space is automatically complete, and every function f : V [0, +]
is measurable, with the integral simply being the average:
Z
1 X
f dV =
f (v).
|V |
V
vV

3There are technical difficulties with the theory when X or Y is not -finite, but in this
section we will only be dealing with probability spaces, which are clearly -finite, so this difficulty
will not concern us.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.10. Loeb measure and the triangle removal lemma

285

Of course, Tonellis theorem is obvious for these discrete spaces; the deeper
content of that theorem is only apparent at the level of continuous measure
spaces.
Among other things, this probability space structure on finite sets can
be used to describe various statistics of dense graphs. Recall that a graph
G = (V, E) is a finite vertex set V , together with a set of edges E, which
we will think of as a symmetric subset of the Cartesian product V V .
(If one wishes, one can prohibit loops in E, so that E is disjoint from the
diagonal V := {(v, v) : v V } of V V , but this will not make much
difference for the discussion below.) Then, if V is non-empty, and ignoring
some minor errors coming from the diagonal V , the edge density of the
graph is essentially
Z
e(G) := V V (E) =
1E (v, w) dV V (v, w),
V V

the triangle density of the graph is basically

Z
t(G) :=
1E (u, v)1E (v, w)1E (w, u) dV V V (u, v, w),
V V V

and so forth.
In [Ru1978], Ruzsa and Szemeredi established the triangle removal
lemma concerning triangle densities, which informally asserts that a graph
with few triangles can be made completely triangle-free by removing a small
number of edges:
Lemma 2.10.1 (Triangle removal lemma). Let G = (V, E) be a graph on a
non-empty finite set V , such that t(G) for some > 0. Then there exists
a subgraph G0 = (V, E 0 ) of G with t(G0 ) = 0, such that e(G\G0 ) = o0 (1),
where o0 (1) denotes a quantity bounded by c() for some function c() of
that goes to zero as 0.
The original proof of the triangle removal lemma was a finitary one,
and proceeded via the Szemeredi regularity lemma [Sz1978]. It has a number of consequences; for instance, as already noted in that paper, the triangle
removal lemma implies as a corollary the famous theorem of Roth [Ro1953]
that subsets of Z of positive upper density contain infinitely many arithmetic
progressions of length three.
It is however also possible to establish this lemma by infinitary means.
There are at least three basic approaches for this. One is via a correspondence principle between questions about dense finite graphs, and questions
about exchangeable random infinite graphs, as was pursued in [Ta2007]. A
second (closely related to the first) is to use the machinery of graph limits, as developed in [LoSz2006], [BoChLoSoVe2008]. The third is via

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

286

2. Related articles

nonstandard analysis (or equivalently, by using ultraproducts), as was pursued in [ElSz2006]. These three approaches differ in the technical details
of their execution, but the net effect of all of these approaches is broadly
the same, in that they both convert statements about large dense graphs
(such as the triangle removal lemma) to measure theoretic statements on
infinitary measure spaces. (This is analogous to how the Furstenberg correspondence principle converts combinatorial statements about dense sets of
integers into ergodic-theoretic statements on measure-preserving systems, as
discussed for instance in [Ta2009].)
In this section we will illustrate the nonstandard analysis approach from
[ElSz2006] by providing a nonstandard proof of the triangle removal lemma.
The main technical tool used here (besides the basic machinery of nonstandard analysis) is that Loeb measure,[Lo1975] which gives a probability
Q space
structure (V, BV , V ) to nonstandard finite non-empty sets V = n Vn
that is an infinitary analogue of the discrete probability space structures
V = (V, 2V , V ) one has on standard finite non-empty sets. The nonstandard analogue of quantities such as triangle densities then become the integrals of various nonstandard functions with respect to Loeb measure. With
this approach, the epsilons and deltas that are so prevalent in the finitary
approach to these subjects disappear almost completely; but to compensate
for this, one now must pay much more attention to questions of measurability, which were automatic in the finitary setting but now require some care
in the infinitary one.
The nonstandard analysis approaches are also related to the regularity
lemma approach; see [Ta2011c, 4.4] for a proof of the regularity lemma
using Loeb measure.
As usual, the nonstandard approach offers a complexity tradeoff: there
is more effort expended in building the foundational mathematical structures of the argument (in this case, ultraproducts and Loeb measure), but
once these foundations are completed, the actual arguments are shorter than
their finitary counterparts. In the case of the triangle removal lemma, this
tradeoff does not lead to a particularly significant reduction in complexity
(and arguably leads in fact to an increase in the length of the arguments,
when written out in full), but the gain becomes more apparent when proving more complicated results, such as the hypergraph removal lemma, in
which the initial investment in foundations leads to a greater savings in net
complexity, as can be seen in [ElSz2006].

2.10.1. Loeb measure. We use the usual setup of nonstandard analysis

(as reviewed in Section 1.7). Thus, we will fix a non-principal ultrafilter
N\N on the natural numbers N.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.10. Loeb measure and the triangle removal lemma

287

Consider
a nonstandard finite non-empty set V , i.e. an ultraproduct
Q
V = n Vn of standard finite non-empty setsQVn . Define an internal
subset of V to be a subset of V of the form A = n An , where each An
is a subset of Vn . It is easy to see that the collection AV of all internal
subsets of V is a boolean algebra. In general, though, AV will not be a algebra. For instance, suppose that the Vn are the standard discrete intervals
Vn := [1, n] := {i N : i n}, then V is the non-standard discrete interval
V = [1, N ] := {i N : i N }, where N is the unbounded nonstandard
natural number N := limn n. For any standard integer m, the subinterval
[1, N/m] is an internal subset of V ; but the intersection
\
[1, N/m] = {i N : i = o(N )}
[1, o(N )] :=
mN

is not an internal subset of V . (This can be seen, for instance, by noting that
all non-empty internal subsets of [1, N ] have a maximal element, whereas
[1, o(N )] does not.)
Q
Given any internal subset A = n An of V , we can define the cardinality |A| of A, which is the nonstandard natural number |A| := limn |An |.
|A|
We then have the nonstandard density |V
| , which is a nonstandard real
number between 0 and 1. By Exercise 1.7.18, this bounded nonstandard
|A|
|A|
real number |V
| has a unique standard part st( |V | ), which is a standard real
number in [0, 1] such that
|A|
|A|
= st(
) + o(1),
|V |
|V |
where o(1) denotes a nonstandard infinitesimal (i.e. a nonstandard number
which is smaller in magnitude than any standard > 0).
In [Lo1975], Loeb observed that this standard density can be extended
to a complete probability measure:
Theorem 2.10.2 (Construction of Loeb measure). Let V be a nonstandard finite non-empty set. Then there exists a complete probability space
(V, LV , V ), with the following properties:
(Internal sets are Loeb measurable) If A is an internal subset of V ,
then A LV and
V (A) = st(

|A|
).
|V |

(Loeb measurable sets are almost internal) If E is a subset of V ,

then E is Loeb measurable if and only if, for every standard > 0,

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

288

2. Related articles

there exists internal subsets A, B1 , B2 , . . . of V such that

[
Bn
EA
n=1

and

V (Bn ) .

n=1
|A|
Proof. The map V : A 7 st( |V
| ) is a finitely additive probability measure
on AV . We claim that this map V is in fact a pre-measure on AV , thus
one has

X
(2.41)
V (A) =
V (An )
n=1

whenever A is an internal set that is partitioned into a disjoint sequence

of internal sets An . But the countable sequence of sets A\(A1 . . . An )
are internal, and have empty intersection, so by the countable saturation
property of ultraproducts (Lemma 1.7.13), one of the A\(A1 An )
must be empty. The pre-measure property (2.41) then follows from the
finite additivity of V .
Invoking the Hahn-Kolmogorov extension theorem (see e.g. [Ta2011,
Theorem 1.7.8]), we conclude that V extends to a countably additive probability measure on the -algebra hAV i generated by the internal sets. This
measure need not be complete, but we can then pass to the completion
LV := hAV i of that -algebra. This probability space certainly obeys the
first property. The only if portion of second property asserts that all
Loeb measurable sets differ from an internal set by sets of arbitrarily small
outer measure, but this is easily seen since the space of all sets that have
this property is easily verified to be a complete -algebra that contain the
algebra of internal sets. The if portion follows easily from the fact that
LV is a complete -algebra containing the internal sets. (These facts are
very similar to the more familiar facts that a bounded subset of a Euclidean
space is Lebesgue measurable if and only if it differs from an elementary set
by a set of arbitrarily small outer measure.)

Now we turn to the analogue of Tonellis theorem for Loeb measure,
which will be a fundamental tool when it comes to prove the triangle removal
lemma. Let V, W be two nonstandard finite non-empty sets, then V
W is also a nonstandard finite non-empty set. We then have three Loeb
probability spaces
(V, LV , V ),
(W, LW , W ),

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.10. Loeb measure and the triangle removal lemma

289

and
(V W, LV W , V W ),

(2.42)

and we also have the product space

(2.43)

(V W, LV LW , V W ).

It is then natural to ask how the two probability spaces (2.42) and (2.43)
are related. There is one easy relationship, which shows that (2.42) extends
(2.43):
Exercise 2.10.1. Show that (2.42) is a refinement of (2.43), thus LV LW ,
and V W extends V W . (Hint: first recall why the product of Lebesgue
measurable sets is Lebesgue measurable, and mimic that proof to show that
the product of a LV -measurable set and a LW -measurable set is LV W measurable, and that the two measures V W and V W agree in this
case.)
In the converse direction, (2.42) enjoys the type of Tonelli theorem that
(2.43) does:
Theorem 2.10.3 (Tonelli theorem for Loeb measure). Let V, W be two
nonstandard finite non-empty sets, and let f : V W [0, +] be an
unsigned LV W -measurable function. Then the expressions
Z Z
(2.44)
(
f (v, w) dW (w)) dV (v)
V

(2.45)

(
W

f (v, w) dW (w)) dV (v)

V

and
Z
(2.46)

f (v, w) dV W (v, w)
V W

are well-defined (thus all integrands are almost everywhere well-defined and
appropriately measurable) and equal to each other.
This result is sometines referred to as the Keisler-Fubini theorem.
Proof. By the monotone convergence theorem it suffices to verify this when
f is a simple function; by linearity we may then take f to be an indicator
function f = 1E . Using Theorem 2.10.2 and an approximation argument
(and many further applications of monotone convergence) we may assume
without loss of generality that E is an internal set. We then have
Z
|E|
f (v, w) dV W (v, w) = st(
)
|V ||W |
V W

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

290

2. Related articles

and for every v V , we have



Z
|Ev |
f (v, w) dW (w) = st
,
|W |
W
where Ev is the internal set
Ev := {w W : (v, w) E}.
Let n be a standard natural number, then we can partition V into the
internal sets V = V1 Vn , where


i1
|Ev |
i
Vi := v V :
<

.
n
|W |
n
On each Vi , we have
Z
(2.47)
W

i
f (v, w) dW (w) = + O
n

 
1
n

and
(2.48)

|Ev |
i
= +O
|W |
n

 
1
.
n

From (2.47), we see that the upper and lower integrals of

are both of the form
 
n
X
i |Vi |
1
+O
.
n |V |
n

R
W

f (v, w) dW (w)

i=1

Meanwhile, using the nonstandard double counting identity

|E|
1 X |Ev |
=
|V |
|W |
|V ||W |
vV

(where all arithmetic operations are interpreted in the nonstandard sense,

of course) and (2.48), we see that
 
n
X
|E|
i |Vi |
1
=
+O
.
|V ||W |
n |V |
n
i=1
R
Thus we see that the upper and lower integrals of W f (v, w) dW (w) are


1
equal to |V|E|
||W R
| + O n for every standard n. Sending n to infinity, we
conclude that W f (v, w) dW (w) is measurable, and that


Z Z
|E|
(
f (v, w) dW (w)) dV (v) = st
|V ||W |
V
W
showing that (2.44) and (2.46) are well-defined and equal. A similar argument holds for (2.45) and (2.46), and the claim follows.


Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.10. Loeb measure and the triangle removal lemma

291

Remark 2.10.4. It is well known that the product of two Lebesgue measure spaces Rn , Rm , upon completion, becomes the Lebesgue measure space
on Rn+m . Drawing the analogy between Loeb measure and Lebesgue measure, it is then natural to ask whether (2.42) is simply the completion of
(2.43). But while (2.42) certainly contains the completion of (2.43), it is a
significantly larger space
Q in general (a fact
Q first observed by Doug Hoover).
Indeed, suppose V = n Vn , W = n Wn , where the cardinality of
Vn , Wn goes to infinity at some reasonable rate, e.g. |Vn |, |Wn | n for all
n. For each n, let En be a random subset of Vn Wn , with each element of
Vn Wn having an independent probability of 1/2 of lying in En . Then, as
is well known, the sequence of sets En is almost surely asymptotically regular
in the sense that almost surely, we have the bound
sup
An Vn ,Bn Wn

||En (An Bn )| 21 |An ||Bn ||

0
|Vn ||Wn |

as n . Let us condition on the event that this asymptotic regularity

Q
holds. Taking ultralimits, we conclude that the internal set E := np En
obeys the property
1
V W (E (A B)) = V W (A B)
2
for all internal A V, B W ; in particular, E has Loeb measure 1/2.
Using Theorem 2.10.2 we conclude that
1
V W (E F ) = V W (F )
2
for all LV LW -measurable F , which implies in particular that E cannot be
LV LW -measurable. (Indeed, 1E 21 is anti-measurable in the sense that
it is orthogonal to all functions in L2 (LV LW ); or equivalently, we have the
conditional expectation formula E(1E |LV LW ) = 12 almost everywhere.)
Intuitively, a LV LW -measurable set corresponds to a subset of V W
that is of almost bounded complexity, in that it can be approximated by
a bounded boolean combination of Cartesian products. In contrast, LV W measurable sets (such as the set E given above) have no bound on their
complexity.
2.10.2. The triangle removal lemma. Now we can prove the triangle
removal lemma, Lemma 2.10.1. We will deduce it from the following nonstandard (and tripartite) counterpart (a special case of a result established
in [Ta2007]):
Lemma 2.10.5 (Nonstandard triangle removal lemma). Let V be a nonstandard finite non-empty set, and let E12 , E23 , E31 V V be Loebmeasurable subsets of V V which are almost triangle-free in the sense

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

292

2. Related articles

that
Z
(2.49)
V V V

1E12 (u, v)1E23 (v, w)1E31 (w, u) dV V V (u, v, w) = 0.

Then for any standard > 0, there exists a internal subsets Fij V V for
ij = 12, 23, 31 with V V (Eij \Fij ) < , which are completely triangle-free
in the sense that
(2.50)

1F12 (u, v)1F23 (v, w)1F31 (w, u) = 0

for all u, v, w V .
Let us first see why Lemma 2.10.5 implies Lemma 2.10.1. We use the
usual compactness and contradiction argument. Suppose for contradiction
that Lemma 1.2 failed. Carefully negating the quantifiers, we can find a
(standard) > 0, and a sequence Gn = (Vn , En ) of graphs with t(Gn ) 1/n,
such that for each n, there does not exist a subgraph G0n = (Vn , En0 ) of n
with |En \En0 | |Vn |2 with t(G0n ) = 0. Clearly we may assume the Vn are
non-empty.
Q
We form
the
ultraproduct
G
=
(V,
E)
of
the
G
,
thus
V
=
n
n Vn
Q
and E = n En . By construction, E is a symmetric internal subset of
V V and we have
Z
1E (u, v)1E (v, w)1E (w, u) dV V V (u, v, w) = st lim t(Gn ) = 0.
n

V V V

Thus, by Lemma 2.10.5, we may find internal subsets F12 , F23 , F31 of V V
with V V (E\Fij ) < /6 (say) for ij = 12, 23, 31 such that (2.50) holds for
all u, v, w V . By letting E 0 be the intersection of all E with all the Fij
and their reflections, we see that E 0 is a symmetric internal subset of E with
V V (E\E 0 ) < , and we still have
1E 0 (u, v)1E 0 (v, w)1E 0 (w, u) = 0
for all u, v, w V . If we write E 0 = limn En0 for some sets En0 , then for n
sufficiently close to p, one has En0 a symmetric subset of En with
Vn Vn (En \En0 ) <
and
1En0 (u, v)1En0 (v, w)1En0 (w, u) = 0.
G0n

:= (Vn , En ), we thus have |En \En0 | |Vn |2 and t(G0n ) = 0,

If we then set
which contradicts the construction of Gn by taking n sufficiently large.
Now we prove Lemma 2.10.5. The idea (similar to that used to prove
the Furstenberg recurrence theorem, as discussed for instance in [Ta2009,
2.10]) is to first prove the lemma for very simple examples of sets Eij , and
then work ones way towards the general case. Readers who are familiar
with the traditional proof of the triangle removal lemma using the regularity

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.10. Loeb measure and the triangle removal lemma

293

lemma will see strong similarities between that argument and the one given
here (and, on some level, they are essentially the same argument).
To begin with, we suppose first that the Eij are all elementary sets, in
the sense that they are finite boolean combinations of products of internal
sets. (At the finitary level, this corresponds to graphs that are bounded
combinations of bipartite graphs.) This implies that there is an internal
partition V = V1 Vn of the vertex set V , such that each Eij is the
union of some of the Va Vb .
Let Fij be the union of all the Va Vb in Eij for which Va and Vb have
positive Loeb measure; then V V (Eij \Fij ) = 0. We claim that (2.50)
holds for all u, v, w V , which gives Theorem 2.10.5 in this case. Indeed, if
u Va , v Vb , w Vc were such that (2.50) failed, then E12 would contain
Va Vb , E23 would contain Vb Vc , and E31 would contain Vc Va . The
integrand in (2.49) is then equal to 1 on Va Vb Vc , which has Loeb
measure V (Va )V (Vb )V (Vc ) which is non-zero, contradicting (2.49). This
gives Theorem 2.10.5 in the elementary set case.
Next, we increase the level of generality by assuming that the Eij are
all LV LV -measurable. (The finitary equivalent of this is a little difficult
to pin down; roughly speaking, it is dealing with graphs that are not quite
bounded combinations of bounded graphs, but can be well approximated
by such bounded combinations; a good example is the half-graph, which is
a bipartite graph between two copies of {1, . . . , N }, which joins an edge
between the first copy of i and the second copy of j iff i < j.) Then each Eij
can be approximated to within an error of /3 in V V by elementary sets.
In particular, we can find a finite partition V = V1 Vn of V , and sets
0 that are unions of some of the V V , such that
0
Eij
a
V V (Eij Eij ) < /3.
b
0 such that V , V
Let Fij be the union of all the Va Vb contained in Eij
a
b
have positive Loeb measure, and such that
2
V V (Eij (Va Vb )) > V V (Va Vb ).
3
Then the Fij are internal subsets of V V , and V V (Eij \Fij ) < .

We now claim that the Fij obey (2.50) for all u, v, w, which gives Theorem 2.10.5 in this case. Indeed, if u Va , v Vb , w Vc were such that
(2.50) failed, then E12 occupies more than 23 of Va Vb , and thus
Z
2
1E12 (u, v) dV V V (u, v, w) > V V V (Va Vb Vc ).
3
Va Vb Vc
Similarly for 1E23 (v, w) and 1E31 (w, u). From the inclusion-exclusion formula, we conclude that
Z
1E12 (u, v)1E23 (v, w)1E31 (w, u) dV V V (u, v, w) > 0,
Va Vb Vc

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

294

2. Related articles

contradicting (2.49), and the claim follows.

Finally, we turn to the general case, when the Eij are merely LV V measurable. Here, we split
1Eij = fij + gij
where fij := E(1Eij |LV LV ) is the conditional expectation of 1Eij onto
LV LV , and gij := 1Eij fij is the remainder. We observe that each
gij (u, v) is orthogonal to any tensor product f (u)g(v) with f, g bounded
and LV -measurable. From this and Tonellis theorem for Loeb measure
(Theorem 2.10.3) we conclude that each of the gij make a zero contribution
to (2.49), and thus
Z
f12 (u, v)f23 (v, w)f31 (w, u) dV V V (u, v, w) = 0.
V V V
0 := {(u, v) V V : f (u, v) /2}, then the E 0 are L L Now let Eij
ij
V
V
ij
measurable, and we have
Z
0 (u, v)1E 0 (v, w)1E 0 (w, u) dV V V (u, v, w) = 0.
1E12
31
23
V V V

Also, we have
0
V V (Eij \Eij
)

Z
=
V V

1Eij (1 1Eij0 )

Z
=
V V

fij (1 1Eij0 )

/2.
Applying the already established cases of Theorem 2.10.5, we can find
0 \F ) < /2, and hence
internal sets Fij obeying (2.50) with V V (Eij
ij
V V (Eij \Fij ) < , and Theorem 2.10.5 follows.
Remark 2.10.6. The full hypergraph removal lemma can be proven using
similar techniques, but with a longer tower of generalisations than the three
cases given here; see [Ta2007] or [ElSz2006].

2.11. Two notes on Lie groups

In this section we record two small miscellaneous facts about Lie groups.
The first fact concerns the exponential map exp : g G from a Lie algebra g of a Lie group G to that group. (For this discuss we will only consider
finite-dimensional Lie groups and Lie algebras over the reals R.) A basic
fact in the subject is that the exponential map is locally a homeomorphism:
there is a neighbourhood of the origin in g that is mapped homeomorphically by the exponential map to a neighbourhood of the identity in G. This

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.11. Two notes on Lie groups

295

local homeomorphism property is the foundation of an important dictionary

between Lie groups and Lie algebras.
It is natural to ask whether the exponential map is globally a homeomorphism, and not just locally: in particular, whether the exponential map
remains both injective and surjective. For instance, this is the case for connected, simply connected, nilpotent Lie groups (as can be seen from the
Baker-Campbell-Hausdorff formula from Section 1.2.5).
The circle group S 1 , which has R as its Lie algebra, already shows that
global injectivity fails for any group that contains a circle subgroup, which
is a huge class of examples (including, for instance, the positive dimensional
compact Lie groups, or non-simply-connected Lie groups). Surjectivity also
obviously fails for disconnected groups, since the Lie algebra is necessarily
connected, and so the image under the exponential map must be connected
also. However, even for connected Lie groups, surjectivity can fail. To see
this, first observe that if the exponential map was surjective, then every
group element g G would have a square root (i.e. an element h G with
h2 = g), since exp(x) has exp(x/2) as a square root for any x g. However,
there exist elements in connected Lie groups without square roots. A simple
example is provided by the matrix


4
0
g=
0 1/4
in the connected Lie group SL2 (R). This matrix has eigenvalues 4, 1/4.
Thus, if h SL2 (R) is a square root of g, we see (from the Jordan normal
form) that it must have at least one eigenvalue in {2i, +2i}, and at least
one eigenvalue in {i/2, i/2}. On the other hand, as h has real coefficients,
the complex eigenvalues must come in conjugate pairs {a + bi, a bi}. Since
h can only have at most 2 eigenvalues, we obtain a contradiction.
However, there is an important case where surjectivity is recovered:
Proposition 2.11.1. If G is a compact connected Lie group, then the exponential map is surjective.
Proof. The idea here is to relate the exponential map in Lie theory to the
exponential map in Riemannian geometry. We first observe that every compact Lie group G can be given the structure of a Riemannian manifold with
a bi-invariant metric. This can be seen in one of two ways. Firstly, one can
put an arbitrary positive definite inner product on g and average it against
the adjoint action of G using Haar probability measure (which is available
since G is compact); this gives an ad-invariant positive-definite inner product on g that one can then translate by either left or right translation to
give a bi-invariant Riemannian structure on G. Alternatively, one can use
Theorem 1.4.14 to embed G in a unitary group U (n), at which point one can

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

296

2. Related articles

induce a bi-invariant metric on G from the one on the space Mn (C) Cn

of n n complex matrices.

As G is connected and compact and thus complete, we can apply the

Hopf-Rinow theorem and conclude that any two points are connected by
at least one geodesic, so that the Riemannian exponential map from g to
G formed by following geodesics from the origin is surjective. But one can
check that the Lie exponential map and Riemannian exponential map agree;
for instance, this can be seen by noting that the group structure naturally
defines a connection on the tangent bundle which is both torsion-free and
preserves the bi-invariant metric, and must therefore agree with the LeviCivita metric. (Alternatively, one can embed into a unitary group U (n) and
observe that G is totally geodesic inside U (n), because the geodesics in U (n)
can be described explicitly in terms of one-parameter subgroups.) The claim
follows.


The other basic fact we will present here concerns the algebraic nature
of Lie groups and Lie algebras. An important family of examples of Lie
groups are the algebraic groups - algebraic varieties with a group law given
by algebraic maps. Given that one can always automatically upgrade the
smooth structure on a Lie group to analytic structure (by using the BakerCampbell-Hausdorff formula), it is natural to ask whether one can upgrade
the structure further to an algebraic structure. Unfortunately, this is not always the case. A prototypical example of this is given by the one-parameter
subgroup

(2.51)

G :=

t 0
0 t

:tR

of GL2 (R). This is a Lie group for any exponent R, but if is irrational,
then the curve that G traces out is not an algebraic subset of GL2 (R) (as
one can see by playing around with Puiseux series).
This is not a true counterexample to the claim that every Lie group
can be given the structure of an algebraic group, because one can give G
a different algebraic structure than one inherited from the ambient group
GL2 (R). Indeed, G is clearly isomorphic to the additive group R, which is of
course an algebraic group. However, a modification of the above construction
works:
Proposition 2.11.2. There exists a Lie group G that cannot be given the
structure of an algebraic group.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

2.11. Two notes on Lie groups

297

Proof. We use an example from [TaYu2005]. Consider the subgroup

1 0 0

G := x t 0 : x, y R; t R+

y 0 t
of GL3 (R), with an irrational number. This is a three-dimensional (metabelian)
Lie group, whose Lie algebra g gl3 (R) is spanned by the elements

0 0 0
X := 0 1 0
0 0

0 0 0
Y := 1 0 0
0 0 0

0 0 0
Z := 0 0 0
0 0
with the Lie bracket given by
[Y, X] = Y ; [Z, X] = Z; [Y, Z] = 0.
As such, we see that if we use the basis X, Y, Z to identify g to R3 , then the
adjoint representation of G is the identity map.
If G is an algebraic group, it is easy to see that the adjoint representation
Ad : G GL(g) is also algebraic, and so Ad(G) = G is algebraic in GL(g).
Specialising to our specific example, in which adjoint representation is the
identity, we conclude that if G has any algebraic structure, then it must also
be an algebraic subgroup of GL3 (R); but G projects to the group (2.51)
which is not algebraic, a contradiction.

A slight modification of the same argument also shows that not every Lie
algebra is algebraic, in the sense that it is isomorphic to a Lie algebra of an
algebraic group. (However, there are important classes of Lie algebras that
are automatically algebraic, such as nilpotent or semisimple Lie algebras.)

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Bibliography

[Ar1950] K. Arrow, A Difficulty in the Concept of Social Welfare, Journal of Political

Economy 58 (1950), 328346.
[BoChLoSoVe2008] C. Borgs, J. Chayes, L. Lov
asz, V. S
os, K. Vesztergombi, Convergent
sequences of dense graphs. I. Subgraph frequencies, metric properties and testing, Adv.
Math. 219 (2008), no. 6, 18011851.
ements de mathematique. Fasc. XXXIV. Groupes et alg`ebres
[Bo1968] N. Bourbaki, El
de Lie. Chapitre IV: Groupes de Coxeter et syst`emes de Tits. Chapitre V: Groupes
engendres par des reflexions. Chapitre VI: syst`emes de racines. Actualites Scientifiques
et Industrielles, No. 1337 Hermann, Paris 1968.
[BrGrTa2011] E. Breuillard, B. Green, T. Tao, The structure of approximate groups, Publ.

Math. Inst. Hautes Etudes

Sci. 116 (2012), 115221.
[Ch2003] M. Chang, Factorization in generalized arithmetic progressions and applications
to the Erdos-Szemeredi sum-product problems, Geom. Funct. Anal. 13 (2003), no. 4,
720736.
[ChCo1996] J. Cheeger, T. Colding, Lower bounds on Ricci curvature and the almost
rigidity of warped products, Ann. of Math. 144 (1996), no. 1, 189237.
[ChGr1972] J. Cheeger, D. Gromoll, On the structure of complete manifolds of nonnegative
curvature, Ann. of Math. 96 (1972), 413443.
[CoMi1997] T. Colding, W. Minicozzi, Harmonic functions on manifolds, Ann. of Math.
146 (1997), no. 3, 725747.
[CrSi2010] E. Croot, O. Sisask, A probabilistic technique for finding almost-periods of
convolutions, Geom. Funct. Anal. 20 (2010), 13671396.
[de2006] T. de la Rue, 2-fold and 3-fold mixing: why 3-dot-type counterexamples are impossible in one dimension, Bull. Braz. Math. Soc. (N.S.) 37 (2006), no. 4, 503521.
[Dr1969] A. Dress, Newmans theorems on transformation groups, Topology 8 (1969),
203207.
[ElSz2006] G. Elek, B. Szegedy, A measure-theoretic approach to the theory of dense hypergraphs, Adv. Math. 231 (2012), 17311772.
[ElSh2011] G. Elekes, M. Sharir, Incidences in three dimensions and distinct distances in
the plane, Combin. Probab. Comput. 20 (2011), no. 4, 571608.

299

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

300

Bibliography

[FuHa1991] W. Fulton, J. Harris, Representation theory. A first course. Graduate Texts

in Mathematics, 129. Readings in Mathematics. Springer-Verlag, New York, 1991.
[Fr1973] G. A. Freiman, Foundations of a structural theory of set addition, American
Mathematical Society, Providence, R. I., 1973. Translated from the Russian, Translations of Mathematical Monographs, Vol 37.
[Fr1936] H. Freudenthal, Einige S
atze u
ber topologische Gruppen, Annals of Math. 37
(1936), 4656.
[Fu1977] H. Furstenberg, Ergodic behavior of diagonal measures and a theorem of Szemerdi
on arithmetic progressions, J. Analyse Math. 31 (1977), 204256.
[Gl1950] A. M. Gleason, Spaces with a compact Lie group of transformations, Proc. Amer.
Math. Soc. 1 (1950), 3543.
[Gl1952] A. M. Gleason, Groups without small subgroups, Ann. of Math. 56 (1952), 193
212.
[Gl1951] A. M. Gleason, The structure of locally compact groups, Duke Math. J. 18 (1951),
85104.
[Go1938] K. G
odel, Consistency of the axiom of choice and of the generalized continuumhypothesis with the axioms of set theory, Proc. Nat. Acad. Sci, 24 (1938), 556557.
[Go2009] I. Goldbring, Nonstandard methods in lie theory. Ph.D. Thesis, University of
Illinois at Urbana-Champaign, 2009.
[Go2010] I. Goldbring, Hilberts fifth problem for local groups, Ann. of Math. 172 (2010),
no. 2, 12691314.
[GoSa1964] E. D. Golod, I. R. Shafarevich, On the class field tower, Izv. Akad. Nauk
SSSR Ser. Mat. 28 (1964) 261272.
[GrRu2007] B. Green, I. Z. Ruzsa, Freimans theorem in an arbitrary abelian group, J.
Lond. Math. Soc. 75 (2007), 163175.
[Gr1984] R.I. Grigorchuk, Degrees of growth of finitely generated groups and the theory of
invariant means, Math. USSR Izv. 25 (1985), 259300; Russian original: Izv. Akad.
Nauk SSSR Sr. Mat. 48 (1984), 939985.
[Gr1981] M. Gromov, Groups of polynomial growth and expanding maps, Inst. Hautes

Etudes
Sci. Publ. Math. No. 53 (1981), 5373.
[GuKa2010] L. Guth, N. Katz, On the Erdos distinct distance problem in the plane,
preprint.
[He1983] J. Heintz, Definability and fast quantifier elimination in algebraically closed
fields, Theoret. Comput. Sci. 24 (1983), no. 3, 239277.
[HeRo1979] E. Hewitt, K. Ross, Abstract harmonic analysis. Vol. I. Structure of topological groups, integration theory, group representations. Second edition. Grundlehren der
Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences],
115. Springer-Verlag, Berlin-New York, 1979.
[Hi1990] J. Hirschfeld, The nonstandard treatment of Hilberts fifth problem, Trans. Amer.
Math. Soc. 321 (1990), no. 1, 379400.
[Ho1991] B. Host, Mixing of all orders and pairwise independent joinings of systems with
singular spectrum, Israel J. Math. 76 (1991), no. 3, 289298.
[Hr2012] E. Hrushovski, Stable group theory and approximate subgroups, J. Amer. Math.
Soc. 25 (2012), no. 1, 189243.
[IoRoRu2011] A. Iosevich, O. Roche-Newton, M. Rudnev, On an application of Guth-Katz
theorem, Math. Res. Lett. 18 (2011), no. 4, 691697.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Bibliography

301

[Jo1878] C. Jordan, Memoire sur les equations differentielles lineaires a

` integrale
algebrique, J. Reine Angew. Math. 84 (1878), 89215.
[Ka1984] S. Kalikow, Twofold mixing implies threefold mixing for rank one transformations, Ergodic Theory Dynam. Systems 4 (1984), no. 2, 237259.
[KaPeTu2010] V. Kapovitch, A. Petrunin, W. Tuschmann, Nilpotency, almost nonnegative
curvature, and the gradient flow on Alexandrov spaces, Ann. of Math. 171 (2010), no.
1, 343373.
[KaWi2011] V. Kapovitch, B. Wilking, Structure of fundamental groups of manifolds with
Ricci curvature bounded below, preprint.
[KiSo1972] A. Kirman, D. Sondermann, Arrows theorem, many agents, and invisible
dictators, Journal of Economic Theory 5 (1972), 267277.
[Kl2010] B. Kleiner, A new proof of Gromovs theorem on groups of polynomial growth, J.
Amer. Math. Soc., 23 (2010), 815829.
[Ku1998] W. Kulpa, Poincare and domain invariance theorem, Acta Univ. Carolin. Math.
Phys. 39 (1998), no. 1-2, 127136.
[Ku1950] M. Kuranishi, On Euclidean local groups satisfying certain conditions, Proc.
Amer. Math. Soc. 1 (1950), 372380.
[Le1978] F. Ledrappier, Un champ markovien peut etre dentropie nulle et melangeant, C.
R. Acad. Sci. Paris Ser. A-B 287 (1978), no. 7, A561-A563.
[Le1997] J. S. Lee, Totally disconnected groups, p-adic groups and the Hilbert-Smith conjecture, Commun. Korean Math. Soc. 12 (1997), no. 3, 691699.
[LeMa2005] G. Leonardi, S. Masnou, On the isoperimetric problem in the Heisenberg group
H n , Ann. Mat. Pura Appl. (4) 184 (2005), no. 4, 533553.
[Lo1975] P. Loeb, Conversion from nonstandard to standard measure spaces and applications in probability theory, Trans. Amer. Math. Soc. 211 (1975), 113122.
[Lo1955] J. Los, Quelques remarques, theor`emes et probl`emes sur les classes definissables
dalgebres, in: Mathematical Interpretation of Formal Systems, North-Holland, Amsterdam, 1955, 98113.
[LoSz2006] L. Lov
asz, B. Szegedy, Limits of dense graph sequences, J. Combin. Theory
Ser. B 96 (2006), no. 6, 933957.
[Ma1941] A. Malcev, Sur les groupes topologiques locaux et complets, C. R. (Doklady)
Acad. Sci. URSS (N.S.) 32, (1941). 606608.
[Mi1956] J. Milnor, On manifolds homeomorphic to the 7-sphere, Annals of Mathematics
64 (1956), 399405.
[MoZi1952] D. Montgomery, L. Zippin, Small subgroups of finite-dimensional groups, Ann.
of Math. 56 (1952), 213241.
[MoZi1974] D. Montgomery, L. Zippin, Topological transformation groups. Reprint of the
1955 original. Robert E. Krieger Publishing Co., Huntington, N.Y., 1974. xi+289 pp.
[Mo2003] R. Monti, Brunn-Minkowski and isoperimetric inequality in the Heisenberg
group, Ann. Acad. Sci. Fenn. Math. 28 (2003), no. 1, 99109.
[My1941] S. B. Myers, Riemannian manifolds with positive mean curvature, Duke Mathematical Journal 8 (1941), 401404.
[Ne1931] M. H. A. Newman, A theorem on periodic transformations of spaces, Quart. J.
Math 2 (1931), 19.
[Ol1996] P. Olver, Non-associative local Lie groups, J. Lie Theory 6 (1996), no. 1, 2351.
[Pa2013] J. Pardon, The HilbertSmith conjecture for three-manifolds, J. Amer. Math.
Soc. 26 (2013), no. 3, 879899.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

302

Bibliography

[Pe2006] P. Petersen, Riemannian geometry. Second edition. Graduate Texts in Mathematics, 171. Springer, New York, 2006.

[ReSc97] D. Repovs, E. Scepin,

A proof of the Hilbert-Smith conjecture for actions by
Lipschitz maps, Math. Ann. 308 (1997), no. 2, 361364.
[Ro1966] A. Robinson, Non-standard analysis. North-Holland Publishing Co., Amsterdam
1966.
[Ro1947] V. A. Rohlin, On endomorphisms of compact commutative groups, Izvestiya
Akad. Nauk SSSR. Ser. Mat. 13, (1949). 329340.
[Ro1953] K.F. Roth, On certain sets of integers, J. London Math. Soc. 28 (1953), 245-252.
[Ru1962] W. Rudin, Fourier analysis on groups. Interscience Tracts in Pure and Applied
Mathematics, No. 12 Interscience Publishers (a division of John Wiley and Sons),
New York-London 1962 ix+285 pp.
[Ru1978] I. Ruzsa, E. Szemeredi, Triple systems with no six points carrying three triangles,
Combinatorics (Proc. Fifth Hungarian Colloq., Keszthely, 1976), Vol. II, pp. 939-945,
Colloq. Math. Soc. Jnos Bolyai, 18, North-Holland, Amsterdam-New York, 1978.
[Sa2009] T. Sanders, On a non-abelian Balog-Szemeredi-type lemma, J. Aust. Math. Soc.
89 (2010), no. 1, 127132.
[Sa2010] T. Sanders, On the Bogolyubov-Ruzsa lemma, Anal. PDE 5 (2012), 627655.
[Sc1995] J. Schmid, On the affine Bezout inequality, Manuscripta Math. 88 (1995), no. 2,
225232.
[Sc1989] K. Schmidt-G
ottsch, Polynomial bounds in polynomial rings over fields, J. Algebra 125 (1989), no. 1, 164180.
[Se1931] H. Seifert, Konstruction drei dimensionaler geschlossener Raume, Berichte
Sachs. Akad. Leipzig, Math.-Phys. Kl. 83 (1931) 2666.
[Se1964] J.-P. Serre, Lie algebras and Lie groups. 1964 lectures given at Harvard University. Corrected fifth printing of the second (1992) edition. Lecture Notes in Mathematics, 1500. Springer-Verlag, Berlin, 2006.
[ShTa2010] Y. Shalom, T. Tao, A finitary version of Gromovs polynomial growth theorem,
Geom. Funct. Anal. 20 (2010), no. 6, 15021547.
[Sm1941] P. A. Smith, Transformations of finite period. III. Newmans theorem, Ann. of
Math. 42 (1941), 446458.
[SoTa2012] J. Solymosi, T. Tao, An Incidence Theorem in Higher Dimensions, Discrete
Comput. Geom. 48 (2012), no. 2, 255280.
[Sz1975] E. Szemeredi, On sets of integers containing no k elements in arithmetic progression, ollection of articles in memory of Jurii Vladimirovic Linnik. Acta Arith. 27
(1975), 199-245.
[Sz1978] E. Szemeredi, Regular partitions of graphs, in Problemes Combinatoires et
Theorie des Graphes, Proc. Colloque Inter. CNRS, (Bermond, Fournier, Las Vergnas,
Sotteau, eds.), CNRS Paris, 1978, 399401.
[Ta2007] T. Tao, A correspondence principle between (hyper)graph theory and probability
theory, and the (hyper)graph removal lemma, J. Anal. Math. 103 (2007), 145.
[Ta2008] T. Tao, Structure and Randomness: pages from year one of a mathematical blog,
American Mathematical Society, Providence RI, 2008.
[Ta2008b] T. Tao, Product set estimates for non-commutative groups, Combinatorica 28
(2008), 547594.
[Ta2009] T. Tao, Poincares Legacies: pages from year two of a mathematical blog, Vol.
I, American Mathematical Society, Providence RI, 2009.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Bibliography

303

[Ta2009b] T. Tao, Poincares Legacies: pages from year two of a mathematical blog, Vol.
II, American Mathematical Society, Providence RI, 2009.
[Ta2010] T. Tao, An epsilon of room, Vol. I, American Mathematical Society, Providence
RI, 2010.
[Ta2010b] T. Tao, An epsilon of room, Vol. II, American Mathematical Society, Providence RI, 2010.
[Ta2011] T. Tao, An introduction to measure theory, American Mathematical Society,
Providence RI, 2011.
[Ta2011b] T. Tao, Higher order Fourier analysis, American Mathematical Society, Providence RI, 2011.
[Ta2011c] T. Tao, Compactness and contradiction, in preparation.
[TaVu2006] T. Tao, V. Vu, Additive combinatorics. Cambridge Studies in Advanced
Mathematics, 105. Cambridge University Press, Cambridge, 2006.
[TaYu2005] P. Tauvel, R. Yu, Lie algebras and algebraic groups. Springer Monographs in
Mathematics. Springer-Verlag, Berlin, 2005.
[vDa1936] D. van Dantzig, Zur topologischen Algebra. III. Brouwersche und Cantorsche
Gruppen, Compositio Math. 3 (1936), 408426.
[vdDrGo2010] L. van den Dries, I. Goldbring, Globalizing locally compact local groups, J.
Lie Theory 20 (2010), no. 3, 519524.
[vKa1833] E. R. van Kampen, On the connection between the fundamental groups of some
related spaces, Amer. J. Math., 55 (1933), 261267.
[vdW1927] B.L. van der Waerden, Beweis einer Baudetschen Vermutung, Nieuw. Arch.
Wisk. 15 (1927), 212216.
[vdDrWi1984] L. van den Dries, A. J. Wilkie, Gromovs theorem on groups of polynomial
growth and elementary logic, J. Algebra, 89 (1984), 349374.
[vEsKo1964] W. T. van Est, Th. J.Korthagen, Non-enlargible Lie algebras, Nederl. Akad.
Wetensch. Proc. Ser. A 67, Indag. Math. 26 (1964), 1531.
[We1973] B. A. F. Wehrfritz, Infinite linear groups. An account of the group-theoretic
properties of infinite groups of matrices. Ergebnisse der Matematik und ihrer Grenzgebiete, Band 76. Springer-Verlag, New York-Heidelberg, 1973.
[Ya1953] H. Yamabe, On the conjecture of Iwasawa and Gleason, Ann. of Math. 58,
(1953), 4854.
[Ya1953b] H. Yamabe, A generalization of a theorem of Gleason, Ann. of Math. 58 (1953),
351365.
[Ze1990] E. I. Zelmanov, Solution of the restricted Burnside problem for groups of odd
exponent, Izv. Akad. Nauk SSSR Ser. Mat. 54 (1990), no. 1, 4259.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Index

C 1,1 local group, 32

p-adic norm, 10
*-homomorphism, 133
Loss theorem, 141
A group is virtually nilpotent if it
contains a finite index subgroup
which is nilpotent., 20
abstract Lie algebra, 228
additive combinatorics, 14
adjoint representation, 41, 78
Ados theorem, 229
almost period, 165
approximate group, 148, 196
arrows theorem, 134
associated neighbourhood, 110
asymptotic notation, xi, 146
baby Peter-Weyl theorem, 84
Baker-Campbell-Hausdorff formula, 40
Baker-Campbell-Hausdorff-Dynkin
formula, 44
Birkoff-Kakutani theorem, 92
Borel measure, 70
Brouwer fixed point theorem, 120
Burnsides theorem, 215
cancellative local group, 109
Cartans theorem, 49
Cayley graph, 19
central element, 5
centraliser, 5
chain rule, 28
character, 279

characteristic subgroup, 220

cheap structure theorem, 194
class function, 279
Classification of one-parameter
subgroups, 52
clopen set, 116
closed graph theorem, 51
cocompact subgroup, 87
cocompact topology, 56
cocycle, 257
cocycle equation, 257
cohomologous, 258
commutator, 4, 219
compact operator, 82
compact-open topology, 60
compactness of integral operators, 82
compactness theorem, 144
complete measure space, 283
concretely nilpotent Lie algebra, 233
conjugate, 219
connected component, 116
connected space, 116
continuous action, 116
convolution, 82
coset nilprogression, 152
coset progression, 16, 151
countable saturation, 138
degree, 267
derivation, 235
derivations and vector fields, 29
derivative map, 27
directional derivative, 28

305

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

306

edge density, 285

elementary set, 293
equivariant map, 273
escape norm, 184
existence and uniqueness of Haar
measure, 71
exponential coordinates, 37
exponential map, 64
extension problem, 254
faithful action, 116
finite trapping, 105
formal commutator word, 223
formal predicate, 144
Fourier analysis on compact abelian
groups, 87
Fourier inversion formula, 277
Freimans theorem, 16, 151
Fubini-Tonelli theorem, 284
gauge transformation, 258
general linear group, 8
generalised arithmetic progression, 15,
151
Gleason lemma, 186
Gleason metric, 50
Gleason-Yamabe theorem for abelian
groups, 89
Gleason-Yamabe theorem for compact
groups, 85
good model, 170, 196
Gromovs theorem, 20, 206
Gromov-Hausdorff distance, 157
group cohomology, 258
group germ, 244
group of polynomial growth, 19
Haar measure, 70, 80
Hadamard lemma, 42
half-open topology, 56
Hall-Witt identity, 220
hard analysis, 126
Hausdorff distance, 107, 155
height, 281
Heisenberg group, 8
Helfgott-Lindenstrauss conjecture, 200
Hilberts fifth problem, 9, 122
Hilbert-Smith conjecture, 265
homogeneous space, 80
homomorphism, 26
Hrushovski Lie model theorem, 197

Index

Hrushovskis Lie model theorem, 177,

180
Hrushovskis structure theorem, 179
hyperreal, 136
indicator function, xi
infinitesimal, 146
infinitesimal part, 146
inner regularity, 70
internal function, 139
internal set, 135
invariance of dimension, 115
invariance of domain, 115, 120
inverse limit, 86
irreducible representation, 272
isotypic component, 274
Jacobi identity, 228
Jordans theorem, 4, 214
Jordan-Schur theorem, 214
Keisler-Fubini theorem, 289
lamplighter group, 176
left-invariant vector field, 30
Leibniz rule, 29
Levis theorem, 237
Lie algebra, 30
Lie algebra of a Lie group, 30
Lie bracket, 29
Lie group, 7
Lie groups are analytic, 46
Lies first theorem, 36
Lies second theorem, 46
Lies third theorem, 32, 47, 247
lifting lemma, 198
local finiteness, 70
local Gleason metric, 112
local Gleason-Yamabe theorem, 112
local group, 23
local Haar measure, 80
local homomorphism, 26
local Lie group, 23
Local Lie implies Lie, 39
local section, 261
local topological group, 23
Loeb measure, 288
lower central series, 221
Margulis-type lemma, 211
matrix exponential, 2
measure space, 283
monad, 245

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

Index

Newmans first theorem, 266

Newmans second theorem, 268
nilpotent Lie algebra, 230
nilprogression, 152
nilradical, 235
no small subgroups, 97
noncommutative progression, 17, 152
nonstandard finite set, 145
nonstandard function, 139
nonstandard object, 135
nonstandard space, 135
nonstandard universe, 136
normal Sanders lemma, 167
normal sub-local group, 110
normalising neighbourhood, 110
notation, x
NSS, 97
NSS approximate group, 191
one-parameter subgroup, 38, 52
open mapping theorem, 51
orthogonal group, 8
outer regularity, 70
overspill principle, 142
periodic group, 214
Peter-Weyl theorem, 272
Plancherel identity, 277
Poincare-Birkhoff-Witt theorem, 228
pointed Gromov-Hausdorff convergence,
160
pointed metric space, 160
probability space, 283
profinite group, 119, 123
projective limit, 86
pseudometric, 92
pushforward, 29
quotient space, 111
quotiented commutator map, 222
radially homogeneous C 1,1 local group,
36
radical, 234
Radon measure, 70
regular representation, 67, 81
restricted Burnside problem, 216
restriction of a local group, 24
Riesz lemma, 58
Riesz representation theorem, 72
Ruzsa covering lemma, 168
Ruzsa triangle inequality, 203

307

Sanders lemma, 153, 196

Schurs lemma, 273
Schurs theorem, 214
smooth action, 80
smooth atlas, 6
smooth manifold, 6
smooth structure, 6
soft analysis, 126
solenoid, 10
special orthogonal group, 8
special unitary group, 8
spectral theorem, 83
stabiliser subgroup, 80
standard function, 139
standard object, 135
standard part, 146
standard universe, 135
Stone-Cech compactification, 133
strictly nonstandard, 136
sub-local group, 110
subgroup trapping, 103, 113
submanifold, 48
syndetic set, 88
tangent bundle, 27
tangent space, 27
tangent vector, 27
Tonelli theorem, 289
topological group, 9
topological manifold, 6
topological vector space, 55
torsion group, 149
totally disconnected space, 116
transitive action, 80, 116
triangle removal lemma, 285, 292
ultra approximate group, 149, 196
ultrafilter, 132
ultralimit, 135
ultrapower, 136
ultraproduct, 135
uniform total boundedness, 158
Uniqueness of Lie structure, 39
unitary group, 8
universal cover, 208
universal enveloping algebra, 228
van Dantzigs theorem, 117
vector field, 28
von Neumanns theorem, 49
weak Gleason metric, 95
Weil-Heisenberg group, 8

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.

308

Index

well-defined word, 26
well-formed formula, 143
word metric, 19

Author's preliminary version made available with permission of the publisher, the American Mathematical Society.